diff --git a/ChangeLog b/ChangeLog
index bb3fc62..fc95786 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -2,6 +2,14 @@
 magma R package changes
 --------------------------------------------------------------------------------
 
+2014-10-06  Serguei Sokol <sokol@insa-toulouse.fr>
+
+   * magma_1.5.0-1.tar.gz: Package upgrade release.
+
+   * src/*.c:  Updated to magma 1.5.0 and cublas 6.0 (4.0+ is required) libraries
+   
+   * fixed few bugs in magQRSolve()
+
 2013-04-02  Brian Smith <brian-j-smith@uiowa.edu>
 
    * magma_1.3.0-1.tar.gz: Package upgrade release.
diff --git a/DESCRIPTION b/DESCRIPTION
old mode 100644
new mode 100755
index f5e1a1e..1d9ec5c
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -1,19 +1,24 @@
 Package: magma
 Title: Matrix Algebra on GPU and Multicore Architectures
-Version: 1.3.0-2
-Date: 2013-04-03
-Author: Brian J Smith <brian-j-smith@uiowa.edu>
-Maintainer: Brian J Smith <brian-j-smith@uiowa.edu>
+Version: 1.5.0-1
+Date: 2014-10-07
+Authors@R: c(person("Brian", "Smith", middle="J.", role = c("aut", "cre"),
+		email = "brian-j-smith@uiowa.edu"),
+	person("Serguei", "Sokol", role = c("com", "ctb"),
+		email="sokol@insa-toulouse.fr"))
 Depends: R (>= 2.11.0), methods
-SystemRequirements: MAGMA shared libraries (== release 1.3.0), LAPACK shared
-   library, NVIDIA CUDA Toolkit (>= release 4.0)
-Description: Magma matrix classes and methods for parallel processing of matrix
-   algebra operations.  Operations are performed with algorithms developed by
-   the MAGMA research project.  MAGMA aims to achieve the fastest possible
-   linear algebra libraries on hybrid multicore CPU and GPU architectures by
-   exploiting their massive parallelism and minimizing communication latencies.
+SystemRequirements: MAGMA software (== release 1.5.0), LAPACK
+        shared library, NVIDIA CUDA Toolkit (>= release 4.0)
+Description: Magma matrix classes and methods for parallel processing
+        of matrix algebra operations.  Operations are performed with
+        algorithms developed by the MAGMA research project.  MAGMA aims
+        to achieve the fastest possible linear algebra libraries on
+        hybrid multicore CPU and GPU architectures by exploiting their
+        massive parallelism and minimizing communication latencies.
 License: GPL-3
 URL: http://icl.cs.utk.edu/magma
 OS_type: unix
-
-
+Packaged: 
+NeedsCompilation: yes
+Repository: CRAN
+Date/Publication: 
diff --git a/INSTALL b/INSTALL
index f7bbfb0..923106c 100644
--- a/INSTALL
+++ b/INSTALL
@@ -1,8 +1,10 @@
 --------------------------------------------------------------------------------
 --------------------------------------------------------------------------------
-magma R Package Version 1.3.0
-Installation and Release Notes (4/2/2013)
+magma R Package Version 1.5.0
+Installation and Release Notes (2014-10-07)
+Package Author: Brian J. Smith <brian-j-smith@uiowa.edu>
 Package Maintainer: Brian J. Smith <brian-j-smith@uiowa.edu>
+Package Contributor: Serguei Sokol <sokol@insa-toulouse.fr>
 --------------------------------------------------------------------------------
 --------------------------------------------------------------------------------
 
@@ -60,14 +62,14 @@ I. (b) Hardware requirements
 I. (c) Supported operating systems
 --------------------------------------------------------------------------------
 
-* Development and testing of the package is performed on 64-bit Ubuntu Linux
-  10.04 with.  Other 64-bit Linux distributions should work as well.
-  Development is conducted with release 4.3 of the GCC tools; release 4.4 is
-  not supported and may cause errors.
+* Development and testing of the package is performed on 64-bit Mageia Linux
+  4.0.  Other 64-bit Linux distributions should work as well.
+  Development is conducted with release 4.8.2 of the GCC tools and 6.0
+  of the cuda tools.
 
 * At this time, the package has not been tried on OS X, although input from any
-  attempts to do so are welcome.  Windows is not supported currently due to a
-  lack of shared MAGMA libraries for that platform.
+  attempts to do so are welcome. Windows should be supported too as the
+  library magma is reported to copile on this platform.
 
 
 --------------------------------------------------------------------------------
@@ -137,11 +139,7 @@ Next, the MAGMA software (http://icl.cs.utk.edu/magma) must be installed.  To do
 so,
 
 * Extract the source code to a local directory, and follow the instructions
-  supplied with the code to make shared libraries 'libmagmablas.so' and
-  'libmagma.so'.  Additional help on making the shared magma libraries can be
-  found at http://www.hiplar.org/hiplar-M-installation.html#MAGMA.  After making
-  the shared libraries, refer to the previous section for instructions on
-  installing the libraries.
+  supplied with the code to make shared library 'libmagma.so'.
 
 
 --------------------------------------------------------------------------------
@@ -165,7 +163,7 @@ where <option> are optional configuration arguments that can be supplied to
 change default files and locations, and include
 
   --with-r-home=DIR       full path to R installation []
-  --with-magma-lib=DIR    full path to MAGMA shared libraries []
+  --with-magma-home=DIR   full path to MAGMA installation directory []
   --with-lapack=LIBS      LAPACK shared library [-llapack]
   --with-cuda-home=DIR    full path to CUDA installation [/usr/local/cuda]
 
@@ -182,24 +180,25 @@ the particular LAPACK implementation being used.
 
 Example:
 
-The following illustrate installation of magma package release 0.2.2-1 on a
-system in which CUDA is installed in '/opt/cuda', MAGMA shared libraries in
-'/opt/magma/lib', and with a ldconfig-registered LAPACK library named
-'liblapack.so' (the default for --with-lapack):
+The following illustrate installation of magma package release 1.5.0-1 on a
+system in which CUDA is installed in '/usr/local/cuda' (default), MAGMA home directory in
+'/usr/local/src/magma-1.5.0', and LAPACK library compiled by ATLAS project and named
+/usr/local/atlas/lib64/libsatlas.so
 
   Terminal command prompt
   -----------------------
-  R CMD INSTALL \
-    --configure-args="
-       --with-cuda-home=/opt/cuda \
-       --with-magma-lib=/opt/magma/lib" \
-    magma_0.2.2-1.tar.gz
+  R CMD INSTALL --configure-args="\
+       --with-lapack=/usr/local/atlas/lib64/libsatlas.so \
+       --with-magma-home=/usr/local/src/magma-1.5.0 \
+       --with-r-home=/usr/local/src/R-3.1.1" \
+    magma_1.5.0-1.tar.gz
 
   R command prompt
   ----------------
   install.packages("magma",
-    configure.args = "--with-cuda-home=/opt/cuda
-      --with-magma-lib=/opt/magma/lib"
+    configure.args = "--with-lapack=/usr/local/atlas/lib64/libsatlas.so \
+       --with-magma-home=/usr/local/src/magma-1.5.0 \
+       --with-r-home=/usr/local/src/R-3.1.1"
   )
 
 
@@ -207,9 +206,6 @@ system in which CUDA is installed in '/opt/cuda', MAGMA shared libraries in
 III. Known Issues
 --------------------------------------------------------------------------------
 
-* Performance (speed) of the QR solver may be suboptimal due to workarounds for
-  a bug in version 1.3.0 of the magma_dgeqrs_gpu routine.
-
 
 --------------------------------------------------------------------------------
 IV. Troubleshooting
@@ -248,11 +244,11 @@ cannot find -llapack
 - If an alternative, installed LAPACK library is desired, then specify it using
   the configuration option '--with-lapack'.  See Section II(d).
 
-cannot find -lmagmablas or -lmagma
-- Check that the shared MAGMA libraries 'libmagmablas.so' and 'libmagma.so' are
+cannot find -lmagma
+- Check that the shared MAGMA sources are
   installed on your system.  Install them, if not.  Refer to Section II(c).
 - If they are installed, try supplying their directory path using the
-  configuration option '--with-magma-lib'.  See Section II(d).
+  configuration option '--with-magma-home'.  See Section II(d).
 
 
 R Run-time Errors
diff --git a/MD5 b/MD5
new file mode 100644
index 0000000..093b6db
--- /dev/null
+++ b/MD5
@@ -0,0 +1,67 @@
+8f0e2cd40e05189ec81232da84bd6e1a *COPYING
+73ab637c95c8ac53586ad63383f31cdd *ChangeLog
+41354a359a950d8b01fd7015da692c9c *DESCRIPTION
+32407b0000dca1262d5ae79eb6b2efe3 *INSTALL
+2280c65898c35e3cc4406de62d5591cc *NAMESPACE
+7ae7c3651a63677b5d19f0f79b5c2ac4 *R/AllGenerics.r
+0fd7bd1e34a61517905b0ac7648edadb *R/Classes.r
+83db8b129155879edd5394abe0c5dbcf *R/magAriths.r
+26c0a2607d6070cf27c18dd537c87e36 *R/magFactors.r
+b3542f41432a2b897de3292401665399 *R/magSolvers.r
+1d81b96b1ee67a180921872426beaafd *R/magUtils.r
+148260b02ac4fb6f5282dac388f04df5 *cleanup
+3dec06adf97f1edf973b0172bc32db86 *configure
+378499169b30549fea419e41b0a90fb5 *configure.ac
+dc328f38f8241e9b06baa73dc0a4cffc *inst/CITATION
+653f8fb438453be3e2abf132cf2d3db3 *man/backsolve.Rd
+62299a88fde6682670012392fc09cba9 *man/gpu.Rd
+5d5d438c4b47d8e5362af1474e4b0ad4 *man/lu.Rd
+2ab45972f6f7a5021dd6f7c4d1cd9fbf *man/magma-class.Rd
+fa1c83d80ce326f1e03fe1762dbf8650 *man/magma.Rd
+522aa9e97a32ce0c110aa075496b2fe5 *man/magmaLU-class.Rd
+e56bf7401e61d779303af3354f85054a *man/magmaQR-class.Rd
+9d9736f21d2a4ad3616165e3ad08a9f0 *src/Makefile.in
+7d5fb6df5d668708ae0adc3063073920 *src/include/cblas.h
+b7a6db7b7113f0fada181c6551ed7628 *src/include/gnu/stubs-32.h
+583f7b4ad3da81a6abd04eb0bedda2da *src/include/magma.h
+634e9149c1a9a561e45d8a1bbe4899e5 *src/include/magma_bulge.h
+3acd63edfe6c7b07dd2e7f0f63ea3f78 *src/include/magma_c.h
+b0a28f624ee0a44c93d54cc60df8f405 *src/include/magma_cbulge.h
+77c9ffe11af54fbb8cedbb1cd7b50a94 *src/include/magma_cbulgeinc.h
+5e03b80ca7395d39c9d341d2c4d8ab7e *src/include/magma_clapack.h
+926bbe7507c0d675c59fbf7acc855c14 *src/include/magma_d.h
+f58ec1510fa2fc29a915813b2ab885c2 *src/include/magma_dbulge.h
+55a551bb76ec432bd1a5f190701ba898 *src/include/magma_dbulgeinc.h
+4196ba980162736f10641930cf8ddca2 *src/include/magma_dlapack.h
+ed7021241404fb0fe6df2b3d99cddc86 *src/include/magma_ds.h
+925fd142dad8503fd1ec4d36c655208d *src/include/magma_lapack.h
+d587d1f3cb23db4c72dbd0a9e7307a93 *src/include/magma_s.h
+3224f475fc75a0248030ce6993ebaf20 *src/include/magma_sbulge.h
+d915a843903b8d713d46b4966cd76a95 *src/include/magma_sbulgeinc.h
+3d5560d8ee310d7328a2b9ab3014d6d6 *src/include/magma_slapack.h
+65b1d335a4f22e641ab44dc6184f7322 *src/include/magma_z.h
+07b558b924b08b26568e5685179420ed *src/include/magma_zbulge.h
+a1b4f0be8c8257ef66e53c38b4b98685 *src/include/magma_zbulgeinc.h
+8128878591d7e4058ff7818b5817e527 *src/include/magma_zc.h
+47c2b0df91b0e7e49c06f52e21f4d81f *src/include/magma_zlapack.h
+7fc5789de32e359e8711b0b0dd9b6b12 *src/include/magmablas.h
+35b8347b6671eb5ee9a31cfba958e4c7 *src/include/magmablas_c.h
+dfda4612b57e30da40515d43385ab51b *src/include/magmablas_d.h
+f9ac54a3d30ef1265bf38b5bde74e7ab *src/include/magmablas_ds.h
+97b832fef3ba6f487ab928965a120c55 *src/include/magmablas_s.h
+ada11fe61165197b33ecd2a110baca08 *src/include/magmablas_z.h
+456c856a727c5403d14b6b3ec6f564b3 *src/include/magmablas_zc.h
+94eb0f34cf187bcebab8f7a92887bf2e *src/magAriths.c
+579bdb18c98658bd8ec131dc9e88b0e8 *src/magAriths.h
+1a2287e375f765e59db9643a9a0f7071 *src/magFactors.c
+5971e4b06c2ad948878cca54324793c7 *src/magFactors.h
+1133c23184c3bad1701e7d19c4465dc5 *src/magSolvers.c
+47d396d5bdd3353ae0f7bc847bc3a980 *src/magSolvers.h
+4b6b4b4be94df1a3c1acb0c0fe828e0e *src/magUtils.c
+02a2617f6d7a1144a7b80da991361956 *src/magUtils.h
+eb1af5b1366c148007a2aada94f6f905 *tests/testing.r
+cc4680349521e99e797af60e980642c6 *tools/config.guess
+6adca7f3405b57aeecc1df1704f27ba1 *tools/config.sub
+5cc51dc6eb45a90231a4e31efbc3f220 *tools/install-sh
+b65c3f66a47925985fbf5ffecb22c48c *tools/mdate-sh
+c7645c248f924f07869059b76300fea9 *tools/missing
diff --git a/NAMESPACE b/NAMESPACE
index b81a7de..496d36c 100644
--- a/NAMESPACE
+++ b/NAMESPACE
@@ -1,31 +1,30 @@
-useDynLib(magma)
-
-export(
-   "forwardsolve",
-   "magma"
-)
-
-exportClasses(
-   "magma",
-   "magmaLU",
-   "magmaQR"
-)
-
-exportMethods(
-   "Arith",
-   "%*%",
-
-   "backsolve",
-   "chol",
-   "chol2inv",
-   "crossprod",
-   "gpu",
-   "gpu<-",
-   "lu",
-   "qr",
-   "show",
-   "solve",
-   "tcrossprod"
-)
-
-  
+useDynLib(magma)
+
+export(
+   "forwardsolve",
+   "magma"
+)
+
+exportClasses(
+   "magma",
+   "magmaLU",
+   "magmaQR"
+)
+
+exportMethods(
+   "Arith",
+   "%*%",
+
+   "backsolve",
+   "chol",
+   "chol2inv",
+   "crossprod",
+   "gpu",
+   "gpu<-",
+   "lu",
+   "qr",
+   "show",
+   "solve",
+   "tcrossprod"
+)
+
diff --git a/R/magAriths.r b/R/magAriths.r
index bb0738e..8f98d44 100644
--- a/R/magAriths.r
+++ b/R/magAriths.r
@@ -50,26 +50,28 @@ setMethod("Arith", signature(e1 = "numeric", e2 = "magma"),
 #################### Multiplication Operators ####################
 
 setMethod("%*%", signature(x = "magma", y = "magma"),
-   function(x, y) .Call("magMultmm", x, FALSE, y, FALSE)
+   function(x, y) { .Call("magMultmm", x, FALSE, y, FALSE) }
 )
 
 setMethod("%*%", signature(x = "magma", y = "matrix"),
-   function(x, y) .Call("magMultmm", x, FALSE, y, FALSE)
+   function(x, y) { .Call("magMultmm", x, FALSE, y, FALSE) }
 )
 
 setMethod("%*%", signature(x = "magma", y = "numeric"),
    function(x, y) {
+      #cat("mmn\n")
       if(ncol(x) == 1) .Call("magMultmm", x, FALSE, as.matrix(y), TRUE)
       else .Call("magMultmv", x, FALSE, y, TRUE)
    }
 )
 
 setMethod("%*%", signature(x = "matrix", y = "magma"),
-   function(x, y) .Call("magMultmm", x, FALSE, y, FALSE)
+   function(x, y) { .Call("magMultmm", x, FALSE, y, FALSE) }
 )
 
 setMethod("%*%", signature(x = "numeric", y = "magma"),
    function(x, y) {
+      #cat("nmm\n"); 
       if(nrow(y) == 1) .Call("magMultmm", as.matrix(x), FALSE, y, FALSE)
       else .Call("magMultmv", y, FALSE, x, FALSE)
    }
@@ -131,4 +133,3 @@ setMethod("tcrossprod", signature(x = "numeric", y = "magma"),
       else .Call("magMultmv", y, TRUE, x, FALSE)
    }
 )
-
diff --git a/R/magSolvers.r b/R/magSolvers.r
index cf431e4..e8df0b9 100644
--- a/R/magSolvers.r
+++ b/R/magSolvers.r
@@ -56,12 +56,12 @@ setMethod("solve", signature(a = "magmaQR", b = "missing"),
       n <- nrow(a$qr)
       if(n != ncol(a$qr))
          stop("only square matrices can be inverted in 'solve'", call.=FALSE)
-      callGeneric(a, diag(n))
+      .Call("magQRSolve", a, diag(n))
    }
 )
 
 setMethod("solve", signature(a = "magmaQR", b = "numeric"),
-   function(a, b) as.numeric(callGeneric(a, as.matrix(b)))
+   function(a, b) as.numeric(.Call("magQRSolve", a, as.matrix(b)))
 )
 
 
@@ -92,6 +92,7 @@ setMethod("solve", signature(a = "matrix", b = "magma"),
 
 setMethod("backsolve", signature(r = "magma", x = "magma"),
    function(r, x, k, upper.tri, transpose) {
+      #cat("utri=", upper.tri, "trans=", transpose, "\n", sep="")
       .Call("magTriSolve", r, x, k, upper.tri, transpose)
    }
 )
@@ -115,4 +116,3 @@ forwardsolve <- function(l, x, k = ncol(l), upper.tri=FALSE, transpose=FALSE)
 {
    backsolve(l, x, k, upper.tri, transpose)
 }
-
diff --git a/R/magUtils.r b/R/magUtils.r
index 6e79108..fd55e48 100644
--- a/R/magUtils.r
+++ b/R/magUtils.r
@@ -29,6 +29,3 @@
 .onUnload <- function(libname) {
    .C("magUnload", PACKAGE="magma")
 }
-
-
-
diff --git a/cleanup b/cleanup
old mode 100644
new mode 100755
index df5aa5d..28b2155
--- a/cleanup
+++ b/cleanup
@@ -5,13 +5,11 @@
 #
 echo "Cleaning up after installing the magma package"
 
-for f in config.log config.status config.cache ; do
+for f in config.log config.status config.cache *.bck */*.bck; do
   if test -w $f ; then
     rm -f $f
   fi
 done
 
 cd src
-make clean
-rm -f Makefile
-
+if test -e Makefile; then make clean; rm -f Makefile; fi
diff --git a/configure b/configure
old mode 100644
new mode 100755
index ebaa22a..b7cf4c6
--- a/configure
+++ b/configure
@@ -1,6 +1,6 @@
 #! /bin/sh
 # Guess values for system-dependent variables and create Makefiles.
-# Generated by GNU Autoconf 2.65 for magma 1.3.0.
+# Generated by GNU Autoconf 2.65 for magma 1.5.0.
 #
 #
 # Copyright (C) 1992, 1993, 1994, 1995, 1996, 1998, 1999, 2000, 2001,
@@ -548,8 +548,8 @@ MAKEFLAGS=
 # Identity of this package.
 PACKAGE_NAME='magma'
 PACKAGE_TARNAME='magma'
-PACKAGE_VERSION='1.3.0'
-PACKAGE_STRING='magma 1.3.0'
+PACKAGE_VERSION='1.5.0'
+PACKAGE_STRING='magma 1.5.0'
 PACKAGE_BUGREPORT=''
 PACKAGE_URL=''
 
@@ -608,7 +608,7 @@ ac_user_opts='
 enable_option_checking
 with_cuda_home
 with_lapack
-with_magma_lib
+with_magma_home
 with_r_home
 '
       ac_precious_vars='build_alias
@@ -1160,7 +1160,7 @@ if test "$ac_init_help" = "long"; then
   # Omit some internal or obsolete options to make the list less imposing.
   # This message is too long to be a string in the A/UX 3.1 sh.
   cat <<_ACEOF
-\`configure' configures magma 1.3.0 to adapt to many kinds of systems.
+\`configure' configures magma 1.5.0 to adapt to many kinds of systems.
 
 Usage: $0 [OPTION]... [VAR=VALUE]...
 
@@ -1221,7 +1221,7 @@ fi
 
 if test -n "$ac_init_help"; then
   case $ac_init_help in
-     short | recursive ) echo "Configuration of magma 1.3.0:";;
+     short | recursive ) echo "Configuration of magma 1.5.0:";;
    esac
   cat <<\_ACEOF
 
@@ -1230,7 +1230,7 @@ Optional Packages:
   --without-PACKAGE       do not use PACKAGE (same as --with-PACKAGE=no)
   --with-cuda-home=DIR    full path to CUDA installation [/usr/local/cuda]
   --with-lapack=LIBS      LAPACK shared library [-llapack]
-  --with-magma-lib=DIR    full path to MAGMA shared libraries []
+  --with-magma-home=DIR    full path to MAGMA shared libraries []
   --with-r-home=DIR       full path to R installation []
 
 Some influential environment variables:
@@ -1308,7 +1308,7 @@ fi
 test -n "$ac_init_help" && exit $ac_status
 if $ac_init_version; then
   cat <<\_ACEOF
-magma configure 1.3.0
+magma configure 1.5.0
 generated by GNU Autoconf 2.65
 
 Copyright (C) 2009 Free Software Foundation, Inc.
@@ -1363,7 +1363,7 @@ cat >config.log <<_ACEOF
 This file contains any messages produced by compilers while
 running configure, to aid debugging if configure makes a mistake.
 
-It was created by magma $as_me 1.3.0, which was
+It was created by magma $as_me 1.5.0, which was
 generated by GNU Autoconf 2.65.  Invocation command line was
 
   $ $0 $@
@@ -2376,7 +2376,7 @@ else
 fi
 
 
-NCFLAGS="-arch=sm_13 -O2 -DADD_"
+NCFLAGS="-arch=sm_20 -O3 -DADD_"
 CUDA_INC=${CUDA_HOME}/include
 { $as_echo "$as_me:${as_lineno-$LINENO}: checking for CUDA headers" >&5
 $as_echo_n "checking for CUDA headers... " >&6; }
@@ -2452,19 +2452,18 @@ fi
 
 
 
-#################### MAGMA Libraries ####################
+#################### MAGMA Home ####################
 
-NCFLAGS="${NCFLAGS} -I./include"
 
 
-# Check whether --with-magma-lib was given.
-if test "${with_magma_lib+set}" = set; then :
-  withval=$with_magma_lib; LDFLAGS="${LDFLAGS} -L$withval"
+# Check whether --with-magma-home was given.
+if test "${with_magma_home+set}" = set; then :
+  withval=$with_magma_home; LDFLAGS="${LDFLAGS} -L$withval/lib"; NCFLAGS="${NCFLAGS} -I$withval/include"
 
 fi
 
 
-LIBS="${LIBS} -lmagmablas -lmagma"
+LIBS="${LIBS} -lmagma"
 
 
 #################### R Software ####################
@@ -2473,9 +2472,10 @@ LIBS="${LIBS} -lmagmablas -lmagma"
 
 # Check whether --with-r-home was given.
 if test "${with_r_home+set}" = set; then :
-  withval=$with_r_home; R_HOME=$withval
+  withval=$with_r_home; R_HOME="$withval"
 
 fi
+R_INCLUDE_DIR="${R_HOME}/include"
 
 { $as_echo "$as_me:${as_lineno-$LINENO}: result: setting R home directory to ${R_HOME}" >&5
 $as_echo "setting R home directory to ${R_HOME}" >&6; }
@@ -3074,7 +3074,7 @@ cat >>$CONFIG_STATUS <<\_ACEOF || ac_write_fail=1
 # report actual input values of CONFIG_FILES etc. instead of their
 # values after options handling.
 ac_log="
-This file was extended by magma $as_me 1.3.0, which was
+This file was extended by magma $as_me 1.5.0, which was
 generated by GNU Autoconf 2.65.  Invocation command line was
 
   CONFIG_FILES    = $CONFIG_FILES
@@ -3127,7 +3127,7 @@ _ACEOF
 cat >>$CONFIG_STATUS <<_ACEOF || ac_write_fail=1
 ac_cs_config="`$as_echo "$ac_configure_args" | sed 's/^ //; s/[\\""\`\$]/\\\\&/g'`"
 ac_cs_version="\\
-magma config.status 1.3.0
+magma config.status 1.5.0
 configured by $0, generated by GNU Autoconf 2.65,
   with options \\"\$ac_cs_config\\"
 
diff --git a/configure.ac b/configure.ac
old mode 100644
new mode 100755
index a220e9d..6c956ff
--- a/configure.ac
+++ b/configure.ac
@@ -1,4 +1,4 @@
-AC_INIT([magma], 1.3.0)
+AC_INIT([magma], 1.5.0)
 AC_CONFIG_AUX_DIR([tools])
 
 #################### GCC Compiler ####################
@@ -18,7 +18,7 @@ NVCC=${CUDA_HOME}/bin/nvcc
 AC_CHECK_FILE([${NVCC}],,
 	[AC_MSG_ERROR([nvcc not found!])])
 
-NCFLAGS="-arch=sm_13 -O2 -DADD_"
+NCFLAGS="-arch=sm_20 -O3 -DADD_"
 CUDA_INC=${CUDA_HOME}/include
 AC_MSG_CHECKING([for CUDA headers])
 AC_CHECK_FILE([${CUDA_INC}/cublas.h],
@@ -46,17 +46,39 @@ AC_ARG_WITH([lapack],
 )
 
 
-#################### MAGMA Libraries ####################
+#################### MAGMA home ####################
+AC_ARG_WITH([cuda-home],
+	[AC_HELP_STRING([--with-cuda-home=DIR],
+		[full path to CUDA installation @<:@/usr/local/cuda@:>@])],
+	[CUDA_HOME="$withval"], [CUDA_HOME=/usr/local/cuda])
+AC_MSG_RESULT([setting CUDA home directory to ${CUDA_HOME}])
 
-NCFLAGS="${NCFLAGS} -I./include"
+NVCC=${CUDA_HOME}/bin/nvcc
+AC_CHECK_FILE([${NVCC}],,
+	[AC_MSG_ERROR([nvcc not found!])])
 
-AC_ARG_WITH([magma-lib],
-	[AC_HELP_STRING([--with-magma-lib=DIR],
-		[full path to MAGMA shared libraries @<:@@:>@])],
-	[LDFLAGS="${LDFLAGS} -L$withval"]
+NCFLAGS="-arch=sm_20 -O3 -DADD_"
+CUDA_INC=${CUDA_HOME}/include
+AC_MSG_CHECKING([for CUDA headers])
+AC_CHECK_FILE([${CUDA_INC}/cublas.h],
+	[NCFLAGS="${NCFLAGS} -I${CUDA_INC}"],
+	[AC_MSG_ERROR([CUDA headers not found!])]
+)
+
+
+AC_ARG_WITH([magma-home],
+	[AC_HELP_STRING([--with-magma-home=DIR],
+		[full path to MAGMA installation @<:@@:>@])],
+	[MAGMA_HOME="$withval"], [MAGMA_HOME=/usr/local/magma]
+	[LDFLAGS="${LDFLAGS} -L$withval/lib"]
+)
+AC_MSG_CHECKING([for MAGMA headers])
+AC_CHECK_FILE([${MAGMA_HOME}/include/magma.h],
+	[NCFLAGS="${NCFLAGS} -I${MAGMA_HOME}/include"],
+	[AC_MSG_ERROR([MAGMA headers not found!])]
 )
 
-LIBS="${LIBS} -lmagmablas -lmagma"
+LIBS="${LIBS} -lmagma"
 
 
 #################### R Software ####################
diff --git a/inst/CITATION b/inst/CITATION
old mode 100644
new mode 100755
index ae3cdb8..b856b6b
--- a/inst/CITATION
+++ b/inst/CITATION
@@ -1,12 +1,19 @@
-citEntry(entry="misc",
-  title = "R package magma: Matrix Algebra on GPU and Multicore Architectures, Version 0.2.2",
-  author = personList(person("Brian", "Smith", "J.", "brian-j-smith@uiowa.edu")),
-  year = "September 3, 2010",
-  note = "[On-line] http://cran.r-project.org/package=magma",
-  textVersion = paste("Smith, B.J. (September 3, 2010)",
-    "R package magma: Matrix Algebra on GPU and Multicore Architectures, Version 0.2.2,",
-    "[On-line] http://cran.r-project.org/package=magma."),
-  header = "To cite magma in publications use"
-)
-
-
+citHeader("To cite magma in publications use:")
+citEntry(entry="misc",
+  title = "R package magma: Matrix Algebra on GPU and Multicore Architectures, Version 0.2.2",
+  author = personList(person("Brian", "Smith", "J.", "brian-j-smith@uiowa.edu")),
+  year = "September 3, 2010",
+  note = "[On-line] http://cran.r-project.org/package=magma",
+  textVersion = paste("Smith, B.J. (September 3, 2010)",
+    "R package magma: Matrix Algebra on GPU and Multicore Architectures, Version 0.2.2,",
+    "[On-line] http://cran.r-project.org/package=magma")
+)
+citEntry(entry="misc",
+  title = "R package magma: Matrix Algebra on GPU and Multicore Architectures, Version 1.5.0",
+  author = personList(person("Serguei", "Sokol", "", "sokol@insa-toulouse.fr")),
+  year = "October 7, 2014",
+  note = "[On-line] http://cran.r-project.org/package=magma",
+  textVersion = paste("Sokol, S. (October 7, 2014)",
+    "R package magma: Matrix Algebra on GPU and Multicore Architectures, Version 1.5.0,",
+    "[On-line] http://cran.r-project.org/package=magma"),
+)
diff --git a/man/magma.Rd b/man/magma.Rd
index a430926..b772f20 100644
--- a/man/magma.Rd
+++ b/man/magma.Rd
@@ -1,62 +1,62 @@
-\name{magma}
-\alias{magma}
-
-\title{Matrix Class "magma" Constructor}
-
-\description{
-Constructs a \linkS4class{magma} class object that inherits from \linkS4class{matrix}.
-}
-
-\usage{
-magma(data = NA, nrow = 1, ncol = 1, byrow = FALSE,
-      dimnames = NULL, gpu = TRUE)
-}
-
-\arguments{
-  \item{data}{a numeric data vector or matrix.}
-  \item{nrow}{when \code{data} is not a matrix, the desired number of rows}
-  \item{ncol}{when \code{data} is not a matrix, the desired number of columns}
-  \item{byrow}{logical.  If \code{FALSE} (the default) the matrix is filled by
-  columns, otherwise the matrix is filled by rows.}
-  \item{dimnames}{a \code{\link{dimnames}} attribute for the matrix: a
-  \code{list} of two character components.  They are set if not
-  \code{\link{NULL}} (as per default).}
-  \item{gpu}{logical.  If \code{TRUE}, the MAGMA library GPU interface is used
-  for matrix operations, when available (default).  Otherwise, the CPU interface
-  is used.}
-}
-
-\value{
-Returns a \linkS4class{magma} class matrix.  The arguments \code{nrow}, \code{ncol}, \code{byrow}, and \code{dimnames} are made use of only if \code{data} is not a matrix or magma object.
-}
-
-\details{
-If either of \code{nrow} or \code{ncol} is not given, an attempt is made to infer its value from the length of \code{data} and the other parameter.
-
-The MAGMA library of C functions offers two LAPACK-style interfaces, referred to
-as the \bold{GPU interface} and the \bold{CPU interface}.  The GPU interface
-takes input and produces results in the GPU's memory, whereas the CPU interface
-produces results in the CPU's memory.  The GPU and CPU interfaces, although
-similar, are not derivatives of each other, but instead have different
-communication patters.  The \code{gpu} argument allows specification of the
-interface to be used for caluculations involving the resulting magma matrix.
-}
-
-\author{
-Brian J. Smith <brian-j-smith@uiowa.edu>
-}
-
-\seealso{
-\code{\link[magma]{gpu}},
-\code{\linkS4class{magma}},
-\code{\linkS4class{matrix}}
-}
-
-\examples{
-magma(1:6, nrow = 2, ncol = 3, byrow = TRUE,
-      dimnames = list(c("row1", "row2"), c("col1", "col2", "col3")),
-      gpu = TRUE)
-}
-
-\keyword{array}
-\keyword{algebra}
+\name{magma}
+\alias{magma}
+
+\title{Matrix Class "magma" Constructor}
+
+\description{
+Constructs a \linkS4class{magma} class object that inherits from \linkS4class{matrix}.
+}
+
+\usage{
+magma(data = NA, nrow = 1, ncol = 1, byrow = FALSE,
+      dimnames = NULL, gpu = TRUE)
+}
+
+\arguments{
+  \item{data}{a numeric data vector or matrix.}
+  \item{nrow}{when \code{data} is not a matrix, the desired number of rows}
+  \item{ncol}{when \code{data} is not a matrix, the desired number of columns}
+  \item{byrow}{logical.  If \code{FALSE} (the default) the matrix is filled by
+  columns, otherwise the matrix is filled by rows.}
+  \item{dimnames}{a \code{\link{dimnames}} attribute for the matrix: a
+  \code{list} of two character components.  They are set if not
+  \code{\link{NULL}} (as per default).}
+  \item{gpu}{logical.  If \code{TRUE}, the MAGMA library GPU interface is used
+  for matrix operations, when available (default).  Otherwise, the CPU interface
+  is used.}
+}
+
+\value{
+Returns a \linkS4class{magma} class matrix.  The arguments \code{nrow}, \code{ncol}, \code{byrow}, and \code{dimnames} are made use of only if \code{data} is not a matrix or magma object.
+}
+
+\details{
+If either of \code{nrow} or \code{ncol} is not given, an attempt is made to infer its value from the length of \code{data} and the other parameter.
+
+The MAGMA library of C functions offers two LAPACK-style interfaces, referred to
+as the \bold{GPU interface} and the \bold{CPU interface}.  The GPU interface
+takes input and produces results in the GPU's memory, whereas the CPU interface
+produces results in the CPU's memory.  The GPU and CPU interfaces, although
+similar, are not derivatives of each other, but instead have different
+communication patters.  The \code{gpu} argument allows specification of the
+interface to be used for caluculations involving the resulting magma matrix.
+}
+
+\author{
+Brian J. Smith <brian-j-smith@uiowa.edu>
+}
+
+\seealso{
+\code{\link[magma]{gpu}},
+\code{\linkS4class{magma}},
+\code{\linkS4class{matrix}}
+}
+
+\examples{
+magma(1:6, nrow = 2, ncol = 3, byrow = TRUE,
+      dimnames = list(c("row1", "row2"), c("col1", "col2", "col3")),
+      gpu = TRUE)
+}
+
+\keyword{array}
+\keyword{algebra}
diff --git a/src/Makefile.in b/src/Makefile.in
old mode 100644
new mode 100755
diff --git a/src/include/cblas.h b/src/include/cblas.h
deleted file mode 100644
index b748819..0000000
--- a/src/include/cblas.h
+++ /dev/null
@@ -1,579 +0,0 @@
-#ifndef CBLAS_H
-#define CBLAS_H
-#include <stddef.h>
-
-/*
- * Enumerated and derived types
- */
-#define CBLAS_INDEX size_t  /* this may vary between platforms */
-
-enum CBLAS_ORDER {CblasRowMajor=101, CblasColMajor=102};
-enum CBLAS_TRANSPOSE {CblasNoTrans=111, CblasTrans=112, CblasConjTrans=113};
-enum CBLAS_UPLO {CblasUpper=121, CblasLower=122};
-enum CBLAS_DIAG {CblasNonUnit=131, CblasUnit=132};
-enum CBLAS_SIDE {CblasLeft=141, CblasRight=142};
-
-/*
- * ===========================================================================
- * Prototypes for level 1 BLAS functions (complex are recast as routines)
- * ===========================================================================
- */
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-float  cblas_sdsdot(const int N, const float alpha, const float *X,
-                    const int incX, const float *Y, const int incY);
-double cblas_dsdot(const int N, const float *X, const int incX, const float *Y,
-                   const int incY);
-float  cblas_sdot(const int N, const float  *X, const int incX,
-                  const float  *Y, const int incY);
-double cblas_ddot(const int N, const double *X, const int incX,
-                  const double *Y, const int incY);
-
-/*
- * Functions having prefixes Z and C only
- */
-void   cblas_cdotu_sub(const int N, const void *X, const int incX,
-                       const void *Y, const int incY, void *dotu);
-void   cblas_cdotc_sub(const int N, const void *X, const int incX,
-                       const void *Y, const int incY, void *dotc);
-
-void   cblas_zdotu_sub(const int N, const void *X, const int incX,
-                       const void *Y, const int incY, void *dotu);
-void   cblas_zdotc_sub(const int N, const void *X, const int incX,
-                       const void *Y, const int incY, void *dotc);
-
-
-/*
- * Functions having prefixes S D SC DZ
- */
-float  cblas_snrm2(const int N, const float *X, const int incX);
-float  cblas_sasum(const int N, const float *X, const int incX);
-
-double cblas_dnrm2(const int N, const double *X, const int incX);
-double cblas_dasum(const int N, const double *X, const int incX);
-
-float  cblas_scnrm2(const int N, const void *X, const int incX);
-float  cblas_scasum(const int N, const void *X, const int incX);
-
-double cblas_dznrm2(const int N, const void *X, const int incX);
-double cblas_dzasum(const int N, const void *X, const int incX);
-
-
-/*
- * Functions having standard 4 prefixes (S D C Z)
- */
-CBLAS_INDEX cblas_isamax(const int N, const float  *X, const int incX);
-CBLAS_INDEX cblas_idamax(const int N, const double *X, const int incX);
-CBLAS_INDEX cblas_icamax(const int N, const void   *X, const int incX);
-CBLAS_INDEX cblas_izamax(const int N, const void   *X, const int incX);
-
-/*
- * ===========================================================================
- * Prototypes for level 1 BLAS routines
- * ===========================================================================
- */
-
-/*
- * Routines with standard 4 prefixes (s, d, c, z)
- */
-void cblas_sswap(const int N, float *X, const int incX,
-                 float *Y, const int incY);
-void cblas_scopy(const int N, const float *X, const int incX,
-                 float *Y, const int incY);
-void cblas_saxpy(const int N, const float alpha, const float *X,
-                 const int incX, float *Y, const int incY);
-
-void cblas_dswap(const int N, double *X, const int incX,
-                 double *Y, const int incY);
-void cblas_dcopy(const int N, const double *X, const int incX,
-                 double *Y, const int incY);
-void cblas_daxpy(const int N, const double alpha, const double *X,
-                 const int incX, double *Y, const int incY);
-
-void cblas_cswap(const int N, void *X, const int incX,
-                 void *Y, const int incY);
-void cblas_ccopy(const int N, const void *X, const int incX,
-                 void *Y, const int incY);
-void cblas_caxpy(const int N, const void *alpha, const void *X,
-                 const int incX, void *Y, const int incY);
-
-void cblas_zswap(const int N, void *X, const int incX,
-                 void *Y, const int incY);
-void cblas_zcopy(const int N, const void *X, const int incX,
-                 void *Y, const int incY);
-void cblas_zaxpy(const int N, const void *alpha, const void *X,
-                 const int incX, void *Y, const int incY);
-
-
-/*
- * Routines with S and D prefix only
- */
-void cblas_srotg(float *a, float *b, float *c, float *s);
-void cblas_srotmg(float *d1, float *d2, float *b1, const float b2, float *P);
-void cblas_srot(const int N, float *X, const int incX,
-                float *Y, const int incY, const float c, const float s);
-void cblas_srotm(const int N, float *X, const int incX,
-                float *Y, const int incY, const float *P);
-
-void cblas_drotg(double *a, double *b, double *c, double *s);
-void cblas_drotmg(double *d1, double *d2, double *b1, const double b2, double *P);
-void cblas_drot(const int N, double *X, const int incX,
-                double *Y, const int incY, const double c, const double  s);
-void cblas_drotm(const int N, double *X, const int incX,
-                double *Y, const int incY, const double *P);
-
-
-/*
- * Routines with S D C Z CS and ZD prefixes
- */
-void cblas_sscal(const int N, const float alpha, float *X, const int incX);
-void cblas_dscal(const int N, const double alpha, double *X, const int incX);
-void cblas_cscal(const int N, const void *alpha, void *X, const int incX);
-void cblas_zscal(const int N, const void *alpha, void *X, const int incX);
-void cblas_csscal(const int N, const float alpha, void *X, const int incX);
-void cblas_zdscal(const int N, const double alpha, void *X, const int incX);
-
-/*
- * ===========================================================================
- * Prototypes for level 2 BLAS
- * ===========================================================================
- */
-
-/*
- * Routines with standard 4 prefixes (S, D, C, Z)
- */
-void cblas_sgemv(const enum CBLAS_ORDER order,
-                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
-                 const float alpha, const float *A, const int lda,
-                 const float *X, const int incX, const float beta,
-                 float *Y, const int incY);
-void cblas_sgbmv(const enum CBLAS_ORDER order,
-                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
-                 const int KL, const int KU, const float alpha,
-                 const float *A, const int lda, const float *X,
-                 const int incX, const float beta, float *Y, const int incY);
-void cblas_strmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const float *A, const int lda,
-                 float *X, const int incX);
-void cblas_stbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const int K, const float *A, const int lda,
-                 float *X, const int incX);
-void cblas_stpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const float *Ap, float *X, const int incX);
-void cblas_strsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const float *A, const int lda, float *X,
-                 const int incX);
-void cblas_stbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const int K, const float *A, const int lda,
-                 float *X, const int incX);
-void cblas_stpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const float *Ap, float *X, const int incX);
-
-void cblas_dgemv(const enum CBLAS_ORDER order,
-                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
-                 const double alpha, const double *A, const int lda,
-                 const double *X, const int incX, const double beta,
-                 double *Y, const int incY);
-void cblas_dgbmv(const enum CBLAS_ORDER order,
-                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
-                 const int KL, const int KU, const double alpha,
-                 const double *A, const int lda, const double *X,
-                 const int incX, const double beta, double *Y, const int incY);
-void cblas_dtrmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const double *A, const int lda,
-                 double *X, const int incX);
-void cblas_dtbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const int K, const double *A, const int lda,
-                 double *X, const int incX);
-void cblas_dtpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const double *Ap, double *X, const int incX);
-void cblas_dtrsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const double *A, const int lda, double *X,
-                 const int incX);
-void cblas_dtbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const int K, const double *A, const int lda,
-                 double *X, const int incX);
-void cblas_dtpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const double *Ap, double *X, const int incX);
-
-void cblas_cgemv(const enum CBLAS_ORDER order,
-                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
-                 const void *alpha, const void *A, const int lda,
-                 const void *X, const int incX, const void *beta,
-                 void *Y, const int incY);
-void cblas_cgbmv(const enum CBLAS_ORDER order,
-                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
-                 const int KL, const int KU, const void *alpha,
-                 const void *A, const int lda, const void *X,
-                 const int incX, const void *beta, void *Y, const int incY);
-void cblas_ctrmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const void *A, const int lda,
-                 void *X, const int incX);
-void cblas_ctbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const int K, const void *A, const int lda,
-                 void *X, const int incX);
-void cblas_ctpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const void *Ap, void *X, const int incX);
-void cblas_ctrsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const void *A, const int lda, void *X,
-                 const int incX);
-void cblas_ctbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const int K, const void *A, const int lda,
-                 void *X, const int incX);
-void cblas_ctpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const void *Ap, void *X, const int incX);
-
-void cblas_zgemv(const enum CBLAS_ORDER order,
-                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
-                 const void *alpha, const void *A, const int lda,
-                 const void *X, const int incX, const void *beta,
-                 void *Y, const int incY);
-void cblas_zgbmv(const enum CBLAS_ORDER order,
-                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
-                 const int KL, const int KU, const void *alpha,
-                 const void *A, const int lda, const void *X,
-                 const int incX, const void *beta, void *Y, const int incY);
-void cblas_ztrmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const void *A, const int lda,
-                 void *X, const int incX);
-void cblas_ztbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const int K, const void *A, const int lda,
-                 void *X, const int incX);
-void cblas_ztpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const void *Ap, void *X, const int incX);
-void cblas_ztrsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const void *A, const int lda, void *X,
-                 const int incX);
-void cblas_ztbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const int K, const void *A, const int lda,
-                 void *X, const int incX);
-void cblas_ztpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
-                 const int N, const void *Ap, void *X, const int incX);
-
-
-/*
- * Routines with S and D prefixes only
- */
-void cblas_ssymv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const float alpha, const float *A,
-                 const int lda, const float *X, const int incX,
-                 const float beta, float *Y, const int incY);
-void cblas_ssbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const int K, const float alpha, const float *A,
-                 const int lda, const float *X, const int incX,
-                 const float beta, float *Y, const int incY);
-void cblas_sspmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const float alpha, const float *Ap,
-                 const float *X, const int incX,
-                 const float beta, float *Y, const int incY);
-void cblas_sger(const enum CBLAS_ORDER order, const int M, const int N,
-                const float alpha, const float *X, const int incX,
-                const float *Y, const int incY, float *A, const int lda);
-void cblas_ssyr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const float alpha, const float *X,
-                const int incX, float *A, const int lda);
-void cblas_sspr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const float alpha, const float *X,
-                const int incX, float *Ap);
-void cblas_ssyr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const float alpha, const float *X,
-                const int incX, const float *Y, const int incY, float *A,
-                const int lda);
-void cblas_sspr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const float alpha, const float *X,
-                const int incX, const float *Y, const int incY, float *A);
-
-void cblas_dsymv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const double alpha, const double *A,
-                 const int lda, const double *X, const int incX,
-                 const double beta, double *Y, const int incY);
-void cblas_dsbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const int K, const double alpha, const double *A,
-                 const int lda, const double *X, const int incX,
-                 const double beta, double *Y, const int incY);
-void cblas_dspmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const double alpha, const double *Ap,
-                 const double *X, const int incX,
-                 const double beta, double *Y, const int incY);
-void cblas_dger(const enum CBLAS_ORDER order, const int M, const int N,
-                const double alpha, const double *X, const int incX,
-                const double *Y, const int incY, double *A, const int lda);
-void cblas_dsyr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const double alpha, const double *X,
-                const int incX, double *A, const int lda);
-void cblas_dspr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const double alpha, const double *X,
-                const int incX, double *Ap);
-void cblas_dsyr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const double alpha, const double *X,
-                const int incX, const double *Y, const int incY, double *A,
-                const int lda);
-void cblas_dspr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const double alpha, const double *X,
-                const int incX, const double *Y, const int incY, double *A);
-
-
-/*
- * Routines with C and Z prefixes only
- */
-void cblas_chemv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const void *alpha, const void *A,
-                 const int lda, const void *X, const int incX,
-                 const void *beta, void *Y, const int incY);
-void cblas_chbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const int K, const void *alpha, const void *A,
-                 const int lda, const void *X, const int incX,
-                 const void *beta, void *Y, const int incY);
-void cblas_chpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const void *alpha, const void *Ap,
-                 const void *X, const int incX,
-                 const void *beta, void *Y, const int incY);
-void cblas_cgeru(const enum CBLAS_ORDER order, const int M, const int N,
-                 const void *alpha, const void *X, const int incX,
-                 const void *Y, const int incY, void *A, const int lda);
-void cblas_cgerc(const enum CBLAS_ORDER order, const int M, const int N,
-                 const void *alpha, const void *X, const int incX,
-                 const void *Y, const int incY, void *A, const int lda);
-void cblas_cher(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const float alpha, const void *X, const int incX,
-                void *A, const int lda);
-void cblas_chpr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const float alpha, const void *X,
-                const int incX, void *A);
-void cblas_cher2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo, const int N,
-                const void *alpha, const void *X, const int incX,
-                const void *Y, const int incY, void *A, const int lda);
-void cblas_chpr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo, const int N,
-                const void *alpha, const void *X, const int incX,
-                const void *Y, const int incY, void *Ap);
-
-void cblas_zhemv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const void *alpha, const void *A,
-                 const int lda, const void *X, const int incX,
-                 const void *beta, void *Y, const int incY);
-void cblas_zhbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const int K, const void *alpha, const void *A,
-                 const int lda, const void *X, const int incX,
-                 const void *beta, void *Y, const int incY);
-void cblas_zhpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                 const int N, const void *alpha, const void *Ap,
-                 const void *X, const int incX,
-                 const void *beta, void *Y, const int incY);
-void cblas_zgeru(const enum CBLAS_ORDER order, const int M, const int N,
-                 const void *alpha, const void *X, const int incX,
-                 const void *Y, const int incY, void *A, const int lda);
-void cblas_zgerc(const enum CBLAS_ORDER order, const int M, const int N,
-                 const void *alpha, const void *X, const int incX,
-                 const void *Y, const int incY, void *A, const int lda);
-void cblas_zher(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const double alpha, const void *X, const int incX,
-                void *A, const int lda);
-void cblas_zhpr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
-                const int N, const double alpha, const void *X,
-                const int incX, void *A);
-void cblas_zher2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo, const int N,
-                const void *alpha, const void *X, const int incX,
-                const void *Y, const int incY, void *A, const int lda);
-void cblas_zhpr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo, const int N,
-                const void *alpha, const void *X, const int incX,
-                const void *Y, const int incY, void *Ap);
-
-/*
- * ===========================================================================
- * Prototypes for level 3 BLAS
- * ===========================================================================
- */
-
-/*
- * Routines with standard 4 prefixes (S, D, C, Z)
- */
-void cblas_sgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
-                 const int K, const float alpha, const float *A,
-                 const int lda, const float *B, const int ldb,
-                 const float beta, float *C, const int ldc);
-void cblas_ssymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const int M, const int N,
-                 const float alpha, const float *A, const int lda,
-                 const float *B, const int ldb, const float beta,
-                 float *C, const int ldc);
-void cblas_ssyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                 const float alpha, const float *A, const int lda,
-                 const float beta, float *C, const int ldc);
-void cblas_ssyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                  const float alpha, const float *A, const int lda,
-                  const float *B, const int ldb, const float beta,
-                  float *C, const int ldc);
-void cblas_strmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_DIAG Diag, const int M, const int N,
-                 const float alpha, const float *A, const int lda,
-                 float *B, const int ldb);
-void cblas_strsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_DIAG Diag, const int M, const int N,
-                 const float alpha, const float *A, const int lda,
-                 float *B, const int ldb);
-
-void cblas_dgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
-                 const int K, const double alpha, const double *A,
-                 const int lda, const double *B, const int ldb,
-                 const double beta, double *C, const int ldc);
-void cblas_dsymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const int M, const int N,
-                 const double alpha, const double *A, const int lda,
-                 const double *B, const int ldb, const double beta,
-                 double *C, const int ldc);
-void cblas_dsyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                 const double alpha, const double *A, const int lda,
-                 const double beta, double *C, const int ldc);
-void cblas_dsyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                  const double alpha, const double *A, const int lda,
-                  const double *B, const int ldb, const double beta,
-                  double *C, const int ldc);
-void cblas_dtrmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_DIAG Diag, const int M, const int N,
-                 const double alpha, const double *A, const int lda,
-                 double *B, const int ldb);
-void cblas_dtrsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_DIAG Diag, const int M, const int N,
-                 const double alpha, const double *A, const int lda,
-                 double *B, const int ldb);
-
-void cblas_cgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
-                 const int K, const void *alpha, const void *A,
-                 const int lda, const void *B, const int ldb,
-                 const void *beta, void *C, const int ldc);
-void cblas_csymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const int M, const int N,
-                 const void *alpha, const void *A, const int lda,
-                 const void *B, const int ldb, const void *beta,
-                 void *C, const int ldc);
-void cblas_csyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                 const void *alpha, const void *A, const int lda,
-                 const void *beta, void *C, const int ldc);
-void cblas_csyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                  const void *alpha, const void *A, const int lda,
-                  const void *B, const int ldb, const void *beta,
-                  void *C, const int ldc);
-void cblas_ctrmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_DIAG Diag, const int M, const int N,
-                 const void *alpha, const void *A, const int lda,
-                 void *B, const int ldb);
-void cblas_ctrsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_DIAG Diag, const int M, const int N,
-                 const void *alpha, const void *A, const int lda,
-                 void *B, const int ldb);
-
-void cblas_zgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
-                 const int K, const void *alpha, const void *A,
-                 const int lda, const void *B, const int ldb,
-                 const void *beta, void *C, const int ldc);
-void cblas_zsymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const int M, const int N,
-                 const void *alpha, const void *A, const int lda,
-                 const void *B, const int ldb, const void *beta,
-                 void *C, const int ldc);
-void cblas_zsyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                 const void *alpha, const void *A, const int lda,
-                 const void *beta, void *C, const int ldc);
-void cblas_zsyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                  const void *alpha, const void *A, const int lda,
-                  const void *B, const int ldb, const void *beta,
-                  void *C, const int ldc);
-void cblas_ztrmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_DIAG Diag, const int M, const int N,
-                 const void *alpha, const void *A, const int lda,
-                 void *B, const int ldb);
-void cblas_ztrsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
-                 const enum CBLAS_DIAG Diag, const int M, const int N,
-                 const void *alpha, const void *A, const int lda,
-                 void *B, const int ldb);
-
-
-/*
- * Routines with prefixes C and Z only
- */
-void cblas_chemm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const int M, const int N,
-                 const void *alpha, const void *A, const int lda,
-                 const void *B, const int ldb, const void *beta,
-                 void *C, const int ldc);
-void cblas_cherk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                 const float alpha, const void *A, const int lda,
-                 const float beta, void *C, const int ldc);
-void cblas_cher2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                  const void *alpha, const void *A, const int lda,
-                  const void *B, const int ldb, const float beta,
-                  void *C, const int ldc);
-
-void cblas_zhemm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
-                 const enum CBLAS_UPLO Uplo, const int M, const int N,
-                 const void *alpha, const void *A, const int lda,
-                 const void *B, const int ldb, const void *beta,
-                 void *C, const int ldc);
-void cblas_zherk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                 const double alpha, const void *A, const int lda,
-                 const double beta, void *C, const int ldc);
-void cblas_zher2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
-                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
-                  const void *alpha, const void *A, const int lda,
-                  const void *B, const int ldb, const double beta,
-                  void *C, const int ldc);
-
-void cblas_xerbla(int p, const char *rout, const char *form, ...);
-
-
-#ifdef __cplusplus
-}
-#endif
-
-
-#endif
diff --git a/src/include/gnu/stubs-32.h b/src/include/gnu/stubs-32.h
deleted file mode 100644
index 1404dcd..0000000
--- a/src/include/gnu/stubs-32.h
+++ /dev/null
@@ -1,24 +0,0 @@
-/* This file is automatically generated.
-   It defines a symbol `__stub_FUNCTION' for each function
-   in the C library which is a stub, meaning it will fail
-   every time called, usually setting errno to ENOSYS.  */
-
-#ifdef _LIBC
- #error Applications may not define the macro _LIBC
-#endif
-
-#define __stub___kernel_cosl
-#define __stub___kernel_sinl
-#define __stub___kernel_tanl
-#define __stub_chflags
-#define __stub_fattach
-#define __stub_fchflags
-#define __stub_fdetach
-#define __stub_gtty
-#define __stub_lchmod
-#define __stub_lutimes
-#define __stub_revoke
-#define __stub_setlogin
-#define __stub_sigreturn
-#define __stub_sstk
-#define __stub_stty
diff --git a/src/include/magma.h b/src/include/magma.h
deleted file mode 100644
index be63567..0000000
--- a/src/include/magma.h
+++ /dev/null
@@ -1,336 +0,0 @@
-/*
-    -- MAGMA (version 1.3.0) --
-       Univ. of Tennessee, Knoxville
-       Univ. of California, Berkeley
-       Univ. of Colorado, Denver
-       November 2012
-*/
-
-/*#include <quark.h>*/
-
-#ifndef _MAGMA_
-#define _MAGMA_
-
-/* ------------------------------------------------------------
- * MAGMA Blas Functions
- * --------------------------------------------------------- */
-#include "magmablas.h"
-
-/* ------------------------------------------------------------
- * MAGMA functions
- * --------------------------------------------------------- */
-#include "magma_z.h"
-#include "magma_c.h"
-#include "magma_d.h"
-#include "magma_s.h"
-#include "magma_zc.h"
-#include "magma_ds.h"
-
-/* ------------------------------------------------------------
- * MAGMA constants
- * --------------------------------------------------------- */
-#define MAGMA_VERSION_MAJOR 1
-#define MAGMA_VERSION_MINOR 3
-#define MAGMA_VERSION_MICRO 0
-
-#define MagmaNoTrans       'N'
-#define MagmaTrans         'T'
-#define MagmaConjTrans     'C'
-
-#define MagmaUpper         'U'
-#define MagmaLower         'L'
-#define MagmaUpperLower    'A'
-
-#define MagmaNonUnit       'N'
-#define MagmaUnit          'U'
-
-#define MagmaLeft          'L'
-#define MagmaRight         'R'
-
-#define MagmaForward       'F'
-#define MagmaBackward      'B'
-
-#define MagmaColumnwise    'C'
-#define MagmaRowwise       'R'
-
-#define MagmaNoVectors     'N'
-#define MagmaVectors       'V'
-
-#define MagmaNoTransStr    "NonTrans"
-#define MagmaTransStr      "Trans"
-#define MagmaConjTransStr  "Conj"
-
-#define MagmaUpperStr      "Upper"
-#define MagmaLowerStr      "Lower"
-#define MagmaUpperLowerStr "All"
-
-#define MagmaNonUnitStr    "NonUnit"
-#define MagmaUnitStr       "Unit"
-
-#define MagmaLeftStr       "Left"
-#define MagmaRightStr      "Right"
-
-#define MagmaForwardStr    "Forward"
-#define MagmaBackwardStr   "Backward"
-
-#define MagmaColumnwiseStr "Columnwise"
-#define MagmaRowwiseStr    "Rowwise"
-
-#define MagmaNoVectorsStr  "NoVectors"
-#define MagmaVectorsStr    "Vectors"
-
-/* ------------------------------------------------------------
- * Return codes
- * LAPACK argument errors are < 0 but > MAGMA_ERR.
- * MAGMA errors are < MAGMA_ERR.
- * --------------------------------------------------------- */
-#define MAGMA_SUCCESS                 0
-#define MAGMA_ERR                  -100
-#define MAGMA_ERR_NOT_INITIALIZED  -101
-#define MAGMA_ERR_REINITIALIZED    -102
-#define MAGMA_ERR_NOT_SUPPORTED    -103
-#define MAGMA_ERR_ILLEGAL_VALUE    -104
-#define MAGMA_ERR_NOT_FOUND        -105
-#define MAGMA_ERR_OUT_OF_RESOURCES -106
-#define MAGMA_ERR_ALLOCATION       -106
-#define MAGMA_ERR_INTERNAL_LIMIT   -107
-#define MAGMA_ERR_UNALLOCATED      -108
-#define MAGMA_ERR_FILESYSTEM       -109
-#define MAGMA_ERR_UNEXPECTED       -110
-#define MAGMA_ERR_SEQUENCE_FLUSHED -111
-#define MAGMA_ERR_HOST_ALLOC       -112
-#define MAGMA_ERR_DEVICE_ALLOC     -113
-#define MAGMA_ERR_CUDASTREAM       -114
-#define MAGMA_ERR_INVALID_PTR      -115
-
-/* ------------------------------------------------------------
- *   Define new type that will not be changed by the generator
- * --------------------------------------------------------- */
-typedef double real_Double_t;
-
-/* ------------------------------------------------------------
- *   Macros to deal with cuda complex
- * --------------------------------------------------------- */
-#define MAGMA_Z_SET2REAL(v, t)    {(v).x = (t); (v).y = 0.0;}
-#define MAGMA_Z_EQUAL(u,v)        (((u).x == (v).x) && ((u).y == (v).y))
-#define MAGMA_Z_DSCALE(v, t, s)   {(v).x = (t).x/(s); (v).y = (t).y/(s);}
-#define MAGMA_Z_MAKE(r, i)        make_cuDoubleComplex((r), (i))
-#define MAGMA_Z_REAL(a)           cuCreal(a)
-#define MAGMA_Z_IMAG(a)           cuCimag(a)
-#define MAGMA_Z_ADD(a, b)         cuCadd((a), (b))
-#define MAGMA_Z_SUB(a, b)         cuCsub((a), (b))
-#define MAGMA_Z_MUL(a, b)         cuCmul((a), (b))
-#define MAGMA_Z_DIV(a, b)         cuCdiv((a), (b))
-#define MAGMA_Z_ABS(a)            cuCabs((a))
-#define MAGMA_Z_CNJG(a)           cuConj(a)
-#define MAGMA_Z_NEGATE(a)         make_cuDoubleComplex( -(a).x, -(a).y )
-#define MAGMA_Z_ZERO              make_cuDoubleComplex(0.0, 0.0)
-#define MAGMA_Z_ONE               make_cuDoubleComplex(1.0, 0.0)
-#define MAGMA_Z_HALF              make_cuDoubleComplex(0.5, 0.0)
-#define MAGMA_Z_NEG_ONE           make_cuDoubleComplex(-1.0, 0.0)
-#define MAGMA_Z_NEG_HALF          make_cuDoubleComplex(-0.5, 0.0)
-
-#define MAGMA_C_SET2REAL(v, t)    {(v).x = (t); (v).y = 0.0;}
-#define MAGMA_C_EQUAL(u,v)        (((u).x == (v).x) && ((u).y == (v).y))
-#define MAGMA_C_SSCALE(v, t, s)   {(v).x = (t).x/(s); (v).y = (t).y/(s);}
-#define MAGMA_C_MAKE(r, i)        make_cuFloatComplex((r), (i))
-#define MAGMA_C_REAL(a)           cuCrealf(a)
-#define MAGMA_C_IMAG(a)           cuCimagf(a)
-#define MAGMA_C_ADD(a, b)         cuCaddf((a), (b))
-#define MAGMA_C_SUB(a, b)         cuCsubf((a), (b))
-#define MAGMA_C_MUL(a, b)         cuCmulf((a), (b))
-#define MAGMA_C_DIV(a, b)         cuCdivf((a), (b))
-#define MAGMA_C_ABS(a)            cuCabsf((a))
-#define MAGMA_C_CNJG(a)           cuConjf(a)
-#define MAGMA_C_NEGATE(a)         make_cuFloatComplex( -(a).x, -(a).y )
-#define MAGMA_C_ZERO              make_cuFloatComplex(0.0, 0.0)
-#define MAGMA_C_ONE               make_cuFloatComplex(1.0, 0.0)
-#define MAGMA_C_HALF              make_cuFloatComplex(0.5, 0.0)
-#define MAGMA_C_NEG_ONE           make_cuFloatComplex(-1.0, 0.0)
-#define MAGMA_C_NEG_HALF          make_cuFloatComplex(-0.5, 0.0)
-
-#define MAGMA_D_SET2REAL(v, t)    (v) = (t)
-#define MAGMA_D_OP_NEG_ASGN(t, z) (t) = -(z)
-#define MAGMA_D_EQUAL(u,v)        ((u) == (v))
-#define MAGMA_D_DSCALE(v, t, s)   (v) = (t)/(s)
-#define MAGMA_D_MAKE(r, i)        (r)
-#define MAGMA_D_REAL(a)           (a)
-#define MAGMA_D_IMAG(a)           (a)
-#define MAGMA_D_ADD(a, b)         ( (a) + (b) )
-#define MAGMA_D_SUB(a, b)         ( (a) - (b) )
-#define MAGMA_D_MUL(a, b)         ( (a) * (b) )
-#define MAGMA_D_DIV(a, b)         ( (a) / (b) )
-#define MAGMA_D_ABS(a)            ((a)>0?(a):-(a))
-#define MAGMA_D_CNJG(a)           (a)
-#define MAGMA_D_NEGATE(a)         (-(a))
-#define MAGMA_D_ZERO              (0.0)
-#define MAGMA_D_ONE               (1.0)
-#define MAGMA_D_HALF              (0.5)
-#define MAGMA_D_NEG_ONE           (-1.0)
-#define MAGMA_D_NEG_HALF          (-0.5)
-
-#define MAGMA_S_SET2REAL(v, t)    (v) = (t)
-#define MAGMA_S_OP_NEG_ASGN(t, z) (t) = -(z)
-#define MAGMA_S_EQUAL(u,v)        ((u) == (v))
-#define MAGMA_S_SSCALE(v, t, s)   (v) = (t)/(s)
-#define MAGMA_S_MAKE(r, i)        (r)
-#define MAGMA_S_REAL(a)           (a)
-#define MAGMA_S_IMAG(a)           (a)
-#define MAGMA_S_ADD(a, b)         ( (a) + (b) )
-#define MAGMA_S_SUB(a, b)         ( (a) - (b) )
-#define MAGMA_S_MUL(a, b)         ( (a) * (b) )
-#define MAGMA_S_DIV(a, b)         ( (a) / (b) )
-#define MAGMA_S_ABS(a)            ((a)>0?(a):-(a))
-#define MAGMA_S_CNJG(a)           (a)
-#define MAGMA_S_NEGATE(a)         (-(a))
-#define MAGMA_S_ZERO              (0.0)
-#define MAGMA_S_ONE               (1.0)
-#define MAGMA_S_HALF              (0.5)
-#define MAGMA_S_NEG_ONE           (-1.0)
-#define MAGMA_S_NEG_HALF          (-0.5)
-
-#ifndef CBLAS_SADDR
-#define CBLAS_SADDR(a)  &(a)
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* ------------------------------------------------------------
- *   -- MAGMA function definitions
- * --------------------------------------------------------- */
-
-// ========================================
-// initialization
-void magma_init( void );
-void magma_finalize( void );
-void magma_version( int* major, int* minor, int* micro );
-
-// ========================================
-// memory allocation
-magma_err_t magma_malloc( magma_devptr *ptrPtr, size_t bytes );
-magma_err_t magma_free  ( magma_devptr ptr );
-
-magma_err_t magma_malloc_cpu( void **ptrPtr, size_t bytes );
-magma_err_t magma_free_cpu  ( void *ptr );
-
-magma_err_t magma_malloc_pinned( void **ptrPtr, size_t bytes );
-magma_err_t magma_free_pinned  ( void *ptr );
-
-// type-safe convenience functions to avoid using (void**) cast and sizeof(...)
-// here n is the number of elements (floats, doubles, etc.) not the number of bytes.
-static inline magma_err_t magma_smalloc( float           **ptrPtr, size_t n ) { return magma_malloc( (void**) ptrPtr, n*sizeof(float)           ); }
-static inline magma_err_t magma_dmalloc( double          **ptrPtr, size_t n ) { return magma_malloc( (void**) ptrPtr, n*sizeof(double)          ); }
-static inline magma_err_t magma_cmalloc( cuFloatComplex  **ptrPtr, size_t n ) { return magma_malloc( (void**) ptrPtr, n*sizeof(cuFloatComplex)  ); }
-static inline magma_err_t magma_zmalloc( cuDoubleComplex **ptrPtr, size_t n ) { return magma_malloc( (void**) ptrPtr, n*sizeof(cuDoubleComplex) ); }
-
-static inline magma_err_t magma_smalloc_cpu( float           **ptrPtr, size_t n ) { return magma_malloc_cpu( (void**) ptrPtr, n*sizeof(float)           ); }
-static inline magma_err_t magma_dmalloc_cpu( double          **ptrPtr, size_t n ) { return magma_malloc_cpu( (void**) ptrPtr, n*sizeof(double)          ); }
-static inline magma_err_t magma_cmalloc_cpu( cuFloatComplex  **ptrPtr, size_t n ) { return magma_malloc_cpu( (void**) ptrPtr, n*sizeof(cuFloatComplex)  ); }
-static inline magma_err_t magma_zmalloc_cpu( cuDoubleComplex **ptrPtr, size_t n ) { return magma_malloc_cpu( (void**) ptrPtr, n*sizeof(cuDoubleComplex) ); }
-
-static inline magma_err_t magma_smalloc_pinned( float           **ptrPtr, size_t n ) { return magma_malloc_pinned( (void**) ptrPtr, n*sizeof(float)           ); }
-static inline magma_err_t magma_dmalloc_pinned( double          **ptrPtr, size_t n ) { return magma_malloc_pinned( (void**) ptrPtr, n*sizeof(double)          ); }
-static inline magma_err_t magma_cmalloc_pinned( cuFloatComplex  **ptrPtr, size_t n ) { return magma_malloc_pinned( (void**) ptrPtr, n*sizeof(cuFloatComplex)  ); }
-static inline magma_err_t magma_zmalloc_pinned( cuDoubleComplex **ptrPtr, size_t n ) { return magma_malloc_pinned( (void**) ptrPtr, n*sizeof(cuDoubleComplex) ); }
-
-
-// ========================================
-// device & queue support
-void magma_getdevices(
-    magma_device_t* devices,
-    magma_int_t     size,
-    magma_int_t*    numPtr );
-
-void magma_getdevice( magma_device_t* dev );
-
-void magma_setdevice( magma_device_t dev );
-
-void magma_device_sync();
-
-
-// ========================================
-// queue support
-void magma_queue_create( /*magma_device_t device,*/ magma_queue_t* queuePtr );
-
-void magma_queue_destroy( magma_queue_t queue );
-
-void magma_queue_sync( magma_queue_t queue );
-
-
-// ========================================
-// event support
-void magma_event_create( magma_event_t* eventPtr );
-
-void magma_event_destroy( magma_event_t event );
-
-void magma_event_record( magma_event_t event, magma_queue_t queue );
-
-// blocks CPU until event occurs
-void magma_event_sync( magma_event_t event );
-
-// blocks queue (but not CPU) until event occurs
-void magma_queue_wait_event( magma_queue_t queue, magma_event_t event );
-
-// magma GPU-complex connection
-int magma_buildconnection_mgpu(  magma_int_t gnode[MagmaMaxGPUs+2][MagmaMaxGPUs+2], magma_int_t *nbcmplx, magma_int_t ngpu);
-
-
-
-// ========================================
-// generic, type-independent routines to copy data.
-// type-safe versions which avoid the user needing sizeof(...) are in [sdcz]set_get.cpp
-void magma_setvector(
-    magma_int_t n, size_t elemSize,
-    void const *hx_src, magma_int_t incx,
-    void       *dy_dst, magma_int_t incy );
-
-void magma_getvector(
-    magma_int_t n, size_t elemSize,
-    void const *dx_src, magma_int_t incx,
-    void       *hy_dst, magma_int_t incy );
-
-void magma_setvector_async(
-    magma_int_t n, size_t elemSize,
-    void const *hx_src, magma_int_t incx,
-    void       *dy_dst, magma_int_t incy,
-    magma_stream_t stream );
-
-void magma_getvector_async(
-    magma_int_t n, size_t elemSize,
-    void const *dx_src, magma_int_t incx,
-    void       *hy_dst, magma_int_t incy,
-    magma_stream_t stream );
-
-
-// ========================================
-// error handler
-void magma_xerbla( const char *name, magma_int_t info );
-
-/* ------------------------------------------------------------
- *   -- MAGMA Auxiliary structures and functions
- * --------------------------------------------------------- */
-typedef struct magma_timestr_s
-{
-  unsigned int sec;
-  unsigned int usec;
-} magma_timestr_t;
-
-magma_timestr_t get_current_time(void);
-double GetTimerValue(magma_timestr_t time_1, magma_timestr_t time_2);
-void printout_devices();
-void swp2pswp(char trans, magma_int_t n, magma_int_t *ipiv, magma_int_t *newipiv);
-float getv(float *da);
-
-double magma_wtime( void );
-double magma_sync_wtime( magma_queue_t queue );
-size_t magma_strlcpy(char *dst, const char *src, size_t siz);
-int magma_num_gpus( void );
-int magma_is_devptr( const void* A );
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
diff --git a/src/include/magma_bulge.h b/src/include/magma_bulge.h
deleted file mode 100644
index 0fbbb1c..0000000
--- a/src/include/magma_bulge.h
+++ /dev/null
@@ -1,40 +0,0 @@
-/*
-    -- MAGMA (version 1.3.0) --
-       Univ. of Tennessee, Knoxville
-       Univ. of California, Berkeley
-       Univ. of Colorado, Denver
-       November 2012
-*/
-
-#ifndef _MAGMA_BULGE_H_
-#define _MAGMA_BULGE_H_
-
-#ifdef __cplusplus
-extern "C"{
-#endif
-
-    inline magma_int_t magma_ceildiv(magma_int_t a, magma_int_t b)
-    {
-        return (a+b-1)/b;
-    }
-
-    magma_int_t magma_bulge_get_nb(magma_int_t n);
-
-    void cmp_vals(int n, double *wr1, double *wr2, double *nrmI, double *nrm1, double *nrm2);
-
-    void magma_bulge_findVTAUpos(magma_int_t n, magma_int_t nb, magma_int_t Vblksiz, magma_int_t sweep, magma_int_t st, magma_int_t ldv,
-                                 magma_int_t *Vpos, magma_int_t *TAUpos);
-
-    void magma_bulge_findVTpos(magma_int_t n, magma_int_t nb, magma_int_t Vblksiz, magma_int_t sweep, magma_int_t st, magma_int_t ldv, magma_int_t ldt,
-                               magma_int_t *Vpos, magma_int_t *Tpos);
-
-    void magma_bulge_findVTAUTpos(magma_int_t n, magma_int_t nb, magma_int_t Vblksiz, magma_int_t sweep, magma_int_t st, magma_int_t ldv, magma_int_t ldt,
-                                  magma_int_t *Vpos, magma_int_t *TAUpos, magma_int_t *Tpos, magma_int_t *blkid);
-
-    magma_int_t magma_bulge_get_blkcnt(magma_int_t n, magma_int_t nb, magma_int_t Vblksiz);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
diff --git a/src/include/magma_c.h b/src/include/magma_c.h
deleted file mode 100644
index 5a114fc..0000000
--- a/src/include/magma_c.h
+++ /dev/null
@@ -1,514 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated c Wed Nov 14 22:52:26 2012
- */
-
-#ifndef _MAGMA_C_H_
-#define _MAGMA_C_H_
-#define PRECISION_c
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- MAGMA Auxiliary functions to get the NB used
-*/
-int magma_get_cpotrf_nb(int m);
-int magma_get_cgetrf_nb(int m);
-int magma_get_cgetri_nb(int m);
-int magma_get_cgeqp3_nb(int m);
-int magma_get_cgeqrf_nb(int m);
-int magma_get_cgeqlf_nb(int m);
-int magma_get_cgehrd_nb(int m);
-int magma_get_chetrd_nb(int m);
-int magma_get_cgelqf_nb(int m);
-int magma_get_cgebrd_nb(int m);
-int magma_get_chegst_nb(int m);
-int magma_get_cgesvd_nb(int m);
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- MAGMA function definitions / Data on CPU
-*/
-magma_int_t magma_cgebrd( magma_int_t m, magma_int_t n, cuFloatComplex *A, 
-                          magma_int_t lda, float *d, float *e,
-                          cuFloatComplex *tauq,  cuFloatComplex *taup, 
-                          cuFloatComplex *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_cgehrd2(magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          cuFloatComplex *A, magma_int_t lda, cuFloatComplex *tau, 
-                          cuFloatComplex *work, magma_int_t *lwork, magma_int_t *info);
-magma_int_t magma_cgehrd( magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          cuFloatComplex *A, magma_int_t lda, cuFloatComplex *tau,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          cuFloatComplex *d_T, magma_int_t *info);
-magma_int_t magma_cgelqf( magma_int_t m, magma_int_t n, 
-                          cuFloatComplex *A,    magma_int_t lda,   cuFloatComplex *tau, 
-                          cuFloatComplex *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_cgeqlf( magma_int_t m, magma_int_t n, 
-                          cuFloatComplex *A,    magma_int_t lda,   cuFloatComplex *tau, 
-                          cuFloatComplex *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_cgeqrf( magma_int_t m, magma_int_t n, cuFloatComplex *A, 
-                          magma_int_t lda, cuFloatComplex *tau, cuFloatComplex *work, 
-                          magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_cgeqr2( magma_int_t *m, magma_int_t *n, cuFloatComplex *a,
-                          magma_int_t *lda, cuFloatComplex *tau, cuFloatComplex *work, 
-                          magma_int_t *info);
-magma_int_t magma_cgeqrf4(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
-                          cuFloatComplex *a,    magma_int_t lda, cuFloatComplex *tau,
-                          cuFloatComplex *work, magma_int_t lwork, magma_int_t *info );
-magma_int_t magma_cgeqrf_ooc( magma_int_t m, magma_int_t n, cuFloatComplex *A,
-                          magma_int_t lda, cuFloatComplex *tau, cuFloatComplex *work,
-                          magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_cgesv ( magma_int_t n, magma_int_t nrhs, 
-                          cuFloatComplex *A, magma_int_t lda, magma_int_t *ipiv, 
-                          cuFloatComplex *B, magma_int_t ldb, magma_int_t *info);
-magma_int_t magma_cgetrf( magma_int_t m, magma_int_t n, cuFloatComplex *A, 
-                          magma_int_t lda, magma_int_t *ipiv, 
-                          magma_int_t *info);
-magma_int_t magma_cgetrf2(magma_int_t m, magma_int_t n, cuFloatComplex *a, 
-                          magma_int_t lda, magma_int_t *ipiv, magma_int_t *info);
-
-magma_int_t magma_claqps( magma_int_t m, magma_int_t n, magma_int_t offset, 
-                          magma_int_t nb, magma_int_t *kb, 
-                          cuFloatComplex *A,  magma_int_t lda,
-                          cuFloatComplex *dA, magma_int_t ldda,
-                          magma_int_t *jpvt, cuFloatComplex *tau, float *vn1, float *vn2, 
-                          cuFloatComplex *auxv, 
-                          cuFloatComplex *F,  magma_int_t ldf,
-                          cuFloatComplex *dF, magma_int_t lddf );
-magma_int_t magma_clarf(  char *side, magma_int_t *m, magma_int_t *n,
-                          cuFloatComplex *v, magma_int_t *incv, cuFloatComplex *tau,
-                          cuFloatComplex *c__, magma_int_t *ldc, cuFloatComplex *work);
-magma_int_t magma_clarfg( magma_int_t *n, cuFloatComplex *alpha, cuFloatComplex *x,
-                          magma_int_t *incx, cuFloatComplex *tau);
-magma_int_t magma_clatrd( char uplo, magma_int_t n, magma_int_t nb, cuFloatComplex *a, 
-                          magma_int_t lda, float *e, cuFloatComplex *tau, 
-                          cuFloatComplex *w, magma_int_t ldw,
-                          cuFloatComplex *da, magma_int_t ldda, 
-                          cuFloatComplex *dw, magma_int_t lddw);
-magma_int_t magma_clatrd2(char uplo, magma_int_t n, magma_int_t nb,
-                          cuFloatComplex *a,  magma_int_t lda,
-                          float *e, cuFloatComplex *tau,
-                          cuFloatComplex *w,  magma_int_t ldw,
-                          cuFloatComplex *da, magma_int_t ldda,
-                          cuFloatComplex *dw, magma_int_t lddw,
-                          cuFloatComplex *dwork, magma_int_t ldwork);
-magma_int_t magma_clahr2( magma_int_t m, magma_int_t n, magma_int_t nb, 
-                          cuFloatComplex *da, cuFloatComplex *dv, cuFloatComplex *a, 
-                          magma_int_t lda, cuFloatComplex *tau, cuFloatComplex *t, 
-                          magma_int_t ldt, cuFloatComplex *y, magma_int_t ldy);
-magma_int_t magma_clahru( magma_int_t n, magma_int_t ihi, magma_int_t k, magma_int_t nb, 
-                          cuFloatComplex *a, magma_int_t lda, 
-                          cuFloatComplex *da, cuFloatComplex *y, 
-                          cuFloatComplex *v, cuFloatComplex *t, 
-                          cuFloatComplex *dwork);
-magma_int_t magma_cposv ( char uplo, magma_int_t n, magma_int_t nrhs, 
-                          cuFloatComplex *A, magma_int_t lda, 
-                          cuFloatComplex *B, magma_int_t ldb, magma_int_t *info);
-magma_int_t magma_cpotrf( char uplo, magma_int_t n, cuFloatComplex *A, 
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_cpotri( char uplo, magma_int_t n, cuFloatComplex *A,
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_clauum( char uplo, magma_int_t n, cuFloatComplex *A,
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_ctrtri( char uplo, char diag, magma_int_t n, cuFloatComplex *A, 
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_chetrd( char uplo, magma_int_t n, cuFloatComplex *A, 
-                          magma_int_t lda, float *d, float *e, 
-                          cuFloatComplex *tau, cuFloatComplex *work, magma_int_t lwork, 
-                          magma_int_t *info);
-magma_int_t magma_cungqr( magma_int_t m, magma_int_t n, magma_int_t k,
-                          cuFloatComplex *a, magma_int_t lda,
-                          cuFloatComplex *tau, cuFloatComplex *dwork,
-                          magma_int_t nb, magma_int_t *info );
-magma_int_t magma_cunmql( const char side, const char trans,
-                          magma_int_t m, magma_int_t n, magma_int_t k,
-                          cuFloatComplex *a, magma_int_t lda,
-                          cuFloatComplex *tau,
-                          cuFloatComplex *c, magma_int_t ldc,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_cunmqr( char side, char trans, 
-                          magma_int_t m, magma_int_t n, magma_int_t k, 
-                          cuFloatComplex *a, magma_int_t lda, cuFloatComplex *tau, 
-                          cuFloatComplex *c, magma_int_t ldc, 
-                          cuFloatComplex *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_cunmtr( char side, char uplo, char trans,
-                          magma_int_t m, magma_int_t n,
-                          cuFloatComplex *a,    magma_int_t lda,
-                          cuFloatComplex *tau,
-                          cuFloatComplex *c,    magma_int_t ldc,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_cunghr( magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          cuFloatComplex *a, magma_int_t lda,
-                          cuFloatComplex *tau,
-                          cuFloatComplex *dT, magma_int_t nb,
-                          magma_int_t *info);
-magma_int_t magma_cheev( char jobz, char uplo, magma_int_t n,
-                         cuFloatComplex *a, magma_int_t lda, float *w,
-                         cuFloatComplex *work, magma_int_t lwork,
-                         float *rwork, magma_int_t *info);
-magma_int_t magma_cheevx(char jobz, char range, char uplo, magma_int_t n,
-                         cuFloatComplex *a, magma_int_t lda, float vl, float vu,
-                         magma_int_t il, magma_int_t iu, float abstol, magma_int_t *m,
-                         float *w, cuFloatComplex *z, magma_int_t ldz, 
-                         cuFloatComplex *work, magma_int_t lwork,
-                         float *rwork, magma_int_t *iwork, 
-                         magma_int_t *ifail, magma_int_t *info);
-#if defined(PRECISION_z) || defined(PRECISION_c)
-magma_int_t  magma_cgeev( char jobvl, char jobvr, magma_int_t n,
-                          cuFloatComplex *a, magma_int_t lda,
-                          cuFloatComplex *w,
-                          cuFloatComplex *vl, magma_int_t ldvl,
-                          cuFloatComplex *vr, magma_int_t ldvr,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          float *rwork, magma_int_t *info);
-magma_int_t magma_cgeqp3( magma_int_t m, magma_int_t n,
-                          cuFloatComplex *a, magma_int_t lda, 
-                          magma_int_t *jpvt, cuFloatComplex *tau,
-                          cuFloatComplex *work, magma_int_t lwork, 
-                          float *rwork, magma_int_t *info);
-magma_int_t magma_cgesvd( char jobu, char jobvt, magma_int_t m, magma_int_t n,
-                          cuFloatComplex *a,    magma_int_t lda, float *s, 
-                          cuFloatComplex *u,    magma_int_t ldu, 
-                          cuFloatComplex *vt,   magma_int_t ldvt,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          float *rwork, magma_int_t *info );
-magma_int_t magma_cheevd( char jobz, char uplo, magma_int_t n,
-                          cuFloatComplex *a, magma_int_t lda, float *w,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          float *rwork, magma_int_t lrwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_cheevr( char jobz, char range, char uplo, magma_int_t n,
-                          cuFloatComplex *a, magma_int_t lda, float vl, float vu,
-                          magma_int_t il, magma_int_t iu, float abstol, magma_int_t *m,
-                          float *w, cuFloatComplex *z, magma_int_t ldz, 
-                          magma_int_t *isuppz,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          float *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_chegvd( magma_int_t itype, char jobz, char uplo, magma_int_t n,
-                          cuFloatComplex *a, magma_int_t lda,
-                          cuFloatComplex *b, magma_int_t ldb,
-                          float *w, cuFloatComplex *work, magma_int_t lwork,
-                          float *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_chegvdx(magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, cuFloatComplex *a, magma_int_t lda,
-                          cuFloatComplex *b, magma_int_t ldb,
-                          float vl, float vu, magma_int_t il, magma_int_t iu,
-                          magma_int_t *m, float *w, cuFloatComplex *work, 
-                          magma_int_t lwork, float *rwork,
-                          magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_chegvx( magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, cuFloatComplex *a, magma_int_t lda, 
-                          cuFloatComplex *b, magma_int_t ldb,
-                          float vl, float vu, magma_int_t il, magma_int_t iu,
-                          float abstol, magma_int_t *m, float *w, 
-                          cuFloatComplex *z, magma_int_t ldz,
-                          cuFloatComplex *work, magma_int_t lwork, float *rwork,
-                          magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info);
-magma_int_t magma_chegvr( magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, cuFloatComplex *a, magma_int_t lda,
-                          cuFloatComplex *b, magma_int_t ldb,
-                          float vl, float vu, magma_int_t il, magma_int_t iu,
-                          float abstol, magma_int_t *m, float *w, 
-                          cuFloatComplex *z, magma_int_t ldz,
-                          magma_int_t *isuppz, cuFloatComplex *work, magma_int_t lwork,
-                          float *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_cstedx( char range, magma_int_t n, float vl, float vu,
-                          magma_int_t il, magma_int_t iu, float *D, float *E,
-                          cuFloatComplex *Z, magma_int_t ldz,
-                          float *rwork, magma_int_t ldrwork, magma_int_t *iwork,
-                          magma_int_t liwork, float* dwork, magma_int_t *info);
-#else
-magma_int_t  magma_cgeev( char jobvl, char jobvr, magma_int_t n,
-                          cuFloatComplex *a,    magma_int_t lda,
-                          cuFloatComplex *wr, cuFloatComplex *wi,
-                          cuFloatComplex *vl,   magma_int_t ldvl,
-                          cuFloatComplex *vr,   magma_int_t ldvr,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_cgeqp3( magma_int_t m, magma_int_t n,
-                          cuFloatComplex *a, magma_int_t lda, 
-                          magma_int_t *jpvt, cuFloatComplex *tau,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_cgesvd( char jobu, char jobvt, magma_int_t m, magma_int_t n,
-                          cuFloatComplex *a,    magma_int_t lda, float *s, 
-                          cuFloatComplex *u,    magma_int_t ldu, 
-                          cuFloatComplex *vt,   magma_int_t ldvt,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          magma_int_t *info );
-magma_int_t magma_cheevd( char jobz, char uplo, magma_int_t n,
-                          cuFloatComplex *a, magma_int_t lda, float *w,
-                          cuFloatComplex *work, magma_int_t lwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_chegvd( magma_int_t itype, char jobz, char uplo, magma_int_t n,
-                          cuFloatComplex *a, magma_int_t lda,
-                          cuFloatComplex *b, magma_int_t ldb,
-                          float *w, cuFloatComplex *work, magma_int_t lwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_cstedx( char range, magma_int_t n, float vl, float vu,
-                          magma_int_t il, magma_int_t iu, float* d, float* e,
-                          float* z, magma_int_t ldz,
-                          float* work, magma_int_t lwork,
-                          magma_int_t* iwork, magma_int_t liwork,
-                          float* dwork, magma_int_t* info);
-magma_int_t magma_claex0( magma_int_t n, float* d, float* e, float* q, magma_int_t ldq,
-                          float* work, magma_int_t* iwork, float* dwork,
-                          char range, float vl, float vu,
-                          magma_int_t il, magma_int_t iu, magma_int_t* info);
-magma_int_t magma_claex1( magma_int_t n, float* d, float* q, magma_int_t ldq,
-                          magma_int_t* indxq, float rho, magma_int_t cutpnt,
-                          float* work, magma_int_t* iwork, float* dwork,
-                          char range, float vl, float vu,
-                          magma_int_t il, magma_int_t iu, magma_int_t* info);
-magma_int_t magma_claex3( magma_int_t k, magma_int_t n, magma_int_t n1, float* d,
-                          float* q, magma_int_t ldq, float rho,
-                          float* dlamda, float* q2, magma_int_t* indx,
-                          magma_int_t* ctot, float* w, float* s, magma_int_t* indxq,
-                          float* dwork,
-                          char range, float vl, float vu, magma_int_t il, magma_int_t iu,
-                          magma_int_t* info );
-#endif
-
-magma_int_t magma_chegst( magma_int_t itype, char uplo, magma_int_t n,
-                          cuFloatComplex *a, magma_int_t lda,
-                          cuFloatComplex *b, magma_int_t ldb, magma_int_t *info);
-
-/* //////////////////////////////////////////////////////////////////////////// 
- -- MAGMA function definitions / Data on GPU
-*/
-magma_int_t magma_cgels_gpu(  char trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
-                              cuFloatComplex *dA,    magma_int_t ldda, 
-                              cuFloatComplex *dB,    magma_int_t lddb, 
-                              cuFloatComplex *hwork, magma_int_t lwork, 
-                              magma_int_t *info);
-magma_int_t magma_cgels3_gpu( char trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
-                              cuFloatComplex *dA,    magma_int_t ldda,
-                              cuFloatComplex *dB,    magma_int_t lddb,
-                              cuFloatComplex *hwork, magma_int_t lwork,
-                              magma_int_t *info);
-magma_int_t magma_cgelqf_gpu( magma_int_t m, magma_int_t n,
-                              cuFloatComplex *dA,    magma_int_t ldda,   cuFloatComplex *tau,
-                              cuFloatComplex *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_cgeqrf_gpu( magma_int_t m, magma_int_t n, 
-                              cuFloatComplex *dA,  magma_int_t ldda, 
-                              cuFloatComplex *tau, cuFloatComplex *dT, 
-                              magma_int_t *info);
-magma_int_t magma_cgeqrf2_gpu(magma_int_t m, magma_int_t n, 
-                              cuFloatComplex *dA,  magma_int_t ldda, 
-                              cuFloatComplex *tau, magma_int_t *info);
-magma_int_t magma_cgeqrf2_mgpu(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
-                               cuFloatComplex **dlA, magma_int_t ldda,
-                               cuFloatComplex *tau, magma_int_t *info );
-magma_int_t magma_cgeqrf3_gpu(magma_int_t m, magma_int_t n, 
-                              cuFloatComplex *dA,  magma_int_t ldda, 
-                              cuFloatComplex *tau, cuFloatComplex *dT, 
-                              magma_int_t *info);
-magma_int_t magma_cgeqrs_gpu( magma_int_t m, magma_int_t n, magma_int_t nrhs, 
-                              cuFloatComplex *dA,     magma_int_t ldda, 
-                              cuFloatComplex *tau,   cuFloatComplex *dT,
-                              cuFloatComplex *dB,    magma_int_t lddb,
-                              cuFloatComplex *hwork, magma_int_t lhwork, 
-                              magma_int_t *info);
-magma_int_t magma_cgeqrs3_gpu( magma_int_t m, magma_int_t n, magma_int_t nrhs, 
-                              cuFloatComplex *dA,     magma_int_t ldda, 
-                              cuFloatComplex *tau,   cuFloatComplex *dT,
-                              cuFloatComplex *dB,    magma_int_t lddb,
-                              cuFloatComplex *hwork, magma_int_t lhwork, 
-                              magma_int_t *info);
-magma_int_t magma_cgessm_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t k, magma_int_t ib, 
-                              magma_int_t *ipiv, 
-                              cuFloatComplex *dL1, magma_int_t lddl1, 
-                              cuFloatComplex *dL,  magma_int_t lddl, 
-                              cuFloatComplex *dA,  magma_int_t ldda, 
-                              magma_int_t *info);
-magma_int_t magma_cgesv_gpu(  magma_int_t n, magma_int_t nrhs, 
-                              cuFloatComplex *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              cuFloatComplex *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_cgetrf_incpiv_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t ib,
-                              cuFloatComplex *hA, magma_int_t ldha, cuFloatComplex *dA, magma_int_t ldda,
-                              cuFloatComplex *hL, magma_int_t ldhl, cuFloatComplex *dL, magma_int_t lddl,
-                              magma_int_t *ipiv, 
-                              cuFloatComplex *dwork, magma_int_t lddwork,
-                              magma_int_t *info);
-magma_int_t magma_cgetrf_gpu( magma_int_t m, magma_int_t n, 
-                              cuFloatComplex *dA, magma_int_t ldda, 
-                              magma_int_t *ipiv, magma_int_t *info);
-magma_int_t 
-magma_cgetrf_nopiv_gpu      ( magma_int_t m, magma_int_t n,
-                              cuFloatComplex *dA, magma_int_t ldda,
-                              magma_int_t *info);
-magma_int_t magma_cgetri_gpu( magma_int_t n, 
-                              cuFloatComplex *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              cuFloatComplex *dwork, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_cgetrs_gpu( char trans, magma_int_t n, magma_int_t nrhs, 
-                              cuFloatComplex *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              cuFloatComplex *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_clabrd_gpu( magma_int_t m, magma_int_t n, magma_int_t nb, 
-                              cuFloatComplex *a, magma_int_t lda, cuFloatComplex *da, magma_int_t ldda,
-                              float *d, float *e, cuFloatComplex *tauq, cuFloatComplex *taup,  
-                              cuFloatComplex *x, magma_int_t ldx, cuFloatComplex *dx, magma_int_t lddx, 
-                              cuFloatComplex *y, magma_int_t ldy, cuFloatComplex *dy, magma_int_t lddy);
-magma_int_t magma_clarfb_gpu( char side, char trans, char direct, char storev, 
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              const cuFloatComplex *dv, magma_int_t ldv,
-                              const cuFloatComplex *dt, magma_int_t ldt, 
-                              cuFloatComplex *dc,       magma_int_t ldc,
-                              cuFloatComplex *dwork,    magma_int_t ldwork );
-magma_int_t magma_cposv_gpu(  char uplo, magma_int_t n, magma_int_t nrhs, 
-                              cuFloatComplex *dA, magma_int_t ldda, 
-                              cuFloatComplex *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_cpotrf_gpu( char uplo,  magma_int_t n, 
-                              cuFloatComplex *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_cpotri_gpu( char uplo,  magma_int_t n,
-                              cuFloatComplex *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_clauum_gpu( char uplo,  magma_int_t n,
-                              cuFloatComplex *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_ctrtri_gpu( char uplo,  char diag, magma_int_t n,
-                              cuFloatComplex *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_chetrd_gpu( char uplo, magma_int_t n,
-                              cuFloatComplex *da, magma_int_t ldda,
-                              float *d, float *e, cuFloatComplex *tau,
-                              cuFloatComplex *wa,  magma_int_t ldwa,
-                              cuFloatComplex *work, magma_int_t lwork,
-                              magma_int_t *info);
-magma_int_t magma_chetrd2_gpu(char uplo, magma_int_t n,
-                              cuFloatComplex *da, magma_int_t ldda,
-                              float *d, float *e, cuFloatComplex *tau,
-                              cuFloatComplex *wa,  magma_int_t ldwa,
-                              cuFloatComplex *work, magma_int_t lwork,
-                              cuFloatComplex *dwork, magma_int_t ldwork,
-                              magma_int_t *info);
-magma_int_t magma_cpotrs_gpu( char uplo,  magma_int_t n, magma_int_t nrhs, 
-                              cuFloatComplex *dA, magma_int_t ldda, 
-                              cuFloatComplex *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_cssssm_gpu( char storev, magma_int_t m1, magma_int_t n1, 
-                              magma_int_t m2, magma_int_t n2, magma_int_t k, magma_int_t ib, 
-                              cuFloatComplex *dA1, magma_int_t ldda1, 
-                              cuFloatComplex *dA2, magma_int_t ldda2, 
-                              cuFloatComplex *dL1, magma_int_t lddl1, 
-                              cuFloatComplex *dL2, magma_int_t lddl2,
-                              magma_int_t *IPIV, magma_int_t *info);
-magma_int_t magma_ctstrf_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t ib, magma_int_t nb,
-                              cuFloatComplex *hU, magma_int_t ldhu, cuFloatComplex *dU, magma_int_t lddu, 
-                              cuFloatComplex *hA, magma_int_t ldha, cuFloatComplex *dA, magma_int_t ldda, 
-                              cuFloatComplex *hL, magma_int_t ldhl, cuFloatComplex *dL, magma_int_t lddl,
-                              magma_int_t *ipiv, 
-                              cuFloatComplex *hwork, magma_int_t ldhwork, 
-                              cuFloatComplex *dwork, magma_int_t lddwork,
-                              magma_int_t *info);
-magma_int_t magma_cungqr_gpu( magma_int_t m, magma_int_t n, magma_int_t k, 
-                              cuFloatComplex *da, magma_int_t ldda, 
-                              cuFloatComplex *tau, cuFloatComplex *dwork, 
-                              magma_int_t nb, magma_int_t *info );
-magma_int_t magma_cunmql2_gpu(const char side, const char trans,
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              cuFloatComplex *da, magma_int_t ldda,
-                              cuFloatComplex *tau,
-                              cuFloatComplex *dc, magma_int_t lddc,
-                              cuFloatComplex *wa, magma_int_t ldwa,
-                              magma_int_t *info);
-magma_int_t magma_cunmqr_gpu( char side, char trans, 
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              cuFloatComplex *a,    magma_int_t lda, cuFloatComplex *tau, 
-                              cuFloatComplex *c,    magma_int_t ldc,
-                              cuFloatComplex *work, magma_int_t lwork, 
-                              cuFloatComplex *td,   magma_int_t nb, magma_int_t *info);
-magma_int_t magma_cunmqr2_gpu(const char side, const char trans,
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              cuFloatComplex *da,   magma_int_t ldda,
-                              cuFloatComplex *tau,
-                              cuFloatComplex *dc,    magma_int_t lddc,
-                              cuFloatComplex *wa,    magma_int_t ldwa,
-                              magma_int_t *info);
-magma_int_t magma_cunmtr_gpu( char side, char uplo, char trans,
-                              magma_int_t m, magma_int_t n,
-                              cuFloatComplex *da,    magma_int_t ldda,
-                              cuFloatComplex *tau,
-                              cuFloatComplex *dc,    magma_int_t lddc,
-                              cuFloatComplex *wa,    magma_int_t ldwa,
-                              magma_int_t *info);
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-magma_int_t magma_cheevd_gpu( char jobz, char uplo,
-                              magma_int_t n,
-                              cuFloatComplex *da, magma_int_t ldda,
-                              float *w,
-                              cuFloatComplex *wa,  magma_int_t ldwa,
-                              cuFloatComplex *work, magma_int_t lwork,
-                              float *rwork, magma_int_t lrwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-magma_int_t magma_cheevdx_gpu(char jobz, char range, char uplo,
-                              magma_int_t n, cuFloatComplex *da, 
-                              magma_int_t ldda, float vl, float vu, 
-                              magma_int_t il, magma_int_t iu,
-                              magma_int_t *m, float *w,
-                              cuFloatComplex *wa,  magma_int_t ldwa,
-                              cuFloatComplex *work, magma_int_t lwork,
-                              float *rwork, magma_int_t lrwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-magma_int_t magma_cheevr_gpu( char jobz, char range, char uplo, magma_int_t n,
-                              cuFloatComplex *da, magma_int_t ldda, float vl, float vu,
-                              magma_int_t il, magma_int_t iu, float abstol, magma_int_t *m,
-                              float *w, cuFloatComplex *dz, magma_int_t lddz,
-                              magma_int_t *isuppz,
-                              cuFloatComplex *wa, magma_int_t ldwa,
-                              cuFloatComplex *wz, magma_int_t ldwz,
-                              cuFloatComplex *work, magma_int_t lwork,
-                              float *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                              magma_int_t liwork, magma_int_t *info);
-#else
-magma_int_t magma_cheevd_gpu( char jobz, char uplo,
-                              magma_int_t n,
-                              cuFloatComplex *da, magma_int_t ldda,
-                              cuFloatComplex *w,
-                              cuFloatComplex *wa,  magma_int_t ldwa,
-                              cuFloatComplex *work, magma_int_t lwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-#endif
-
-magma_int_t magma_cheevx_gpu( char jobz, char range, char uplo, magma_int_t n,
-                              cuFloatComplex *da, magma_int_t ldda, float vl, 
-                              float vu, magma_int_t il, magma_int_t iu, 
-                              float abstol, magma_int_t *m,
-                              float *w, cuFloatComplex *dz, magma_int_t lddz,
-                              cuFloatComplex *wa, magma_int_t ldwa,
-                              cuFloatComplex *wz, magma_int_t ldwz,
-                              cuFloatComplex *work, magma_int_t lwork,
-                              float *rwork, magma_int_t *iwork, 
-                              magma_int_t *ifail, magma_int_t *info);
-magma_int_t magma_chegst_gpu(magma_int_t itype, char uplo, magma_int_t n,
-                             cuFloatComplex *da, magma_int_t ldda,
-                             cuFloatComplex *db, magma_int_t lddb, magma_int_t *info);
-
-
-/* //////////////////////////////////////////////////////////////////////////// 
- -- MAGMA utility function definitions
-*/
-
-void magma_cprint    ( magma_int_t m, magma_int_t n, const cuFloatComplex  *A, magma_int_t lda  );
-void magma_cprint_gpu( magma_int_t m, magma_int_t n, const cuFloatComplex *dA, magma_int_t ldda );
-
-void cpanel_to_q( char uplo, magma_int_t ib, cuFloatComplex *A, magma_int_t lda, cuFloatComplex *work );
-void cq_to_panel( char uplo, magma_int_t ib, cuFloatComplex *A, magma_int_t lda, cuFloatComplex *work );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_c
-#endif /* _MAGMA_C_H_ */
diff --git a/src/include/magma_cbulge.h b/src/include/magma_cbulge.h
deleted file mode 100644
index 2c4d808..0000000
--- a/src/include/magma_cbulge.h
+++ /dev/null
@@ -1,35 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated c Wed Nov 14 22:52:27 2012
- */
-
-#ifndef MAGMA_CBULGEINC_H
-#define MAGMA_CBULGEINC_H
-
-#define PRECISION_c
-
-extern "C" {
-    magma_int_t magma_chetrd_he2hb(char uplo, magma_int_t n, magma_int_t NB, cuFloatComplex *a, magma_int_t lda,
-                                   cuFloatComplex *tau, cuFloatComplex *work, magma_int_t lwork, cuFloatComplex *dT, magma_int_t threads, magma_int_t *info);
-
-    magma_int_t magma_chetrd_hb2st(magma_int_t threads, char uplo, magma_int_t n, magma_int_t nb, magma_int_t Vblksiz,
-                                   cuFloatComplex *A, magma_int_t lda, float *D, float *E,
-                                   cuFloatComplex *V, magma_int_t ldv, cuFloatComplex *TAU, magma_int_t compT, cuFloatComplex *T, magma_int_t ldt);
-
-    magma_int_t magma_cbulge_back(magma_int_t threads, char uplo, magma_int_t n, magma_int_t nb, magma_int_t ne, magma_int_t Vblksiz,
-                                  cuFloatComplex *Z, magma_int_t ldz, cuFloatComplex *dZ, magma_int_t lddz,
-                                  cuFloatComplex *V, magma_int_t ldv, cuFloatComplex *TAU, cuFloatComplex *T, magma_int_t ldt, magma_int_t* info);
-
-    magma_int_t magma_cunmqr_gpu_2stages(char side, char trans, magma_int_t m, magma_int_t n, magma_int_t k, cuFloatComplex *dA, magma_int_t ldda,
-                                         cuFloatComplex *dC, magma_int_t lddc, cuFloatComplex *dT, magma_int_t nb, magma_int_t *info);
-
-    magma_int_t magma_cbulge_get_lq2(magma_int_t n);
-
-    magma_int_t magma_cbulge_get_Vblksiz(magma_int_t n, magma_int_t nb);
-}
-#endif
diff --git a/src/include/magma_cbulgeinc.h b/src/include/magma_cbulgeinc.h
deleted file mode 100644
index c5a8165..0000000
--- a/src/include/magma_cbulgeinc.h
+++ /dev/null
@@ -1,141 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated c Wed Nov 14 22:52:28 2012
- */
-
-#ifndef MAGMA_CBULGEINC_H
-#define MAGMA_CBULGEINC_H
-
-#define PRECISION_c
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-
-/***************************************************************************//**
- *  Configuration
- **/
-
- // maximum contexts
-#define MAX_THREADS_BLG         256
-
-void findVTpos(magma_int_t N, magma_int_t NB, magma_int_t Vblksiz, magma_int_t sweep, magma_int_t st, magma_int_t *Vpos, magma_int_t *TAUpos, magma_int_t *Tpos, magma_int_t *myblkid);
-void findVTsiz(magma_int_t N, magma_int_t NB, magma_int_t Vblksiz, magma_int_t *blkcnt, magma_int_t *LDV);
-magma_int_t plasma_ceildiv(magma_int_t a, magma_int_t b);
-
-
-/*
-extern volatile magma_int_t barrier_in[MAX_THREADS_BLG];
-extern volatile magma_int_t barrier_out[MAX_THREADS_BLG];
-extern volatile magma_int_t *ss_prog;
-*/
-
- /***************************************************************************//**
- *  Static scheduler
- **/
-/*
-#define ssched_init(nbtiles) \
-{ \
-        volatile int   prog_ol[2*nbtiles+10];\
-                 int   iamdone[MAX_THREADS_BLG]; \
-                 int   thread_num[MAX_THREADS_BLG];\
-        pthread_t      thread_id[MAX_THREADS_BLG];\
-        pthread_attr_t thread_attr;\
-}
-*/
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-
-
- struct gbstrct_blg {
-    cuFloatComplex *dQ1;
-    cuFloatComplex *dT1;
-    cuFloatComplex *dT2;
-    cuFloatComplex *dV2;
-    cuFloatComplex *dE;
-    cuFloatComplex *T;
-    cuFloatComplex *A;
-    cuFloatComplex *V;
-    cuFloatComplex *TAU;
-    cuFloatComplex *E;
-    cuFloatComplex *E_CPU;
-    int cores_num;
-    int locores_num;
-    int overlapQ1;
-    int usemulticpu;
-    int NB;
-    int NBTILES;
-    int N;
-    int NE;
-    int N_CPU;
-    int N_GPU;
-    int LDA;
-    int LDE;
-    int BAND;
-    int grsiz;
-    int Vblksiz;
-    int WANTZ;
-    char SIDE;
-    real_Double_t *timeblg;
-    real_Double_t *timeaplQ;
-    volatile int *ss_prog;
-} ;
-
-// declare globals here; defined in chetrd_bhe2trc.cpp
-extern struct gbstrct_blg core_in_all;
-
-
-
-
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-#define MAX_EVENTSBLG 163840
-//#define MAX_EVENTSBLG 1048576
-
-// If we're not using GNU C, elide __attribute__
-#ifndef __GNUC__
-#define __attribute__(x)  /*NOTHING*/
-#endif
-
-// declare globals here; defined in chetrd_bhe2trc.cpp
-extern int           event_numblg        [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_start_timeblg [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_end_timeblg   [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_logblg        [MAX_THREADS_BLG][MAX_EVENTSBLG]  __attribute__ ((aligned (128)));
-extern int           log_eventsblg;
-
-#ifndef __GNUC__
-#undef  __attribute__
-#endif
-
-#define core_event_startblg(my_core_id)\
-    event_start_timeblg[my_core_id] = magma_wtime();
-
-#define core_event_endblg(my_core_id)\
-    event_end_timeblg[my_core_id] = magma_wtime();
-
-#define core_log_eventblg(event, my_core_id)\
-    event_logblg[my_core_id][event_numblg[my_core_id]+0] = my_core_id;\
-    event_logblg[my_core_id][event_numblg[my_core_id]+1] = event_start_timeblg[my_core_id];\
-    event_logblg[my_core_id][event_numblg[my_core_id]+2] = event_end_timeblg[my_core_id];\
-    event_logblg[my_core_id][event_numblg[my_core_id]+3] = (event);\
-    event_numblg[my_core_id] += (log_eventsblg << 2);\
-    event_numblg[my_core_id] &= (MAX_EVENTSBLG-1);
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_c
-
-#endif /* MAGMA_CBULGEINC_H */
diff --git a/src/include/magma_clapack.h b/src/include/magma_clapack.h
deleted file mode 100644
index 02cd46b..0000000
--- a/src/include/magma_clapack.h
+++ /dev/null
@@ -1,953 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated c Wed Nov 14 22:52:27 2012
- */
-
-#ifndef MAGMA_CLAPACK_H
-#define MAGMA_CLAPACK_H
-
-#define PRECISION_c
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- BLAS and LAPACK functions (alphabetical order)
-*/
-#define blasf77_icamax     FORTRAN_NAME( icamax, ICAMAX )
-#define blasf77_caxpy      FORTRAN_NAME( caxpy,  CAXPY  )
-#define blasf77_ccopy      FORTRAN_NAME( ccopy,  CCOPY  )
-#define blasf77_cgemm      FORTRAN_NAME( cgemm,  CGEMM  )
-#define blasf77_cgemv      FORTRAN_NAME( cgemv,  CGEMV  )
-#define blasf77_cgerc      FORTRAN_NAME( cgerc,  CGERC  )
-#define blasf77_cgeru      FORTRAN_NAME( cgeru,  CGERU  )
-#define blasf77_chemm      FORTRAN_NAME( chemm,  CHEMM  )
-#define blasf77_chemv      FORTRAN_NAME( chemv,  CHEMV  )
-#define blasf77_cher2      FORTRAN_NAME( cher2,  CHER2  )
-#define blasf77_cher2k     FORTRAN_NAME( cher2k, CHER2K )
-#define blasf77_cherk      FORTRAN_NAME( cherk,  CHERK  )
-#define blasf77_cscal      FORTRAN_NAME( cscal,  CSCAL  )
-#define blasf77_csscal     FORTRAN_NAME( csscal, CSSCAL )
-#define blasf77_cswap      FORTRAN_NAME( cswap,  CSWAP  )
-#define blasf77_csymm      FORTRAN_NAME( csymm,  CSYMM  )
-#define blasf77_csyr2k     FORTRAN_NAME( csyr2k, CSYR2K )
-#define blasf77_csyrk      FORTRAN_NAME( csyrk,  CSYRK  )
-#define blasf77_ctrmm      FORTRAN_NAME( ctrmm,  CTRMM  )
-#define blasf77_ctrmv      FORTRAN_NAME( ctrmv,  CTRMV  )
-#define blasf77_ctrsm      FORTRAN_NAME( ctrsm,  CTRSM  )
-#define blasf77_ctrsv      FORTRAN_NAME( ctrsv,  CTRSV  )
-
-#define lapackf77_slaed4   FORTRAN_NAME( slaed4, SLAED4 )
-#define lapackf77_slamc3   FORTRAN_NAME( slamc3, SLAMC3 )
-#define lapackf77_slamrg   FORTRAN_NAME( slamrg, SLAMRG )
-#define lapackf77_sstebz   FORTRAN_NAME( sstebz, SSTEBZ )
-
-#define lapackf77_cbdsqr   FORTRAN_NAME( cbdsqr, CBDSQR )
-#define lapackf77_cgebak   FORTRAN_NAME( cgebak, CGEBAK )
-#define lapackf77_cgebal   FORTRAN_NAME( cgebal, CGEBAL )
-#define lapackf77_cgebd2   FORTRAN_NAME( cgebd2, CGEBD2 )
-#define lapackf77_cgebrd   FORTRAN_NAME( cgebrd, CGEBRD )
-#define lapackf77_cgeev    FORTRAN_NAME( cgeev,  CGEEV  )
-#define lapackf77_cgehd2   FORTRAN_NAME( cgehd2, CGEHD2 )
-#define lapackf77_cgehrd   FORTRAN_NAME( cgehrd, CGEHRD )
-#define lapackf77_cgelqf   FORTRAN_NAME( cgelqf, CGELQF )
-#define lapackf77_cgels    FORTRAN_NAME( cgels,  CGELS  )
-#define lapackf77_cgeqlf   FORTRAN_NAME( cgeqlf, CGEQLF )
-#define lapackf77_cgeqp3   FORTRAN_NAME( cgeqp3, CGEQP3 )
-#define lapackf77_cgeqrf   FORTRAN_NAME( cgeqrf, CGEQRF )
-#define lapackf77_cgesvd   FORTRAN_NAME( cgesvd, CGESVD )
-#define lapackf77_cgetrf   FORTRAN_NAME( cgetrf, CGETRF )
-#define lapackf77_cgetri   FORTRAN_NAME( cgetri, CGETRI )
-#define lapackf77_cgetrs   FORTRAN_NAME( cgetrs, CGETRS )
-#define lapackf77_chbtrd   FORTRAN_NAME( chbtrd, CHBTRD )
-#define lapackf77_cheev    FORTRAN_NAME( cheev,  CHEEV  )
-#define lapackf77_cheevd   FORTRAN_NAME( cheevd, CHEEVD )
-#define lapackf77_chegs2   FORTRAN_NAME( chegs2, CHEGS2 )
-#define lapackf77_chegvd   FORTRAN_NAME( chegvd, CHEGVD )
-#define lapackf77_chetd2   FORTRAN_NAME( chetd2, CHETD2 )
-#define lapackf77_chetrd   FORTRAN_NAME( chetrd, CHETRD )
-#define lapackf77_chetrf   FORTRAN_NAME( chetrf, CHETRF )
-#define lapackf77_chseqr   FORTRAN_NAME( chseqr, CHSEQR )
-#define lapackf77_clabrd   FORTRAN_NAME( clabrd, CLABRD )
-#define lapackf77_cladiv   FORTRAN_NAME( cladiv, ZLADIV )
-#define lapackf77_clacgv   FORTRAN_NAME( clacgv, CLACGV )
-#define lapackf77_clacpy   FORTRAN_NAME( clacpy, CLACPY )
-#define lapackf77_clahef   FORTRAN_NAME( clahef, CLAHEF )
-#define lapackf77_clange   FORTRAN_NAME( clange, CLANGE )
-#define lapackf77_clanhe   FORTRAN_NAME( clanhe, CLANHE )
-#define lapackf77_clanht   FORTRAN_NAME( clanht, CLANHT )
-#define lapackf77_clansy   FORTRAN_NAME( clansy, CLANSY )
-#define lapackf77_slapy3   FORTRAN_NAME( slapy3, DLAPY3 )
-#define lapackf77_claqp2   FORTRAN_NAME( claqp2, CLAQP2 )
-#define lapackf77_clarf    FORTRAN_NAME( clarf,  ZLARF  )
-#define lapackf77_clarfb   FORTRAN_NAME( clarfb, CLARFB )
-#define lapackf77_clarfg   FORTRAN_NAME( clarfg, CLARFG )
-#define lapackf77_clarft   FORTRAN_NAME( clarft, CLARFT )
-#define lapackf77_clarnv   FORTRAN_NAME( clarnv, CLARNV )
-#define lapackf77_clartg   FORTRAN_NAME( clartg, CLARTG )
-#define lapackf77_clascl   FORTRAN_NAME( clascl, CLASCL )
-#define lapackf77_claset   FORTRAN_NAME( claset, CLASET )
-#define lapackf77_claswp   FORTRAN_NAME( claswp, CLASWP )
-#define lapackf77_clatrd   FORTRAN_NAME( clatrd, CLATRD )
-#define lapackf77_clauum   FORTRAN_NAME( clauum, CLAUUM )
-#define lapackf77_clavhe   FORTRAN_NAME( clavhe, CLAVHE )
-#define lapackf77_cpotrf   FORTRAN_NAME( cpotrf, CPOTRF )
-#define lapackf77_cpotri   FORTRAN_NAME( cpotri, CPOTRI )
-#define lapackf77_cpotrs   FORTRAN_NAME( cpotrs, CPOTRS )
-#define lapackf77_cstedc   FORTRAN_NAME( cstedc, CSTEDC )
-#define lapackf77_cstein   FORTRAN_NAME( cstein, CSTEIN )
-#define lapackf77_cstemr   FORTRAN_NAME( cstemr, CSTEMR )
-#define lapackf77_csteqr   FORTRAN_NAME( csteqr, CSTEQR )
-#define lapackf77_csymv    FORTRAN_NAME( csymv,  CSYMV  )
-#define lapackf77_ctrevc   FORTRAN_NAME( ctrevc, CTREVC )
-#define lapackf77_ctrtri   FORTRAN_NAME( ctrtri, CTRTRI )
-#define lapackf77_cung2r   FORTRAN_NAME( cung2r, CUNG2R )
-#define lapackf77_cungbr   FORTRAN_NAME( cungbr, CUNGBR )
-#define lapackf77_cunghr   FORTRAN_NAME( cunghr, CUNGHR )
-#define lapackf77_cunglq   FORTRAN_NAME( cunglq, CUNGLQ )
-#define lapackf77_cungql   FORTRAN_NAME( cungql, CUNGQL )
-#define lapackf77_cungqr   FORTRAN_NAME( cungqr, CUNGQR )
-#define lapackf77_cungtr   FORTRAN_NAME( cungtr, CUNGTR )
-#define lapackf77_cunm2r   FORTRAN_NAME( cunm2r, CUNM2R )
-#define lapackf77_cunmbr   FORTRAN_NAME( cunmbr, CUNMBR )
-#define lapackf77_cunmlq   FORTRAN_NAME( cunmlq, CUNMLQ )
-#define lapackf77_cunmql   FORTRAN_NAME( cunmql, CUNMQL )
-#define lapackf77_cunmqr   FORTRAN_NAME( cunmqr, CUNMQR )
-#define lapackf77_cunmtr   FORTRAN_NAME( cunmtr, CUNMTR )
-
-/* testing functions */
-#define lapackf77_cbdt01   FORTRAN_NAME( cbdt01, CBDT01 )
-#define lapackf77_cget22   FORTRAN_NAME( cget22, CGET22 )
-#define lapackf77_cqpt01   FORTRAN_NAME( cqpt01, CQPT01 )
-#define lapackf77_chet21   FORTRAN_NAME( chet21, CHET21 )
-#define lapackf77_chst01   FORTRAN_NAME( chst01, CHST01 )
-#define lapackf77_cqrt02   FORTRAN_NAME( cqrt02, CQRT02 )
-#define lapackf77_cunt01   FORTRAN_NAME( cunt01, CUNT01 )
-#define lapackf77_clarfy   FORTRAN_NAME( clarfy, CLARFY )
-#define lapackf77_clarfx   FORTRAN_NAME( clarfx, CLARFX )
-#define lapackf77_cstt21   FORTRAN_NAME( cstt21, CSTT21 )
-
-// macros to handle differences in arguments between complex and real versions of routines.
-#if defined(PRECISION_z) || defined(PRECISION_c)
-#define DWORKFORZ        float *rwork,
-#define DWORKFORZ_AND_LD float *rwork, const magma_int_t *ldrwork,
-#define WSPLIT           cuFloatComplex *w
-#else
-#define DWORKFORZ
-#define DWORKFORZ_AND_LD
-#define WSPLIT           float *wr, float *wi
-#endif
-
-/*
- * BLAS functions (alphabetical order)
- */
-magma_int_t blasf77_icamax(
-                     const magma_int_t *n,
-                     const cuFloatComplex *x, const magma_int_t *incx);
-
-void blasf77_caxpy(  const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *x, const magma_int_t *incx,
-                           cuFloatComplex *y, const magma_int_t *incy );
-
-void blasf77_ccopy(  const magma_int_t *n,
-                     const cuFloatComplex *x, const magma_int_t *incx,
-                           cuFloatComplex *y, const magma_int_t *incy );
-
-void blasf77_cgemm(  const char *transa, const char *transb,
-                     const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                     const cuFloatComplex *B, const magma_int_t *ldb,
-                     const cuFloatComplex *beta,
-                           cuFloatComplex *C, const magma_int_t *ldc );
-
-void blasf77_cgemv(  const char *transa,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                     const cuFloatComplex *x, const magma_int_t *incx,
-                     const cuFloatComplex *beta,
-                           cuFloatComplex *y, const magma_int_t *incy );
-
-void blasf77_cgerc(  const magma_int_t *m, const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *x, const magma_int_t *incx,
-                     const cuFloatComplex *y, const magma_int_t *incy,
-                           cuFloatComplex *A, const magma_int_t *lda );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void blasf77_cgeru(  const magma_int_t *m, const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *x, const magma_int_t *incx,
-                     const cuFloatComplex *y, const magma_int_t *incy,
-                           cuFloatComplex *A, const magma_int_t *lda );
-#endif
-
-void blasf77_chemm(  const char *side, const char *uplo,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                     const cuFloatComplex *B, const magma_int_t *ldb,
-                     const cuFloatComplex *beta,
-                           cuFloatComplex *C, const magma_int_t *ldc );
-
-void blasf77_chemv(  const char *uplo,
-                     const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                     const cuFloatComplex *x, const magma_int_t *incx,
-                     const cuFloatComplex *beta,
-                           cuFloatComplex *y, const magma_int_t *incy );
-
-void blasf77_cher2(  const char *uplo,
-                     const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *x, const magma_int_t *incx,
-                     const cuFloatComplex *y, const magma_int_t *incy,
-                           cuFloatComplex *A, const magma_int_t *lda );
-
-void blasf77_cher2k(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                     const cuFloatComplex *B, const magma_int_t *ldb,
-                     const float *beta,
-                           cuFloatComplex *C, const magma_int_t *ldc );
-
-void blasf77_cherk(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const float *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                     const float *beta,
-                           cuFloatComplex *C, const magma_int_t *ldc );
-
-void blasf77_cscal(  const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                           cuFloatComplex *x, const magma_int_t *incx );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void blasf77_csscal( const magma_int_t *n,
-                     const float *alpha,
-                           cuFloatComplex *x, const magma_int_t *incx );
-#endif
-
-void blasf77_cswap(  const magma_int_t *n,
-                     cuFloatComplex *x, const magma_int_t *incx,
-                     cuFloatComplex *y, const magma_int_t *incy );
-
-void blasf77_csymm(  const char *side, const char *uplo,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                     const cuFloatComplex *B, const magma_int_t *ldb,
-                     const cuFloatComplex *beta,
-                           cuFloatComplex *C, const magma_int_t *ldc );
-
-void blasf77_csyr2k( const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                     const cuFloatComplex *B, const magma_int_t *ldb,
-                     const cuFloatComplex *beta,
-                           cuFloatComplex *C, const magma_int_t *ldc );
-
-void blasf77_csyrk(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                     const cuFloatComplex *beta,
-                           cuFloatComplex *C, const magma_int_t *ldc );
-
-void blasf77_ctrmm(  const char *side, const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                           cuFloatComplex *B, const magma_int_t *ldb );
-
-void blasf77_ctrmv(  const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *n,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                           cuFloatComplex *x, const magma_int_t *incx );
-
-void blasf77_ctrsm(  const char *side, const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const cuFloatComplex *alpha,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                           cuFloatComplex *B, const magma_int_t *ldb );
-
-void blasf77_ctrsv(  const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *n,
-                     const cuFloatComplex *A, const magma_int_t *lda,
-                           cuFloatComplex *x, const magma_int_t *incx );
-
-/*
- * LAPACK functions (alphabetical order)
- */
-void   lapackf77_cbdsqr( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *ncvt, const magma_int_t *nru,  const magma_int_t *ncc,
-                         float *d, float *e,
-                         cuFloatComplex *Vt, const magma_int_t *ldvt,
-                         cuFloatComplex *U, const magma_int_t *ldu,
-                         cuFloatComplex *C, const magma_int_t *ldc,
-                         float *work,
-                         magma_int_t *info );
-
-void   lapackf77_cgebak( const char *job, const char *side,
-                         const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         const float *scale, const magma_int_t *m,
-                         cuFloatComplex *V, const magma_int_t *ldv,
-                         magma_int_t *info );
-
-void   lapackf77_cgebal( const char *job,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         magma_int_t *ilo, magma_int_t *ihi,
-                         float *scale,
-                         magma_int_t *info );
-
-void   lapackf77_cgebd2( const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         cuFloatComplex *tauq,
-                         cuFloatComplex *taup,
-                         cuFloatComplex *work,
-                         magma_int_t *info );
-
-void   lapackf77_cgebrd( const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         cuFloatComplex *tauq,
-                         cuFloatComplex *taup,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cgeev(  const char *jobvl, const char *jobvr,
-                         const magma_int_t *n,
-                         cuFloatComplex *A,    const magma_int_t *lda,
-                         WSPLIT,
-                         cuFloatComplex *Vl,   const magma_int_t *ldvl,
-                         cuFloatComplex *Vr,   const magma_int_t *ldvr,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_cgehd2( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work,
-                         magma_int_t *info );
-
-void   lapackf77_cgehrd( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cgelqf( const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cgels(  const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *nrhs,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *B, const magma_int_t *ldb,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cgeqlf( const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cgeqp3( const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         magma_int_t *jpvt,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_cgeqrf( const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cgesvd( const char *jobu, const char *jobvt,
-                         const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *s,
-                         cuFloatComplex *U,  const magma_int_t *ldu,
-                         cuFloatComplex *Vt, const magma_int_t *ldvt,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_cgetrf( const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         magma_int_t *info );
-
-void   lapackf77_cgetri( const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         const magma_int_t *ipiv,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cgetrs( const char* trans,
-                         const magma_int_t *n, const magma_int_t *nrhs,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         const magma_int_t *ipiv,
-                         cuFloatComplex *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_chbtrd( const char *vect, const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kd,
-                         cuFloatComplex *Ab, const magma_int_t *ldab,
-                         float *d, float *e,
-                         cuFloatComplex *Q, const magma_int_t *ldq,
-                         cuFloatComplex *work,
-                         magma_int_t *info );
-
-void   lapackf77_cheev(  const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *w,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_cheevd( const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *w,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_chegs2( const magma_int_t *itype, const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_chegvd( const magma_int_t *itype, const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *B, const magma_int_t *ldb,
-                         float *w,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_chetd2( const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         cuFloatComplex *tau,
-                         magma_int_t *info );
-
-void   lapackf77_chetrd( const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_chetrf( const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_chseqr( const char *job, const char *compz,
-                         const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuFloatComplex *H, const magma_int_t *ldh,
-                         WSPLIT,
-                         cuFloatComplex *Z, const magma_int_t *ldz,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_clabrd( const magma_int_t *m, const magma_int_t *n, const magma_int_t *nb,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         cuFloatComplex *tauq,
-                         cuFloatComplex *taup,
-                         cuFloatComplex *X, const magma_int_t *ldx,
-                         cuFloatComplex *Y, const magma_int_t *ldy );
-
-void   lapackf77_cladiv( cuFloatComplex *ret_val, cuFloatComplex *x, 
-                         cuFloatComplex *y );
-
-void   lapackf77_clacgv( const magma_int_t *n,
-                         cuFloatComplex *x, const magma_int_t *incx );
-
-void   lapackf77_clacpy( const char *uplo,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *B, const magma_int_t *ldb );
-
-void   lapackf77_clahef( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kn,
-                         magma_int_t *kb,
-                         cuFloatComplex *A, const magma_int_t lda,
-                         magma_int_t *ipiv,
-                         cuFloatComplex *work, const magma_int_t *ldwork,
-                         magma_int_t *info );
-
-float lapackf77_clange( const char *norm,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         float *work );
-
-float lapackf77_clanhe( const char *norm, const char *uplo,
-                         const magma_int_t *n,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         float * work );
-
-float lapackf77_clanht( const char* norm, const magma_int_t* n,
-                         const float* d, const cuFloatComplex* e );
-
-float lapackf77_clansy( const char *norm, const char *uplo,
-                         const magma_int_t *n,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         float * work );
-
-void lapackf77_claqp2 (  magma_int_t *m, magma_int_t *n, magma_int_t *offset,
-                         cuFloatComplex *a, magma_int_t *lda, magma_int_t *jpvt, 
-                         cuFloatComplex *tau,
-                         float *vn1, float *vn2, cuFloatComplex *work);
-
-void lapackf77_clarf  (  char *, magma_int_t *, magma_int_t *,
-                         cuFloatComplex *, magma_int_t *, cuFloatComplex *, cuFloatComplex *,
-                         magma_int_t *, cuFloatComplex *);
-
-void   lapackf77_clarfb( const char *side, const char *trans, const char *direct, const char *storev,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuFloatComplex *V, const magma_int_t *ldv,
-                         const cuFloatComplex *T, const magma_int_t *ldt,
-                         cuFloatComplex *C, const magma_int_t *ldc,
-                         cuFloatComplex *work, const magma_int_t *ldwork );
-
-void   lapackf77_clarfg( const magma_int_t *n,
-                         cuFloatComplex *alpha,
-                         cuFloatComplex *x, const magma_int_t *incx,
-                         cuFloatComplex *tau );
-
-void   lapackf77_clarft( const char *direct, const char *storev,
-                         const magma_int_t *n, const magma_int_t *k,
-                         cuFloatComplex *V, const magma_int_t *ldv,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *T, const magma_int_t *ldt );
-
-void   lapackf77_clarnv( const magma_int_t *idist, magma_int_t *iseed, const magma_int_t *n,
-                         cuFloatComplex *x );
-
-void   lapackf77_clartg( cuFloatComplex *F,
-                         cuFloatComplex *G,
-                         float *cs,
-                         cuFloatComplex *SN,
-                         cuFloatComplex *R );
-
-void   lapackf77_clascl( const char *type,
-                         const magma_int_t *kl, const magma_int_t *ku,
-                         float *cfrom,
-                         float *cto,
-                         const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_claset( const char *uplo,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const cuFloatComplex *alpha,
-                         const cuFloatComplex *beta,
-                         cuFloatComplex *A, const magma_int_t *lda );
-
-void   lapackf77_claswp( const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         const magma_int_t *k1, const magma_int_t *k2,
-                         magma_int_t *ipiv,
-                         const magma_int_t *incx );
-
-void   lapackf77_clatrd( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *nb,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *e,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *ldwork );
-
-void   lapackf77_clauum( const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_clavhe( const char *uplo, const char *trans, const char *diag,
-                         magma_int_t *n, magma_int_t *nrhs,
-                         cuFloatComplex *A, magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         cuFloatComplex *B, magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_cpotrf( const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_cpotri( const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_cpotrs( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *nrhs,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_cstedc( const char *compz,
-                         const magma_int_t *n,
-                         float *d, float *e,
-                         cuFloatComplex *Z, const magma_int_t *ldz,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_cstein( const magma_int_t *n,
-                         const float *d, const float *e,
-                         const magma_int_t *m,
-                         const float *w,
-                         const magma_int_t *iblock,
-                         const magma_int_t *isplit,
-                         cuFloatComplex *Z, const magma_int_t *ldz,
-                         float *work, magma_int_t *iwork, magma_int_t *ifailv,
-                         magma_int_t *info );
-
-void   lapackf77_cstemr( const char *jobz, const char *range,
-                         const magma_int_t *n,
-                         float *d, float *e,
-                         const float *vl, const float *vu,
-                         const magma_int_t *il, const magma_int_t *iu,
-                         magma_int_t *m,
-                         float *w,
-                         cuFloatComplex *Z, const magma_int_t *ldz,
-                         const magma_int_t *nzc, magma_int_t *isuppz, magma_int_t *tryrac,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_csteqr( const char *compz,
-                         const magma_int_t *n,
-                         float *d, float *e,
-                         cuFloatComplex *Z, const magma_int_t *ldz,
-                         float *work,
-                         magma_int_t *info );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void   lapackf77_csymv(  const char *uplo,
-                         const magma_int_t *n,
-                         const cuFloatComplex *alpha,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *x, const magma_int_t *incx,
-                         const cuFloatComplex *beta,
-                               cuFloatComplex *y, const magma_int_t *incy );
-#endif
-
-void   lapackf77_ctrevc( const char *side, const char *howmny,
-                         magma_int_t *select, const magma_int_t *n,
-                         cuFloatComplex *T,  const magma_int_t *ldt,
-                         cuFloatComplex *Vl, const magma_int_t *ldvl,
-                         cuFloatComplex *Vr, const magma_int_t *ldvr,
-                         const magma_int_t *mm, magma_int_t *m,
-                         cuFloatComplex *work,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_ctrtri( const char *uplo, const char *diag,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_cung2r( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *work,
-                         magma_int_t *info );
-
-void   lapackf77_cungbr( const char *vect,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cunghr( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cunglq( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cungql( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cungqr( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cungtr( const char *uplo,
-                         const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cunm2r( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *C, const magma_int_t *ldc,
-                         cuFloatComplex *work,
-                         magma_int_t *info );
-
-void   lapackf77_cunmbr( const char *vect, const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *C, const magma_int_t *ldc,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cunmlq( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *C, const magma_int_t *ldc,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cunmql( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *C, const magma_int_t *ldc,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cunmqr( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *C, const magma_int_t *ldc,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_cunmtr( const char *side, const char *uplo, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const cuFloatComplex *A, const magma_int_t *lda,
-                         const cuFloatComplex *tau,
-                         cuFloatComplex *C, const magma_int_t *ldc,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-/*
- * Real precision extras
- */
-void   lapackf77_sstebz( const char *range, const char *order,
-                         const magma_int_t *n,
-                         float *vl, float *vu,
-                         magma_int_t *il, magma_int_t *iu,
-                         float *abstol,
-                         float *d, float *e,
-                         const magma_int_t *m, const magma_int_t *nsplit,
-                         float *w,
-                         magma_int_t *iblock, magma_int_t *isplit,
-                         float *work,
-                         magma_int_t *iwork,
-                         magma_int_t *info );
-
-float lapackf77_slamc3( float* a, float* b );
-
-void   lapackf77_slamrg( magma_int_t* n1, magma_int_t* n2,
-                         float* a,
-                         magma_int_t* dtrd1, magma_int_t* dtrd2, magma_int_t* index );
-
-float lapackf77_slapy3(float *, float *, float *);
-
-void   lapackf77_slaed4( magma_int_t* n, magma_int_t* i,
-                         float* d,
-                         float* z,
-                         float* delta,
-                         float* rho,
-                         float* dlam, magma_int_t* info );
-
-/*
- * Testing functions
- */
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void   lapackf77_cbdt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *kd,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *Q, const magma_int_t *ldq,
-                         float *d, float *e,
-                         cuFloatComplex *Pt, const magma_int_t *ldpt,
-                         cuFloatComplex *work,
-                         float *rwork,
-                         float *resid );
-
-void   lapackf77_cget22( const char *transa, const char *transe, const char *transw, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *E, const magma_int_t *lde,
-                         cuFloatComplex *w,
-                         cuFloatComplex *work,
-                         float *rwork,
-                         float *result );
-
-void   lapackf77_chet21( const magma_int_t *itype, const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kband,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         cuFloatComplex *U, const magma_int_t *ldu,
-                         cuFloatComplex *V, const magma_int_t *ldv,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work,
-                         float *rwork,
-                         float *result );
-
-void   lapackf77_chst01( const magma_int_t *n, const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *H, const magma_int_t *ldh,
-                         cuFloatComplex *Q, const magma_int_t *ldq,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         float *rwork,
-                         float *result );
-
-void   lapackf77_cstt21( const magma_int_t *n, const magma_int_t *kband,
-                         float *AD,
-                         float *AE,
-                         float *SD,
-                         float *SE,
-                         cuFloatComplex *U, const magma_int_t *ldu,
-                         cuFloatComplex *work,
-                         float *rwork,
-                         float *result );
-
-void   lapackf77_cunt01( const char *rowcol, const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *U, const magma_int_t *ldu,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         float *rwork,
-                         float *resid );
-#else
-void   lapackf77_cbdt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *kd,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *Q, const magma_int_t *ldq,
-                         float *d, float *e,
-                         cuFloatComplex *Pt, const magma_int_t *ldpt,
-                         cuFloatComplex *work,
-                         float *resid );
-
-void   lapackf77_cget22( const char *transa, const char *transe, const char *transw, const magma_int_t *n,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *E, const magma_int_t *lde,
-                         cuFloatComplex *wr,
-                         cuFloatComplex *wi,
-                         float *work,
-                         float *result );
-
-void   lapackf77_chet21( magma_int_t *itype, const char *uplo, const magma_int_t *n, const magma_int_t *kband,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         cuFloatComplex *U, const magma_int_t *ldu,
-                         cuFloatComplex *V, const magma_int_t *ldv,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work,
-                         float *result );
-
-void   lapackf77_chst01( const magma_int_t *n, const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuFloatComplex *A, const magma_int_t *lda,
-                         cuFloatComplex *H, const magma_int_t *ldh,
-                         cuFloatComplex *Q, const magma_int_t *ldq,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         float *result );
-
-void   lapackf77_cstt21( const magma_int_t *n, const magma_int_t *kband,
-                         float *AD,
-                         float *AE,
-                         float *SD,
-                         float *SE,
-                         cuFloatComplex *U, const magma_int_t *ldu,
-                         cuFloatComplex *work,
-                         float *result );
-
-void   lapackf77_cunt01( const char *rowcol, const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *U, const magma_int_t *ldu,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         float *resid );
-#endif
-
-void   lapackf77_clarfy( const char *uplo, const magma_int_t *n,
-                         cuFloatComplex *V, const magma_int_t *incv,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *C, const magma_int_t *ldc,
-                         cuFloatComplex *work );
-
-void   lapackf77_clarfx( const char *side, const magma_int_t *m, const magma_int_t *n,
-                         cuFloatComplex *V,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *C, const magma_int_t *ldc,
-                         cuFloatComplex *work );
-
-float lapackf77_cqpt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuFloatComplex *A,
-                         cuFloatComplex *Af, const magma_int_t *lda,
-                         cuFloatComplex *tau, magma_int_t *jpvt,
-                         cuFloatComplex *work, const magma_int_t *lwork );
-
-void   lapackf77_cqrt02( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuFloatComplex *A,
-                         cuFloatComplex *AF,
-                         cuFloatComplex *Q,
-                         cuFloatComplex *R, const magma_int_t *lda,
-                         cuFloatComplex *tau,
-                         cuFloatComplex *work, const magma_int_t *lwork,
-                         float *rwork,
-                         float *result );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef DWORKFORZ
-#undef DWORKFORZ_AND_LD
-#undef WSPLIT
-#undef PRECISION_c
-
-#endif /* MAGMA_CLAPACK_H */
diff --git a/src/include/magma_d.h b/src/include/magma_d.h
deleted file mode 100644
index 83e6675..0000000
--- a/src/include/magma_d.h
+++ /dev/null
@@ -1,514 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated d Wed Nov 14 22:52:26 2012
- */
-
-#ifndef _MAGMA_D_H_
-#define _MAGMA_D_H_
-#define PRECISION_d
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- MAGMA Auxiliary functions to get the NB used
-*/
-int magma_get_dpotrf_nb(int m);
-int magma_get_dgetrf_nb(int m);
-int magma_get_dgetri_nb(int m);
-int magma_get_dgeqp3_nb(int m);
-int magma_get_dgeqrf_nb(int m);
-int magma_get_dgeqlf_nb(int m);
-int magma_get_dgehrd_nb(int m);
-int magma_get_dsytrd_nb(int m);
-int magma_get_dgelqf_nb(int m);
-int magma_get_dgebrd_nb(int m);
-int magma_get_dsygst_nb(int m);
-int magma_get_dgesvd_nb(int m);
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- MAGMA function definitions / Data on CPU
-*/
-magma_int_t magma_dgebrd( magma_int_t m, magma_int_t n, double *A, 
-                          magma_int_t lda, double *d, double *e,
-                          double *tauq,  double *taup, 
-                          double *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_dgehrd2(magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          double *A, magma_int_t lda, double *tau, 
-                          double *work, magma_int_t *lwork, magma_int_t *info);
-magma_int_t magma_dgehrd( magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          double *A, magma_int_t lda, double *tau,
-                          double *work, magma_int_t lwork,
-                          double *d_T, magma_int_t *info);
-magma_int_t magma_dgelqf( magma_int_t m, magma_int_t n, 
-                          double *A,    magma_int_t lda,   double *tau, 
-                          double *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_dgeqlf( magma_int_t m, magma_int_t n, 
-                          double *A,    magma_int_t lda,   double *tau, 
-                          double *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_dgeqrf( magma_int_t m, magma_int_t n, double *A, 
-                          magma_int_t lda, double *tau, double *work, 
-                          magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_dgeqr2( magma_int_t *m, magma_int_t *n, double *a,
-                          magma_int_t *lda, double *tau, double *work, 
-                          magma_int_t *info);
-magma_int_t magma_dgeqrf4(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
-                          double *a,    magma_int_t lda, double *tau,
-                          double *work, magma_int_t lwork, magma_int_t *info );
-magma_int_t magma_dgeqrf_ooc( magma_int_t m, magma_int_t n, double *A,
-                          magma_int_t lda, double *tau, double *work,
-                          magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_dgesv ( magma_int_t n, magma_int_t nrhs, 
-                          double *A, magma_int_t lda, magma_int_t *ipiv, 
-                          double *B, magma_int_t ldb, magma_int_t *info);
-magma_int_t magma_dgetrf( magma_int_t m, magma_int_t n, double *A, 
-                          magma_int_t lda, magma_int_t *ipiv, 
-                          magma_int_t *info);
-magma_int_t magma_dgetrf2(magma_int_t m, magma_int_t n, double *a, 
-                          magma_int_t lda, magma_int_t *ipiv, magma_int_t *info);
-
-magma_int_t magma_dlaqps( magma_int_t m, magma_int_t n, magma_int_t offset, 
-                          magma_int_t nb, magma_int_t *kb, 
-                          double *A,  magma_int_t lda,
-                          double *dA, magma_int_t ldda,
-                          magma_int_t *jpvt, double *tau, double *vn1, double *vn2, 
-                          double *auxv, 
-                          double *F,  magma_int_t ldf,
-                          double *dF, magma_int_t lddf );
-magma_int_t magma_dlarf(  char *side, magma_int_t *m, magma_int_t *n,
-                          double *v, magma_int_t *incv, double *tau,
-                          double *c__, magma_int_t *ldc, double *work);
-magma_int_t magma_dlarfg( magma_int_t *n, double *alpha, double *x,
-                          magma_int_t *incx, double *tau);
-magma_int_t magma_dlatrd( char uplo, magma_int_t n, magma_int_t nb, double *a, 
-                          magma_int_t lda, double *e, double *tau, 
-                          double *w, magma_int_t ldw,
-                          double *da, magma_int_t ldda, 
-                          double *dw, magma_int_t lddw);
-magma_int_t magma_dlatrd2(char uplo, magma_int_t n, magma_int_t nb,
-                          double *a,  magma_int_t lda,
-                          double *e, double *tau,
-                          double *w,  magma_int_t ldw,
-                          double *da, magma_int_t ldda,
-                          double *dw, magma_int_t lddw,
-                          double *dwork, magma_int_t ldwork);
-magma_int_t magma_dlahr2( magma_int_t m, magma_int_t n, magma_int_t nb, 
-                          double *da, double *dv, double *a, 
-                          magma_int_t lda, double *tau, double *t, 
-                          magma_int_t ldt, double *y, magma_int_t ldy);
-magma_int_t magma_dlahru( magma_int_t n, magma_int_t ihi, magma_int_t k, magma_int_t nb, 
-                          double *a, magma_int_t lda, 
-                          double *da, double *y, 
-                          double *v, double *t, 
-                          double *dwork);
-magma_int_t magma_dposv ( char uplo, magma_int_t n, magma_int_t nrhs, 
-                          double *A, magma_int_t lda, 
-                          double *B, magma_int_t ldb, magma_int_t *info);
-magma_int_t magma_dpotrf( char uplo, magma_int_t n, double *A, 
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_dpotri( char uplo, magma_int_t n, double *A,
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_dlauum( char uplo, magma_int_t n, double *A,
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_dtrtri( char uplo, char diag, magma_int_t n, double *A, 
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_dsytrd( char uplo, magma_int_t n, double *A, 
-                          magma_int_t lda, double *d, double *e, 
-                          double *tau, double *work, magma_int_t lwork, 
-                          magma_int_t *info);
-magma_int_t magma_dorgqr( magma_int_t m, magma_int_t n, magma_int_t k,
-                          double *a, magma_int_t lda,
-                          double *tau, double *dwork,
-                          magma_int_t nb, magma_int_t *info );
-magma_int_t magma_dormql( const char side, const char trans,
-                          magma_int_t m, magma_int_t n, magma_int_t k,
-                          double *a, magma_int_t lda,
-                          double *tau,
-                          double *c, magma_int_t ldc,
-                          double *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_dormqr( char side, char trans, 
-                          magma_int_t m, magma_int_t n, magma_int_t k, 
-                          double *a, magma_int_t lda, double *tau, 
-                          double *c, magma_int_t ldc, 
-                          double *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_dormtr( char side, char uplo, char trans,
-                          magma_int_t m, magma_int_t n,
-                          double *a,    magma_int_t lda,
-                          double *tau,
-                          double *c,    magma_int_t ldc,
-                          double *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_dorghr( magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          double *a, magma_int_t lda,
-                          double *tau,
-                          double *dT, magma_int_t nb,
-                          magma_int_t *info);
-magma_int_t magma_dsyev( char jobz, char uplo, magma_int_t n,
-                         double *a, magma_int_t lda, double *w,
-                         double *work, magma_int_t lwork,
-                         double *rwork, magma_int_t *info);
-magma_int_t magma_dsyevx(char jobz, char range, char uplo, magma_int_t n,
-                         double *a, magma_int_t lda, double vl, double vu,
-                         magma_int_t il, magma_int_t iu, double abstol, magma_int_t *m,
-                         double *w, double *z, magma_int_t ldz, 
-                         double *work, magma_int_t lwork,
-                         double *rwork, magma_int_t *iwork, 
-                         magma_int_t *ifail, magma_int_t *info);
-#if defined(PRECISION_z) || defined(PRECISION_c)
-magma_int_t  magma_dgeev( char jobvl, char jobvr, magma_int_t n,
-                          double *a, magma_int_t lda,
-                          double *w,
-                          double *vl, magma_int_t ldvl,
-                          double *vr, magma_int_t ldvr,
-                          double *work, magma_int_t lwork,
-                          double *rwork, magma_int_t *info);
-magma_int_t magma_dgeqp3( magma_int_t m, magma_int_t n,
-                          double *a, magma_int_t lda, 
-                          magma_int_t *jpvt, double *tau,
-                          double *work, magma_int_t lwork, 
-                          double *rwork, magma_int_t *info);
-magma_int_t magma_dgesvd( char jobu, char jobvt, magma_int_t m, magma_int_t n,
-                          double *a,    magma_int_t lda, double *s, 
-                          double *u,    magma_int_t ldu, 
-                          double *vt,   magma_int_t ldvt,
-                          double *work, magma_int_t lwork,
-                          double *rwork, magma_int_t *info );
-magma_int_t magma_dsyevd( char jobz, char uplo, magma_int_t n,
-                          double *a, magma_int_t lda, double *w,
-                          double *work, magma_int_t lwork,
-                          double *rwork, magma_int_t lrwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_dsyevr( char jobz, char range, char uplo, magma_int_t n,
-                          double *a, magma_int_t lda, double vl, double vu,
-                          magma_int_t il, magma_int_t iu, double abstol, magma_int_t *m,
-                          double *w, double *z, magma_int_t ldz, 
-                          magma_int_t *isuppz,
-                          double *work, magma_int_t lwork,
-                          double *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_dsygvd( magma_int_t itype, char jobz, char uplo, magma_int_t n,
-                          double *a, magma_int_t lda,
-                          double *b, magma_int_t ldb,
-                          double *w, double *work, magma_int_t lwork,
-                          double *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_dsygvdx(magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, double *a, magma_int_t lda,
-                          double *b, magma_int_t ldb,
-                          double vl, double vu, magma_int_t il, magma_int_t iu,
-                          magma_int_t *m, double *w, double *work, 
-                          magma_int_t lwork, double *rwork,
-                          magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_dsygvx( magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, double *a, magma_int_t lda, 
-                          double *b, magma_int_t ldb,
-                          double vl, double vu, magma_int_t il, magma_int_t iu,
-                          double abstol, magma_int_t *m, double *w, 
-                          double *z, magma_int_t ldz,
-                          double *work, magma_int_t lwork, double *rwork,
-                          magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info);
-magma_int_t magma_dsygvr( magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, double *a, magma_int_t lda,
-                          double *b, magma_int_t ldb,
-                          double vl, double vu, magma_int_t il, magma_int_t iu,
-                          double abstol, magma_int_t *m, double *w, 
-                          double *z, magma_int_t ldz,
-                          magma_int_t *isuppz, double *work, magma_int_t lwork,
-                          double *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_dstedx( char range, magma_int_t n, double vl, double vu,
-                          magma_int_t il, magma_int_t iu, double *D, double *E,
-                          double *Z, magma_int_t ldz,
-                          double *rwork, magma_int_t ldrwork, magma_int_t *iwork,
-                          magma_int_t liwork, double* dwork, magma_int_t *info);
-#else
-magma_int_t  magma_dgeev( char jobvl, char jobvr, magma_int_t n,
-                          double *a,    magma_int_t lda,
-                          double *wr, double *wi,
-                          double *vl,   magma_int_t ldvl,
-                          double *vr,   magma_int_t ldvr,
-                          double *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_dgeqp3( magma_int_t m, magma_int_t n,
-                          double *a, magma_int_t lda, 
-                          magma_int_t *jpvt, double *tau,
-                          double *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_dgesvd( char jobu, char jobvt, magma_int_t m, magma_int_t n,
-                          double *a,    magma_int_t lda, double *s, 
-                          double *u,    magma_int_t ldu, 
-                          double *vt,   magma_int_t ldvt,
-                          double *work, magma_int_t lwork,
-                          magma_int_t *info );
-magma_int_t magma_dsyevd( char jobz, char uplo, magma_int_t n,
-                          double *a, magma_int_t lda, double *w,
-                          double *work, magma_int_t lwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_dsygvd( magma_int_t itype, char jobz, char uplo, magma_int_t n,
-                          double *a, magma_int_t lda,
-                          double *b, magma_int_t ldb,
-                          double *w, double *work, magma_int_t lwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_dstedx( char range, magma_int_t n, double vl, double vu,
-                          magma_int_t il, magma_int_t iu, double* d, double* e,
-                          double* z, magma_int_t ldz,
-                          double* work, magma_int_t lwork,
-                          magma_int_t* iwork, magma_int_t liwork,
-                          double* dwork, magma_int_t* info);
-magma_int_t magma_dlaex0( magma_int_t n, double* d, double* e, double* q, magma_int_t ldq,
-                          double* work, magma_int_t* iwork, double* dwork,
-                          char range, double vl, double vu,
-                          magma_int_t il, magma_int_t iu, magma_int_t* info);
-magma_int_t magma_dlaex1( magma_int_t n, double* d, double* q, magma_int_t ldq,
-                          magma_int_t* indxq, double rho, magma_int_t cutpnt,
-                          double* work, magma_int_t* iwork, double* dwork,
-                          char range, double vl, double vu,
-                          magma_int_t il, magma_int_t iu, magma_int_t* info);
-magma_int_t magma_dlaex3( magma_int_t k, magma_int_t n, magma_int_t n1, double* d,
-                          double* q, magma_int_t ldq, double rho,
-                          double* dlamda, double* q2, magma_int_t* indx,
-                          magma_int_t* ctot, double* w, double* s, magma_int_t* indxq,
-                          double* dwork,
-                          char range, double vl, double vu, magma_int_t il, magma_int_t iu,
-                          magma_int_t* info );
-#endif
-
-magma_int_t magma_dsygst( magma_int_t itype, char uplo, magma_int_t n,
-                          double *a, magma_int_t lda,
-                          double *b, magma_int_t ldb, magma_int_t *info);
-
-/* //////////////////////////////////////////////////////////////////////////// 
- -- MAGMA function definitions / Data on GPU
-*/
-magma_int_t magma_dgels_gpu(  char trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
-                              double *dA,    magma_int_t ldda, 
-                              double *dB,    magma_int_t lddb, 
-                              double *hwork, magma_int_t lwork, 
-                              magma_int_t *info);
-magma_int_t magma_dgels3_gpu( char trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
-                              double *dA,    magma_int_t ldda,
-                              double *dB,    magma_int_t lddb,
-                              double *hwork, magma_int_t lwork,
-                              magma_int_t *info);
-magma_int_t magma_dgelqf_gpu( magma_int_t m, magma_int_t n,
-                              double *dA,    magma_int_t ldda,   double *tau,
-                              double *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_dgeqrf_gpu( magma_int_t m, magma_int_t n, 
-                              double *dA,  magma_int_t ldda, 
-                              double *tau, double *dT, 
-                              magma_int_t *info);
-magma_int_t magma_dgeqrf2_gpu(magma_int_t m, magma_int_t n, 
-                              double *dA,  magma_int_t ldda, 
-                              double *tau, magma_int_t *info);
-magma_int_t magma_dgeqrf2_mgpu(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
-                               double **dlA, magma_int_t ldda,
-                               double *tau, magma_int_t *info );
-magma_int_t magma_dgeqrf3_gpu(magma_int_t m, magma_int_t n, 
-                              double *dA,  magma_int_t ldda, 
-                              double *tau, double *dT, 
-                              magma_int_t *info);
-magma_int_t magma_dgeqrs_gpu( magma_int_t m, magma_int_t n, magma_int_t nrhs, 
-                              double *dA,     magma_int_t ldda, 
-                              double *tau,   double *dT,
-                              double *dB,    magma_int_t lddb,
-                              double *hwork, magma_int_t lhwork, 
-                              magma_int_t *info);
-magma_int_t magma_dgeqrs3_gpu( magma_int_t m, magma_int_t n, magma_int_t nrhs, 
-                              double *dA,     magma_int_t ldda, 
-                              double *tau,   double *dT,
-                              double *dB,    magma_int_t lddb,
-                              double *hwork, magma_int_t lhwork, 
-                              magma_int_t *info);
-magma_int_t magma_dgessm_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t k, magma_int_t ib, 
-                              magma_int_t *ipiv, 
-                              double *dL1, magma_int_t lddl1, 
-                              double *dL,  magma_int_t lddl, 
-                              double *dA,  magma_int_t ldda, 
-                              magma_int_t *info);
-magma_int_t magma_dgesv_gpu(  magma_int_t n, magma_int_t nrhs, 
-                              double *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              double *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_dgetrf_incpiv_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t ib,
-                              double *hA, magma_int_t ldha, double *dA, magma_int_t ldda,
-                              double *hL, magma_int_t ldhl, double *dL, magma_int_t lddl,
-                              magma_int_t *ipiv, 
-                              double *dwork, magma_int_t lddwork,
-                              magma_int_t *info);
-magma_int_t magma_dgetrf_gpu( magma_int_t m, magma_int_t n, 
-                              double *dA, magma_int_t ldda, 
-                              magma_int_t *ipiv, magma_int_t *info);
-magma_int_t 
-magma_dgetrf_nopiv_gpu      ( magma_int_t m, magma_int_t n,
-                              double *dA, magma_int_t ldda,
-                              magma_int_t *info);
-magma_int_t magma_dgetri_gpu( magma_int_t n, 
-                              double *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              double *dwork, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_dgetrs_gpu( char trans, magma_int_t n, magma_int_t nrhs, 
-                              double *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              double *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_dlabrd_gpu( magma_int_t m, magma_int_t n, magma_int_t nb, 
-                              double *a, magma_int_t lda, double *da, magma_int_t ldda,
-                              double *d, double *e, double *tauq, double *taup,  
-                              double *x, magma_int_t ldx, double *dx, magma_int_t lddx, 
-                              double *y, magma_int_t ldy, double *dy, magma_int_t lddy);
-magma_int_t magma_dlarfb_gpu( char side, char trans, char direct, char storev, 
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              const double *dv, magma_int_t ldv,
-                              const double *dt, magma_int_t ldt, 
-                              double *dc,       magma_int_t ldc,
-                              double *dwork,    magma_int_t ldwork );
-magma_int_t magma_dposv_gpu(  char uplo, magma_int_t n, magma_int_t nrhs, 
-                              double *dA, magma_int_t ldda, 
-                              double *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_dpotrf_gpu( char uplo,  magma_int_t n, 
-                              double *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_dpotri_gpu( char uplo,  magma_int_t n,
-                              double *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_dlauum_gpu( char uplo,  magma_int_t n,
-                              double *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_dtrtri_gpu( char uplo,  char diag, magma_int_t n,
-                              double *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_dsytrd_gpu( char uplo, magma_int_t n,
-                              double *da, magma_int_t ldda,
-                              double *d, double *e, double *tau,
-                              double *wa,  magma_int_t ldwa,
-                              double *work, magma_int_t lwork,
-                              magma_int_t *info);
-magma_int_t magma_dsytrd2_gpu(char uplo, magma_int_t n,
-                              double *da, magma_int_t ldda,
-                              double *d, double *e, double *tau,
-                              double *wa,  magma_int_t ldwa,
-                              double *work, magma_int_t lwork,
-                              double *dwork, magma_int_t ldwork,
-                              magma_int_t *info);
-magma_int_t magma_dpotrs_gpu( char uplo,  magma_int_t n, magma_int_t nrhs, 
-                              double *dA, magma_int_t ldda, 
-                              double *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_dssssm_gpu( char storev, magma_int_t m1, magma_int_t n1, 
-                              magma_int_t m2, magma_int_t n2, magma_int_t k, magma_int_t ib, 
-                              double *dA1, magma_int_t ldda1, 
-                              double *dA2, magma_int_t ldda2, 
-                              double *dL1, magma_int_t lddl1, 
-                              double *dL2, magma_int_t lddl2,
-                              magma_int_t *IPIV, magma_int_t *info);
-magma_int_t magma_dtstrf_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t ib, magma_int_t nb,
-                              double *hU, magma_int_t ldhu, double *dU, magma_int_t lddu, 
-                              double *hA, magma_int_t ldha, double *dA, magma_int_t ldda, 
-                              double *hL, magma_int_t ldhl, double *dL, magma_int_t lddl,
-                              magma_int_t *ipiv, 
-                              double *hwork, magma_int_t ldhwork, 
-                              double *dwork, magma_int_t lddwork,
-                              magma_int_t *info);
-magma_int_t magma_dorgqr_gpu( magma_int_t m, magma_int_t n, magma_int_t k, 
-                              double *da, magma_int_t ldda, 
-                              double *tau, double *dwork, 
-                              magma_int_t nb, magma_int_t *info );
-magma_int_t magma_dormql2_gpu(const char side, const char trans,
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              double *da, magma_int_t ldda,
-                              double *tau,
-                              double *dc, magma_int_t lddc,
-                              double *wa, magma_int_t ldwa,
-                              magma_int_t *info);
-magma_int_t magma_dormqr_gpu( char side, char trans, 
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              double *a,    magma_int_t lda, double *tau, 
-                              double *c,    magma_int_t ldc,
-                              double *work, magma_int_t lwork, 
-                              double *td,   magma_int_t nb, magma_int_t *info);
-magma_int_t magma_dormqr2_gpu(const char side, const char trans,
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              double *da,   magma_int_t ldda,
-                              double *tau,
-                              double *dc,    magma_int_t lddc,
-                              double *wa,    magma_int_t ldwa,
-                              magma_int_t *info);
-magma_int_t magma_dormtr_gpu( char side, char uplo, char trans,
-                              magma_int_t m, magma_int_t n,
-                              double *da,    magma_int_t ldda,
-                              double *tau,
-                              double *dc,    magma_int_t lddc,
-                              double *wa,    magma_int_t ldwa,
-                              magma_int_t *info);
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-magma_int_t magma_dsyevd_gpu( char jobz, char uplo,
-                              magma_int_t n,
-                              double *da, magma_int_t ldda,
-                              double *w,
-                              double *wa,  magma_int_t ldwa,
-                              double *work, magma_int_t lwork,
-                              double *rwork, magma_int_t lrwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-magma_int_t magma_dsyevdx_gpu(char jobz, char range, char uplo,
-                              magma_int_t n, double *da, 
-                              magma_int_t ldda, double vl, double vu, 
-                              magma_int_t il, magma_int_t iu,
-                              magma_int_t *m, double *w,
-                              double *wa,  magma_int_t ldwa,
-                              double *work, magma_int_t lwork,
-                              double *rwork, magma_int_t lrwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-magma_int_t magma_dsyevr_gpu( char jobz, char range, char uplo, magma_int_t n,
-                              double *da, magma_int_t ldda, double vl, double vu,
-                              magma_int_t il, magma_int_t iu, double abstol, magma_int_t *m,
-                              double *w, double *dz, magma_int_t lddz,
-                              magma_int_t *isuppz,
-                              double *wa, magma_int_t ldwa,
-                              double *wz, magma_int_t ldwz,
-                              double *work, magma_int_t lwork,
-                              double *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                              magma_int_t liwork, magma_int_t *info);
-#else
-magma_int_t magma_dsyevd_gpu( char jobz, char uplo,
-                              magma_int_t n,
-                              double *da, magma_int_t ldda,
-                              double *w,
-                              double *wa,  magma_int_t ldwa,
-                              double *work, magma_int_t lwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-#endif
-
-magma_int_t magma_dsyevx_gpu( char jobz, char range, char uplo, magma_int_t n,
-                              double *da, magma_int_t ldda, double vl, 
-                              double vu, magma_int_t il, magma_int_t iu, 
-                              double abstol, magma_int_t *m,
-                              double *w, double *dz, magma_int_t lddz,
-                              double *wa, magma_int_t ldwa,
-                              double *wz, magma_int_t ldwz,
-                              double *work, magma_int_t lwork,
-                              double *rwork, magma_int_t *iwork, 
-                              magma_int_t *ifail, magma_int_t *info);
-magma_int_t magma_dsygst_gpu(magma_int_t itype, char uplo, magma_int_t n,
-                             double *da, magma_int_t ldda,
-                             double *db, magma_int_t lddb, magma_int_t *info);
-
-
-/* //////////////////////////////////////////////////////////////////////////// 
- -- MAGMA utility function definitions
-*/
-
-void magma_dprint    ( magma_int_t m, magma_int_t n, const double  *A, magma_int_t lda  );
-void magma_dprint_gpu( magma_int_t m, magma_int_t n, const double *dA, magma_int_t ldda );
-
-void dpanel_to_q( char uplo, magma_int_t ib, double *A, magma_int_t lda, double *work );
-void dq_to_panel( char uplo, magma_int_t ib, double *A, magma_int_t lda, double *work );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_d
-#endif /* _MAGMA_D_H_ */
diff --git a/src/include/magma_dbulge.h b/src/include/magma_dbulge.h
deleted file mode 100644
index 049079e..0000000
--- a/src/include/magma_dbulge.h
+++ /dev/null
@@ -1,35 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated d Wed Nov 14 22:52:27 2012
- */
-
-#ifndef MAGMA_DBULGEINC_H
-#define MAGMA_DBULGEINC_H
-
-#define PRECISION_d
-
-extern "C" {
-    magma_int_t magma_dsytrd_sy2sb(char uplo, magma_int_t n, magma_int_t NB, double *a, magma_int_t lda,
-                                   double *tau, double *work, magma_int_t lwork, double *dT, magma_int_t threads, magma_int_t *info);
-
-    magma_int_t magma_dsytrd_hb2st(magma_int_t threads, char uplo, magma_int_t n, magma_int_t nb, magma_int_t Vblksiz,
-                                   double *A, magma_int_t lda, double *D, double *E,
-                                   double *V, magma_int_t ldv, double *TAU, magma_int_t compT, double *T, magma_int_t ldt);
-
-    magma_int_t magma_dbulge_back(magma_int_t threads, char uplo, magma_int_t n, magma_int_t nb, magma_int_t ne, magma_int_t Vblksiz,
-                                  double *Z, magma_int_t ldz, double *dZ, magma_int_t lddz,
-                                  double *V, magma_int_t ldv, double *TAU, double *T, magma_int_t ldt, magma_int_t* info);
-
-    magma_int_t magma_dormqr_gpu_2stages(char side, char trans, magma_int_t m, magma_int_t n, magma_int_t k, double *dA, magma_int_t ldda,
-                                         double *dC, magma_int_t lddc, double *dT, magma_int_t nb, magma_int_t *info);
-
-    magma_int_t magma_dbulge_get_lq2(magma_int_t n);
-
-    magma_int_t magma_dbulge_get_Vblksiz(magma_int_t n, magma_int_t nb);
-}
-#endif
diff --git a/src/include/magma_dbulgeinc.h b/src/include/magma_dbulgeinc.h
deleted file mode 100644
index aeb4aef..0000000
--- a/src/include/magma_dbulgeinc.h
+++ /dev/null
@@ -1,141 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated d Wed Nov 14 22:52:27 2012
- */
-
-#ifndef MAGMA_DBULGEINC_H
-#define MAGMA_DBULGEINC_H
-
-#define PRECISION_d
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-
-/***************************************************************************//**
- *  Configuration
- **/
-
- // maximum contexts
-#define MAX_THREADS_BLG         256
-
-void findVTpos(magma_int_t N, magma_int_t NB, magma_int_t Vblksiz, magma_int_t sweep, magma_int_t st, magma_int_t *Vpos, magma_int_t *TAUpos, magma_int_t *Tpos, magma_int_t *myblkid);
-void findVTsiz(magma_int_t N, magma_int_t NB, magma_int_t Vblksiz, magma_int_t *blkcnt, magma_int_t *LDV);
-magma_int_t plasma_ceildiv(magma_int_t a, magma_int_t b);
-
-
-/*
-extern volatile magma_int_t barrier_in[MAX_THREADS_BLG];
-extern volatile magma_int_t barrier_out[MAX_THREADS_BLG];
-extern volatile magma_int_t *ss_prog;
-*/
-
- /***************************************************************************//**
- *  Static scheduler
- **/
-/*
-#define ssched_init(nbtiles) \
-{ \
-        volatile int   prog_ol[2*nbtiles+10];\
-                 int   iamdone[MAX_THREADS_BLG]; \
-                 int   thread_num[MAX_THREADS_BLG];\
-        pthread_t      thread_id[MAX_THREADS_BLG];\
-        pthread_attr_t thread_attr;\
-}
-*/
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-
-
- struct gbstrct_blg {
-    double *dQ1;
-    double *dT1;
-    double *dT2;
-    double *dV2;
-    double *dE;
-    double *T;
-    double *A;
-    double *V;
-    double *TAU;
-    double *E;
-    double *E_CPU;
-    int cores_num;
-    int locores_num;
-    int overlapQ1;
-    int usemulticpu;
-    int NB;
-    int NBTILES;
-    int N;
-    int NE;
-    int N_CPU;
-    int N_GPU;
-    int LDA;
-    int LDE;
-    int BAND;
-    int grsiz;
-    int Vblksiz;
-    int WANTZ;
-    char SIDE;
-    real_Double_t *timeblg;
-    real_Double_t *timeaplQ;
-    volatile int *ss_prog;
-} ;
-
-// declare globals here; defined in dsytrd_bsy2trc.cpp
-extern struct gbstrct_blg core_in_all;
-
-
-
-
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-#define MAX_EVENTSBLG 163840
-//#define MAX_EVENTSBLG 1048576
-
-// If we're not using GNU C, elide __attribute__
-#ifndef __GNUC__
-#define __attribute__(x)  /*NOTHING*/
-#endif
-
-// declare globals here; defined in dsytrd_bsy2trc.cpp
-extern int           event_numblg        [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_start_timeblg [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_end_timeblg   [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_logblg        [MAX_THREADS_BLG][MAX_EVENTSBLG]  __attribute__ ((aligned (128)));
-extern int           log_eventsblg;
-
-#ifndef __GNUC__
-#undef  __attribute__
-#endif
-
-#define core_event_startblg(my_core_id)\
-    event_start_timeblg[my_core_id] = magma_wtime();
-
-#define core_event_endblg(my_core_id)\
-    event_end_timeblg[my_core_id] = magma_wtime();
-
-#define core_log_eventblg(event, my_core_id)\
-    event_logblg[my_core_id][event_numblg[my_core_id]+0] = my_core_id;\
-    event_logblg[my_core_id][event_numblg[my_core_id]+1] = event_start_timeblg[my_core_id];\
-    event_logblg[my_core_id][event_numblg[my_core_id]+2] = event_end_timeblg[my_core_id];\
-    event_logblg[my_core_id][event_numblg[my_core_id]+3] = (event);\
-    event_numblg[my_core_id] += (log_eventsblg << 2);\
-    event_numblg[my_core_id] &= (MAX_EVENTSBLG-1);
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_d
-
-#endif /* MAGMA_DBULGEINC_H */
diff --git a/src/include/magma_dlapack.h b/src/include/magma_dlapack.h
deleted file mode 100644
index 02e2d3f..0000000
--- a/src/include/magma_dlapack.h
+++ /dev/null
@@ -1,953 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated d Wed Nov 14 22:52:27 2012
- */
-
-#ifndef MAGMA_DLAPACK_H
-#define MAGMA_DLAPACK_H
-
-#define PRECISION_d
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- BLAS and LAPACK functions (alphabetical order)
-*/
-#define blasf77_idamax     FORTRAN_NAME( idamax, IDAMAX )
-#define blasf77_daxpy      FORTRAN_NAME( daxpy,  DAXPY  )
-#define blasf77_dcopy      FORTRAN_NAME( dcopy,  DCOPY  )
-#define blasf77_dgemm      FORTRAN_NAME( dgemm,  DGEMM  )
-#define blasf77_dgemv      FORTRAN_NAME( dgemv,  DGEMV  )
-#define blasf77_dger      FORTRAN_NAME( dger,  DGER  )
-#define blasf77_dger      FORTRAN_NAME( dger,  DGER  )
-#define blasf77_dsymm      FORTRAN_NAME( dsymm,  DSYMM  )
-#define blasf77_dsymv      FORTRAN_NAME( dsymv,  DSYMV  )
-#define blasf77_dsyr2      FORTRAN_NAME( dsyr2,  DSYR2  )
-#define blasf77_dsyr2k     FORTRAN_NAME( dsyr2k, DSYR2K )
-#define blasf77_dsyrk      FORTRAN_NAME( dsyrk,  DSYRK  )
-#define blasf77_dscal      FORTRAN_NAME( dscal,  DSCAL  )
-#define blasf77_ddscal     FORTRAN_NAME( dscal, DSCAL )
-#define blasf77_dswap      FORTRAN_NAME( dswap,  DSWAP  )
-#define blasf77_dsymm      FORTRAN_NAME( dsymm,  DSYMM  )
-#define blasf77_dsyr2k     FORTRAN_NAME( dsyr2k, DSYR2K )
-#define blasf77_dsyrk      FORTRAN_NAME( dsyrk,  DSYRK  )
-#define blasf77_dtrmm      FORTRAN_NAME( dtrmm,  DTRMM  )
-#define blasf77_dtrmv      FORTRAN_NAME( dtrmv,  DTRMV  )
-#define blasf77_dtrsm      FORTRAN_NAME( dtrsm,  DTRSM  )
-#define blasf77_dtrsv      FORTRAN_NAME( dtrsv,  DTRSV  )
-
-#define lapackf77_dlaed4   FORTRAN_NAME( dlaed4, DLAED4 )
-#define lapackf77_dlamc3   FORTRAN_NAME( dlamc3, DLAMC3 )
-#define lapackf77_dlamrg   FORTRAN_NAME( dlamrg, DLAMRG )
-#define lapackf77_dstebz   FORTRAN_NAME( dstebz, DSTEBZ )
-
-#define lapackf77_dbdsqr   FORTRAN_NAME( dbdsqr, DBDSQR )
-#define lapackf77_dgebak   FORTRAN_NAME( dgebak, DGEBAK )
-#define lapackf77_dgebal   FORTRAN_NAME( dgebal, DGEBAL )
-#define lapackf77_dgebd2   FORTRAN_NAME( dgebd2, DGEBD2 )
-#define lapackf77_dgebrd   FORTRAN_NAME( dgebrd, DGEBRD )
-#define lapackf77_dgeev    FORTRAN_NAME( dgeev,  DGEEV  )
-#define lapackf77_dgehd2   FORTRAN_NAME( dgehd2, DGEHD2 )
-#define lapackf77_dgehrd   FORTRAN_NAME( dgehrd, DGEHRD )
-#define lapackf77_dgelqf   FORTRAN_NAME( dgelqf, DGELQF )
-#define lapackf77_dgels    FORTRAN_NAME( dgels,  DGELS  )
-#define lapackf77_dgeqlf   FORTRAN_NAME( dgeqlf, DGEQLF )
-#define lapackf77_dgeqp3   FORTRAN_NAME( dgeqp3, DGEQP3 )
-#define lapackf77_dgeqrf   FORTRAN_NAME( dgeqrf, DGEQRF )
-#define lapackf77_dgesvd   FORTRAN_NAME( dgesvd, DGESVD )
-#define lapackf77_dgetrf   FORTRAN_NAME( dgetrf, DGETRF )
-#define lapackf77_dgetri   FORTRAN_NAME( dgetri, DGETRI )
-#define lapackf77_dgetrs   FORTRAN_NAME( dgetrs, DGETRS )
-#define lapackf77_dsbtrd   FORTRAN_NAME( dsbtrd, DSBTRD )
-#define lapackf77_dsyev    FORTRAN_NAME( dsyev,  DSYEV  )
-#define lapackf77_dsyevd   FORTRAN_NAME( dsyevd, DSYEVD )
-#define lapackf77_dsygs2   FORTRAN_NAME( dsygs2, DSYGS2 )
-#define lapackf77_dsygvd   FORTRAN_NAME( dsygvd, DSYGVD )
-#define lapackf77_dsytd2   FORTRAN_NAME( dsytd2, DSYTD2 )
-#define lapackf77_dsytrd   FORTRAN_NAME( dsytrd, DSYTRD )
-#define lapackf77_dsytrf   FORTRAN_NAME( dsytrf, DSYTRF )
-#define lapackf77_dhseqr   FORTRAN_NAME( dhseqr, DHSEQR )
-#define lapackf77_dlabrd   FORTRAN_NAME( dlabrd, DLABRD )
-#define lapackf77_dladiv   FORTRAN_NAME( dladiv, ZLADIV )
-#define lapackf77_dlacgv   FORTRAN_NAME( dlacgv, DLACGV )
-#define lapackf77_dlacpy   FORTRAN_NAME( dlacpy, DLACPY )
-#define lapackf77_dlasyf   FORTRAN_NAME( dlasyf, DLASYF )
-#define lapackf77_dlange   FORTRAN_NAME( dlange, DLANGE )
-#define lapackf77_dlansy   FORTRAN_NAME( dlansy, DLANSY )
-#define lapackf77_dlanst   FORTRAN_NAME( dlanst, DLANST )
-#define lapackf77_dlansy   FORTRAN_NAME( dlansy, DLANSY )
-#define lapackf77_dlapy3   FORTRAN_NAME( dlapy3, DLAPY3 )
-#define lapackf77_dlaqp2   FORTRAN_NAME( dlaqp2, DLAQP2 )
-#define lapackf77_dlarf    FORTRAN_NAME( dlarf,  ZLARF  )
-#define lapackf77_dlarfb   FORTRAN_NAME( dlarfb, DLARFB )
-#define lapackf77_dlarfg   FORTRAN_NAME( dlarfg, DLARFG )
-#define lapackf77_dlarft   FORTRAN_NAME( dlarft, DLARFT )
-#define lapackf77_dlarnv   FORTRAN_NAME( dlarnv, DLARNV )
-#define lapackf77_dlartg   FORTRAN_NAME( dlartg, DLARTG )
-#define lapackf77_dlascl   FORTRAN_NAME( dlascl, DLASCL )
-#define lapackf77_dlaset   FORTRAN_NAME( dlaset, DLASET )
-#define lapackf77_dlaswp   FORTRAN_NAME( dlaswp, DLASWP )
-#define lapackf77_dlatrd   FORTRAN_NAME( dlatrd, DLATRD )
-#define lapackf77_dlauum   FORTRAN_NAME( dlauum, DLAUUM )
-#define lapackf77_dlavsy   FORTRAN_NAME( dlavsy, DLAVSY )
-#define lapackf77_dpotrf   FORTRAN_NAME( dpotrf, DPOTRF )
-#define lapackf77_dpotri   FORTRAN_NAME( dpotri, DPOTRI )
-#define lapackf77_dpotrs   FORTRAN_NAME( dpotrs, DPOTRS )
-#define lapackf77_dstedc   FORTRAN_NAME( dstedc, DSTEDC )
-#define lapackf77_dstein   FORTRAN_NAME( dstein, DSTEIN )
-#define lapackf77_dstemr   FORTRAN_NAME( dstemr, DSTEMR )
-#define lapackf77_dsteqr   FORTRAN_NAME( dsteqr, DSTEQR )
-#define lapackf77_dsymv    FORTRAN_NAME( dsymv,  DSYMV  )
-#define lapackf77_dtrevc   FORTRAN_NAME( dtrevc, DTREVC )
-#define lapackf77_dtrtri   FORTRAN_NAME( dtrtri, DTRTRI )
-#define lapackf77_dorg2r   FORTRAN_NAME( dorg2r, DORG2R )
-#define lapackf77_dorgbr   FORTRAN_NAME( dorgbr, DORGBR )
-#define lapackf77_dorghr   FORTRAN_NAME( dorghr, DORGHR )
-#define lapackf77_dorglq   FORTRAN_NAME( dorglq, DORGLQ )
-#define lapackf77_dorgql   FORTRAN_NAME( dorgql, DORGQL )
-#define lapackf77_dorgqr   FORTRAN_NAME( dorgqr, DORGQR )
-#define lapackf77_dorgtr   FORTRAN_NAME( dorgtr, DORGTR )
-#define lapackf77_dorm2r   FORTRAN_NAME( dorm2r, DORM2R )
-#define lapackf77_dormbr   FORTRAN_NAME( dormbr, DORMBR )
-#define lapackf77_dormlq   FORTRAN_NAME( dormlq, DORMLQ )
-#define lapackf77_dormql   FORTRAN_NAME( dormql, DORMQL )
-#define lapackf77_dormqr   FORTRAN_NAME( dormqr, DORMQR )
-#define lapackf77_dormtr   FORTRAN_NAME( dormtr, DORMTR )
-
-/* testing functions */
-#define lapackf77_dbdt01   FORTRAN_NAME( dbdt01, DBDT01 )
-#define lapackf77_dget22   FORTRAN_NAME( dget22, DGET22 )
-#define lapackf77_dqpt01   FORTRAN_NAME( dqpt01, DQPT01 )
-#define lapackf77_dsyt21   FORTRAN_NAME( dsyt21, DSYT21 )
-#define lapackf77_dhst01   FORTRAN_NAME( dhst01, DHST01 )
-#define lapackf77_dqrt02   FORTRAN_NAME( dqrt02, DQRT02 )
-#define lapackf77_dort01   FORTRAN_NAME( dort01, DORT01 )
-#define lapackf77_dlarfy   FORTRAN_NAME( dlarfy, DLARFY )
-#define lapackf77_dlarfx   FORTRAN_NAME( dlarfx, DLARFX )
-#define lapackf77_dstt21   FORTRAN_NAME( dstt21, DSTT21 )
-
-// macros to handle differences in arguments between real and real versions of routines.
-#if defined(PRECISION_z) || defined(PRECISION_c)
-#define DWORKFORZ        double *rwork,
-#define DWORKFORZ_AND_LD double *rwork, const magma_int_t *ldrwork,
-#define WSPLIT           double *w
-#else
-#define DWORKFORZ
-#define DWORKFORZ_AND_LD
-#define WSPLIT           double *wr, double *wi
-#endif
-
-/*
- * BLAS functions (alphabetical order)
- */
-magma_int_t blasf77_idamax(
-                     const magma_int_t *n,
-                     const double *x, const magma_int_t *incx);
-
-void blasf77_daxpy(  const magma_int_t *n,
-                     const double *alpha,
-                     const double *x, const magma_int_t *incx,
-                           double *y, const magma_int_t *incy );
-
-void blasf77_dcopy(  const magma_int_t *n,
-                     const double *x, const magma_int_t *incx,
-                           double *y, const magma_int_t *incy );
-
-void blasf77_dgemm(  const char *transa, const char *transb,
-                     const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                     const double *B, const magma_int_t *ldb,
-                     const double *beta,
-                           double *C, const magma_int_t *ldc );
-
-void blasf77_dgemv(  const char *transa,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                     const double *x, const magma_int_t *incx,
-                     const double *beta,
-                           double *y, const magma_int_t *incy );
-
-void blasf77_dger(  const magma_int_t *m, const magma_int_t *n,
-                     const double *alpha,
-                     const double *x, const magma_int_t *incx,
-                     const double *y, const magma_int_t *incy,
-                           double *A, const magma_int_t *lda );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void blasf77_dger(  const magma_int_t *m, const magma_int_t *n,
-                     const double *alpha,
-                     const double *x, const magma_int_t *incx,
-                     const double *y, const magma_int_t *incy,
-                           double *A, const magma_int_t *lda );
-#endif
-
-void blasf77_dsymm(  const char *side, const char *uplo,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                     const double *B, const magma_int_t *ldb,
-                     const double *beta,
-                           double *C, const magma_int_t *ldc );
-
-void blasf77_dsymv(  const char *uplo,
-                     const magma_int_t *n,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                     const double *x, const magma_int_t *incx,
-                     const double *beta,
-                           double *y, const magma_int_t *incy );
-
-void blasf77_dsyr2(  const char *uplo,
-                     const magma_int_t *n,
-                     const double *alpha,
-                     const double *x, const magma_int_t *incx,
-                     const double *y, const magma_int_t *incy,
-                           double *A, const magma_int_t *lda );
-
-void blasf77_dsyr2k(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                     const double *B, const magma_int_t *ldb,
-                     const double *beta,
-                           double *C, const magma_int_t *ldc );
-
-void blasf77_dsyrk(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                     const double *beta,
-                           double *C, const magma_int_t *ldc );
-
-void blasf77_dscal(  const magma_int_t *n,
-                     const double *alpha,
-                           double *x, const magma_int_t *incx );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void blasf77_ddscal( const magma_int_t *n,
-                     const double *alpha,
-                           double *x, const magma_int_t *incx );
-#endif
-
-void blasf77_dswap(  const magma_int_t *n,
-                     double *x, const magma_int_t *incx,
-                     double *y, const magma_int_t *incy );
-
-void blasf77_dsymm(  const char *side, const char *uplo,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                     const double *B, const magma_int_t *ldb,
-                     const double *beta,
-                           double *C, const magma_int_t *ldc );
-
-void blasf77_dsyr2k( const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                     const double *B, const magma_int_t *ldb,
-                     const double *beta,
-                           double *C, const magma_int_t *ldc );
-
-void blasf77_dsyrk(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                     const double *beta,
-                           double *C, const magma_int_t *ldc );
-
-void blasf77_dtrmm(  const char *side, const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                           double *B, const magma_int_t *ldb );
-
-void blasf77_dtrmv(  const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *n,
-                     const double *A, const magma_int_t *lda,
-                           double *x, const magma_int_t *incx );
-
-void blasf77_dtrsm(  const char *side, const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const double *alpha,
-                     const double *A, const magma_int_t *lda,
-                           double *B, const magma_int_t *ldb );
-
-void blasf77_dtrsv(  const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *n,
-                     const double *A, const magma_int_t *lda,
-                           double *x, const magma_int_t *incx );
-
-/*
- * LAPACK functions (alphabetical order)
- */
-void   lapackf77_dbdsqr( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *ncvt, const magma_int_t *nru,  const magma_int_t *ncc,
-                         double *d, double *e,
-                         double *Vt, const magma_int_t *ldvt,
-                         double *U, const magma_int_t *ldu,
-                         double *C, const magma_int_t *ldc,
-                         double *work,
-                         magma_int_t *info );
-
-void   lapackf77_dgebak( const char *job, const char *side,
-                         const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         const double *scale, const magma_int_t *m,
-                         double *V, const magma_int_t *ldv,
-                         magma_int_t *info );
-
-void   lapackf77_dgebal( const char *job,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         magma_int_t *ilo, magma_int_t *ihi,
-                         double *scale,
-                         magma_int_t *info );
-
-void   lapackf77_dgebd2( const magma_int_t *m, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         double *tauq,
-                         double *taup,
-                         double *work,
-                         magma_int_t *info );
-
-void   lapackf77_dgebrd( const magma_int_t *m, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         double *tauq,
-                         double *taup,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dgeev(  const char *jobvl, const char *jobvr,
-                         const magma_int_t *n,
-                         double *A,    const magma_int_t *lda,
-                         WSPLIT,
-                         double *Vl,   const magma_int_t *ldvl,
-                         double *Vr,   const magma_int_t *ldvr,
-                         double *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_dgehd2( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         double *A, const magma_int_t *lda,
-                         double *tau,
-                         double *work,
-                         magma_int_t *info );
-
-void   lapackf77_dgehrd( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         double *A, const magma_int_t *lda,
-                         double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dgelqf( const magma_int_t *m, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dgels(  const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *nrhs,
-                         double *A, const magma_int_t *lda,
-                         double *B, const magma_int_t *ldb,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dgeqlf( const magma_int_t *m, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dgeqp3( const magma_int_t *m, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         magma_int_t *jpvt,
-                         double *tau,
-                         double *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_dgeqrf( const magma_int_t *m, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dgesvd( const char *jobu, const char *jobvt,
-                         const magma_int_t *m, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *s,
-                         double *U,  const magma_int_t *ldu,
-                         double *Vt, const magma_int_t *ldvt,
-                         double *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_dgetrf( const magma_int_t *m, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         magma_int_t *info );
-
-void   lapackf77_dgetri( const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         const magma_int_t *ipiv,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dgetrs( const char* trans,
-                         const magma_int_t *n, const magma_int_t *nrhs,
-                         const double *A, const magma_int_t *lda,
-                         const magma_int_t *ipiv,
-                         double *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_dsbtrd( const char *vect, const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kd,
-                         double *Ab, const magma_int_t *ldab,
-                         double *d, double *e,
-                         double *Q, const magma_int_t *ldq,
-                         double *work,
-                         magma_int_t *info );
-
-void   lapackf77_dsyev(  const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *w,
-                         double *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_dsyevd( const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *w,
-                         double *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_dsygs2( const magma_int_t *itype, const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_dsygvd( const magma_int_t *itype, const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *B, const magma_int_t *ldb,
-                         double *w,
-                         double *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_dsytd2( const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         double *tau,
-                         magma_int_t *info );
-
-void   lapackf77_dsytrd( const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dsytrf( const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dhseqr( const char *job, const char *compz,
-                         const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         double *H, const magma_int_t *ldh,
-                         WSPLIT,
-                         double *Z, const magma_int_t *ldz,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dlabrd( const magma_int_t *m, const magma_int_t *n, const magma_int_t *nb,
-                         double *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         double *tauq,
-                         double *taup,
-                         double *X, const magma_int_t *ldx,
-                         double *Y, const magma_int_t *ldy );
-
-void   lapackf77_dladiv( double *ret_val, double *x, 
-                         double *y );
-
-void   lapackf77_dlacgv( const magma_int_t *n,
-                         double *x, const magma_int_t *incx );
-
-void   lapackf77_dlacpy( const char *uplo,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const double *A, const magma_int_t *lda,
-                         double *B, const magma_int_t *ldb );
-
-void   lapackf77_dlasyf( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kn,
-                         magma_int_t *kb,
-                         double *A, const magma_int_t lda,
-                         magma_int_t *ipiv,
-                         double *work, const magma_int_t *ldwork,
-                         magma_int_t *info );
-
-double lapackf77_dlange( const char *norm,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const double *A, const magma_int_t *lda,
-                         double *work );
-
-double lapackf77_dlansy( const char *norm, const char *uplo,
-                         const magma_int_t *n,
-                         const double *A, const magma_int_t *lda,
-                         double * work );
-
-double lapackf77_dlanst( const char* norm, const magma_int_t* n,
-                         const double* d, const double* e );
-
-double lapackf77_dlansy( const char *norm, const char *uplo,
-                         const magma_int_t *n,
-                         const double *A, const magma_int_t *lda,
-                         double * work );
-
-void lapackf77_dlaqp2 (  magma_int_t *m, magma_int_t *n, magma_int_t *offset,
-                         double *a, magma_int_t *lda, magma_int_t *jpvt, 
-                         double *tau,
-                         double *vn1, double *vn2, double *work);
-
-void lapackf77_dlarf  (  char *, magma_int_t *, magma_int_t *,
-                         double *, magma_int_t *, double *, double *,
-                         magma_int_t *, double *);
-
-void   lapackf77_dlarfb( const char *side, const char *trans, const char *direct, const char *storev,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const double *V, const magma_int_t *ldv,
-                         const double *T, const magma_int_t *ldt,
-                         double *C, const magma_int_t *ldc,
-                         double *work, const magma_int_t *ldwork );
-
-void   lapackf77_dlarfg( const magma_int_t *n,
-                         double *alpha,
-                         double *x, const magma_int_t *incx,
-                         double *tau );
-
-void   lapackf77_dlarft( const char *direct, const char *storev,
-                         const magma_int_t *n, const magma_int_t *k,
-                         double *V, const magma_int_t *ldv,
-                         const double *tau,
-                         double *T, const magma_int_t *ldt );
-
-void   lapackf77_dlarnv( const magma_int_t *idist, magma_int_t *iseed, const magma_int_t *n,
-                         double *x );
-
-void   lapackf77_dlartg( double *F,
-                         double *G,
-                         double *cs,
-                         double *SN,
-                         double *R );
-
-void   lapackf77_dlascl( const char *type,
-                         const magma_int_t *kl, const magma_int_t *ku,
-                         double *cfrom,
-                         double *cto,
-                         const magma_int_t *m, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_dlaset( const char *uplo,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const double *alpha,
-                         const double *beta,
-                         double *A, const magma_int_t *lda );
-
-void   lapackf77_dlaswp( const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         const magma_int_t *k1, const magma_int_t *k2,
-                         magma_int_t *ipiv,
-                         const magma_int_t *incx );
-
-void   lapackf77_dlatrd( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *nb,
-                         double *A, const magma_int_t *lda,
-                         double *e,
-                         double *tau,
-                         double *work, const magma_int_t *ldwork );
-
-void   lapackf77_dlauum( const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_dlavsy( const char *uplo, const char *trans, const char *diag,
-                         magma_int_t *n, magma_int_t *nrhs,
-                         double *A, magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         double *B, magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_dpotrf( const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_dpotri( const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_dpotrs( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *nrhs,
-                         const double *A, const magma_int_t *lda,
-                         double *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_dstedc( const char *compz,
-                         const magma_int_t *n,
-                         double *d, double *e,
-                         double *Z, const magma_int_t *ldz,
-                         double *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_dstein( const magma_int_t *n,
-                         const double *d, const double *e,
-                         const magma_int_t *m,
-                         const double *w,
-                         const magma_int_t *iblock,
-                         const magma_int_t *isplit,
-                         double *Z, const magma_int_t *ldz,
-                         double *work, magma_int_t *iwork, magma_int_t *ifailv,
-                         magma_int_t *info );
-
-void   lapackf77_dstemr( const char *jobz, const char *range,
-                         const magma_int_t *n,
-                         double *d, double *e,
-                         const double *vl, const double *vu,
-                         const magma_int_t *il, const magma_int_t *iu,
-                         magma_int_t *m,
-                         double *w,
-                         double *Z, const magma_int_t *ldz,
-                         const magma_int_t *nzc, magma_int_t *isuppz, magma_int_t *tryrac,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_dsteqr( const char *compz,
-                         const magma_int_t *n,
-                         double *d, double *e,
-                         double *Z, const magma_int_t *ldz,
-                         double *work,
-                         magma_int_t *info );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void   lapackf77_dsymv(  const char *uplo,
-                         const magma_int_t *n,
-                         const double *alpha,
-                         const double *A, const magma_int_t *lda,
-                         const double *x, const magma_int_t *incx,
-                         const double *beta,
-                               double *y, const magma_int_t *incy );
-#endif
-
-void   lapackf77_dtrevc( const char *side, const char *howmny,
-                         magma_int_t *select, const magma_int_t *n,
-                         double *T,  const magma_int_t *ldt,
-                         double *Vl, const magma_int_t *ldvl,
-                         double *Vr, const magma_int_t *ldvr,
-                         const magma_int_t *mm, magma_int_t *m,
-                         double *work,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_dtrtri( const char *uplo, const char *diag,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_dorg2r( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *work,
-                         magma_int_t *info );
-
-void   lapackf77_dorgbr( const char *vect,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dorghr( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dorglq( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dorgql( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dorgqr( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dorgtr( const char *uplo,
-                         const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dorm2r( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *C, const magma_int_t *ldc,
-                         double *work,
-                         magma_int_t *info );
-
-void   lapackf77_dormbr( const char *vect, const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *C, const magma_int_t *ldc,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dormlq( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *C, const magma_int_t *ldc,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dormql( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *C, const magma_int_t *ldc,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dormqr( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *C, const magma_int_t *ldc,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_dormtr( const char *side, const char *uplo, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const double *A, const magma_int_t *lda,
-                         const double *tau,
-                         double *C, const magma_int_t *ldc,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-/*
- * Real precision extras
- */
-void   lapackf77_dstebz( const char *range, const char *order,
-                         const magma_int_t *n,
-                         double *vl, double *vu,
-                         magma_int_t *il, magma_int_t *iu,
-                         double *abstol,
-                         double *d, double *e,
-                         const magma_int_t *m, const magma_int_t *nsplit,
-                         double *w,
-                         magma_int_t *iblock, magma_int_t *isplit,
-                         double *work,
-                         magma_int_t *iwork,
-                         magma_int_t *info );
-
-double lapackf77_dlamc3( double* a, double* b );
-
-void   lapackf77_dlamrg( magma_int_t* n1, magma_int_t* n2,
-                         double* a,
-                         magma_int_t* dtrd1, magma_int_t* dtrd2, magma_int_t* index );
-
-double lapackf77_dlapy3(double *, double *, double *);
-
-void   lapackf77_dlaed4( magma_int_t* n, magma_int_t* i,
-                         double* d,
-                         double* z,
-                         double* delta,
-                         double* rho,
-                         double* dlam, magma_int_t* info );
-
-/*
- * Testing functions
- */
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void   lapackf77_dbdt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *kd,
-                         double *A, const magma_int_t *lda,
-                         double *Q, const magma_int_t *ldq,
-                         double *d, double *e,
-                         double *Pt, const magma_int_t *ldpt,
-                         double *work,
-                         double *rwork,
-                         double *resid );
-
-void   lapackf77_dget22( const char *transa, const char *transe, const char *transw, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *E, const magma_int_t *lde,
-                         double *w,
-                         double *work,
-                         double *rwork,
-                         double *result );
-
-void   lapackf77_dsyt21( const magma_int_t *itype, const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kband,
-                         double *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         double *U, const magma_int_t *ldu,
-                         double *V, const magma_int_t *ldv,
-                         double *tau,
-                         double *work,
-                         double *rwork,
-                         double *result );
-
-void   lapackf77_dhst01( const magma_int_t *n, const magma_int_t *ilo, const magma_int_t *ihi,
-                         double *A, const magma_int_t *lda,
-                         double *H, const magma_int_t *ldh,
-                         double *Q, const magma_int_t *ldq,
-                         double *work, const magma_int_t *lwork,
-                         double *rwork,
-                         double *result );
-
-void   lapackf77_dstt21( const magma_int_t *n, const magma_int_t *kband,
-                         double *AD,
-                         double *AE,
-                         double *SD,
-                         double *SE,
-                         double *U, const magma_int_t *ldu,
-                         double *work,
-                         double *rwork,
-                         double *result );
-
-void   lapackf77_dort01( const char *rowcol, const magma_int_t *m, const magma_int_t *n,
-                         double *U, const magma_int_t *ldu,
-                         double *work, const magma_int_t *lwork,
-                         double *rwork,
-                         double *resid );
-#else
-void   lapackf77_dbdt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *kd,
-                         double *A, const magma_int_t *lda,
-                         double *Q, const magma_int_t *ldq,
-                         double *d, double *e,
-                         double *Pt, const magma_int_t *ldpt,
-                         double *work,
-                         double *resid );
-
-void   lapackf77_dget22( const char *transa, const char *transe, const char *transw, const magma_int_t *n,
-                         double *A, const magma_int_t *lda,
-                         double *E, const magma_int_t *lde,
-                         double *wr,
-                         double *wi,
-                         double *work,
-                         double *result );
-
-void   lapackf77_dsyt21( magma_int_t *itype, const char *uplo, const magma_int_t *n, const magma_int_t *kband,
-                         double *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         double *U, const magma_int_t *ldu,
-                         double *V, const magma_int_t *ldv,
-                         double *tau,
-                         double *work,
-                         double *result );
-
-void   lapackf77_dhst01( const magma_int_t *n, const magma_int_t *ilo, const magma_int_t *ihi,
-                         double *A, const magma_int_t *lda,
-                         double *H, const magma_int_t *ldh,
-                         double *Q, const magma_int_t *ldq,
-                         double *work, const magma_int_t *lwork,
-                         double *result );
-
-void   lapackf77_dstt21( const magma_int_t *n, const magma_int_t *kband,
-                         double *AD,
-                         double *AE,
-                         double *SD,
-                         double *SE,
-                         double *U, const magma_int_t *ldu,
-                         double *work,
-                         double *result );
-
-void   lapackf77_dort01( const char *rowcol, const magma_int_t *m, const magma_int_t *n,
-                         double *U, const magma_int_t *ldu,
-                         double *work, const magma_int_t *lwork,
-                         double *resid );
-#endif
-
-void   lapackf77_dlarfy( const char *uplo, const magma_int_t *n,
-                         double *V, const magma_int_t *incv,
-                         double *tau,
-                         double *C, const magma_int_t *ldc,
-                         double *work );
-
-void   lapackf77_dlarfx( const char *side, const magma_int_t *m, const magma_int_t *n,
-                         double *V,
-                         double *tau,
-                         double *C, const magma_int_t *ldc,
-                         double *work );
-
-double lapackf77_dqpt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         double *A,
-                         double *Af, const magma_int_t *lda,
-                         double *tau, magma_int_t *jpvt,
-                         double *work, const magma_int_t *lwork );
-
-void   lapackf77_dqrt02( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         double *A,
-                         double *AF,
-                         double *Q,
-                         double *R, const magma_int_t *lda,
-                         double *tau,
-                         double *work, const magma_int_t *lwork,
-                         double *rwork,
-                         double *result );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef DWORKFORZ
-#undef DWORKFORZ_AND_LD
-#undef WSPLIT
-#undef PRECISION_d
-
-#endif /* MAGMA_DLAPACK_H */
diff --git a/src/include/magma_ds.h b/src/include/magma_ds.h
deleted file mode 100644
index ce57163..0000000
--- a/src/include/magma_ds.h
+++ /dev/null
@@ -1,53 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated ds Wed Nov 14 22:52:26 2012
- */
-
-#ifndef _MAGMA_DS_H_
-#define _MAGMA_DS_H_
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* Mixed precision */
-magma_int_t magma_dsgesv_gpu(  char trans, magma_int_t N, magma_int_t NRHS, 
-                               double *dA, magma_int_t ldda, 
-                               magma_int_t *IPIV, magma_int_t *dIPIV,
-                               double *dB, magma_int_t lddb, 
-                               double *dX, magma_int_t lddx, 
-                               double *dworkd, float *dworks,
-                               magma_int_t *iter, magma_int_t *info);
-
-magma_int_t magma_dsgetrs_gpu( char trans, magma_int_t n, magma_int_t nrhs, 
-                               float  *dA, magma_int_t ldda,
-                               magma_int_t *ipiv, 
-                               double *dB, magma_int_t lddb,
-                               double *dX, magma_int_t lddx,
-                               float  *dSX, 
-                               magma_int_t *info );
-
-magma_int_t magma_dsposv_gpu( char uplo, magma_int_t n, magma_int_t nrhs, 
-                              double *dA, magma_int_t ldda, 
-                              double *dB, magma_int_t lddb, 
-                              double *dX, magma_int_t lddx, 
-                              double *dworkd, float *dworks,
-                              magma_int_t *iter, magma_int_t *info);
-
-magma_int_t magma_dsgeqrsv_gpu(magma_int_t M, magma_int_t N, magma_int_t NRHS, 
-                               double *dA,  magma_int_t ldda, 
-                               double *dB,  magma_int_t lddb, 
-                               double *dX,  magma_int_t lddx,
-                               magma_int_t *iter,    magma_int_t *info);
-  
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif /* _MAGMA_D_H_ */
diff --git a/src/include/magma_lapack.h b/src/include/magma_lapack.h
deleted file mode 100644
index e168b2d..0000000
--- a/src/include/magma_lapack.h
+++ /dev/null
@@ -1,113 +0,0 @@
-#ifndef MAGMA_LAPACK_H
-#define MAGMA_LAPACK_H
-
-#ifndef FORTRAN_NAME
-#if defined(ADD_)
-#define FORTRAN_NAME(lcname, UCNAME)  lcname##_
-#elif defined(NOCHANGE)
-#define FORTRAN_NAME(lcname, UCNAME)  lcname
-#elif defined(UPCASE)
-#define FORTRAN_NAME(lcname, UCNAME)  UCNAME
-#else
-#error Define one of ADD_, NOCHANGE, or UPCASE for how Fortran functions are name mangled.
-#endif
-#endif
-
-#include "magma_zlapack.h"
-#include "magma_clapack.h"
-#include "magma_dlapack.h"
-#include "magma_slapack.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#define lapackf77_ieeeck FORTRAN_NAME( ieeeck, IEEECK )
-#define lapackf77_lsame  FORTRAN_NAME( lsame,  LSAME  )
-#define lapackf77_xerbla FORTRAN_NAME( xerbla, XERBLA )
-
-#define lapackf77_slamch FORTRAN_NAME( slamch, SLAMCH )
-#define lapackf77_dlamch FORTRAN_NAME( dlamch, DLAMCH )
-#define lapackf77_slabad FORTRAN_NAME( slabad, SLABAD )
-#define lapackf77_dlabad FORTRAN_NAME( dlabad, DLABAD )
-#define lapackf77_zcgesv FORTRAN_NAME( zcgesv, ZCGESV )
-#define lapackf77_dsgesv FORTRAN_NAME( dsgesv, DSGESV )
-
-#define lapackf77_dsterf FORTRAN_NAME( dsterf, DSTERF )
-#define lapackf77_ssterf FORTRAN_NAME( ssterf, SSTERF )
-
-#define lapackf77_zlag2c FORTRAN_NAME( zlag2c, ZLAG2C )
-#define lapackf77_clag2z FORTRAN_NAME( clag2z, CLAG2Z )
-#define lapackf77_dlag2s FORTRAN_NAME( dlag2s, DLAG2S )
-#define lapackf77_slag2d FORTRAN_NAME( slag2d, SLAG2D )
-
-#define lapackf77_dlapy2 FORTRAN_NAME( dlapy2, DLAPY2 )
-#define lapackf77_slapy2 FORTRAN_NAME( slapy2, SLAPY2 )
-
-magma_int_t lapackf77_ieeeck( magma_int_t* ispec, float* zero, float* one );
-
-long   lapackf77_lsame(  const char *ca, const char *cb );
-
-void   lapackf77_xerbla( const char *name, magma_int_t *info, magma_int_t name_len );
-
-float  lapackf77_slamch( const char *cmach );
-double lapackf77_dlamch( const char *cmach );
-
-// "small" (lowercase) defined as char on Windows (reported by MathWorks)
-void   lapackf77_slabad( float  *Small, float  *large );
-void   lapackf77_dlabad( double *Small, double *large );
-
-void   lapackf77_zcgesv( const magma_int_t *n, const magma_int_t *nrhs,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         const cuDoubleComplex *B, const magma_int_t *ldb,
-                               cuDoubleComplex *X, const magma_int_t *ldx,
-                         cuDoubleComplex *work, cuFloatComplex *swork, double *rwork,
-                         magma_int_t *iter,
-                         magma_int_t *info );
-
-void   lapackf77_dsgesv( const magma_int_t *n, const magma_int_t *nrhs,
-                         double *A, const magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         const double *B, const magma_int_t *ldb,
-                               double *X, const magma_int_t *ldx,
-                         double *work, float *swork,
-                         magma_int_t *iter,
-                         magma_int_t *info );
-
-void   lapackf77_dsterf( const magma_int_t *n,
-                         double *d, double *e,
-                         magma_int_t *info );
-
-void   lapackf77_ssterf( const magma_int_t *n,
-                         float *d, float *e,
-                         magma_int_t *info );
-
-void   lapackf77_zlag2c( magma_int_t *m, magma_int_t *n,
-                         const cuDoubleComplex *A,  magma_int_t *lda,
-                               cuFloatComplex  *SA, magma_int_t *ldsa,
-                         magma_int_t *info );
-
-void   lapackf77_clag2z( magma_int_t *m, magma_int_t *n,
-                         const cuFloatComplex  *SA, magma_int_t *ldsa,
-                               cuDoubleComplex *A,  magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_dlag2s( magma_int_t *m, magma_int_t *n,
-                         const double *A,  magma_int_t *lda,
-                               float  *SA, magma_int_t *ldsa,
-                         magma_int_t *info );
-
-void   lapackf77_slag2d( magma_int_t *m, magma_int_t *n,
-                         const float  *SA, magma_int_t *ldsa,
-                               double *A,  magma_int_t *lda,
-                         magma_int_t *info );
-
-double lapackf77_dlapy2( const double *x, const double *y );
-float  lapackf77_slapy2( const float  *x, const float  *y );
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif /* MAGMA LAPACK */
diff --git a/src/include/magma_s.h b/src/include/magma_s.h
deleted file mode 100644
index ae14f01..0000000
--- a/src/include/magma_s.h
+++ /dev/null
@@ -1,514 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated s Wed Nov 14 22:52:26 2012
- */
-
-#ifndef _MAGMA_S_H_
-#define _MAGMA_S_H_
-#define PRECISION_s
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- MAGMA Auxiliary functions to get the NB used
-*/
-int magma_get_spotrf_nb(int m);
-int magma_get_sgetrf_nb(int m);
-int magma_get_sgetri_nb(int m);
-int magma_get_sgeqp3_nb(int m);
-int magma_get_sgeqrf_nb(int m);
-int magma_get_sgeqlf_nb(int m);
-int magma_get_sgehrd_nb(int m);
-int magma_get_ssytrd_nb(int m);
-int magma_get_sgelqf_nb(int m);
-int magma_get_sgebrd_nb(int m);
-int magma_get_ssygst_nb(int m);
-int magma_get_sgesvd_nb(int m);
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- MAGMA function definitions / Data on CPU
-*/
-magma_int_t magma_sgebrd( magma_int_t m, magma_int_t n, float *A, 
-                          magma_int_t lda, float *d, float *e,
-                          float *tauq,  float *taup, 
-                          float *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_sgehrd2(magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          float *A, magma_int_t lda, float *tau, 
-                          float *work, magma_int_t *lwork, magma_int_t *info);
-magma_int_t magma_sgehrd( magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          float *A, magma_int_t lda, float *tau,
-                          float *work, magma_int_t lwork,
-                          float *d_T, magma_int_t *info);
-magma_int_t magma_sgelqf( magma_int_t m, magma_int_t n, 
-                          float *A,    magma_int_t lda,   float *tau, 
-                          float *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_sgeqlf( magma_int_t m, magma_int_t n, 
-                          float *A,    magma_int_t lda,   float *tau, 
-                          float *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_sgeqrf( magma_int_t m, magma_int_t n, float *A, 
-                          magma_int_t lda, float *tau, float *work, 
-                          magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_sgeqr2( magma_int_t *m, magma_int_t *n, float *a,
-                          magma_int_t *lda, float *tau, float *work, 
-                          magma_int_t *info);
-magma_int_t magma_sgeqrf4(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
-                          float *a,    magma_int_t lda, float *tau,
-                          float *work, magma_int_t lwork, magma_int_t *info );
-magma_int_t magma_sgeqrf_ooc( magma_int_t m, magma_int_t n, float *A,
-                          magma_int_t lda, float *tau, float *work,
-                          magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_sgesv ( magma_int_t n, magma_int_t nrhs, 
-                          float *A, magma_int_t lda, magma_int_t *ipiv, 
-                          float *B, magma_int_t ldb, magma_int_t *info);
-magma_int_t magma_sgetrf( magma_int_t m, magma_int_t n, float *A, 
-                          magma_int_t lda, magma_int_t *ipiv, 
-                          magma_int_t *info);
-magma_int_t magma_sgetrf2(magma_int_t m, magma_int_t n, float *a, 
-                          magma_int_t lda, magma_int_t *ipiv, magma_int_t *info);
-
-magma_int_t magma_slaqps( magma_int_t m, magma_int_t n, magma_int_t offset, 
-                          magma_int_t nb, magma_int_t *kb, 
-                          float *A,  magma_int_t lda,
-                          float *dA, magma_int_t ldda,
-                          magma_int_t *jpvt, float *tau, float *vn1, float *vn2, 
-                          float *auxv, 
-                          float *F,  magma_int_t ldf,
-                          float *dF, magma_int_t lddf );
-magma_int_t magma_slarf(  char *side, magma_int_t *m, magma_int_t *n,
-                          float *v, magma_int_t *incv, float *tau,
-                          float *c__, magma_int_t *ldc, float *work);
-magma_int_t magma_slarfg( magma_int_t *n, float *alpha, float *x,
-                          magma_int_t *incx, float *tau);
-magma_int_t magma_slatrd( char uplo, magma_int_t n, magma_int_t nb, float *a, 
-                          magma_int_t lda, float *e, float *tau, 
-                          float *w, magma_int_t ldw,
-                          float *da, magma_int_t ldda, 
-                          float *dw, magma_int_t lddw);
-magma_int_t magma_slatrd2(char uplo, magma_int_t n, magma_int_t nb,
-                          float *a,  magma_int_t lda,
-                          float *e, float *tau,
-                          float *w,  magma_int_t ldw,
-                          float *da, magma_int_t ldda,
-                          float *dw, magma_int_t lddw,
-                          float *dwork, magma_int_t ldwork);
-magma_int_t magma_slahr2( magma_int_t m, magma_int_t n, magma_int_t nb, 
-                          float *da, float *dv, float *a, 
-                          magma_int_t lda, float *tau, float *t, 
-                          magma_int_t ldt, float *y, magma_int_t ldy);
-magma_int_t magma_slahru( magma_int_t n, magma_int_t ihi, magma_int_t k, magma_int_t nb, 
-                          float *a, magma_int_t lda, 
-                          float *da, float *y, 
-                          float *v, float *t, 
-                          float *dwork);
-magma_int_t magma_sposv ( char uplo, magma_int_t n, magma_int_t nrhs, 
-                          float *A, magma_int_t lda, 
-                          float *B, magma_int_t ldb, magma_int_t *info);
-magma_int_t magma_spotrf( char uplo, magma_int_t n, float *A, 
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_spotri( char uplo, magma_int_t n, float *A,
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_slauum( char uplo, magma_int_t n, float *A,
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_strtri( char uplo, char diag, magma_int_t n, float *A, 
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_ssytrd( char uplo, magma_int_t n, float *A, 
-                          magma_int_t lda, float *d, float *e, 
-                          float *tau, float *work, magma_int_t lwork, 
-                          magma_int_t *info);
-magma_int_t magma_sorgqr( magma_int_t m, magma_int_t n, magma_int_t k,
-                          float *a, magma_int_t lda,
-                          float *tau, float *dwork,
-                          magma_int_t nb, magma_int_t *info );
-magma_int_t magma_sormql( const char side, const char trans,
-                          magma_int_t m, magma_int_t n, magma_int_t k,
-                          float *a, magma_int_t lda,
-                          float *tau,
-                          float *c, magma_int_t ldc,
-                          float *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_sormqr( char side, char trans, 
-                          magma_int_t m, magma_int_t n, magma_int_t k, 
-                          float *a, magma_int_t lda, float *tau, 
-                          float *c, magma_int_t ldc, 
-                          float *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_sormtr( char side, char uplo, char trans,
-                          magma_int_t m, magma_int_t n,
-                          float *a,    magma_int_t lda,
-                          float *tau,
-                          float *c,    magma_int_t ldc,
-                          float *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_sorghr( magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          float *a, magma_int_t lda,
-                          float *tau,
-                          float *dT, magma_int_t nb,
-                          magma_int_t *info);
-magma_int_t magma_ssyev( char jobz, char uplo, magma_int_t n,
-                         float *a, magma_int_t lda, float *w,
-                         float *work, magma_int_t lwork,
-                         float *rwork, magma_int_t *info);
-magma_int_t magma_ssyevx(char jobz, char range, char uplo, magma_int_t n,
-                         float *a, magma_int_t lda, float vl, float vu,
-                         magma_int_t il, magma_int_t iu, float abstol, magma_int_t *m,
-                         float *w, float *z, magma_int_t ldz, 
-                         float *work, magma_int_t lwork,
-                         float *rwork, magma_int_t *iwork, 
-                         magma_int_t *ifail, magma_int_t *info);
-#if defined(PRECISION_z) || defined(PRECISION_c)
-magma_int_t  magma_sgeev( char jobvl, char jobvr, magma_int_t n,
-                          float *a, magma_int_t lda,
-                          float *w,
-                          float *vl, magma_int_t ldvl,
-                          float *vr, magma_int_t ldvr,
-                          float *work, magma_int_t lwork,
-                          float *rwork, magma_int_t *info);
-magma_int_t magma_sgeqp3( magma_int_t m, magma_int_t n,
-                          float *a, magma_int_t lda, 
-                          magma_int_t *jpvt, float *tau,
-                          float *work, magma_int_t lwork, 
-                          float *rwork, magma_int_t *info);
-magma_int_t magma_sgesvd( char jobu, char jobvt, magma_int_t m, magma_int_t n,
-                          float *a,    magma_int_t lda, float *s, 
-                          float *u,    magma_int_t ldu, 
-                          float *vt,   magma_int_t ldvt,
-                          float *work, magma_int_t lwork,
-                          float *rwork, magma_int_t *info );
-magma_int_t magma_ssyevd( char jobz, char uplo, magma_int_t n,
-                          float *a, magma_int_t lda, float *w,
-                          float *work, magma_int_t lwork,
-                          float *rwork, magma_int_t lrwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_ssyevr( char jobz, char range, char uplo, magma_int_t n,
-                          float *a, magma_int_t lda, float vl, float vu,
-                          magma_int_t il, magma_int_t iu, float abstol, magma_int_t *m,
-                          float *w, float *z, magma_int_t ldz, 
-                          magma_int_t *isuppz,
-                          float *work, magma_int_t lwork,
-                          float *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_ssygvd( magma_int_t itype, char jobz, char uplo, magma_int_t n,
-                          float *a, magma_int_t lda,
-                          float *b, magma_int_t ldb,
-                          float *w, float *work, magma_int_t lwork,
-                          float *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_ssygvdx(magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, float *a, magma_int_t lda,
-                          float *b, magma_int_t ldb,
-                          float vl, float vu, magma_int_t il, magma_int_t iu,
-                          magma_int_t *m, float *w, float *work, 
-                          magma_int_t lwork, float *rwork,
-                          magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_ssygvx( magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, float *a, magma_int_t lda, 
-                          float *b, magma_int_t ldb,
-                          float vl, float vu, magma_int_t il, magma_int_t iu,
-                          float abstol, magma_int_t *m, float *w, 
-                          float *z, magma_int_t ldz,
-                          float *work, magma_int_t lwork, float *rwork,
-                          magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info);
-magma_int_t magma_ssygvr( magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, float *a, magma_int_t lda,
-                          float *b, magma_int_t ldb,
-                          float vl, float vu, magma_int_t il, magma_int_t iu,
-                          float abstol, magma_int_t *m, float *w, 
-                          float *z, magma_int_t ldz,
-                          magma_int_t *isuppz, float *work, magma_int_t lwork,
-                          float *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_sstedx( char range, magma_int_t n, float vl, float vu,
-                          magma_int_t il, magma_int_t iu, float *D, float *E,
-                          float *Z, magma_int_t ldz,
-                          float *rwork, magma_int_t ldrwork, magma_int_t *iwork,
-                          magma_int_t liwork, float* dwork, magma_int_t *info);
-#else
-magma_int_t  magma_sgeev( char jobvl, char jobvr, magma_int_t n,
-                          float *a,    magma_int_t lda,
-                          float *wr, float *wi,
-                          float *vl,   magma_int_t ldvl,
-                          float *vr,   magma_int_t ldvr,
-                          float *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_sgeqp3( magma_int_t m, magma_int_t n,
-                          float *a, magma_int_t lda, 
-                          magma_int_t *jpvt, float *tau,
-                          float *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_sgesvd( char jobu, char jobvt, magma_int_t m, magma_int_t n,
-                          float *a,    magma_int_t lda, float *s, 
-                          float *u,    magma_int_t ldu, 
-                          float *vt,   magma_int_t ldvt,
-                          float *work, magma_int_t lwork,
-                          magma_int_t *info );
-magma_int_t magma_ssyevd( char jobz, char uplo, magma_int_t n,
-                          float *a, magma_int_t lda, float *w,
-                          float *work, magma_int_t lwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_ssygvd( magma_int_t itype, char jobz, char uplo, magma_int_t n,
-                          float *a, magma_int_t lda,
-                          float *b, magma_int_t ldb,
-                          float *w, float *work, magma_int_t lwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_sstedx( char range, magma_int_t n, float vl, float vu,
-                          magma_int_t il, magma_int_t iu, float* d, float* e,
-                          float* z, magma_int_t ldz,
-                          float* work, magma_int_t lwork,
-                          magma_int_t* iwork, magma_int_t liwork,
-                          float* dwork, magma_int_t* info);
-magma_int_t magma_slaex0( magma_int_t n, float* d, float* e, float* q, magma_int_t ldq,
-                          float* work, magma_int_t* iwork, float* dwork,
-                          char range, float vl, float vu,
-                          magma_int_t il, magma_int_t iu, magma_int_t* info);
-magma_int_t magma_slaex1( magma_int_t n, float* d, float* q, magma_int_t ldq,
-                          magma_int_t* indxq, float rho, magma_int_t cutpnt,
-                          float* work, magma_int_t* iwork, float* dwork,
-                          char range, float vl, float vu,
-                          magma_int_t il, magma_int_t iu, magma_int_t* info);
-magma_int_t magma_slaex3( magma_int_t k, magma_int_t n, magma_int_t n1, float* d,
-                          float* q, magma_int_t ldq, float rho,
-                          float* dlamda, float* q2, magma_int_t* indx,
-                          magma_int_t* ctot, float* w, float* s, magma_int_t* indxq,
-                          float* dwork,
-                          char range, float vl, float vu, magma_int_t il, magma_int_t iu,
-                          magma_int_t* info );
-#endif
-
-magma_int_t magma_ssygst( magma_int_t itype, char uplo, magma_int_t n,
-                          float *a, magma_int_t lda,
-                          float *b, magma_int_t ldb, magma_int_t *info);
-
-/* //////////////////////////////////////////////////////////////////////////// 
- -- MAGMA function definitions / Data on GPU
-*/
-magma_int_t magma_sgels_gpu(  char trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
-                              float *dA,    magma_int_t ldda, 
-                              float *dB,    magma_int_t lddb, 
-                              float *hwork, magma_int_t lwork, 
-                              magma_int_t *info);
-magma_int_t magma_sgels3_gpu( char trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
-                              float *dA,    magma_int_t ldda,
-                              float *dB,    magma_int_t lddb,
-                              float *hwork, magma_int_t lwork,
-                              magma_int_t *info);
-magma_int_t magma_sgelqf_gpu( magma_int_t m, magma_int_t n,
-                              float *dA,    magma_int_t ldda,   float *tau,
-                              float *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_sgeqrf_gpu( magma_int_t m, magma_int_t n, 
-                              float *dA,  magma_int_t ldda, 
-                              float *tau, float *dT, 
-                              magma_int_t *info);
-magma_int_t magma_sgeqrf2_gpu(magma_int_t m, magma_int_t n, 
-                              float *dA,  magma_int_t ldda, 
-                              float *tau, magma_int_t *info);
-magma_int_t magma_sgeqrf2_mgpu(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
-                               float **dlA, magma_int_t ldda,
-                               float *tau, magma_int_t *info );
-magma_int_t magma_sgeqrf3_gpu(magma_int_t m, magma_int_t n, 
-                              float *dA,  magma_int_t ldda, 
-                              float *tau, float *dT, 
-                              magma_int_t *info);
-magma_int_t magma_sgeqrs_gpu( magma_int_t m, magma_int_t n, magma_int_t nrhs, 
-                              float *dA,     magma_int_t ldda, 
-                              float *tau,   float *dT,
-                              float *dB,    magma_int_t lddb,
-                              float *hwork, magma_int_t lhwork, 
-                              magma_int_t *info);
-magma_int_t magma_sgeqrs3_gpu( magma_int_t m, magma_int_t n, magma_int_t nrhs, 
-                              float *dA,     magma_int_t ldda, 
-                              float *tau,   float *dT,
-                              float *dB,    magma_int_t lddb,
-                              float *hwork, magma_int_t lhwork, 
-                              magma_int_t *info);
-magma_int_t magma_sgessm_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t k, magma_int_t ib, 
-                              magma_int_t *ipiv, 
-                              float *dL1, magma_int_t lddl1, 
-                              float *dL,  magma_int_t lddl, 
-                              float *dA,  magma_int_t ldda, 
-                              magma_int_t *info);
-magma_int_t magma_sgesv_gpu(  magma_int_t n, magma_int_t nrhs, 
-                              float *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              float *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_sgetrf_incpiv_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t ib,
-                              float *hA, magma_int_t ldha, float *dA, magma_int_t ldda,
-                              float *hL, magma_int_t ldhl, float *dL, magma_int_t lddl,
-                              magma_int_t *ipiv, 
-                              float *dwork, magma_int_t lddwork,
-                              magma_int_t *info);
-magma_int_t magma_sgetrf_gpu( magma_int_t m, magma_int_t n, 
-                              float *dA, magma_int_t ldda, 
-                              magma_int_t *ipiv, magma_int_t *info);
-magma_int_t 
-magma_sgetrf_nopiv_gpu      ( magma_int_t m, magma_int_t n,
-                              float *dA, magma_int_t ldda,
-                              magma_int_t *info);
-magma_int_t magma_sgetri_gpu( magma_int_t n, 
-                              float *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              float *dwork, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_sgetrs_gpu( char trans, magma_int_t n, magma_int_t nrhs, 
-                              float *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              float *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_slabrd_gpu( magma_int_t m, magma_int_t n, magma_int_t nb, 
-                              float *a, magma_int_t lda, float *da, magma_int_t ldda,
-                              float *d, float *e, float *tauq, float *taup,  
-                              float *x, magma_int_t ldx, float *dx, magma_int_t lddx, 
-                              float *y, magma_int_t ldy, float *dy, magma_int_t lddy);
-magma_int_t magma_slarfb_gpu( char side, char trans, char direct, char storev, 
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              const float *dv, magma_int_t ldv,
-                              const float *dt, magma_int_t ldt, 
-                              float *dc,       magma_int_t ldc,
-                              float *dwork,    magma_int_t ldwork );
-magma_int_t magma_sposv_gpu(  char uplo, magma_int_t n, magma_int_t nrhs, 
-                              float *dA, magma_int_t ldda, 
-                              float *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_spotrf_gpu( char uplo,  magma_int_t n, 
-                              float *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_spotri_gpu( char uplo,  magma_int_t n,
-                              float *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_slauum_gpu( char uplo,  magma_int_t n,
-                              float *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_strtri_gpu( char uplo,  char diag, magma_int_t n,
-                              float *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_ssytrd_gpu( char uplo, magma_int_t n,
-                              float *da, magma_int_t ldda,
-                              float *d, float *e, float *tau,
-                              float *wa,  magma_int_t ldwa,
-                              float *work, magma_int_t lwork,
-                              magma_int_t *info);
-magma_int_t magma_ssytrd2_gpu(char uplo, magma_int_t n,
-                              float *da, magma_int_t ldda,
-                              float *d, float *e, float *tau,
-                              float *wa,  magma_int_t ldwa,
-                              float *work, magma_int_t lwork,
-                              float *dwork, magma_int_t ldwork,
-                              magma_int_t *info);
-magma_int_t magma_spotrs_gpu( char uplo,  magma_int_t n, magma_int_t nrhs, 
-                              float *dA, magma_int_t ldda, 
-                              float *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_sssssm_gpu( char storev, magma_int_t m1, magma_int_t n1, 
-                              magma_int_t m2, magma_int_t n2, magma_int_t k, magma_int_t ib, 
-                              float *dA1, magma_int_t ldda1, 
-                              float *dA2, magma_int_t ldda2, 
-                              float *dL1, magma_int_t lddl1, 
-                              float *dL2, magma_int_t lddl2,
-                              magma_int_t *IPIV, magma_int_t *info);
-magma_int_t magma_ststrf_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t ib, magma_int_t nb,
-                              float *hU, magma_int_t ldhu, float *dU, magma_int_t lddu, 
-                              float *hA, magma_int_t ldha, float *dA, magma_int_t ldda, 
-                              float *hL, magma_int_t ldhl, float *dL, magma_int_t lddl,
-                              magma_int_t *ipiv, 
-                              float *hwork, magma_int_t ldhwork, 
-                              float *dwork, magma_int_t lddwork,
-                              magma_int_t *info);
-magma_int_t magma_sorgqr_gpu( magma_int_t m, magma_int_t n, magma_int_t k, 
-                              float *da, magma_int_t ldda, 
-                              float *tau, float *dwork, 
-                              magma_int_t nb, magma_int_t *info );
-magma_int_t magma_sormql2_gpu(const char side, const char trans,
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              float *da, magma_int_t ldda,
-                              float *tau,
-                              float *dc, magma_int_t lddc,
-                              float *wa, magma_int_t ldwa,
-                              magma_int_t *info);
-magma_int_t magma_sormqr_gpu( char side, char trans, 
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              float *a,    magma_int_t lda, float *tau, 
-                              float *c,    magma_int_t ldc,
-                              float *work, magma_int_t lwork, 
-                              float *td,   magma_int_t nb, magma_int_t *info);
-magma_int_t magma_sormqr2_gpu(const char side, const char trans,
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              float *da,   magma_int_t ldda,
-                              float *tau,
-                              float *dc,    magma_int_t lddc,
-                              float *wa,    magma_int_t ldwa,
-                              magma_int_t *info);
-magma_int_t magma_sormtr_gpu( char side, char uplo, char trans,
-                              magma_int_t m, magma_int_t n,
-                              float *da,    magma_int_t ldda,
-                              float *tau,
-                              float *dc,    magma_int_t lddc,
-                              float *wa,    magma_int_t ldwa,
-                              magma_int_t *info);
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-magma_int_t magma_ssyevd_gpu( char jobz, char uplo,
-                              magma_int_t n,
-                              float *da, magma_int_t ldda,
-                              float *w,
-                              float *wa,  magma_int_t ldwa,
-                              float *work, magma_int_t lwork,
-                              float *rwork, magma_int_t lrwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-magma_int_t magma_ssyevdx_gpu(char jobz, char range, char uplo,
-                              magma_int_t n, float *da, 
-                              magma_int_t ldda, float vl, float vu, 
-                              magma_int_t il, magma_int_t iu,
-                              magma_int_t *m, float *w,
-                              float *wa,  magma_int_t ldwa,
-                              float *work, magma_int_t lwork,
-                              float *rwork, magma_int_t lrwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-magma_int_t magma_ssyevr_gpu( char jobz, char range, char uplo, magma_int_t n,
-                              float *da, magma_int_t ldda, float vl, float vu,
-                              magma_int_t il, magma_int_t iu, float abstol, magma_int_t *m,
-                              float *w, float *dz, magma_int_t lddz,
-                              magma_int_t *isuppz,
-                              float *wa, magma_int_t ldwa,
-                              float *wz, magma_int_t ldwz,
-                              float *work, magma_int_t lwork,
-                              float *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                              magma_int_t liwork, magma_int_t *info);
-#else
-magma_int_t magma_ssyevd_gpu( char jobz, char uplo,
-                              magma_int_t n,
-                              float *da, magma_int_t ldda,
-                              float *w,
-                              float *wa,  magma_int_t ldwa,
-                              float *work, magma_int_t lwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-#endif
-
-magma_int_t magma_ssyevx_gpu( char jobz, char range, char uplo, magma_int_t n,
-                              float *da, magma_int_t ldda, float vl, 
-                              float vu, magma_int_t il, magma_int_t iu, 
-                              float abstol, magma_int_t *m,
-                              float *w, float *dz, magma_int_t lddz,
-                              float *wa, magma_int_t ldwa,
-                              float *wz, magma_int_t ldwz,
-                              float *work, magma_int_t lwork,
-                              float *rwork, magma_int_t *iwork, 
-                              magma_int_t *ifail, magma_int_t *info);
-magma_int_t magma_ssygst_gpu(magma_int_t itype, char uplo, magma_int_t n,
-                             float *da, magma_int_t ldda,
-                             float *db, magma_int_t lddb, magma_int_t *info);
-
-
-/* //////////////////////////////////////////////////////////////////////////// 
- -- MAGMA utility function definitions
-*/
-
-void magma_sprint    ( magma_int_t m, magma_int_t n, const float  *A, magma_int_t lda  );
-void magma_sprint_gpu( magma_int_t m, magma_int_t n, const float *dA, magma_int_t ldda );
-
-void spanel_to_q( char uplo, magma_int_t ib, float *A, magma_int_t lda, float *work );
-void sq_to_panel( char uplo, magma_int_t ib, float *A, magma_int_t lda, float *work );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_s
-#endif /* _MAGMA_S_H_ */
diff --git a/src/include/magma_sbulge.h b/src/include/magma_sbulge.h
deleted file mode 100644
index 9e519c2..0000000
--- a/src/include/magma_sbulge.h
+++ /dev/null
@@ -1,35 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated s Wed Nov 14 22:52:27 2012
- */
-
-#ifndef MAGMA_SBULGEINC_H
-#define MAGMA_SBULGEINC_H
-
-#define PRECISION_s
-
-extern "C" {
-    magma_int_t magma_ssytrd_sy2sb(char uplo, magma_int_t n, magma_int_t NB, float *a, magma_int_t lda,
-                                   float *tau, float *work, magma_int_t lwork, float *dT, magma_int_t threads, magma_int_t *info);
-
-    magma_int_t magma_ssytrd_hb2st(magma_int_t threads, char uplo, magma_int_t n, magma_int_t nb, magma_int_t Vblksiz,
-                                   float *A, magma_int_t lda, float *D, float *E,
-                                   float *V, magma_int_t ldv, float *TAU, magma_int_t compT, float *T, magma_int_t ldt);
-
-    magma_int_t magma_sbulge_back(magma_int_t threads, char uplo, magma_int_t n, magma_int_t nb, magma_int_t ne, magma_int_t Vblksiz,
-                                  float *Z, magma_int_t ldz, float *dZ, magma_int_t lddz,
-                                  float *V, magma_int_t ldv, float *TAU, float *T, magma_int_t ldt, magma_int_t* info);
-
-    magma_int_t magma_sormqr_gpu_2stages(char side, char trans, magma_int_t m, magma_int_t n, magma_int_t k, float *dA, magma_int_t ldda,
-                                         float *dC, magma_int_t lddc, float *dT, magma_int_t nb, magma_int_t *info);
-
-    magma_int_t magma_sbulge_get_lq2(magma_int_t n);
-
-    magma_int_t magma_sbulge_get_Vblksiz(magma_int_t n, magma_int_t nb);
-}
-#endif
diff --git a/src/include/magma_sbulgeinc.h b/src/include/magma_sbulgeinc.h
deleted file mode 100644
index dc73b93..0000000
--- a/src/include/magma_sbulgeinc.h
+++ /dev/null
@@ -1,141 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated s Wed Nov 14 22:52:27 2012
- */
-
-#ifndef MAGMA_SBULGEINC_H
-#define MAGMA_SBULGEINC_H
-
-#define PRECISION_s
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-
-/***************************************************************************//**
- *  Configuration
- **/
-
- // maximum contexts
-#define MAX_THREADS_BLG         256
-
-void findVTpos(magma_int_t N, magma_int_t NB, magma_int_t Vblksiz, magma_int_t sweep, magma_int_t st, magma_int_t *Vpos, magma_int_t *TAUpos, magma_int_t *Tpos, magma_int_t *myblkid);
-void findVTsiz(magma_int_t N, magma_int_t NB, magma_int_t Vblksiz, magma_int_t *blkcnt, magma_int_t *LDV);
-magma_int_t plasma_ceildiv(magma_int_t a, magma_int_t b);
-
-
-/*
-extern volatile magma_int_t barrier_in[MAX_THREADS_BLG];
-extern volatile magma_int_t barrier_out[MAX_THREADS_BLG];
-extern volatile magma_int_t *ss_prog;
-*/
-
- /***************************************************************************//**
- *  Static scheduler
- **/
-/*
-#define ssched_init(nbtiles) \
-{ \
-        volatile int   prog_ol[2*nbtiles+10];\
-                 int   iamdone[MAX_THREADS_BLG]; \
-                 int   thread_num[MAX_THREADS_BLG];\
-        pthread_t      thread_id[MAX_THREADS_BLG];\
-        pthread_attr_t thread_attr;\
-}
-*/
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-
-
- struct gbstrct_blg {
-    float *dQ1;
-    float *dT1;
-    float *dT2;
-    float *dV2;
-    float *dE;
-    float *T;
-    float *A;
-    float *V;
-    float *TAU;
-    float *E;
-    float *E_CPU;
-    int cores_num;
-    int locores_num;
-    int overlapQ1;
-    int usemulticpu;
-    int NB;
-    int NBTILES;
-    int N;
-    int NE;
-    int N_CPU;
-    int N_GPU;
-    int LDA;
-    int LDE;
-    int BAND;
-    int grsiz;
-    int Vblksiz;
-    int WANTZ;
-    char SIDE;
-    real_Double_t *timeblg;
-    real_Double_t *timeaplQ;
-    volatile int *ss_prog;
-} ;
-
-// declare globals here; defined in ssytrd_bsy2trc.cpp
-extern struct gbstrct_blg core_in_all;
-
-
-
-
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-#define MAX_EVENTSBLG 163840
-//#define MAX_EVENTSBLG 1048576
-
-// If we're not using GNU C, elide __attribute__
-#ifndef __GNUC__
-#define __attribute__(x)  /*NOTHING*/
-#endif
-
-// declare globals here; defined in ssytrd_bsy2trc.cpp
-extern int           event_numblg        [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_start_timeblg [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_end_timeblg   [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_logblg        [MAX_THREADS_BLG][MAX_EVENTSBLG]  __attribute__ ((aligned (128)));
-extern int           log_eventsblg;
-
-#ifndef __GNUC__
-#undef  __attribute__
-#endif
-
-#define core_event_startblg(my_core_id)\
-    event_start_timeblg[my_core_id] = magma_wtime();
-
-#define core_event_endblg(my_core_id)\
-    event_end_timeblg[my_core_id] = magma_wtime();
-
-#define core_log_eventblg(event, my_core_id)\
-    event_logblg[my_core_id][event_numblg[my_core_id]+0] = my_core_id;\
-    event_logblg[my_core_id][event_numblg[my_core_id]+1] = event_start_timeblg[my_core_id];\
-    event_logblg[my_core_id][event_numblg[my_core_id]+2] = event_end_timeblg[my_core_id];\
-    event_logblg[my_core_id][event_numblg[my_core_id]+3] = (event);\
-    event_numblg[my_core_id] += (log_eventsblg << 2);\
-    event_numblg[my_core_id] &= (MAX_EVENTSBLG-1);
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_s
-
-#endif /* MAGMA_SBULGEINC_H */
diff --git a/src/include/magma_slapack.h b/src/include/magma_slapack.h
deleted file mode 100644
index 95c1f97..0000000
--- a/src/include/magma_slapack.h
+++ /dev/null
@@ -1,953 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated s Wed Nov 14 22:52:27 2012
- */
-
-#ifndef MAGMA_SLAPACK_H
-#define MAGMA_SLAPACK_H
-
-#define PRECISION_s
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- BLAS and LAPACK functions (alphabetical order)
-*/
-#define blasf77_isamax     FORTRAN_NAME( isamax, ISAMAX )
-#define blasf77_saxpy      FORTRAN_NAME( saxpy,  SAXPY  )
-#define blasf77_scopy      FORTRAN_NAME( scopy,  SCOPY  )
-#define blasf77_sgemm      FORTRAN_NAME( sgemm,  SGEMM  )
-#define blasf77_sgemv      FORTRAN_NAME( sgemv,  SGEMV  )
-#define blasf77_sger      FORTRAN_NAME( sger,  SGER  )
-#define blasf77_sger      FORTRAN_NAME( sger,  SGER  )
-#define blasf77_ssymm      FORTRAN_NAME( ssymm,  SSYMM  )
-#define blasf77_ssymv      FORTRAN_NAME( ssymv,  SSYMV  )
-#define blasf77_ssyr2      FORTRAN_NAME( ssyr2,  SSYR2  )
-#define blasf77_ssyr2k     FORTRAN_NAME( ssyr2k, SSYR2K )
-#define blasf77_ssyrk      FORTRAN_NAME( ssyrk,  SSYRK  )
-#define blasf77_sscal      FORTRAN_NAME( sscal,  SSCAL  )
-#define blasf77_ssscal     FORTRAN_NAME( sscal, SSCAL )
-#define blasf77_sswap      FORTRAN_NAME( sswap,  SSWAP  )
-#define blasf77_ssymm      FORTRAN_NAME( ssymm,  SSYMM  )
-#define blasf77_ssyr2k     FORTRAN_NAME( ssyr2k, SSYR2K )
-#define blasf77_ssyrk      FORTRAN_NAME( ssyrk,  SSYRK  )
-#define blasf77_strmm      FORTRAN_NAME( strmm,  STRMM  )
-#define blasf77_strmv      FORTRAN_NAME( strmv,  STRMV  )
-#define blasf77_strsm      FORTRAN_NAME( strsm,  STRSM  )
-#define blasf77_strsv      FORTRAN_NAME( strsv,  STRSV  )
-
-#define lapackf77_slaed4   FORTRAN_NAME( slaed4, SLAED4 )
-#define lapackf77_slamc3   FORTRAN_NAME( slamc3, SLAMC3 )
-#define lapackf77_slamrg   FORTRAN_NAME( slamrg, SLAMRG )
-#define lapackf77_sstebz   FORTRAN_NAME( sstebz, SSTEBZ )
-
-#define lapackf77_sbdsqr   FORTRAN_NAME( sbdsqr, SBDSQR )
-#define lapackf77_sgebak   FORTRAN_NAME( sgebak, SGEBAK )
-#define lapackf77_sgebal   FORTRAN_NAME( sgebal, SGEBAL )
-#define lapackf77_sgebd2   FORTRAN_NAME( sgebd2, SGEBD2 )
-#define lapackf77_sgebrd   FORTRAN_NAME( sgebrd, SGEBRD )
-#define lapackf77_sgeev    FORTRAN_NAME( sgeev,  SGEEV  )
-#define lapackf77_sgehd2   FORTRAN_NAME( sgehd2, SGEHD2 )
-#define lapackf77_sgehrd   FORTRAN_NAME( sgehrd, SGEHRD )
-#define lapackf77_sgelqf   FORTRAN_NAME( sgelqf, SGELQF )
-#define lapackf77_sgels    FORTRAN_NAME( sgels,  SGELS  )
-#define lapackf77_sgeqlf   FORTRAN_NAME( sgeqlf, SGEQLF )
-#define lapackf77_sgeqp3   FORTRAN_NAME( sgeqp3, SGEQP3 )
-#define lapackf77_sgeqrf   FORTRAN_NAME( sgeqrf, SGEQRF )
-#define lapackf77_sgesvd   FORTRAN_NAME( sgesvd, SGESVD )
-#define lapackf77_sgetrf   FORTRAN_NAME( sgetrf, SGETRF )
-#define lapackf77_sgetri   FORTRAN_NAME( sgetri, SGETRI )
-#define lapackf77_sgetrs   FORTRAN_NAME( sgetrs, SGETRS )
-#define lapackf77_ssbtrd   FORTRAN_NAME( ssbtrd, SSBTRD )
-#define lapackf77_ssyev    FORTRAN_NAME( ssyev,  SSYEV  )
-#define lapackf77_ssyevd   FORTRAN_NAME( ssyevd, SSYEVD )
-#define lapackf77_ssygs2   FORTRAN_NAME( ssygs2, SSYGS2 )
-#define lapackf77_ssygvd   FORTRAN_NAME( ssygvd, SSYGVD )
-#define lapackf77_ssytd2   FORTRAN_NAME( ssytd2, SSYTD2 )
-#define lapackf77_ssytrd   FORTRAN_NAME( ssytrd, SSYTRD )
-#define lapackf77_ssytrf   FORTRAN_NAME( ssytrf, SSYTRF )
-#define lapackf77_shseqr   FORTRAN_NAME( shseqr, SHSEQR )
-#define lapackf77_slabrd   FORTRAN_NAME( slabrd, SLABRD )
-#define lapackf77_sladiv   FORTRAN_NAME( sladiv, ZLADIV )
-#define lapackf77_slacgv   FORTRAN_NAME( slacgv, SLACGV )
-#define lapackf77_slacpy   FORTRAN_NAME( slacpy, SLACPY )
-#define lapackf77_slasyf   FORTRAN_NAME( slasyf, SLASYF )
-#define lapackf77_slange   FORTRAN_NAME( slange, SLANGE )
-#define lapackf77_slansy   FORTRAN_NAME( slansy, SLANSY )
-#define lapackf77_slanst   FORTRAN_NAME( slanst, SLANST )
-#define lapackf77_slansy   FORTRAN_NAME( slansy, SLANSY )
-#define lapackf77_slapy3   FORTRAN_NAME( slapy3, DLAPY3 )
-#define lapackf77_slaqp2   FORTRAN_NAME( slaqp2, SLAQP2 )
-#define lapackf77_slarf    FORTRAN_NAME( slarf,  ZLARF  )
-#define lapackf77_slarfb   FORTRAN_NAME( slarfb, SLARFB )
-#define lapackf77_slarfg   FORTRAN_NAME( slarfg, SLARFG )
-#define lapackf77_slarft   FORTRAN_NAME( slarft, SLARFT )
-#define lapackf77_slarnv   FORTRAN_NAME( slarnv, SLARNV )
-#define lapackf77_slartg   FORTRAN_NAME( slartg, SLARTG )
-#define lapackf77_slascl   FORTRAN_NAME( slascl, SLASCL )
-#define lapackf77_slaset   FORTRAN_NAME( slaset, SLASET )
-#define lapackf77_slaswp   FORTRAN_NAME( slaswp, SLASWP )
-#define lapackf77_slatrd   FORTRAN_NAME( slatrd, SLATRD )
-#define lapackf77_slauum   FORTRAN_NAME( slauum, SLAUUM )
-#define lapackf77_slavsy   FORTRAN_NAME( slavsy, SLAVSY )
-#define lapackf77_spotrf   FORTRAN_NAME( spotrf, SPOTRF )
-#define lapackf77_spotri   FORTRAN_NAME( spotri, SPOTRI )
-#define lapackf77_spotrs   FORTRAN_NAME( spotrs, SPOTRS )
-#define lapackf77_sstedc   FORTRAN_NAME( sstedc, SSTEDC )
-#define lapackf77_sstein   FORTRAN_NAME( sstein, SSTEIN )
-#define lapackf77_sstemr   FORTRAN_NAME( sstemr, SSTEMR )
-#define lapackf77_ssteqr   FORTRAN_NAME( ssteqr, SSTEQR )
-#define lapackf77_ssymv    FORTRAN_NAME( ssymv,  SSYMV  )
-#define lapackf77_strevc   FORTRAN_NAME( strevc, STREVC )
-#define lapackf77_strtri   FORTRAN_NAME( strtri, STRTRI )
-#define lapackf77_sorg2r   FORTRAN_NAME( sorg2r, SORG2R )
-#define lapackf77_sorgbr   FORTRAN_NAME( sorgbr, SORGBR )
-#define lapackf77_sorghr   FORTRAN_NAME( sorghr, SORGHR )
-#define lapackf77_sorglq   FORTRAN_NAME( sorglq, SORGLQ )
-#define lapackf77_sorgql   FORTRAN_NAME( sorgql, SORGQL )
-#define lapackf77_sorgqr   FORTRAN_NAME( sorgqr, SORGQR )
-#define lapackf77_sorgtr   FORTRAN_NAME( sorgtr, SORGTR )
-#define lapackf77_sorm2r   FORTRAN_NAME( sorm2r, SORM2R )
-#define lapackf77_sormbr   FORTRAN_NAME( sormbr, SORMBR )
-#define lapackf77_sormlq   FORTRAN_NAME( sormlq, SORMLQ )
-#define lapackf77_sormql   FORTRAN_NAME( sormql, SORMQL )
-#define lapackf77_sormqr   FORTRAN_NAME( sormqr, SORMQR )
-#define lapackf77_sormtr   FORTRAN_NAME( sormtr, SORMTR )
-
-/* testing functions */
-#define lapackf77_sbdt01   FORTRAN_NAME( sbdt01, SBDT01 )
-#define lapackf77_sget22   FORTRAN_NAME( sget22, SGET22 )
-#define lapackf77_sqpt01   FORTRAN_NAME( sqpt01, SQPT01 )
-#define lapackf77_ssyt21   FORTRAN_NAME( ssyt21, SSYT21 )
-#define lapackf77_shst01   FORTRAN_NAME( shst01, SHST01 )
-#define lapackf77_sqrt02   FORTRAN_NAME( sqrt02, SQRT02 )
-#define lapackf77_sort01   FORTRAN_NAME( sort01, SORT01 )
-#define lapackf77_slarfy   FORTRAN_NAME( slarfy, SLARFY )
-#define lapackf77_slarfx   FORTRAN_NAME( slarfx, SLARFX )
-#define lapackf77_sstt21   FORTRAN_NAME( sstt21, SSTT21 )
-
-// macros to handle differences in arguments between real and real versions of routines.
-#if defined(PRECISION_z) || defined(PRECISION_c)
-#define DWORKFORZ        float *rwork,
-#define DWORKFORZ_AND_LD float *rwork, const magma_int_t *ldrwork,
-#define WSPLIT           float *w
-#else
-#define DWORKFORZ
-#define DWORKFORZ_AND_LD
-#define WSPLIT           float *wr, float *wi
-#endif
-
-/*
- * BLAS functions (alphabetical order)
- */
-magma_int_t blasf77_isamax(
-                     const magma_int_t *n,
-                     const float *x, const magma_int_t *incx);
-
-void blasf77_saxpy(  const magma_int_t *n,
-                     const float *alpha,
-                     const float *x, const magma_int_t *incx,
-                           float *y, const magma_int_t *incy );
-
-void blasf77_scopy(  const magma_int_t *n,
-                     const float *x, const magma_int_t *incx,
-                           float *y, const magma_int_t *incy );
-
-void blasf77_sgemm(  const char *transa, const char *transb,
-                     const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                     const float *B, const magma_int_t *ldb,
-                     const float *beta,
-                           float *C, const magma_int_t *ldc );
-
-void blasf77_sgemv(  const char *transa,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                     const float *x, const magma_int_t *incx,
-                     const float *beta,
-                           float *y, const magma_int_t *incy );
-
-void blasf77_sger(  const magma_int_t *m, const magma_int_t *n,
-                     const float *alpha,
-                     const float *x, const magma_int_t *incx,
-                     const float *y, const magma_int_t *incy,
-                           float *A, const magma_int_t *lda );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void blasf77_sger(  const magma_int_t *m, const magma_int_t *n,
-                     const float *alpha,
-                     const float *x, const magma_int_t *incx,
-                     const float *y, const magma_int_t *incy,
-                           float *A, const magma_int_t *lda );
-#endif
-
-void blasf77_ssymm(  const char *side, const char *uplo,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                     const float *B, const magma_int_t *ldb,
-                     const float *beta,
-                           float *C, const magma_int_t *ldc );
-
-void blasf77_ssymv(  const char *uplo,
-                     const magma_int_t *n,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                     const float *x, const magma_int_t *incx,
-                     const float *beta,
-                           float *y, const magma_int_t *incy );
-
-void blasf77_ssyr2(  const char *uplo,
-                     const magma_int_t *n,
-                     const float *alpha,
-                     const float *x, const magma_int_t *incx,
-                     const float *y, const magma_int_t *incy,
-                           float *A, const magma_int_t *lda );
-
-void blasf77_ssyr2k(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                     const float *B, const magma_int_t *ldb,
-                     const float *beta,
-                           float *C, const magma_int_t *ldc );
-
-void blasf77_ssyrk(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                     const float *beta,
-                           float *C, const magma_int_t *ldc );
-
-void blasf77_sscal(  const magma_int_t *n,
-                     const float *alpha,
-                           float *x, const magma_int_t *incx );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void blasf77_ssscal( const magma_int_t *n,
-                     const float *alpha,
-                           float *x, const magma_int_t *incx );
-#endif
-
-void blasf77_sswap(  const magma_int_t *n,
-                     float *x, const magma_int_t *incx,
-                     float *y, const magma_int_t *incy );
-
-void blasf77_ssymm(  const char *side, const char *uplo,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                     const float *B, const magma_int_t *ldb,
-                     const float *beta,
-                           float *C, const magma_int_t *ldc );
-
-void blasf77_ssyr2k( const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                     const float *B, const magma_int_t *ldb,
-                     const float *beta,
-                           float *C, const magma_int_t *ldc );
-
-void blasf77_ssyrk(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                     const float *beta,
-                           float *C, const magma_int_t *ldc );
-
-void blasf77_strmm(  const char *side, const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                           float *B, const magma_int_t *ldb );
-
-void blasf77_strmv(  const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *n,
-                     const float *A, const magma_int_t *lda,
-                           float *x, const magma_int_t *incx );
-
-void blasf77_strsm(  const char *side, const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const float *alpha,
-                     const float *A, const magma_int_t *lda,
-                           float *B, const magma_int_t *ldb );
-
-void blasf77_strsv(  const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *n,
-                     const float *A, const magma_int_t *lda,
-                           float *x, const magma_int_t *incx );
-
-/*
- * LAPACK functions (alphabetical order)
- */
-void   lapackf77_sbdsqr( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *ncvt, const magma_int_t *nru,  const magma_int_t *ncc,
-                         float *d, float *e,
-                         float *Vt, const magma_int_t *ldvt,
-                         float *U, const magma_int_t *ldu,
-                         float *C, const magma_int_t *ldc,
-                         float *work,
-                         magma_int_t *info );
-
-void   lapackf77_sgebak( const char *job, const char *side,
-                         const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         const float *scale, const magma_int_t *m,
-                         float *V, const magma_int_t *ldv,
-                         magma_int_t *info );
-
-void   lapackf77_sgebal( const char *job,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         magma_int_t *ilo, magma_int_t *ihi,
-                         float *scale,
-                         magma_int_t *info );
-
-void   lapackf77_sgebd2( const magma_int_t *m, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         float *tauq,
-                         float *taup,
-                         float *work,
-                         magma_int_t *info );
-
-void   lapackf77_sgebrd( const magma_int_t *m, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         float *tauq,
-                         float *taup,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sgeev(  const char *jobvl, const char *jobvr,
-                         const magma_int_t *n,
-                         float *A,    const magma_int_t *lda,
-                         WSPLIT,
-                         float *Vl,   const magma_int_t *ldvl,
-                         float *Vr,   const magma_int_t *ldvr,
-                         float *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_sgehd2( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         float *A, const magma_int_t *lda,
-                         float *tau,
-                         float *work,
-                         magma_int_t *info );
-
-void   lapackf77_sgehrd( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         float *A, const magma_int_t *lda,
-                         float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sgelqf( const magma_int_t *m, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sgels(  const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *nrhs,
-                         float *A, const magma_int_t *lda,
-                         float *B, const magma_int_t *ldb,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sgeqlf( const magma_int_t *m, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sgeqp3( const magma_int_t *m, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         magma_int_t *jpvt,
-                         float *tau,
-                         float *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_sgeqrf( const magma_int_t *m, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sgesvd( const char *jobu, const char *jobvt,
-                         const magma_int_t *m, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *s,
-                         float *U,  const magma_int_t *ldu,
-                         float *Vt, const magma_int_t *ldvt,
-                         float *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_sgetrf( const magma_int_t *m, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         magma_int_t *info );
-
-void   lapackf77_sgetri( const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         const magma_int_t *ipiv,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sgetrs( const char* trans,
-                         const magma_int_t *n, const magma_int_t *nrhs,
-                         const float *A, const magma_int_t *lda,
-                         const magma_int_t *ipiv,
-                         float *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_ssbtrd( const char *vect, const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kd,
-                         float *Ab, const magma_int_t *ldab,
-                         float *d, float *e,
-                         float *Q, const magma_int_t *ldq,
-                         float *work,
-                         magma_int_t *info );
-
-void   lapackf77_ssyev(  const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *w,
-                         float *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_ssyevd( const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *w,
-                         float *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_ssygs2( const magma_int_t *itype, const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_ssygvd( const magma_int_t *itype, const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *B, const magma_int_t *ldb,
-                         float *w,
-                         float *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_ssytd2( const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         float *tau,
-                         magma_int_t *info );
-
-void   lapackf77_ssytrd( const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_ssytrf( const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_shseqr( const char *job, const char *compz,
-                         const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         float *H, const magma_int_t *ldh,
-                         WSPLIT,
-                         float *Z, const magma_int_t *ldz,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_slabrd( const magma_int_t *m, const magma_int_t *n, const magma_int_t *nb,
-                         float *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         float *tauq,
-                         float *taup,
-                         float *X, const magma_int_t *ldx,
-                         float *Y, const magma_int_t *ldy );
-
-void   lapackf77_sladiv( float *ret_val, float *x, 
-                         float *y );
-
-void   lapackf77_slacgv( const magma_int_t *n,
-                         float *x, const magma_int_t *incx );
-
-void   lapackf77_slacpy( const char *uplo,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const float *A, const magma_int_t *lda,
-                         float *B, const magma_int_t *ldb );
-
-void   lapackf77_slasyf( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kn,
-                         magma_int_t *kb,
-                         float *A, const magma_int_t lda,
-                         magma_int_t *ipiv,
-                         float *work, const magma_int_t *ldwork,
-                         magma_int_t *info );
-
-float lapackf77_slange( const char *norm,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const float *A, const magma_int_t *lda,
-                         float *work );
-
-float lapackf77_slansy( const char *norm, const char *uplo,
-                         const magma_int_t *n,
-                         const float *A, const magma_int_t *lda,
-                         float * work );
-
-float lapackf77_slanst( const char* norm, const magma_int_t* n,
-                         const float* d, const float* e );
-
-float lapackf77_slansy( const char *norm, const char *uplo,
-                         const magma_int_t *n,
-                         const float *A, const magma_int_t *lda,
-                         float * work );
-
-void lapackf77_slaqp2 (  magma_int_t *m, magma_int_t *n, magma_int_t *offset,
-                         float *a, magma_int_t *lda, magma_int_t *jpvt, 
-                         float *tau,
-                         float *vn1, float *vn2, float *work);
-
-void lapackf77_slarf  (  char *, magma_int_t *, magma_int_t *,
-                         float *, magma_int_t *, float *, float *,
-                         magma_int_t *, float *);
-
-void   lapackf77_slarfb( const char *side, const char *trans, const char *direct, const char *storev,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const float *V, const magma_int_t *ldv,
-                         const float *T, const magma_int_t *ldt,
-                         float *C, const magma_int_t *ldc,
-                         float *work, const magma_int_t *ldwork );
-
-void   lapackf77_slarfg( const magma_int_t *n,
-                         float *alpha,
-                         float *x, const magma_int_t *incx,
-                         float *tau );
-
-void   lapackf77_slarft( const char *direct, const char *storev,
-                         const magma_int_t *n, const magma_int_t *k,
-                         float *V, const magma_int_t *ldv,
-                         const float *tau,
-                         float *T, const magma_int_t *ldt );
-
-void   lapackf77_slarnv( const magma_int_t *idist, magma_int_t *iseed, const magma_int_t *n,
-                         float *x );
-
-void   lapackf77_slartg( float *F,
-                         float *G,
-                         float *cs,
-                         float *SN,
-                         float *R );
-
-void   lapackf77_slascl( const char *type,
-                         const magma_int_t *kl, const magma_int_t *ku,
-                         float *cfrom,
-                         float *cto,
-                         const magma_int_t *m, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_slaset( const char *uplo,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const float *alpha,
-                         const float *beta,
-                         float *A, const magma_int_t *lda );
-
-void   lapackf77_slaswp( const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         const magma_int_t *k1, const magma_int_t *k2,
-                         magma_int_t *ipiv,
-                         const magma_int_t *incx );
-
-void   lapackf77_slatrd( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *nb,
-                         float *A, const magma_int_t *lda,
-                         float *e,
-                         float *tau,
-                         float *work, const magma_int_t *ldwork );
-
-void   lapackf77_slauum( const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_slavsy( const char *uplo, const char *trans, const char *diag,
-                         magma_int_t *n, magma_int_t *nrhs,
-                         float *A, magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         float *B, magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_spotrf( const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_spotri( const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_spotrs( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *nrhs,
-                         const float *A, const magma_int_t *lda,
-                         float *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_sstedc( const char *compz,
-                         const magma_int_t *n,
-                         float *d, float *e,
-                         float *Z, const magma_int_t *ldz,
-                         float *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_sstein( const magma_int_t *n,
-                         const float *d, const float *e,
-                         const magma_int_t *m,
-                         const float *w,
-                         const magma_int_t *iblock,
-                         const magma_int_t *isplit,
-                         float *Z, const magma_int_t *ldz,
-                         float *work, magma_int_t *iwork, magma_int_t *ifailv,
-                         magma_int_t *info );
-
-void   lapackf77_sstemr( const char *jobz, const char *range,
-                         const magma_int_t *n,
-                         float *d, float *e,
-                         const float *vl, const float *vu,
-                         const magma_int_t *il, const magma_int_t *iu,
-                         magma_int_t *m,
-                         float *w,
-                         float *Z, const magma_int_t *ldz,
-                         const magma_int_t *nzc, magma_int_t *isuppz, magma_int_t *tryrac,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_ssteqr( const char *compz,
-                         const magma_int_t *n,
-                         float *d, float *e,
-                         float *Z, const magma_int_t *ldz,
-                         float *work,
-                         magma_int_t *info );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void   lapackf77_ssymv(  const char *uplo,
-                         const magma_int_t *n,
-                         const float *alpha,
-                         const float *A, const magma_int_t *lda,
-                         const float *x, const magma_int_t *incx,
-                         const float *beta,
-                               float *y, const magma_int_t *incy );
-#endif
-
-void   lapackf77_strevc( const char *side, const char *howmny,
-                         magma_int_t *select, const magma_int_t *n,
-                         float *T,  const magma_int_t *ldt,
-                         float *Vl, const magma_int_t *ldvl,
-                         float *Vr, const magma_int_t *ldvr,
-                         const magma_int_t *mm, magma_int_t *m,
-                         float *work,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_strtri( const char *uplo, const char *diag,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_sorg2r( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *work,
-                         magma_int_t *info );
-
-void   lapackf77_sorgbr( const char *vect,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sorghr( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sorglq( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sorgql( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sorgqr( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sorgtr( const char *uplo,
-                         const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sorm2r( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *C, const magma_int_t *ldc,
-                         float *work,
-                         magma_int_t *info );
-
-void   lapackf77_sormbr( const char *vect, const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *C, const magma_int_t *ldc,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sormlq( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *C, const magma_int_t *ldc,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sormql( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *C, const magma_int_t *ldc,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sormqr( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *C, const magma_int_t *ldc,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_sormtr( const char *side, const char *uplo, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const float *A, const magma_int_t *lda,
-                         const float *tau,
-                         float *C, const magma_int_t *ldc,
-                         float *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-/*
- * Real precision extras
- */
-void   lapackf77_sstebz( const char *range, const char *order,
-                         const magma_int_t *n,
-                         float *vl, float *vu,
-                         magma_int_t *il, magma_int_t *iu,
-                         float *abstol,
-                         float *d, float *e,
-                         const magma_int_t *m, const magma_int_t *nsplit,
-                         float *w,
-                         magma_int_t *iblock, magma_int_t *isplit,
-                         float *work,
-                         magma_int_t *iwork,
-                         magma_int_t *info );
-
-float lapackf77_slamc3( float* a, float* b );
-
-void   lapackf77_slamrg( magma_int_t* n1, magma_int_t* n2,
-                         float* a,
-                         magma_int_t* dtrd1, magma_int_t* dtrd2, magma_int_t* index );
-
-float lapackf77_slapy3(float *, float *, float *);
-
-void   lapackf77_slaed4( magma_int_t* n, magma_int_t* i,
-                         float* d,
-                         float* z,
-                         float* delta,
-                         float* rho,
-                         float* dlam, magma_int_t* info );
-
-/*
- * Testing functions
- */
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void   lapackf77_sbdt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *kd,
-                         float *A, const magma_int_t *lda,
-                         float *Q, const magma_int_t *ldq,
-                         float *d, float *e,
-                         float *Pt, const magma_int_t *ldpt,
-                         float *work,
-                         float *rwork,
-                         float *resid );
-
-void   lapackf77_sget22( const char *transa, const char *transe, const char *transw, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *E, const magma_int_t *lde,
-                         float *w,
-                         float *work,
-                         float *rwork,
-                         float *result );
-
-void   lapackf77_ssyt21( const magma_int_t *itype, const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kband,
-                         float *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         float *U, const magma_int_t *ldu,
-                         float *V, const magma_int_t *ldv,
-                         float *tau,
-                         float *work,
-                         float *rwork,
-                         float *result );
-
-void   lapackf77_shst01( const magma_int_t *n, const magma_int_t *ilo, const magma_int_t *ihi,
-                         float *A, const magma_int_t *lda,
-                         float *H, const magma_int_t *ldh,
-                         float *Q, const magma_int_t *ldq,
-                         float *work, const magma_int_t *lwork,
-                         float *rwork,
-                         float *result );
-
-void   lapackf77_sstt21( const magma_int_t *n, const magma_int_t *kband,
-                         float *AD,
-                         float *AE,
-                         float *SD,
-                         float *SE,
-                         float *U, const magma_int_t *ldu,
-                         float *work,
-                         float *rwork,
-                         float *result );
-
-void   lapackf77_sort01( const char *rowcol, const magma_int_t *m, const magma_int_t *n,
-                         float *U, const magma_int_t *ldu,
-                         float *work, const magma_int_t *lwork,
-                         float *rwork,
-                         float *resid );
-#else
-void   lapackf77_sbdt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *kd,
-                         float *A, const magma_int_t *lda,
-                         float *Q, const magma_int_t *ldq,
-                         float *d, float *e,
-                         float *Pt, const magma_int_t *ldpt,
-                         float *work,
-                         float *resid );
-
-void   lapackf77_sget22( const char *transa, const char *transe, const char *transw, const magma_int_t *n,
-                         float *A, const magma_int_t *lda,
-                         float *E, const magma_int_t *lde,
-                         float *wr,
-                         float *wi,
-                         float *work,
-                         float *result );
-
-void   lapackf77_ssyt21( magma_int_t *itype, const char *uplo, const magma_int_t *n, const magma_int_t *kband,
-                         float *A, const magma_int_t *lda,
-                         float *d, float *e,
-                         float *U, const magma_int_t *ldu,
-                         float *V, const magma_int_t *ldv,
-                         float *tau,
-                         float *work,
-                         float *result );
-
-void   lapackf77_shst01( const magma_int_t *n, const magma_int_t *ilo, const magma_int_t *ihi,
-                         float *A, const magma_int_t *lda,
-                         float *H, const magma_int_t *ldh,
-                         float *Q, const magma_int_t *ldq,
-                         float *work, const magma_int_t *lwork,
-                         float *result );
-
-void   lapackf77_sstt21( const magma_int_t *n, const magma_int_t *kband,
-                         float *AD,
-                         float *AE,
-                         float *SD,
-                         float *SE,
-                         float *U, const magma_int_t *ldu,
-                         float *work,
-                         float *result );
-
-void   lapackf77_sort01( const char *rowcol, const magma_int_t *m, const magma_int_t *n,
-                         float *U, const magma_int_t *ldu,
-                         float *work, const magma_int_t *lwork,
-                         float *resid );
-#endif
-
-void   lapackf77_slarfy( const char *uplo, const magma_int_t *n,
-                         float *V, const magma_int_t *incv,
-                         float *tau,
-                         float *C, const magma_int_t *ldc,
-                         float *work );
-
-void   lapackf77_slarfx( const char *side, const magma_int_t *m, const magma_int_t *n,
-                         float *V,
-                         float *tau,
-                         float *C, const magma_int_t *ldc,
-                         float *work );
-
-float lapackf77_sqpt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         float *A,
-                         float *Af, const magma_int_t *lda,
-                         float *tau, magma_int_t *jpvt,
-                         float *work, const magma_int_t *lwork );
-
-void   lapackf77_sqrt02( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         float *A,
-                         float *AF,
-                         float *Q,
-                         float *R, const magma_int_t *lda,
-                         float *tau,
-                         float *work, const magma_int_t *lwork,
-                         float *rwork,
-                         float *result );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef DWORKFORZ
-#undef DWORKFORZ_AND_LD
-#undef WSPLIT
-#undef PRECISION_s
-
-#endif /* MAGMA_SLAPACK_H */
diff --git a/src/include/magma_z.h b/src/include/magma_z.h
deleted file mode 100644
index 7d9465e..0000000
--- a/src/include/magma_z.h
+++ /dev/null
@@ -1,514 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @precisions normal z -> s d c
- */
-
-#ifndef _MAGMA_Z_H_
-#define _MAGMA_Z_H_
-#define PRECISION_z
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- MAGMA Auxiliary functions to get the NB used
-*/
-int magma_get_zpotrf_nb(int m);
-int magma_get_zgetrf_nb(int m);
-int magma_get_zgetri_nb(int m);
-int magma_get_zgeqp3_nb(int m);
-int magma_get_zgeqrf_nb(int m);
-int magma_get_zgeqlf_nb(int m);
-int magma_get_zgehrd_nb(int m);
-int magma_get_zhetrd_nb(int m);
-int magma_get_zgelqf_nb(int m);
-int magma_get_zgebrd_nb(int m);
-int magma_get_zhegst_nb(int m);
-int magma_get_zgesvd_nb(int m);
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- MAGMA function definitions / Data on CPU
-*/
-magma_int_t magma_zgebrd( magma_int_t m, magma_int_t n, cuDoubleComplex *A, 
-                          magma_int_t lda, double *d, double *e,
-                          cuDoubleComplex *tauq,  cuDoubleComplex *taup, 
-                          cuDoubleComplex *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_zgehrd2(magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          cuDoubleComplex *A, magma_int_t lda, cuDoubleComplex *tau, 
-                          cuDoubleComplex *work, magma_int_t *lwork, magma_int_t *info);
-magma_int_t magma_zgehrd( magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          cuDoubleComplex *A, magma_int_t lda, cuDoubleComplex *tau,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          cuDoubleComplex *d_T, magma_int_t *info);
-magma_int_t magma_zgelqf( magma_int_t m, magma_int_t n, 
-                          cuDoubleComplex *A,    magma_int_t lda,   cuDoubleComplex *tau, 
-                          cuDoubleComplex *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_zgeqlf( magma_int_t m, magma_int_t n, 
-                          cuDoubleComplex *A,    magma_int_t lda,   cuDoubleComplex *tau, 
-                          cuDoubleComplex *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_zgeqrf( magma_int_t m, magma_int_t n, cuDoubleComplex *A, 
-                          magma_int_t lda, cuDoubleComplex *tau, cuDoubleComplex *work, 
-                          magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_zgeqr2( magma_int_t *m, magma_int_t *n, cuDoubleComplex *a,
-                          magma_int_t *lda, cuDoubleComplex *tau, cuDoubleComplex *work, 
-                          magma_int_t *info);
-magma_int_t magma_zgeqrf4(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
-                          cuDoubleComplex *a,    magma_int_t lda, cuDoubleComplex *tau,
-                          cuDoubleComplex *work, magma_int_t lwork, magma_int_t *info );
-magma_int_t magma_zgeqrf_ooc( magma_int_t m, magma_int_t n, cuDoubleComplex *A,
-                          magma_int_t lda, cuDoubleComplex *tau, cuDoubleComplex *work,
-                          magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_zgesv ( magma_int_t n, magma_int_t nrhs, 
-                          cuDoubleComplex *A, magma_int_t lda, magma_int_t *ipiv, 
-                          cuDoubleComplex *B, magma_int_t ldb, magma_int_t *info);
-magma_int_t magma_zgetrf( magma_int_t m, magma_int_t n, cuDoubleComplex *A, 
-                          magma_int_t lda, magma_int_t *ipiv, 
-                          magma_int_t *info);
-magma_int_t magma_zgetrf2(magma_int_t m, magma_int_t n, cuDoubleComplex *a, 
-                          magma_int_t lda, magma_int_t *ipiv, magma_int_t *info);
-
-magma_int_t magma_zlaqps( magma_int_t m, magma_int_t n, magma_int_t offset, 
-                          magma_int_t nb, magma_int_t *kb, 
-                          cuDoubleComplex *A,  magma_int_t lda,
-                          cuDoubleComplex *dA, magma_int_t ldda,
-                          magma_int_t *jpvt, cuDoubleComplex *tau, double *vn1, double *vn2, 
-                          cuDoubleComplex *auxv, 
-                          cuDoubleComplex *F,  magma_int_t ldf,
-                          cuDoubleComplex *dF, magma_int_t lddf );
-magma_int_t magma_zlarf(  char *side, magma_int_t *m, magma_int_t *n,
-                          cuDoubleComplex *v, magma_int_t *incv, cuDoubleComplex *tau,
-                          cuDoubleComplex *c__, magma_int_t *ldc, cuDoubleComplex *work);
-magma_int_t magma_zlarfg( magma_int_t *n, cuDoubleComplex *alpha, cuDoubleComplex *x,
-                          magma_int_t *incx, cuDoubleComplex *tau);
-magma_int_t magma_zlatrd( char uplo, magma_int_t n, magma_int_t nb, cuDoubleComplex *a, 
-                          magma_int_t lda, double *e, cuDoubleComplex *tau, 
-                          cuDoubleComplex *w, magma_int_t ldw,
-                          cuDoubleComplex *da, magma_int_t ldda, 
-                          cuDoubleComplex *dw, magma_int_t lddw);
-magma_int_t magma_zlatrd2(char uplo, magma_int_t n, magma_int_t nb,
-                          cuDoubleComplex *a,  magma_int_t lda,
-                          double *e, cuDoubleComplex *tau,
-                          cuDoubleComplex *w,  magma_int_t ldw,
-                          cuDoubleComplex *da, magma_int_t ldda,
-                          cuDoubleComplex *dw, magma_int_t lddw,
-                          cuDoubleComplex *dwork, magma_int_t ldwork);
-magma_int_t magma_zlahr2( magma_int_t m, magma_int_t n, magma_int_t nb, 
-                          cuDoubleComplex *da, cuDoubleComplex *dv, cuDoubleComplex *a, 
-                          magma_int_t lda, cuDoubleComplex *tau, cuDoubleComplex *t, 
-                          magma_int_t ldt, cuDoubleComplex *y, magma_int_t ldy);
-magma_int_t magma_zlahru( magma_int_t n, magma_int_t ihi, magma_int_t k, magma_int_t nb, 
-                          cuDoubleComplex *a, magma_int_t lda, 
-                          cuDoubleComplex *da, cuDoubleComplex *y, 
-                          cuDoubleComplex *v, cuDoubleComplex *t, 
-                          cuDoubleComplex *dwork);
-magma_int_t magma_zposv ( char uplo, magma_int_t n, magma_int_t nrhs, 
-                          cuDoubleComplex *A, magma_int_t lda, 
-                          cuDoubleComplex *B, magma_int_t ldb, magma_int_t *info);
-magma_int_t magma_zpotrf( char uplo, magma_int_t n, cuDoubleComplex *A, 
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_zpotri( char uplo, magma_int_t n, cuDoubleComplex *A,
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_zlauum( char uplo, magma_int_t n, cuDoubleComplex *A,
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_ztrtri( char uplo, char diag, magma_int_t n, cuDoubleComplex *A, 
-                          magma_int_t lda, magma_int_t *info);
-magma_int_t magma_zhetrd( char uplo, magma_int_t n, cuDoubleComplex *A, 
-                          magma_int_t lda, double *d, double *e, 
-                          cuDoubleComplex *tau, cuDoubleComplex *work, magma_int_t lwork, 
-                          magma_int_t *info);
-magma_int_t magma_zungqr( magma_int_t m, magma_int_t n, magma_int_t k,
-                          cuDoubleComplex *a, magma_int_t lda,
-                          cuDoubleComplex *tau, cuDoubleComplex *dwork,
-                          magma_int_t nb, magma_int_t *info );
-magma_int_t magma_zunmql( const char side, const char trans,
-                          magma_int_t m, magma_int_t n, magma_int_t k,
-                          cuDoubleComplex *a, magma_int_t lda,
-                          cuDoubleComplex *tau,
-                          cuDoubleComplex *c, magma_int_t ldc,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_zunmqr( char side, char trans, 
-                          magma_int_t m, magma_int_t n, magma_int_t k, 
-                          cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *tau, 
-                          cuDoubleComplex *c, magma_int_t ldc, 
-                          cuDoubleComplex *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_zunmtr( char side, char uplo, char trans,
-                          magma_int_t m, magma_int_t n,
-                          cuDoubleComplex *a,    magma_int_t lda,
-                          cuDoubleComplex *tau,
-                          cuDoubleComplex *c,    magma_int_t ldc,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_zunghr( magma_int_t n, magma_int_t ilo, magma_int_t ihi,
-                          cuDoubleComplex *a, magma_int_t lda,
-                          cuDoubleComplex *tau,
-                          cuDoubleComplex *dT, magma_int_t nb,
-                          magma_int_t *info);
-magma_int_t magma_zheev( char jobz, char uplo, magma_int_t n,
-                         cuDoubleComplex *a, magma_int_t lda, double *w,
-                         cuDoubleComplex *work, magma_int_t lwork,
-                         double *rwork, magma_int_t *info);
-magma_int_t magma_zheevx(char jobz, char range, char uplo, magma_int_t n,
-                         cuDoubleComplex *a, magma_int_t lda, double vl, double vu,
-                         magma_int_t il, magma_int_t iu, double abstol, magma_int_t *m,
-                         double *w, cuDoubleComplex *z, magma_int_t ldz, 
-                         cuDoubleComplex *work, magma_int_t lwork,
-                         double *rwork, magma_int_t *iwork, 
-                         magma_int_t *ifail, magma_int_t *info);
-#if defined(PRECISION_z) || defined(PRECISION_c)
-magma_int_t  magma_zgeev( char jobvl, char jobvr, magma_int_t n,
-                          cuDoubleComplex *a, magma_int_t lda,
-                          cuDoubleComplex *w,
-                          cuDoubleComplex *vl, magma_int_t ldvl,
-                          cuDoubleComplex *vr, magma_int_t ldvr,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          double *rwork, magma_int_t *info);
-magma_int_t magma_zgeqp3( magma_int_t m, magma_int_t n,
-                          cuDoubleComplex *a, magma_int_t lda, 
-                          magma_int_t *jpvt, cuDoubleComplex *tau,
-                          cuDoubleComplex *work, magma_int_t lwork, 
-                          double *rwork, magma_int_t *info);
-magma_int_t magma_zgesvd( char jobu, char jobvt, magma_int_t m, magma_int_t n,
-                          cuDoubleComplex *a,    magma_int_t lda, double *s, 
-                          cuDoubleComplex *u,    magma_int_t ldu, 
-                          cuDoubleComplex *vt,   magma_int_t ldvt,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          double *rwork, magma_int_t *info );
-magma_int_t magma_zheevd( char jobz, char uplo, magma_int_t n,
-                          cuDoubleComplex *a, magma_int_t lda, double *w,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          double *rwork, magma_int_t lrwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_zheevr( char jobz, char range, char uplo, magma_int_t n,
-                          cuDoubleComplex *a, magma_int_t lda, double vl, double vu,
-                          magma_int_t il, magma_int_t iu, double abstol, magma_int_t *m,
-                          double *w, cuDoubleComplex *z, magma_int_t ldz, 
-                          magma_int_t *isuppz,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          double *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_zhegvd( magma_int_t itype, char jobz, char uplo, magma_int_t n,
-                          cuDoubleComplex *a, magma_int_t lda,
-                          cuDoubleComplex *b, magma_int_t ldb,
-                          double *w, cuDoubleComplex *work, magma_int_t lwork,
-                          double *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_zhegvdx(magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, cuDoubleComplex *a, magma_int_t lda,
-                          cuDoubleComplex *b, magma_int_t ldb,
-                          double vl, double vu, magma_int_t il, magma_int_t iu,
-                          magma_int_t *m, double *w, cuDoubleComplex *work, 
-                          magma_int_t lwork, double *rwork,
-                          magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_zhegvx( magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, cuDoubleComplex *a, magma_int_t lda, 
-                          cuDoubleComplex *b, magma_int_t ldb,
-                          double vl, double vu, magma_int_t il, magma_int_t iu,
-                          double abstol, magma_int_t *m, double *w, 
-                          cuDoubleComplex *z, magma_int_t ldz,
-                          cuDoubleComplex *work, magma_int_t lwork, double *rwork,
-                          magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info);
-magma_int_t magma_zhegvr( magma_int_t itype, char jobz, char range, char uplo, 
-                          magma_int_t n, cuDoubleComplex *a, magma_int_t lda,
-                          cuDoubleComplex *b, magma_int_t ldb,
-                          double vl, double vu, magma_int_t il, magma_int_t iu,
-                          double abstol, magma_int_t *m, double *w, 
-                          cuDoubleComplex *z, magma_int_t ldz,
-                          magma_int_t *isuppz, cuDoubleComplex *work, magma_int_t lwork,
-                          double *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                          magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_zstedx( char range, magma_int_t n, double vl, double vu,
-                          magma_int_t il, magma_int_t iu, double *D, double *E,
-                          cuDoubleComplex *Z, magma_int_t ldz,
-                          double *rwork, magma_int_t ldrwork, magma_int_t *iwork,
-                          magma_int_t liwork, double* dwork, magma_int_t *info);
-#else
-magma_int_t  magma_zgeev( char jobvl, char jobvr, magma_int_t n,
-                          cuDoubleComplex *a,    magma_int_t lda,
-                          cuDoubleComplex *wr, cuDoubleComplex *wi,
-                          cuDoubleComplex *vl,   magma_int_t ldvl,
-                          cuDoubleComplex *vr,   magma_int_t ldvr,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_zgeqp3( magma_int_t m, magma_int_t n,
-                          cuDoubleComplex *a, magma_int_t lda, 
-                          magma_int_t *jpvt, cuDoubleComplex *tau,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          magma_int_t *info);
-magma_int_t magma_zgesvd( char jobu, char jobvt, magma_int_t m, magma_int_t n,
-                          cuDoubleComplex *a,    magma_int_t lda, double *s, 
-                          cuDoubleComplex *u,    magma_int_t ldu, 
-                          cuDoubleComplex *vt,   magma_int_t ldvt,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          magma_int_t *info );
-magma_int_t magma_zheevd( char jobz, char uplo, magma_int_t n,
-                          cuDoubleComplex *a, magma_int_t lda, double *w,
-                          cuDoubleComplex *work, magma_int_t lwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_zhegvd( magma_int_t itype, char jobz, char uplo, magma_int_t n,
-                          cuDoubleComplex *a, magma_int_t lda,
-                          cuDoubleComplex *b, magma_int_t ldb,
-                          double *w, cuDoubleComplex *work, magma_int_t lwork,
-                          magma_int_t *iwork, magma_int_t liwork, magma_int_t *info);
-magma_int_t magma_zstedx( char range, magma_int_t n, double vl, double vu,
-                          magma_int_t il, magma_int_t iu, double* d, double* e,
-                          double* z, magma_int_t ldz,
-                          double* work, magma_int_t lwork,
-                          magma_int_t* iwork, magma_int_t liwork,
-                          double* dwork, magma_int_t* info);
-magma_int_t magma_zlaex0( magma_int_t n, double* d, double* e, double* q, magma_int_t ldq,
-                          double* work, magma_int_t* iwork, double* dwork,
-                          char range, double vl, double vu,
-                          magma_int_t il, magma_int_t iu, magma_int_t* info);
-magma_int_t magma_zlaex1( magma_int_t n, double* d, double* q, magma_int_t ldq,
-                          magma_int_t* indxq, double rho, magma_int_t cutpnt,
-                          double* work, magma_int_t* iwork, double* dwork,
-                          char range, double vl, double vu,
-                          magma_int_t il, magma_int_t iu, magma_int_t* info);
-magma_int_t magma_zlaex3( magma_int_t k, magma_int_t n, magma_int_t n1, double* d,
-                          double* q, magma_int_t ldq, double rho,
-                          double* dlamda, double* q2, magma_int_t* indx,
-                          magma_int_t* ctot, double* w, double* s, magma_int_t* indxq,
-                          double* dwork,
-                          char range, double vl, double vu, magma_int_t il, magma_int_t iu,
-                          magma_int_t* info );
-#endif
-
-magma_int_t magma_zhegst( magma_int_t itype, char uplo, magma_int_t n,
-                          cuDoubleComplex *a, magma_int_t lda,
-                          cuDoubleComplex *b, magma_int_t ldb, magma_int_t *info);
-
-/* //////////////////////////////////////////////////////////////////////////// 
- -- MAGMA function definitions / Data on GPU
-*/
-magma_int_t magma_zgels_gpu(  char trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
-                              cuDoubleComplex *dA,    magma_int_t ldda, 
-                              cuDoubleComplex *dB,    magma_int_t lddb, 
-                              cuDoubleComplex *hwork, magma_int_t lwork, 
-                              magma_int_t *info);
-magma_int_t magma_zgels3_gpu( char trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
-                              cuDoubleComplex *dA,    magma_int_t ldda,
-                              cuDoubleComplex *dB,    magma_int_t lddb,
-                              cuDoubleComplex *hwork, magma_int_t lwork,
-                              magma_int_t *info);
-magma_int_t magma_zgelqf_gpu( magma_int_t m, magma_int_t n,
-                              cuDoubleComplex *dA,    magma_int_t ldda,   cuDoubleComplex *tau,
-                              cuDoubleComplex *work, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_zgeqrf_gpu( magma_int_t m, magma_int_t n, 
-                              cuDoubleComplex *dA,  magma_int_t ldda, 
-                              cuDoubleComplex *tau, cuDoubleComplex *dT, 
-                              magma_int_t *info);
-magma_int_t magma_zgeqrf2_gpu(magma_int_t m, magma_int_t n, 
-                              cuDoubleComplex *dA,  magma_int_t ldda, 
-                              cuDoubleComplex *tau, magma_int_t *info);
-magma_int_t magma_zgeqrf2_mgpu(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
-                               cuDoubleComplex **dlA, magma_int_t ldda,
-                               cuDoubleComplex *tau, magma_int_t *info );
-magma_int_t magma_zgeqrf3_gpu(magma_int_t m, magma_int_t n, 
-                              cuDoubleComplex *dA,  magma_int_t ldda, 
-                              cuDoubleComplex *tau, cuDoubleComplex *dT, 
-                              magma_int_t *info);
-magma_int_t magma_zgeqrs_gpu( magma_int_t m, magma_int_t n, magma_int_t nrhs, 
-                              cuDoubleComplex *dA,     magma_int_t ldda, 
-                              cuDoubleComplex *tau,   cuDoubleComplex *dT,
-                              cuDoubleComplex *dB,    magma_int_t lddb,
-                              cuDoubleComplex *hwork, magma_int_t lhwork, 
-                              magma_int_t *info);
-magma_int_t magma_zgeqrs3_gpu( magma_int_t m, magma_int_t n, magma_int_t nrhs, 
-                              cuDoubleComplex *dA,     magma_int_t ldda, 
-                              cuDoubleComplex *tau,   cuDoubleComplex *dT,
-                              cuDoubleComplex *dB,    magma_int_t lddb,
-                              cuDoubleComplex *hwork, magma_int_t lhwork, 
-                              magma_int_t *info);
-magma_int_t magma_zgessm_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t k, magma_int_t ib, 
-                              magma_int_t *ipiv, 
-                              cuDoubleComplex *dL1, magma_int_t lddl1, 
-                              cuDoubleComplex *dL,  magma_int_t lddl, 
-                              cuDoubleComplex *dA,  magma_int_t ldda, 
-                              magma_int_t *info);
-magma_int_t magma_zgesv_gpu(  magma_int_t n, magma_int_t nrhs, 
-                              cuDoubleComplex *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              cuDoubleComplex *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_zgetrf_incpiv_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t ib,
-                              cuDoubleComplex *hA, magma_int_t ldha, cuDoubleComplex *dA, magma_int_t ldda,
-                              cuDoubleComplex *hL, magma_int_t ldhl, cuDoubleComplex *dL, magma_int_t lddl,
-                              magma_int_t *ipiv, 
-                              cuDoubleComplex *dwork, magma_int_t lddwork,
-                              magma_int_t *info);
-magma_int_t magma_zgetrf_gpu( magma_int_t m, magma_int_t n, 
-                              cuDoubleComplex *dA, magma_int_t ldda, 
-                              magma_int_t *ipiv, magma_int_t *info);
-magma_int_t 
-magma_zgetrf_nopiv_gpu      ( magma_int_t m, magma_int_t n,
-                              cuDoubleComplex *dA, magma_int_t ldda,
-                              magma_int_t *info);
-magma_int_t magma_zgetri_gpu( magma_int_t n, 
-                              cuDoubleComplex *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              cuDoubleComplex *dwork, magma_int_t lwork, magma_int_t *info);
-magma_int_t magma_zgetrs_gpu( char trans, magma_int_t n, magma_int_t nrhs, 
-                              cuDoubleComplex *dA, magma_int_t ldda, magma_int_t *ipiv, 
-                              cuDoubleComplex *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_zlabrd_gpu( magma_int_t m, magma_int_t n, magma_int_t nb, 
-                              cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *da, magma_int_t ldda,
-                              double *d, double *e, cuDoubleComplex *tauq, cuDoubleComplex *taup,  
-                              cuDoubleComplex *x, magma_int_t ldx, cuDoubleComplex *dx, magma_int_t lddx, 
-                              cuDoubleComplex *y, magma_int_t ldy, cuDoubleComplex *dy, magma_int_t lddy);
-magma_int_t magma_zlarfb_gpu( char side, char trans, char direct, char storev, 
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              const cuDoubleComplex *dv, magma_int_t ldv,
-                              const cuDoubleComplex *dt, magma_int_t ldt, 
-                              cuDoubleComplex *dc,       magma_int_t ldc,
-                              cuDoubleComplex *dwork,    magma_int_t ldwork );
-magma_int_t magma_zposv_gpu(  char uplo, magma_int_t n, magma_int_t nrhs, 
-                              cuDoubleComplex *dA, magma_int_t ldda, 
-                              cuDoubleComplex *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_zpotrf_gpu( char uplo,  magma_int_t n, 
-                              cuDoubleComplex *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_zpotri_gpu( char uplo,  magma_int_t n,
-                              cuDoubleComplex *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_zlauum_gpu( char uplo,  magma_int_t n,
-                              cuDoubleComplex *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_ztrtri_gpu( char uplo,  char diag, magma_int_t n,
-                              cuDoubleComplex *dA, magma_int_t ldda, magma_int_t *info);
-magma_int_t magma_zhetrd_gpu( char uplo, magma_int_t n,
-                              cuDoubleComplex *da, magma_int_t ldda,
-                              double *d, double *e, cuDoubleComplex *tau,
-                              cuDoubleComplex *wa,  magma_int_t ldwa,
-                              cuDoubleComplex *work, magma_int_t lwork,
-                              magma_int_t *info);
-magma_int_t magma_zhetrd2_gpu(char uplo, magma_int_t n,
-                              cuDoubleComplex *da, magma_int_t ldda,
-                              double *d, double *e, cuDoubleComplex *tau,
-                              cuDoubleComplex *wa,  magma_int_t ldwa,
-                              cuDoubleComplex *work, magma_int_t lwork,
-                              cuDoubleComplex *dwork, magma_int_t ldwork,
-                              magma_int_t *info);
-magma_int_t magma_zpotrs_gpu( char uplo,  magma_int_t n, magma_int_t nrhs, 
-                              cuDoubleComplex *dA, magma_int_t ldda, 
-                              cuDoubleComplex *dB, magma_int_t lddb, magma_int_t *info);
-magma_int_t magma_zssssm_gpu( char storev, magma_int_t m1, magma_int_t n1, 
-                              magma_int_t m2, magma_int_t n2, magma_int_t k, magma_int_t ib, 
-                              cuDoubleComplex *dA1, magma_int_t ldda1, 
-                              cuDoubleComplex *dA2, magma_int_t ldda2, 
-                              cuDoubleComplex *dL1, magma_int_t lddl1, 
-                              cuDoubleComplex *dL2, magma_int_t lddl2,
-                              magma_int_t *IPIV, magma_int_t *info);
-magma_int_t magma_ztstrf_gpu( char storev, magma_int_t m, magma_int_t n, magma_int_t ib, magma_int_t nb,
-                              cuDoubleComplex *hU, magma_int_t ldhu, cuDoubleComplex *dU, magma_int_t lddu, 
-                              cuDoubleComplex *hA, magma_int_t ldha, cuDoubleComplex *dA, magma_int_t ldda, 
-                              cuDoubleComplex *hL, magma_int_t ldhl, cuDoubleComplex *dL, magma_int_t lddl,
-                              magma_int_t *ipiv, 
-                              cuDoubleComplex *hwork, magma_int_t ldhwork, 
-                              cuDoubleComplex *dwork, magma_int_t lddwork,
-                              magma_int_t *info);
-magma_int_t magma_zungqr_gpu( magma_int_t m, magma_int_t n, magma_int_t k, 
-                              cuDoubleComplex *da, magma_int_t ldda, 
-                              cuDoubleComplex *tau, cuDoubleComplex *dwork, 
-                              magma_int_t nb, magma_int_t *info );
-magma_int_t magma_zunmql2_gpu(const char side, const char trans,
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              cuDoubleComplex *da, magma_int_t ldda,
-                              cuDoubleComplex *tau,
-                              cuDoubleComplex *dc, magma_int_t lddc,
-                              cuDoubleComplex *wa, magma_int_t ldwa,
-                              magma_int_t *info);
-magma_int_t magma_zunmqr_gpu( char side, char trans, 
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              cuDoubleComplex *a,    magma_int_t lda, cuDoubleComplex *tau, 
-                              cuDoubleComplex *c,    magma_int_t ldc,
-                              cuDoubleComplex *work, magma_int_t lwork, 
-                              cuDoubleComplex *td,   magma_int_t nb, magma_int_t *info);
-magma_int_t magma_zunmqr2_gpu(const char side, const char trans,
-                              magma_int_t m, magma_int_t n, magma_int_t k,
-                              cuDoubleComplex *da,   magma_int_t ldda,
-                              cuDoubleComplex *tau,
-                              cuDoubleComplex *dc,    magma_int_t lddc,
-                              cuDoubleComplex *wa,    magma_int_t ldwa,
-                              magma_int_t *info);
-magma_int_t magma_zunmtr_gpu( char side, char uplo, char trans,
-                              magma_int_t m, magma_int_t n,
-                              cuDoubleComplex *da,    magma_int_t ldda,
-                              cuDoubleComplex *tau,
-                              cuDoubleComplex *dc,    magma_int_t lddc,
-                              cuDoubleComplex *wa,    magma_int_t ldwa,
-                              magma_int_t *info);
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-magma_int_t magma_zheevd_gpu( char jobz, char uplo,
-                              magma_int_t n,
-                              cuDoubleComplex *da, magma_int_t ldda,
-                              double *w,
-                              cuDoubleComplex *wa,  magma_int_t ldwa,
-                              cuDoubleComplex *work, magma_int_t lwork,
-                              double *rwork, magma_int_t lrwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-magma_int_t magma_zheevdx_gpu(char jobz, char range, char uplo,
-                              magma_int_t n, cuDoubleComplex *da, 
-                              magma_int_t ldda, double vl, double vu, 
-                              magma_int_t il, magma_int_t iu,
-                              magma_int_t *m, double *w,
-                              cuDoubleComplex *wa,  magma_int_t ldwa,
-                              cuDoubleComplex *work, magma_int_t lwork,
-                              double *rwork, magma_int_t lrwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-magma_int_t magma_zheevr_gpu( char jobz, char range, char uplo, magma_int_t n,
-                              cuDoubleComplex *da, magma_int_t ldda, double vl, double vu,
-                              magma_int_t il, magma_int_t iu, double abstol, magma_int_t *m,
-                              double *w, cuDoubleComplex *dz, magma_int_t lddz,
-                              magma_int_t *isuppz,
-                              cuDoubleComplex *wa, magma_int_t ldwa,
-                              cuDoubleComplex *wz, magma_int_t ldwz,
-                              cuDoubleComplex *work, magma_int_t lwork,
-                              double *rwork, magma_int_t lrwork, magma_int_t *iwork,
-                              magma_int_t liwork, magma_int_t *info);
-#else
-magma_int_t magma_zheevd_gpu( char jobz, char uplo,
-                              magma_int_t n,
-                              cuDoubleComplex *da, magma_int_t ldda,
-                              cuDoubleComplex *w,
-                              cuDoubleComplex *wa,  magma_int_t ldwa,
-                              cuDoubleComplex *work, magma_int_t lwork,
-                              magma_int_t *iwork, magma_int_t liwork,
-                              magma_int_t *info);
-#endif
-
-magma_int_t magma_zheevx_gpu( char jobz, char range, char uplo, magma_int_t n,
-                              cuDoubleComplex *da, magma_int_t ldda, double vl, 
-                              double vu, magma_int_t il, magma_int_t iu, 
-                              double abstol, magma_int_t *m,
-                              double *w, cuDoubleComplex *dz, magma_int_t lddz,
-                              cuDoubleComplex *wa, magma_int_t ldwa,
-                              cuDoubleComplex *wz, magma_int_t ldwz,
-                              cuDoubleComplex *work, magma_int_t lwork,
-                              double *rwork, magma_int_t *iwork, 
-                              magma_int_t *ifail, magma_int_t *info);
-magma_int_t magma_zhegst_gpu(magma_int_t itype, char uplo, magma_int_t n,
-                             cuDoubleComplex *da, magma_int_t ldda,
-                             cuDoubleComplex *db, magma_int_t lddb, magma_int_t *info);
-
-
-/* //////////////////////////////////////////////////////////////////////////// 
- -- MAGMA utility function definitions
-*/
-
-void magma_zprint    ( magma_int_t m, magma_int_t n, const cuDoubleComplex  *A, magma_int_t lda  );
-void magma_zprint_gpu( magma_int_t m, magma_int_t n, const cuDoubleComplex *dA, magma_int_t ldda );
-
-void zpanel_to_q( char uplo, magma_int_t ib, cuDoubleComplex *A, magma_int_t lda, cuDoubleComplex *work );
-void zq_to_panel( char uplo, magma_int_t ib, cuDoubleComplex *A, magma_int_t lda, cuDoubleComplex *work );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_z
-#endif /* _MAGMA_Z_H_ */
diff --git a/src/include/magma_zbulge.h b/src/include/magma_zbulge.h
deleted file mode 100644
index c4e947b..0000000
--- a/src/include/magma_zbulge.h
+++ /dev/null
@@ -1,35 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @precisions normal z -> s d c
- */
-
-#ifndef MAGMA_ZBULGEINC_H
-#define MAGMA_ZBULGEINC_H
-
-#define PRECISION_z
-
-extern "C" {
-    magma_int_t magma_zhetrd_he2hb(char uplo, magma_int_t n, magma_int_t NB, cuDoubleComplex *a, magma_int_t lda,
-                                   cuDoubleComplex *tau, cuDoubleComplex *work, magma_int_t lwork, cuDoubleComplex *dT, magma_int_t threads, magma_int_t *info);
-
-    magma_int_t magma_zhetrd_hb2st(magma_int_t threads, char uplo, magma_int_t n, magma_int_t nb, magma_int_t Vblksiz,
-                                   cuDoubleComplex *A, magma_int_t lda, double *D, double *E,
-                                   cuDoubleComplex *V, magma_int_t ldv, cuDoubleComplex *TAU, magma_int_t compT, cuDoubleComplex *T, magma_int_t ldt);
-
-    magma_int_t magma_zbulge_back(magma_int_t threads, char uplo, magma_int_t n, magma_int_t nb, magma_int_t ne, magma_int_t Vblksiz,
-                                  cuDoubleComplex *Z, magma_int_t ldz, cuDoubleComplex *dZ, magma_int_t lddz,
-                                  cuDoubleComplex *V, magma_int_t ldv, cuDoubleComplex *TAU, cuDoubleComplex *T, magma_int_t ldt, magma_int_t* info);
-
-    magma_int_t magma_zunmqr_gpu_2stages(char side, char trans, magma_int_t m, magma_int_t n, magma_int_t k, cuDoubleComplex *dA, magma_int_t ldda,
-                                         cuDoubleComplex *dC, magma_int_t lddc, cuDoubleComplex *dT, magma_int_t nb, magma_int_t *info);
-
-    magma_int_t magma_zbulge_get_lq2(magma_int_t n);
-
-    magma_int_t magma_zbulge_get_Vblksiz(magma_int_t n, magma_int_t nb);
-}
-#endif
diff --git a/src/include/magma_zbulgeinc.h b/src/include/magma_zbulgeinc.h
deleted file mode 100644
index 94f85f1..0000000
--- a/src/include/magma_zbulgeinc.h
+++ /dev/null
@@ -1,141 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @precisions normal z -> s d c
- */
-
-#ifndef MAGMA_ZBULGEINC_H
-#define MAGMA_ZBULGEINC_H
-
-#define PRECISION_z
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-
-/***************************************************************************//**
- *  Configuration
- **/
-
- // maximum contexts
-#define MAX_THREADS_BLG         256
-
-void findVTpos(magma_int_t N, magma_int_t NB, magma_int_t Vblksiz, magma_int_t sweep, magma_int_t st, magma_int_t *Vpos, magma_int_t *TAUpos, magma_int_t *Tpos, magma_int_t *myblkid);
-void findVTsiz(magma_int_t N, magma_int_t NB, magma_int_t Vblksiz, magma_int_t *blkcnt, magma_int_t *LDV);
-magma_int_t plasma_ceildiv(magma_int_t a, magma_int_t b);
-
-
-/*
-extern volatile magma_int_t barrier_in[MAX_THREADS_BLG];
-extern volatile magma_int_t barrier_out[MAX_THREADS_BLG];
-extern volatile magma_int_t *ss_prog;
-*/
-
- /***************************************************************************//**
- *  Static scheduler
- **/
-/*
-#define ssched_init(nbtiles) \
-{ \
-        volatile int   prog_ol[2*nbtiles+10];\
-                 int   iamdone[MAX_THREADS_BLG]; \
-                 int   thread_num[MAX_THREADS_BLG];\
-        pthread_t      thread_id[MAX_THREADS_BLG];\
-        pthread_attr_t thread_attr;\
-}
-*/
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-
-
- struct gbstrct_blg {
-    cuDoubleComplex *dQ1;
-    cuDoubleComplex *dT1;
-    cuDoubleComplex *dT2;
-    cuDoubleComplex *dV2;
-    cuDoubleComplex *dE;
-    cuDoubleComplex *T;
-    cuDoubleComplex *A;
-    cuDoubleComplex *V;
-    cuDoubleComplex *TAU;
-    cuDoubleComplex *E;
-    cuDoubleComplex *E_CPU;
-    int cores_num;
-    int locores_num;
-    int overlapQ1;
-    int usemulticpu;
-    int NB;
-    int NBTILES;
-    int N;
-    int NE;
-    int N_CPU;
-    int N_GPU;
-    int LDA;
-    int LDE;
-    int BAND;
-    int grsiz;
-    int Vblksiz;
-    int WANTZ;
-    char SIDE;
-    real_Double_t *timeblg;
-    real_Double_t *timeaplQ;
-    volatile int *ss_prog;
-} ;
-
-// declare globals here; defined in zhetrd_bhe2trc.cpp
-extern struct gbstrct_blg core_in_all;
-
-
-
-
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-#define MAX_EVENTSBLG 163840
-//#define MAX_EVENTSBLG 1048576
-
-// If we're not using GNU C, elide __attribute__
-#ifndef __GNUC__
-#define __attribute__(x)  /*NOTHING*/
-#endif
-
-// declare globals here; defined in zhetrd_bhe2trc.cpp
-extern int           event_numblg        [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_start_timeblg [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_end_timeblg   [MAX_THREADS_BLG]                 __attribute__ ((aligned (128)));
-extern real_Double_t event_logblg        [MAX_THREADS_BLG][MAX_EVENTSBLG]  __attribute__ ((aligned (128)));
-extern int           log_eventsblg;
-
-#ifndef __GNUC__
-#undef  __attribute__
-#endif
-
-#define core_event_startblg(my_core_id)\
-    event_start_timeblg[my_core_id] = magma_wtime();
-
-#define core_event_endblg(my_core_id)\
-    event_end_timeblg[my_core_id] = magma_wtime();
-
-#define core_log_eventblg(event, my_core_id)\
-    event_logblg[my_core_id][event_numblg[my_core_id]+0] = my_core_id;\
-    event_logblg[my_core_id][event_numblg[my_core_id]+1] = event_start_timeblg[my_core_id];\
-    event_logblg[my_core_id][event_numblg[my_core_id]+2] = event_end_timeblg[my_core_id];\
-    event_logblg[my_core_id][event_numblg[my_core_id]+3] = (event);\
-    event_numblg[my_core_id] += (log_eventsblg << 2);\
-    event_numblg[my_core_id] &= (MAX_EVENTSBLG-1);
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_z
-
-#endif /* MAGMA_ZBULGEINC_H */
diff --git a/src/include/magma_zc.h b/src/include/magma_zc.h
deleted file mode 100644
index b34a524..0000000
--- a/src/include/magma_zc.h
+++ /dev/null
@@ -1,53 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @precisions mixed zc -> ds
- */
-
-#ifndef _MAGMA_ZC_H_
-#define _MAGMA_ZC_H_
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* Mixed precision */
-magma_int_t magma_zcgesv_gpu(  char trans, magma_int_t N, magma_int_t NRHS, 
-                               cuDoubleComplex *dA, magma_int_t ldda, 
-                               magma_int_t *IPIV, magma_int_t *dIPIV,
-                               cuDoubleComplex *dB, magma_int_t lddb, 
-                               cuDoubleComplex *dX, magma_int_t lddx, 
-                               cuDoubleComplex *dworkd, cuFloatComplex *dworks,
-                               magma_int_t *iter, magma_int_t *info);
-
-magma_int_t magma_zcgetrs_gpu( char trans, magma_int_t n, magma_int_t nrhs, 
-                               cuFloatComplex  *dA, magma_int_t ldda,
-                               magma_int_t *ipiv, 
-                               cuDoubleComplex *dB, magma_int_t lddb,
-                               cuDoubleComplex *dX, magma_int_t lddx,
-                               cuFloatComplex  *dSX, 
-                               magma_int_t *info );
-
-magma_int_t magma_zcposv_gpu( char uplo, magma_int_t n, magma_int_t nrhs, 
-                              cuDoubleComplex *dA, magma_int_t ldda, 
-                              cuDoubleComplex *dB, magma_int_t lddb, 
-                              cuDoubleComplex *dX, magma_int_t lddx, 
-                              cuDoubleComplex *dworkd, cuFloatComplex *dworks,
-                              magma_int_t *iter, magma_int_t *info);
-
-magma_int_t magma_zcgeqrsv_gpu(magma_int_t M, magma_int_t N, magma_int_t NRHS, 
-                               cuDoubleComplex *dA,  magma_int_t ldda, 
-                               cuDoubleComplex *dB,  magma_int_t lddb, 
-                               cuDoubleComplex *dX,  magma_int_t lddx,
-                               magma_int_t *iter,    magma_int_t *info);
-  
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif /* _MAGMA_Z_H_ */
diff --git a/src/include/magma_zlapack.h b/src/include/magma_zlapack.h
deleted file mode 100644
index f0a6a38..0000000
--- a/src/include/magma_zlapack.h
+++ /dev/null
@@ -1,953 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @precisions normal z -> s d c
- */
-
-#ifndef MAGMA_ZLAPACK_H
-#define MAGMA_ZLAPACK_H
-
-#define PRECISION_z
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* ////////////////////////////////////////////////////////////////////////////
-   -- BLAS and LAPACK functions (alphabetical order)
-*/
-#define blasf77_izamax     FORTRAN_NAME( izamax, IZAMAX )
-#define blasf77_zaxpy      FORTRAN_NAME( zaxpy,  ZAXPY  )
-#define blasf77_zcopy      FORTRAN_NAME( zcopy,  ZCOPY  )
-#define blasf77_zgemm      FORTRAN_NAME( zgemm,  ZGEMM  )
-#define blasf77_zgemv      FORTRAN_NAME( zgemv,  ZGEMV  )
-#define blasf77_zgerc      FORTRAN_NAME( zgerc,  ZGERC  )
-#define blasf77_zgeru      FORTRAN_NAME( zgeru,  ZGERU  )
-#define blasf77_zhemm      FORTRAN_NAME( zhemm,  ZHEMM  )
-#define blasf77_zhemv      FORTRAN_NAME( zhemv,  ZHEMV  )
-#define blasf77_zher2      FORTRAN_NAME( zher2,  ZHER2  )
-#define blasf77_zher2k     FORTRAN_NAME( zher2k, ZHER2K )
-#define blasf77_zherk      FORTRAN_NAME( zherk,  ZHERK  )
-#define blasf77_zscal      FORTRAN_NAME( zscal,  ZSCAL  )
-#define blasf77_zdscal     FORTRAN_NAME( zdscal, ZDSCAL )
-#define blasf77_zswap      FORTRAN_NAME( zswap,  ZSWAP  )
-#define blasf77_zsymm      FORTRAN_NAME( zsymm,  ZSYMM  )
-#define blasf77_zsyr2k     FORTRAN_NAME( zsyr2k, ZSYR2K )
-#define blasf77_zsyrk      FORTRAN_NAME( zsyrk,  ZSYRK  )
-#define blasf77_ztrmm      FORTRAN_NAME( ztrmm,  ZTRMM  )
-#define blasf77_ztrmv      FORTRAN_NAME( ztrmv,  ZTRMV  )
-#define blasf77_ztrsm      FORTRAN_NAME( ztrsm,  ZTRSM  )
-#define blasf77_ztrsv      FORTRAN_NAME( ztrsv,  ZTRSV  )
-
-#define lapackf77_dlaed4   FORTRAN_NAME( dlaed4, DLAED4 )
-#define lapackf77_dlamc3   FORTRAN_NAME( dlamc3, DLAMC3 )
-#define lapackf77_dlamrg   FORTRAN_NAME( dlamrg, DLAMRG )
-#define lapackf77_dstebz   FORTRAN_NAME( dstebz, DSTEBZ )
-
-#define lapackf77_zbdsqr   FORTRAN_NAME( zbdsqr, ZBDSQR )
-#define lapackf77_zgebak   FORTRAN_NAME( zgebak, ZGEBAK )
-#define lapackf77_zgebal   FORTRAN_NAME( zgebal, ZGEBAL )
-#define lapackf77_zgebd2   FORTRAN_NAME( zgebd2, ZGEBD2 )
-#define lapackf77_zgebrd   FORTRAN_NAME( zgebrd, ZGEBRD )
-#define lapackf77_zgeev    FORTRAN_NAME( zgeev,  ZGEEV  )
-#define lapackf77_zgehd2   FORTRAN_NAME( zgehd2, ZGEHD2 )
-#define lapackf77_zgehrd   FORTRAN_NAME( zgehrd, ZGEHRD )
-#define lapackf77_zgelqf   FORTRAN_NAME( zgelqf, ZGELQF )
-#define lapackf77_zgels    FORTRAN_NAME( zgels,  ZGELS  )
-#define lapackf77_zgeqlf   FORTRAN_NAME( zgeqlf, ZGEQLF )
-#define lapackf77_zgeqp3   FORTRAN_NAME( zgeqp3, ZGEQP3 )
-#define lapackf77_zgeqrf   FORTRAN_NAME( zgeqrf, ZGEQRF )
-#define lapackf77_zgesvd   FORTRAN_NAME( zgesvd, ZGESVD )
-#define lapackf77_zgetrf   FORTRAN_NAME( zgetrf, ZGETRF )
-#define lapackf77_zgetri   FORTRAN_NAME( zgetri, ZGETRI )
-#define lapackf77_zgetrs   FORTRAN_NAME( zgetrs, ZGETRS )
-#define lapackf77_zhbtrd   FORTRAN_NAME( zhbtrd, ZHBTRD )
-#define lapackf77_zheev    FORTRAN_NAME( zheev,  ZHEEV  )
-#define lapackf77_zheevd   FORTRAN_NAME( zheevd, ZHEEVD )
-#define lapackf77_zhegs2   FORTRAN_NAME( zhegs2, ZHEGS2 )
-#define lapackf77_zhegvd   FORTRAN_NAME( zhegvd, ZHEGVD )
-#define lapackf77_zhetd2   FORTRAN_NAME( zhetd2, ZHETD2 )
-#define lapackf77_zhetrd   FORTRAN_NAME( zhetrd, ZHETRD )
-#define lapackf77_zhetrf   FORTRAN_NAME( zhetrf, ZHETRF )
-#define lapackf77_zhseqr   FORTRAN_NAME( zhseqr, ZHSEQR )
-#define lapackf77_zlabrd   FORTRAN_NAME( zlabrd, ZLABRD )
-#define lapackf77_zladiv   FORTRAN_NAME( zladiv, ZLADIV )
-#define lapackf77_zlacgv   FORTRAN_NAME( zlacgv, ZLACGV )
-#define lapackf77_zlacpy   FORTRAN_NAME( zlacpy, ZLACPY )
-#define lapackf77_zlahef   FORTRAN_NAME( zlahef, ZLAHEF )
-#define lapackf77_zlange   FORTRAN_NAME( zlange, ZLANGE )
-#define lapackf77_zlanhe   FORTRAN_NAME( zlanhe, ZLANHE )
-#define lapackf77_zlanht   FORTRAN_NAME( zlanht, ZLANHT )
-#define lapackf77_zlansy   FORTRAN_NAME( zlansy, ZLANSY )
-#define lapackf77_dlapy3   FORTRAN_NAME( dlapy3, DLAPY3 )
-#define lapackf77_zlaqp2   FORTRAN_NAME( zlaqp2, ZLAQP2 )
-#define lapackf77_zlarf    FORTRAN_NAME( zlarf,  ZLARF  )
-#define lapackf77_zlarfb   FORTRAN_NAME( zlarfb, ZLARFB )
-#define lapackf77_zlarfg   FORTRAN_NAME( zlarfg, ZLARFG )
-#define lapackf77_zlarft   FORTRAN_NAME( zlarft, ZLARFT )
-#define lapackf77_zlarnv   FORTRAN_NAME( zlarnv, ZLARNV )
-#define lapackf77_zlartg   FORTRAN_NAME( zlartg, ZLARTG )
-#define lapackf77_zlascl   FORTRAN_NAME( zlascl, ZLASCL )
-#define lapackf77_zlaset   FORTRAN_NAME( zlaset, ZLASET )
-#define lapackf77_zlaswp   FORTRAN_NAME( zlaswp, ZLASWP )
-#define lapackf77_zlatrd   FORTRAN_NAME( zlatrd, ZLATRD )
-#define lapackf77_zlauum   FORTRAN_NAME( zlauum, ZLAUUM )
-#define lapackf77_zlavhe   FORTRAN_NAME( zlavhe, ZLAVHE )
-#define lapackf77_zpotrf   FORTRAN_NAME( zpotrf, ZPOTRF )
-#define lapackf77_zpotri   FORTRAN_NAME( zpotri, ZPOTRI )
-#define lapackf77_zpotrs   FORTRAN_NAME( zpotrs, ZPOTRS )
-#define lapackf77_zstedc   FORTRAN_NAME( zstedc, ZSTEDC )
-#define lapackf77_zstein   FORTRAN_NAME( zstein, ZSTEIN )
-#define lapackf77_zstemr   FORTRAN_NAME( zstemr, ZSTEMR )
-#define lapackf77_zsteqr   FORTRAN_NAME( zsteqr, ZSTEQR )
-#define lapackf77_zsymv    FORTRAN_NAME( zsymv,  ZSYMV  )
-#define lapackf77_ztrevc   FORTRAN_NAME( ztrevc, ZTREVC )
-#define lapackf77_ztrtri   FORTRAN_NAME( ztrtri, ZTRTRI )
-#define lapackf77_zung2r   FORTRAN_NAME( zung2r, ZUNG2R )
-#define lapackf77_zungbr   FORTRAN_NAME( zungbr, ZUNGBR )
-#define lapackf77_zunghr   FORTRAN_NAME( zunghr, ZUNGHR )
-#define lapackf77_zunglq   FORTRAN_NAME( zunglq, ZUNGLQ )
-#define lapackf77_zungql   FORTRAN_NAME( zungql, ZUNGQL )
-#define lapackf77_zungqr   FORTRAN_NAME( zungqr, ZUNGQR )
-#define lapackf77_zungtr   FORTRAN_NAME( zungtr, ZUNGTR )
-#define lapackf77_zunm2r   FORTRAN_NAME( zunm2r, ZUNM2R )
-#define lapackf77_zunmbr   FORTRAN_NAME( zunmbr, ZUNMBR )
-#define lapackf77_zunmlq   FORTRAN_NAME( zunmlq, ZUNMLQ )
-#define lapackf77_zunmql   FORTRAN_NAME( zunmql, ZUNMQL )
-#define lapackf77_zunmqr   FORTRAN_NAME( zunmqr, ZUNMQR )
-#define lapackf77_zunmtr   FORTRAN_NAME( zunmtr, ZUNMTR )
-
-/* testing functions */
-#define lapackf77_zbdt01   FORTRAN_NAME( zbdt01, ZBDT01 )
-#define lapackf77_zget22   FORTRAN_NAME( zget22, ZGET22 )
-#define lapackf77_zqpt01   FORTRAN_NAME( zqpt01, ZQPT01 )
-#define lapackf77_zhet21   FORTRAN_NAME( zhet21, ZHET21 )
-#define lapackf77_zhst01   FORTRAN_NAME( zhst01, ZHST01 )
-#define lapackf77_zqrt02   FORTRAN_NAME( zqrt02, ZQRT02 )
-#define lapackf77_zunt01   FORTRAN_NAME( zunt01, ZUNT01 )
-#define lapackf77_zlarfy   FORTRAN_NAME( zlarfy, ZLARFY )
-#define lapackf77_zlarfx   FORTRAN_NAME( zlarfx, ZLARFX )
-#define lapackf77_zstt21   FORTRAN_NAME( zstt21, ZSTT21 )
-
-// macros to handle differences in arguments between complex and real versions of routines.
-#if defined(PRECISION_z) || defined(PRECISION_c)
-#define DWORKFORZ        double *rwork,
-#define DWORKFORZ_AND_LD double *rwork, const magma_int_t *ldrwork,
-#define WSPLIT           cuDoubleComplex *w
-#else
-#define DWORKFORZ
-#define DWORKFORZ_AND_LD
-#define WSPLIT           double *wr, double *wi
-#endif
-
-/*
- * BLAS functions (alphabetical order)
- */
-magma_int_t blasf77_izamax(
-                     const magma_int_t *n,
-                     const cuDoubleComplex *x, const magma_int_t *incx);
-
-void blasf77_zaxpy(  const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *x, const magma_int_t *incx,
-                           cuDoubleComplex *y, const magma_int_t *incy );
-
-void blasf77_zcopy(  const magma_int_t *n,
-                     const cuDoubleComplex *x, const magma_int_t *incx,
-                           cuDoubleComplex *y, const magma_int_t *incy );
-
-void blasf77_zgemm(  const char *transa, const char *transb,
-                     const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                     const cuDoubleComplex *B, const magma_int_t *ldb,
-                     const cuDoubleComplex *beta,
-                           cuDoubleComplex *C, const magma_int_t *ldc );
-
-void blasf77_zgemv(  const char *transa,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                     const cuDoubleComplex *x, const magma_int_t *incx,
-                     const cuDoubleComplex *beta,
-                           cuDoubleComplex *y, const magma_int_t *incy );
-
-void blasf77_zgerc(  const magma_int_t *m, const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *x, const magma_int_t *incx,
-                     const cuDoubleComplex *y, const magma_int_t *incy,
-                           cuDoubleComplex *A, const magma_int_t *lda );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void blasf77_zgeru(  const magma_int_t *m, const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *x, const magma_int_t *incx,
-                     const cuDoubleComplex *y, const magma_int_t *incy,
-                           cuDoubleComplex *A, const magma_int_t *lda );
-#endif
-
-void blasf77_zhemm(  const char *side, const char *uplo,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                     const cuDoubleComplex *B, const magma_int_t *ldb,
-                     const cuDoubleComplex *beta,
-                           cuDoubleComplex *C, const magma_int_t *ldc );
-
-void blasf77_zhemv(  const char *uplo,
-                     const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                     const cuDoubleComplex *x, const magma_int_t *incx,
-                     const cuDoubleComplex *beta,
-                           cuDoubleComplex *y, const magma_int_t *incy );
-
-void blasf77_zher2(  const char *uplo,
-                     const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *x, const magma_int_t *incx,
-                     const cuDoubleComplex *y, const magma_int_t *incy,
-                           cuDoubleComplex *A, const magma_int_t *lda );
-
-void blasf77_zher2k(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                     const cuDoubleComplex *B, const magma_int_t *ldb,
-                     const double *beta,
-                           cuDoubleComplex *C, const magma_int_t *ldc );
-
-void blasf77_zherk(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const double *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                     const double *beta,
-                           cuDoubleComplex *C, const magma_int_t *ldc );
-
-void blasf77_zscal(  const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                           cuDoubleComplex *x, const magma_int_t *incx );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void blasf77_zdscal( const magma_int_t *n,
-                     const double *alpha,
-                           cuDoubleComplex *x, const magma_int_t *incx );
-#endif
-
-void blasf77_zswap(  const magma_int_t *n,
-                     cuDoubleComplex *x, const magma_int_t *incx,
-                     cuDoubleComplex *y, const magma_int_t *incy );
-
-void blasf77_zsymm(  const char *side, const char *uplo,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                     const cuDoubleComplex *B, const magma_int_t *ldb,
-                     const cuDoubleComplex *beta,
-                           cuDoubleComplex *C, const magma_int_t *ldc );
-
-void blasf77_zsyr2k( const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                     const cuDoubleComplex *B, const magma_int_t *ldb,
-                     const cuDoubleComplex *beta,
-                           cuDoubleComplex *C, const magma_int_t *ldc );
-
-void blasf77_zsyrk(  const char *uplo, const char *trans,
-                     const magma_int_t *n, const magma_int_t *k,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                     const cuDoubleComplex *beta,
-                           cuDoubleComplex *C, const magma_int_t *ldc );
-
-void blasf77_ztrmm(  const char *side, const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                           cuDoubleComplex *B, const magma_int_t *ldb );
-
-void blasf77_ztrmv(  const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *n,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                           cuDoubleComplex *x, const magma_int_t *incx );
-
-void blasf77_ztrsm(  const char *side, const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *m, const magma_int_t *n,
-                     const cuDoubleComplex *alpha,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                           cuDoubleComplex *B, const magma_int_t *ldb );
-
-void blasf77_ztrsv(  const char *uplo, const char *transa, const char *diag,
-                     const magma_int_t *n,
-                     const cuDoubleComplex *A, const magma_int_t *lda,
-                           cuDoubleComplex *x, const magma_int_t *incx );
-
-/*
- * LAPACK functions (alphabetical order)
- */
-void   lapackf77_zbdsqr( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *ncvt, const magma_int_t *nru,  const magma_int_t *ncc,
-                         double *d, double *e,
-                         cuDoubleComplex *Vt, const magma_int_t *ldvt,
-                         cuDoubleComplex *U, const magma_int_t *ldu,
-                         cuDoubleComplex *C, const magma_int_t *ldc,
-                         double *work,
-                         magma_int_t *info );
-
-void   lapackf77_zgebak( const char *job, const char *side,
-                         const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         const double *scale, const magma_int_t *m,
-                         cuDoubleComplex *V, const magma_int_t *ldv,
-                         magma_int_t *info );
-
-void   lapackf77_zgebal( const char *job,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         magma_int_t *ilo, magma_int_t *ihi,
-                         double *scale,
-                         magma_int_t *info );
-
-void   lapackf77_zgebd2( const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         cuDoubleComplex *tauq,
-                         cuDoubleComplex *taup,
-                         cuDoubleComplex *work,
-                         magma_int_t *info );
-
-void   lapackf77_zgebrd( const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         cuDoubleComplex *tauq,
-                         cuDoubleComplex *taup,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zgeev(  const char *jobvl, const char *jobvr,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A,    const magma_int_t *lda,
-                         WSPLIT,
-                         cuDoubleComplex *Vl,   const magma_int_t *ldvl,
-                         cuDoubleComplex *Vr,   const magma_int_t *ldvr,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_zgehd2( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work,
-                         magma_int_t *info );
-
-void   lapackf77_zgehrd( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zgelqf( const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zgels(  const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *nrhs,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *B, const magma_int_t *ldb,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zgeqlf( const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zgeqp3( const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         magma_int_t *jpvt,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_zgeqrf( const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zgesvd( const char *jobu, const char *jobvt,
-                         const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *s,
-                         cuDoubleComplex *U,  const magma_int_t *ldu,
-                         cuDoubleComplex *Vt, const magma_int_t *ldvt,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_zgetrf( const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         magma_int_t *info );
-
-void   lapackf77_zgetri( const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         const magma_int_t *ipiv,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zgetrs( const char* trans,
-                         const magma_int_t *n, const magma_int_t *nrhs,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         const magma_int_t *ipiv,
-                         cuDoubleComplex *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_zhbtrd( const char *vect, const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kd,
-                         cuDoubleComplex *Ab, const magma_int_t *ldab,
-                         double *d, double *e,
-                         cuDoubleComplex *Q, const magma_int_t *ldq,
-                         cuDoubleComplex *work,
-                         magma_int_t *info );
-
-void   lapackf77_zheev(  const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *w,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_zheevd( const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *w,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_zhegs2( const magma_int_t *itype, const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_zhegvd( const magma_int_t *itype, const char *jobz, const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *B, const magma_int_t *ldb,
-                         double *w,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_zhetd2( const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         cuDoubleComplex *tau,
-                         magma_int_t *info );
-
-void   lapackf77_zhetrd( const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zhetrf( const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zhseqr( const char *job, const char *compz,
-                         const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuDoubleComplex *H, const magma_int_t *ldh,
-                         WSPLIT,
-                         cuDoubleComplex *Z, const magma_int_t *ldz,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zlabrd( const magma_int_t *m, const magma_int_t *n, const magma_int_t *nb,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         cuDoubleComplex *tauq,
-                         cuDoubleComplex *taup,
-                         cuDoubleComplex *X, const magma_int_t *ldx,
-                         cuDoubleComplex *Y, const magma_int_t *ldy );
-
-void   lapackf77_zladiv( cuDoubleComplex *ret_val, cuDoubleComplex *x, 
-                         cuDoubleComplex *y );
-
-void   lapackf77_zlacgv( const magma_int_t *n,
-                         cuDoubleComplex *x, const magma_int_t *incx );
-
-void   lapackf77_zlacpy( const char *uplo,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *B, const magma_int_t *ldb );
-
-void   lapackf77_zlahef( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kn,
-                         magma_int_t *kb,
-                         cuDoubleComplex *A, const magma_int_t lda,
-                         magma_int_t *ipiv,
-                         cuDoubleComplex *work, const magma_int_t *ldwork,
-                         magma_int_t *info );
-
-double lapackf77_zlange( const char *norm,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         double *work );
-
-double lapackf77_zlanhe( const char *norm, const char *uplo,
-                         const magma_int_t *n,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         double * work );
-
-double lapackf77_zlanht( const char* norm, const magma_int_t* n,
-                         const double* d, const cuDoubleComplex* e );
-
-double lapackf77_zlansy( const char *norm, const char *uplo,
-                         const magma_int_t *n,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         double * work );
-
-void lapackf77_zlaqp2 (  magma_int_t *m, magma_int_t *n, magma_int_t *offset,
-                         cuDoubleComplex *a, magma_int_t *lda, magma_int_t *jpvt, 
-                         cuDoubleComplex *tau,
-                         double *vn1, double *vn2, cuDoubleComplex *work);
-
-void lapackf77_zlarf  (  char *, magma_int_t *, magma_int_t *,
-                         cuDoubleComplex *, magma_int_t *, cuDoubleComplex *, cuDoubleComplex *,
-                         magma_int_t *, cuDoubleComplex *);
-
-void   lapackf77_zlarfb( const char *side, const char *trans, const char *direct, const char *storev,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuDoubleComplex *V, const magma_int_t *ldv,
-                         const cuDoubleComplex *T, const magma_int_t *ldt,
-                         cuDoubleComplex *C, const magma_int_t *ldc,
-                         cuDoubleComplex *work, const magma_int_t *ldwork );
-
-void   lapackf77_zlarfg( const magma_int_t *n,
-                         cuDoubleComplex *alpha,
-                         cuDoubleComplex *x, const magma_int_t *incx,
-                         cuDoubleComplex *tau );
-
-void   lapackf77_zlarft( const char *direct, const char *storev,
-                         const magma_int_t *n, const magma_int_t *k,
-                         cuDoubleComplex *V, const magma_int_t *ldv,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *T, const magma_int_t *ldt );
-
-void   lapackf77_zlarnv( const magma_int_t *idist, magma_int_t *iseed, const magma_int_t *n,
-                         cuDoubleComplex *x );
-
-void   lapackf77_zlartg( cuDoubleComplex *F,
-                         cuDoubleComplex *G,
-                         double *cs,
-                         cuDoubleComplex *SN,
-                         cuDoubleComplex *R );
-
-void   lapackf77_zlascl( const char *type,
-                         const magma_int_t *kl, const magma_int_t *ku,
-                         double *cfrom,
-                         double *cto,
-                         const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_zlaset( const char *uplo,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const cuDoubleComplex *alpha,
-                         const cuDoubleComplex *beta,
-                         cuDoubleComplex *A, const magma_int_t *lda );
-
-void   lapackf77_zlaswp( const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         const magma_int_t *k1, const magma_int_t *k2,
-                         magma_int_t *ipiv,
-                         const magma_int_t *incx );
-
-void   lapackf77_zlatrd( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *nb,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *e,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *ldwork );
-
-void   lapackf77_zlauum( const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_zlavhe( const char *uplo, const char *trans, const char *diag,
-                         magma_int_t *n, magma_int_t *nrhs,
-                         cuDoubleComplex *A, magma_int_t *lda,
-                         magma_int_t *ipiv,
-                         cuDoubleComplex *B, magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_zpotrf( const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_zpotri( const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_zpotrs( const char *uplo,
-                         const magma_int_t *n, const magma_int_t *nrhs,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *B, const magma_int_t *ldb,
-                         magma_int_t *info );
-
-void   lapackf77_zstedc( const char *compz,
-                         const magma_int_t *n,
-                         double *d, double *e,
-                         cuDoubleComplex *Z, const magma_int_t *ldz,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         DWORKFORZ_AND_LD
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_zstein( const magma_int_t *n,
-                         const double *d, const double *e,
-                         const magma_int_t *m,
-                         const double *w,
-                         const magma_int_t *iblock,
-                         const magma_int_t *isplit,
-                         cuDoubleComplex *Z, const magma_int_t *ldz,
-                         double *work, magma_int_t *iwork, magma_int_t *ifailv,
-                         magma_int_t *info );
-
-void   lapackf77_zstemr( const char *jobz, const char *range,
-                         const magma_int_t *n,
-                         double *d, double *e,
-                         const double *vl, const double *vu,
-                         const magma_int_t *il, const magma_int_t *iu,
-                         magma_int_t *m,
-                         double *w,
-                         cuDoubleComplex *Z, const magma_int_t *ldz,
-                         const magma_int_t *nzc, magma_int_t *isuppz, magma_int_t *tryrac,
-                         double *work, const magma_int_t *lwork,
-                         magma_int_t *iwork, const magma_int_t *liwork,
-                         magma_int_t *info );
-
-void   lapackf77_zsteqr( const char *compz,
-                         const magma_int_t *n,
-                         double *d, double *e,
-                         cuDoubleComplex *Z, const magma_int_t *ldz,
-                         double *work,
-                         magma_int_t *info );
-
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void   lapackf77_zsymv(  const char *uplo,
-                         const magma_int_t *n,
-                         const cuDoubleComplex *alpha,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *x, const magma_int_t *incx,
-                         const cuDoubleComplex *beta,
-                               cuDoubleComplex *y, const magma_int_t *incy );
-#endif
-
-void   lapackf77_ztrevc( const char *side, const char *howmny,
-                         magma_int_t *select, const magma_int_t *n,
-                         cuDoubleComplex *T,  const magma_int_t *ldt,
-                         cuDoubleComplex *Vl, const magma_int_t *ldvl,
-                         cuDoubleComplex *Vr, const magma_int_t *ldvr,
-                         const magma_int_t *mm, magma_int_t *m,
-                         cuDoubleComplex *work,
-                         DWORKFORZ
-                         magma_int_t *info );
-
-void   lapackf77_ztrtri( const char *uplo, const char *diag,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         magma_int_t *info );
-
-void   lapackf77_zung2r( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *work,
-                         magma_int_t *info );
-
-void   lapackf77_zungbr( const char *vect,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zunghr( const magma_int_t *n,
-                         const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zunglq( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zungql( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zungqr( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zungtr( const char *uplo,
-                         const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zunm2r( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *C, const magma_int_t *ldc,
-                         cuDoubleComplex *work,
-                         magma_int_t *info );
-
-void   lapackf77_zunmbr( const char *vect, const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *C, const magma_int_t *ldc,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zunmlq( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *C, const magma_int_t *ldc,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zunmql( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *C, const magma_int_t *ldc,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zunmqr( const char *side, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *C, const magma_int_t *ldc,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-void   lapackf77_zunmtr( const char *side, const char *uplo, const char *trans,
-                         const magma_int_t *m, const magma_int_t *n,
-                         const cuDoubleComplex *A, const magma_int_t *lda,
-                         const cuDoubleComplex *tau,
-                         cuDoubleComplex *C, const magma_int_t *ldc,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         magma_int_t *info );
-
-/*
- * Real precision extras
- */
-void   lapackf77_dstebz( const char *range, const char *order,
-                         const magma_int_t *n,
-                         double *vl, double *vu,
-                         magma_int_t *il, magma_int_t *iu,
-                         double *abstol,
-                         double *d, double *e,
-                         const magma_int_t *m, const magma_int_t *nsplit,
-                         double *w,
-                         magma_int_t *iblock, magma_int_t *isplit,
-                         double *work,
-                         magma_int_t *iwork,
-                         magma_int_t *info );
-
-double lapackf77_dlamc3( double* a, double* b );
-
-void   lapackf77_dlamrg( magma_int_t* n1, magma_int_t* n2,
-                         double* a,
-                         magma_int_t* dtrd1, magma_int_t* dtrd2, magma_int_t* index );
-
-double lapackf77_dlapy3(double *, double *, double *);
-
-void   lapackf77_dlaed4( magma_int_t* n, magma_int_t* i,
-                         double* d,
-                         double* z,
-                         double* delta,
-                         double* rho,
-                         double* dlam, magma_int_t* info );
-
-/*
- * Testing functions
- */
-#if defined(PRECISION_z) || defined(PRECISION_c)
-void   lapackf77_zbdt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *kd,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *Q, const magma_int_t *ldq,
-                         double *d, double *e,
-                         cuDoubleComplex *Pt, const magma_int_t *ldpt,
-                         cuDoubleComplex *work,
-                         double *rwork,
-                         double *resid );
-
-void   lapackf77_zget22( const char *transa, const char *transe, const char *transw, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *E, const magma_int_t *lde,
-                         cuDoubleComplex *w,
-                         cuDoubleComplex *work,
-                         double *rwork,
-                         double *result );
-
-void   lapackf77_zhet21( const magma_int_t *itype, const char *uplo,
-                         const magma_int_t *n, const magma_int_t *kband,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         cuDoubleComplex *U, const magma_int_t *ldu,
-                         cuDoubleComplex *V, const magma_int_t *ldv,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work,
-                         double *rwork,
-                         double *result );
-
-void   lapackf77_zhst01( const magma_int_t *n, const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *H, const magma_int_t *ldh,
-                         cuDoubleComplex *Q, const magma_int_t *ldq,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         double *rwork,
-                         double *result );
-
-void   lapackf77_zstt21( const magma_int_t *n, const magma_int_t *kband,
-                         double *AD,
-                         double *AE,
-                         double *SD,
-                         double *SE,
-                         cuDoubleComplex *U, const magma_int_t *ldu,
-                         cuDoubleComplex *work,
-                         double *rwork,
-                         double *result );
-
-void   lapackf77_zunt01( const char *rowcol, const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *U, const magma_int_t *ldu,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         double *rwork,
-                         double *resid );
-#else
-void   lapackf77_zbdt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *kd,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *Q, const magma_int_t *ldq,
-                         double *d, double *e,
-                         cuDoubleComplex *Pt, const magma_int_t *ldpt,
-                         cuDoubleComplex *work,
-                         double *resid );
-
-void   lapackf77_zget22( const char *transa, const char *transe, const char *transw, const magma_int_t *n,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *E, const magma_int_t *lde,
-                         cuDoubleComplex *wr,
-                         cuDoubleComplex *wi,
-                         double *work,
-                         double *result );
-
-void   lapackf77_zhet21( magma_int_t *itype, const char *uplo, const magma_int_t *n, const magma_int_t *kband,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         double *d, double *e,
-                         cuDoubleComplex *U, const magma_int_t *ldu,
-                         cuDoubleComplex *V, const magma_int_t *ldv,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work,
-                         double *result );
-
-void   lapackf77_zhst01( const magma_int_t *n, const magma_int_t *ilo, const magma_int_t *ihi,
-                         cuDoubleComplex *A, const magma_int_t *lda,
-                         cuDoubleComplex *H, const magma_int_t *ldh,
-                         cuDoubleComplex *Q, const magma_int_t *ldq,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         double *result );
-
-void   lapackf77_zstt21( const magma_int_t *n, const magma_int_t *kband,
-                         double *AD,
-                         double *AE,
-                         double *SD,
-                         double *SE,
-                         cuDoubleComplex *U, const magma_int_t *ldu,
-                         cuDoubleComplex *work,
-                         double *result );
-
-void   lapackf77_zunt01( const char *rowcol, const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *U, const magma_int_t *ldu,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         double *resid );
-#endif
-
-void   lapackf77_zlarfy( const char *uplo, const magma_int_t *n,
-                         cuDoubleComplex *V, const magma_int_t *incv,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *C, const magma_int_t *ldc,
-                         cuDoubleComplex *work );
-
-void   lapackf77_zlarfx( const char *side, const magma_int_t *m, const magma_int_t *n,
-                         cuDoubleComplex *V,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *C, const magma_int_t *ldc,
-                         cuDoubleComplex *work );
-
-double lapackf77_zqpt01( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuDoubleComplex *A,
-                         cuDoubleComplex *Af, const magma_int_t *lda,
-                         cuDoubleComplex *tau, magma_int_t *jpvt,
-                         cuDoubleComplex *work, const magma_int_t *lwork );
-
-void   lapackf77_zqrt02( const magma_int_t *m, const magma_int_t *n, const magma_int_t *k,
-                         cuDoubleComplex *A,
-                         cuDoubleComplex *AF,
-                         cuDoubleComplex *Q,
-                         cuDoubleComplex *R, const magma_int_t *lda,
-                         cuDoubleComplex *tau,
-                         cuDoubleComplex *work, const magma_int_t *lwork,
-                         double *rwork,
-                         double *result );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef DWORKFORZ
-#undef DWORKFORZ_AND_LD
-#undef WSPLIT
-#undef PRECISION_z
-
-#endif /* MAGMA_ZLAPACK_H */
diff --git a/src/include/magmablas.h b/src/include/magmablas.h
deleted file mode 100644
index 4c46851..0000000
--- a/src/include/magmablas.h
+++ /dev/null
@@ -1,63 +0,0 @@
-/*
-    -- MAGMA (version 1.3.0) --
-       Univ. of Tennessee, Knoxville
-       Univ. of California, Berkeley
-       Univ. of Colorado, Denver
-       November 2012
-*/
-
-#ifndef _MAGMABLAS_
-#define _MAGMABLAS_
-
-#include <cublas.h>
-
-typedef int  magma_int_t;
-typedef int  magma_err_t;
-typedef void* magma_devptr;
-
-// For now, make these compatible with old cublas v1 prototypes.
-// In the future, we will redefine these data types and
-// add queues (opencl queues, cublas handles).
-typedef char magma_trans_t;
-typedef char magma_side_t ;
-typedef char magma_uplo_t ;
-typedef char magma_diag_t ;
-
-typedef cudaStream_t magma_stream_t;
-typedef cudaStream_t magma_queue_t;
-typedef cudaEvent_t  magma_event_t;
-typedef int          magma_device_t;
-
-// needed by magmablas*.h, but should eventually go in magma_types.h (see clMAGMA)
-#define MagmaMaxGPUs       8
-
-#include "magmablas_z.h"
-#include "magmablas_c.h"
-#include "magmablas_d.h"
-#include "magmablas_s.h"
-#include "magmablas_zc.h"
-#include "magmablas_ds.h"
-
-#if (GPUSHMEM < 200)  
-#define magmablas_zgemm cublasZgemm
-#endif
-#define magmablas_cgemm cublasCgemm
-
-/** ****************************************************************************
- *  Define magma streams
- */
-
-extern cudaStream_t magma_stream;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-cublasStatus_t magmablasSetKernelStream( cudaStream_t stream );
-cublasStatus_t magmablasGetKernelStream( cudaStream_t *stream );
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
diff --git a/src/include/magmablas_c.h b/src/include/magmablas_c.h
deleted file mode 100644
index 92f2ef1..0000000
--- a/src/include/magmablas_c.h
+++ /dev/null
@@ -1,617 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated c Wed Nov 14 22:52:26 2012
- */
-
-#ifndef MAGMABLAS_C_H
-#define MAGMABLAS_C_H
-
-#define PRECISION_c
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-  /*
-   * Interface to clean
-   */
-float cpu_gpu_cdiff(
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *hA, magma_int_t lda,
-    const cuFloatComplex *dA, magma_int_t ldda );
-
-// see also claset
-void czero_32x32_block(
-    cuFloatComplex *dA, magma_int_t ldda );
-
-void czero_nbxnb_block(
-    magma_int_t nb, cuFloatComplex *dA, magma_int_t ldda );
-
-// see also claswp
-void magmablas_cpermute_long2(
-    magma_int_t n, cuFloatComplex *dAT, magma_int_t ldda,
-    magma_int_t *ipiv, magma_int_t nb, magma_int_t ind );
-
-void magmablas_cpermute_long3(
-    /*magma_int_t n,*/ cuFloatComplex *dAT, magma_int_t ldda,
-    const magma_int_t *ipiv, magma_int_t nb, magma_int_t ind );
-
-  /*
-   * Transpose functions
-   */
-void magmablas_cinplace_transpose(
-    cuFloatComplex *dA, magma_int_t ldda, magma_int_t n );
-
-void magmablas_ctranspose(
-    cuFloatComplex       *odata, magma_int_t ldo,
-    const cuFloatComplex *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n );
-
-void magmablas_ctranspose2(
-    cuFloatComplex       *odata, magma_int_t ldo,
-    const cuFloatComplex *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n );
-
-void magmablas_ctranspose2s(
-    cuFloatComplex       *odata, magma_int_t ldo,
-    const cuFloatComplex *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n,
-    cudaStream_t *stream );
-
-void magmablas_cgetmatrix_transpose(
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *dAT,   magma_int_t ldda,
-    cuFloatComplex       *hA,    magma_int_t lda,
-    cuFloatComplex       *dwork, magma_int_t lddwork, magma_int_t nb );
-
-void magmablas_csetmatrix_transpose(
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *hA,    magma_int_t lda,
-    cuFloatComplex       *dAT,   magma_int_t ldda,
-    cuFloatComplex       *dwork, magma_int_t lddwork, magma_int_t nb );
-
-  /*
-   * Multi-GPU functions
-   */
-void magmablas_cgetmatrix_transpose_mgpu(
-    magma_int_t ngpu, cudaStream_t stream[][2],
-    cuFloatComplex **dAT, magma_int_t ldda,
-    cuFloatComplex  *hA,  magma_int_t lda,
-    cuFloatComplex **dB,  magma_int_t lddb,
-    magma_int_t m, magma_int_t n, magma_int_t nb );
-
-void magmablas_csetmatrix_transpose_mgpu(
-    magma_int_t ngpu, cudaStream_t stream[][2],
-    const cuFloatComplex  *hA,  magma_int_t lda,
-    cuFloatComplex       **dAT, magma_int_t ldda,
-    cuFloatComplex       **dB,  magma_int_t lddb,
-    magma_int_t m, magma_int_t n, magma_int_t nb );
-
-void magmablas_cgetmatrix_1D_bcyclic(
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex *dA[], magma_int_t ldda,
-    cuFloatComplex       *hA,   magma_int_t lda,
-    magma_int_t ngpu, magma_int_t nb );
-
-void magmablas_csetmatrix_1D_bcyclic(
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *hA,   magma_int_t lda,
-    cuFloatComplex       *dA[], magma_int_t ldda,
-    magma_int_t ngpu, magma_int_t nb );
-
-void magmablas_chemm_1gpu_old(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    cuFloatComplex *dA[], magma_int_t ldda,  magma_int_t offset,
-    cuFloatComplex *dB[], magma_int_t lddb,
-    cuFloatComplex beta,
-    cuFloatComplex *dC[], magma_int_t lddc,
-    cuFloatComplex *C,    magma_int_t ldc,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-void magmablas_chemm_1gpu(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    cuFloatComplex *dA[], magma_int_t ldda,  magma_int_t offset,
-    cuFloatComplex *dB[], magma_int_t lddb,
-    cuFloatComplex beta,
-    cuFloatComplex *dC[], magma_int_t lddc,
-    cuFloatComplex *C,    magma_int_t ldc,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-
-void magmablas_chemm_mgpu(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha, cuFloatComplex *dA[], magma_int_t ldda,  magma_int_t offset,
-                           cuFloatComplex *dB[], magma_int_t lddb,
-    cuFloatComplex beta,  cuFloatComplex *dC[], magma_int_t lddc,
-                           cuFloatComplex *dwork[],    magma_int_t lddwork,
-                           cuFloatComplex *C,    magma_int_t ldc,
-                           cuFloatComplex *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][20],magma_int_t nbevents );
-
-void magmablas_chemm_mgpu_com(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha, cuFloatComplex *dA[], magma_int_t ldda,  magma_int_t offset,
-                           cuFloatComplex *dB[], magma_int_t lddb,
-    cuFloatComplex beta,  cuFloatComplex *dC[], magma_int_t lddc,
-                           cuFloatComplex *dwork[],    magma_int_t lddwork,
-                           cuFloatComplex *C,    magma_int_t ldc,
-                           cuFloatComplex *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][20],magma_int_t nbevents, 
-                           magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2], magma_int_t nbcmplx );
-
-void magmablas_chemm_mgpu_spec(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha, cuFloatComplex *dA[], magma_int_t ldda,  magma_int_t offset,
-                           cuFloatComplex *dB[], magma_int_t lddb,
-    cuFloatComplex beta,  cuFloatComplex *dC[], magma_int_t lddc,
-                           cuFloatComplex *dwork[],    magma_int_t lddwork,
-                           cuFloatComplex *C,    magma_int_t ldc,
-                           cuFloatComplex *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][MagmaMaxGPUs*MagmaMaxGPUs],magma_int_t nbevents, 
-                           magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2], magma_int_t nbcmplx );
-
-
-
-void magmablas_cher2k_mgpu2(
-    char uplo, char trans, magma_int_t n, magma_int_t k,
-    cuFloatComplex alpha,
-    cuFloatComplex *dA[], magma_int_t ldda, magma_int_t aoff,
-    cuFloatComplex *dB[], magma_int_t lddb, magma_int_t boff,
-    float beta,
-    cuFloatComplex *dC[], magma_int_t lddc, magma_int_t offset,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-void magmablas_cher2k_mgpu_spec(
-    char uplo, char trans, magma_int_t n, magma_int_t k,
-    cuFloatComplex alpha, cuFloatComplex *dA[], magma_int_t lda, magma_int_t aoff,
-                           cuFloatComplex *dB[], magma_int_t ldb, magma_int_t boff,
-    float beta,           cuFloatComplex *dC[], magma_int_t ldc,  magma_int_t offset,
-    magma_int_t ngpu, magma_int_t nb, cudaStream_t streams[][20], magma_int_t nstream );
-
-  /*
-   * LAPACK auxiliary functions
-   */
-void magmablas_cgeadd(
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    cuFloatComplex       *dB, magma_int_t lddb );
-
-void magmablas_clacpy(
-    char uplo,
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    cuFloatComplex       *dB, magma_int_t lddb );
-
-float magmablas_clange(
-    char norm,
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *dA, magma_int_t ldda, float *dwork );
-
-float magmablas_clanhe(
-    char norm, char uplo,
-    magma_int_t n,
-    const cuFloatComplex *dA, magma_int_t ldda, float *dwork );
-
-float magmablas_clansy(
-    char norm, char uplo,
-    magma_int_t n,
-    const cuFloatComplex *dA, magma_int_t ldda, float *dwork );
-
-void magmablas_clascl(
-    char type, magma_int_t kl, magma_int_t ku,
-    float cfrom, float cto,
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex *dA, magma_int_t ldda, magma_int_t *info );
-
-void magmablas_claset(
-    char uplo, magma_int_t m, magma_int_t n,
-    cuFloatComplex *dA, magma_int_t ldda );
-
-void magmablas_claset_identity(
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex *dA, magma_int_t ldda );
-
-void magmablas_claswp(
-    magma_int_t n,
-    cuFloatComplex *dAT, magma_int_t ldda,
-    magma_int_t i1,  magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci );
-
-void magmablas_claswpx(
-    magma_int_t n,
-    cuFloatComplex *dAT, magma_int_t ldx, magma_int_t ldy,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci );
-
-void magmablas_claswp2(
-    magma_int_t n,
-    cuFloatComplex* dAT, magma_int_t ldda,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *d_ipiv );
-
-void magmablas_csymmetrize(
-    char uplo, magma_int_t m, cuFloatComplex *dA, magma_int_t ldda );
-
-void magmablas_csymmetrize_tiles(
-    char uplo, magma_int_t m, cuFloatComplex *dA, magma_int_t ldda,
-    magma_int_t ntile, magma_int_t mstride, magma_int_t nstride );
-
-  /*
-   * Level 1 BLAS
-   */
-void magmablas_cswap(
-    magma_int_t n,
-    cuFloatComplex *dA, magma_int_t ldda,
-    cuFloatComplex *dB, magma_int_t lddb );
-
-void magmablas_cswapblk(
-    char storev,
-    magma_int_t n,
-    cuFloatComplex *dA, magma_int_t ldda,
-    cuFloatComplex *dB, magma_int_t lddb,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci,
-    magma_int_t offset );
-
-void magmablas_cswapdblk(
-    magma_int_t n, magma_int_t nb,
-    cuFloatComplex *dA, magma_int_t ldda, magma_int_t inca,
-    cuFloatComplex *dB, magma_int_t lddb, magma_int_t incb );
-
-  /*
-   * Level 2 BLAS
-   */
-void magmablas_cgemv(
-    char t, magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dx, magma_int_t incx,
-    cuFloatComplex beta,
-    cuFloatComplex       *dy, magma_int_t incy );
-
-#if defined(PRECISION_z ) || defined(PRECISION_c )
-magma_int_t magmablas_chemv(
-    char u, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dx, magma_int_t incx,
-    cuFloatComplex beta,
-    cuFloatComplex       *dy, magma_int_t incy );
-#endif
-
-magma_int_t magmablas_csymv(
-    char u, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dx, magma_int_t incx,
-    cuFloatComplex beta,
-    cuFloatComplex       *dy, magma_int_t incy );
-
-  /*
-   * Level 3 BLAS
-   */
-void magmablas_cgemm(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dB, magma_int_t lddb,
-    cuFloatComplex beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magmablas_cgemm_fermi80(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dB, magma_int_t lddb,
-    cuFloatComplex beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magmablas_cgemm_fermi64(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dB, magma_int_t lddb,
-    cuFloatComplex beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magmablas_chemm(
-    char s, char u,
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dB, magma_int_t lddb,
-    cuFloatComplex beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magmablas_csymm(
-    char s, char u,
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dB, magma_int_t lddb,
-    cuFloatComplex beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magmablas_csyrk(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    cuFloatComplex beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magmablas_cherk(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    float  alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    float  beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magmablas_csyr2k(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dB, magma_int_t lddb,
-    cuFloatComplex beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magmablas_cher2k(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dB, magma_int_t lddb,
-    float  beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magmablas_ctrmm(
-    char s, char u, char t,  char d,
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    cuFloatComplex       *dB, magma_int_t lddb );
-
-void magmablas_ctrsm(
-    char s, char u, char t, char d,
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    cuFloatComplex       *dB, magma_int_t lddb );
-
-
-  /*
-   * Wrappers for platform independence.
-   * These wrap CUBLAS or AMD OpenCL BLAS functions.
-   */
-
-// ========================================
-// copying vectors
-// set copies host to device
-// get copies device to host
-// Add the function, file, and line for error-reporting purposes.
-
-#define magma_csetvector(           n, hx_src, incx, dy_dst, incy ) \
-        magma_csetvector_internal(  n, hx_src, incx, dy_dst, incy, __func__, __FILE__, __LINE__ )
-
-#define magma_cgetvector(           n, dx_src, incx, hy_dst, incy ) \
-        magma_cgetvector_internal(  n, dx_src, incx, hy_dst, incy, __func__, __FILE__, __LINE__ )
-
-#define magma_csetvector_async(           n, hx_src, incx, dy_dst, incy, stream ) \
-        magma_csetvector_async_internal(  n, hx_src, incx, dy_dst, incy, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_cgetvector_async(           n, dx_src, incx, hy_dst, incy, stream ) \
-        magma_cgetvector_async_internal(  n, dx_src, incx, hy_dst, incy, stream, __func__, __FILE__, __LINE__ )
-
-void magma_csetvector_internal(
-    magma_int_t n,
-    const cuFloatComplex *hx_src, magma_int_t incx,
-    cuFloatComplex       *dy_dst, magma_int_t incy,
-    const char* func, const char* file, int line );
-
-void magma_cgetvector_internal(
-    magma_int_t n,
-    const cuFloatComplex *dx_src, magma_int_t incx,
-    cuFloatComplex       *hy_dst, magma_int_t incy,
-    const char* func, const char* file, int line );
-
-void magma_csetvector_async_internal(
-    magma_int_t n,
-    const cuFloatComplex *hx_src, magma_int_t incx,
-    cuFloatComplex       *dy_dst, magma_int_t incy,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_cgetvector_async_internal(
-    magma_int_t n,
-    const cuFloatComplex *dx_src, magma_int_t incx,
-    cuFloatComplex       *hy_dst, magma_int_t incy,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-
-// ========================================
-// copying sub-matrices (contiguous columns )
-// set  copies host to device
-// get  copies device to host
-// copy copies device to device
-// (with CUDA unified addressing, copy can be between same or different devices )
-// Add the function, file, and line for error-reporting purposes.
-
-#define magma_csetmatrix(           m, n, hA_src, lda, dB_dst, lddb ) \
-        magma_csetmatrix_internal(  m, n, hA_src, lda, dB_dst, lddb, __func__, __FILE__, __LINE__ )
-
-#define magma_cgetmatrix(           m, n, dA_src, ldda, hB_dst, ldb ) \
-        magma_cgetmatrix_internal(  m, n, dA_src, ldda, hB_dst, ldb, __func__, __FILE__, __LINE__ )
-
-#define magma_ccopymatrix(          m, n, dA_src, ldda, dB_dst, lddb ) \
-        magma_ccopymatrix_internal( m, n, dA_src, ldda, dB_dst, lddb, __func__, __FILE__, __LINE__ )
-
-#define magma_csetmatrix_async(           m, n, hA_src, lda, dB_dst, lddb, stream ) \
-        magma_csetmatrix_async_internal(  m, n, hA_src, lda, dB_dst, lddb, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_cgetmatrix_async(           m, n, dA_src, ldda, hB_dst, ldb, stream ) \
-        magma_cgetmatrix_async_internal(  m, n, dA_src, ldda, hB_dst, ldb, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_ccopymatrix_async(          m, n, dA_src, ldda, dB_dst, lddb, stream ) \
-        magma_ccopymatrix_async_internal( m, n, dA_src, ldda, dB_dst, lddb, stream, __func__, __FILE__, __LINE__ )
-
-void magma_csetmatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *hA_src, magma_int_t lda,
-    cuFloatComplex       *dB_dst, magma_int_t lddb,
-    const char* func, const char* file, int line );
-
-void magma_cgetmatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *dA_src, magma_int_t ldda,
-    cuFloatComplex       *hB_dst, magma_int_t ldb,
-    const char* func, const char* file, int line );
-
-void magma_ccopymatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *dA_src, magma_int_t ldda,
-    cuFloatComplex       *dB_dst, magma_int_t lddb,
-    const char* func, const char* file, int line );
-
-void magma_csetmatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *hA_src, magma_int_t lda,
-    cuFloatComplex       *dB_dst, magma_int_t lddb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_cgetmatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *dA_src, magma_int_t ldda,
-    cuFloatComplex       *hB_dst, magma_int_t ldb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_ccopymatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const cuFloatComplex *dA_src, magma_int_t ldda,
-    cuFloatComplex       *dB_dst, magma_int_t lddb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-
-// ========================================
-// Level 1 BLAS
-
-void magma_cswap(
-    magma_int_t n,
-    cuFloatComplex *dx, magma_int_t incx,
-    cuFloatComplex *dy, magma_int_t incy );
-
-magma_int_t magma_icamax(
-    magma_int_t n,
-    cuFloatComplex *dx, magma_int_t incx );
-
-// ========================================
-// Level 2 BLAS
-
-void magma_cgemv(
-    magma_trans_t transA,
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dx, magma_int_t incx,
-    cuFloatComplex beta,
-    cuFloatComplex       *dy, magma_int_t incy );
-
-void magma_chemv(
-    magma_uplo_t uplo,
-    magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dx, magma_int_t incx,
-    cuFloatComplex beta,
-    cuFloatComplex       *dy, magma_int_t incy );
-
-void magma_ctrsv(
-    magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t n,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    cuFloatComplex       *dx, magma_int_t incx );
-
-// ========================================
-// Level 3 BLAS
-
-void magma_cgemm(
-    magma_trans_t transA, magma_trans_t transB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dB, magma_int_t lddb,
-    cuFloatComplex beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magma_chemm(
-    magma_side_t side, magma_uplo_t uplo,
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dB, magma_int_t lddb,
-    cuFloatComplex beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magma_cherk(
-    magma_uplo_t uplo, magma_trans_t trans,
-    magma_int_t n, magma_int_t k,
-    float alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    float beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magma_cher2k(
-    magma_uplo_t uplo, magma_trans_t trans,
-    magma_int_t n, magma_int_t k,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    const cuFloatComplex *dB, magma_int_t lddb,
-    float beta,
-    cuFloatComplex       *dC, magma_int_t lddc );
-
-void magma_ctrmm(
-    magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    cuFloatComplex       *dB, magma_int_t lddb );
-
-void magma_ctrsm(
-    magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t m, magma_int_t n,
-    cuFloatComplex alpha,
-    const cuFloatComplex *dA, magma_int_t ldda,
-    cuFloatComplex       *dB, magma_int_t lddb );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_c
-
-#endif  // MAGMABLAS_C_H
diff --git a/src/include/magmablas_d.h b/src/include/magmablas_d.h
deleted file mode 100644
index f81e6f2..0000000
--- a/src/include/magmablas_d.h
+++ /dev/null
@@ -1,617 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated d Wed Nov 14 22:52:26 2012
- */
-
-#ifndef MAGMABLAS_D_H
-#define MAGMABLAS_D_H
-
-#define PRECISION_d
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-  /*
-   * Interface to clean
-   */
-double cpu_gpu_ddiff(
-    magma_int_t m, magma_int_t n,
-    const double *hA, magma_int_t lda,
-    const double *dA, magma_int_t ldda );
-
-// see also dlaset
-void dzero_32x32_block(
-    double *dA, magma_int_t ldda );
-
-void dzero_nbxnb_block(
-    magma_int_t nb, double *dA, magma_int_t ldda );
-
-// see also dlaswp
-void magmablas_dpermute_long2(
-    magma_int_t n, double *dAT, magma_int_t ldda,
-    magma_int_t *ipiv, magma_int_t nb, magma_int_t ind );
-
-void magmablas_dpermute_long3(
-    /*magma_int_t n,*/ double *dAT, magma_int_t ldda,
-    const magma_int_t *ipiv, magma_int_t nb, magma_int_t ind );
-
-  /*
-   * Transpose functions
-   */
-void magmablas_dinplace_transpose(
-    double *dA, magma_int_t ldda, magma_int_t n );
-
-void magmablas_dtranspose(
-    double       *odata, magma_int_t ldo,
-    const double *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n );
-
-void magmablas_dtranspose2(
-    double       *odata, magma_int_t ldo,
-    const double *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n );
-
-void magmablas_dtranspose2s(
-    double       *odata, magma_int_t ldo,
-    const double *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n,
-    cudaStream_t *stream );
-
-void magmablas_dgetmatrix_transpose(
-    magma_int_t m, magma_int_t n,
-    const double *dAT,   magma_int_t ldda,
-    double       *hA,    magma_int_t lda,
-    double       *dwork, magma_int_t lddwork, magma_int_t nb );
-
-void magmablas_dsetmatrix_transpose(
-    magma_int_t m, magma_int_t n,
-    const double *hA,    magma_int_t lda,
-    double       *dAT,   magma_int_t ldda,
-    double       *dwork, magma_int_t lddwork, magma_int_t nb );
-
-  /*
-   * Multi-GPU functions
-   */
-void magmablas_dgetmatrix_transpose_mgpu(
-    magma_int_t ngpu, cudaStream_t stream[][2],
-    double **dAT, magma_int_t ldda,
-    double  *hA,  magma_int_t lda,
-    double **dB,  magma_int_t lddb,
-    magma_int_t m, magma_int_t n, magma_int_t nb );
-
-void magmablas_dsetmatrix_transpose_mgpu(
-    magma_int_t ngpu, cudaStream_t stream[][2],
-    const double  *hA,  magma_int_t lda,
-    double       **dAT, magma_int_t ldda,
-    double       **dB,  magma_int_t lddb,
-    magma_int_t m, magma_int_t n, magma_int_t nb );
-
-void magmablas_dgetmatrix_1D_bcyclic(
-    magma_int_t m, magma_int_t n,
-    double *dA[], magma_int_t ldda,
-    double       *hA,   magma_int_t lda,
-    magma_int_t ngpu, magma_int_t nb );
-
-void magmablas_dsetmatrix_1D_bcyclic(
-    magma_int_t m, magma_int_t n,
-    const double *hA,   magma_int_t lda,
-    double       *dA[], magma_int_t ldda,
-    magma_int_t ngpu, magma_int_t nb );
-
-void magmablas_dsymm_1gpu_old(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    double alpha,
-    double *dA[], magma_int_t ldda,  magma_int_t offset,
-    double *dB[], magma_int_t lddb,
-    double beta,
-    double *dC[], magma_int_t lddc,
-    double *C,    magma_int_t ldc,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-void magmablas_dsymm_1gpu(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    double alpha,
-    double *dA[], magma_int_t ldda,  magma_int_t offset,
-    double *dB[], magma_int_t lddb,
-    double beta,
-    double *dC[], magma_int_t lddc,
-    double *C,    magma_int_t ldc,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-
-void magmablas_dsymm_mgpu(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    double alpha, double *dA[], magma_int_t ldda,  magma_int_t offset,
-                           double *dB[], magma_int_t lddb,
-    double beta,  double *dC[], magma_int_t lddc,
-                           double *dwork[],    magma_int_t lddwork,
-                           double *C,    magma_int_t ldc,
-                           double *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][20],magma_int_t nbevents );
-
-void magmablas_dsymm_mgpu_com(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    double alpha, double *dA[], magma_int_t ldda,  magma_int_t offset,
-                           double *dB[], magma_int_t lddb,
-    double beta,  double *dC[], magma_int_t lddc,
-                           double *dwork[],    magma_int_t lddwork,
-                           double *C,    magma_int_t ldc,
-                           double *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][20],magma_int_t nbevents, 
-                           magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2], magma_int_t nbcmplx );
-
-void magmablas_dsymm_mgpu_spec(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    double alpha, double *dA[], magma_int_t ldda,  magma_int_t offset,
-                           double *dB[], magma_int_t lddb,
-    double beta,  double *dC[], magma_int_t lddc,
-                           double *dwork[],    magma_int_t lddwork,
-                           double *C,    magma_int_t ldc,
-                           double *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][MagmaMaxGPUs*MagmaMaxGPUs],magma_int_t nbevents, 
-                           magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2], magma_int_t nbcmplx );
-
-
-
-void magmablas_dsyr2k_mgpu2(
-    char uplo, char trans, magma_int_t n, magma_int_t k,
-    double alpha,
-    double *dA[], magma_int_t ldda, magma_int_t aoff,
-    double *dB[], magma_int_t lddb, magma_int_t boff,
-    double beta,
-    double *dC[], magma_int_t lddc, magma_int_t offset,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-void magmablas_dsyr2k_mgpu_spec(
-    char uplo, char trans, magma_int_t n, magma_int_t k,
-    double alpha, double *dA[], magma_int_t lda, magma_int_t aoff,
-                           double *dB[], magma_int_t ldb, magma_int_t boff,
-    double beta,           double *dC[], magma_int_t ldc,  magma_int_t offset,
-    magma_int_t ngpu, magma_int_t nb, cudaStream_t streams[][20], magma_int_t nstream );
-
-  /*
-   * LAPACK auxiliary functions
-   */
-void magmablas_dgeadd(
-    magma_int_t m, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    double       *dB, magma_int_t lddb );
-
-void magmablas_dlacpy(
-    char uplo,
-    magma_int_t m, magma_int_t n,
-    const double *dA, magma_int_t ldda,
-    double       *dB, magma_int_t lddb );
-
-double magmablas_dlange(
-    char norm,
-    magma_int_t m, magma_int_t n,
-    const double *dA, magma_int_t ldda, double *dwork );
-
-double magmablas_dlansy(
-    char norm, char uplo,
-    magma_int_t n,
-    const double *dA, magma_int_t ldda, double *dwork );
-
-double magmablas_dlansy(
-    char norm, char uplo,
-    magma_int_t n,
-    const double *dA, magma_int_t ldda, double *dwork );
-
-void magmablas_dlascl(
-    char type, magma_int_t kl, magma_int_t ku,
-    double cfrom, double cto,
-    magma_int_t m, magma_int_t n,
-    double *dA, magma_int_t ldda, magma_int_t *info );
-
-void magmablas_dlaset(
-    char uplo, magma_int_t m, magma_int_t n,
-    double *dA, magma_int_t ldda );
-
-void magmablas_dlaset_identity(
-    magma_int_t m, magma_int_t n,
-    double *dA, magma_int_t ldda );
-
-void magmablas_dlaswp(
-    magma_int_t n,
-    double *dAT, magma_int_t ldda,
-    magma_int_t i1,  magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci );
-
-void magmablas_dlaswpx(
-    magma_int_t n,
-    double *dAT, magma_int_t ldx, magma_int_t ldy,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci );
-
-void magmablas_dlaswp2(
-    magma_int_t n,
-    double* dAT, magma_int_t ldda,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *d_ipiv );
-
-void magmablas_dsymmetrize(
-    char uplo, magma_int_t m, double *dA, magma_int_t ldda );
-
-void magmablas_dsymmetrize_tiles(
-    char uplo, magma_int_t m, double *dA, magma_int_t ldda,
-    magma_int_t ntile, magma_int_t mstride, magma_int_t nstride );
-
-  /*
-   * Level 1 BLAS
-   */
-void magmablas_dswap(
-    magma_int_t n,
-    double *dA, magma_int_t ldda,
-    double *dB, magma_int_t lddb );
-
-void magmablas_dswapblk(
-    char storev,
-    magma_int_t n,
-    double *dA, magma_int_t ldda,
-    double *dB, magma_int_t lddb,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci,
-    magma_int_t offset );
-
-void magmablas_dswapdblk(
-    magma_int_t n, magma_int_t nb,
-    double *dA, magma_int_t ldda, magma_int_t inca,
-    double *dB, magma_int_t lddb, magma_int_t incb );
-
-  /*
-   * Level 2 BLAS
-   */
-void magmablas_dgemv(
-    char t, magma_int_t m, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dx, magma_int_t incx,
-    double beta,
-    double       *dy, magma_int_t incy );
-
-#if defined(PRECISION_z ) || defined(PRECISION_c )
-magma_int_t magmablas_dsymv(
-    char u, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dx, magma_int_t incx,
-    double beta,
-    double       *dy, magma_int_t incy );
-#endif
-
-magma_int_t magmablas_dsymv(
-    char u, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dx, magma_int_t incx,
-    double beta,
-    double       *dy, magma_int_t incy );
-
-  /*
-   * Level 3 BLAS
-   */
-void magmablas_dgemm(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dB, magma_int_t lddb,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magmablas_dgemm_fermi80(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dB, magma_int_t lddb,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magmablas_dgemm_fermi64(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dB, magma_int_t lddb,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magmablas_dsymm(
-    char s, char u,
-    magma_int_t m, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dB, magma_int_t lddb,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magmablas_dsymm(
-    char s, char u,
-    magma_int_t m, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dB, magma_int_t lddb,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magmablas_dsyrk(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magmablas_dsyrk(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    double  alpha,
-    const double *dA, magma_int_t ldda,
-    double  beta,
-    double       *dC, magma_int_t lddc );
-
-void magmablas_dsyr2k(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dB, magma_int_t lddb,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magmablas_dsyr2k(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dB, magma_int_t lddb,
-    double  beta,
-    double       *dC, magma_int_t lddc );
-
-void magmablas_dtrmm(
-    char s, char u, char t,  char d,
-    magma_int_t m, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    double       *dB, magma_int_t lddb );
-
-void magmablas_dtrsm(
-    char s, char u, char t, char d,
-    magma_int_t m, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    double       *dB, magma_int_t lddb );
-
-
-  /*
-   * Wrappers for platform independence.
-   * These wrap CUBLAS or AMD OpenCL BLAS functions.
-   */
-
-// ========================================
-// copying vectors
-// set copies host to device
-// get copies device to host
-// Add the function, file, and line for error-reporting purposes.
-
-#define magma_dsetvector(           n, hx_src, incx, dy_dst, incy ) \
-        magma_dsetvector_internal(  n, hx_src, incx, dy_dst, incy, __func__, __FILE__, __LINE__ )
-
-#define magma_dgetvector(           n, dx_src, incx, hy_dst, incy ) \
-        magma_dgetvector_internal(  n, dx_src, incx, hy_dst, incy, __func__, __FILE__, __LINE__ )
-
-#define magma_dsetvector_async(           n, hx_src, incx, dy_dst, incy, stream ) \
-        magma_dsetvector_async_internal(  n, hx_src, incx, dy_dst, incy, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_dgetvector_async(           n, dx_src, incx, hy_dst, incy, stream ) \
-        magma_dgetvector_async_internal(  n, dx_src, incx, hy_dst, incy, stream, __func__, __FILE__, __LINE__ )
-
-void magma_dsetvector_internal(
-    magma_int_t n,
-    const double *hx_src, magma_int_t incx,
-    double       *dy_dst, magma_int_t incy,
-    const char* func, const char* file, int line );
-
-void magma_dgetvector_internal(
-    magma_int_t n,
-    const double *dx_src, magma_int_t incx,
-    double       *hy_dst, magma_int_t incy,
-    const char* func, const char* file, int line );
-
-void magma_dsetvector_async_internal(
-    magma_int_t n,
-    const double *hx_src, magma_int_t incx,
-    double       *dy_dst, magma_int_t incy,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_dgetvector_async_internal(
-    magma_int_t n,
-    const double *dx_src, magma_int_t incx,
-    double       *hy_dst, magma_int_t incy,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-
-// ========================================
-// copying sub-matrices (contiguous columns )
-// set  copies host to device
-// get  copies device to host
-// copy copies device to device
-// (with CUDA unified addressing, copy can be between same or different devices )
-// Add the function, file, and line for error-reporting purposes.
-
-#define magma_dsetmatrix(           m, n, hA_src, lda, dB_dst, lddb ) \
-        magma_dsetmatrix_internal(  m, n, hA_src, lda, dB_dst, lddb, __func__, __FILE__, __LINE__ )
-
-#define magma_dgetmatrix(           m, n, dA_src, ldda, hB_dst, ldb ) \
-        magma_dgetmatrix_internal(  m, n, dA_src, ldda, hB_dst, ldb, __func__, __FILE__, __LINE__ )
-
-#define magma_dcopymatrix(          m, n, dA_src, ldda, dB_dst, lddb ) \
-        magma_dcopymatrix_internal( m, n, dA_src, ldda, dB_dst, lddb, __func__, __FILE__, __LINE__ )
-
-#define magma_dsetmatrix_async(           m, n, hA_src, lda, dB_dst, lddb, stream ) \
-        magma_dsetmatrix_async_internal(  m, n, hA_src, lda, dB_dst, lddb, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_dgetmatrix_async(           m, n, dA_src, ldda, hB_dst, ldb, stream ) \
-        magma_dgetmatrix_async_internal(  m, n, dA_src, ldda, hB_dst, ldb, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_dcopymatrix_async(          m, n, dA_src, ldda, dB_dst, lddb, stream ) \
-        magma_dcopymatrix_async_internal( m, n, dA_src, ldda, dB_dst, lddb, stream, __func__, __FILE__, __LINE__ )
-
-void magma_dsetmatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const double *hA_src, magma_int_t lda,
-    double       *dB_dst, magma_int_t lddb,
-    const char* func, const char* file, int line );
-
-void magma_dgetmatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const double *dA_src, magma_int_t ldda,
-    double       *hB_dst, magma_int_t ldb,
-    const char* func, const char* file, int line );
-
-void magma_dcopymatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const double *dA_src, magma_int_t ldda,
-    double       *dB_dst, magma_int_t lddb,
-    const char* func, const char* file, int line );
-
-void magma_dsetmatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const double *hA_src, magma_int_t lda,
-    double       *dB_dst, magma_int_t lddb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_dgetmatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const double *dA_src, magma_int_t ldda,
-    double       *hB_dst, magma_int_t ldb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_dcopymatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const double *dA_src, magma_int_t ldda,
-    double       *dB_dst, magma_int_t lddb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-
-// ========================================
-// Level 1 BLAS
-
-void magma_dswap(
-    magma_int_t n,
-    double *dx, magma_int_t incx,
-    double *dy, magma_int_t incy );
-
-magma_int_t magma_idamax(
-    magma_int_t n,
-    double *dx, magma_int_t incx );
-
-// ========================================
-// Level 2 BLAS
-
-void magma_dgemv(
-    magma_trans_t transA,
-    magma_int_t m, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dx, magma_int_t incx,
-    double beta,
-    double       *dy, magma_int_t incy );
-
-void magma_dsymv(
-    magma_uplo_t uplo,
-    magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dx, magma_int_t incx,
-    double beta,
-    double       *dy, magma_int_t incy );
-
-void magma_dtrsv(
-    magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t n,
-    const double *dA, magma_int_t ldda,
-    double       *dx, magma_int_t incx );
-
-// ========================================
-// Level 3 BLAS
-
-void magma_dgemm(
-    magma_trans_t transA, magma_trans_t transB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dB, magma_int_t lddb,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magma_dsymm(
-    magma_side_t side, magma_uplo_t uplo,
-    magma_int_t m, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dB, magma_int_t lddb,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magma_dsyrk(
-    magma_uplo_t uplo, magma_trans_t trans,
-    magma_int_t n, magma_int_t k,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magma_dsyr2k(
-    magma_uplo_t uplo, magma_trans_t trans,
-    magma_int_t n, magma_int_t k,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    const double *dB, magma_int_t lddb,
-    double beta,
-    double       *dC, magma_int_t lddc );
-
-void magma_dtrmm(
-    magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t m, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    double       *dB, magma_int_t lddb );
-
-void magma_dtrsm(
-    magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t m, magma_int_t n,
-    double alpha,
-    const double *dA, magma_int_t ldda,
-    double       *dB, magma_int_t lddb );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_d
-
-#endif  // MAGMABLAS_D_H
diff --git a/src/include/magmablas_ds.h b/src/include/magmablas_ds.h
deleted file mode 100644
index d96e54a..0000000
--- a/src/include/magmablas_ds.h
+++ /dev/null
@@ -1,55 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated ds Wed Nov 14 22:52:26 2012
- */
-
-#ifndef MAGMABLAS_DS_H
-#define MAGMABLAS_DS_H
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-  /* Mixed precision */
-void magmablas_dsaxpycp(
-    float *R, double *X,
-    magma_int_t m, double *B, double *W );
-
-void magmablas_daxpycp(
-    double *R, double *X,
-    magma_int_t m, double *B );
-
-void magmablas_dslaswp(
-    magma_int_t n,
-    double *A, magma_int_t lda,
-    float *SA, magma_int_t m,
-    const magma_int_t *ipiv, magma_int_t incx );
-
-void magmablas_dlag2s(
-    magma_int_t m, magma_int_t n,
-    const double *A,  magma_int_t lda,
-    float        *SA, magma_int_t ldsa,
-    magma_int_t *info );
-
-void magmablas_slag2d(
-    magma_int_t m, magma_int_t n, 
-    const float  *SA, magma_int_t ldsa, 
-    double       *A,  magma_int_t lda, 
-    magma_int_t *info );
-
-void magmablas_dlat2s(
-    char uplo, magma_int_t n, 
-    const double *A,  magma_int_t lda, 
-    float        *SA, magma_int_t ldsa, 
-    magma_int_t *info );
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // MAGMABLAS_DS_H
diff --git a/src/include/magmablas_s.h b/src/include/magmablas_s.h
deleted file mode 100644
index 47bc5b6..0000000
--- a/src/include/magmablas_s.h
+++ /dev/null
@@ -1,617 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @generated s Wed Nov 14 22:52:26 2012
- */
-
-#ifndef MAGMABLAS_S_H
-#define MAGMABLAS_S_H
-
-#define PRECISION_s
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-  /*
-   * Interface to clean
-   */
-float cpu_gpu_sdiff(
-    magma_int_t m, magma_int_t n,
-    const float *hA, magma_int_t lda,
-    const float *dA, magma_int_t ldda );
-
-// see also slaset
-void szero_32x32_block(
-    float *dA, magma_int_t ldda );
-
-void szero_nbxnb_block(
-    magma_int_t nb, float *dA, magma_int_t ldda );
-
-// see also slaswp
-void magmablas_spermute_long2(
-    magma_int_t n, float *dAT, magma_int_t ldda,
-    magma_int_t *ipiv, magma_int_t nb, magma_int_t ind );
-
-void magmablas_spermute_long3(
-    /*magma_int_t n,*/ float *dAT, magma_int_t ldda,
-    const magma_int_t *ipiv, magma_int_t nb, magma_int_t ind );
-
-  /*
-   * Transpose functions
-   */
-void magmablas_sinplace_transpose(
-    float *dA, magma_int_t ldda, magma_int_t n );
-
-void magmablas_stranspose(
-    float       *odata, magma_int_t ldo,
-    const float *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n );
-
-void magmablas_stranspose2(
-    float       *odata, magma_int_t ldo,
-    const float *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n );
-
-void magmablas_stranspose2s(
-    float       *odata, magma_int_t ldo,
-    const float *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n,
-    cudaStream_t *stream );
-
-void magmablas_sgetmatrix_transpose(
-    magma_int_t m, magma_int_t n,
-    const float *dAT,   magma_int_t ldda,
-    float       *hA,    magma_int_t lda,
-    float       *dwork, magma_int_t lddwork, magma_int_t nb );
-
-void magmablas_ssetmatrix_transpose(
-    magma_int_t m, magma_int_t n,
-    const float *hA,    magma_int_t lda,
-    float       *dAT,   magma_int_t ldda,
-    float       *dwork, magma_int_t lddwork, magma_int_t nb );
-
-  /*
-   * Multi-GPU functions
-   */
-void magmablas_sgetmatrix_transpose_mgpu(
-    magma_int_t ngpu, cudaStream_t stream[][2],
-    float **dAT, magma_int_t ldda,
-    float  *hA,  magma_int_t lda,
-    float **dB,  magma_int_t lddb,
-    magma_int_t m, magma_int_t n, magma_int_t nb );
-
-void magmablas_ssetmatrix_transpose_mgpu(
-    magma_int_t ngpu, cudaStream_t stream[][2],
-    const float  *hA,  magma_int_t lda,
-    float       **dAT, magma_int_t ldda,
-    float       **dB,  magma_int_t lddb,
-    magma_int_t m, magma_int_t n, magma_int_t nb );
-
-void magmablas_sgetmatrix_1D_bcyclic(
-    magma_int_t m, magma_int_t n,
-    float *dA[], magma_int_t ldda,
-    float       *hA,   magma_int_t lda,
-    magma_int_t ngpu, magma_int_t nb );
-
-void magmablas_ssetmatrix_1D_bcyclic(
-    magma_int_t m, magma_int_t n,
-    const float *hA,   magma_int_t lda,
-    float       *dA[], magma_int_t ldda,
-    magma_int_t ngpu, magma_int_t nb );
-
-void magmablas_ssymm_1gpu_old(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    float alpha,
-    float *dA[], magma_int_t ldda,  magma_int_t offset,
-    float *dB[], magma_int_t lddb,
-    float beta,
-    float *dC[], magma_int_t lddc,
-    float *C,    magma_int_t ldc,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-void magmablas_ssymm_1gpu(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    float alpha,
-    float *dA[], magma_int_t ldda,  magma_int_t offset,
-    float *dB[], magma_int_t lddb,
-    float beta,
-    float *dC[], magma_int_t lddc,
-    float *C,    magma_int_t ldc,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-
-void magmablas_ssymm_mgpu(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    float alpha, float *dA[], magma_int_t ldda,  magma_int_t offset,
-                           float *dB[], magma_int_t lddb,
-    float beta,  float *dC[], magma_int_t lddc,
-                           float *dwork[],    magma_int_t lddwork,
-                           float *C,    magma_int_t ldc,
-                           float *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][20],magma_int_t nbevents );
-
-void magmablas_ssymm_mgpu_com(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    float alpha, float *dA[], magma_int_t ldda,  magma_int_t offset,
-                           float *dB[], magma_int_t lddb,
-    float beta,  float *dC[], magma_int_t lddc,
-                           float *dwork[],    magma_int_t lddwork,
-                           float *C,    magma_int_t ldc,
-                           float *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][20],magma_int_t nbevents, 
-                           magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2], magma_int_t nbcmplx );
-
-void magmablas_ssymm_mgpu_spec(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    float alpha, float *dA[], magma_int_t ldda,  magma_int_t offset,
-                           float *dB[], magma_int_t lddb,
-    float beta,  float *dC[], magma_int_t lddc,
-                           float *dwork[],    magma_int_t lddwork,
-                           float *C,    magma_int_t ldc,
-                           float *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][MagmaMaxGPUs*MagmaMaxGPUs],magma_int_t nbevents, 
-                           magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2], magma_int_t nbcmplx );
-
-
-
-void magmablas_ssyr2k_mgpu2(
-    char uplo, char trans, magma_int_t n, magma_int_t k,
-    float alpha,
-    float *dA[], magma_int_t ldda, magma_int_t aoff,
-    float *dB[], magma_int_t lddb, magma_int_t boff,
-    float beta,
-    float *dC[], magma_int_t lddc, magma_int_t offset,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-void magmablas_ssyr2k_mgpu_spec(
-    char uplo, char trans, magma_int_t n, magma_int_t k,
-    float alpha, float *dA[], magma_int_t lda, magma_int_t aoff,
-                           float *dB[], magma_int_t ldb, magma_int_t boff,
-    float beta,           float *dC[], magma_int_t ldc,  magma_int_t offset,
-    magma_int_t ngpu, magma_int_t nb, cudaStream_t streams[][20], magma_int_t nstream );
-
-  /*
-   * LAPACK auxiliary functions
-   */
-void magmablas_sgeadd(
-    magma_int_t m, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    float       *dB, magma_int_t lddb );
-
-void magmablas_slacpy(
-    char uplo,
-    magma_int_t m, magma_int_t n,
-    const float *dA, magma_int_t ldda,
-    float       *dB, magma_int_t lddb );
-
-float magmablas_slange(
-    char norm,
-    magma_int_t m, magma_int_t n,
-    const float *dA, magma_int_t ldda, float *dwork );
-
-float magmablas_slansy(
-    char norm, char uplo,
-    magma_int_t n,
-    const float *dA, magma_int_t ldda, float *dwork );
-
-float magmablas_slansy(
-    char norm, char uplo,
-    magma_int_t n,
-    const float *dA, magma_int_t ldda, float *dwork );
-
-void magmablas_slascl(
-    char type, magma_int_t kl, magma_int_t ku,
-    float cfrom, float cto,
-    magma_int_t m, magma_int_t n,
-    float *dA, magma_int_t ldda, magma_int_t *info );
-
-void magmablas_slaset(
-    char uplo, magma_int_t m, magma_int_t n,
-    float *dA, magma_int_t ldda );
-
-void magmablas_slaset_identity(
-    magma_int_t m, magma_int_t n,
-    float *dA, magma_int_t ldda );
-
-void magmablas_slaswp(
-    magma_int_t n,
-    float *dAT, magma_int_t ldda,
-    magma_int_t i1,  magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci );
-
-void magmablas_slaswpx(
-    magma_int_t n,
-    float *dAT, magma_int_t ldx, magma_int_t ldy,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci );
-
-void magmablas_slaswp2(
-    magma_int_t n,
-    float* dAT, magma_int_t ldda,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *d_ipiv );
-
-void magmablas_ssymmetrize(
-    char uplo, magma_int_t m, float *dA, magma_int_t ldda );
-
-void magmablas_ssymmetrize_tiles(
-    char uplo, magma_int_t m, float *dA, magma_int_t ldda,
-    magma_int_t ntile, magma_int_t mstride, magma_int_t nstride );
-
-  /*
-   * Level 1 BLAS
-   */
-void magmablas_sswap(
-    magma_int_t n,
-    float *dA, magma_int_t ldda,
-    float *dB, magma_int_t lddb );
-
-void magmablas_sswapblk(
-    char storev,
-    magma_int_t n,
-    float *dA, magma_int_t ldda,
-    float *dB, magma_int_t lddb,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci,
-    magma_int_t offset );
-
-void magmablas_sswapdblk(
-    magma_int_t n, magma_int_t nb,
-    float *dA, magma_int_t ldda, magma_int_t inca,
-    float *dB, magma_int_t lddb, magma_int_t incb );
-
-  /*
-   * Level 2 BLAS
-   */
-void magmablas_sgemv(
-    char t, magma_int_t m, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dx, magma_int_t incx,
-    float beta,
-    float       *dy, magma_int_t incy );
-
-#if defined(PRECISION_z ) || defined(PRECISION_c )
-magma_int_t magmablas_ssymv(
-    char u, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dx, magma_int_t incx,
-    float beta,
-    float       *dy, magma_int_t incy );
-#endif
-
-magma_int_t magmablas_ssymv(
-    char u, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dx, magma_int_t incx,
-    float beta,
-    float       *dy, magma_int_t incy );
-
-  /*
-   * Level 3 BLAS
-   */
-void magmablas_sgemm(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dB, magma_int_t lddb,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magmablas_sgemm_fermi80(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dB, magma_int_t lddb,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magmablas_sgemm_fermi64(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dB, magma_int_t lddb,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magmablas_ssymm(
-    char s, char u,
-    magma_int_t m, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dB, magma_int_t lddb,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magmablas_ssymm(
-    char s, char u,
-    magma_int_t m, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dB, magma_int_t lddb,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magmablas_ssyrk(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magmablas_ssyrk(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    float  alpha,
-    const float *dA, magma_int_t ldda,
-    float  beta,
-    float       *dC, magma_int_t lddc );
-
-void magmablas_ssyr2k(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dB, magma_int_t lddb,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magmablas_ssyr2k(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dB, magma_int_t lddb,
-    float  beta,
-    float       *dC, magma_int_t lddc );
-
-void magmablas_strmm(
-    char s, char u, char t,  char d,
-    magma_int_t m, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    float       *dB, magma_int_t lddb );
-
-void magmablas_strsm(
-    char s, char u, char t, char d,
-    magma_int_t m, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    float       *dB, magma_int_t lddb );
-
-
-  /*
-   * Wrappers for platform independence.
-   * These wrap CUBLAS or AMD OpenCL BLAS functions.
-   */
-
-// ========================================
-// copying vectors
-// set copies host to device
-// get copies device to host
-// Add the function, file, and line for error-reporting purposes.
-
-#define magma_ssetvector(           n, hx_src, incx, dy_dst, incy ) \
-        magma_ssetvector_internal(  n, hx_src, incx, dy_dst, incy, __func__, __FILE__, __LINE__ )
-
-#define magma_sgetvector(           n, dx_src, incx, hy_dst, incy ) \
-        magma_sgetvector_internal(  n, dx_src, incx, hy_dst, incy, __func__, __FILE__, __LINE__ )
-
-#define magma_ssetvector_async(           n, hx_src, incx, dy_dst, incy, stream ) \
-        magma_ssetvector_async_internal(  n, hx_src, incx, dy_dst, incy, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_sgetvector_async(           n, dx_src, incx, hy_dst, incy, stream ) \
-        magma_sgetvector_async_internal(  n, dx_src, incx, hy_dst, incy, stream, __func__, __FILE__, __LINE__ )
-
-void magma_ssetvector_internal(
-    magma_int_t n,
-    const float *hx_src, magma_int_t incx,
-    float       *dy_dst, magma_int_t incy,
-    const char* func, const char* file, int line );
-
-void magma_sgetvector_internal(
-    magma_int_t n,
-    const float *dx_src, magma_int_t incx,
-    float       *hy_dst, magma_int_t incy,
-    const char* func, const char* file, int line );
-
-void magma_ssetvector_async_internal(
-    magma_int_t n,
-    const float *hx_src, magma_int_t incx,
-    float       *dy_dst, magma_int_t incy,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_sgetvector_async_internal(
-    magma_int_t n,
-    const float *dx_src, magma_int_t incx,
-    float       *hy_dst, magma_int_t incy,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-
-// ========================================
-// copying sub-matrices (contiguous columns )
-// set  copies host to device
-// get  copies device to host
-// copy copies device to device
-// (with CUDA unified addressing, copy can be between same or different devices )
-// Add the function, file, and line for error-reporting purposes.
-
-#define magma_ssetmatrix(           m, n, hA_src, lda, dB_dst, lddb ) \
-        magma_ssetmatrix_internal(  m, n, hA_src, lda, dB_dst, lddb, __func__, __FILE__, __LINE__ )
-
-#define magma_sgetmatrix(           m, n, dA_src, ldda, hB_dst, ldb ) \
-        magma_sgetmatrix_internal(  m, n, dA_src, ldda, hB_dst, ldb, __func__, __FILE__, __LINE__ )
-
-#define magma_scopymatrix(          m, n, dA_src, ldda, dB_dst, lddb ) \
-        magma_scopymatrix_internal( m, n, dA_src, ldda, dB_dst, lddb, __func__, __FILE__, __LINE__ )
-
-#define magma_ssetmatrix_async(           m, n, hA_src, lda, dB_dst, lddb, stream ) \
-        magma_ssetmatrix_async_internal(  m, n, hA_src, lda, dB_dst, lddb, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_sgetmatrix_async(           m, n, dA_src, ldda, hB_dst, ldb, stream ) \
-        magma_sgetmatrix_async_internal(  m, n, dA_src, ldda, hB_dst, ldb, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_scopymatrix_async(          m, n, dA_src, ldda, dB_dst, lddb, stream ) \
-        magma_scopymatrix_async_internal( m, n, dA_src, ldda, dB_dst, lddb, stream, __func__, __FILE__, __LINE__ )
-
-void magma_ssetmatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const float *hA_src, magma_int_t lda,
-    float       *dB_dst, magma_int_t lddb,
-    const char* func, const char* file, int line );
-
-void magma_sgetmatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const float *dA_src, magma_int_t ldda,
-    float       *hB_dst, magma_int_t ldb,
-    const char* func, const char* file, int line );
-
-void magma_scopymatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const float *dA_src, magma_int_t ldda,
-    float       *dB_dst, magma_int_t lddb,
-    const char* func, const char* file, int line );
-
-void magma_ssetmatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const float *hA_src, magma_int_t lda,
-    float       *dB_dst, magma_int_t lddb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_sgetmatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const float *dA_src, magma_int_t ldda,
-    float       *hB_dst, magma_int_t ldb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_scopymatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const float *dA_src, magma_int_t ldda,
-    float       *dB_dst, magma_int_t lddb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-
-// ========================================
-// Level 1 BLAS
-
-void magma_sswap(
-    magma_int_t n,
-    float *dx, magma_int_t incx,
-    float *dy, magma_int_t incy );
-
-magma_int_t magma_isamax(
-    magma_int_t n,
-    float *dx, magma_int_t incx );
-
-// ========================================
-// Level 2 BLAS
-
-void magma_sgemv(
-    magma_trans_t transA,
-    magma_int_t m, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dx, magma_int_t incx,
-    float beta,
-    float       *dy, magma_int_t incy );
-
-void magma_ssymv(
-    magma_uplo_t uplo,
-    magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dx, magma_int_t incx,
-    float beta,
-    float       *dy, magma_int_t incy );
-
-void magma_strsv(
-    magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t n,
-    const float *dA, magma_int_t ldda,
-    float       *dx, magma_int_t incx );
-
-// ========================================
-// Level 3 BLAS
-
-void magma_sgemm(
-    magma_trans_t transA, magma_trans_t transB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dB, magma_int_t lddb,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magma_ssymm(
-    magma_side_t side, magma_uplo_t uplo,
-    magma_int_t m, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dB, magma_int_t lddb,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magma_ssyrk(
-    magma_uplo_t uplo, magma_trans_t trans,
-    magma_int_t n, magma_int_t k,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magma_ssyr2k(
-    magma_uplo_t uplo, magma_trans_t trans,
-    magma_int_t n, magma_int_t k,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    const float *dB, magma_int_t lddb,
-    float beta,
-    float       *dC, magma_int_t lddc );
-
-void magma_strmm(
-    magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t m, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    float       *dB, magma_int_t lddb );
-
-void magma_strsm(
-    magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t m, magma_int_t n,
-    float alpha,
-    const float *dA, magma_int_t ldda,
-    float       *dB, magma_int_t lddb );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_s
-
-#endif  // MAGMABLAS_S_H
diff --git a/src/include/magmablas_z.h b/src/include/magmablas_z.h
deleted file mode 100644
index 08bbb91..0000000
--- a/src/include/magmablas_z.h
+++ /dev/null
@@ -1,617 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @precisions normal z -> s d c
- */
-
-#ifndef MAGMABLAS_Z_H
-#define MAGMABLAS_Z_H
-
-#define PRECISION_z
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-  /*
-   * Interface to clean
-   */
-double cpu_gpu_zdiff(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *hA, magma_int_t lda,
-    const cuDoubleComplex *dA, magma_int_t ldda );
-
-// see also zlaset
-void zzero_32x32_block(
-    cuDoubleComplex *dA, magma_int_t ldda );
-
-void zzero_nbxnb_block(
-    magma_int_t nb, cuDoubleComplex *dA, magma_int_t ldda );
-
-// see also zlaswp
-void magmablas_zpermute_long2(
-    magma_int_t n, cuDoubleComplex *dAT, magma_int_t ldda,
-    magma_int_t *ipiv, magma_int_t nb, magma_int_t ind );
-
-void magmablas_zpermute_long3(
-    /*magma_int_t n,*/ cuDoubleComplex *dAT, magma_int_t ldda,
-    const magma_int_t *ipiv, magma_int_t nb, magma_int_t ind );
-
-  /*
-   * Transpose functions
-   */
-void magmablas_zinplace_transpose(
-    cuDoubleComplex *dA, magma_int_t ldda, magma_int_t n );
-
-void magmablas_ztranspose(
-    cuDoubleComplex       *odata, magma_int_t ldo,
-    const cuDoubleComplex *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n );
-
-void magmablas_ztranspose2(
-    cuDoubleComplex       *odata, magma_int_t ldo,
-    const cuDoubleComplex *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n );
-
-void magmablas_ztranspose2s(
-    cuDoubleComplex       *odata, magma_int_t ldo,
-    const cuDoubleComplex *idata, magma_int_t ldi,
-    magma_int_t m, magma_int_t n,
-    cudaStream_t *stream );
-
-void magmablas_zgetmatrix_transpose(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *dAT,   magma_int_t ldda,
-    cuDoubleComplex       *hA,    magma_int_t lda,
-    cuDoubleComplex       *dwork, magma_int_t lddwork, magma_int_t nb );
-
-void magmablas_zsetmatrix_transpose(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *hA,    magma_int_t lda,
-    cuDoubleComplex       *dAT,   magma_int_t ldda,
-    cuDoubleComplex       *dwork, magma_int_t lddwork, magma_int_t nb );
-
-  /*
-   * Multi-GPU functions
-   */
-void magmablas_zgetmatrix_transpose_mgpu(
-    magma_int_t ngpu, cudaStream_t stream[][2],
-    cuDoubleComplex **dAT, magma_int_t ldda,
-    cuDoubleComplex  *hA,  magma_int_t lda,
-    cuDoubleComplex **dB,  magma_int_t lddb,
-    magma_int_t m, magma_int_t n, magma_int_t nb );
-
-void magmablas_zsetmatrix_transpose_mgpu(
-    magma_int_t ngpu, cudaStream_t stream[][2],
-    const cuDoubleComplex  *hA,  magma_int_t lda,
-    cuDoubleComplex       **dAT, magma_int_t ldda,
-    cuDoubleComplex       **dB,  magma_int_t lddb,
-    magma_int_t m, magma_int_t n, magma_int_t nb );
-
-void magmablas_zgetmatrix_1D_bcyclic(
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex *dA[], magma_int_t ldda,
-    cuDoubleComplex       *hA,   magma_int_t lda,
-    magma_int_t ngpu, magma_int_t nb );
-
-void magmablas_zsetmatrix_1D_bcyclic(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *hA,   magma_int_t lda,
-    cuDoubleComplex       *dA[], magma_int_t ldda,
-    magma_int_t ngpu, magma_int_t nb );
-
-void magmablas_zhemm_1gpu_old(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    cuDoubleComplex *dA[], magma_int_t ldda,  magma_int_t offset,
-    cuDoubleComplex *dB[], magma_int_t lddb,
-    cuDoubleComplex beta,
-    cuDoubleComplex *dC[], magma_int_t lddc,
-    cuDoubleComplex *C,    magma_int_t ldc,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-void magmablas_zhemm_1gpu(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    cuDoubleComplex *dA[], magma_int_t ldda,  magma_int_t offset,
-    cuDoubleComplex *dB[], magma_int_t lddb,
-    cuDoubleComplex beta,
-    cuDoubleComplex *dC[], magma_int_t lddc,
-    cuDoubleComplex *C,    magma_int_t ldc,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-
-void magmablas_zhemm_mgpu(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha, cuDoubleComplex *dA[], magma_int_t ldda,  magma_int_t offset,
-                           cuDoubleComplex *dB[], magma_int_t lddb,
-    cuDoubleComplex beta,  cuDoubleComplex *dC[], magma_int_t lddc,
-                           cuDoubleComplex *dwork[],    magma_int_t lddwork,
-                           cuDoubleComplex *C,    magma_int_t ldc,
-                           cuDoubleComplex *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][20],magma_int_t nbevents );
-
-void magmablas_zhemm_mgpu_com(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha, cuDoubleComplex *dA[], magma_int_t ldda,  magma_int_t offset,
-                           cuDoubleComplex *dB[], magma_int_t lddb,
-    cuDoubleComplex beta,  cuDoubleComplex *dC[], magma_int_t lddc,
-                           cuDoubleComplex *dwork[],    magma_int_t lddwork,
-                           cuDoubleComplex *C,    magma_int_t ldc,
-                           cuDoubleComplex *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][20],magma_int_t nbevents, 
-                           magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2], magma_int_t nbcmplx );
-
-void magmablas_zhemm_mgpu_spec(
-    char side, char uplo, magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha, cuDoubleComplex *dA[], magma_int_t ldda,  magma_int_t offset,
-                           cuDoubleComplex *dB[], magma_int_t lddb,
-    cuDoubleComplex beta,  cuDoubleComplex *dC[], magma_int_t lddc,
-                           cuDoubleComplex *dwork[],    magma_int_t lddwork,
-                           cuDoubleComplex *C,    magma_int_t ldc,
-                           cuDoubleComplex *work[], magma_int_t ldwork,
-                           magma_int_t ngpu, magma_int_t nb, 
-                           cudaStream_t streams[][20], magma_int_t nstream, 
-                           cudaEvent_t redevents[][MagmaMaxGPUs*MagmaMaxGPUs],magma_int_t nbevents, 
-                           magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2], magma_int_t nbcmplx );
-
-
-
-void magmablas_zher2k_mgpu2(
-    char uplo, char trans, magma_int_t n, magma_int_t k,
-    cuDoubleComplex alpha,
-    cuDoubleComplex *dA[], magma_int_t ldda, magma_int_t aoff,
-    cuDoubleComplex *dB[], magma_int_t lddb, magma_int_t boff,
-    double beta,
-    cuDoubleComplex *dC[], magma_int_t lddc, magma_int_t offset,
-    magma_int_t ngpu, magma_int_t nb,
-    cudaStream_t streams[][20], magma_int_t nstream );
-
-void magmablas_zher2k_mgpu_spec(
-    char uplo, char trans, magma_int_t n, magma_int_t k,
-    cuDoubleComplex alpha, cuDoubleComplex *dA[], magma_int_t lda, magma_int_t aoff,
-                           cuDoubleComplex *dB[], magma_int_t ldb, magma_int_t boff,
-    double beta,           cuDoubleComplex *dC[], magma_int_t ldc,  magma_int_t offset,
-    magma_int_t ngpu, magma_int_t nb, cudaStream_t streams[][20], magma_int_t nstream );
-
-  /*
-   * LAPACK auxiliary functions
-   */
-void magmablas_zgeadd(
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    cuDoubleComplex       *dB, magma_int_t lddb );
-
-void magmablas_zlacpy(
-    char uplo,
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    cuDoubleComplex       *dB, magma_int_t lddb );
-
-double magmablas_zlange(
-    char norm,
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *dA, magma_int_t ldda, double *dwork );
-
-double magmablas_zlanhe(
-    char norm, char uplo,
-    magma_int_t n,
-    const cuDoubleComplex *dA, magma_int_t ldda, double *dwork );
-
-double magmablas_zlansy(
-    char norm, char uplo,
-    magma_int_t n,
-    const cuDoubleComplex *dA, magma_int_t ldda, double *dwork );
-
-void magmablas_zlascl(
-    char type, magma_int_t kl, magma_int_t ku,
-    double cfrom, double cto,
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex *dA, magma_int_t ldda, magma_int_t *info );
-
-void magmablas_zlaset(
-    char uplo, magma_int_t m, magma_int_t n,
-    cuDoubleComplex *dA, magma_int_t ldda );
-
-void magmablas_zlaset_identity(
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex *dA, magma_int_t ldda );
-
-void magmablas_zlaswp(
-    magma_int_t n,
-    cuDoubleComplex *dAT, magma_int_t ldda,
-    magma_int_t i1,  magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci );
-
-void magmablas_zlaswpx(
-    magma_int_t n,
-    cuDoubleComplex *dAT, magma_int_t ldx, magma_int_t ldy,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci );
-
-void magmablas_zlaswp2(
-    magma_int_t n,
-    cuDoubleComplex* dAT, magma_int_t ldda,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *d_ipiv );
-
-void magmablas_zsymmetrize(
-    char uplo, magma_int_t m, cuDoubleComplex *dA, magma_int_t ldda );
-
-void magmablas_zsymmetrize_tiles(
-    char uplo, magma_int_t m, cuDoubleComplex *dA, magma_int_t ldda,
-    magma_int_t ntile, magma_int_t mstride, magma_int_t nstride );
-
-  /*
-   * Level 1 BLAS
-   */
-void magmablas_zswap(
-    magma_int_t n,
-    cuDoubleComplex *dA, magma_int_t ldda,
-    cuDoubleComplex *dB, magma_int_t lddb );
-
-void magmablas_zswapblk(
-    char storev,
-    magma_int_t n,
-    cuDoubleComplex *dA, magma_int_t ldda,
-    cuDoubleComplex *dB, magma_int_t lddb,
-    magma_int_t i1, magma_int_t i2,
-    const magma_int_t *ipiv, magma_int_t inci,
-    magma_int_t offset );
-
-void magmablas_zswapdblk(
-    magma_int_t n, magma_int_t nb,
-    cuDoubleComplex *dA, magma_int_t ldda, magma_int_t inca,
-    cuDoubleComplex *dB, magma_int_t lddb, magma_int_t incb );
-
-  /*
-   * Level 2 BLAS
-   */
-void magmablas_zgemv(
-    char t, magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dx, magma_int_t incx,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dy, magma_int_t incy );
-
-#if defined(PRECISION_z ) || defined(PRECISION_c )
-magma_int_t magmablas_zhemv(
-    char u, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dx, magma_int_t incx,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dy, magma_int_t incy );
-#endif
-
-magma_int_t magmablas_zsymv(
-    char u, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dx, magma_int_t incx,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dy, magma_int_t incy );
-
-  /*
-   * Level 3 BLAS
-   */
-void magmablas_zgemm(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dB, magma_int_t lddb,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magmablas_zgemm_fermi80(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dB, magma_int_t lddb,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magmablas_zgemm_fermi64(
-    char tA, char tB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dB, magma_int_t lddb,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magmablas_zhemm(
-    char s, char u,
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dB, magma_int_t lddb,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magmablas_zsymm(
-    char s, char u,
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dB, magma_int_t lddb,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magmablas_zsyrk(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magmablas_zherk(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    double  alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    double  beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magmablas_zsyr2k(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dB, magma_int_t lddb,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magmablas_zher2k(
-    char u, char t,
-    magma_int_t n, magma_int_t k,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dB, magma_int_t lddb,
-    double  beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magmablas_ztrmm(
-    char s, char u, char t,  char d,
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    cuDoubleComplex       *dB, magma_int_t lddb );
-
-void magmablas_ztrsm(
-    char s, char u, char t, char d,
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    cuDoubleComplex       *dB, magma_int_t lddb );
-
-
-  /*
-   * Wrappers for platform independence.
-   * These wrap CUBLAS or AMD OpenCL BLAS functions.
-   */
-
-// ========================================
-// copying vectors
-// set copies host to device
-// get copies device to host
-// Add the function, file, and line for error-reporting purposes.
-
-#define magma_zsetvector(           n, hx_src, incx, dy_dst, incy ) \
-        magma_zsetvector_internal(  n, hx_src, incx, dy_dst, incy, __func__, __FILE__, __LINE__ )
-
-#define magma_zgetvector(           n, dx_src, incx, hy_dst, incy ) \
-        magma_zgetvector_internal(  n, dx_src, incx, hy_dst, incy, __func__, __FILE__, __LINE__ )
-
-#define magma_zsetvector_async(           n, hx_src, incx, dy_dst, incy, stream ) \
-        magma_zsetvector_async_internal(  n, hx_src, incx, dy_dst, incy, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_zgetvector_async(           n, dx_src, incx, hy_dst, incy, stream ) \
-        magma_zgetvector_async_internal(  n, dx_src, incx, hy_dst, incy, stream, __func__, __FILE__, __LINE__ )
-
-void magma_zsetvector_internal(
-    magma_int_t n,
-    const cuDoubleComplex *hx_src, magma_int_t incx,
-    cuDoubleComplex       *dy_dst, magma_int_t incy,
-    const char* func, const char* file, int line );
-
-void magma_zgetvector_internal(
-    magma_int_t n,
-    const cuDoubleComplex *dx_src, magma_int_t incx,
-    cuDoubleComplex       *hy_dst, magma_int_t incy,
-    const char* func, const char* file, int line );
-
-void magma_zsetvector_async_internal(
-    magma_int_t n,
-    const cuDoubleComplex *hx_src, magma_int_t incx,
-    cuDoubleComplex       *dy_dst, magma_int_t incy,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_zgetvector_async_internal(
-    magma_int_t n,
-    const cuDoubleComplex *dx_src, magma_int_t incx,
-    cuDoubleComplex       *hy_dst, magma_int_t incy,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-
-// ========================================
-// copying sub-matrices (contiguous columns )
-// set  copies host to device
-// get  copies device to host
-// copy copies device to device
-// (with CUDA unified addressing, copy can be between same or different devices )
-// Add the function, file, and line for error-reporting purposes.
-
-#define magma_zsetmatrix(           m, n, hA_src, lda, dB_dst, lddb ) \
-        magma_zsetmatrix_internal(  m, n, hA_src, lda, dB_dst, lddb, __func__, __FILE__, __LINE__ )
-
-#define magma_zgetmatrix(           m, n, dA_src, ldda, hB_dst, ldb ) \
-        magma_zgetmatrix_internal(  m, n, dA_src, ldda, hB_dst, ldb, __func__, __FILE__, __LINE__ )
-
-#define magma_zcopymatrix(          m, n, dA_src, ldda, dB_dst, lddb ) \
-        magma_zcopymatrix_internal( m, n, dA_src, ldda, dB_dst, lddb, __func__, __FILE__, __LINE__ )
-
-#define magma_zsetmatrix_async(           m, n, hA_src, lda, dB_dst, lddb, stream ) \
-        magma_zsetmatrix_async_internal(  m, n, hA_src, lda, dB_dst, lddb, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_zgetmatrix_async(           m, n, dA_src, ldda, hB_dst, ldb, stream ) \
-        magma_zgetmatrix_async_internal(  m, n, dA_src, ldda, hB_dst, ldb, stream, __func__, __FILE__, __LINE__ )
-
-#define magma_zcopymatrix_async(          m, n, dA_src, ldda, dB_dst, lddb, stream ) \
-        magma_zcopymatrix_async_internal( m, n, dA_src, ldda, dB_dst, lddb, stream, __func__, __FILE__, __LINE__ )
-
-void magma_zsetmatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *hA_src, magma_int_t lda,
-    cuDoubleComplex       *dB_dst, magma_int_t lddb,
-    const char* func, const char* file, int line );
-
-void magma_zgetmatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *dA_src, magma_int_t ldda,
-    cuDoubleComplex       *hB_dst, magma_int_t ldb,
-    const char* func, const char* file, int line );
-
-void magma_zcopymatrix_internal(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *dA_src, magma_int_t ldda,
-    cuDoubleComplex       *dB_dst, magma_int_t lddb,
-    const char* func, const char* file, int line );
-
-void magma_zsetmatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *hA_src, magma_int_t lda,
-    cuDoubleComplex       *dB_dst, magma_int_t lddb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_zgetmatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *dA_src, magma_int_t ldda,
-    cuDoubleComplex       *hB_dst, magma_int_t ldb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-void magma_zcopymatrix_async_internal(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *dA_src, magma_int_t ldda,
-    cuDoubleComplex       *dB_dst, magma_int_t lddb,
-    magma_stream_t stream,
-    const char* func, const char* file, int line );
-
-
-// ========================================
-// Level 1 BLAS
-
-void magma_zswap(
-    magma_int_t n,
-    cuDoubleComplex *dx, magma_int_t incx,
-    cuDoubleComplex *dy, magma_int_t incy );
-
-magma_int_t magma_izamax(
-    magma_int_t n,
-    cuDoubleComplex *dx, magma_int_t incx );
-
-// ========================================
-// Level 2 BLAS
-
-void magma_zgemv(
-    magma_trans_t transA,
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dx, magma_int_t incx,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dy, magma_int_t incy );
-
-void magma_zhemv(
-    magma_uplo_t uplo,
-    magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dx, magma_int_t incx,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dy, magma_int_t incy );
-
-void magma_ztrsv(
-    magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t n,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    cuDoubleComplex       *dx, magma_int_t incx );
-
-// ========================================
-// Level 3 BLAS
-
-void magma_zgemm(
-    magma_trans_t transA, magma_trans_t transB,
-    magma_int_t m, magma_int_t n, magma_int_t k,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dB, magma_int_t lddb,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magma_zhemm(
-    magma_side_t side, magma_uplo_t uplo,
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dB, magma_int_t lddb,
-    cuDoubleComplex beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magma_zherk(
-    magma_uplo_t uplo, magma_trans_t trans,
-    magma_int_t n, magma_int_t k,
-    double alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    double beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magma_zher2k(
-    magma_uplo_t uplo, magma_trans_t trans,
-    magma_int_t n, magma_int_t k,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    const cuDoubleComplex *dB, magma_int_t lddb,
-    double beta,
-    cuDoubleComplex       *dC, magma_int_t lddc );
-
-void magma_ztrmm(
-    magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    cuDoubleComplex       *dB, magma_int_t lddb );
-
-void magma_ztrsm(
-    magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
-    magma_int_t m, magma_int_t n,
-    cuDoubleComplex alpha,
-    const cuDoubleComplex *dA, magma_int_t ldda,
-    cuDoubleComplex       *dB, magma_int_t lddb );
-
-#ifdef __cplusplus
-}
-#endif
-
-#undef PRECISION_z
-
-#endif  // MAGMABLAS_Z_H
diff --git a/src/include/magmablas_zc.h b/src/include/magmablas_zc.h
deleted file mode 100644
index 8d4ae4d..0000000
--- a/src/include/magmablas_zc.h
+++ /dev/null
@@ -1,55 +0,0 @@
-/*
- *   -- MAGMA (version 1.3.0) --
- *      Univ. of Tennessee, Knoxville
- *      Univ. of California, Berkeley
- *      Univ. of Colorado, Denver
- *      November 2012
- *
- * @precisions mixed zc -> ds
- */
-
-#ifndef MAGMABLAS_ZC_H
-#define MAGMABLAS_ZC_H
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-  /* Mixed precision */
-void magmablas_zcaxpycp(
-    cuFloatComplex *R, cuDoubleComplex *X,
-    magma_int_t m, cuDoubleComplex *B, cuDoubleComplex *W );
-
-void magmablas_zaxpycp(
-    cuDoubleComplex *R, cuDoubleComplex *X,
-    magma_int_t m, cuDoubleComplex *B );
-
-void magmablas_zclaswp(
-    magma_int_t n,
-    cuDoubleComplex *A, magma_int_t lda,
-    cuFloatComplex *SA, magma_int_t m,
-    const magma_int_t *ipiv, magma_int_t incx );
-
-void magmablas_zlag2c(
-    magma_int_t m, magma_int_t n,
-    const cuDoubleComplex *A,  magma_int_t lda,
-    cuFloatComplex        *SA, magma_int_t ldsa,
-    magma_int_t *info );
-
-void magmablas_clag2z(
-    magma_int_t m, magma_int_t n, 
-    const cuFloatComplex  *SA, magma_int_t ldsa, 
-    cuDoubleComplex       *A,  magma_int_t lda, 
-    magma_int_t *info );
-
-void magmablas_zlat2c(
-    char uplo, magma_int_t n, 
-    const cuDoubleComplex *A,  magma_int_t lda, 
-    cuFloatComplex        *SA, magma_int_t ldsa, 
-    magma_int_t *info );
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // MAGMABLAS_ZC_H
diff --git a/src/magAriths.c b/src/magAriths.c
index dab7dde..80c8f05 100644
--- a/src/magAriths.c
+++ b/src/magAriths.c
@@ -23,8 +23,9 @@
 #include "magAriths.h"
 #include "magUtils.h"
 
-#include <cublas.h>
-#include <magmablas.h>
+#include <cublas_v2.h>
+#include <magma_types.h>
+#include <magma.h>
 
 
 SEXP magMultmm(SEXP a, SEXP transa, SEXP b, SEXP transb)
@@ -37,9 +38,10 @@ SEXP magMultmm(SEXP a, SEXP transa, SEXP b, SEXP transb)
        LDA = DIMA[0], LDB = DIMB[0], LDC = M;
    char TRANSA = (TA ? 'T' : 'N'), TRANSB = (TB ? 'T' : 'N');
    double *A = REAL(PROTECT(AS_NUMERIC(a))), *B = REAL(PROTECT(AS_NUMERIC(b))),
-          *dA, *dB, *dC;
+          *dA, *dB, *dC, one=1., zero=0.;
  
    if(DIMB[TB] != K) error("non-conformable matrices");
+   /*warning("here magMultmm");*/
 
    c = SET_SLOT(c, install(".Data"), allocMatrix(REALSXP, M, N));
    SET_SLOT(c, install("gpu"), duplicate(gpu));
@@ -51,10 +53,13 @@ SEXP magMultmm(SEXP a, SEXP transa, SEXP b, SEXP transb)
    magma_dsetmatrix(DIMA[0], DIMA[1], A, LDA, dA, LDA);
    magma_dsetmatrix(DIMB[0], DIMB[1], B, LDB, dB, LDB);
 
-   if(LOGICAL_VALUE(gpu))
-      magmablas_dgemm(TRANSA, TRANSB, M, N, K, 1.0, dA, LDA, dB, LDB, 0.0, dC, LDC);
-   else
-      cublasDgemm(TRANSA, TRANSB, M, N, K, 1.0, dA, LDA, dB, LDB, 0.0, dC, LDC);
+   if(LOGICAL_VALUE(gpu)) {
+      /*warning("here magma");*/
+      magmablas_dgemm(magma_trans_const(TRANSA), magma_trans_const(TRANSB), M, N, K, 1.0, dA, LDA, dB, LDB, 0.0, dC, LDC);
+   } else {
+      /*warning("here cublas");*/
+      cublasDgemm(handle, cublas_trans_const(TRANSA), cublas_trans_const(TRANSB), M, N, K, &one, dA, LDA, dB, LDB, &zero, dC, LDC);
+   }
 
    magma_dgetmatrix(M, N, dC, LDC, REAL(c), LDC);
 
@@ -77,7 +82,7 @@ SEXP magMultmv(SEXP a, SEXP transa, SEXP x, SEXP right)
        M = DIMA[0], N = DIMA[1], LENX = LENGTH(x), LENY = DIMA[TA], LDA=M;
    char TRANSA = (TA ? 'T' : 'N');
    double *A = REAL(PROTECT(AS_NUMERIC(a))), *X = REAL(PROTECT(AS_NUMERIC(x))),
-          *dA, *dX, *dY;
+          *dA, *dX, *dY, one=1., zero=0.;
 
    if(DIMA[!TA] != LENX) error("non-conformable matrices");
 
@@ -93,9 +98,9 @@ SEXP magMultmv(SEXP a, SEXP transa, SEXP x, SEXP right)
    magma_dsetvector(LENX, X, 1, dX, 1);
 
    if(LOGICAL_VALUE(gpu)) {
-      magmablas_dgemv(TRANSA, M, N, 1.0, dA, LDA, dX, 1, 0.0, dY, 1);
+      magmablas_dgemv(magma_trans_const(TRANSA), M, N, 1.0, dA, LDA, dX, 1, 0.0, dY, 1);
    } else {
-      cublasDgemv(TRANSA, M, N, 1.0, dA, LDA, dX, 1, 0.0, dY, 1);
+      cublasDgemv(handle, cublas_trans_const(TRANSA), M, N, &one, dA, LDA, dX, 1, &zero, dY, 1);
    }
 
    magma_dgetvector(LENY, dY, 1, REAL(y), 1);
diff --git a/src/magFactors.c b/src/magFactors.c
index 6632448..565610b 100644
--- a/src/magFactors.c
+++ b/src/magFactors.c
@@ -23,7 +23,7 @@
 #include "magUtils.h"
 #include "magFactors.h"
 
-#include <cublas.h>
+#include <cublas_v2.h>
 #include <cuda.h>
 #include <magma.h>
 #include <magma_lapack.h>
@@ -48,7 +48,7 @@ SEXP magChol(SEXP a)
       magma_malloc((void**)&dB, N2*sizeof(double));
 
       magma_dsetmatrix(N, N, B, LDB, dB, LDB);
-      magma_dpotrf_gpu('U', N, dB, LDB, &info);
+      magma_dpotrf_gpu(magma_uplo_const('U'), N, dB, LDB, &info);
       magma_dgetmatrix(N, N, dB, LDB, B, LDB);
 
       magma_free(dB);
@@ -57,7 +57,7 @@ SEXP magChol(SEXP a)
 
       magma_malloc_pinned((void**)&hB, N2*sizeof(double));
       lapackf77_dlacpy(MagmaUpperStr, &N, &N, B, &LDB, hB, &LDB);
-      magma_dpotrf('U', N, hB, N, &info);
+      magma_dpotrf(magma_uplo_const('U'), N, hB, N, &info);
       lapackf77_dlacpy(MagmaUpperStr, &N, &N, hB, &LDB, B, &LDB);
 
       magma_free_pinned(hB);
@@ -150,7 +150,7 @@ SEXP magQR(SEXP a)
       magma_malloc((void**)&dT, LENT*sizeof(double));
 
       magma_dsetmatrix(M, N, A, M, dA, M);
-      magma_dgeqrf_gpu(M, N, dA, M, tau, dT, &info);
+      magma_dgeqrf3_gpu(M, N, dA, M, tau, dT, &info);
       magma_dgetmatrix(M, N, dA, M, A, M);
       magma_dgetvector(LENT, dT, 1, work, 1);
 
@@ -164,7 +164,7 @@ SEXP magQR(SEXP a)
       magma_malloc_pinned((void**)&hwork, LWORK*sizeof(double));
 
       lapackf77_dlacpy(MagmaUpperLowerStr, &M, &N, A, &M, hA, &M);
-      magma_dgeqrf(M, N, hA, M, tau, hwork, LWORK, &info);
+      magma_dgeqrf_ooc(M, N, hA, M, tau, hwork, LWORK, &info);
       lapackf77_dlacpy(MagmaUpperLowerStr, &M, &N, hA, &M, A, &M);
 
       magma_free_pinned(hA);
diff --git a/src/magSolvers.c b/src/magSolvers.c
index 58f722b..de021a1 100644
--- a/src/magSolvers.c
+++ b/src/magSolvers.c
@@ -23,7 +23,7 @@
 #include "magUtils.h"
 #include "magSolvers.h"
 
-#include <cublas.h>
+#include <cublas_v2.h>
 #include <magmablas.h>
 #include <magma.h>
 #include <magma_lapack.h>
@@ -50,7 +50,7 @@ SEXP magCholSolve(SEXP a, SEXP b)
 
       magma_dsetmatrix(N, N, A, N, dA, N);
       magma_dsetmatrix(N, NRHS, REAL(c), N, dB, N);
-      magma_dpotrs_gpu('U', N, NRHS, dA, N, dB, N, &info);
+      magma_dpotrs_gpu(magma_uplo_const('U'), N, NRHS, dA, N, dB, N, &info);
       magma_dgetmatrix(N, NRHS, dB, N, REAL(c), N);
 
       magma_free(dA);
@@ -90,7 +90,7 @@ SEXP magLUSolve(SEXP a, SEXP b)
 
       magma_dsetmatrix(N, N, A, N, dA, LDA);
       magma_dsetmatrix(N, NRHS, REAL(c), N, dB, LDB);
-      magma_dgetrs_gpu('N', N, NRHS, dA, LDA, ipiv, dB, LDB, &info);
+      magma_dgetrs_gpu(magma_trans_const('N'), N, NRHS, dA, LDA, ipiv, dB, LDB, &info);
       magma_dgetmatrix(N, NRHS, dB, LDB, REAL(c), N);
 
       magma_free(dA);
@@ -127,7 +127,7 @@ SEXP magQRSolve(SEXP a, SEXP b)
 
    magma_malloc_pinned((void**)&hwork, LWORK*sizeof(double));
 
-   if(LOGICAL_VALUE(gpu) && 0) {
+   if(LOGICAL_VALUE(gpu)) {
       SEXP workS = GET_SLOT(a, install("work"));
       int LENT = LENGTH(workS);
       double *dA, *dB, *dT, *work = REAL(workS);
@@ -140,9 +140,9 @@ SEXP magQRSolve(SEXP a, SEXP b)
       magma_dsetmatrix(M, NRHS, B, M, dB, M);
       magma_dsetvector(LENT, work, 1, dT, 1);
 
-      magma_dgeqrs_gpu(M, N, NRHS, dA, M, tau, dB, dT, M, hwork, LWORK, &info);
+      magma_dgeqrs3_gpu(M, N, NRHS, dA, M, tau, dT, dB, M, hwork, LWORK, &info);
 
-      magma_dgetmatrix(N, NRHS, dA, M, REAL(c), N);
+      magma_dgetmatrix(N, NRHS, dB, M, REAL(c), N);
 
       magma_free(dA);
       magma_free(dB);
@@ -192,7 +192,7 @@ SEXP magSolve(SEXP a, SEXP b)
       if(info < 0) error("illegal argument %d in 'magSolve'", -1 * info);
       else if(info > 0) error("non-singular matrix");
 
-      magma_dgetrs_gpu('N', N, NRHS, dA, N, ipiv, dB, N, &info);
+      magma_dgetrs_gpu(magma_trans_const('N'), N, NRHS, dA, N, ipiv, dB, N, &info);
       if(info < 0) error("illegal argument %d in 'magSolve'", -1 * info);
 
       magma_dgetmatrix(N, NRHS, dB, N, REAL(c), N);
@@ -226,7 +226,7 @@ SEXP magTriSolve(SEXP a, SEXP b, SEXP k, SEXP uprtri, SEXP transa)
    char UPLO = (LOGICAL_VALUE(uprtri) ? 'U' : 'L'),
         TRANSA = (LOGICAL_VALUE(transa) ? 'T' : 'N');
    double *A = REAL(PROTECT(AS_NUMERIC(a))), *B = REAL(PROTECT(AS_NUMERIC(b))),
-          *dA, *dB;
+          *dA, *dB, one=1., zero=0.;
 
    if((K <= 0) || (K > M)) error("invalid number of equations");
 
@@ -240,9 +240,9 @@ SEXP magTriSolve(SEXP a, SEXP b, SEXP k, SEXP uprtri, SEXP transa)
    magma_dsetmatrix(M, N, B, M, dB, M);
 
    if(LOGICAL_VALUE(gpu))
-      magma_dtrsm('L', UPLO, TRANSA, 'N', K, N, 1.0, dA, M, dB, M);
+      magma_dtrsm(magma_side_const('L'), magma_uplo_const(UPLO), magma_trans_const(TRANSA), magma_diag_const('N'), K, N, 1.0, dA, M, dB, M);
    else
-      cublasDtrsm('L', UPLO, TRANSA, 'N', K, N, 1.0, dA, M, dB, M);
+      cublasDtrsm(handle, cublas_side_const('L'), cublas_uplo_const(UPLO), cublas_trans_const(TRANSA), cublas_trans_const('N'), K, N, &one, dA, M, dB, M);
 
    magma_dgetmatrix(K, N, dB, M, REAL(c), K);
 
diff --git a/src/magUtils.c b/src/magUtils.c
index 1bb2e33..89ca74a 100644
--- a/src/magUtils.c
+++ b/src/magUtils.c
@@ -21,8 +21,9 @@
  ********************************************************************************/
 
 #include "magUtils.h"
+cublasHandle_t handle;
 
-#include <cublas.h>
+#include <cublas_v2.h>
 #include <cuda.h>
 #include <cuda_runtime_api.h>
 
@@ -52,14 +53,17 @@ SEXP magGetGPU(SEXP a, SEXP b)
 
 void magLoad()
 {
-   cublasInit();
+   magma_init();
+   cudaSetDevice( 0 );
+   cublasCreate_v2( &handle );
    checkCublasError("initialization failed");
 }
 
 
 void magUnload()
 {
-   cublasShutdown();
+   cublasDestroy(handle);
+   magma_finalize();
 }
 
 
@@ -71,7 +75,7 @@ int checkCudaError(const char *msg)
 	return 0;
 }
 
-char *cublasGetErrorString(cublasStatus err)
+char *cublasGetErrorString(cublasStatus_t err)
 {
 	switch(err) {
 		case CUBLAS_STATUS_SUCCESS :
@@ -98,7 +102,7 @@ char *cublasGetErrorString(cublasStatus err)
 
 int checkCublasError(const char *msg)
 {
-	cublasStatus err = cublasGetError();
+	cublasStatus_t err = cublasGetError();
 	if(err != CUBLAS_STATUS_SUCCESS)
 		error("CUBLAS %s : %s\n", msg, cublasGetErrorString(err));
 	return 0;
diff --git a/src/magUtils.h b/src/magUtils.h
index 68768d3..c6e1b66 100644
--- a/src/magUtils.h
+++ b/src/magUtils.h
@@ -42,5 +42,6 @@ extern "C" {
 #ifdef	__cplusplus
 }
 #endif
-
+#include <cublas_v2.h>
+extern cublasHandle_t handle;
 #endif	/* _MAGUTILS_H */
diff --git a/src/todo.txt b/src/todo.txt
new file mode 100644
index 0000000..c91e8d0
--- /dev/null
+++ b/src/todo.txt
@@ -0,0 +1,8 @@
+- cublas_v2.h
+- cublasStatus -> cublasStatus_t
+- add cublas_handle_t handle as global variable in magmaUtils.[ch];
+- add handle to cublas...() calls as the first parameter
+- in cublas...() one and zero are sent by ref.
+- magma_trans_const(TRANSA), cublas_trans_const(TRANSA)
+- cublas_uplo_const(UPLO), magma_uplo_const(UPLO)
+- cublas_side_const('L'), magma_side_const('L')