AVX512 support

.ci_support/linux_64_cuda_compiler_version12.8microarch_level1.yaml → .ci_support/linux_64_cuda_compiler_version12.9microarch_level1.yaml

@@ -13,7 +13,7 @@ channel_targets:
 cuda_compiler:
 - cuda-nvcc
 cuda_compiler_version:
-- '12.8'
+- '12.9'
 cudnn:
 - '9'
 cxx_compiler:

.ci_support/linux_64_cuda_compiler_version12.8microarch_level3.yaml → .ci_support/linux_64_cuda_compiler_version12.9microarch_level3.yaml

@@ -13,7 +13,7 @@ channel_targets:
 cuda_compiler:
 - cuda-nvcc
 cuda_compiler_version:
-- '12.8'
+- '12.9'
 cudnn:
 - '9'
 cxx_compiler:

.ci_support/linux_64_cuda_compiler_version12.9microarch_level4.yaml

@@ -0,0 +1,80 @@
+c_compiler:
+- gcc
+c_compiler_version:
+- '14'
+c_stdlib:
+- sysroot
+c_stdlib_version:
+- '2.17'
+channel_sources:
+- conda-forge
+channel_targets:
+- conda-forge main
+cuda_compiler:
+- cuda-nvcc
+cuda_compiler_version:
+- '12.9'
+cudnn:
+- '9'
+cxx_compiler:
+- gxx
+cxx_compiler_version:
+- '14'
+docker_image:
+- quay.io/condaforge/linux-anvil-x86_64:alma9
+flatbuffers:
+- 25.9.23
+giflib:
+- '5.2'
+github_actions_labels:
+- cirun-openstack-cpu-2xlarge
+icu:
+- '78'
+libabseil:
+- '20250512'
+libcurl:
+- '8'
+libgrpc:
+- '1.73'
+libjpeg_turbo:
+- '3'
+libpng:
+- '1.6'
+libprotobuf:
+- 6.31.1
+libtensorflow_cc:
+- '2.16'
+microarch_level:
+- '4'
+nccl:
+- '2'
+numpy:
+- '2'
+openssl:
+- '3.5'
+perl:
+- 5.32.1
+pin_run_as_build:
+  python:
+    min_pin: x.x
+    max_pin: x.x
+python:
+- 3.10.* *_cpython
+- 3.11.* *_cpython
+- 3.12.* *_cpython
+- 3.13.* *_cp313
+snappy:
+- '1.2'
+sqlite:
+- '3'
+target_platform:
+- linux-64
+tensorflow:
+- '2.16'
+zip_keys:
+- - c_compiler_version
+  - cxx_compiler_version
+  - c_stdlib_version
+  - cuda_compiler_version
+zlib:
+- '1'

.ci_support/linux_64_cuda_compiler_versionNonemicroarch_level4.yaml

@@ -0,0 +1,80 @@
+c_compiler:
+- gcc
+c_compiler_version:
+- '14'
+c_stdlib:
+- sysroot
+c_stdlib_version:
+- '2.17'
+channel_sources:
+- conda-forge
+channel_targets:
+- conda-forge main
+cuda_compiler:
+- cuda-nvcc
+cuda_compiler_version:
+- None
+cudnn:
+- '9'
+cxx_compiler:
+- gxx
+cxx_compiler_version:
+- '14'
+docker_image:
+- quay.io/condaforge/linux-anvil-x86_64:alma9
+flatbuffers:
+- 25.9.23
+giflib:
+- '5.2'
+github_actions_labels:
+- cirun-openstack-cpu-2xlarge
+icu:
+- '78'
+libabseil:
+- '20250512'
+libcurl:
+- '8'
+libgrpc:
+- '1.73'
+libjpeg_turbo:
+- '3'
+libpng:
+- '1.6'
+libprotobuf:
+- 6.31.1
+libtensorflow_cc:
+- '2.16'
+microarch_level:
+- '4'
+nccl:
+- '2'
+numpy:
+- '2'
+openssl:
+- '3.5'
+perl:
+- 5.32.1
+pin_run_as_build:
+  python:
+    min_pin: x.x
+    max_pin: x.x
+python:
+- 3.10.* *_cpython
+- 3.11.* *_cpython
+- 3.12.* *_cpython
+- 3.13.* *_cp313
+snappy:
+- '1.2'
+sqlite:
+- '3'
+target_platform:
+- linux-64
+tensorflow:
+- '2.16'
+zip_keys:
+- - c_compiler_version
+  - cxx_compiler_version
+  - c_stdlib_version
+  - cuda_compiler_version
+zlib:
+- '1'

.github/workflows/conda-build.yml

@@ -21,15 +21,20 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - CONFIG: linux_64_cuda_compiler_version12.8microarch_level1
+          - CONFIG: linux_64_cuda_compiler_version12.9microarch_level1
             UPLOAD_PACKAGES: True
             os: ubuntu
-            runs_on: ['cirun-openstack-cpu-2xlarge--${{ github.run_id }}-linux_64_cuda_compiler_version12.8microa_hc8cea6d2', 'linux', 'x64', 'self-hosted']
+            runs_on: ['cirun-openstack-cpu-2xlarge--${{ github.run_id }}-linux_64_cuda_compiler_version12.9microa_h28a6e8b1', 'linux', 'x64', 'self-hosted']
             DOCKER_IMAGE: quay.io/condaforge/linux-anvil-x86_64:alma9
-          - CONFIG: linux_64_cuda_compiler_version12.8microarch_level3
+          - CONFIG: linux_64_cuda_compiler_version12.9microarch_level3
             UPLOAD_PACKAGES: True
             os: ubuntu
-            runs_on: ['cirun-openstack-cpu-2xlarge--${{ github.run_id }}-linux_64_cuda_compiler_version12.8microa_h82e5350b', 'linux', 'x64', 'self-hosted']
+            runs_on: ['cirun-openstack-cpu-2xlarge--${{ github.run_id }}-linux_64_cuda_compiler_version12.9microa_h764a83b2', 'linux', 'x64', 'self-hosted']
+            DOCKER_IMAGE: quay.io/condaforge/linux-anvil-x86_64:alma9
+          - CONFIG: linux_64_cuda_compiler_version12.9microarch_level4
+            UPLOAD_PACKAGES: True
+            os: ubuntu
+            runs_on: ['cirun-openstack-cpu-2xlarge--${{ github.run_id }}-linux_64_cuda_compiler_version12.9microa_h5fa35844', 'linux', 'x64', 'self-hosted']
             DOCKER_IMAGE: quay.io/condaforge/linux-anvil-x86_64:alma9
           - CONFIG: linux_64_cuda_compiler_versionNonemicroarch_level1
             UPLOAD_PACKAGES: True
@@ -41,6 +46,11 @@ jobs:
             os: ubuntu
             runs_on: ['cirun-openstack-cpu-2xlarge--${{ github.run_id }}-linux_64_cuda_compiler_versionNonemicroa_hb1134425', 'linux', 'x64', 'self-hosted']
             DOCKER_IMAGE: quay.io/condaforge/linux-anvil-x86_64:alma9
+          - CONFIG: linux_64_cuda_compiler_versionNonemicroarch_level4
+            UPLOAD_PACKAGES: True
+            os: ubuntu
+            runs_on: ['cirun-openstack-cpu-2xlarge--${{ github.run_id }}-linux_64_cuda_compiler_versionNonemicroa_h0ea2bc29', 'linux', 'x64', 'self-hosted']
+            DOCKER_IMAGE: quay.io/condaforge/linux-anvil-x86_64:alma9
           - CONFIG: linux_aarch64_cuda_compiler_versionNone
             UPLOAD_PACKAGES: True
             os: ubuntu
@@ -49,7 +59,7 @@ jobs:
     steps:
 
     - name: Checkout code
-      uses: actions/checkout@0ad4b8fadaa221de15dcec353f45205ec38ea70b # v4.1.4
+      uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
     - name: Build on Linux
       id: build-linux

recipe/build.sh

@@ -2,6 +2,9 @@
 
 set -ex
 
+# See https://github.com/conda-forge/bazel-feedstock/issues/273
+find $BUILD_PREFIX/share/bazel/install | xargs -n 1 touch -mt 203601010101
+
 for ver in 3.9 3.10 3.11 3.12; do
   export PY_VER=$ver
   echo "Building for $PY_VER"

recipe/build_common.sh

@@ -80,12 +80,17 @@ export BAZEL_MKL_OPT=""
 mkdir -p ./bazel_output_base
 export BAZEL_OPTS=""
 # Set this to something as otherwise, it would include CFLAGS which itself contains a host path and this then breaks bazel's include path validation.
-if [[ "${target_platform}" != *-64 ]]; then
-  export CC_OPT_FLAGS="-O2"
-elif [[ "${microarch_level}" == "1" ]]; then
-  export CC_OPT_FLAGS="-O2 -march=nocona -mtune=haswell"
-else
-  export CC_OPT_FLAGS="-O2 -march=x86-64-v${microarch_level}"
+export CC_OPT_FLAGS="-O2"
+
+if [[ "${target_platform}" == "linux-64" && "${microarch_level}" == "4" &&  "${CROSSCOMPILING_EMULATOR:-}" != "" ]]; then
+  # we can't use the x86-64-v4 emulator from Intel with bazel. Instead build with
+  # - v4 for TF libraries
+  # - v3 for external libraries
+  # - v3 for TF executables
+  CFLAGS=$(echo "$CFLAGS" | sed 's/-march=x86-64-v4/-march=x86-64-v3/g')
+  CXXFLAGS=$(echo "$CXXFLAGS" | sed 's/-march=x86-64-v4/-march=x86-64-v3/g')
+  CPPFLAGS=$(echo "$CPPFLAGS" | sed 's/-march=x86-64-v4/-march=x86-64-v3/g')
+  BAZEL_OPTS="${BAZEL_OPTS} --copt=-march=x86-64-v${microarch_level} --host_copt=-march=x86-64-v3"
 fi
 
 # Quick debug:
@@ -250,7 +255,7 @@ rm -f tensorflow/lite/acceleration/configuration/configuration_generated.h
 sed -ie "s;BUILD_PREFIX;${BUILD_PREFIX};g" tensorflow/tools/pip_package/build_pip_package.py
 
 # build using bazel
-bazel ${BAZEL_OPTS} build ${BUILD_TARGET}
+bazel build ${BUILD_TARGET} ${BAZEL_OPTS}
 
 # build a whl file
 mkdir -p $SRC_DIR/tensorflow_pkg

recipe/conda_build_config.yaml

@@ -13,3 +13,4 @@ microarch_level:
 - 1
 # hmaarrfk -- temporary, uncomment during actual build
 - 3   # [linux and x86_64]
+- 4   # [linux and x86_64]