Add native CPU build toggle

Author: 0xShug0

CMakeLists.txt

@@ -19,6 +19,7 @@ set(ENGINE_DEFAULT_ENABLE_VULKAN OFF)
 set(ENGINE_DEFAULT_ENABLE_METAL OFF)
 set(ENGINE_DEFAULT_ENABLE_LLAMAFILE ON)
 set(ENGINE_DEFAULT_ENABLE_CUDA_GRAPHS ON)
+set(ENGINE_DEFAULT_ENABLE_NATIVE_CPU ON)
 if (APPLE)
     set(ENGINE_DEFAULT_ENABLE_METAL ON)
 endif()
@@ -43,6 +44,7 @@ option(ENGINE_ENABLE_VULKAN "Build ggml with Vulkan backend support" ${ENGINE_DE
 option(ENGINE_ENABLE_METAL "Build ggml with Metal backend support" ${ENGINE_DEFAULT_ENABLE_METAL})
 option(ENGINE_ENABLE_LLAMAFILE "Build ggml with llamafile SGEMM support" ${ENGINE_DEFAULT_ENABLE_LLAMAFILE})
 option(ENGINE_ENABLE_CUDA_GRAPHS "Enable ggml CUDA graphs support" ${ENGINE_DEFAULT_ENABLE_CUDA_GRAPHS})
+option(ENGINE_ENABLE_NATIVE_CPU "Build ggml CPU kernels with native host ISA flags" ${ENGINE_DEFAULT_ENABLE_NATIVE_CPU})
 option(ENGINE_ENABLE_OPENMP "Build host code with OpenMP support" ON)
 option(ENGINE_BUILD_EXAMPLES "Build example binaries" OFF)
 option(ENGINE_BUILD_TESTS "Build framework unit tests" OFF)
@@ -66,6 +68,7 @@ set(GGML_VULKAN ${ENGINE_ENABLE_VULKAN} CACHE BOOL "Build ggml with Vulkan backe
 set(GGML_METAL ${ENGINE_ENABLE_METAL} CACHE BOOL "Build ggml with Metal backend support" FORCE)
 set(GGML_LLAMAFILE ${ENGINE_ENABLE_LLAMAFILE} CACHE BOOL "Build ggml with llamafile SGEMM support" FORCE)
 set(GGML_CUDA_GRAPHS ${ENGINE_ENABLE_CUDA_GRAPHS} CACHE BOOL "Enable ggml CUDA graphs support" FORCE)
+set(GGML_NATIVE ${ENGINE_ENABLE_NATIVE_CPU} CACHE BOOL "Build ggml CPU kernels with native host ISA flags" FORCE)
 
 set(SPM_ENABLE_SHARED OFF CACHE BOOL "Build sentencepiece shared libs" FORCE)
 set(SPM_BUILD_TEST OFF CACHE BOOL "Do not build sentencepiece tests" FORCE)

Makefile.windows

@@ -7,16 +7,17 @@ PRESET ?= windows-cuda-release
 TARGET ?= audiocpp_cli
 JOBS ?=
 VS_INSTALL ?=
+NATIVE_CPU ?=
 
 .PHONY: all build configure cpu cpu-release cuda cuda-debug cuda-release clean presets help
 
 all: cuda
 
 build:
-	$(POWERSHELL) -NoProfile -ExecutionPolicy Bypass -File "$(BUILD_SCRIPT)" -Preset "$(PRESET)" $(if $(TARGET),-Target "$(TARGET)",) $(if $(JOBS),-Jobs "$(JOBS)",) $(if $(VS_INSTALL),-VsInstall "$(VS_INSTALL)",)
+	$(POWERSHELL) -NoProfile -ExecutionPolicy Bypass -File "$(BUILD_SCRIPT)" -Preset "$(PRESET)" $(if $(TARGET),-Target "$(TARGET)",) $(if $(JOBS),-Jobs "$(JOBS)",) $(if $(NATIVE_CPU),-NativeCpu "$(NATIVE_CPU)",) $(if $(VS_INSTALL),-VsInstall "$(VS_INSTALL)",)
 
 configure:
-	$(POWERSHELL) -NoProfile -ExecutionPolicy Bypass -File "$(BUILD_SCRIPT)" -Preset "$(PRESET)" -ConfigureOnly $(if $(VS_INSTALL),-VsInstall "$(VS_INSTALL)",)
+	$(POWERSHELL) -NoProfile -ExecutionPolicy Bypass -File "$(BUILD_SCRIPT)" -Preset "$(PRESET)" -ConfigureOnly $(if $(NATIVE_CPU),-NativeCpu "$(NATIVE_CPU)",) $(if $(VS_INSTALL),-VsInstall "$(VS_INSTALL)",)
 
 cpu:
 	$(MAKE) -f "$(THIS_MAKEFILE)" build PRESET=windows-cpu-release
@@ -52,5 +53,6 @@ help:
 	@echo Examples:
 	@echo   make -f Makefile.windows cpu JOBS=8
 	@echo   make -f Makefile.windows cuda JOBS=8
+	@echo   make -f Makefile.windows cuda NATIVE_CPU=OFF JOBS=8
 	@echo   make -f Makefile.windows cuda-debug TARGET=audiocpp_cli JOBS=8
 	@echo   make -f Makefile.windows cuda VS_INSTALL="C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools"

README.md

@@ -93,6 +93,8 @@ That default configure is a CPU build unless you enable an accelerator backend e
 
 Use GCC 13 or newer for Linux builds.
 
+Native ggml CPU optimization is enabled by default for local performance. If your compiler or assembler rejects a generated CPU instruction such as `vpdpbusd`, reconfigure with `-DENGINE_ENABLE_NATIVE_CPU=OFF` to build portable CPU kernels.
+
 Common Linux configure examples:
 
 CPU-only:
@@ -113,6 +115,12 @@ Vulkan:
 cmake -S . -B build -DENGINE_ENABLE_VULKAN=ON
 ```
 
+Portable CPU-kernel fallback:
+
+```bash
+cmake -S . -B build -DENGINE_ENABLE_NATIVE_CPU=OFF
+```
+
 Build the CLI and server from the configured tree:
 
 ```bash
@@ -133,6 +141,7 @@ Examples:
 scripts/build_linux.sh --backend cuda --target audiocpp_cli --target audiocpp_server
 scripts/build_linux.sh --backend vulkan --target audiocpp_cli --target audiocpp_server
 scripts/build_linux.sh --backend cpu --target audiocpp_cli --target audiocpp_server
+scripts/build_linux.sh --backend cuda --native-cpu OFF --target audiocpp_cli --target audiocpp_server
 ```
 
 Use `--build-dir <dir>` only when you intentionally want a custom output directory.
@@ -170,16 +179,18 @@ If GNU Make is available on Windows:
 ```bash
 make -f Makefile.windows cpu JOBS=16
 make -f Makefile.windows cuda JOBS=16
+make -f Makefile.windows cuda NATIVE_CPU=OFF JOBS=16
 ```
 
-The Windows script configures `build/windows-cuda-release` by default and builds `audiocpp_cli`. CUDA presets enable CUDA, CUDA graphs, OpenMP, Ninja, `/utf-8`, `/EHsc`, MSVC OpenMP SIMD support with `/openmp:experimental`, and the same portable CPU optimization baseline used for the Windows CUDA path. The CPU preset uses the same MSVC/Ninja/OpenMP setup without requiring CUDA. CUDA presets auto-detect the local GPU CUDA architecture when `nvidia-smi` is available.
+The Windows script configures `build/windows-cuda-release` by default and builds `audiocpp_cli`. CUDA presets enable CUDA, CUDA graphs, OpenMP, Ninja, `/utf-8`, `/EHsc`, MSVC OpenMP SIMD support with `/openmp:experimental`, and native CPU optimization by default. The CPU preset uses the same MSVC/Ninja/OpenMP setup without requiring CUDA. CUDA presets auto-detect the local GPU CUDA architecture when `nvidia-smi` is available. Pass `-NativeCpu OFF` or `NATIVE_CPU=OFF` to use portable CPU kernels.
 
 Useful variants:
 
 ```powershell
 .\scripts\build_windows.ps1 -Target audiocpp_server -Jobs 16
 .\scripts\build_windows.ps1 -Preset windows-cpu-release -Target audiocpp_cli
 .\scripts\build_windows.ps1 -Preset windows-cuda-debug -Target audiocpp_cli
+.\scripts\build_windows.ps1 -NativeCpu OFF -Target audiocpp_cli
 .\scripts\build_windows.ps1 -ConfigureOnly
 .\scripts\build_windows.ps1 -CudaArchitectures 120a-real
 ```
@@ -214,6 +225,7 @@ scripts/build_metal.sh --target audiocpp_server
 scripts/build_metal.sh --build-type Release --archs arm64 --target audiocpp_cli
 scripts/build_metal.sh --with-tests --target audio_dsp_test
 scripts/build_metal.sh --openmp auto --target audiocpp_cli
+scripts/build_metal.sh --native-cpu OFF --target audiocpp_cli
 ```
 
 The built CLI is written to:
@@ -231,6 +243,7 @@ Build options:
 | `ENGINE_ENABLE_METAL` | Enable the ggml Metal backend. Required for `--backend metal`. | `OFF` on most platforms, `ON` on Apple |
 | `ENGINE_ENABLE_LLAMAFILE` | Enable llamafile SGEMM support in ggml CPU builds. | `ON` |
 | `ENGINE_ENABLE_CUDA_GRAPHS` | Enable ggml CUDA graphs support when CUDA is enabled. | `ON` |
+| `ENGINE_ENABLE_NATIVE_CPU` | Build ggml CPU kernels with native host ISA flags such as `-march=native`. Disable this for portable CPU kernels or toolchains that reject generated CPU instructions. | `ON` |
 | `ENGINE_ENABLE_OPENMP` | Enable OpenMP for host-side parallel work. | `ON` |
 | `ENGINE_BUILD_EXAMPLES` | Build example binaries. | `OFF` |
 | `ENGINE_BUILD_TESTS` | Build framework unit tests. | `OFF` |

scripts/build_linux.sh

@@ -10,6 +10,7 @@ VULKAN_MODE="off"
 WITH_TESTS="OFF"
 WITH_EXAMPLES="OFF"
 WITH_WARMBENCH="OFF"
+NATIVE_CPU="ON"
 TARGETS=()
 JOBS=""
 
@@ -63,6 +64,21 @@ while [[ $# -gt 0 ]]; do
             WITH_WARMBENCH="ON"
             shift
             ;;
+        --native-cpu)
+            case "$2" in
+                ON|on|On|1|true|TRUE|yes|YES)
+                    NATIVE_CPU="ON"
+                    ;;
+                OFF|off|Off|0|false|FALSE|no|NO)
+                    NATIVE_CPU="OFF"
+                    ;;
+                *)
+                    echo "--native-cpu must be ON or OFF" >&2
+                    exit 1
+                    ;;
+            esac
+            shift 2
+            ;;
         --target)
             TARGETS+=("$2")
             shift 2
@@ -161,6 +177,7 @@ echo "Using generator: $GENERATOR"
 echo "Using build dir: $BUILD_DIR"
 echo "Including CUDA backend: $ENGINE_ENABLE_CUDA"
 echo "Including Vulkan backend: $ENGINE_ENABLE_VULKAN"
+echo "Native CPU optimization: $NATIVE_CPU"
 echo "Building examples: $WITH_EXAMPLES"
 echo "Building tests: $WITH_TESTS"
 echo "Building warmbench: $WITH_WARMBENCH"
@@ -172,6 +189,7 @@ echo "Building warmbench: $WITH_WARMBENCH"
     -DCMAKE_BUILD_TYPE="$BUILD_TYPE" \
     -DENGINE_ENABLE_CUDA="$ENGINE_ENABLE_CUDA" \
     -DENGINE_ENABLE_VULKAN="$ENGINE_ENABLE_VULKAN" \
+    -DENGINE_ENABLE_NATIVE_CPU="$NATIVE_CPU" \
     -DENGINE_BUILD_EXAMPLES="$WITH_EXAMPLES" \
     -DENGINE_BUILD_TESTS="$WITH_TESTS" \
     -DENGINE_BUILD_WARMBENCH="$WITH_WARMBENCH"

scripts/build_metal.sh

@@ -12,6 +12,7 @@ WITH_EXAMPLES="OFF"
 WITH_WARMBENCH="OFF"
 OPENMP_MODE="off"
 LLAMAFILE="ON"
+NATIVE_CPU="ON"
 METAL_EMBED_LIBRARY="ON"
 DEPLOYMENT_TARGET=""
 ARCHS=""
@@ -36,6 +37,8 @@ Options:
                            Default: OFF
   --llamafile ON|OFF       Enable ggml llamafile SGEMM.
                            Default: ON
+  --native-cpu ON|OFF      Build ggml CPU kernels with native host ISA flags.
+                           Default: ON
   --embed-metal ON|OFF     Embed ggml Metal shader library.
                            Default: ON
   --with-tests             Build framework unit tests.
@@ -95,6 +98,10 @@ while [[ $# -gt 0 ]]; do
             LLAMAFILE="$(normalize_on_off --llamafile "$2")"
             shift 2
             ;;
+        --native-cpu)
+            NATIVE_CPU="$(normalize_on_off --native-cpu "$2")"
+            shift 2
+            ;;
         --embed-metal)
             METAL_EMBED_LIBRARY="$(normalize_on_off --embed-metal "$2")"
             shift 2
@@ -194,6 +201,7 @@ CMAKE_CMD=(
     -DENGINE_ENABLE_METAL=ON
     -DENGINE_ENABLE_OPENMP="$ENGINE_ENABLE_OPENMP"
     -DENGINE_ENABLE_LLAMAFILE="$LLAMAFILE"
+    -DENGINE_ENABLE_NATIVE_CPU="$NATIVE_CPU"
     -DGGML_OPENMP="$ENGINE_ENABLE_OPENMP"
     -DGGML_METAL_EMBED_LIBRARY="$METAL_EMBED_LIBRARY"
     -DENGINE_BUILD_TESTS="$WITH_TESTS"
@@ -225,6 +233,7 @@ echo "Including Metal backend: ON"
 echo "Embedding Metal library: $METAL_EMBED_LIBRARY"
 echo "Including OpenMP: $ENGINE_ENABLE_OPENMP"
 echo "Including llamafile: $LLAMAFILE"
+echo "Native CPU optimization: $NATIVE_CPU"
 echo "Building examples: $WITH_EXAMPLES"
 echo "Building tests: $WITH_TESTS"
 echo "Building warmbench: $WITH_WARMBENCH"

scripts/build_windows.ps1

@@ -6,6 +6,8 @@ param(
     [switch]$ConfigureOnly,
     [switch]$Clean,
     [string]$CudaArchitectures = "auto",
+    [ValidateSet("ON", "OFF")]
+    [string]$NativeCpu = $null,
     [string]$VsInstall = ""
 )
 
@@ -258,7 +260,7 @@ function Get-PresetSettings {
             return @{
                 BuildType = "Release"
                 BuildTests = "OFF"
-                Native = "OFF"
+                Native = "ON"
                 EnableCuda = "OFF"
                 EnableCudaGraphs = "OFF"
                 CFlagsDebug = ""
@@ -269,7 +271,7 @@ function Get-PresetSettings {
             return @{
                 BuildType = "Debug"
                 BuildTests = "ON"
-                Native = "OFF"
+                Native = "ON"
                 EnableCuda = "ON"
                 EnableCudaGraphs = "ON"
                 CFlagsDebug = "/O2 /Zi"
@@ -280,7 +282,7 @@ function Get-PresetSettings {
             return @{
                 BuildType = "Release"
                 BuildTests = "OFF"
-                Native = "OFF"
+                Native = "ON"
                 EnableCuda = "ON"
                 EnableCudaGraphs = "ON"
                 CFlagsDebug = ""
@@ -305,6 +307,9 @@ function Get-PresetSettings {
 }
 
 $settings = Get-PresetSettings $Preset
+if ($null -ne $NativeCpu) {
+    $settings.Native = $NativeCpu
+}
 $isCudaPreset = $settings.EnableCuda -eq "ON"
 
 if ($isCudaPreset) {
@@ -350,6 +355,7 @@ Write-Host "Windows SDK: $(Split-Path $mt -Parent)"
 if ($arch -ne "") {
     Write-Host "CUDA architectures: $arch"
 }
+Write-Host "Native CPU optimization: $($settings.Native)"
 
 if ($Clean) {
     $buildDirForClean = Join-Path (Join-Path (Split-Path $PSScriptRoot -Parent) "build") $Preset
@@ -380,7 +386,7 @@ $configureArgs = @(
     "-DENGINE_ENABLE_VULKAN=OFF",
     "-DENGINE_ENABLE_METAL=OFF",
     "-DGGML_OPENMP=ON",
-    "-DGGML_NATIVE=$($settings.Native)",
+    "-DENGINE_ENABLE_NATIVE_CPU=$($settings.Native)",
     "-DENGINE_BUILD_TESTS=$($settings.BuildTests)"
 )
 if ($settings.CFlagsDebug -ne "") {

scripts/build_xcframework.sh

@@ -13,6 +13,7 @@ DEPLOYMENT_TARGET="13.3"
 JOBS="$(sysctl -n hw.logicalcpu 2>/dev/null || echo 8)"
 CLEAN="OFF"
 LLAMAFILE="ON"
+NATIVE_CPU="ON"
 
 usage() {
     cat <<'EOF'
@@ -36,6 +37,8 @@ Options:
   --clean               Remove intermediate/output directories first.
   --llamafile ON|OFF    Enable ggml llamafile SGEMM.
                         Default: ON
+  --native-cpu ON|OFF   Build ggml CPU kernels with native host ISA flags.
+                        Default: ON
   -j, --jobs <n>        Parallel build jobs.
   -h, --help            Show this help.
 
@@ -80,6 +83,10 @@ while [[ $# -gt 0 ]]; do
             LLAMAFILE="$2"
             shift 2
             ;;
+        --native-cpu)
+            NATIVE_CPU="$2"
+            shift 2
+            ;;
         -j|--jobs)
             JOBS="$2"
             shift 2
@@ -118,6 +125,16 @@ case "$LLAMAFILE" in
         ;;
 esac
 
+case "$NATIVE_CPU" in
+    ON|OFF) ;;
+    on) NATIVE_CPU="ON" ;;
+    off) NATIVE_CPU="OFF" ;;
+    *)
+        echo "--native-cpu must be ON or OFF" >&2
+        exit 1
+        ;;
+esac
+
 ARCHS="${ARCHS//,/ }"
 read -r -a ARCH_LIST <<< "$ARCHS"
 if [[ ${#ARCH_LIST[@]} -eq 0 ]]; then
@@ -165,6 +182,7 @@ archive_for_arch() {
         -DENGINE_ENABLE_VULKAN=OFF \
         -DENGINE_ENABLE_METAL=ON \
         -DENGINE_ENABLE_LLAMAFILE="$LLAMAFILE" \
+        -DENGINE_ENABLE_NATIVE_CPU="$NATIVE_CPU" \
         -DGGML_METAL_EMBED_LIBRARY=ON \
         -DENGINE_BUILD_TESTS=OFF \
         -DENGINE_BUILD_EXAMPLES=OFF