ggml : add predefined list of CPU backend variants to build (llama/10626)

* ggml : add predefined list of CPU backend variants to build

* update CPU dockerfiles

ggml/CMakeLists.txt

@@ -92,30 +92,33 @@ else()
     set(INS_ENB ON)
 endif()
 
-option(GGML_CPU_HBM     "ggml: use memkind for CPU HBM" OFF)
-option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
-
-option(GGML_AVX         "ggml: enable AVX"              ${INS_ENB})
-option(GGML_AVX_VNNI    "ggml: enable AVX-VNNI"         OFF)
-option(GGML_AVX2        "ggml: enable AVX2"             ${INS_ENB})
-option(GGML_AVX512      "ggml: enable AVX512"           OFF)
-option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI"      OFF)
-option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI"      OFF)
-option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16"      OFF)
-option(GGML_AMX_TILE    "ggml: enable AMX-TILE"         OFF)
-option(GGML_AMX_INT8    "ggml: enable AMX-INT8"         OFF)
-option(GGML_AMX_BF16    "ggml: enable AMX-BF16"         OFF)
-option(GGML_FMA         "ggml: enable FMA"              ${INS_ENB})
+option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
+option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
+option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
+option(GGML_AVX_VNNI         "ggml: enable AVX-VNNI"         OFF)
+option(GGML_AVX2             "ggml: enable AVX2"             ${INS_ENB})
+option(GGML_AVX512           "ggml: enable AVX512F"          OFF)
+option(GGML_AVX512_VBMI      "ggml: enable AVX512-VBMI"      OFF)
+option(GGML_AVX512_VNNI      "ggml: enable AVX512-VNNI"      OFF)
+option(GGML_AVX512_BF16      "ggml: enable AVX512-BF16"      OFF)
 if (NOT MSVC)
-    option(GGML_F16C    "ggml: enable F16C"             ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512
+    # in MSVC F16C and FMA is implied with AVX2/AVX512
+    option(GGML_FMA          "ggml: enable FMA"              ${INS_ENB})
+    option(GGML_F16C         "ggml: enable F16C"             ${INS_ENB})
+    # MSVC does not seem to support AMX
+    option(GGML_AMX_TILE     "ggml: enable AMX-TILE"         OFF)
+    option(GGML_AMX_INT8     "ggml: enable AMX-INT8"         OFF)
+    option(GGML_AMX_BF16     "ggml: enable AMX-BF16"         OFF)
 endif()
-option(GGML_LASX        "ggml: enable lasx"             ON)
-option(GGML_LSX         "ggml: enable lsx"              ON)
-option(GGML_RVV         "ggml: enable rvv"              ON)
-option(GGML_SVE         "ggml: enable SVE"              OFF)
+option(GGML_LASX             "ggml: enable lasx"             ON)
+option(GGML_LSX              "ggml: enable lsx"              ON)
+option(GGML_RVV              "ggml: enable rvv"              ON)
+option(GGML_SVE              "ggml: enable SVE"              OFF)
+option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
+
 
 if (WIN32)
-    set(GGML_WIN_VER "0x602" CACHE STRING "ggml: Windows Version")
+    set(GGML_WIN_VER "0x602" CACHE STRING   "ggml: Windows version")
 endif()
 
 # ggml core
@@ -180,11 +183,7 @@ option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 
-if (GGML_SYCL)
-    set(CMAKE_CXX_STANDARD 17)
-else()
-    set(CMAKE_CXX_STANDARD 11)
-endif()
+set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED true)
 
 set(THREADS_PREFER_PTHREAD_FLAG ON)

ggml/src/CMakeLists.txt

@@ -269,7 +269,42 @@ function(ggml_add_backend backend)
     endif()
 endfunction()
 
+function(ggml_add_cpu_backend_variant tag_name)
+    set(GGML_CPU_TAG_NAME ${tag_name})
+    # other: OPENMP LLAMAFILE CPU_HBM
+    foreach (feat NATIVE
+                  AVX AVX2 AVX_VNNI FMA F16C
+                  AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
+                  AMX_TILE AMX_INT8 AMX_BF16)
+        set(GGML_${feat} OFF)
+    endforeach()
+
+    foreach (feat ${ARGN})
+        set(GGML_${feat} ON)
+    endforeach()
+
+    ggml_add_cpu_backend_variant_impl(${tag_name})
+endfunction()
+
 ggml_add_backend(CPU)
+
+if (GGML_CPU_ALL_VARIANTS)
+    if (NOT GGML_BACKEND_DL)
+        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
+    endif()
+    ggml_add_cpu_backend_variant(sandybridge    AVX)
+    ggml_add_cpu_backend_variant(haswell        AVX F16C AVX2 FMA)
+    ggml_add_cpu_backend_variant(skylakex       AVX F16C AVX2 FMA AVX512)
+    ggml_add_cpu_backend_variant(icelake        AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
+    if (NOT MSVC)
+        # MSVC doesn't support AVX-VNNI or AMX
+        ggml_add_cpu_backend_variant(alderlake      AVX F16C AVX2 FMA AVX_VNNI)
+        ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+    endif()
+else ()
+    ggml_add_cpu_backend_variant_impl("")
+endif()
+
 ggml_add_backend(BLAS)
 ggml_add_backend(CANN)
 ggml_add_backend(CUDA)

ggml/src/ggml-backend-reg.cpp

@@ -483,6 +483,10 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent)
                                 best_score = s;
                                 best_path = entry.path().string();
                             }
+                        } else {
+                            if (!silent) {
+                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, entry.path().string().c_str());
+                            }
                         }
                     }
                 }
@@ -505,15 +509,21 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent)
 }
 
 void ggml_backend_load_all() {
-    ggml_backend_load_best("blas", true);
-    ggml_backend_load_best("cann", true);
-    ggml_backend_load_best("cuda", true);
-    ggml_backend_load_best("hip", true);
-    ggml_backend_load_best("kompute", true);
-    ggml_backend_load_best("metal", true);
-    ggml_backend_load_best("rpc", true);
-    ggml_backend_load_best("sycl", true);
-    ggml_backend_load_best("vulkan", true);
-    ggml_backend_load_best("musa", true);
-    ggml_backend_load_best("cpu", true);
+#ifdef NDEBUG
+    bool silent = true;
+#else
+    bool silent = false;
+#endif
+
+    ggml_backend_load_best("blas", silent);
+    ggml_backend_load_best("cann", silent);
+    ggml_backend_load_best("cuda", silent);
+    ggml_backend_load_best("hip", silent);
+    ggml_backend_load_best("kompute", silent);
+    ggml_backend_load_best("metal", silent);
+    ggml_backend_load_best("rpc", silent);
+    ggml_backend_load_best("sycl", silent);
+    ggml_backend_load_best("vulkan", silent);
+    ggml_backend_load_best("musa", silent);
+    ggml_backend_load_best("cpu", silent);
 }

ggml/src/ggml-cpu/CMakeLists.txt

@@ -1,319 +1,354 @@
-ggml_add_backend_library(ggml-cpu)
-
-list (APPEND GGML_CPU_SOURCES
-    ggml-cpu.c
-    ggml-cpu.cpp
-    ggml-cpu-aarch64.c
-    ggml-cpu-aarch64.h
-    ggml-cpu-quants.c
-    ggml-cpu-quants.h
-    amx/amx.cpp
-    amx/amx.h
-    amx/mmq.cpp
-    amx/mmq.h
-    ggml-cpu-impl.h
-    )
-
-target_compile_features(ggml-cpu PRIVATE c_std_11 cxx_std_17)
-target_include_directories(ggml-cpu PRIVATE .)
-
-if (APPLE AND GGML_ACCELERATE)
-    find_library(ACCELERATE_FRAMEWORK Accelerate)
-    if (ACCELERATE_FRAMEWORK)
-        message(STATUS "Accelerate framework found")
-
-        target_compile_definitions(ggml-cpu PRIVATE GGML_USE_ACCELERATE)
-        target_compile_definitions(ggml-cpu PRIVATE ACCELERATE_NEW_LAPACK)
-        target_compile_definitions(ggml-cpu PRIVATE ACCELERATE_LAPACK_ILP64)
-
-        target_link_libraries(ggml-cpu PRIVATE ${ACCELERATE_FRAMEWORK})
+function(ggml_add_cpu_backend_variant_impl tag_name)
+    if (tag_name)
+        set(GGML_CPU_NAME ggml-cpu-${tag_name})
     else()
-        message(WARNING "Accelerate framework not found")
+        set(GGML_CPU_NAME ggml-cpu)
     endif()
-endif()
 
-if (GGML_OPENMP)
-    find_package(OpenMP)
-    if (OpenMP_FOUND)
-        message(STATUS "OpenMP found")
+    ggml_add_backend_library(${GGML_CPU_NAME})
+
+    list (APPEND GGML_CPU_SOURCES
+        ggml-cpu/ggml-cpu.c
+        ggml-cpu/ggml-cpu.cpp
+        ggml-cpu/ggml-cpu-aarch64.c
+        ggml-cpu/ggml-cpu-aarch64.h
+        ggml-cpu/ggml-cpu-quants.c
+        ggml-cpu/ggml-cpu-quants.h
+        ggml-cpu/amx/amx.cpp
+        ggml-cpu/amx/amx.h
+        ggml-cpu/amx/mmq.cpp
+        ggml-cpu/amx/mmq.h
+        ggml-cpu/ggml-cpu-impl.h
+        )
+
+    target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17)
+    target_include_directories(${GGML_CPU_NAME} PRIVATE . ggml-cpu)
+
+    if (APPLE AND GGML_ACCELERATE)
+        find_library(ACCELERATE_FRAMEWORK Accelerate)
+        if (ACCELERATE_FRAMEWORK)
+            message(STATUS "Accelerate framework found")
+
+            target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_ACCELERATE)
+            target_compile_definitions(${GGML_CPU_NAME} PRIVATE ACCELERATE_NEW_LAPACK)
+            target_compile_definitions(${GGML_CPU_NAME} PRIVATE ACCELERATE_LAPACK_ILP64)
+
+            target_link_libraries(${GGML_CPU_NAME} PRIVATE ${ACCELERATE_FRAMEWORK})
+        else()
+            message(WARNING "Accelerate framework not found")
+        endif()
+    endif()
 
-        target_compile_definitions(ggml-cpu PRIVATE GGML_USE_OPENMP)
+    if (GGML_OPENMP)
+        find_package(OpenMP)
+        if (OpenMP_FOUND)
+            target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
 
-        target_link_libraries(ggml-cpu PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
-    else()
-        message(WARNING "OpenMP not found")
+            target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+        else()
+            message(WARNING "OpenMP not found")
+        endif()
     endif()
-endif()
-
-if (GGML_LLAMAFILE)
-    message(STATUS "Using llamafile")
 
-    target_compile_definitions(ggml-cpu PRIVATE GGML_USE_LLAMAFILE)
+    if (GGML_LLAMAFILE)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_LLAMAFILE)
 
-    list(APPEND GGML_CPU_SOURCES
-                llamafile/sgemm.cpp
-                llamafile/sgemm.h)
-endif()
+        list(APPEND GGML_CPU_SOURCES
+                    ggml-cpu/llamafile/sgemm.cpp
+                    ggml-cpu/llamafile/sgemm.h)
+    endif()
 
-if (GGML_CPU_HBM)
-    find_library(memkind memkind REQUIRED)
+    if (GGML_CPU_HBM)
+        find_library(memkind memkind REQUIRED)
 
-    message(STATUS "Using memkind for CPU HBM")
+        message(STATUS "Using memkind for CPU HBM")
 
-    target_compile_definitions(ggml-cpu PRIVATE GGML_USE_CPU_HBM)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_HBM)
 
-    target_link_libraries(ggml-cpu PUBLIC memkind)
-endif()
+        target_link_libraries(${GGML_CPU_NAME} PUBLIC memkind)
+    endif()
 
-if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
-    CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
-    (NOT CMAKE_OSX_ARCHITECTURES      AND
-     NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-         CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
+    if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
+        CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
+        (NOT CMAKE_OSX_ARCHITECTURES      AND
+        NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+            CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
 
-    message(STATUS "ARM detected")
+        message(STATUS "ARM detected")
 
-    if (MSVC)
-        list(APPEND ARCH_DEFINITIONS __aarch64__) # MSVC defines _M_ARM64 instead
-        list(APPEND ARCH_DEFINITIONS __ARM_NEON)
-        list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FMA)
+        if (MSVC)
+            list(APPEND ARCH_DEFINITIONS __aarch64__) # MSVC defines _M_ARM64 instead
+            list(APPEND ARCH_DEFINITIONS __ARM_NEON)
+            list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FMA)
 
-        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
-        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
+            set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
+            string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
 
-        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
-        if (GGML_COMPILER_SUPPORT_DOTPROD)
-            list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+            if (GGML_COMPILER_SUPPORT_DOTPROD)
+                list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
 
-            message(STATUS "ARM feature DOTPROD enabled")
-        endif ()
+                message(STATUS "ARM feature DOTPROD enabled")
+            endif ()
 
-        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
 
-        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
-            list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
+            if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+                list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
 
-            message(STATUS "ARM feature MATMUL_INT8 enabled")
-        endif ()
+                message(STATUS "ARM feature MATMUL_INT8 enabled")
+            endif ()
 
-        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-            list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+            if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+                list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
 
-            message(STATUS "ARM feature FP16_VECTOR_ARITHMETIC enabled")
-        endif ()
+                message(STATUS "ARM feature FP16_VECTOR_ARITHMETIC enabled")
+            endif ()
 
-        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
-    elseif (APPLE)
-        if (GGML_NATIVE)
-            set(USER_PROVIDED_MARCH FALSE)
-            foreach(flag_var IN ITEMS CMAKE_C_FLAGS CMAKE_CXX_FLAGS CMAKE_REQUIRED_FLAGS)
-                if ("${${flag_var}}" MATCHES "-march=[a-zA-Z0-9+._-]+")
-                    set(USER_PROVIDED_MARCH TRUE)
-                    break()
-                endif()
-            endforeach()
+            set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
+        elseif (APPLE)
+            if (GGML_NATIVE)
+                set(USER_PROVIDED_MARCH FALSE)
+                foreach(flag_var IN ITEMS CMAKE_C_FLAGS CMAKE_CXX_FLAGS CMAKE_REQUIRED_FLAGS)
+                    if ("${${flag_var}}" MATCHES "-march=[a-zA-Z0-9+._-]+")
+                        set(USER_PROVIDED_MARCH TRUE)
+                        break()
+                    endif()
+                endforeach()
 
-            if (NOT USER_PROVIDED_MARCH)
-                set(MARCH_FLAGS "-march=armv8.2a")
+                if (NOT USER_PROVIDED_MARCH)
+                    set(MARCH_FLAGS "-march=armv8.2a")
 
-                check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
-                if (GGML_COMPILER_SUPPORT_DOTPROD)
-                    set(MARCH_FLAGS "${MARCH_FLAGS}+dotprod")
-                    list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
+                    check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+                    if (GGML_COMPILER_SUPPORT_DOTPROD)
+                        set(MARCH_FLAGS "${MARCH_FLAGS}+dotprod")
+                        list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
 
-                    message(STATUS "ARM feature DOTPROD enabled")
-                endif ()
+                        message(STATUS "ARM feature DOTPROD enabled")
+                    endif ()
 
-                set(TEST_I8MM_FLAGS "-march=armv8.2a+i8mm")
+                    set(TEST_I8MM_FLAGS "-march=armv8.2a+i8mm")
 
-                set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
-                set(CMAKE_REQUIRED_FLAGS     "${CMAKE_REQUIRED_FLAGS} ${TEST_I8MM_FLAGS}")
+                    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+                    set(CMAKE_REQUIRED_FLAGS     "${CMAKE_REQUIRED_FLAGS} ${TEST_I8MM_FLAGS}")
 
-                check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
-                if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
-                    set(MARCH_FLAGS "${MARCH_FLAGS}+i8mm")
-                    list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
+                    check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+                    if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+                        set(MARCH_FLAGS "${MARCH_FLAGS}+i8mm")
+                        list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
 
-                    message(STATUS "ARM feature MATMUL_INT8 enabled")
-                endif ()
+                        message(STATUS "ARM feature MATMUL_INT8 enabled")
+                    endif ()
 
-                set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+                    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
 
-                list(APPEND ARCH_FLAGS "${MARCH_FLAGS}")
+                    list(APPEND ARCH_FLAGS "${MARCH_FLAGS}")
+                endif ()
             endif ()
-        endif ()
-    else()
-        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
-        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
-            list(APPEND ARCH_FLAGS -mfp16-format=ieee)
-        endif()
-        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
-            # Raspberry Pi 1, Zero
-            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
-        endif()
-        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
-                # Android armeabi-v7a
-                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
-            else()
-                # Raspberry Pi 2
-                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+        else()
+            check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+            if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+                list(APPEND ARCH_FLAGS -mfp16-format=ieee)
             endif()
-        endif()
-        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
-            # Android arm64-v8a
-            # Raspberry Pi 3, 4, Zero 2 (32-bit)
-            list(APPEND ARCH_FLAGS -mno-unaligned-access)
-        endif()
-        if (GGML_SVE)
-            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
-        endif()
-    endif()
-elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
-        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
-    message(STATUS "x86 detected")
-    if (MSVC)
-        # instruction set detection for MSVC only
-        if (GGML_NATIVE)
-            include(cmake/FindSIMD.cmake)
-        endif ()
-        if (GGML_AVX512)
-            list(APPEND ARCH_FLAGS /arch:AVX512)
-            # MSVC has no compile-time flags enabling specific
-            # AVX512 extensions, neither it defines the
-            # macros corresponding to the extensions.
-            # Do it manually.
-            if (GGML_AVX512_VBMI)
-                list(APPEND ARCH_DEFINITIONS __AVX512VBMI__)
-                if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
-                    list(APPEND ARCH_FLAGS -mavx512vbmi)
-                endif()
+            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
+                # Raspberry Pi 1, Zero
+                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
             endif()
-            if (GGML_AVX512_VNNI)
-                list(APPEND ARCH_DEFINITIONS __AVX512VNNI__)
-                if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
-                    list(APPEND ARCH_FLAGS -mavx512vnni)
+            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
+                if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
+                    # Android armeabi-v7a
+                    list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
+                else()
+                    # Raspberry Pi 2
+                    list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
                 endif()
             endif()
-            if (GGML_AVX512_BF16)
-                list(APPEND ARCH_DEFINITIONS __AVX512BF16__)
-                if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
-                    list(APPEND ARCH_FLAGS -mavx512bf16)
-                endif()
+            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
+                # Android arm64-v8a
+                # Raspberry Pi 3, 4, Zero 2 (32-bit)
+                list(APPEND ARCH_FLAGS -mno-unaligned-access)
             endif()
-            if (GGML_AMX_TILE)
-                list(APPEND ARCH_DEFINITIONS __AMX_TILE__)
+            if (GGML_SVE)
+                list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
             endif()
-            if (GGML_AMX_INT8)
-                list(APPEND ARCH_DEFINITIONS __AMX_INT8__)
+        endif()
+    elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+            (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+            CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
+        if (MSVC)
+            # instruction set detection for MSVC only
+            if (GGML_NATIVE)
+                include(ggml-cpu/cmake/FindSIMD.cmake)
+            endif ()
+            if (GGML_AVX512)
+                list(APPEND ARCH_FLAGS /arch:AVX512)
+                # /arch:AVX512 includes: __AVX512F__, __AVX512CD__, __AVX512BW__, __AVX512DQ__, and __AVX512VL__
+                # MSVC has no compile-time flags enabling specific
+                # AVX512 extensions, neither it defines the
+                # macros corresponding to the extensions.
+                # Do it manually.
+                list(APPEND ARCH_DEFINITIONS GGML_AVX512)
+                if (GGML_AVX512_VBMI)
+                    list(APPEND ARCH_DEFINITIONS __AVX512VBMI__)
+                    if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                        list(APPEND ARCH_FLAGS -mavx512vbmi)
+                    endif()
+                endif()
+                if (GGML_AVX512_VNNI)
+                    list(APPEND ARCH_DEFINITIONS __AVX512VNNI__ GGML_AVX512_VNNI)
+                    if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                        list(APPEND ARCH_FLAGS -mavx512vnni)
+                    endif()
+                endif()
+                if (GGML_AVX512_BF16)
+                    list(APPEND ARCH_DEFINITIONS __AVX512BF16__ GGML_AVX512_BF16)
+                    if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                        list(APPEND ARCH_FLAGS -mavx512bf16)
+                    endif()
+                endif()
+                if (GGML_AMX_TILE)
+                    list(APPEND ARCH_DEFINITIONS __AMX_TILE__ GGML_AMX_TILE)
+                endif()
+                if (GGML_AMX_INT8)
+                    list(APPEND ARCH_DEFINITIONS __AMX_INT8__ GGML_AMX_INT8)
+                endif()
+                if (GGML_AMX_BF16)
+                    list(APPEND ARCH_DEFINITIONS __AMX_BF16__ GGML_AMX_BF16)
+                endif()
+            elseif (GGML_AVX2)
+                list(APPEND ARCH_FLAGS /arch:AVX2)
+                list(APPEND ARCH_DEFINITIONS GGML_AVX2 GGML_FMA GGML_F16C)
+            elseif (GGML_AVX)
+                list(APPEND ARCH_FLAGS /arch:AVX)
+                list(APPEND ARCH_DEFINITIONS GGML_AVX)
+            else ()
+                list(APPEND ARCH_FLAGS /arch:SSE4.2)
+                list(APPEND ARCH_DEFINITIONS GGML_SSE42)
             endif()
-            if (GGML_AMX_BF16)
-                list(APPEND ARCH_DEFINITIONS __AMX_BF16__)
+            if (GGML_AVX_VNNI)
+                # MSVC generates AVX512 with AVX-VNNI intrinsics even with /arch:AVX2
+                #list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
             endif()
-        elseif (GGML_AVX2)
-            list(APPEND ARCH_FLAGS /arch:AVX2)
-        elseif (GGML_AVX)
-            list(APPEND ARCH_FLAGS /arch:AVX)
-        endif()
-        if (GGML_AVX_VNNI)
-            list(APPEND ARCH_DEFINITIONS __AVXVNNI__)
-            if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
-                list(APPEND ARCH_FLAGS -mavxvnni)
+        else ()
+            if (GGML_NATIVE)
+                list(APPEND ARCH_FLAGS -march=native)
+            else ()
+                list(APPEND ARCH_FLAGS -msse4.2)
+                list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+                if (GGML_F16C)
+                    list(APPEND ARCH_FLAGS -mf16c)
+                    list(APPEND ARCH_DEFINITIONS GGML_F16C)
+                endif()
+                if (GGML_FMA)
+                    list(APPEND ARCH_FLAGS -mfma)
+                    list(APPEND ARCH_DEFINITIONS GGML_FMA)
+                endif()
+                if (GGML_AVX)
+                    list(APPEND ARCH_FLAGS -mavx)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX)
+                endif()
+                if (GGML_AVX2)
+                    list(APPEND ARCH_FLAGS -mavx2)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX2)
+                endif()
+                if (GGML_AVX_VNNI)
+                    list(APPEND ARCH_FLAGS -mavxvnni)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX_VNNI)
+                endif()
+                if (GGML_AVX512)
+                    list(APPEND ARCH_FLAGS -mavx512f)
+                    list(APPEND ARCH_FLAGS -mavx512cd)
+                    list(APPEND ARCH_FLAGS -mavx512vl)
+                    list(APPEND ARCH_FLAGS -mavx512dq)
+                    list(APPEND ARCH_FLAGS -mavx512bw)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX512)
+                endif()
+                if (GGML_AVX512_VBMI)
+                    list(APPEND ARCH_FLAGS -mavx512vbmi)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX512_VBMI)
+                endif()
+                if (GGML_AVX512_VNNI)
+                    list(APPEND ARCH_FLAGS -mavx512vnni)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX512_VNNI)
+                endif()
+                if (GGML_AVX512_BF16)
+                    list(APPEND ARCH_FLAGS -mavx512bf16)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX512_BF16)
+                endif()
+                if (GGML_AMX_TILE)
+                    list(APPEND ARCH_FLAGS -mamx-tile)
+                    list(APPEND ARCH_DEFINITIONS GGML_AMX_TILE)
+                endif()
+                if (GGML_AMX_INT8)
+                    list(APPEND ARCH_FLAGS -mamx-int8)
+                    list(APPEND ARCH_DEFINITIONS GGML_AMX_INT8)
+                endif()
+                if (GGML_AMX_BF16)
+                    list(APPEND ARCH_FLAGS -mamx-bf16)
+                    list(APPEND ARCH_DEFINITIONS GGML_AMX_BF16)
+                endif()
             endif()
         endif()
-    else()
-        if (GGML_NATIVE)
-            list(APPEND ARCH_FLAGS -march=native)
-        endif()
-        if (GGML_F16C)
-            list(APPEND ARCH_FLAGS -mf16c)
-        endif()
-        if (GGML_FMA)
-            list(APPEND ARCH_FLAGS -mfma)
-        endif()
-        if (GGML_AVX)
-            list(APPEND ARCH_FLAGS -mavx)
-        endif()
-        if (GGML_AVX2)
-            list(APPEND ARCH_FLAGS -mavx2)
-        endif()
-        if (GGML_AVX_VNNI)
-            list(APPEND ARCH_FLAGS -mavxvnni)
-        endif()
-        if (GGML_AVX512)
-            list(APPEND ARCH_FLAGS -mavx512f)
-            list(APPEND ARCH_FLAGS -mavx512dq)
-            list(APPEND ARCH_FLAGS -mavx512bw)
-        endif()
-        if (GGML_AVX512_VBMI)
-            list(APPEND ARCH_FLAGS -mavx512vbmi)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+        message(STATUS "PowerPC detected")
+        execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M)
+        string(FIND "${POWER10_M}" "POWER10" substring_index)
+        if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "")
+            set(substring_index -1)
         endif()
-        if (GGML_AVX512_VNNI)
-            list(APPEND ARCH_FLAGS -mavx512vnni)
-        endif()
-        if (GGML_AVX512_BF16)
-            list(APPEND ARCH_FLAGS -mavx512bf16)
+
+        if (${substring_index} GREATER_EQUAL 0)
+        list(APPEND ARCH_FLAGS -mcpu=power10)
+        elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
+        else()
+            list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
+            # TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
         endif()
-        if (GGML_AMX_TILE)
-            list(APPEND ARCH_FLAGS -mamx-tile)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+        message(STATUS "loongarch64 detected")
+
+        list(APPEND ARCH_FLAGS -march=loongarch64)
+        if (GGML_LASX)
+            list(APPEND ARCH_FLAGS -mlasx)
         endif()
-        if (GGML_AMX_INT8)
-            list(APPEND ARCH_FLAGS -mamx-int8)
+        if (GGML_LSX)
+            list(APPEND ARCH_FLAGS -mlsx)
         endif()
-        if (GGML_AMX_BF16)
-            list(APPEND ARCH_FLAGS -mamx-bf16)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
+        message(STATUS "RISC-V detected")
+        if (GGML_RVV)
+            list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
         endif()
-    endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
-    message(STATUS "PowerPC detected")
-    execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M)
-    string(FIND "${POWER10_M}" "POWER10" substring_index)
-    if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "")
-        set(substring_index -1)
-    endif()
-
-    if (${substring_index} GREATER_EQUAL 0)
-       list(APPEND ARCH_FLAGS -mcpu=power10)
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-       list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
     else()
-        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
-        # TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+        message(STATUS "Unknown architecture")
     endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
-    message(STATUS "loongarch64 detected")
 
-    list(APPEND ARCH_FLAGS -march=loongarch64)
-    if (GGML_LASX)
-        list(APPEND ARCH_FLAGS -mlasx)
+    if (GGML_CPU_AARCH64)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
     endif()
-    if (GGML_LSX)
-        list(APPEND ARCH_FLAGS -mlsx)
+
+    message(STATUS "Adding CPU backend variant ${GGML_CPU_NAME}: ${ARCH_FLAGS} ${ARCH_DEFINITIONS}")
+    target_sources(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_SOURCES})
+    target_compile_options(${GGML_CPU_NAME} PRIVATE ${ARCH_FLAGS})
+    target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS})
+
+    if (GGML_BACKEND_DL)
+        # The feature detection code is compiled as a separate target so that
+        # it can be built without the architecture flags
+        # Since multiple variants of the CPU backend may be included in the same
+        # build, using set_source_files_properties() to set the arch flags is not possible
+        set(GGML_CPU_FEATS_NAME ${GGML_CPU_NAME}-feats)
+        add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp)
+        target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
+        target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
+        target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
+        set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_FEATS_NAME})
     endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
-    message(STATUS "RISC-V detected")
-    if (GGML_RVV)
-        list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
+
+    if (EMSCRIPTEN)
+        set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
     endif()
-else()
-    message(STATUS "Unknown architecture")
-endif()
-
-if (GGML_CPU_AARCH64)
-    message(STATUS "Using runtime weight conversion of Q4_0 to Q4_0_x_x to enable optimized GEMM/GEMV kernels")
-    target_compile_definitions(ggml-cpu PRIVATE GGML_USE_CPU_AARCH64)
-endif()
-
-target_sources(ggml-cpu PRIVATE ${GGML_CPU_SOURCES})
-set_source_files_properties(${GGML_CPU_SOURCES} PROPERTIES COMPILE_OPTIONS     "${ARCH_FLAGS}")
-set_source_files_properties(${GGML_CPU_SOURCES} PROPERTIES COMPILE_DEFINITIONS "${ARCH_DEFINITIONS}")
-
-# the feature detection code must be compiled without any architecture flags
-target_sources(ggml-cpu PRIVATE cpu-feats-x86.cpp)
-# target_sources(ggml-cpu PRIVATE cpu-feats-arm.cpp) # TODO: ARM feature detection
-
-if (EMSCRIPTEN)
-    set_target_properties(ggml-cpu PROPERTIES COMPILE_FLAGS "-msimd128")
-endif()
+endfunction()

ggml/src/ggml-cpu/cpu-feats-x86.cpp

@@ -1,4 +1,3 @@
-#include "ggml-cpu.h"
 #include "ggml-backend-impl.h"
 
 #if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
@@ -13,6 +12,7 @@
 #include <array>
 #include <string>
 
+// ref: https://cdrdv2-public.intel.com/782156/325383-sdm-vol-2abcd.pdf
 struct cpuid_x86 {
     bool SSE3(void) { return f_1_ecx[0]; }
     bool PCLMULQDQ(void) { return f_1_ecx[1]; }
@@ -50,11 +50,15 @@ struct cpuid_x86 {
     bool INVPCID(void) { return f_7_ebx[10]; }
     bool RTM(void) { return is_intel && f_7_ebx[11]; }
     bool AVX512F(void) { return f_7_ebx[16]; }
+    bool AVX512DQ(void) { return f_7_ebx[17]; }
     bool RDSEED(void) { return f_7_ebx[18]; }
     bool ADX(void) { return f_7_ebx[19]; }
     bool AVX512PF(void) { return f_7_ebx[26]; }
     bool AVX512ER(void) { return f_7_ebx[27]; }
     bool AVX512CD(void) { return f_7_ebx[28]; }
+    bool AVX512BW(void) { return f_7_ebx[30]; }
+    bool AVX512VL(void) { return f_7_ebx[31]; }
+
     bool SHA(void) { return f_7_ebx[29]; }
 
     bool PREFETCHWT1(void) { return f_7_ecx[0]; }
@@ -259,36 +263,57 @@ void test_x86_is() {
 static int ggml_backend_cpu_x86_score() {
     // FIXME: this does not check for OS support
 
+    int score = 0;
     cpuid_x86 is;
-    // if the CPU backend was built with any features not supported by the current CPU, it cannot be used
-    if (ggml_cpu_has_fma() && !is.FMA()) { return 0; }
-    if (ggml_cpu_has_f16c() && !is.F16C()) { return 0; }
-    if (ggml_cpu_has_ssse3() && !is.SSSE3()) { return 0; }
-    if (ggml_cpu_has_sse3() && !is.SSE3()) { return 0; }
-    if (ggml_cpu_has_avx() && !is.AVX()) { return 0; }
-    if (ggml_cpu_has_avx_vnni() && !is.AVX_VNNI()) { return 0; }
-    if (ggml_cpu_has_avx2() && !is.AVX2()) { return 0; }
-    if (ggml_cpu_has_avx512() && !is.AVX512F()) { return 0; }
-    if (ggml_cpu_has_avx512_vbmi() && !is.AVX512_VBMI()) { return 0; }
-    if (ggml_cpu_has_avx512_bf16() && !is.AVX512_BF16()) { return 0; }
-    if (ggml_cpu_has_avx512_vnni() && !is.AVX512_VNNI()) { return 0; }
-    if (ggml_cpu_has_amx_int8() && !is.AMX_INT8()) { return 0; }
 
-    // calculate a backend score based on the supported features
-    // more important features have a higher weight
-    int score = 0;
-    score +=  ggml_cpu_has_fma        () * 1;
-    score +=  ggml_cpu_has_f16c       () * 1<<1;
-    score +=  ggml_cpu_has_ssse3      () * 1<<2;
-    score +=  ggml_cpu_has_sse3       () * 1<<3;
-    score +=  ggml_cpu_has_avx_vnni   () * 1<<4;
-    score +=  ggml_cpu_has_avx        () * 1<<5;
-    score +=  ggml_cpu_has_avx2       () * 1<<6;
-    score +=  ggml_cpu_has_avx512     () * 1<<7;
-    // score +=  ggml_cpu_has_avx512_vbmi() * 1<<8; // not used
-    score +=  ggml_cpu_has_avx512_bf16() * 1<<9;
-    score +=  ggml_cpu_has_avx512_vnni() * 1<<10;
-    score +=  ggml_cpu_has_amx_int8   () * 1<<11;
+#ifdef GGML_FMA
+    if (!is.FMA()) { return 0; }
+    score += 1;
+#endif
+#ifdef GGML_F16C
+    if (!is.F16C()) { return 0; }
+    score += 1<<1;
+#endif
+#ifdef GGML_SSE42
+    if (!is.SSE42()) { return 0; }
+    score += 1<<2;
+#endif
+#ifdef GGML_AVX
+    if (!is.AVX()) { return 0; }
+    score += 1<<4;
+#endif
+#ifdef GGML_AVX2
+    if (!is.AVX2()) { return 0; }
+    score += 1<<5;
+#endif
+#ifdef GGML_AVX_VNNI
+    if (!is.AVX_VNNI()) { return 0; }
+    score += 1<<6;
+#endif
+#ifdef GGML_AVX512
+    if (!is.AVX512F()) { return 0; }
+    if (!is.AVX512CD()) { return 0; }
+    if (!is.AVX512VL()) { return 0; }
+    if (!is.AVX512DQ()) { return 0; }
+    if (!is.AVX512BW()) { return 0; }
+    score += 1<<7;
+#endif
+#ifdef GGML_AVX512_VBMI
+    if (!is.AVX512_VBMI()) { return 0; }
+    score += 1<<8;
+#endif
+#ifdef GGML_AVX512_BF16
+    if (!is.AVX512_BF16()) { return 0; }
+    score += 1<<9;
+#endif
+#ifdef GGML_AVX512_VNNI
+    if (!is.AVX512_VNNI()) { return 0; }
+    score += 1<<10;
+#endif
+#ifdef GGML_AMX_INT8
+    if (!is.AMX_INT8()) { return 0; }
+    score += 1<<11;
+#endif
 
     return score;
 }

ggml/src/ggml-cpu/ggml-cpu.c

@@ -756,7 +756,7 @@ do {                                                              \
 #define GGML_F32Cx8_LOAD(x)     _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x)))
 #define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))
 #else
-static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
+static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) {
     float tmp[8];
 
     for (int i = 0; i < 8; i++) {

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -641,7 +641,15 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_llamafile()) {
             features.push_back({ "LLAMAFILE", "1" });
         }
-        // TODO: rename this
+    #ifdef GGML_USE_ACCELERATE
+        features.push_back({ "ACCELERATE", "1" });
+    #endif
+    #ifdef GGML_USE_CPU_HBM
+        features.push_back({ "CPU_HBM", "1" });
+    #endif
+    #ifdef GGML_USE_OPENMP
+        features.push_back({ "OPENMP", "1" });
+    #endif
     #ifdef GGML_USE_CPU_AARCH64
         features.push_back({ "AARCH64_REPACK", "1" });
     #endif