add GGML_AVX_VNNI to enable avx-vnni, fix checks

Author: slaren

.devops/llama-server.Dockerfile

@@ -15,7 +15,7 @@ RUN \
     scripts/build-cpu.sh avx         -DGGML_AVX=ON -DGGML_AVX2=OFF && \
     scripts/build-cpu.sh avx2        -DGGML_AVX=ON -DGGML_AVX2=ON && \
     scripts/build-cpu.sh avx512      -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON && \
-    scripts/build-cpu.sh amx         -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON -DGGML_AVX512_VNNI=ON -DGGML_AMX_TILE=ON -DGGML_AMX_INT8=ON && \
+    scripts/build-cpu.sh amx         -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON -DGGML_AVX_VNNI=ON -DGGML_AVX512_VNNI=ON -DGGML_AMX_TILE=ON -DGGML_AMX_INT8=ON && \
     # Build llama-server
     cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \
     cmake --build build --target llama-server -j $(nproc) && \

CMakeLists.txt

@@ -96,10 +96,6 @@ if (NOT DEFINED GGML_LLAMAFILE)
     set(GGML_LLAMAFILE_DEFAULT ON)
 endif()
 
-if (NOT DEFINED GGML_AMX)
-    set(GGML_AMX ON)
-endif()
-
 if (NOT DEFINED GGML_CUDA_GRAPHS)
     set(GGML_CUDA_GRAPHS_DEFAULT ON)
 endif()

ggml/CMakeLists.txt

@@ -96,6 +96,7 @@ 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)

ggml/src/ggml-cpu/CMakeLists.txt

@@ -217,6 +217,12 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
         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)
+            endif()
+        endif()
     else()
         if (GGML_NATIVE)
             list(APPEND ARCH_FLAGS -march=native)
@@ -233,6 +239,9 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
         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)

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

@@ -281,8 +281,8 @@ static int ggml_backend_cpu_x86_score() {
     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        () * 1<<5;
     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

ggml/src/ggml-cpu/ggml-cpu-aarch64.c

@@ -128,7 +128,7 @@ static inline __m512i sum_i16_pairs_int_32x16(const __m512i x) {
 }
 
 static inline __m512i mul_sum_us8_pairs_int32x16(const __m512i ax, const __m512i sy) {
-#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+#if defined(__AVX512VNNI__)
     const __m512i zero = _mm512_setzero_si512();
     return _mm512_dpbusd_epi32(zero, ax, sy);
 #else