build

Tencent · May 23, 2024 · 8ef9f05 · 8ef9f05
1 parent b2da715
commit 8ef9f05
Show file tree

Hide file tree

Showing 3 changed files with 36 additions and 36 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -476,19 +476,19 @@ else()
 
         check_cxx_compiler_flag("/arch:AVX -mxop" NCNN_COMPILER_SUPPORT_X86_XOP)
         check_cxx_compiler_flag("/arch:AVX -mf16c" NCNN_COMPILER_SUPPORT_X86_F16C)
-        check_cxx_compiler_flag("/arch:AVX2" NCNN_COMPILER_SUPPORT_X86_AVX2)
-        check_cxx_compiler_flag("/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl" NCNN_COMPILER_SUPPORT_X86_AVX512)
+        check_cxx_compiler_flag("/arch:AVX2 -mfma -mf16c" NCNN_COMPILER_SUPPORT_X86_AVX2)
+        check_cxx_compiler_flag("/arch:AVX512 -mfma -mf16c -mavx512cd -mavx512bw -mavx512dq -mavx512vl" NCNN_COMPILER_SUPPORT_X86_AVX512)
 
-        set(CMAKE_REQUIRED_FLAGS "/arch:AVX2 -mavxvnni")
+        set(CMAKE_REQUIRED_FLAGS "/arch:AVX2 -mfma -mf16c -mavxvnni")
         check_cxx_source_compiles("#include <immintrin.h>\nint main() { __m256i _s, _a, _b; _s = _mm256_dpwssd_epi32(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI)
 
-        set(CMAKE_REQUIRED_FLAGS "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512vnni")
+        set(CMAKE_REQUIRED_FLAGS "/arch:AVX512 -mfma -mf16c -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512vnni")
         check_cxx_source_compiles("#include <immintrin.h>\nint main() { __m512i _s, _a, _b; _s = _mm512_dpwssd_epi32(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_X86_AVX512_VNNI)
 
-        set(CMAKE_REQUIRED_FLAGS "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512bf16")
+        set(CMAKE_REQUIRED_FLAGS "/arch:AVX512 -mfma -mf16c -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512bf16")
         check_cxx_source_compiles("#include <immintrin.h>\nint main() { __m256bh _s; __m512bh _a, _b; _s = _mm512_cvtneps_pbh(_mm512_dpbf16_ps(_mm512_cvtpbh_ps(_s), _a, _b)); return 0; }" NCNN_COMPILER_SUPPORT_X86_AVX512_BF16)
 
-        set(CMAKE_REQUIRED_FLAGS "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512fp16")
+        set(CMAKE_REQUIRED_FLAGS "/arch:AVX512 -mfma -mf16c -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512fp16")
         check_cxx_source_compiles("#include <immintrin.h>\nint main() { __m512h _s, _a, _b; _s = _mm512_fmadd_ph(_s, _a, _b); __m512 _s2; _s2 = _mm512_cvtxph_ps(_mm512_cvtxps_ph(_s2)); return 0; }" NCNN_COMPILER_SUPPORT_X86_AVX512_FP16)
 
         unset(CMAKE_REQUIRED_FLAGS)

diff --git a/cmake/ncnn_add_layer.cmake b/cmake/ncnn_add_layer.cmake
@@ -167,38 +167,38 @@ macro(ncnn_add_layer class)
             endif()
         elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC")
             if(NCNN_RUNTIME_CPU AND NCNN_AVX512)
-                ncnn_add_arch_opt_layer(${class} avx512 "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl")
+                ncnn_add_arch_opt_layer(${class} avx512 "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__")
             endif()
             if(NCNN_RUNTIME_CPU AND NCNN_FMA)
-                ncnn_add_arch_opt_layer(${class} fma "/arch:AVX -mfma -mf16c")
+                ncnn_add_arch_opt_layer(${class} fma "/arch:AVX -mfma -mf16c /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__")
             endif()
             if(NCNN_RUNTIME_CPU AND NCNN_AVX)
-                ncnn_add_arch_opt_layer(${class} avx "/arch:AVX")
+                ncnn_add_arch_opt_layer(${class} avx "/arch:AVX /D__SSSE3__ /D__SSE4_1__")
             endif()
             if(NCNN_AVX512VNNI)
-                ncnn_add_arch_opt_source(${class} avx512vnni "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512vnni")
+                ncnn_add_arch_opt_source(${class} avx512vnni "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512vnni /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512VNNI__")
             endif()
             if(NCNN_AVX512BF16)
-                ncnn_add_arch_opt_source(${class} avx512bf16 "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512bf16")
+                ncnn_add_arch_opt_source(${class} avx512bf16 "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512bf16 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512BF16__")
             endif()
             if(NCNN_AVX512FP16)
-                ncnn_add_arch_opt_source(${class} avx512fp16 "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512fp16")
+                ncnn_add_arch_opt_source(${class} avx512fp16 "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512fp16 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512FP16__")
             endif()
             if(NCNN_AVXVNNI)
-                ncnn_add_arch_opt_source(${class} avxvnni "/arch:AVX2 -mavxvnni")
+                ncnn_add_arch_opt_source(${class} avxvnni "/arch:AVX2 -mfma -mf16c -mavxvnni /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVXVNNI__")
             endif()
             if(NCNN_AVX2)
-                ncnn_add_arch_opt_source(${class} avx2 "/arch:AVX2")
+                ncnn_add_arch_opt_source(${class} avx2 "/arch:AVX2 -mfma -mf16c /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__")
             endif()
             if(NCNN_XOP)
-                ncnn_add_arch_opt_source(${class} xop "/arch:AVX -mxop")
+                ncnn_add_arch_opt_source(${class} xop "/arch:AVX -mxop /D__SSSE3__ /D__SSE4_1__ /D__XOP__")
             endif()
             if(NCNN_F16C)
-                ncnn_add_arch_opt_source(${class} f16c "/arch:AVX -mf16c")
+                ncnn_add_arch_opt_source(${class} f16c "/arch:AVX -mf16c /D__SSSE3__ /D__SSE4_1__ /D__F16C__")
             endif()
         else()
             if(NCNN_RUNTIME_CPU AND NCNN_AVX512)
-                ncnn_add_arch_opt_layer(${class} avx512 "-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl")
+                ncnn_add_arch_opt_layer(${class} avx512 "-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c")
             endif()
             if(NCNN_RUNTIME_CPU AND NCNN_FMA)
                 ncnn_add_arch_opt_layer(${class} fma "-mavx -mfma -mf16c")
@@ -207,19 +207,19 @@ macro(ncnn_add_layer class)
                 ncnn_add_arch_opt_layer(${class} avx "-mavx")
             endif()
             if(NCNN_AVX512VNNI)
-                ncnn_add_arch_opt_source(${class} avx512vnni "-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512vnni")
+                ncnn_add_arch_opt_source(${class} avx512vnni "-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512vnni")
             endif()
             if(NCNN_AVX512BF16)
-                ncnn_add_arch_opt_source(${class} avx512bf16 "-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512bf16")
+                ncnn_add_arch_opt_source(${class} avx512bf16 "-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512bf16")
             endif()
             if(NCNN_AVX512FP16)
-                ncnn_add_arch_opt_source(${class} avx512fp16 "-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512fp16")
+                ncnn_add_arch_opt_source(${class} avx512fp16 "-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512fp16")
             endif()
             if(NCNN_AVXVNNI)
-                ncnn_add_arch_opt_source(${class} avxvnni "-mavx2 -mavxvnni")
+                ncnn_add_arch_opt_source(${class} avxvnni "-mavx2 -mfma -mf16c -mavxvnni")
             endif()
             if(NCNN_AVX2)
-                ncnn_add_arch_opt_source(${class} avx2 "-mavx2")
+                ncnn_add_arch_opt_source(${class} avx2 "-mavx2 -mfma -mf16c")
             endif()
             if(NCNN_XOP)
                 ncnn_add_arch_opt_source(${class} xop "-mavx -mxop")

diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -409,18 +409,18 @@ if(NCNN_TARGET_ARCH STREQUAL "x86")
                 target_compile_options(ncnn PRIVATE /D__AVX512VNNI__)
             endif()
         elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC")
-            target_compile_options(ncnn PRIVATE /arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl)
+            target_compile_options(ncnn PRIVATE /arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__)
             if(NCNN_AVX512VNNI)
-                target_compile_options(ncnn PRIVATE -mavx512vnni)
+                target_compile_options(ncnn PRIVATE -mavx512vnni /D__AVX512VNNI__)
             endif()
             if(NCNN_AVX512BF16)
-                target_compile_options(ncnn PRIVATE -mavx512bf16)
+                target_compile_options(ncnn PRIVATE -mavx512bf16 /D__AVX512BF16__)
             endif()
             if(NCNN_AVX512FP16)
-                target_compile_options(ncnn PRIVATE -mavx512fp16)
+                target_compile_options(ncnn PRIVATE -mavx512fp16 /D__AVX512FP16__)
             endif()
         else()
-            target_compile_options(ncnn PRIVATE -mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl)
+            target_compile_options(ncnn PRIVATE -mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c)
             if(NCNN_AVX512VNNI)
                 target_compile_options(ncnn PRIVATE -mavx512vnni)
             endif()
@@ -448,21 +448,21 @@ if(NCNN_TARGET_ARCH STREQUAL "x86")
             endif()
         elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC")
             if(NCNN_AVX2)
-                target_compile_options(ncnn PRIVATE /arch:AVX2)
+                target_compile_options(ncnn PRIVATE /arch:AVX2 -mfma /D__SSSE3__ /D__SSE4_1__ /D__FMA__)
             else()
-                target_compile_options(ncnn PRIVATE /arch:AVX -mfma)
+                target_compile_options(ncnn PRIVATE /arch:AVX -mfma /D__SSSE3__ /D__SSE4_1__ /D__FMA__)
             endif()
             if(NCNN_AVXVNNI)
-                target_compile_options(ncnn PRIVATE -mavxvnni)
+                target_compile_options(ncnn PRIVATE -mavxvnni /D__AVXVNNI__)
             elseif(NCNN_XOP)
-                target_compile_options(ncnn PRIVATE -mxop)
+                target_compile_options(ncnn PRIVATE -mxop /D__XOP__)
             endif()
             if(NCNN_F16C)
-                target_compile_options(ncnn PRIVATE -mf16c)
+                target_compile_options(ncnn PRIVATE -mf16c /D__F16C__)
             endif()
         else()
             if(NCNN_AVX2)
-                target_compile_options(ncnn PRIVATE -mavx2)
+                target_compile_options(ncnn PRIVATE -mavx2 -mfma)
             else()
                 target_compile_options(ncnn PRIVATE -mavx -mfma)
             endif()
@@ -485,12 +485,12 @@ if(NCNN_TARGET_ARCH STREQUAL "x86")
                 target_compile_options(ncnn PRIVATE /D__F16C__)
             endif()
         elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC")
-            target_compile_options(ncnn PRIVATE /arch:AVX)
+            target_compile_options(ncnn PRIVATE /arch:AVX /D__SSSE3__ /D__SSE4_1__)
             if(NCNN_XOP)
-                target_compile_options(ncnn PRIVATE -mxop)
+                target_compile_options(ncnn PRIVATE -mxop /D__XOP__)
             endif()
             if(NCNN_F16C)
-                target_compile_options(ncnn PRIVATE -mf16c)
+                target_compile_options(ncnn PRIVATE -mf16c /D__F16C__)
             endif()
         else()
             target_compile_options(ncnn PRIVATE -mavx)