windows clang ci

.github/workflows/windows-clang.yml

@@ -0,0 +1,85 @@
+name: windows-clang
+on:
+  push:
+    branches: [master]
+    paths:
+    - '.github/workflows/windows-clang.yml'
+    - 'CMakeLists.txt'
+    - 'cmake/**'
+    - 'src/*'
+    - 'src/layer/*'
+    - 'src/layer/arm/**'
+    - 'src/layer/x86/**'
+    - 'src/layer/vulkan/**'
+    - 'tests/**'
+  pull_request:
+    branches: [master]
+    paths:
+    - '.github/workflows/windows-clang.yml'
+    - 'CMakeLists.txt'
+    - 'cmake/**'
+    - 'src/*'
+    - 'src/layer/*'
+    - 'src/layer/arm/**'
+    - 'src/layer/x86/**'
+    - 'src/layer/vulkan/**'
+    - 'tests/**'
+concurrency:
+  group: windows-clang-${{ github.ref }}
+  cancel-in-progress: true
+permissions:
+  contents: read
+
+jobs:
+  windows:
+    name: ClangCL
+    runs-on: windows-2022
+
+    env:
+      UseMultiToolTask: true
+      NCNN_CMAKE_OPTIONS: -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF
+
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        submodules: true
+
+    - name: arm64
+      run: |
+        mkdir build-arm64; cd build-arm64
+        cmake -T ClangCL -A arm64 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_VULKAN=OFF ..
+        cmake --build . --config Release -j 4
+
+    - name: arm64-vulkan
+      run: |
+        mkdir build-arm64-vulkan; cd build-arm64-vulkan
+        cmake -T ClangCL -A arm64 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_VULKAN=ON -DNCNN_SHARED_LIB=ON ..
+        cmake --build . --config Release -j 4
+
+    - name: x86
+      run: |
+        mkdir build-x86; cd build-x86
+        cmake -T ClangCL -A Win32 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_BUILD_TESTS=ON -DNCNN_VULKAN=OFF ..
+        cmake --build . --config Release -j 4
+    - name: x86-test
+      run: cd build-x86; ctest -C Release --output-on-failure -j 4
+
+    - name: x86-vulkan
+      run: |
+        mkdir build-x86-vulkan; cd build-x86-vulkan
+        cmake -T ClangCL -A Win32 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_VULKAN=ON -DNCNN_SHARED_LIB=ON ..
+        cmake --build . --config Release -j 4
+
+    - name: x64
+      run: |
+        mkdir build-x64; cd build-x64
+        cmake -T ClangCL -A x64 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_BUILD_TESTS=ON -DNCNN_VULKAN=OFF ..
+        cmake --build . --config Release -j 4
+    - name: x64-test
+      run: cd build-x64; ctest -C Release --output-on-failure -j 4
+
+    - name: x64-vulkan
+      run: |
+        mkdir build-x64-vulkan; cd build-x64-vulkan
+        cmake -T ClangCL -A x64 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_VULKAN=ON -DNCNN_SHARED_LIB=ON ..
+        cmake --build . --config Release -j 4

CMakeLists.txt

@@ -134,11 +134,15 @@ endif()
 ##############################################
 
 include(CheckCXXCompilerFlag)
-check_cxx_source_compiles("int main() { int a = 0; asm volatile(\"\" : \"=r\"(a) : \"0\"(a) : \"memory\"); return 0; }" NCNN_COMPILER_SUPPORT_GNU_INLINE_ASM)
-if(NCNN_COMPILER_SUPPORT_GNU_INLINE_ASM)
-    option(NCNN_GNU_INLINE_ASM "optimize platform with gnu style inline assembly" ON)
-else()
-    message(WARNING "The compiler does not support gnu style inline assembly. NCNN_GNU_INLINE_ASM will be OFF.")
+
+# gnu inline assembly in clang msvc does not work actually
+if(NOT (CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC")))
+    check_cxx_source_compiles("int main() { int a = 0; asm volatile(\"\" : \"=r\"(a) : \"0\"(a) : \"memory\"); return 0; }" NCNN_COMPILER_SUPPORT_GNU_INLINE_ASM)
+    if(NCNN_COMPILER_SUPPORT_GNU_INLINE_ASM)
+        option(NCNN_GNU_INLINE_ASM "optimize platform with gnu style inline assembly" ON)
+    else()
+        message(WARNING "The compiler does not support gnu style inline assembly. NCNN_GNU_INLINE_ASM will be OFF.")
+    endif()
 endif()
 
 if((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm")
@@ -178,7 +182,7 @@ if((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm")
     endif()
 
     if(CMAKE_SIZEOF_VOID_P EQUAL 8 OR NCNN_TARGET_ILP32)
-        if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
+        if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC")
             set(CMAKE_REQUIRED_FLAGS "/arch:armv8.0")
             check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float32x4_t _a; float16x4_t _s = vcvt_f16_f32(_a); return 0; }" NCNN_COMPILER_SUPPORT_ARM_VFPV4)
 
@@ -212,6 +216,41 @@ if((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm")
             set(CMAKE_REQUIRED_FLAGS "/arch:armv8.6")
             check_cxx_source_compiles("#include <arm_sve.h>\nint main() { svfloat32_t _s, _a, _b; _s = svmmla_f32(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM86_SVEF32MM)
 
+            unset(CMAKE_REQUIRED_FLAGS)
+        elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC")
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.0")
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float32x4_t _a; float16x4_t _s = vcvt_f16_f32(_a); return 0; }" NCNN_COMPILER_SUPPORT_ARM_VFPV4)
+
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.2 -march=armv8.2-a+fp16")
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16x8_t _s, _a, _b; _s = vfmaq_f16(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM82_FP16)
+
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.2 -march=armv8.2-a+dotprod")
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int32x4_t _s; int8x16_t _a, _b; _s = vdotq_s32(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM82_DOTPROD)
+
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.2 -march=armv8.2-a+fp16fml")
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float32x4_t _s; float16x8_t _a, _b; _s = vfmlalq_low_f16(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM82_FP16FML)
+
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.4 -march=armv8.4-a+bf16")
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float32x4_t _s; bfloat16x8_t _a, _b; _s = vcvt_f32_bf16(vcvt_bf16_f32(vbfmmlaq_f32(_s, _a, _b))); return 0; }" NCNN_COMPILER_SUPPORT_ARM84_BF16)
+
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.4 -march=armv8.4-a+i8mm")
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int32x4_t _s; int8x16_t _a, _b; _s = vmmlaq_s32(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM84_I8MM)
+
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.6 -march=armv8.6-a+sve")
+            check_cxx_source_compiles("#include <arm_sve.h>\nint main() { svfloat16_t _s, _a, _b; svbool_t bp; _s = svmla_f16_z(bp, _s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM86_SVE)
+
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.6 -march=armv8.6-a+sve2")
+            check_cxx_source_compiles("#include <arm_sve.h>\nint main() { svint16_t _s; svint8_t _a, _b; _s = svmlslb_s16(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM86_SVE2)
+
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.6 -march=armv8.6-a+sve+bf16")
+            check_cxx_source_compiles("#include <arm_sve.h>\nint main() { svfloat32_t _s; svbfloat16_t _a, _b; _s = svbfmmla_f32(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM86_SVEBF16)
+
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.6 -march=armv8.6-a+sve+i8mm")
+            check_cxx_source_compiles("#include <arm_sve.h>\nint main() { svint32_t _s; svint8_t _a, _b; _s = svmmla_s32(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM86_SVEI8MM)
+
+            set(CMAKE_REQUIRED_FLAGS "/arch:armv8.6 -march=armv8.6-a+sve+f32mm")
+            check_cxx_source_compiles("#include <arm_sve.h>\nint main() { svfloat32_t _s, _a, _b; _s = svmmla_f32(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM86_SVEF32MM)
+
             unset(CMAKE_REQUIRED_FLAGS)
         else()
             set(CMAKE_REQUIRED_FLAGS "-march=armv8-a")
@@ -447,7 +486,7 @@ else()
 
     option(NCNN_SSE2 "optimize x86 platform with sse2 extension" ON)
 
-    if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
+    if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC")
         check_cxx_compiler_flag("/arch:AVX" NCNN_COMPILER_SUPPORT_X86_AVX)
         check_cxx_compiler_flag("/arch:AVX" NCNN_COMPILER_SUPPORT_X86_FMA)
         check_cxx_compiler_flag("/arch:AVX" NCNN_COMPILER_SUPPORT_X86_XOP)
@@ -467,6 +506,32 @@ else()
         set(CMAKE_REQUIRED_FLAGS "/arch:AVX512")
         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)
+    elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC")
+        check_cxx_compiler_flag("/arch:AVX" NCNN_COMPILER_SUPPORT_X86_AVX)
+
+        set(CMAKE_REQUIRED_FLAGS "/arch:AVX -mfma -mf16c")
+        check_cxx_source_compiles("#include <immintrin.h>\nint main() { __m256 _s, _a, _b; _s = _mm256_fmadd_ps(_a, _b, _s); return 0; }" NCNN_COMPILER_SUPPORT_X86_FMA)
+
+        set(CMAKE_REQUIRED_FLAGS "/arch:AVX -mxop")
+        check_cxx_source_compiles("#include <x86intrin.h>\nint main() { __m128 _s, _a, _b; _s = _mm_maddd_epi16(_a, _b, _s); return 0; }" NCNN_COMPILER_SUPPORT_X86_XOP)
+
+        check_cxx_compiler_flag("/arch:AVX -mf16c" NCNN_COMPILER_SUPPORT_X86_F16C)
+        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 -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 -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 -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 -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)
     else()
         check_cxx_compiler_flag("-mavx" NCNN_COMPILER_SUPPORT_X86_AVX)

cmake/ncnn_add_layer.cmake

@@ -134,7 +134,7 @@ macro(ncnn_add_layer class)
     endif()
 
     if(NCNN_TARGET_ARCH STREQUAL "x86")
-        if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
+        if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC")
             if(NCNN_RUNTIME_CPU AND NCNN_AVX512)
                 ncnn_add_arch_opt_layer(${class} avx512 "/arch:AVX512 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__")
             endif()
@@ -165,6 +165,37 @@ macro(ncnn_add_layer class)
             if(NCNN_F16C)
                 ncnn_add_arch_opt_source(${class} f16c "/arch:AVX /D__SSSE3__ /D__SSE4_1__ /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")
+            if(NCNN_RUNTIME_CPU AND NCNN_AVX512)
+                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 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__")
+            endif()
+            if(NCNN_RUNTIME_CPU AND NCNN_AVX)
+                ncnn_add_arch_opt_layer(${class} avx "/arch:AVX /D__SSSE3__ /D__SSE4_1__")
+            endif()
+            if(NCNN_AVX512VNNI)
+                ncnn_add_arch_opt_source(${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(${class} avx512bf16 "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512bf16 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512BF16__")
+            endif()
+            if(NCNN_AVX512FP16)
+                ncnn_add_arch_opt_source(${class} avx512fp16 "/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512fp16 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512FP16__")
+            endif()
+            if(NCNN_AVXVNNI)
+                ncnn_add_arch_opt_source(${class} avxvnni "/arch:AVX2 -mfma -mf16c -mavxvnni /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVXVNNI__")
+            endif()
+            if(NCNN_AVX2)
+                ncnn_add_arch_opt_source(${class} avx2 "/arch:AVX2 -mfma -mf16c /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__")
+            endif()
+            if(NCNN_XOP)
+                ncnn_add_arch_opt_source(${class} xop "/arch:AVX -mxop /D__SSSE3__ /D__SSE4_1__ /D__XOP__")
+            endif()
+            if(NCNN_F16C)
+                ncnn_add_arch_opt_source(${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 -mfma -mf16c")
@@ -216,7 +247,7 @@ macro(ncnn_add_layer class)
     endif()
 
     if(NCNN_TARGET_ARCH STREQUAL "arm" AND (CMAKE_SIZEOF_VOID_P EQUAL 8 OR NCNN_TARGET_ILP32))
-        if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
+        if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC")
             if(NCNN_VFPV4)
                 ncnn_add_arch_opt_source(${class} vfpv4 " ")
             endif()
@@ -246,6 +277,36 @@ macro(ncnn_add_layer class)
             endif()
             if(NCNN_ARM86SVEF32MM)
             endif()
+        elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC")
+            if(NCNN_VFPV4)
+                ncnn_add_arch_opt_source(${class} vfpv4 " ")
+            endif()
+            if(NCNN_ARM82)
+                ncnn_add_arch_opt_source(${class} asimdhp "/arch:armv8.2 -march=armv8.2-a+fp16 /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC")
+            endif()
+            if(NCNN_ARM82DOT)
+                ncnn_add_arch_opt_source(${class} asimddp "/arch:armv8.2 -march=armv8.2-a+fp16+dotprod /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_DOTPROD")
+            endif()
+            if(NCNN_ARM82FP16FML)
+                ncnn_add_arch_opt_source(${class} asimdfhm "/arch:armv8.2 -march=armv8.2-a+fp16+fp16fml /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_FP16_FML")
+            endif()
+            if(NCNN_ARM84BF16)
+                ncnn_add_arch_opt_source(${class} bf16 "/arch:armv8.4 -march=armv8.4-a+fp16+dotprod+bf16 /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_DOTPROD /D__ARM_FEATURE_FP16_FML /D__ARM_FEATURE_BF16_VECTOR_ARITHMETIC")
+            endif()
+            if(NCNN_ARM84I8MM)
+                ncnn_add_arch_opt_source(${class} i8mm "/arch:armv8.4 -march=armv8.4-a+fp16+dotprod+i8mm /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_DOTPROD /D__ARM_FEATURE_FP16_FML /D__ARM_FEATURE_MATMUL_INT8")
+            endif()
+            # TODO add support for sve family
+            if(NCNN_ARM86SVE)
+            endif()
+            if(NCNN_ARM86SVE2)
+            endif()
+            if(NCNN_ARM86SVEBF16)
+            endif()
+            if(NCNN_ARM86SVEI8MM)
+            endif()
+            if(NCNN_ARM86SVEF32MM)
+            endif()
         else()
             if(NCNN_VFPV4)
                 ncnn_add_arch_opt_source(${class} vfpv4 " ")