check msvc arm target support

Tencent · Nov 15, 2023 · 266c88a · 266c88a
1 parent feb5732
commit 266c88a
Showing 1 changed file with 76 additions and 32 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -153,16 +153,23 @@ if((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm")
     OR ((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")) AND (${CMAKE_GENERATOR_PLATFORM} MATCHES "^(arm|arm64)")))
     set(NCNN_TARGET_ARCH arm)
 
-    if(CMAKE_SIZEOF_VOID_P EQUAL 4 AND 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")))
-        set(CMAKE_REQUIRED_FLAGS "-mfpu=neon-vfpv4")
-        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)
+    if(CMAKE_SIZEOF_VOID_P EQUAL 4)
+        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"))
+            set(CMAKE_REQUIRED_FLAGS "/arch:VFPv4")
+            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)
 
-        if(NOT NCNN_COMPILER_SUPPORT_ARM_VFPV4)
-            set(CMAKE_REQUIRED_FLAGS "-mfpu=neon-vfpv4 -mfp16-format=ieee")
-            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_FP16)
-        endif()
+            unset(CMAKE_REQUIRED_FLAGS)
+        else()
+            set(CMAKE_REQUIRED_FLAGS "-mfpu=neon-vfpv4")
+            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)
 
-        unset(CMAKE_REQUIRED_FLAGS)
+            if(NOT NCNN_COMPILER_SUPPORT_ARM_VFPV4)
+                set(CMAKE_REQUIRED_FLAGS "-mfpu=neon-vfpv4 -mfp16-format=ieee")
+                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_FP16)
+            endif()
+
+            unset(CMAKE_REQUIRED_FLAGS)
+        endif()
 
         if(NCNN_COMPILER_SUPPORT_ARM_VFPV4 OR NCNN_COMPILER_SUPPORT_ARM_VFPV4_FP16)
             option(NCNN_VFPV4 "optimize armv7 platform with vfpv4" ON)
@@ -171,41 +178,78 @@ if((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm")
         endif()
     endif()
 
-    if(CMAKE_SIZEOF_VOID_P EQUAL 8 AND 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")))
-        set(CMAKE_REQUIRED_FLAGS "-march=armv8-a")
-        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)
+    if(CMAKE_SIZEOF_VOID_P EQUAL 8)
+        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"))
+            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 "-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")
+            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 "-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")
+            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 "-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:armv8.2")
+            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 "-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")
+            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 "-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:armv8.4")
+            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 "-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")
+            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 "-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")
+            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 "-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")
+            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 "-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")
+            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 "-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)
+            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)
+            unset(CMAKE_REQUIRED_FLAGS)
+        else()
+            set(CMAKE_REQUIRED_FLAGS "-march=armv8-a")
+            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 "-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 "-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 "-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 "-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 "-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 "-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 "-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 "-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 "-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 "-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)
+        endif()
 
         if(NCNN_COMPILER_SUPPORT_ARM_VFPV4)
             option(NCNN_VFPV4 "optimize aarch64 platform with vfpv4" ON)