enable vfpv4 on neon build only

Tencent · Oct 10, 2024 · 73e7364 · 73e7364
1 parent 061205a
commit 73e7364
Showing 1 changed file with 16 additions and 12 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -162,21 +162,25 @@ if((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm")
     endif()
 
     if(CMAKE_SIZEOF_VOID_P EQUAL 4 AND NOT 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"))
-            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)
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float32x4_t _s, _a, _b; _s = vmlaq_f32(_s, _a, _b); return 0; }" NCNN_COMPILER_SUPPORT_ARM_NEON)
 
-            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)
+        if(NCNN_COMPILER_SUPPORT_ARM_NEON)
+            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()
         endif()
 
         if(NCNN_COMPILER_SUPPORT_ARM_VFPV4 OR NCNN_COMPILER_SUPPORT_ARM_VFPV4_FP16)