Only translate SSE4.1 to VSX if _mm_packus_epi32 available

_mm_packus_epi32 was added in GCC v12.1.

CMakeLists.txt

@@ -361,6 +361,18 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(powerpc|ppc)")
             set(NCNN_SSE2 ON)
         endif()
 
+        set(CMAKE_REQUIRED_FLAGS "-DNO_WARN_X86_INTRINSICS -D__SSE4_1__")
+        check_cxx_source_compiles("#include <smmintrin.h>\nint main() { __m128i _v, _a, _b; _v = _mm_packus_epi32(_a, _b); return 0; }" NCNN_COMPILER_SUPPORT_PPC64LE_SSE4_1)
+        unset(CMAKE_REQUIRED_FLAGS)
+
+        if(NCNN_COMPILER_SUPPORT_PPC64LE_SSE4_1)
+            if(NCNN_VSX)
+                option(NCNN_SSE4_1 "optimize ppc64le platform with sse4.1 extension" ON)
+            endif()
+        else()
+            message(WARNING "The compiler does not support sse4.1 extension. NCNN_SSE4_1 will be OFF.")
+        endif()
+
         check_cxx_compiler_flag("-mcpu=native -mtune=native" NCNN_COMPILER_SUPPORT_CPU_NATIVE)
         if(NCNN_COMPILER_SUPPORT_CPU_NATIVE)
             set(NCNN_PPC64LE_DEFAULT_MACHINE_TYPE "native")

src/CMakeLists.txt

@@ -513,7 +513,12 @@ if(NCNN_TARGET_ARCH STREQUAL "powerpc")
     # Auto-translate SSE2 to VSX
     if(NCNN_SSE2)
         # Don't define __SSE3__ because it triggers usage of _MM_DENORMALS_ZERO_OFF, which isn't supported on ppc64le yet.
-        target_compile_options(ncnn PRIVATE -DNO_WARN_X86_INTRINSICS -D__MMX__ -D__SSE__ -D__SSE2__ -D__SSSE3__ -D__SSE4_1__)
+        target_compile_options(ncnn PRIVATE -DNO_WARN_X86_INTRINSICS -D__MMX__ -D__SSE__ -D__SSE2__ -D__SSSE3__)
+    endif()
+
+    # Auto-translate SSE4.1 to VSX
+    if(NCNN_SSE4_1)
+        target_compile_options(ncnn PRIVATE -DNO_WARN_X86_INTRINSICS -D__SSE4_1__)
     endif()
 endif()