Add more generic AVX feature detection as well as flags for if avx512… (#1041)

Author: JonathanHenson

CMakeLists.txt

@@ -24,7 +24,6 @@ include(AwsFeatureTests)
 include(AwsSanitizers)
 include(AwsThreadAffinity)
 include(AwsThreadName)
-include(AwsSIMD)
 include(CTest)
 
 set(GENERATED_ROOT_DIR "${CMAKE_CURRENT_BINARY_DIR}/generated")
@@ -224,12 +223,9 @@ target_include_directories(${PROJECT_NAME} PUBLIC
 
 target_compile_definitions(${PROJECT_NAME} PRIVATE -DCJSON_HIDE_SYMBOLS)
 
-# Enable SIMD encoder if the compiler supports the right features
-simd_add_definitions(${PROJECT_NAME})
-
-if (HAVE_AVX2_INTRINSICS)
+if (AWS_HAVE_AVX2_INTRINSICS)
     target_compile_definitions(${PROJECT_NAME} PRIVATE -DUSE_SIMD_ENCODING)
-    simd_add_source_avx2(${PROJECT_NAME} "source/arch/intel/encoding_avx2.c")
+    simd_add_source_avx(${PROJECT_NAME} "source/arch/intel/encoding_avx2.c")
     message(STATUS "Building SIMD base64 decoder")
 endif()
 

README.md

@@ -158,7 +158,7 @@ Example:
 * Avoid C99 features in header files. For some types such as bool, uint32_t etc..., these are defined if not available for the language
 standard being used in `aws/common/common.h`, so feel free to use them.
 * For C++ compatibility, don't put const members in structs.
-* Avoid C++ style comments e.g. `//`.
+* Avoid C++ style comments e.g. `//` in header files and prefer block style  (`/* */`) for long blocks of text. C++ style comments are fine in C files.
 * All public API functions need C++ guards and Windows dll semantics.
 * Use Unix line endings.
 * Where implementation hiding is desired for either ABI or runtime polymorphism reasons, use the `void *impl` pattern. v-tables

cmake/AwsFeatureTests.cmake

@@ -117,3 +117,7 @@ if(MSVC)
         return 0;
     }" AWS_HAVE_MSVC_INTRINSICS_X64)
 endif()
+
+# This does a lot to detect when intrinsics are available and has to set cflags to do so.
+# leave it in its own file for ease of managing it.
+include(AwsSIMD)

cmake/AwsSIMD.cmake

@@ -8,18 +8,31 @@ if (USE_CPU_EXTENSIONS)
     if (MSVC)
         check_c_compiler_flag("/arch:AVX2" HAVE_M_AVX2_FLAG)
         if (HAVE_M_AVX2_FLAG)
-            set(AVX2_CFLAGS "/arch:AVX2")
+            set(AVX_CFLAGS "/arch:AVX2")
         endif()
     else()
         check_c_compiler_flag(-mavx2 HAVE_M_AVX2_FLAG)
         if (HAVE_M_AVX2_FLAG)
-            set(AVX2_CFLAGS "-mavx -mavx2")
+            set(AVX_CFLAGS "-mavx -mavx2")
         endif()
     endif()
 
+    if (MSVC)
+        check_c_compiler_flag("/arch:AVX512" HAVE_M_AVX512_FLAG)
+        if (HAVE_M_AVX512_FLAG)
+            # docs imply AVX512 brings in AVX2. And it will compile, but it will break at runtime on
+            # instructions such as _mm256_load_si256(). Leave it on.
+            set(AVX_CFLAGS "/arch:AVX512 /arch:AVX2")
+        endif()
+    else()
+        check_c_compiler_flag("-mavx512f -mvpclmulqdq" HAVE_M_AVX512_FLAG)
+        if (HAVE_M_AVX512_FLAG)
+            set(AVX_CFLAGS "-mavx512f -mvpclmulqdq -mpclmul -mavx -mavx2 -msse4.2")
+        endif()
+    endif()
 
     set(old_flags "${CMAKE_REQUIRED_FLAGS}")
-    set(CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS} ${AVX2_CFLAGS}")
+    set(CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS} ${AVX_CFLAGS}")
 
     check_c_source_compiles("
         #include <immintrin.h>
@@ -35,7 +48,15 @@ if (USE_CPU_EXTENSIONS)
             _mm256_permutevar8x32_epi32(vec, vec);
 
             return 0;
-        }"  HAVE_AVX2_INTRINSICS)
+        }"  AWS_HAVE_AVX2_INTRINSICS)          
+
+    check_c_source_compiles("
+        #include <immintrin.h>
+
+        int main() {
+            __m512 a = _mm512_setzero_ps();
+            return 0;
+        }" AWS_HAVE_AVX512_INTRINSICS)
 
     check_c_source_compiles("
         #include <immintrin.h>
@@ -45,30 +66,20 @@ if (USE_CPU_EXTENSIONS)
             __m256i vec;
             memset(&vec, 0, sizeof(vec));
             return (int)_mm256_extract_epi64(vec, 2);
-        }" HAVE_MM256_EXTRACT_EPI64)
+        }" AWS_HAVE_MM256_EXTRACT_EPI64)
 
     set(CMAKE_REQUIRED_FLAGS "${old_flags}")
 endif() # USE_CPU_EXTENSIONS
 
-macro(simd_add_definition_if target definition)
-    if(${definition})
-        target_compile_definitions(${target} PRIVATE -D${definition})
-    endif(${definition})
-endmacro(simd_add_definition_if)
-
-# Configure private preprocessor definitions for SIMD-related features
-# Does not set any processor feature codegen flags
-function(simd_add_definitions target)
-    simd_add_definition_if(${target} HAVE_AVX2_INTRINSICS)
-    simd_add_definition_if(${target} HAVE_MM256_EXTRACT_EPI64)
-endfunction(simd_add_definitions)
+# The part where the definition is added to the compiler flags has been moved to config.h.in
+# see git history for more details.
 
-# Adds source files only if AVX2 is supported. These files will be built with
-# avx2 intrinsics enabled.
-# Usage: simd_add_source_avx2(target file1.c file2.c ...)
-function(simd_add_source_avx2 target)
+# Adds AVX flags, if any, that are supported. These files will be built with
+# available avx intrinsics enabled.
+# Usage: simd_add_source_avx(target file1.c file2.c ...)
+function(simd_add_source_avx target)
     foreach(file ${ARGN})
         target_sources(${target} PRIVATE ${file})
-        set_source_files_properties(${file} PROPERTIES COMPILE_FLAGS "${AVX2_CFLAGS}")
+        set_source_files_properties(${file} PROPERTIES COMPILE_FLAGS "${AVX_CFLAGS}")
     endforeach()
-endfunction(simd_add_source_avx2)
+endfunction(simd_add_source_avx)

include/aws/common/config.h.in

@@ -19,5 +19,8 @@
 #cmakedefine AWS_HAVE_EXECINFO
 #cmakedefine AWS_HAVE_WINAPI_DESKTOP
 #cmakedefine AWS_HAVE_LINUX_IF_LINK_H
+#cmakedefine AWS_HAVE_AVX2_INTRINSICS
+#cmakedefine AWS_HAVE_AVX512_INTRINSICS
+#cmakedefine AWS_HAVE_MM256_EXTRACT_EPI64
 
 #endif

include/aws/common/cpuid.h

@@ -17,6 +17,7 @@ enum aws_cpu_feature_name {
     AWS_CPU_FEATURE_AVX512,
     AWS_CPU_FEATURE_ARM_CRC,
     AWS_CPU_FEATURE_BMI2,
+    AWS_CPU_FEATURE_VPCLMULQDQ,
     AWS_CPU_FEATURE_COUNT,
 };
 

source/arch/intel/cpuid.c

@@ -113,13 +113,26 @@ static bool s_has_bmi2(void) {
     return true;
 }
 
+static bool s_has_vpclmulqdq(void) {
+    uint32_t abcd[4];
+    /* Check VPCLMULQDQ:
+     * CPUID.(EAX=07H, ECX=0H):ECX.VPCLMULQDQ[bit 20]==1 */
+    uint32_t vpclmulqdq_mask = (1 << 20);
+    aws_run_cpuid(7, 0, abcd);
+    if ((abcd[2] & vpclmulqdq_mask) != vpclmulqdq_mask) {
+        return false;
+    }
+    return true;
+}
+
 has_feature_fn *s_check_cpu_feature[AWS_CPU_FEATURE_COUNT] = {
     [AWS_CPU_FEATURE_CLMUL] = s_has_clmul,
     [AWS_CPU_FEATURE_SSE_4_1] = s_has_sse41,
     [AWS_CPU_FEATURE_SSE_4_2] = s_has_sse42,
     [AWS_CPU_FEATURE_AVX2] = s_has_avx2,
     [AWS_CPU_FEATURE_AVX512] = s_has_avx512,
     [AWS_CPU_FEATURE_BMI2] = s_has_bmi2,
+    [AWS_CPU_FEATURE_VPCLMULQDQ] = s_has_vpclmulqdq,
 };
 
 bool aws_cpu_has_feature(enum aws_cpu_feature_name feature_name) {

source/arch/intel/encoding_avx2.c

@@ -179,7 +179,7 @@ static inline bool decode(const unsigned char *in, unsigned char *out) {
      * so we'll just copy right out of the vector as a fallback
      */
 
-#ifdef HAVE_MM256_EXTRACT_EPI64
+#ifdef AWS_HAVE_MM256_EXTRACT_EPI64
     uint64_t hi = _mm256_extract_epi64(vec, 2);
     const uint64_t *p_hi = &hi;
 #else

tests/CMakeLists.txt

@@ -295,7 +295,7 @@ add_test_case(test_byte_cursor_utf8_parse_u64_hex)
 
 add_test_case(byte_swap_test)
 
-if(HAVE_AVX2_INTRINSICS)
+if(AWS_HAVE_AVX2_INTRINSICS)
     add_test_case(alignment32_test)
 else()
     add_test_case(alignment16_test)