Cherry pick x86 fix in release 0.3.0 (#606)

* merge x86 linux build

* remove wrong argu

Author: Raymond Yang

cmake/CMakeLists.txt

@@ -129,6 +129,11 @@ else()
     string(APPEND CMAKE_CXX_FLAGS_RELWITHDEBINFO " -march=native -mtune=native")
     string(APPEND CMAKE_C_FLAGS_RELWITHDEBINFO " -march=native -mtune=native")
   endif()
+  if(onnxruntime_BUILD_x86)
+    set (CMAKE_SYSTEM_PROCESSOR "x86")
+    set (CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -msse2 -mfpmath=sse -Wno-narrowing")
+    set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse2 -mfpmath=sse -Wno-narrowing")
+  endif()
 endif()
 
 if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")

cmake/onnxruntime_mlas.cmake

@@ -93,6 +93,25 @@ else()
       ${ONNXRUNTIME_ROOT}/core/mlas/lib/aarch64/sgemma.s
     )
 
+  elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "^(i.86|x86?)$")
+
+    enable_language(ASM)
+
+    set(mlas_platform_srcs_sse2
+      ${ONNXRUNTIME_ROOT}/core/mlas/lib/x86/SgemmKernelSse2.S
+    )
+    set_source_files_properties(${mlas_platform_srcs_sse2} PROPERTIES COMPILE_FLAGS "-msse2")
+
+    set(mlas_platform_srcs_avx
+      ${ONNXRUNTIME_ROOT}/core/mlas/lib/x86/SgemmKernelAvx.S
+    )
+    set_source_files_properties(${mlas_platform_srcs_avx} PROPERTIES COMPILE_FLAGS "-mavx")
+
+    set(mlas_platform_srcs
+      ${mlas_platform_srcs_sse2}
+      ${mlas_platform_srcs_avx}
+    )
+
   elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
 
     enable_language(ASM)
@@ -106,7 +125,7 @@ else()
       ${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmKernelSse2.S
       ${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmTransposePackB16x4Sse2.S
     )
-    set_source_files_properties(${mlas_platform_srcs_sse} PROPERTIES COMPILE_FLAGS "-msse2")
+    set_source_files_properties(${mlas_platform_srcs_sse2} PROPERTIES COMPILE_FLAGS "-msse2")
 
     set(mlas_platform_srcs_avx
       ${ONNXRUNTIME_ROOT}/core/mlas/lib/x86_64/SgemmKernelAvx.S

onnxruntime/core/mlas/lib/mlasi.h

@@ -34,9 +34,6 @@ Module Name:
 #include <cpuid.h>
 #include <immintrin.h>
 #endif
-#if defined(__x86_64__)
-#include "x86_64/xgetbv.h"
-#endif
 #endif
 
 //

onnxruntime/core/mlas/lib/platform.cpp

@@ -23,6 +23,41 @@ Module Name:
 
 MLAS_PLATFORM MlasPlatform;
 
+#ifdef MLAS_TARGET_AMD64_IX86
+
+//
+// Reads the processor extended control register to determine platform
+// capabilities.
+//
+
+#if !defined(_XCR_XFEATURE_ENABLED_MASK)
+#define _XCR_XFEATURE_ENABLED_MASK 0
+#endif
+
+inline
+uint64_t
+MlasReadExtendedControlRegister(
+    unsigned int ext_ctrl_reg
+    )
+{
+#if defined(_WIN32)
+    return _xgetbv(ext_ctrl_reg);
+#else
+    uint32_t eax, edx;
+
+    __asm__
+    (
+        "xgetbv"
+        : "=a" (eax), "=d" (edx)
+        : "c" (ext_ctrl_reg)
+    );
+
+    return ((uint64_t)edx << 32) | eax;
+#endif
+}
+
+#endif
+
 MLAS_PLATFORM::MLAS_PLATFORM(
     void
     )
@@ -74,11 +109,7 @@ Return Value:
         // Check if the operating system supports saving SSE and AVX states.
         //
 
-#if defined(_WIN32)
-        uint64_t xcr0 = _xgetbv(_XCR_XFEATURE_ENABLED_MASK);
-#else
-        uint64_t xcr0 = xgetbv(_XCR_XFEATURE_ENABLED_MASK);
-#endif
+        uint64_t xcr0 = MlasReadExtendedControlRegister(_XCR_XFEATURE_ENABLED_MASK);
 
         if ((xcr0 & 0x6) == 0x6) {