diff --git a/cmake/onnxruntime_mlas.cmake b/cmake/onnxruntime_mlas.cmake
index c6204268b2..3fee9c0cb5 100644
--- a/cmake/onnxruntime_mlas.cmake
+++ b/cmake/onnxruntime_mlas.cmake
@@ -201,11 +201,7 @@ else()
     set(mlas_platform_srcs_avx
       ${ONNXRUNTIME_ROOT}/core/mlas/lib/x86/SgemmKernelAvx.S
     )
-    if (CMAKE_SYSTEM_NAME STREQUAL "Android")
-      set_source_files_properties(${mlas_platform_srcs_avx} PROPERTIES COMPILE_FLAGS "-mavx -fno-integrated-as")
-    else()
-      set_source_files_properties(${mlas_platform_srcs_avx} PROPERTIES COMPILE_FLAGS "-mavx")
-    endif()
+    set_source_files_properties(${mlas_platform_srcs_avx} PROPERTIES COMPILE_FLAGS "-mavx")
 
     set(mlas_platform_srcs
       ${mlas_platform_srcs_sse2}
diff --git a/onnxruntime/core/mlas/lib/i386/SgemmKernelAvx.asm b/onnxruntime/core/mlas/lib/i386/SgemmKernelAvx.asm
index 6db5a10302..cbc88c3c8d 100644
--- a/onnxruntime/core/mlas/lib/i386/SgemmKernelAvx.asm
+++ b/onnxruntime/core/mlas/lib/i386/SgemmKernelAvx.asm
@@ -27,7 +27,7 @@ INCLUDE SgemmKernelCommon.inc
 
         ASSUME  DS:FLAT,ES:FLAT,SS:NOTHING,FS:NOTHING,GS:NOTHING
 
-        EXTERN  _MlasMaskMoveAvx:NEAR
+        EXTERN  _MlasMaskMoveTableAvx:NEAR
 
 _TEXT   SEGMENT DWORD PUBLIC 'CODE'
 
@@ -319,11 +319,8 @@ SkipAccumulateMasked16x2Block:
         add     ebp,8                       ; correct for over-subtract above
 
 OutputMasked8x2Block:
-        mov     SgemmKernelFrame.CountN[esp],ebp
-        vbroadcastss xmm0,SgemmKernelFrame.CountN[esp]
-        vpcmpgtd xmm1,xmm0,XMMWORD PTR [_MlasMaskMoveAvx+16]
-        vpcmpgtd xmm0,xmm0,XMMWORD PTR [_MlasMaskMoveAvx]
-        vinsertf128 ymm0,ymm0,xmm1,1
+        neg     ebp
+        vmovdqu ymm0,YMMWORD PTR [_MlasMaskMoveTableAvx+ebp*4+8*4]
         cmp     BYTE PTR SgemmKernelFrame.ZeroMode[esp],0
         jnz     SkipAccumulateMasked8x2Block
         vmaskmovps ymm4,ymm0,YMMWORD PTR [esi]
@@ -398,11 +395,8 @@ SkipAccumulateMasked16x1Block:
         add     ebp,8                       ; correct for over-subtract above
 
 OutputMasked8x1Block:
-        mov     SgemmKernelFrame.CountN[esp],ebp
-        vbroadcastss xmm0,SgemmKernelFrame.CountN[esp]
-        vpcmpgtd xmm1,xmm0,XMMWORD PTR [_MlasMaskMoveAvx+16]
-        vpcmpgtd xmm0,xmm0,XMMWORD PTR [_MlasMaskMoveAvx]
-        vinsertf128 ymm0,ymm0,xmm1,1
+        neg     ebp
+        vmovdqu ymm0,YMMWORD PTR [_MlasMaskMoveTableAvx+ebp*4+8*4]
         cmp     BYTE PTR SgemmKernelFrame.ZeroMode[esp],0
         jnz     SkipAccumulateMasked8x1Block
         vmaskmovps ymm4,ymm0,YMMWORD PTR [esi]
diff --git a/onnxruntime/core/mlas/lib/x86/SgemmKernelAvx.S b/onnxruntime/core/mlas/lib/x86/SgemmKernelAvx.S
index f874710c1a..7af2a9e118 100644
--- a/onnxruntime/core/mlas/lib/x86/SgemmKernelAvx.S
+++ b/onnxruntime/core/mlas/lib/x86/SgemmKernelAvx.S
@@ -238,8 +238,7 @@ Return Value:
 
 --*/
 
-        .globl  C_UNDERSCORE(MlasGemmFloatKernelAvx)
-C_UNDERSCORE(MlasGemmFloatKernelAvx):
+        FUNCTION_ENTRY MlasGemmFloatKernelAvx
 
         push    ebp
         push    ebx
@@ -342,14 +341,10 @@ C_UNDERSCORE(MlasGemmFloatKernelAvx):
         add     ebp,8                       # correct for over-subtract above
 
 .LOutputMasked8x2Block:
-        call    __x86.get_pc_thunk.bx
-        add     ebx,OFFSET _GLOBAL_OFFSET_TABLE_
-        mov     ebx,DWORD PTR C_UNDERSCORE(MlasMaskMoveAvx)@GOT[ebx]
-        mov     .LSgemmKernelFrame_CountN[esp],ebp
-        vbroadcastss xmm0,.LSgemmKernelFrame_CountN[esp]
-        vpcmpgtd xmm1,xmm0,XMMWORD PTR [ebx+16]
-        vpcmpgtd xmm0,xmm0,XMMWORD PTR [ebx]
-        vinsertf128 ymm0,ymm0,xmm1,1
+        neg     ebp
+        LoadGlobalOffsetTable bx
+        mov     ebx,DWORD PTR C_UNDERSCORE(MlasMaskMoveTableAvx)@GOT[ebx]
+        vmovdqu ymm0,YMMWORD PTR [ebx+ebp*4+8*4]
         cmp     BYTE PTR .LSgemmKernelFrame_ZeroMode[esp],0
         jnz     .LSkipAccumulateMasked8x2Block
         vmaskmovps ymm4,ymm0,YMMWORD PTR [esi]
@@ -424,14 +419,10 @@ C_UNDERSCORE(MlasGemmFloatKernelAvx):
         add     ebp,8                       # correct for over-subtract above
 
 .LOutputMasked8x1Block:
-        call    __x86.get_pc_thunk.bx
-        add     ebx,OFFSET _GLOBAL_OFFSET_TABLE_
-        mov     ebx,DWORD PTR C_UNDERSCORE(MlasMaskMoveAvx)@GOT[ebx]
-        mov     .LSgemmKernelFrame_CountN[esp],ebp
-        vbroadcastss xmm0,.LSgemmKernelFrame_CountN[esp]
-        vpcmpgtd xmm1,xmm0,XMMWORD PTR [ebx+16]
-        vpcmpgtd xmm0,xmm0,XMMWORD PTR [ebx]
-        vinsertf128 ymm0,ymm0,xmm1,1
+        neg     ebp
+        LoadGlobalOffsetTable bx
+        mov     ebx,DWORD PTR C_UNDERSCORE(MlasMaskMoveTableAvx)@GOT[ebx]
+        vmovdqu ymm0,YMMWORD PTR [ebx+ebp*4+8*4]
         cmp     BYTE PTR .LSgemmKernelFrame_ZeroMode[esp],0
         jnz     .LSkipAccumulateMasked8x1Block
         vmaskmovps ymm4,ymm0,YMMWORD PTR [esi]
diff --git a/onnxruntime/core/mlas/lib/x86/SgemmKernelSse2.S b/onnxruntime/core/mlas/lib/x86/SgemmKernelSse2.S
index f7718d3cef..f42175ec6c 100644
--- a/onnxruntime/core/mlas/lib/x86/SgemmKernelSse2.S
+++ b/onnxruntime/core/mlas/lib/x86/SgemmKernelSse2.S
@@ -225,8 +225,7 @@ Return Value:
 
 --*/
 
-        .globl  C_UNDERSCORE(MlasGemmFloatKernelSse)
-C_UNDERSCORE(MlasGemmFloatKernelSse):
+        FUNCTION_ENTRY MlasGemmFloatKernelSse
 
         push    ebp
         push    ebx
diff --git a/onnxruntime/core/mlas/lib/x86/asmmacro.h b/onnxruntime/core/mlas/lib/x86/asmmacro.h
index 00f11eea3f..4b80eea735 100644
--- a/onnxruntime/core/mlas/lib/x86/asmmacro.h
+++ b/onnxruntime/core/mlas/lib/x86/asmmacro.h
@@ -19,3 +19,61 @@ Abstract:
 #else
 #define C_UNDERSCORE(symbol) symbol
 #endif
+
+/*++
+
+Macro Description:
+
+    This macro emits the assembler directives to annotate a new function.
+
+Arguments:
+
+    FunctionName - Supplies the name of the function.
+
+--*/
+
+        .macro FUNCTION_ENTRY FunctionName
+
+        .p2align 4
+#if defined(__APPLE__)
+        .globl  _\FunctionName\()
+_\FunctionName\():
+#else
+        .globl  \FunctionName\()
+        .type   \FunctionName\(),@function
+\FunctionName\():
+#endif
+
+        .endm
+
+/*++
+
+Macro Description:
+
+    This macro emits the code to load the global offset table address into the
+    supplied register.
+
+Arguments:
+
+    TargetReg - Specifies the target register.
+
+--*/
+
+        .macro  LoadGlobalOffsetTable, TargetReg
+
+//
+// The LLVM integrated assembler doesn't support the Intel syntax for OFFSET:
+//
+//      add     ebx,OFFSET _GLOBAL_OFFSET_TABLE_
+//
+// Workaround this by temporarily switching to AT&T syntax.
+//
+
+        .att_syntax
+
+        calll   __x86.get_pc_thunk.\TargetReg\()
+        addl    $_GLOBAL_OFFSET_TABLE_,%e\TargetReg\()
+
+        .intel_syntax noprefix
+
+        .endm