dot load 128 adjust

microsoft · Dec 6, 2021 · 50fd111 · 50fd111
1 parent ab7de03
commit 50fd111
Showing 10 changed files with 728 additions and 275 deletions.
diff --git a/cmake/onnxruntime_mlas.cmake b/cmake/onnxruntime_mlas.cmake
@@ -47,8 +47,8 @@ function(setup_mlas_source_for_windows)
       )
 
       set(mlas_platform_preprocess_srcs
-        ${MLAS_SRC_DIR}/arm64/ConvSymKernelNeon.asm
-        ${MLAS_SRC_DIR}/arm64/ConvSymKernelNeonDot.asm
+        ${MLAS_SRC_DIR}/arm64/ConvSymU8KernelDot.asm
+        ${MLAS_SRC_DIR}/arm64/ConvSymU8KernelNeon.asm
         ${MLAS_SRC_DIR}/arm64/DepthwiseConvsymKernelNeon.asm
         ${MLAS_SRC_DIR}/arm64/DepthwiseQConvKernelSize9Neon.asm
         ${MLAS_SRC_DIR}/arm64/QgemmU8X8KernelNeon.asm
@@ -270,7 +270,8 @@ else()
     if(ARM64 AND MLAS_SOURCE_IS_NOT_SET )
         enable_language(ASM)
         set(mlas_platform_srcs
-          ${MLAS_SRC_DIR}/aarch64/ConvSymKernelNeon.S
+          ${MLAS_SRC_DIR}/aarch64/ConvSymU8KernelDot.S
+          ${MLAS_SRC_DIR}/aarch64/ConvSymU8KernelNeon.S
           ${MLAS_SRC_DIR}/aarch64/DepthwiseConvSymKernelNeon.S
           ${MLAS_SRC_DIR}/aarch64/DepthwiseQConvKernelSize9Neon.S
           ${MLAS_SRC_DIR}/aarch64/QgemmU8X8KernelNeon.S

diff --git a/onnxruntime/core/mlas/lib/aarch64/AssembleDotProduct.h b/onnxruntime/core/mlas/lib/aarch64/AssembleDotProduct.h
@@ -83,3 +83,4 @@ Macro Description:
         .inst   Instruction
 
         .endm
+