POWER10: Add optimized dgemm kernel (microsoft#9652)

* POWER10: Add optimized dgemm kernel This patch makes use of POWER10 matrix multiply assist feature and adds new DGEMM kernel. * Indentation update Co-authored-by: Rajalakshmi Srinivasaraghavan <rajis@linux.ibm.com>
garymm · Nov 23, 2021 · 8564fc1 · 8564fc1
1 parent bf5e9a5
commit 8564fc1
Show file tree

Hide file tree

Showing 8 changed files with 799 additions and 599 deletions.
diff --git a/cmake/onnxruntime_mlas.cmake b/cmake/onnxruntime_mlas.cmake
@@ -291,6 +291,7 @@ else()
           ${MLAS_SRC_DIR}/dgemm.cpp
           ${MLAS_SRC_DIR}/power/DgemmKernelPower.cpp
         )
+        set_source_files_properties(${MLAS_SRC_DIR}/power/SgemmKernelPower.cpp PROPERTIES COMPILE_FLAGS "-DSINGLE")
         check_cxx_compiler_flag("-mcpu=power10" HAS_POWER10)
         if(HAS_POWER10)
           set(CMAKE_REQUIRED_FLAGS "-mcpu=power10")
@@ -318,8 +319,10 @@ else()
             endif()
             set(mlas_platform_srcs_power10
               ${MLAS_SRC_DIR}/power/SgemmKernelPOWER10.cpp
+              ${MLAS_SRC_DIR}/power/DgemmKernelPOWER10.cpp
             )
-            set_source_files_properties(${mlas_platform_srcs_power10} PROPERTIES COMPILE_FLAGS "-O2 -mcpu=power10")
+            set_source_files_properties(${MLAS_SRC_DIR}/power/SgemmKernelPOWER10.cpp PROPERTIES COMPILE_FLAGS "-O2 -mcpu=power10 -DSINGLE")
+            set_source_files_properties(${MLAS_SRC_DIR}/power/DgemmKernelPOWER10.cpp PROPERTIES COMPILE_FLAGS "-O2 -mcpu=power10")
             set(mlas_platform_srcs
               ${mlas_platform_srcs}
               ${mlas_platform_srcs_power10}

diff --git a/onnxruntime/core/mlas/lib/mlasi.h b/onnxruntime/core/mlas/lib/mlasi.h
@@ -499,6 +499,7 @@ extern "C" {
     MLAS_GEMM_FLOAT_KERNEL MlasSgemmKernel;
     MLAS_GEMM_FLOAT_KERNEL MlasSgemmKernelPOWER10;
     MLAS_GEMM_DOUBLE_KERNEL MlasDgemmKernel;
+    MLAS_GEMM_DOUBLE_KERNEL MlasDgemmKernelPOWER10;
 #else
     MLAS_GEMM_FLOAT_KERNEL MlasSgemmKernelZero;
     MLAS_GEMM_FLOAT_KERNEL MlasSgemmKernelAdd;
@@ -1886,7 +1887,7 @@ MlasStoreAlignedFloat64x2(double* Buffer, MLAS_FLOAT64X2 Vector)
 #if defined(MLAS_SSE2_INTRINSICS)
     _mm_store_pd(Buffer, Vector);
 #elif defined(MLAS_VSX_INTRINSICS)
-    vec_st(Vector, 0, Buffer);
+    *((MLAS_FLOAT64X2*)Buffer) = Vector;
 #endif
 }
 

diff --git a/onnxruntime/core/mlas/lib/platform.cpp b/onnxruntime/core/mlas/lib/platform.cpp
@@ -379,6 +379,7 @@ Return Value:
     bool HasP10Instructions = ((hwcap2 & PPC_FEATURE2_MMA) && (hwcap2 & PPC_FEATURE2_ARCH_3_1));
     if (HasP10Instructions) {
         this->GemmFloatKernel = MlasSgemmKernelPOWER10;
+        this->GemmDoubleKernel = MlasDgemmKernelPOWER10;
     }
 #endif
 #endif