[xDNN] Update xDNN library to support more intrinsic.

.gitignore

@@ -25,4 +25,4 @@ dist/
 /3rdparty/onednn
 /3rdparty/cmdline
 /3rdparty/sentencepiece
-/3rdparty/ig
+/3rdparty/xdnn

CMakeLists.txt

@@ -56,14 +56,14 @@ endif()
 
 include("cmake/mklml.cmake")
 include("cmake/onednn.cmake")
-include("cmake/ig.cmake")
+include("cmake/xdnn.cmake")
 include(GNUInstallDirs)
 
 include_directories(${CMAKE_SOURCE_DIR}/3rdparty/)
 include_directories(${CMAKE_SOURCE_DIR}/3rdparty/mklml/include)
 include_directories(${CMAKE_SOURCE_DIR}/3rdparty/onednn/include)
 include_directories(${CMAKE_SOURCE_DIR}/3rdparty/onednn/build/include)
-include_directories(${CMAKE_SOURCE_DIR}/3rdparty/ig)
+include_directories(${CMAKE_SOURCE_DIR}/3rdparty/xdnn)
 include_directories(${CMAKE_SOURCE_DIR}/include)
 include_directories(${CMAKE_SOURCE_DIR}/src/kernels)
 include_directories(${CMAKE_SOURCE_DIR}/src/layers)
@@ -75,18 +75,18 @@ include_directories(${CMAKE_SOURCE_DIR}/src/common)
 link_directories(${CMAKE_SOURCE_DIR}/src/kernels)
 link_directories(${CMAKE_SOURCE_DIR}/3rdparty/mklml/lib)
 link_directories(${CMAKE_SOURCE_DIR}/3rdparty/onednn/build/src)
-link_directories(${CMAKE_SOURCE_DIR}/3rdparty/ig)
+link_directories(${CMAKE_SOURCE_DIR}/3rdparty/xdnn)
 
 set(3RDPART_LIB_LIST "MPI::MPI_CXX" "ccl" "dnnl" "numa")
-set(DEPEND_LIST "onednn" "mklml" "ig_lib")
+set(DEPEND_LIST "onednn" "mklml" "xdnn_lib")
 
 option(BUILD_WITH_SHARED_LIBS "Build with shared libraries" OFF)
 if(BUILD_WITH_SHARED_LIBS)
     message("Building with shared libraries.")
-    list(APPEND 3RDPART_LIB_LIST "ig")
+    list(APPEND 3RDPART_LIB_LIST "xdnn")
 else()
     message("Building with static libraries.")
-    list(APPEND 3RDPART_LIB_LIST "ig_static")
+    list(APPEND 3RDPART_LIB_LIST "xdnn_static")
 endif()
 
 # Enable AVX512_FP16 optimization

cmake/ig.cmake → cmake/xdnn.cmake

@@ -25,11 +25,11 @@ project(dependency NONE)
 include(ExternalProject)
 
 # cmake-format: off
-ExternalProject_Add(ig_lib
-  URL               https://github.com/intel/xFasterTransformer/releases/download/IntrinsicGemm/ig_v1.1.tar.gz
-  URL_HASH          MD5=47e5a2cd021caad2b1367c0b71dff2e7
+ExternalProject_Add(xdnn_lib
+  URL               https://github.com/intel/xFasterTransformer/releases/download/IntrinsicGemm/xdnn_v1.1.tar.gz
+  URL_HASH          MD5=b49bf8808d66ea75cfba80a406c9a587
   TIMEOUT           60
-  SOURCE_DIR        ${CMAKE_SOURCE_DIR}/3rdparty/ig
+  SOURCE_DIR        ${CMAKE_SOURCE_DIR}/3rdparty/xdnn
   CONFIGURE_COMMAND ""
   BUILD_COMMAND     ""
   INSTALL_COMMAND   ""

src/layers/attention.h

@@ -496,9 +496,9 @@ class Attention {
                     C = result.Row(b * ctx->inputSeqLen + startSeq) + i * ctx->attHeadSize;
 
                     if constexpr (std::is_same_v<KVCacheT, float>) {
-                        ig_sgemm_single_thread(false, false, m, n, k, 1.0f, A, lda, B, ldb, 0.0f, C, ldc);
+                        xdnn_sgemm_single_thread(false, false, m, n, k, 1.0f, A, lda, B, ldb, 0.0f, C, ldc);
                     } else if constexpr (std::is_same_v<KVCacheT, float16_t>) {
-                        ig_sgemm_f32f16f32_single_thread(false, false, m, n, k, 1.0f, A, lda, B, ldb, 0.0f, C, ldc);
+                        xdnn_sgemm_f32f16f32_single_thread(false, false, m, n, k, 1.0f, A, lda, (const XDNN_FP16 *)B, ldb, 0.0f, C, ldc);
                     }
 
 #ifdef DEBUG

src/utils/decoder_util.h

@@ -542,7 +542,7 @@ class DecoderUtil {
 
     // C = A * B
     // bTranspose: B need to be transposed or not
-    // ig_sgemm_single_thread(transa, transb, m, n, k, alpha, A, lda, B, ldb, beta, C, ldc);
+    // xdnn_sgemm_single_thread(transa, transb, m, n, k, alpha, A, lda, B, ldb, beta, C, ldc);
     static void sgemm(const float* A, const float* B, float* C, int m, int n, int k,
             bool transa, bool transb) {
         int lda = (transa ? m : k);