Add fp32 executor

openvinotoolkit · Nov 14, 2024 · a745ed1 · a745ed1
1 parent 4b52286
commit a745ed1
Show file tree

Hide file tree

Showing 4 changed files with 41 additions and 2 deletions.
diff --git a/src/plugins/intel_cpu/CMakeLists.txt b/src/plugins/intel_cpu/CMakeLists.txt
@@ -148,6 +148,7 @@ if(RISCV64_XUANTIE)
 else()
     set(ENABLE_SHL_FOR_CPU_DEFAULT OFF)
 endif()
+set(ENABLE_KLEIDIAI_FOR_CPU ON)
 ov_dependent_option(ENABLE_SHL_FOR_CPU "Enable SHL for OpenVINO CPU Plugin" ${ENABLE_SHL_FOR_CPU_DEFAULT} "RISCV64" OFF)
 
 add_subdirectory(thirdparty)
@@ -175,6 +176,11 @@ if(DNNL_USE_ACL)
     set(OV_CPU_WITH_ACL ON)
 endif()
 
+if(ENABLE_KLEIDIAI_FOR_CPU)
+    add_definitions(-DOV_CPU_WITH_KLEIDIAI)
+    set(OV_CPU_WITH_KLEIDIAI ON)
+endif()
+
 if(OV_CPU_WITH_ACL)
     set(CMAKE_CXX_STANDARD 14)
 endif()
@@ -273,6 +279,7 @@ if(ENABLE_SHL_FOR_CPU)
 endif()
 if(ENABLE_KLEIDIAI_FOR_CPU)
     target_link_libraries(${TARGET_NAME} PRIVATE kleidiai)
+    target_include_directories(${TARGET_NAME} SYSTEM PRIVATE $<TARGET_PROPERTY:kleidiai,INCLUDE_DIRECTORIES>)
 endif()
 target_include_directories(${TARGET_NAME} SYSTEM PRIVATE $<TARGET_PROPERTY:dnnl,INCLUDE_DIRECTORIES>)
 

diff --git a/src/plugins/intel_cpu/src/nodes/executors/executor.hpp b/src/plugins/intel_cpu/src/nodes/executors/executor.hpp
@@ -51,6 +51,12 @@ namespace intel_cpu {
 #    define OV_CPU_INSTANCE_DNNL(...)
 #endif
 
+#if defined(OV_CPU_WITH_KLEIDIAI)
+#    define OV_CPU_INSTANCE_KLEIDIAI(...) {__VA_ARGS__},
+#else
+#    define OV_CPU_INSTANCE_KLEIDIAI(...)
+#endif
+
 #if defined(OPENVINO_ARCH_X86_64)
 #    define OV_CPU_INSTANCE_X64(...) {__VA_ARGS__},
 #else
@@ -86,7 +92,8 @@ enum class ExecutorType {
     Acl,
     Mlas,
     jit_aarch64,
-    Shl
+    Shl,
+    Kleidiai
 };
 
 enum class OperationType {

diff --git a/src/plugins/intel_cpu/src/nodes/executors/fullyconnected_implementations.cpp b/src/plugins/intel_cpu/src/nodes/executors/fullyconnected_implementations.cpp
@@ -27,6 +27,7 @@
 #include "ov_optional.hpp"
 #include "utils/cpp/maybe_unused.hpp"
 #include "utils/debug_capabilities.h"
+#include "nodes/executors/kleidiai/kleidiai_mm.hpp"
 
 #if defined(OV_CPU_WITH_ACL)
 #include "nodes/executors/acl/acl_fullyconnected.hpp"
@@ -208,6 +209,27 @@ OV_CPU_MAYBE_UNUSED_FUNCTION static inline bool noPostOps(const FCConfig& config
 template <>
 const std::vector<ExecutorImplementation<FCAttrs>>& getImplementations() {
     static const std::vector<ExecutorImplementation<FCAttrs>> fullyconnectedImplementations {
+        OV_CPU_INSTANCE_KLEIDIAI(
+            "fullyconnected_kleidiai",
+            ExecutorType::Kleidiai,
+            OperationType::MatMul,
+            ShapeTolerance::Agnostic,
+            // supports
+            [](const FCConfig& config) -> bool {
+                return MatMulKleidiAIExecutor::supports(config);
+            },
+            // requiresFallback
+            [](const FCConfig& config) -> ov::optional<executor::Config<FCAttrs>> {
+                return {};
+            },
+            // acceptsShapes
+            [](const MemoryArgs& memory) -> bool {
+                return true;
+            },
+            // create
+            [](const FCAttrs& attrs, const PostOps& postOps, const MemoryArgs& memory, ExecutorContext::CPtr context) {
+                return std::make_shared<MatMulKleidiAIExecutor>(attrs, postOps, memory, context);
+            })
         OV_CPU_INSTANCE_MLAS_X64(
             "fullyconnected_mlas",
             ExecutorType::Mlas,

diff --git a/src/plugins/intel_cpu/thirdparty/CMakeLists.txt b/src/plugins/intel_cpu/thirdparty/CMakeLists.txt
@@ -171,8 +171,11 @@ endif()
 
 if(ENABLE_KLEIDIAI_FOR_CPU)
     set(KLEIDIAI_BUILD_TESTS OFF CACHE BOOL "" FORCE)
+    set(TEMP_BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS})
+    set(BUILD_SHARED_LIBS OFF CACHE BOOL "Build shared libs" FORCE)
     add_subdirectory(kleidiai)
-    ov_install_static_lib(kleidiai ${OV_CPACK_COMP_CORE})
+    # ov_install_static_lib(kleidiai ${OV_CPACK_COMP_CORE})
+    set(BUILD_SHARED_LIBS ${TEMP_BUILD_SHARED_LIBS} CACHE BOOL "Build shared libs" FORCE)
 endif()
 
 ov_add_onednn()