[RISCV64] Added parallelism support for FC

cmake/features.cmake

@@ -85,10 +85,11 @@ ov_dependent_option (ENABLE_PKGCONFIG_GEN "Enable openvino.pc pkg-config file ge
 #
 
 # "OneDNN library based on OMP or TBB or Sequential implementation: TBB|OMP|SEQ"
-if(RISCV64 OR ANDROID)
-    # oneDNN does not support non-SEQ for RISC-V architecture
+if(ANDROID)
     # on Android we experience SEGFAULT during compilation
     set(THREADING_DEFAULT "SEQ")
+elseif(RISCV64)
+    set(THREADING_DEFAULT "OMP")
 else()
     set(THREADING_DEFAULT "TBB")
 endif()

cmake/toolchains/riscv64-071-thead-gnu.toolchain.cmake

@@ -33,9 +33,6 @@ set(CMAKE_CXX_COMPILER ${RISCV_TOOLCHAIN_ROOT}/bin/riscv64-unknown-linux-gnu-g++
 set(CMAKE_STRIP ${RISCV_TOOLCHAIN_ROOT}/bin/riscv64-unknown-linux-gnu-strip)
 set(PKG_CONFIG_EXECUTABLE "NOT-FOUND" CACHE PATH "Path to RISC-V pkg-config")
 
-# Don't run the linker on compiler check
-set(CMAKE_TRY_COMPILE_TARGET_TYPE STATIC_LIBRARY)
-
 set(CMAKE_C_FLAGS_INIT "${CMAKE_C_FLAGS_INIT} -march=rv64gcv0p7_zfh_xtheadc -mabi=lp64d")
 set(CMAKE_CXX_FLAGS_INIT "${CMAKE_CXX_FLAGS_INIT} -march=rv64gcv0p7_zfh_xtheadc -mabi=lp64d")
 

cmake/toolchains/riscv64-100-thead-gnu.toolchain.cmake

@@ -33,9 +33,6 @@ set(CMAKE_CXX_COMPILER ${RISCV_TOOLCHAIN_ROOT}/bin/riscv64-unknown-linux-gnu-g++
 set(CMAKE_STRIP ${RISCV_TOOLCHAIN_ROOT}/bin/riscv64-unknown-linux-gnu-strip)
 set(PKG_CONFIG_EXECUTABLE "NOT-FOUND" CACHE PATH "Path to RISC-V pkg-config")
 
-# Don't run the linker on compiler check
-set(CMAKE_TRY_COMPILE_TARGET_TYPE STATIC_LIBRARY)
-
 set(CMAKE_C_FLAGS_INIT "${CMAKE_C_FLAGS_INIT} -march=rv64gcv1p0_zfh -mabi=lp64d")
 set(CMAKE_CXX_FLAGS_INIT "${CMAKE_CXX_FLAGS_INIT} -march=rv64gcv1p0_zfh -mabi=lp64d")
 

docs/dev/build_riscv64.md

@@ -69,7 +69,9 @@ The software was validated on the following devices:
      -DRISCV_TOOLCHAIN_ROOT=/opt/riscv
    ```
    > **NOTE**: The `riscv-gnu-toolchain` is build as there are essential files used for cross compilation under `/opt/riscv/sysroot`. The latest stable versions of Clang or GCC both support compiling source code into RISC-V instructions, so it is acceptable to choose your preferable compilers by specifying `-DCMAKE_C_COMPILER` and `CMAKE_CXX_COMPILER`. But remember to add the key `-DCMAKE_SYSROOT=/opt/riscv/sysroot`, otherwise many fundamental headers and libs could not be found during cross compilation. 
-
+
+   > **NOTE**: By default OpenVINO is built with OpenMP support on RISC-V devices.
+
    Then run `make` to build the project:
    ```sh
    make install -j$(nproc)

src/plugins/intel_cpu/src/nodes/executors/shl/shl_fullyconnected.cpp

@@ -121,19 +121,37 @@ bool ShlFCExecutor::update(const MemoryArgs& memory) {
     src = src.cloneWithNewShape(memory.at(ARG_SRC)->getDescPtr()->getShape().getStaticDims());
     dst = dst.cloneWithNewShape(memory.at(ARG_DST)->getDescPtr()->getShape().getStaticDims());
 
+    const auto src_shape = src.getShape();
+    const auto dst_shape = dst.getShape();
+    dim_M = std::accumulate(dst_shape.rbegin() + 1, dst_shape.rend(), size_t(1), std::multiplies<size_t>());
+    dim_In = src_shape.back();
+    dim_Out = dst_shape.back();
+    LDA = dim_In * memory.at(ARG_SRC)->getPrecision().size();
+    LDC = dim_Out * memory.at(ARG_DST)->getPrecision().size();
+
     return true;
 }
 
 void ShlFCExecutor::execute(const MemoryArgs& memory) {
-    src.setData(memory.at(ARG_SRC)->getData());
     wei.setData(packedWeights->getData());
-    dst.setData(memory.at(ARG_DST)->getData());
     if (with_bias) {
         bias.setData(memory.at(ARG_BIAS)->getData());
     }
 
-    OPENVINO_ASSERT(csinn_fullyconnected(src.get(), dst.get(), wei.get(), bias.get(), static_cast<csinn_fc_params*>(params.get())) == CSINN_TRUE,
-                    "ShlFCExecutor: failed to execute");
+    const auto nthreads = std::min(static_cast<int>(dim_M), parallel_get_max_threads());
+    parallel_nt(nthreads, [&](const int ithr, const int nthr) {
+        size_t dim_M0 = 0, dim_M1 = 0;
+        splitter(dim_M, nthr, ithr, dim_M0, dim_M1);
+
+        const auto M = dim_M1 - dim_M0;
+        auto src_tensor = src.cloneWithNewShape(ov::Shape{ M, dim_In });
+        auto dst_tensor = dst.cloneWithNewShape(ov::Shape{ M, dim_Out });
+        src_tensor.setData(reinterpret_cast<uint8_t*>(memory.at(ARG_SRC)->getData()) + dim_M0 * LDA);
+        dst_tensor.setData(reinterpret_cast<uint8_t*>(memory.at(ARG_DST)->getData()) + dim_M0 * LDC);
+
+        OPENVINO_ASSERT(csinn_fullyconnected(src_tensor.get(), dst_tensor.get(), wei.get(), bias.get(), static_cast<csinn_fc_params*>(params.get())) == CSINN_TRUE,
+                        "ShlFCExecutor: failed to execute");
+    });
 }
 
 }  // namespace intel_cpu

src/plugins/intel_cpu/src/nodes/executors/shl/shl_fullyconnected.hpp

@@ -38,6 +38,12 @@ class ShlFCExecutor : public Executor {
 
     const MemoryCPtr packedWeights;
 
+    size_t dim_M = 0;
+    size_t dim_In = 0;
+    size_t dim_Out = 0;
+    size_t LDA = 0;
+    size_t LDC = 0;
+
     bool with_bias = false;
 };
 using ShlFCExecutorPtr = std::shared_ptr<ShlFCExecutor>;