added inline assembly into thread-wise gemm

build/cmake-cuda.sh

-#!/bin/bash
-
-rm -f CMakeCache.txt
-rm -f *.cmake
-rm -rf CMakeFiles
-
-MY_PROJECT_SOURCE=/home/chao/code/modular_convolution
-MY_PROJECT_INSTALL=../install.dir
-
-cmake                                                                                       \
--D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL}                                               \
--D CMAKE_CXX_COMPILER=clang++                                                               \
--D CMAKE_BUILD_TYPE=Release                                                                 \
--D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                           \
--D DEVICE_BACKEND=CUDA                                                                      \
--D BOOST_ROOT="/package/install/boost_1.67.0"                                               \
--D CUDA_COMMON_INCLUDE_DIR="/home/chao/code/test_feature/cuda_common/cuda_10.0_common/inc"  \
--D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -arch=sm_61" \
-${MY_PROJECT_SOURCE}
-
-
-#-D CMAKE_CUDA_COMPILER="/package/install/cuda_10.0/bin/nvcc"                                \
-#-D CMAKE_CUDA_FLAGS="-G -lineinfo --source-in-ptx -keep -Xptxas -v -arch=sm_61"            \
-#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -arch=sm_61" \
-#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -arch=sm_61 -Xptxas -v -maxrregcount=128" \

build/cmake-hip.sh

-#!/bin/bash
-
-rm -f CMakeCache.txt
-rm -f *.cmake
-rm -rf CMakeFiles
-
-MY_PROJECT_SOURCE=/home/chao/code/modular_convolution
-MY_PROJECT_INSTALL=../install.dir
-
-cmake                                                                                       \
--D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL}                                               \
--D CMAKE_BUILD_TYPE=Release                                                                 \
--D DEVICE_BACKEND="HIP"                                                                     \
--D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                   \
--D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                           \
-${MY_PROJECT_SOURCE}

src/include/threadwise_gemm.hip.hpp

@@ -53,17 +53,39 @@ __device__ void threadwise_gemm(MatrixA,
         constexpr unsigned N = c_mtx.NCol();
         constexpr unsigned K = a_mtx.NRow(); // A is transposed
 
+        assert(M == 8);
+        assert(N == 8);
+        assert(K == 1);
+
         for(unsigned k = 0; k < K; ++k)
         {
+            const unsigned bindex = b_mtx.Get1dIndex(k, 0);
             for(unsigned i = 0; i < M; ++i)
             {
-                for(unsigned j = 0; j < N; ++j)
+                const unsigned aindex = a_mtx.Get1dIndex(k, i); // A is transposed
+                const unsigned cindex = c_mtx.Get1dIndex(i, 0);
+
+                //N = 8
+                //for(unsigned j = 0; j < N; ++j)
                 {
-                    const unsigned aindex = a_mtx.Get1dIndex(k, i); // A is transposed
-                    const unsigned bindex = b_mtx.Get1dIndex(k, j);
-                    const unsigned cindex = c_mtx.Get1dIndex(i, j);
+                    //const unsigned bindex = b_mtx.Get1dIndex(k, j);
+                    //const unsigned cindex = c_mtx.Get1dIndex(i, j);
+                    //f_accum(p_c_thread[cindex], p_a_thread[aindex] * p_b_thread[bindex]);
 
-                    f_accum(p_c_thread[cindex], p_a_thread[aindex] * p_b_thread[bindex]);
+                    asm volatile("\n \
+                            v_mac_f32 %0, %8, %9 \n \
+                            v_mac_f32 %1, %8, %10 \n \
+                            v_mac_f32 %2, %8, %11 \n \
+                            v_mac_f32 %3, %8, %12 \n \
+                            v_mac_f32 %4, %8, %13 \n \
+                            v_mac_f32 %5, %8, %14 \n \
+                            v_mac_f32 %6, %8, %15 \n \
+                            v_mac_f32 %7, %8, %16 \n \
+                            "
+                            : "=v"(p_c_thread[cindex + 0]),"=v"(p_c_thread[cindex + 1]),"=v"(p_c_thread[cindex + 2]),"=v"(p_c_thread[cindex + 3]),"=v"(p_c_thread[cindex + 4]),"=v"(p_c_thread[cindex + 5]),"=v"(p_c_thread[cindex + 6]),"=v"(p_c_thread[cindex + 7])
+                            : "v"(p_a_thread[aindex]), "v"(p_b_thread[bindex + 0]), "v"(p_b_thread[bindex + 1]),"v"(p_b_thread[bindex + 2]),"v"(p_b_thread[bindex + 3]),"v"(p_b_thread[bindex + 4]),"v"(p_b_thread[bindex + 5]),"v"(p_b_thread[bindex + 6]),"v"(p_b_thread[bindex + 7]),
+                            "0"(p_c_thread[cindex + 0]),"1"(p_c_thread[cindex + 1]),"2"(p_c_thread[cindex + 2]),"3"(p_c_thread[cindex + 3]),"4"(p_c_thread[cindex + 4]),"5"(p_c_thread[cindex + 5]),"6"(p_c_thread[cindex + 6]),"7"(p_c_thread[cindex + 7])
+                            );
                 }
             }
         }