refactor

build_script/cmake-hcc.sh

+#!/bin/bash
+
+rm -f CMakeCache.txt
+rm -f *.cmake
+rm -rf CMakeFiles
+
+MY_PROJECT_SOURCE=../../../
+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=hcc                                                                   \
+-D CMAKE_CXX_FLAGS="-gline-tables-only ${HCC_FLAGS}"                                        \
+-D CMAKE_EXE_LINKER_FLAGS="${HCC_FLAGS} -L/opt/rocm/lib --hip-link -lhip_hcc  "    \
+-D CMAKE_PREFIX_PATH="/opt/rocm;/home/package/build/mlopen_dep"                             \
+-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                           \
+${MY_PROJECT_SOURCE}
+

build_script/compile-hip.sh

@@ -2,4 +2,4 @@
 export KMDUMPISA=1
 
 make -j driver
-/opt/rocm/hcc/bin/llvm-objdump -mcpu=gfx900 -source -line-numbers driver/dump-gfx900.isabin > driver/dump-gfx900.isabin.isa
+/opt/rocm/hcc/bin/llvm-objdump -mcpu=gfx906 -source -line-numbers driver/dump-gfx906.isabin > driver/dump-gfx906.isabin.isa

driver/device_implicit_gemm_convolution_2_chwn_cyxk_khwn.hpp

@@ -196,7 +196,7 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
     for(index_t i = 0; i < nrepeat; ++i)
     {
         float time = launch_kernel(
-#if 0
+#if 1
             gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn
 #else
             gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer