Upgrade to hcc 3.0 (#487)

* Upgrade compiler to hcc

* Add missing perl package for hcc

* Update hip and rocblas version

* Add flag to speedup build

* Remove symlink

* Add no-gpu-rdc to linking

* Add missing define

* Disable no unused argument

* Suppress warning

* Add definition to all targets

* Fix compiler flags

* Refer to rocblas

Dockerfile

@@ -30,6 +30,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
     hsakmt-roct-dev \
     lcov \
     libelf-dev \
+    libfile-which-perl \
     libncurses5-dev \
     libnuma-dev \
     libpthread-stubs0-dev \
@@ -65,7 +66,7 @@ ADD doc/requirements.txt /doc-requirements.txt
 RUN pip3 install -r /doc-requirements.txt
 
 # Install hcc
-RUN rclone -b roc-2.6.x -c 0f4c96b7851af2663a7f3ac16ecfb76c7c78a5bf https://github.com/RadeonOpenCompute/hcc.git /hcc
+RUN rclone -b roc-3.0.x -c 286651a04d9c3a8e3052dd84b1822985498cd27d https://github.com/RadeonOpenCompute/hcc.git /hcc
 RUN cget -p $PREFIX install hcc,/hcc
 
 # Use hcc
@@ -86,6 +87,5 @@ ENV MIOPEN_USER_DB_PATH=/tmp/miopen/user-db
 ENV LD_LIBRARY_PATH=$PREFIX/lib
 
 # Setup ubsan environment to printstacktrace
-RUN ln -s /usr/bin/llvm-symbolizer-5.0 /usr/local/bin/llvm-symbolizer
 ENV UBSAN_OPTIONS=print_stacktrace=1
 ENV ASAN_OPTIONS=detect_stack_use_after_return=1:check_initialization_order=1:strict_init_order=1

dev-requirements.txt

 pfultz2/rocm-recipes
 danmar/cppcheck@ef714225bb31e9a76ac2484796763572386955ae -DHAVE_RULES=1
-ROCm-Developer-Tools/HIP@2490e42baa7d90458f0632fd9fbead2d395f41b9
+ROCm-Developer-Tools/HIP@roc-3.0.0
 python/cpython@v3.6.6 -X autotools -H sha256:92aa914572c695c0aeb01b0a214813f414da4b51a371234df514a74761f2bb36
 -f requirements.txt

requirements.txt

 google/protobuf@v3.11.0 -DCMAKE_POSITION_INDEPENDENT_CODE=On -X subdir -Dprotobuf_BUILD_TESTS=Off
 RadeonOpenCompute/rocm-cmake@b29ff83 --build
-ROCmSoftwarePlatform/rocBLAS@7197df74e5a1ba64ff967065872e5f86a3516637
+ROCmSoftwarePlatform/rocBLAS@abd98a2b48b29326ebaef471630786a548622c06
 ROCmSoftwarePlatform/MIOpen@2.0.0
 blaze,https://bitbucket.org/blaze-lib/blaze/get/f0755dea0e03.tar.gz -X header -DHEADER_DIR=blaze
 half,https://github.com/pfultz2/half/archive/1.12.0.tar.gz -X header -H sha256:0a08660b68abb176ebc2a0cdf8de46e3182a7f46c66443bb80dbfaaec98cf969

src/targets/gpu/CMakeLists.txt

@@ -69,8 +69,15 @@ add_library(migraphx_device
 set_target_properties(migraphx_device PROPERTIES EXPORT_NAME device)
 rocm_set_soversion(migraphx_device ${MIGRAPHX_SO_VERSION})
 rocm_clang_tidy_check(migraphx_device)
-target_compile_options(migraphx_device PRIVATE -std=c++17)
-target_link_libraries(migraphx_device migraphx hip::device -Wno-invalid-command-line-argument -amdgpu-target=gfx803 -amdgpu-target=gfx900 -amdgpu-target=gfx906)
+target_compile_options(migraphx_device PRIVATE -std=c++17 -fno-gpu-rdc -Wno-unused-command-line-argument)
+target_link_libraries(migraphx_device migraphx hip::device -fno-gpu-rdc -Wno-invalid-command-line-argument -Wno-unused-command-line-argument)
+if(CMAKE_CXX_COMPILER MATCHES ".*hcc")
+    set(AMDGPU_TARGETS "gfx803;gfx900;gfx906" CACHE STRING "")
+    foreach(AMDGPU_TARGET ${AMDGPU_TARGETS})
+        target_compile_options(migraphx_device PRIVATE -amdgpu-target=${AMDGPU_TARGET})
+        target_link_libraries(migraphx_device -amdgpu-target=${AMDGPU_TARGET})
+    endforeach()
+endif()
 target_include_directories(migraphx_device PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>)
 target_include_directories(migraphx_device PRIVATE $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/device/include>)
 
@@ -112,6 +119,8 @@ add_library(migraphx_gpu
 set_target_properties(migraphx_gpu PROPERTIES EXPORT_NAME gpu)
 rocm_set_soversion(migraphx_gpu ${MIGRAPHX_SO_VERSION})
 rocm_clang_tidy_check(migraphx_gpu)
+# Workaround broken rocblas headers
+target_compile_definitions(migraphx_gpu PUBLIC -D__HIP_PLATFORM_HCC__=1)
 target_link_libraries(migraphx_gpu PUBLIC migraphx MIOpen roc::rocblas)
 target_link_libraries(migraphx_gpu PRIVATE migraphx_device)