Merge branch 'release/rocm-rel-5.3' into jd/master_5.3

ROCm · Oct 4, 2022 · d4272db · d4272db
2 parents 4df6527 + 6c94c7b
commit d4272db
Show file tree

Hide file tree

Showing 531 changed files with 346,607 additions and 32,168 deletions.
diff --git a/.clang-format b/.clang-format
@@ -1,6 +1,6 @@
 ---
 Language:        Cpp
-AccessModifierOffset: 0
+AccessModifierOffset: -4
 AlignAfterOpenBracket: Align
 AlignConsecutiveAssignments: true
 AlignConsecutiveDeclarations: false

diff --git a/.gitattributes b/.gitattributes
@@ -0,0 +1 @@
+*.kdb filter=lfs diff=lfs merge=lfs -text
diff --git a/.lfsconfig b/.lfsconfig
@@ -0,0 +1,2 @@
+[lfs]
+	fetchexclude = *
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -50,14 +50,16 @@ project ( MIOpen C CXX )
 
 enable_testing()
 
-find_package(ROCM REQUIRED PATHS /opt/rocm)
+find_package(ROCM 0.7.3 REQUIRED PATHS /opt/rocm)
 
 include(ROCMInstallTargets)
 include(ROCMPackageConfigHelpers)
 include(ROCMSetupVersion)
 include(ROCMInstallSymlinks)
 include(ROCMCreatePackage)
 include(CheckCXXCompilerFlag)
+include(ROCMHeaderWrapper)
+
 
 set(MIOPEN_ENABLE_SQLITE On CACHE BOOL "")
 # Use SQLITE for compiled kernels, when turned off this will use raw files
@@ -84,7 +86,7 @@ if(NOT WIN32 AND NOT APPLE)
     set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -s")
 endif()
 
-rocm_setup_version(VERSION 2.16.0)
+rocm_setup_version(VERSION 2.18.0)
 
 list( APPEND CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake )
 include(TargetFlags)
@@ -131,6 +133,7 @@ else()
     option(MIOPEN_DISABLE_SYSDB  "Disable sys database access" ${MIOPEN_EMBED_BUILD})
 endif()
 set(MIOPEN_BINCACHE_PATH "" CACHE STRING "URL or path containing binary cache files to embed")
+option(MIOPEN_EMBED_BINCACHE "Embed Binary Cache or KDB" Off)
 option(MIOPEN_EMBED_BUILD "Build with the set of embed flags." Off)
 option(MIOPEN_DISABLE_USERDB "Disable user database access" ${MIOPEN_EMBED_BUILD})
 
@@ -164,7 +167,7 @@ if( MIOPEN_BACKEND STREQUAL "OpenCL")
             ${CMAKE_INSTALL_PREFIX}/llvm
     )
     if(MIOPEN_HIP_COMPILER)
-        message("hip compiler: ${MIOPEN_HIP_COMPILER}")
+        message(STATUS "hip compiler: ${MIOPEN_HIP_COMPILER}")
     else()
         message(FATAL_ERROR "hip compiler not found")
     endif()
@@ -190,9 +193,12 @@ message(STATUS "Build with HIP ${hip_VERSION}")
 target_flags(HIP_COMPILER_FLAGS hip::device)
 # Remove cuda arch flags
 string(REGEX REPLACE --cuda-gpu-arch=[a-z0-9]+ "" HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
-string(REGEX REPLACE --offload-arch=[a-z0-9]+ "" HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
-string(REPLACE "$<LINK_LANGUAGE:CXX>" "1" HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
-string(REPLACE "SHELL:" "" HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
+string(REGEX REPLACE --offload-arch=[a-z0-9:+-]+ "" HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
+# Skip library paths since hip will incorrectly treat it as a source file
+string(APPEND HIP_COMPILER_FLAGS " ")
+foreach(_unused RANGE 2)
+    string(REGEX REPLACE " /[^ ]+\\.(a|so) " " " HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
+endforeach()
 
 # Override HIP version in config.h, if necessary.
 # The variables set by find_package() can't be overwritten,
@@ -251,7 +257,7 @@ if( MIOPEN_BACKEND STREQUAL "HIP" OR MIOPEN_BACKEND STREQUAL "HIPOC" OR MIOPEN_B
         message(STATUS "OpenCL compiler: ${HIP_OC_COMPILER}")
         set(HIP_OC_COMPILER "${HIP_OC_COMPILER}")
     else()
-        message(FATAL_ERROR "OpenCL compiler not found")
+        message(STATUS "OpenCL compiler not found")
     endif()
 
         # Hcc's clang always defines __HCC__ even when not using hcc driver
@@ -263,7 +269,7 @@ if( MIOPEN_BACKEND STREQUAL "HIP" OR MIOPEN_BACKEND STREQUAL "HIPOC" OR MIOPEN_B
     set(MIOPEN_USE_ROCBLAS ON CACHE BOOL "")
     if(MIOPEN_USE_ROCBLAS)
         find_package(rocblas REQUIRED PATHS /opt/rocm)
-        message(STATUS "Build with rocblas")
+        message(STATUS "Build with rocblas ${rocblas_VERSION}")
     else()
         message(STATUS "Build without rocblas")
     endif()
@@ -352,16 +358,18 @@ endif()
 # miopengemm
 if(MIOPEN_USE_MIOPENGEMM)
     find_package(miopengemm PATHS /opt/rocm)
-    if(miopengemm_FOUND)
-        message(STATUS "Build with miopengemm")
-        set(MIOPEN_USE_MIOPENGEMM 1)
-    else()
-        message(STATUS "Build without miopengemm")
+    if(NOT miopengemm_FOUND)
         set(MIOPEN_USE_MIOPENGEMM 0)
     endif()
+endif()
+
+if(MIOPEN_USE_MIOPENGEMM)
+    message(STATUS "Build with miopengemm")
 else()
+    if(MIOPEN_BACKEND_OPENCL)
+        message(FATAL_ERROR "MIOpen OpenCL backend requires MIOpenGEMM")
+    endif()
     message(STATUS "Build without miopengemm")
-    set(MIOPEN_USE_MIOPENGEMM 0)
 endif()
 
 if(MIOPEN_USE_COMGR)
@@ -387,8 +395,11 @@ message(STATUS "HALF_INCLUDE_DIR: ${HALF_INCLUDE_DIR}")
 
 option( MIOPEN_DEBUG_FIND_DB_CACHING "Use system find-db caching" ON)
 
+# FOR HANDLING ENABLE/DISABLE OPTIONAL BACKWARD COMPATIBILITY for FILE/FOLDER REORG
+option(BUILD_FILE_REORG_BACKWARD_COMPATIBILITY "Build with file/folder reorg with backward compatibility enabled" ON)
+
 set( MIOPEN_INSTALL_DIR miopen)
-set( DATA_INSTALL_DIR ${MIOPEN_INSTALL_DIR}/${CMAKE_INSTALL_DATAROOTDIR}/miopen )
+set( DATA_INSTALL_DIR ${CMAKE_INSTALL_DATAROOTDIR}/miopen )
 
 set(MIOPEN_GPU_SYNC Off CACHE BOOL "")
 if(BUILD_DEV)
@@ -405,8 +416,8 @@ else()
 endif()
 set(MIOPEN_SYSTEM_FIND_DB_SUFFIX "${MIOPEN_BACKEND}" CACHE PATH "Filename suffix for the system find-db files")
 
+set(MIOPEN_PACKAGE_REQS "hip-rocclr")
 
-set(MIOPEN_PACKAGE_REQS "rocm-clang-ocl, hip-rocclr")
 if(MIOPEN_USE_MIOPENGEMM)
     set(MIOPEN_PACKAGE_REQS "${MIOPEN_PACKAGE_REQS}, miopengemm")
 endif()
@@ -415,22 +426,79 @@ if(MIOPEN_USE_ROCBLAS)
     set(MIOPEN_PACKAGE_REQS "${MIOPEN_PACKAGE_REQS}, rocblas")
 endif()
 
-set(CPACK_DEBIAN_PACKAGE_DEPENDS "${MIOPEN_PACKAGE_REQS}, rocm-opencl-dev")
-set(CPACK_RPM_PACKAGE_REQUIRES "${MIOPEN_PACKAGE_REQS}, rocm-opencl-devel")
-
-# Make backends explicitly conflict
 if(MIOPEN_BACKEND STREQUAL "HIP")
+    # In HIP backend, there is a posibility of COMGR disabled.
+    # In this case add the clang-ocl as dependency for runtime kernel compilation
+    if(NOT MIOPEN_USE_COMGR)
+        set(MIOPEN_PACKAGE_REQS "${MIOPEN_PACKAGE_REQS}, rocm-clang-ocl")
+    endif()
+    set(CPACK_DEBIAN_PACKAGE_DEPENDS "${MIOPEN_PACKAGE_REQS}")
+    set(CPACK_RPM_PACKAGE_REQUIRES "${MIOPEN_PACKAGE_REQS}")
+
+    # Make backends explicitly conflict
     set(CPACK_DEBIAN_PACKAGE_CONFLICTS miopen-opencl)
     set(CPACK_RPM_PACKAGE_CONFLICTS miopen-opencl)
+
 elseif(MIOPEN_BACKEND STREQUAL "OpenCL")
+    set(CPACK_DEBIAN_PACKAGE_DEPENDS "${MIOPEN_PACKAGE_REQS}, rocm-opencl-dev")
+    set(CPACK_RPM_PACKAGE_REQUIRES "${MIOPEN_PACKAGE_REQS}, rocm-opencl-devel")
+
+    # Make backends explicitly conflict
     set(CPACK_DEBIAN_PACKAGE_CONFLICTS miopen-hip)
     set(CPACK_RPM_PACKAGE_CONFLICTS miopen-hip)
 endif()
+# Begin KDB component install 
+function(install_kdb)
+    set(oneValueArgs NAME COMPONENT_NAME)
+    cmake_parse_arguments(PARSE "" "${oneValueArgs}" "" ${ARGN})
+
+    if(PARSE_NAME)
+        set(FILE_NAME ${PARSE_NAME})
+    endif()
+    STRING(REPLACE - _ FILE_NAME "${ARCH_NAME}")
+    file(READ "${CMAKE_SOURCE_DIR}/src/kernels/${FILE_NAME}.kdb" FILE_CONTENTS LIMIT 7)
+    string(STRIP "${FILE_CONTENTS}" FILE_CONTENTS)
+    if(FILE_CONTENTS STREQUAL "version")
+        set(LFS_MISSING_FILES ${LFS_MISSING_FILES} ${FILE_NAME}.kdb PARENT_SCOPE)
+    else()
+        message("Installing ${FILE_NAME} in component ${PARSE_COMPONENT_NAME}")
+        rocm_install(FILES
+            src/kernels/${FILE_NAME}.kdb
+        DESTINATION ${DATA_INSTALL_DIR}/db
+        COMPONENT ${PARSE_COMPONENT_NAME})
+    endif()
+endfunction()
+
+# Both the lists below should be in sync always
+list(APPEND ARCH_LST gfx90a68 gfx90a6e gfx1030-36 gfx90878 gfx906-64 gfx906-60 gfx900-64 gfx900-56)
+list(APPEND ARCH_FILE_LST gfx90a-104kdb gfx90a-110kdb gfx1030-36kdb gfx908-120kdb gfx906-64kdb gfx906-60kdb gfx900-64kdb gfx900-56kdb)
+
+list(LENGTH ARCH_LST FULL_LEN)
+math(EXPR ARCH_LST_LEN "${FULL_LEN} - 1")
+
+
+foreach(IDX RANGE ${ARCH_LST_LEN})
+    list(GET ARCH_LST ${IDX} ARCH_NAME)
+    list(GET ARCH_FILE_LST ${IDX} ARCH_FILE_NAME)
+    install_kdb(
+        NAME ${ARCH_NAME}
+        COMPONENT_NAME ${ARCH_FILE_NAME}
+    )
+endforeach()
+
+if(LFS_MISSING_FILES)
+    string(REPLACE ";" "; " LFS_MISSING_FILES "${LFS_MISSING_FILES}")
+    message(WARNING "GIT LFS Files not pulled down, skipped: ${LFS_MISSING_FILES}")
+    set(MIOPEN_NO_LFS_PULLED TRUE CACHE INTERNAL "")
+endif()
+
+set(CPACK_COMPONENTS_ALL ${ARCH_FILE_LST})
+#end kdb package creation
 
 rocm_create_package(
     NAME MIOpen-${MIOPEN_BACKEND}
     DESCRIPTION "AMD's DNN Library"
-    MAINTAINER "Paul Fultz II <paul.fultz@amd.com>"
+    MAINTAINER "MIOpen Maintainer <miopen-lib.support@amd.com>"
     LDCONFIG
     # DEPENDS rocm-opencl hip-rocclr tinygemm
 )
@@ -536,6 +604,18 @@ enable_clang_tidy(
         -readability-qualified-auto
         -readability-redundant-string-init
         -readability-uppercase-literal-suffix
+        #TODO Code Quality WORKAROUND ROCm 5.1 update
+        -altera-id-dependent-backward-branch
+        -bugprone-easily-swappable-parameters
+        -bugprone-implicit-widening-of-multiplication-result
+        -cert-err33-c
+        -google-readability-casting
+        -hicpp-use-emplace
+        -modernize-use-emplace
+        -performance-unnecessary-copy-initialization
+        -readability-container-data-pointer
+        -readability-identifier-length
+        -readability-suspicious-call-argument
 
         ${MIOPEN_TIDY_CHECKS}
     ${MIOPEN_TIDY_ERRORS}

diff --git a/Dockerfile b/Dockerfile
@@ -1,11 +1,6 @@
 FROM ubuntu:18.04
 
-ARG PREFIX=/usr/local
-ARG GPU_ARCH=";"
-ARG MIOTENSILE_VER="default"
-ARG USE_TARGETID="OFF"
 ARG USE_MLIR="OFF"
-ARG USE_FIN="OFF"
 
 # Support multiarch
 RUN dpkg --add-architecture i386
@@ -15,24 +10,28 @@ RUN dpkg --add-architecture i386
 # unless MLIR library is incompatible with current ROCm.
 
 RUN if [ "$USE_MLIR" = "ON" ] ; \
-        then export ROCM_APT_VER=.apt_4.3.1;\
-    else export ROCM_APT_VER=.apt_4.3.1;  \
+        then export ROCM_APT_VER=.apt_5.1;\
+    else \
+        export ROCM_APT_VER=.apt_5.1;  \
     fi && \
 echo $ROCM_APT_VER &&\
-sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/$ROCM_APT_VER/ xenial main > /etc/apt/sources.list.d/rocm.list'
-RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu xenial main universe | tee -a /etc/apt/sources.list"
+sh -c 'echo deb [arch=amd64 trusted=yes] http://repo.radeon.com/rocm/apt/$ROCM_APT_VER/ ubuntu main > /etc/apt/sources.list.d/rocm.list'
+RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu bionic main universe | tee -a /etc/apt/sources.list"
 
 #Add gpg keys
 # Install dependencies
-RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
+RUN apt-get update && \
+DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
     wget \
     ca-certificates \
     curl \
     libnuma-dev \
-    gnupg && \
+    gnupg \
+    apt-utils && \
+apt-key adv --keyserver hkp://keyserver.ubuntu.com:80 --recv-keys 9386B48A1A693C5C && \
 wget -q -O - https://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
-apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
-    apt-utils \
+apt-get update && \
+DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
     build-essential \
     cmake \
     comgr \
@@ -78,6 +77,8 @@ ENV LANG=C.UTF-8
 # Install an init system
 RUN wget https://github.com/Yelp/dumb-init/releases/download/v1.2.0/dumb-init_1.2.0_amd64.deb
 RUN dpkg -i dumb-init_*.deb && rm dumb-init_*.deb
+# Install cget
+RUN pip3 install https://github.com/pfultz2/cget/archive/a426e4e5147d87ea421a3101e6a3beca541c8df8.tar.gz
 
 # Install rbuild
 RUN pip3 install https://github.com/RadeonOpenCompute/rbuild/archive/6d78a0553babdaea8d2da5de15cbda7e869594b8.tar.gz
@@ -94,21 +95,37 @@ ADD requirements.txt /requirements.txt
 ADD dev-requirements.txt /dev-requirements.txt
 # Install dependencies
 # TODO: Add --std=c++14
+ARG GPU_ARCH=";"
+ARG PREFIX=/usr/local
+ARG USE_FIN="OFF"
+ARG CCACHE_SECONDARY_STORAGE=""
+ARG CCACHE_DIR="/tmp"
+RUN env
+# RUN cget -p $PREFIX install https://github.com/ccache/ccache/archive/7f1572ae9ca958fa923a66235f6a64a360b03523.tar.gz -DZSTD_FROM_INTERNET=ON -DHIREDIS_FROM_INTERNET=ON
+RUN rm -rf /tmp/ccache* && mkdir /tmp/ccache && wget https://github.com/ccache/ccache/archive/7f1572ae9ca958fa923a66235f6a64a360b03523.tar.gz -O /tmp/ccache.tar.gz && \
+    tar zxvf /tmp/ccache.tar.gz -C /tmp/ && mkdir /tmp/ccache-7f1572ae9ca958fa923a66235f6a64a360b03523/build && \
+    cd /tmp/ccache-7f1572ae9ca958fa923a66235f6a64a360b03523/build && \
+    cmake -DZSTD_FROM_INTERNET=ON -DHIREDIS_FROM_INTERNET=ON .. && make -j install
+RUN ccache -s 
+ARG COMPILER_LAUNCHER=""
 RUN if [ "$USE_FIN" = "ON" ]; then \
-        rbuild prepare -s fin -d $PREFIX -DAMDGPU_TARGETS=${GPU_ARCH}; \
+        rbuild prepare -s fin -d $PREFIX -DAMDGPU_TARGETS=${GPU_ARCH} -DCMAKE_CXX_COMPILER_LAUNCHER="${COMPILER_LAUNCHER}"; \
     else \
-        rbuild prepare -s develop -d $PREFIX -DAMDGPU_TARGETS=${GPU_ARCH}; \
+        rbuild prepare -s develop -d $PREFIX -DAMDGPU_TARGETS=${GPU_ARCH} -DCMAKE_CXX_COMPILER_LAUNCHER="${COMPILER_LAUNCHER}"; \
     fi
 
+RUN ccache -s 
 # Install doc requirements
 ADD doc/requirements.txt /doc-requirements.txt
 RUN pip3 install -r /doc-requirements.txt
 
 # Use parallel job to accelerate tensile build
 # Workaround for Tensile with TargetID feature
+ARG USE_TARGETID="OFF"
 RUN if [ "$USE_TARGETID" = "ON" ] ; then export HIPCC_LINK_FLAGS_APPEND='-O3 -parallel-jobs=4' && export HIPCC_COMPILE_FLAGS_APPEND='-O3 -Wno-format-nonliteral -parallel-jobs=4' && rm -f /usr/bin/hipcc; fi
 
 # install last released miopentensile in default (master), install latest commits when MIOTENSILE_VER="latest" (develop)
+ARG MIOTENSILE_VER="default"
 RUN if [ "$USE_TARGETID" = "OFF" ] ; then echo "MIOpenTensile is not installed."; elif [ "$MIOTENSILE_VER" = "latest" ] ; then cget -p $PREFIX install ROCmSoftwarePlatform/MIOpenTensile@94a9047741d16a8eccd290131b78fb1aa69cdcdf; else cget -p $PREFIX install ROCmSoftwarePlatform/MIOpenTensile@94a9047741d16a8eccd290131b78fb1aa69cdcdf; fi
 
 RUN groupadd -f render