Squashed 'fin/' changes from 53d2563fe..30d699b9e

30d699b9e Perf Eval Update (#60) 3535b948c PerfCompile and PerfEval changes (#59) de79468d2 remove unneccessary solution check, add check for previously modified kernel names (#56) 6924286a2 miopen hash update (#55) 530399575 Refactor googletest infra to align with MIOpen (#53) 71c50d146 Datatype fix for BN (#57) 8abe2f5c6 Perf Eval updates, Add find info (#51) e1c1ef0f5 filter find compile by solver input (#54) 722feea66 sp/chk precomp kernel 264 (#41) b9aba2034 Batch norm find compile (#50) 359f3da80 Fix missing link directives in fin binary (#48) a4020c1ba Cache Miss Fixes (#46) 2ec7ef44d Enable google test and compiling fin in the CI (#47) 8b6b453bc Applicability support for batch norm (#45) 44323aae9 Perf compile/eval for fin (#42) ebd9aa6bd update member name (#43) d6d798efe add cu count (#39) 8e1989a9f Add find option for selecting only dynamic solvers (#38) 0e164bf66 setting json version (#37) f3f7fed18 Remove function redefinition (#36) e1de51a58 Performance DB de-serialize test (#34) 043cdcdaa Layout support in Fin (#33) 3a1d58236 Hotfix (#32) ee3f0d543 4.4 Tuning Bugfixes (#31) 832dbe234 Tunability Reporting (#27) a564a229f include gfx90a_110 (#28) git-subtree-dir: fin git-subtree-split: 30d699b9edc014c6076a9649f849bd3c4588d4ab
ROCm · Aug 3, 2022 · 41b34da · 41b34da
1 parent 8e0a630
commit 41b34da
Show file tree

Hide file tree

Showing 28 changed files with 4,961 additions and 1,096 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -39,8 +39,6 @@ else()
 endif()
 
 project ( fin C CXX )
-option( BUILD_SHARED_LIBS "Build as a shared library" ON )
-enable_testing()
 
 find_package(ROCM REQUIRED PATHS /opt/rocm)
 
@@ -53,12 +51,6 @@ include(CheckCXXCompilerFlag)
 
 option( BUILD_DEV "Build for development only" OFF)
 
-# Strip symbols for release
-if(NOT WIN32 AND NOT APPLE)
-    set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} -s")
-    set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -s")
-endif()
-
 rocm_setup_version(VERSION 1.0.0)
 
 list( APPEND CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake )
@@ -100,15 +92,6 @@ set_property( CACHE FIN_BACKEND PROPERTY STRINGS
 
 # HIP is always required
 find_package(hip REQUIRED PATHS /opt/rocm)
-find_package(rocblas REQUIRED PATHS /opt/rocm)
-# probably not required
-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}")
-
-message(STATUS "Hip compiler flags: ${HIP_COMPILER_FLAGS}")
-
-add_definitions("-DHIP_COMPILER_FLAGS=${HIP_COMPILER_FLAGS}")
 
 # OpenCL 1.2
 if( FIN_BACKEND STREQUAL "OpenCL")
@@ -194,7 +177,7 @@ set(CPACK_RPM_PACKAGE_REQUIRES "${MIOPEN_PACKAGE_REQS}, rocm-opencl-devel")
 rocm_create_package(
     NAME fin-${FIN_BACKEND}
     DESCRIPTION "Fin MIOpen Driver"
-    MAINTAINER "Jehandad Khan <jehandad.khan@amd.com>"
+    MAINTAINER "Tuna Devs <dl.MITunaX@amd.com>"
     LDCONFIG
     # DEPENDS rocm-opencl rocm-utils hip-hcc tinygemm
 )
@@ -208,6 +191,8 @@ elseif(FIN_BACKEND STREQUAL "HIP")
     set(MIOPEN_TIDY_ERRORS ALL)
 endif()
 
+if(NOT MIOPEN_ENABLE_FIN)
+    message("MIOPEN_ENABLE_FIN is NOT SET")
 include(ClangTidy)
 enable_clang_tidy(
     CHECKS
@@ -329,9 +314,13 @@ enable_cppcheck(
         MIOPEN_USE_MIOPENGEMM=1
         __linux__=1
 )
+else()
+    message("MIOPEN_ENABLE_FIN is SET")
+endif()
 
-set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
-set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
-set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/bin)
+find_package(miopen)
+find_package(rocblas) #MIOpen depends on rocBlas
+find_package(Threads REQUIRED)
 
 add_subdirectory(src)
+add_subdirectory(tests)
diff --git a/Dockerfile b/Dockerfile
@@ -1,14 +1,19 @@
 FROM ubuntu:18.04
 
-ARG PREFIX=/usr/local
+ARG PREFIX=/opt/rocm
 
 # Support multiarch
 RUN dpkg --add-architecture i386
 
 # Add rocm repository
 RUN apt-get update
 RUN DEBIAN_FRONTEND=noninteractive apt-get install -y curl apt-utils wget gnupg2
-RUN curl https://raw.githubusercontent.com/RadeonOpenCompute/ROCm-docker/master/add-rocm.sh | bash
+
+#RUN curl https://raw.githubusercontent.com/RadeonOpenCompute/ROCm-docker/master/add-rocm.sh | bash
+ARG ROCMVERSION=4.5
+ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/
+RUN wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -
+RUN sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"
 
 # Install dependencies required to build hcc
 # Ubuntu csomic contains llvm-7 required to build Tensile
@@ -69,21 +74,42 @@ RUN pip install https://github.com/pfultz2/cget/archive/57b3289000fcdb3b7e424c60
 # Install rclone
 RUN pip install https://github.com/pfultz2/rclone/archive/master.tar.gz
 
-# Install hcc from ROCm 3.0
-RUN rclone -b roc-3.0.x -c 286651a04d9c3a8e3052dd84b1822985498cd27d https://github.com/RadeonOpenCompute/hcc.git /hcc
-RUN LDFLAGS=-fuse-ld=gold cget -p $PREFIX install hcc,/hcc  && rm -rf /hcc
-
 # Make sure /opt/rcom is in the paths
 ENV PATH="/opt/rocm:${PATH}"
+# Install MIOpen
+ARG MIOPEN_DIR=/root/dMIOpen
+#Clone MIOpen
+RUN git clone https://github.com/ROCmSoftwarePlatform/MIOpen.git $MIOPEN_DIR
+WORKDIR $MIOPEN_DIR
+
+ARG MIOPEN_DEPS=$MIOPEN_DIR/cget
+# Install dependencies
+RUN wget --no-check-certificate -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | apt-key add -
+RUN sh -c "echo deb https://apt.kitware.com/ubuntu/ bionic main | tee -a /etc/apt/sources.list"
+RUN apt-get update && apt-get install -y --allow-unauthenticated   cmake-data=3.15.1-0kitware1 cmake=3.15.1-0kitware1
+RUN cget install pfultz2/rocm-recipes
+RUN cget install -f min-requirements.txt
+RUN CXXFLAGS='-isystem $PREFIX/include' cget install -f ./mlir-requirements.txt
 
-# Build using hcc
-RUN cget -p $PREFIX init --cxx $PREFIX/bin/hcc --std=c++14
+ARG TUNA_USER=miopenpdb
+ARG BACKEND=HIP
+# Build MIOpen
+WORKDIR $MIOPEN_DIR/build
+ARG MIOPEN_BRANCH=05169b091cf3d0b752c94ba51d63f7af6463d731
+ARG MIOPEN_CACHE_DIR=/tmp/${TUNA_USER}/cache
+ARG MIOPEN_USER_DB_PATH=/tmp/$TUNA_USER/config/miopen
+ARG MIOPEN_USE_MLIR=On
+RUN git pull && git checkout $MIOPEN_BRANCH
+RUN echo "MIOPEN: Selected $BACKEND backend."
+RUN if [ $BACKEND = "OpenCL" ]; then \
+           cmake -DMIOPEN_INSTALL_CXX_HEADERS=On -DMIOPEN_INSTALL_CXX_HEADERS=On -DMIOPEN_USE_COMGR=Off -DMIOPEN_USE_MLIR=${MIOPEN_USE_MLIR} -DMIOPEN_CACHE_DIR=${MIOPEN_CACHE_DIR} -DMIOPEN_USER_DB_PATH={MIOPEN_USER_DB_PATH} -DMIOPEN_BACKEND=OpenCL -DMIOPEN_HIP_COMPILER=/opt/rocm/llvm/bin/clang++ -DCMAKE_PREFIX_PATH="$MIOPEN_DEPS" $MIOPEN_DIR ; \
+    else \
+           CXX=/opt/rocm/llvm/bin/clang++ cmake -DMIOPEN_INSTALL_CXX_HEADERS=On -DMIOPEN_USE_COMGR=Off -DMIOPEN_USE_MLIR=${MIOPEN_USE_MLIR} -DMIOPEN_INSTALL_CXX_HEADERS=On -DMIOPEN_CACHE_DIR=${MIOPEN_CACHE_DIR} -DMIOPEN_USER_DB_PATH=${MIOPEN_USER_DB_PATH} -DMIOPEN_BACKEND=$BACKEND -DCMAKE_PREFIX_PATH=$MIOPEN_DEPS $MIOPEN_DIR ; \
+    fi
+
+RUN make -j $(nproc)
+RUN make install
 
 # Install dependencies
 ADD requirements.txt /requirements.txt
 RUN CXXFLAGS='-isystem $PREFIX/include' cget -p $PREFIX install -f /requirements.txt
-
-# Install doc requirements
-#ADD doc/requirements.txt /doc-requirements.txt
-#RUN pip install -r /doc-requirements.txt
-
diff --git a/Jenkinsfile b/Jenkinsfile
@@ -22,9 +22,8 @@ def rocmnode(name) {
 
 def cmake_build(compiler, flags, prefixpath="/opt/rocm"){
     def workspace_dir = pwd()
-    def vcache = "/var/jenkins/.cache/miopen/vcache"
     def archive = (flags == '-DCMAKE_BUILD_TYPE=release')
-    def config_targets = "all" // "check doc MIOpenDriver"
+    def config_targets = "all" 
     def test_flags = "--disable-verification-cache"
     def debug_flags = "-g -fno-omit-frame-pointer -fsanitize=undefined -fno-sanitize-recover=undefined"
     def compilerpath = ""
@@ -46,8 +45,8 @@ def cmake_build(compiler, flags, prefixpath="/opt/rocm"){
         rm -rf build
         mkdir build
         cd build
-        CXX=${compilerpath} CXXFLAGS='-Werror' cmake ${configargs} -DMIOPEN_GPU_SYNC=On -DMIOPEN_TEST_FLAGS='${test_flags}' -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}' ${flags} ..
-        MIOPEN_DEBUG_CONV_IMPLICIT_GEMM_XDLOPS=1 CTEST_PARALLEL_LEVEL=4 MIOPEN_VERIFY_CACHE_PATH=${vcache} MIOPEN_CONV_PRECISE_ROCBLAS_TIMING=0 dumb-init make -j\$(nproc) ${config_targets}
+        CXX=${compilerpath} cmake ${configargs}  ${flags} ..
+        dumb-init make -j\$(nproc) ${config_targets}
     """
     echo cmd
     sh cmd
@@ -110,26 +109,12 @@ pipeline {
         parallelsAlwaysFailFast()
     }
     environment{
-        image = "miopen"
+        image = "fin"
     }
     stages{
         // Run all static analysis tests
         stage("Static checks"){
             parallel{
-                stage('Clang Tidy') {
-                    agent{  label rocmnode("rocmtest") }
-                    environment{
-                        cmd = "rm -rf build; \
-                                mkdir build; \
-                                cd build; \
-                                CXX='clang++-3.9' cmake -DBUILD_DEV=On ..; \
-                                make -j\$(nproc) -k analyze;"
-                    }
-                    steps{
-                        buildJob('clang++-3.9', '-DCMAKE_BUILD_TYPE=release', image, "", cmd)
-                    }
-                }
-
                 stage('Clang Format') {
                     agent{ label rocmnode("rocmtest") }
                     environment{
@@ -155,11 +140,37 @@ pipeline {
                         cmd = "rm -rf build; \
                                 mkdir build; \
                                 cd build; \
-                                CXX=/usr/local/bin/hcc cmake -DBUILD_DEV=On ..; \
+                                CXX=/opt/rocm/llvm/bin/clang++ cmake -DBUILD_DEV=On ..; \
                                 make -j\$(nproc) -k analyze;"
                     }
                     steps{
-                        buildJob('hcc', '-DCMAKE_BUILD_TYPE=release', image, "", cmd)
+                        buildJob('clang++', '-DCMAKE_BUILD_TYPE=release', image, "", cmd)
+                    }
+                }
+                stage('Build Fin') {
+                    agent{ label rocmnode("rocmtest") }
+                    environment{
+                        cmd = "rm -rf build; \
+                                mkdir build; \
+                                cd build; \
+                                CXX=/opt/rocm/llvm/bin/clang++ cmake -DBUILD_DEV=On -DCMAKE_PREFIX_PATH=/root/dMIOpen/cget ..; \
+                                make -j\$(nproc) all;"
+                    }
+                    steps{
+                        buildJob('clang++', '-DCMAKE_BUILD_TYPE=release', image, "", cmd)
+                    }
+                }
+                stage('Fin Tests') {
+                    agent{ label rocmnode("rocmtest") }
+                    environment{
+                        cmd = "rm -rf build; \
+                                mkdir build; \
+                                cd build; \
+                                CXX=/opt/rocm/llvm/bin/clang++ cmake -DBUILD_DEV=On -DCMAKE_PREFIX_PATH=/root/dMIOpen/cget ..; \
+                                make -j\$(nproc) check;"
+                    }
+                    steps{
+                        buildJob('clang++', '-DCMAKE_BUILD_TYPE=release', image, "", cmd)
                     }
                 }
             }

diff --git a/cmake/EnableCompilerWarnings.cmake b/cmake/EnableCompilerWarnings.cmake
@@ -24,7 +24,6 @@
 #
 ################################################################################
 # - Enable warning all for gcc/clang or use /W4 for visual studio
-
 ## Strict warning level
 if (MSVC)
     # Use the highest warning level for visual studio.

diff --git a/cmake/googletest.cmake b/cmake/googletest.cmake
@@ -0,0 +1,39 @@
+include(FetchContent)
+
+set(GOOGLETEST_DIR "" CACHE STRING "Location of local GoogleTest repo to build against")
+
+if(GOOGLETEST_DIR)
+  set(FETCHCONTENT_SOURCE_DIR_GOOGLETEST ${GOOGLETEST_DIR} CACHE STRING "GoogleTest source directory override")
+endif()
+
+message(STATUS "Fetching GoogleTest")
+
+list(APPEND GTEST_CMAKE_CXX_FLAGS 
+     -Wno-undef
+     -Wno-reserved-identifier
+     -Wno-global-constructors
+     -Wno-missing-noreturn
+     -Wno-disabled-macro-expansion
+     -Wno-used-but-marked-unused
+     -Wno-switch-enum
+     -Wno-zero-as-null-pointer-constant
+     -Wno-unused-member-function
+     -Wno-comma
+     -Wno-old-style-cast
+)
+message(STATUS "Suppressing googltest warnings with flags: ${GTEST_CMAKE_CXX_FLAGS}")
+
+FetchContent_Declare(
+  googletest
+  GIT_REPOSITORY https://github.com/google/googletest.git
+  GIT_TAG        e2239ee6043f73722e7aa812a459f54a28552929
+)
+
+# Will be necessary for windows build
+# set(gtest_force_shared_crt ON CACHE BOOL "" FORCE)
+FetchContent_MakeAvailable(googletest)
+
+target_compile_options(gtest PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gtest_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gmock PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gmock_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
diff --git a/requirements.txt b/requirements.txt
@@ -1,5 +1,3 @@
-RadeonOpenCompute/rocm-cmake@cdd0f632b3a65bd4411593bb827eb664e25c80bc --build
-# ROCmSoftwarePlatform/MIOpen@9b2d37f9f1e4b5492ef8256cf8168363ca5fd2da 
-nlohmann/json
-#ROCmSoftwarePlatform/rocBLAS@9790a8658341bc665c11c311129ad0dfc533d5c4
+# RadeonOpenCompute/rocm-cmake@cdd0f632b3a65bd4411593bb827eb664e25c80bc --build
+nlohmann/json@v3.10.4
 
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -25,16 +25,13 @@
 ################################################################################
 cmake_minimum_required( VERSION 2.8.12 )
 list(APPEND CMAKE_PREFIX_PATH /opt/rocm /opt/rocm/hip /opt/rocm/hcc /opt/rocm/opencl)
-find_package(miopen)
-
-find_package(Threads REQUIRED)
 
 configure_file("${PROJECT_SOURCE_DIR}/src/include/config.h.in" "${PROJECT_BINARY_DIR}/src/include/config.h")
 
 include_directories(include "${PROJECT_BINARY_DIR}/src/include")
 add_executable(fin main.cpp fin.cpp base64.cpp)
 target_compile_definitions( fin PRIVATE -D__HIP_PLATFORM_HCC__=1 )
-target_link_libraries(fin MIOpen ${Boost_LIBRARIES})
+target_link_libraries(fin MIOpen ${Boost_LIBRARIES} hip::host)
 target_link_libraries(fin ${CMAKE_THREAD_LIBS_INIT})
 if(rocblas_FOUND)
     target_link_libraries( fin $<BUILD_INTERFACE:roc::rocblas> )

diff --git a/src/fin.cpp b/src/fin.cpp
@@ -22,7 +22,7 @@ miopenDataType_t GetDataType<bfloat16>()
 {
     return miopenBFloat16;
 }
-[[gnu::noreturn]] void Fin::Usage()
+[[gnu::noreturn]] void BaseFin::Usage()
 {
     std::cout << "Usage: ./MIOpenFin *base_arg* *other_args*\n";
     std::cout << "Supported Base Arguments: conv[fp16][bfp16]\n";
@@ -37,7 +37,7 @@ void PadBufferSize(size_t& sz, int datatype_sz)
     }
 }
 
-std::string Fin::ParseBaseArg(const int argc, const char* argv[])
+std::string BaseFin::ParseBaseArg(const int argc, const char* argv[])
 {
     if(argc < 2)
     {
@@ -57,4 +57,51 @@ std::string Fin::ParseBaseArg(const int argc, const char* argv[])
     else
         return arg;
 }
+void BaseFin::InitNoGpuHandle(miopen::Handle& handle,
+                              const std::string& arch,
+                              const unsigned long num_cu)
+{
+#if MIOPEN_MODE_NOGPU
+    handle.impl->device_name        = arch;
+    handle.impl->num_cu             = num_cu;
+    handle.impl->max_mem_alloc_size = 32UL * 1024 * 1024 * 1024; // 32 GB
+    handle.impl->global_mem_size    = 32UL * 1024 * 1024 * 1024;
+    handle.impl->target_properties.Init(&handle);
+#else
+    std::ignore = handle;
+    std::ignore = arch;
+    std::ignore = num_cu;
+#endif
+}
+
+void BaseFin::VerifyDevProps(const std::string& in_arch, const unsigned long in_num_cu)
+{
+    std::cerr << "Verifying device properties" << std::endl;
+    std::string arch    = in_arch;
+    arch                = arch.substr(0, arch.find(':'));
+    const size_t num_cu = in_num_cu;
+    std::ignore         = num_cu;
+    if(arch == "gfx900")
+    {
+        assert(num_cu == 56 || num_cu == 64);
+    }
+    else if(arch == "gfx906")
+    {
+        assert(num_cu == 60 || num_cu == 64);
+    }
+    else if(arch == "gfx908")
+    {
+        assert(num_cu == 120);
+    }
+    else if(arch == "gfx1030")
+    {
+        assert(num_cu == 72 || num_cu == 36);
+    }
+    else if(arch == "gfx90a")
+    {
+        assert(num_cu == 110 || num_cu == 104);
+    }
+    else
+        throw std::runtime_error("Invalid Arch Name");
+}
 } // namespace fin