Merge pull request #142 from RadeonOpenCompute/rocm-5.2.x

ROCm 5.2.0 updates

Author: skeelyamd

LICENSE.txt

@@ -1,7 +1,7 @@
 The University of Illinois/NCSA
 Open Source License (NCSA)
 
-Copyright (c) 2014-2017, Advanced Micro Devices, Inc. All rights reserved.
+Copyright (c) 2014-2022, Advanced Micro Devices, Inc. All rights reserved.
 
 Developed by:
 

src/CMakeLists.txt

@@ -59,6 +59,8 @@ project( ${CORE_RUNTIME_TARGET} )
 list ( APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake_modules" )
 include ( utils )
 include ( hsa_common )
+# Set default libdir to be "lib" for ROCm, distros will override this anyway:
+set(CMAKE_INSTALL_LIBDIR "lib" CACHE STRING "Library install directory")
 include ( GNUInstallDirs )
 
 ## Expose static library option
@@ -117,7 +119,7 @@ endif()
 target_include_directories( ${CORE_RUNTIME_TARGET}
   PUBLIC
   $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/inc>
-  $<INSTALL_INTERFACE:include/hsa>
+  $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/hsa>
   PRIVATE
   ${CMAKE_CURRENT_SOURCE_DIR}
   ${CMAKE_CURRENT_SOURCE_DIR}/libamdhsacode )
@@ -191,10 +193,10 @@ set ( SRCS core/util/lnx/os_linux.cpp
 
 target_sources( ${CORE_RUNTIME_TARGET} PRIVATE ${SRCS} )
 
-if ( NOT DEFINED IMAGE_SUPPORT )
+if ( NOT DEFINED IMAGE_SUPPORT AND CMAKE_SYSTEM_PROCESSOR MATCHES "i?86|x86_64|amd64|AMD64" )
   set ( IMAGE_SUPPORT ON )
 endif()
-set ( IMAGE_SUPPORT ${IMAGE_SUPPORT} CACHE BOOL "Build with image support (default: ON)." )
+set ( IMAGE_SUPPORT ${IMAGE_SUPPORT} CACHE BOOL "Build with image support (default: ON for x86, OFF elsewise)." )
 
 ## Optional image module defintions.
 if(${IMAGE_SUPPORT})
@@ -203,8 +205,6 @@ if(${IMAGE_SUPPORT})
     HSA_IMAGE_SUPPORT
     UNIX_OS
     LINUX
-    __AMD64__
-    __x86_64__
     AMD_INTERNAL_BUILD
     BRAHMA_BUILD=1 )
 
@@ -275,7 +275,7 @@ if( NOT ${BUILD_SHARED_LIBS} )
 
   ## Bind to source build target interface but not its link requirements.
   target_include_directories( ${CORE_RUNTIME_NAME} INTERFACE $<TARGET_PROPERTY:${CORE_RUNTIME_NAME}::${CORE_RUNTIME_TARGET},INTERFACE_INCLUDE_DIRECTORIES> )
-  target_link_libraries ( ${CORE_RUNTIME_NAME} INTERFACE -Wl,$<INSTALL_PREFIX>/lib/cmake/${CORE_RUNTIME_NAME}/${LNKSCR}
+  target_link_libraries ( ${CORE_RUNTIME_NAME} INTERFACE -Wl,$<INSTALL_PREFIX>/${CMAKE_INSTALL_LIBDIR}/cmake/${CORE_RUNTIME_NAME}/${LNKSCR}
     -Wl,--whole-archive $<TARGET_FILE:${CORE_RUNTIME_NAME}::${CORE_RUNTIME_TARGET}> -Wl,--no-whole-archive)
   add_dependencies( ${CORE_RUNTIME_NAME} ${CORE_RUNTIME_TARGET} )
 
@@ -287,25 +287,23 @@ if( NOT ${BUILD_SHARED_LIBS} )
 endif()
 
 ## Create symlinks for legacy packaging and install
-add_custom_target ( hsa_include_link ALL WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR} COMMAND ${CMAKE_COMMAND} -E create_symlink ../../include/hsa hsa_include_link )
+add_custom_target ( hsa_include_link ALL WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR} COMMAND ${CMAKE_COMMAND} -E create_symlink ../../${CMAKE_INSTALL_INCLUDEDIR}/hsa hsa_include_link )
 if ( ${BUILD_SHARED_LIBS} )
-    add_custom_target ( hsa_lib_link ALL WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR} COMMAND ${CMAKE_COMMAND} -E create_symlink ../../lib/${CORE_RUNTIME_LIBRARY}.so ${CORE_RUNTIME_LIBRARY}-link.so )
-    add_custom_target ( hsa_lib_link2 ALL WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR} COMMAND ${CMAKE_COMMAND} -E create_symlink ../../lib/${CORE_RUNTIME_LIBRARY}.so.${VERSION_MAJOR} ${CORE_RUNTIME_LIBRARY}-link.so.${VERSION_MAJOR} )
+    add_custom_target ( hsa_lib_link ALL WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR} COMMAND ${CMAKE_COMMAND} -E create_symlink ../../${CMAKE_INSTALL_LIBDIR}/${CORE_RUNTIME_LIBRARY}.so ${CORE_RUNTIME_LIBRARY}-link.so )
+    add_custom_target ( hsa_lib_link2 ALL WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR} COMMAND ${CMAKE_COMMAND} -E create_symlink ../../${CMAKE_INSTALL_LIBDIR}/${CORE_RUNTIME_LIBRARY}.so.${VERSION_MAJOR} ${CORE_RUNTIME_LIBRARY}-link.so.${VERSION_MAJOR} )
 endif()
 
 ## Set install information
 # Installs binaries and exports the library usage data to ${HSAKMT_TARGET}Targets
-# TODO: Fix me for flat directory layout.  Should be ${CMAKE_INSTALL_LIBDIR}
 install ( TARGETS ${CORE_RUNTIME_TARGET} EXPORT ${CORE_RUNTIME_NAME}Targets
-  ARCHIVE DESTINATION lib COMPONENT binary
-  LIBRARY DESTINATION lib COMPONENT binary )
+  ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR} COMPONENT binary
+  LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR} COMPONENT binary )
 
 # Install license
 install ( FILES ${CMAKE_CURRENT_SOURCE_DIR}/LICENSE.md DESTINATION ${CMAKE_INSTALL_DOCDIR} COMPONENT binary )
 
 # Install public headers
-# TODO: Fix me for flat directory layout.  Should be ${CMAKE_INSTALL_INCLUDEDIR}
-install ( DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/inc/ DESTINATION include/hsa COMPONENT dev )
+install ( DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/inc/ DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/hsa COMPONENT dev )
 
 # Legacy symlink - not packaged (CPack is bugged until ~3.18, see https://gitlab.kitware.com/cmake/cmake/-/merge_requests/4637)
 install ( FILES ${CMAKE_CURRENT_BINARY_DIR}/hsa_include_link DESTINATION hsa/include PERMISSIONS OWNER_WRITE OWNER_READ RENAME hsa )
@@ -318,11 +316,10 @@ endif ()
 
 ## Configure and install package config file
 # Record our usage data for clients find_package calls.
-# TODO: Fix me for flat directory layout.  Should be ${CMAKE_INSTALL_LIBDIR}
 install ( EXPORT ${CORE_RUNTIME_NAME}Targets
   FILE ${CORE_RUNTIME_NAME}Targets.cmake
   NAMESPACE ${CORE_RUNTIME_NAME}::
-  DESTINATION lib/cmake/${CORE_RUNTIME_NAME}
+  DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/${CORE_RUNTIME_NAME}
   COMPONENT dev)
 
 # Adds the target alias hsa-runtime64::hsa-runtime64 to the local cmake cache.
@@ -336,28 +333,26 @@ add_library( ${CORE_RUNTIME_NAME}::${CORE_RUNTIME_NAME} ALIAS ${CORE_RUNTIME_NAM
 # Create cmake configuration files
 include(CMakePackageConfigHelpers)
 
-# TODO: Fix me for flat directory layout.  Should be ${CMAKE_INSTALL_LIBDIR}
 configure_package_config_file(${CORE_RUNTIME_NAME}-config.cmake.in
   ${CORE_RUNTIME_NAME}-config.cmake
-  INSTALL_DESTINATION lib/cmake/${CORE_RUNTIME_NAME} )
+  INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/${CORE_RUNTIME_NAME} )
 
 write_basic_package_version_file(${CORE_RUNTIME_NAME}-config-version.cmake
   VERSION ${SO_VERSION_STRING} COMPATIBILITY AnyNewerVersion )
 
-# TODO: Fix me for flat directory layout.  Should be ${CMAKE_INSTALL_LIBDIR}
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${CORE_RUNTIME_NAME}-config.cmake ${CMAKE_CURRENT_BINARY_DIR}/${CORE_RUNTIME_NAME}-config-version.cmake
-  DESTINATION lib/cmake/${CORE_RUNTIME_NAME}
+  DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/${CORE_RUNTIME_NAME}
   COMPONENT dev)
 
 # Install build files needed only when using a static build.
 if( NOT ${BUILD_SHARED_LIBS} )
   # libelf find package module
   install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/cmake_modules/FindLibElf.cmake ${CMAKE_CURRENT_SOURCE_DIR}/cmake_modules/COPYING-CMAKE-SCRIPTS
-    DESTINATION lib/cmake/${CORE_RUNTIME_NAME}
+    DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/${CORE_RUNTIME_NAME}
     COMPONENT dev)
   # Linker script (defines function aliases)
   install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/${LNKSCR}
-    DESTINATION lib/cmake/${CORE_RUNTIME_NAME}
+    DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/${CORE_RUNTIME_NAME}
     COMPONENT dev)
 endif()
 

src/DEBIAN/Binary/prerm.in

@@ -53,7 +53,7 @@ rm_ldconfig() {
 }
 
 case "$1" in
-   ( remove )
+   ( remove | upgrade)
        rm_ldconfig
    ;;
    ( * )

src/RPM/Binary/postun.in

@@ -41,7 +41,8 @@
 ################################################################################
 
 # left-hand term originates from @ENABLE_LDCONFIG@ = ON/OFF at package build
-if [ $1 -eq 0 ] && [ "@ENABLE_LDCONFIG@" == "ON" ]; then
+if [ $1 -le 1 ] && [ "@ENABLE_LDCONFIG@" == "ON" ]; then
+    # perform the below actions for rpm remove($1=0) or upgrade($1=1) operations
     rm -f /etc/ld.so.conf.d/hsa-rocr.conf
     ldconfig
 fi

src/core/inc/amd_loader_context.hpp

@@ -48,9 +48,9 @@
 namespace rocr {
 namespace amd {
 
-class LoaderContext final: public amd::hsa::loader::Context {
-public:
-  LoaderContext(): amd::hsa::loader::Context() {}
+class LoaderContext final : public rocr::amd::hsa::loader::Context {
+ public:
+  LoaderContext() : rocr::amd::hsa::loader::Context() {}
 
   ~LoaderContext() {}
 

src/core/runtime/amd_aql_queue.cpp

@@ -866,9 +866,17 @@ bool AqlQueue::DynamicScratchHandler(hsa_signal_value_t error_code, void* arg) {
       const uint64_t rounds = groups / cu_count;
       const uint64_t asymmetricGroups = rounds * asymmetryPerRound;
       const uint64_t symmetricGroups = groups - asymmetricGroups;
-      const uint64_t maxGroupsPerEngine =
+      uint64_t maxGroupsPerEngine =
           ((symmetricGroups + engines - 1) / engines) + (asymmetryPerRound ? rounds : 0);
 
+      // For gfx10+ devices we must attempt to assign the smaller of 256 lanes or 16 groups to each
+      // engine.
+      if (queue->agent_->isa()->GetMajorVersion() >= 10 && maxGroupsPerEngine < 16 &&
+          lanes_per_group * maxGroupsPerEngine < 256) {
+        uint64_t groups_per_interleave = (256 + lanes_per_group - 1) / lanes_per_group;
+        maxGroupsPerEngine = Min(groups_per_interleave, 16ul);
+      }
+
       // Populate all engines at max group occupancy, then clip down to device limits.
       groups = maxGroupsPerEngine * engines;
       scratch.wanted_slots = groups * waves_per_group;
@@ -1042,6 +1050,13 @@ bool AqlQueue::ExceptionHandler(hsa_signal_value_t error_code, void* arg) {
   // Undefined or unexpected code
   assert((errorCode != HSA_STATUS_ERROR) && "Undefined or unexpected queue error code");
 
+  // Suppress VM fault reporting.  This is more useful when reported through the system error
+  // handler.
+  if (errorCode == HSA_STATUS_ERROR_MEMORY_FAULT) {
+    debug_print("Queue error - HSA_STATUS_ERROR_MEMORY_FAULT\n");
+    return false;
+  }
+
   queue->Suspend();
   if (queue->errors_callback_ != nullptr) {
     queue->errors_callback_(errorCode, queue->public_handle(), queue->errors_data_);

src/core/runtime/amd_gpu_agent.cpp

@@ -1015,6 +1015,7 @@ hsa_status_t GpuAgent::GetInfo(hsa_agent_info_t attribute, void* value) const {
     case HSA_AMD_AGENT_INFO_SVM_DIRECT_HOST_ACCESS:
       assert(regions_.size() != 0 && "No device local memory found!");
       *((bool*)value) = properties_.Capability.ui32.CoherentHostAccess == 1;
+      break;
     case HSA_AMD_AGENT_INFO_COOPERATIVE_COMPUTE_UNIT_COUNT:
       if (core::Runtime::runtime_singleton_->flag().coop_cu_count() &&
           (isa_->GetMajorVersion() == 9) && (isa_->GetMinorVersion() == 0) &&