SWDEV-222307 - Add CMake support for PAL

Remove dependency on COMGR headers from backends.
Remove GFX103 PAL defines.

Change-Id: I6ff76dca2a0c84bc191e8fc8b639f2f5872d2085


[ROCm/clr commit: 1be1e69f2a]

projects/clr/rocclr/device/devkernel.cpp

@@ -1123,7 +1123,11 @@ static inline cl_kernel_arg_type_qualifier GetOclTypeQualOCL(const aclArgData* a
 
 // ================================================================================================
 #if defined(USE_COMGR_LIBRARY)
-bool Kernel::GetAttrCodePropMetadata(const amd_comgr_metadata_node_t kernelMetaNode) {
+bool Kernel::GetAttrCodePropMetadata() {
+  amd_comgr_metadata_node_t kernelMetaNode;
+  if (!prog().getKernelMetadata(name(), &kernelMetaNode)) {
+    return false;
+  }
 
   InitParameters(kernelMetaNode);
 

projects/clr/rocclr/device/devkernel.hpp

@@ -511,7 +511,7 @@ class Kernel : public amd::HeapObject {
   void InitParameters(const amd_comgr_metadata_node_t kernelMD);
 
   //! Retrieve kernel attribute and code properties metadata
-  bool GetAttrCodePropMetadata(const amd_comgr_metadata_node_t kernelMetaNode);
+  bool GetAttrCodePropMetadata();
 
   //! Retrieve the available SGPRs and VGPRs
   bool SetAvailableSgprVgpr(const std::string& targetIdent);

projects/clr/rocclr/device/devprogram.hpp

@@ -220,9 +220,13 @@ class Program : public amd::HeapObject {
   amd_comgr_metadata_node_t metadata() const { return metadata_; }
 
   //! Get the kernel metadata
-  const amd_comgr_metadata_node_t* getKernelMetadata(const std::string name) const {
+  const bool getKernelMetadata(const std::string name, amd_comgr_metadata_node_t* meta) const {
     auto it = kernelMetadataMap_.find(name);
-    return (it == kernelMetadataMap_.end()) ? nullptr : &(it->second);
+    if (it != kernelMetadataMap_.end()) {
+      *meta = it->second;
+      return true;
+    }
+    return false;
   }
 
   const uint32_t codeObjectVer() const { return codeObjectVer_; }

projects/clr/rocclr/device/pal/CMakeLists.txt

@@ -72,17 +72,13 @@ target_compile_definitions(vdipal PRIVATE PAL_BUILD_RENOIR=1)
 
 # GFX10 targets
 target_compile_definitions(vdipal PRIVATE PAL_BUILD_GFX10=1)
-target_compile_definitions(vdipal PRIVATE PAL_BUILD_GFX103=1)
 target_compile_definitions(vdipal PRIVATE PAL_BUILD_NAVI12=1)
 target_compile_definitions(vdipal PRIVATE PAL_BUILD_NAVI14=1)
 
 target_compile_definitions(vdipal PRIVATE OPENCL_MAINLINE=1)
 
 if(${USE_COMGR_LIBRARY} MATCHES "yes")
-  target_compile_definitions(vdipal PRIVATE USE_COMGR_LIBRARY -DCOMGR_DYN_DLL)
-  if (DEFINED COMGR_INCLUDES AND NOT ${COMGR_INCLUDES} STREQUAL "")
-    target_include_directories(${COMGR_INCLUDES})
-  endif() # if (DEFINED COMGR_INCLUDES AND NOT ${COMGR_INCLUDES} STREQUAL "")
+  target_compile_definitions(vdipal PRIVATE USE_COMGR_LIBRARY)
 endif()
 
 target_include_directories(vdipal PRIVATE ${CMAKE_SOURCE_DIR}/compiler/sc/HSAIL/include)

projects/clr/rocclr/device/pal/palcounters.cpp

@@ -154,7 +154,7 @@ bool PalCounterReference::finalize() {
   }
 }
 
-static const std::array<PCIndexSelect, 49> blockIdToIndexSelect = {{
+static const std::array<PCIndexSelect, 46> blockIdToIndexSelect = {{
     PCIndexSelect::None,                     // CPF
     PCIndexSelect::ShaderEngine,             // IA
     PCIndexSelect::ShaderEngine,             // VGT
@@ -201,9 +201,11 @@ static const std::array<PCIndexSelect, 49> blockIdToIndexSelect = {{
     PCIndexSelect::None,                     // GCR
     PCIndexSelect::None,                     // PH
     PCIndexSelect::ShaderArray,              // UTCL1
+#if 0
     PCIndexSelect::None,                     // GeDist
     PCIndexSelect::ShaderEngine,             // GeSe
     PCIndexSelect::None,                     // Df
+#endif
 }};
 
 static_assert(blockIdToIndexSelect.size() ==  static_cast<size_t>(Pal::GpuBlock::Count), "size of blockIdToIndexSelect does not match GpuBlock::Count");
@@ -690,7 +692,6 @@ void PerfCounter::convertInfo() {
       }
       break;
     case Pal::GfxIpLevel::GfxIp10_1:
-    case Pal::GfxIpLevel::GfxIp10_3:
       if (info_.blockIndex_ < gfx10BlockIdPal.size()) {
         auto p = gfx10BlockIdPal[info_.blockIndex_];
         info_.blockIndex_ = std::get<0>(p);

projects/clr/rocclr/device/pal/paldevice.cpp

@@ -197,7 +197,7 @@ bool NullDevice::init() {
             ShouldNotReachHere();
             break;
           case 3:
-            ipLevel = Pal::GfxIpLevel::GfxIp10_3;
+            ShouldNotReachHere();
             break;
           case 4:
             ShouldNotReachHere();

projects/clr/rocclr/device/pal/palkernel.cpp

@@ -426,12 +426,7 @@ bool LightningKernel::init() {
   flags_.internalKernel_ =
       (compileOptions_.find("-cl-internal-kernel") != std::string::npos) ? true : false;
 
-  const amd_comgr_metadata_node_t* kernelMetaNode = prog().getKernelMetadata(name());
-  if (kernelMetaNode == nullptr) {
-    return false;
-  }
-
-  if (!GetAttrCodePropMetadata(*kernelMetaNode)) {
+  if (!GetAttrCodePropMetadata()) {
     return false;
   }
 

projects/clr/rocclr/device/pal/palresource.cpp

@@ -472,10 +472,10 @@ void Resource::memTypeToHeap(Pal::GpuMemoryCreateInfo* createInfo) {
   switch (memoryType()) {
     case Local:
     case Scratch:
-      createInfo->mallPolicy = static_cast<Pal::GpuMemMallPolicy>(dev().settings().mallPolicy_);
+//      createInfo->mallPolicy = static_cast<Pal::GpuMemMallPolicy>(dev().settings().mallPolicy_);
       break;
     default:
-      createInfo->mallPolicy = Pal::GpuMemMallPolicy::Never;
+//      createInfo->mallPolicy = Pal::GpuMemMallPolicy::Never;
       break;
   }
 }
@@ -1951,7 +1951,7 @@ bool MemorySubAllocator::CreateChunk(const Pal::IGpuMemory* reserved_va) {
   createInfo.heapCount = 1;
   createInfo.heaps[0] = Pal::GpuHeapInvisible;
   createInfo.flags.peerWritable = device_->P2PAccessAllowed();
-  createInfo.mallPolicy = static_cast<Pal::GpuMemMallPolicy>(device_->settings().mallPolicy_);
+//  createInfo.mallPolicy = static_cast<Pal::GpuMemMallPolicy>(device_->settings().mallPolicy_);
   GpuMemoryReference* mem_ref = GpuMemoryReference::Create(*device_, createInfo);
   if (mem_ref != nullptr) {
     return InitAllocator(mem_ref);
@@ -1971,7 +1971,7 @@ bool CoarseMemorySubAllocator::CreateChunk(const Pal::IGpuMemory* reserved_va) {
   createInfo.heapCount = 2;
   createInfo.heaps[0] = Pal::GpuHeapInvisible;
   createInfo.heaps[1] = Pal::GpuHeapLocal;
-  createInfo.mallPolicy = static_cast<Pal::GpuMemMallPolicy>(device_->settings().mallPolicy_);
+//  createInfo.mallPolicy = static_cast<Pal::GpuMemMallPolicy>(device_->settings().mallPolicy_);
   GpuMemoryReference* mem_ref = GpuMemoryReference::Create(*device_, createInfo);
   if (mem_ref != nullptr) {
     return InitAllocator(mem_ref);
@@ -1987,7 +1987,7 @@ bool FineMemorySubAllocator::CreateChunk(const Pal::IGpuMemory* reserved_va) {
   createInfo.alignment = MaxGpuAlignment;
   createInfo.flags.useReservedGpuVa = (reserved_va != nullptr);
   createInfo.pReservedGpuVaOwner = reserved_va;
-  createInfo.mallPolicy = Pal::GpuMemMallPolicy::Never;
+//  createInfo.mallPolicy = Pal::GpuMemMallPolicy::Never;
   GpuMemoryReference* mem_ref = GpuMemoryReference::Create(*device_, createInfo);
   if ((mem_ref != nullptr) && InitAllocator(mem_ref)) {
     mem_ref->iMem()->Map(&mem_ref->cpuAddress_);

projects/clr/rocclr/device/pal/palsettings.cpp

@@ -493,7 +493,6 @@ bool Settings::create(const Pal::DeviceProperties& palProp,
 
   if (useLightning_) {
     switch (palProp.gfxLevel) {
-      case Pal::GfxIpLevel::GfxIp10_3:
       case Pal::GfxIpLevel::GfxIp10_1:
       case Pal::GfxIpLevel::GfxIp9:
         singleFpDenorm_ = true;

projects/clr/rocclr/device/rocm/CMakeLists.txt

@@ -1,16 +1,10 @@
-include(bc2h)
-
-set(INC_SUFFIX "amdgcn.inc")
-
 include_directories(${CMAKE_SOURCE_DIR}/compiler/lib/loaders/elf/utils/common)
 include_directories(${CMAKE_SOURCE_DIR}/compiler/lib/loaders/elf/utils/libelf)
 include_directories(${CMAKE_CURRENT_BINARY_DIR})
 include_directories(${ROCM_OCL_INCLUDES})
 
 if(${USE_COMGR_LIBRARY} MATCHES "yes")
-  add_definitions(-DUSE_COMGR_LIBRARY -DCOMGR_DYN_DLL)
-  find_package(amd_comgr REQUIRED CONFIG)
-  include_directories("$<TARGET_PROPERTY:amd_comgr,INTERFACE_INCLUDE_DIRECTORIES>")
+  add_definitions(-DUSE_COMGR_LIBRARY)
 endif()
 
 add_library(oclrocm OBJECT
@@ -30,83 +24,3 @@ add_library(oclrocm OBJECT
 )
 set_target_properties(oclrocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
-# FIXME: Remove following if block after enabling COMGR by default
-if (${USE_COMGR_LIBRARY} STREQUAL "no")
-  # generating libraries.amdgcn.inc
-  file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/libraries.amdgcn.inc "// Automatically generated file; DO NOT EDIT.\n")
-
-  foreach(AMDGCN_LIB_TARGET ${AMDGCN_LIB_TARGETS})
-    get_target_property(lib_file ${AMDGCN_LIB_TARGET} OUTPUT_NAME)
-    get_target_property(lib_file_name ${AMDGCN_LIB_TARGET} ARCHIVE_OUTPUT_NAME)
-    get_target_property(lib_file_path ${AMDGCN_LIB_TARGET} ARCHIVE_OUTPUT_DIRECTORY)
-    set(bclib "${lib_file}")
-    set(header "${lib_file_name}.${INC_SUFFIX}")
-    set(symbol "${lib_file_name}_lib")
-    add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/${header}
-      COMMAND bc2h ${bclib} ${CMAKE_CURRENT_BINARY_DIR}/${header} ${symbol}
-      DEPENDS bc2h ${AMDGCN_LIB_TARGET}
-      COMMENT "Generating ${header}"
-    )
-    set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${CMAKE_CURRENT_BINARY_DIR}/${target}.inc)
-
-    add_custom_target(${header}_target ALL DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${header})
-    add_dependencies(oclrocm  ${header}_target)
-
-    file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/libraries.amdgcn.inc "#include \"${header}\"\n")
-  endforeach()
-
-  # Generate function to select libraries for a given GFXIP number.
-  file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/libraries.amdgcn.inc
-    "static inline std::tuple<const char*, const void*, size_t> get_oclc_isa_version(uint gfxip) { \
-     switch (gfxip) {\n")
-  foreach(AMDGCN_LIB_TARGET ${AMDGCN_LIB_TARGETS})
-    if (${AMDGCN_LIB_TARGET} MATCHES "^oclc_isa_version_[0-9]+_lib$")
-      string(REGEX REPLACE "^oclc_isa_version_([0-9]+)_lib$" "\\1" gfxip ${AMDGCN_LIB_TARGET})
-      file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/libraries.amdgcn.inc
-        "case ${gfxip}: return std::make_tuple( \"oclc_isa_version_${gfxip}_lib.bc\","
-                  "  oclc_isa_version_${gfxip}_lib, oclc_isa_version_${gfxip}_lib_size); break;\n")
-    endif()
-  endforeach()
-  file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/libraries.amdgcn.inc
-    "default: return std::make_tuple(\"\",(const void*)0,(size_t)0); }\n}\n")
-
-  foreach(AMDGCN_LIB_TARGET ${AMDGCN_LIB_TARGETS})
-    if (${AMDGCN_LIB_TARGET} MATCHES "oclc_(.*)_on_lib")
-      string(REGEX REPLACE "oclc_(.*)_on_lib" "\\1" function ${AMDGCN_LIB_TARGET})
-      file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/libraries.amdgcn.inc
-        "static inline std::tuple<const char*,const void*,size_t>  get_oclc_${function}(bool on)"
-                  " { return std::make_tuple(  on ? \"oclc_${function}_on_lib.bc\" : \"oclc_${function}_off_lib.bc\","
-                  "  (const void*)(on ? oclc_${function}_on_lib : oclc_${function}_off_lib),"
-                  " on ? oclc_${function}_on_lib_size : oclc_${function}_off_lib_size);  }\n")
-    endif()
-  endforeach()
-
-  # generating opencl*.inc files
-  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/opencl1.2-c.amdgcn.pch
-    COMMAND clang -cc1 -x cl-header -triple amdgcn-amd-amdhsa -Werror -O3 -DNDEBUG -cl-std=CL1.2 -emit-pch -o ${CMAKE_CURRENT_BINARY_DIR}/opencl1.2-c.amdgcn.pch < ${CMAKE_SOURCE_DIR}/compiler/llvm/tools/clang/lib/Headers/opencl-c.h
-    DEPENDS clang ${CMAKE_SOURCE_DIR}/compiler/llvm/tools/clang/lib/Headers/opencl-c.h
-    COMMENT "Generating opencl1.2-c.amdgcn.pch"
-  )
-  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/opencl1.2-c.amdgcn.inc
-    COMMAND bc2h ${CMAKE_CURRENT_BINARY_DIR}/opencl1.2-c.amdgcn.pch ${CMAKE_CURRENT_BINARY_DIR}/opencl1.2-c.amdgcn.inc opencl1_2_c
-    DEPENDS bc2h ${CMAKE_CURRENT_BINARY_DIR}/opencl1.2-c.amdgcn.pch
-    COMMENT "Generating opencl1.2-c.amdgcn.inc"
-  )
-  set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${CMAKE_CURRENT_BINARY_DIR}/opencl1.2-c.amdgcn.inc ${CMAKE_CURRENT_BINARY_DIR}/opencl1.2-c.amdgcn.pch)
-  add_custom_target(opencl1.2-c.amdgcn.inc_target ALL DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/opencl1.2-c.amdgcn.inc)
-  add_dependencies(oclrocm opencl1.2-c.amdgcn.inc_target)
-
-  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/opencl2.0-c.amdgcn.pch
-    COMMAND clang -cc1 -x cl-header -triple amdgcn-amd-amdhsa -Werror -O3 -DNDEBUG -cl-std=CL2.0 -emit-pch -o ${CMAKE_CURRENT_BINARY_DIR}/opencl2.0-c.amdgcn.pch < ${CMAKE_SOURCE_DIR}/compiler/llvm/tools/clang/lib/Headers/opencl-c.h
-    DEPENDS clang ${CMAKE_SOURCE_DIR}/compiler/llvm/tools/clang/lib/Headers/opencl-c.h
-    COMMENT "Generating opencl2.0-c.amdgcn.pch"
-  )
-  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/opencl2.0-c.amdgcn.inc
-    COMMAND bc2h ${CMAKE_CURRENT_BINARY_DIR}/opencl2.0-c.amdgcn.pch ${CMAKE_CURRENT_BINARY_DIR}/opencl2.0-c.amdgcn.inc opencl2_0_c
-    DEPENDS bc2h ${CMAKE_CURRENT_BINARY_DIR}/opencl2.0-c.amdgcn.pch
-    COMMENT "Generating opencl2.0-c.amdgcn.inc"
-  )
-  set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${CMAKE_CURRENT_BINARY_DIR}/opencl2.0-c.amdgcn.inc ${CMAKE_CURRENT_BINARY_DIR}/opencl2.0-c.amdgcn.pch)
-  add_custom_target(opencl2.0-c.amdgcn.inc_target ALL DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/opencl2.0-c.amdgcn.inc)
-  add_dependencies(oclrocm opencl2.0-c.amdgcn.inc_target)
-endif() # if (${USE_COMGR_LIBRARY} STREQUAL "no")

projects/clr/rocclr/device/rocm/rockernel.cpp

@@ -48,13 +48,7 @@ bool LightningKernel::init() {
 
   hsa_agent_t hsaDevice = program()->hsaDevice();
 
-  const amd_comgr_metadata_node_t* kernelMetaNode =
-              static_cast<const LightningProgram*>(program())->getKernelMetadata(name());
-  if (kernelMetaNode == nullptr) {
-    return false;
-  }
-
-  if (!GetAttrCodePropMetadata(*kernelMetaNode)) {
+  if (!GetAttrCodePropMetadata()) {
     return false;
   }