Automatically set CTS device triples based on which adapters are built.

Also adds the capability to detect the amd arch string we pass for
--offload-arch by finding and using rocm_agent_enumerator.

.github/workflows/cmake.yml

@@ -163,10 +163,10 @@ jobs:
  strategy:
  matrix:
  adapter: [
- {name: CUDA, triplet: nvptx64-nvidia-cuda, platform: ""},
- {name: HIP, triplet: amdgcn-amd-amdhsa, platform: ""},
- {name: L0, triplet: spir64, platform: ""},
- {name: OPENCL, triplet: spir64, platform: "Intel(R) OpenCL"}
+ {name: CUDA, platform: ""},
+ {name: HIP, platform: ""},
+ {name: L0, platform: ""},
+ {name: OPENCL, platform: "Intel(R) OpenCL"}
  ]
  build_type: [Debug, Release]
  compiler: [{c: gcc, cxx: g++}, {c: clang, cxx: clang++}]
@@ -198,7 +198,6 @@ jobs:
  -DUR_BUILD_ADAPTER_${{matrix.adapter.name}}=ON
  -DUR_DPCXX=${{github.workspace}}/dpcpp_compiler/bin/clang++
  -DUR_SYCL_LIBRARY_DIR=${{github.workspace}}/dpcpp_compiler/lib
- -DUR_CONFORMANCE_TARGET_TRIPLES=${{matrix.adapter.triplet}}
  ${{ matrix.adapter.name == 'HIP' && '-DAMD_ARCH=gfx1030' || '' }} 
  ${{ matrix.adapter.name == 'HIP' && '-DUR_HIP_PLATFORM=AMD' || '' }}
 

CMakeLists.txt

@@ -47,6 +47,9 @@ option(VAL_USE_LIBBACKTRACE_BACKTRACE "enable libbacktrace validation backtrace
 set(UR_DPCXX "" CACHE FILEPATH "Path of the DPC++ compiler executable")
 set(UR_SYCL_LIBRARY_DIR "" CACHE PATH
  "Path of the SYCL runtime library directory")
+set(UR_CONFORMANCE_TARGET_TRIPLES "" CACHE STRING
+ "List of sycl targets to build CTS device binaries for")
+set(UR_CONFORMANCE_AMD_ARCH "AMD device target ID to build CTS binaries for" CACHE STRING)
 
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
 set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)

README.md

@@ -131,6 +131,7 @@ List of options provided by CMake:
 | UR_USE_MSAN | Enable MemorySanitizer (clang only) | ON/OFF | OFF |
 | UR_ENABLE_TRACING | Enable XPTI-based tracing layer | ON/OFF | OFF |
 | UR_CONFORMANCE_TARGET_TRIPLES | SYCL triples to build CTS device binaries for | Comma-separated list | spir64 |
+| UR_CONFORMANCE_AMD_ARCH | AMD device target ID to build CTS binaries for | string | `""` |
 | UR_BUILD_ADAPTER_L0 | Build the Level-Zero adapter | ON/OFF | OFF |
 | UR_BUILD_ADAPTER_OPENCL | Build the OpenCL adapter | ON/OFF | OFF |
 | UR_BUILD_ADAPTER_CUDA | Build the CUDA adapter | ON/OFF | OFF |

cmake/FindRocmAgentEnumerator.cmake

@@ -0,0 +1,19 @@
+# Copyright (C) 2023 Intel Corporation
+# Part of the Unified-Runtime Project, under the Apache License v2.0 with LLVM Exceptions.
+# See LICENSE.TXT
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#
+# FindRocmAgentEnumerator.cmake -- module searching for rocm_agent_enumerator.
+# ROCM_AGENT_ENUMERATOR_FOUND is set to true if
+# rocm_agent_enumerator is found.
+#
+
+find_program(ROCM_AGENT_ENUMERATOR NAMES not_rocm_agent_enumerator)
+
+if(ROCM_AGENT_ENUMERATOR)
+ set(ROCM_AGENT_ENUMERATOR_FOUND TRUE)
+endif()
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(RocmAgentEnumerator DEFAULT_MSG ROCM_AGENT_ENUMERATOR)

test/conformance/CMakeLists.txt

@@ -103,10 +103,15 @@ if(UR_DPCXX)
  if(NOT "${UR_CONFORMANCE_TARGET_TRIPLES}" STREQUAL "")
  string(REPLACE "," ";" TARGET_TRIPLES ${UR_CONFORMANCE_TARGET_TRIPLES})
  else()
- message(WARNING
- "UR_CONFORMANCE_TARGET_TRIPLES wasn't set, defaulting to only \
- generate spir64 device binaries")
- list(APPEND TARGET_TRIPLES "spir64")
+ if(UR_BUILD_ADAPTER_L0 OR UR_BUILD_ADAPTER_OPENCL OR UR_BUILD_ADAPTER_ALL)
+ list(APPEND TARGET_TRIPLES "spir64")
+ endif()
+ if(UR_BUILD_ADAPTER_CUDA OR UR_BUILD_ADAPTER_ALL)
+ list(APPEND TARGET_TRIPLES "nvptx64-nvidia-cuda")
+ endif()
+ if(UR_BUILD_ADAPTER_HIP OR UR_BUILD_ADAPTER_ALL)
+ list(APPEND TARGET_TRIPLES "amdgcn-amd-amdhsa")
+ endif()
  endif()
 
  add_subdirectory(device_code)

test/conformance/device_code/CMakeLists.txt

@@ -3,6 +3,37 @@
 # See LICENSE.TXT
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
+set(AMD_ARCH "${UR_CONFORMANCE_AMD_ARCH}")
+
+if("${AMD_ARCH}" STREQUAL "" AND "${TARGET_TRIPLES}" MATCHES "amd")
+ find_package(RocmAgentEnumerator)
+ if(NOT ROCM_AGENT_ENUMERATOR_FOUND)
+ message(FATAL_ERROR
+ "rocm_agent_enumerator could not be found so detecting target "
+ "HIP device has failed. Set target device with UR_CONFORMANCE_AMD_ARCH "
+ "or ensure rocm_agent_enumerator is available on the PATH.")
+ endif()
+ execute_process(COMMAND ${ROCM_AGENT_ENUMERATOR} OUTPUT_VARIABLE ROCM_AGENTS)
+ string(REGEX MATCHALL "[A-Za-z0-9]+" ROCM_AGENT_LIST "${ROCM_AGENTS}")
+ # rocm_agent_enumerator will return gfx000 to represent non-amd-gpu devices,
+ # we should skip over these
+ list(GET ROCM_AGENT_LIST 0 FIRST_ROCM_AGENT)
+ if("${FIRST_ROCM_AGENT}" STREQUAL "gfx000")
+ list(POP_FRONT ROCM_AGENT_LIST)
+ endif()
+ list(LENGTH ROCM_AGENT_LIST NUM_ROCM_AGENTS)
+ if(${NUM_ROCM_AGENTS} EQUAL 0)
+ message(FATAL_ERROR
+ "No target HIP devices detected with rocm_agent_enumerator, "
+ "specify a target with the UR_CONFORMANCE_AMD_ARCH variable.")
+ elseif(${NUM_ROCM_AGENTS} GREATER 1)
+ message(FATAL_ERROR
+ "Multiple possible target HIP devices found: ${ROCM_AGENT_LIST} "
+ "please specify which target to use by setting UR_CONFORMANCE_AMD_ARCH.")
+ endif()
+ list(GET ROCM_AGENT_LIST 0 AMD_ARCH)
+endif()
+
 macro(add_device_binary SOURCE_FILE)
  get_filename_component(KERNEL_NAME ${SOURCE_FILE} NAME_WE)
  set(DEVICE_BINARY_DIR "${UR_CONFORMANCE_DEVICE_BINARIES_DIR}/${KERNEL_NAME}")