samples/custom_dispatch/cuda/kernels/CMakeLists.txt

# Copyright 2022 The IREE Authors
#
# Licensed under the Apache License v2.0 with LLVM Exceptions.
# See https://llvm.org/LICENSE.txt for license information.
# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception

if(NOT IREE_TARGET_BACKEND_CUDA OR NOT IREE_HAL_DRIVER_CUDA)
  return()
endif()

# NOTE: this is not how one should actually build their PTX files. Do not use
# this as an authoritative source for compilation settings or CMake goo. If you
# choose to go the route of custom CUDA kernels you must bring your own build
# infrastructure. This sample only demonstrates how to use compiled PTX blobs
# inside of the IREE compiler and this is the minimum amount of hacking that
# could be done to do that.

# Default to using our own clang. The NVCC route is preserved as an example but
# to allow for consistent cross-compiling we default to clang - they should be
# effectively the same for our purposes (device only code to LTO-IR/PTX).
set(_BUILD_WITH_NVCC ON)

if(_BUILD_WITH_NVCC)
  include(CheckLanguage)
  check_language(CUDA)
  if(NOT CMAKE_CUDA_COMPILER)
    message(STATUS "IREE custom_dispatch/cuda/kernels ignored -- nvcc not found")
    return()
  endif()
  enable_language(CUDA)
endif()

# Builds a PTX blob using cmake + nvcc from the CUDA SDK.
function(cuda_kernel_ptx_nvcc _ARCH)
  set(_NAME iree_samples_custom_dispatch_cuda_kernels_ptx_${_ARCH})
  set(_PTX_SRC_NAME "kernels.cu")
  get_filename_component(_PTX_SRC_BASENAME ${_PTX_SRC_NAME} NAME_WE CACHE)
  set(_PTX_OBJ_NAME "${_PTX_SRC_BASENAME}_sm_${_ARCH}")

  add_library(${_NAME}_obj OBJECT)
  target_sources(${_NAME}_obj PRIVATE ${_PTX_SRC_NAME})
  set_source_files_properties(${_PTX_SRC_NAME} PROPERTIES LANGUAGE CUDA)
  set_target_properties(${_NAME}_obj PROPERTIES
    LANGUAGE CUDA
    LINKER_LANGUAGE CUDA
    CUDA_PTX_COMPILATION ON
    CUDA_SEPARABLE_COMPILATION ON
    CUDA_RESOLVE_DEVICE_SYMBOLS ON
    CUDA_ARCHITECTURES "${_ARCH}"
  )

  # This makes my eyes bleed. There is probably a much better way of doing this
  # and I wish the best of luck to those who try. From:
  # https://sourcegraph.com/github.com/NVIDIA/MDL-SDK/-/blob/cmake/utilities.cmake?L1266
  # This sample should probably just invoke nvcc directly.
  get_property(_GENERATOR_IS_MULTI_CONFIG GLOBAL PROPERTY GENERATOR_IS_MULTI_CONFIG)
  if(_GENERATOR_IS_MULTI_CONFIG)
    set(_PTX_CONFIG_FOLDER /$<CONFIG>)
    set(_CMAKEFILES_FOLDER "")
  else()
    set(_PTX_CONFIG_FOLDER "")
    set(_CMAKEFILES_FOLDER /CMakeFiles)
  endif()
  add_custom_command(
    OUTPUT ${_PTX_OBJ_NAME}.ptx
    DEPENDS $<TARGET_OBJECTS:${_NAME}_obj>
    COMMAND ${CMAKE_COMMAND} -E make_directory ${CMAKE_CURRENT_BINARY_DIR}
    COMMAND ${CMAKE_COMMAND} -E copy_if_different
        ${CMAKE_CURRENT_BINARY_DIR}${_CMAKEFILES_FOLDER}/${_NAME}_obj.dir${_PTX_CONFIG_FOLDER}/${_PTX_SRC_BASENAME}.ptx
        ${CMAKE_CURRENT_BINARY_DIR}/${_PTX_OBJ_NAME}.ptx
  )
  add_custom_target(${_NAME} DEPENDS
    ${CMAKE_CURRENT_BINARY_DIR}/${_PTX_OBJ_NAME}.ptx
  )
  add_dependencies(iree-sample-deps "${_NAME}")
endfunction()

# Builds a PTX blob using the clang built by IREE from tip-of-tree LLVM.
function(cuda_kernel_ptx_clang _ARCH)
  set(_NAME iree_samples_custom_dispatch_cuda_kernels_ptx_${_ARCH})
  set(_PTX_SRC_NAME "kernels.cu")
  get_filename_component(_PTX_SRC_BASENAME ${_PTX_SRC_NAME} NAME_WE CACHE)
  set(_PTX_OBJ_NAME "${_PTX_SRC_BASENAME}_sm_${_ARCH}.ptx")
  add_custom_command(
    OUTPUT
      ${_PTX_OBJ_NAME}
    DEPENDS
      ${_PTX_SRC_NAME}
      ${IREE_CLANG_TARGET}
    COMMAND ${IREE_CLANG_BINARY}
      -x cuda
      -Wno-unknown-cuda-version
      --cuda-path=${CUDAToolkit_ROOT}
      --cuda-device-only
      --cuda-gpu-arch=sm_${_ARCH}
      -O2
      -S
      ${CMAKE_CURRENT_SOURCE_DIR}/${_PTX_SRC_NAME}
      -o ${CMAKE_CURRENT_BINARY_DIR}/${_PTX_OBJ_NAME}
    VERBATIM
  )
  add_custom_target(${_NAME} DEPENDS
    ${CMAKE_CURRENT_BINARY_DIR}/${_PTX_OBJ_NAME}
  )
  add_dependencies(iree-sample-deps "${_NAME}")
endfunction()

# Build the kernels_*.ptx files for each architecture we target.
if(_BUILD_WITH_NVCC)
  cuda_kernel_ptx_nvcc(52)
  cuda_kernel_ptx_nvcc(80)
else()
  cuda_kernel_ptx_clang(52)
  cuda_kernel_ptx_clang(80)
endif()

iree_lit_test_suite(
  NAME
    example
  SRCS
    "example.mlir"
  TOOLS
    FileCheck
    iree-compile
    iree-run-module
  LABELS
    "driver=cuda"
    "hostonly"
)

iree_cuda_bitcode_library(
  NAME
    cuda_ukernel
  CUDA_ARCH
    sm_60
  SRCS
    "ukernel.cu"
)

iree_check_single_backend_test_suite(
  NAME
    check_cuda_ukernel
  SRCS
    "ukernel_example.mlir"
  TARGET_BACKEND
    "cuda"
  COMPILER_FLAGS
    "--iree-link-bitcode=cuda_ukernel.bc"
  DRIVER
    "cuda"
  LABELS
    "noasan"
    "nomsan"
    "notsan"
    "noubsan"
    "requires-gpu-nvidia"
  DEPENDS
    ::cuda_ukernel
)