Migrate internal changes

This synchronization round requires squashing a number of commits to
avoid conflicts. Relevant commits messages are as follows:

[docs] NFC: Add documentation for the compiler C++/C/Python APIs
[cmake] NFC: Add CMP0177 policy
[compiler] NFC: Make the extension API uniform across task types
[compiler] Improve the compiler extension API
[compiler] NFC: Fixup TensorRTToExecutable task
[compiler] NFC: reduce the number of overloads in CompilerClient
[compiler] Add mlir-tensorrt-compiler tool
[compiler] Use a Pass option for configuring default backends in StablehloToExecutable
[compiler] Eliminate some unnecessary complexity from compilation task management
[cmake] Further improvements to the CMake build system
[executor] Account for platform minimum alignment in allocation functions
NFC: Add '--use-pep517' to 'pip install' commands
[tensorrt] NFC: remove use of redundant CMake option
[compiler] NFC: more improvements to the CMake build system
[Integration] add nvfp4 Q-DQ into CI tests
[tensorrt] Update support types for ops that expand to `tensorrt.shuffle`
[plan] Fix a bug in `plan-remove-equivalent-buffer-results` pass.
[compiler] NFC: remove unused include in CUDA dialect implementation
[tensorrt] NFC: Fix unguarded CMake inclusion of TRT plugin test passes
[tensorrt] NFC: fix unguarded inclusion in TRT plugin shape inference pass
NFC: remove useless DEPENDS target from the 'mlir-tensorrt-opt' tool
Fix uses of 'long' instead of 'int64_t'
[cmake] More CMake fixes for fixing the TensorRT dependency
NFC: various cleanup/improvements for the CMake build system
[tensorrt] Add FP4 support
[common] Update C++ TensorRT version check utilities to avoid assuming TRT headers are available
[compiler] Fix phase ordering for CUDAToExecutor conversion
[compiler] NFC: Separate Plan attributes and types into separate files
[compiler] NFC: fix typo in naming of helper function
[tensorrt] Update type support for `tensorrt.shuffle` op
[executor] Drop "unpacked memref calling convention"
[executor] Drop opaque ref type from executor
[executor]: Fix static_assert in a template.
[Compiler] Add donation argument support in `compiler`
[compiler] Remove unnecessary canonicalization pass
[executor] NFC: Separate Executor Attributes into a dedicated translation unit
[compiler] Separate the attribute interface for bounds into a common library
NFC: Rename '*ClusterKindAttr' to '*BackendAttr'
NFC: Rename backend-configuration related attribute names in the Plan dialect
NFC: Rename 'PlanClusterKindAttrInterface' to 'CompilerBackendAttrInterface'
NFC: Deduplicate "region closure" utility
NFC: update copybara configurations and add some additional documentation notes
[executor][LuaAllocation]: Use llvm::stable_sort to preserve original order.
[executor][TranslateToLua]: Add Lua variable allocation.
[executor] Enable all Sol2 safety checks
[compiler] Fix missing changes from recent LIT config change
[compiler] Fix emitc integration test configuration issues
[executor] Fix CMake build under minimal configuration
[executor] Eliminate use of CUDA fp16, bf16, fp8 types/headers
[executor] Make the Executor sub-project independently buildable
[cmake] NFC: Fixup dependency provider CMake code
[cmake] NFC: Improve CMake module code organization
[cmake] Better validation of CMake options related to CUDA requirement
NFC: Move TensorRTVersion.h utils header to common project
Update lit.cfg.py to add SM features

GitOrigin-RevId: fb878f289480850a0fef26762422f32ae313e134

Author: christopherbate

mlir-tensorrt/CMakeLists.txt

@@ -1,10 +1,17 @@
 cmake_minimum_required(VERSION 3.25)
 project(mlir-tensorrt LANGUAGES CXX C)
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_LIST_DIR}/build_tools/cmake")
 
-# -------------------------------------------------
-# Set CMake behavior policies
-# -------------------------------------------------
-include(build_tools/cmake/CMakePolicy.cmake NO_POLICY_SCOPE)
+include(MTRTCMakePolicy NO_POLICY_SCOPE)
+
+include(CMakeDependentOption)
+include(CMakePrintHelpers)
+include(MTRTCPM)
+include(MTRTDependencies)
+include(MTRTFeatures)
+include(AddMLIRTensorRT)
+include(Version.cmake)
+include(CMakeOptions.cmake)
 
 # -------------------------------------------------
 # Package cache configuration
@@ -15,46 +22,6 @@ if(NOT CPM_SOURCE_CACHE)
   your build directory (for example '$PWD/.cache.cpm')")
 endif()
 
-# -------------------------------------------------
-# Include utils/tools
-# -------------------------------------------------
-include(build_tools/cmake/CPM.cmake)
-include(build_tools/cmake/CompilationOptions.cmake)
-include(build_tools/cmake/Targets.cmake)
-include(build_tools/cmake/Dependencies.cmake)
-include(Version.cmake)
-
-# -------------------------------------------------
-# Project Options / Configuration
-# -------------------------------------------------
-set(MLIR_TRT_FEATURE_FLAGS)
-macro(mtrt_option name)
-  option(${name} ${ARGN})
-  list(APPEND MLIR_TRT_FEATURE_FLAGS ${name})
-endmacro()
-
-mtrt_option(MLIR_TRT_ENABLE_HLO "Whether to include stablehlo features" ON)
-mtrt_option(MLIR_TRT_ENABLE_ASSERTIONS "Enables assertions" ON)
-mtrt_option(MLIR_TRT_TARGET_TENSORRT "Enable exporting TensorRT dialect IR to a TensorRT engine" ON)
-mtrt_option(MLIR_TRT_TARGET_CPP "Enable exporting TensorRT dialect IR to C++ code that calls the TensorRT API" ON)
-mtrt_option(MLIR_TRT_ENABLE_PYTHON "Enable building the mlir_tensorrt_compiler python package." ON)
-mtrt_option(MLIR_TRT_ENABLE_NVTX "Enable NVTX tracing" ON)
-mtrt_option(MLIR_TRT_ENABLE_TESTING "Enable building optional tests" ON)
-mtrt_option(MLIR_TRT_TARGET_LUA "Enable translating MLIR to the Lua target" ON)
-mtrt_option(MLIR_TRT_ENABLE_EXECUTOR "Build the Executor dialect and MLIR-TensorRT Execution Engine" ON)
-mtrt_option(MLIR_TRT_ENABLE_NCCL "Enable the NCCL runtime module" ON)
-mtrt_option(MLIR_TRT_ENABLE_TORCH "Whether to include torch-mlir features" OFF)
-
-set(MLIR_TRT_TENSORRT_DIR "" CACHE STRING "Path to TensorRT install directory")
-set(MLIR_TRT_DOWNLOAD_TENSORRT_VERSION "10.9" CACHE STRING
-   "Version of TensorRT to download and use. It overrides MLIR_TRT_TENSORRT_DIR.")
-set(MLIR_TRT_PACKAGE_CACHE_DIR "" CACHE STRING "Directory where to cache downloaded C++ packages")
-set(MLIR_TRT_USE_LINKER "" CACHE STRING "Specify a linker to use (e.g. LLD); this is just an alias for LLVM_USE_LINKER")
-
-# Configure dependent Executor options.
-set(MLIR_EXECUTOR_ENABLE_NCCL ${MLIR_TRT_ENABLE_NCCL} CACHE BOOL "" FORCE)
-set(MLIR_EXECUTOR_ENABLE_MPI ${MLIR_TRT_ENABLE_NCCL} CACHE BOOL "" FORCE)
-set(MLIR_EXECUTOR_ENABLE_CUBLAS ${MLIR_TRT_ENABLE_CUBLAS} CACHE BOOL "" FORCE)
 
 # List of dialects to embed in the Python package. This is different between internal
 # and public versions of the code.
@@ -63,6 +30,25 @@ set(MLIR_TRT_PYTHON_UPSTREAM_DIALECTS_EMBED
   tensor
   )
 
+# List of dialects to embed in the Python package. This is different between internal
+# and public versions of the code.
+set(MLIR_TRT_PYTHON_UPSTREAM_DIALECTS_EMBED
+  affine arith bufferization builtin complex index func gpu math memref quant scf
+  tensor linalg
+  )
+
+# Specifies whether we require CUDA at build time.
+set(MLIR_TRT_REQUIRES_CUDA FALSE)
+if(MLIR_TRT_TARGET_TENSORRT OR
+   MLIR_TRT_ENABLE_CUBLAS OR
+   MLIR_TRT_ENABLE_NCCL)
+  set(MLIR_TRT_REQUIRES_CUDA TRUE)
+endif()
+
+if(MLIR_TRT_REQUIRES_CUDA AND NOT MLIR_TRT_ENABLE_CUDA)
+  message(FATAL_ERROR "CUDA is required but MLIR_TRT_ENABLE_CUDA is disabled")
+endif()
+
 # -------------------------------------------------
 # Option validation
 # -------------------------------------------------
@@ -85,48 +71,27 @@ endif()
 # project.
 # Enables generation of compile_commands.json for clangd.
 if(PROJECT_IS_TOP_LEVEL)
-  include(build_tools/cmake/CompilerCache.cmake)
+  include(MTRTCompilerCache)
   mtrt_use_compiler_cache()
   set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+  set(CMAKE_BUILD_WITH_INSTALL_NAME_DIR ON)
 endif()
 
-# -------------------------------------------------
-# Option validation
-# -------------------------------------------------
-# Write out a header file containing convenience macros for each flag.
-function(mtrt_write_feature_flags_header)
-  set(feature_flags_header
-    "${CMAKE_CURRENT_BINARY_DIR}/include/mlir-tensorrt/Features.h")
-
-  # Generate the header at configure time
-  file(WRITE "${feature_flags_header}" [[
-  // Auto-generated feature macros, do not edit.
-  #ifndef MLIR_TENSORRT_FEATURES_H
-  #define MLIR_TENSORRT_FEATURES_H
-
-  ]])
-
-  foreach(FEATURE IN LISTS MLIR_TRT_FEATURE_FLAGS)
-      file(APPEND "${feature_flags_header}" "#ifdef ${FEATURE}\n")
-      file(APPEND "${feature_flags_header}" "#define IF_${FEATURE}(code) do { code } while (0)\n")
-      file(APPEND "${feature_flags_header}" "#else\n")
-      file(APPEND "${feature_flags_header}" "#define IF_${FEATURE}(code) do {} while (0)\n")
-      file(APPEND "${feature_flags_header}" "#endif // ${FEATURE}\n\n")
-  endforeach()
-  file(APPEND "${feature_flags_header}" "#endif // MLIR_TENSORRT_FEATURES_H\n")
-endfunction()
-
-mtrt_write_feature_flags_header()
-include_directories("${CMAKE_CURRENT_BINARY_DIR}/include")
+# The CUDA headers will trigger excessive "C++20 extensions warning" with
+# clang19+ in macro usage.
+include(CheckCXXCompilerFlag)
+check_cxx_compiler_flag("-Wc++20-extensions"
+  CXX_SUPPORTS_C20_EXTENSIONS_FLAG)
+if(CXX_SUPPORTS_C20_EXTENSIONS_FLAG)
+  add_compile_options(-Wno-c++20-extensions)
+endif()
 
 # -------------------------------------------------
 # Setup LLVM/MLIR
 # -------------------------------------------------
 find_package(Python3 REQUIRED
   COMPONENTS Interpreter Development)
 
-set(CMAKE_BUILD_WITH_INSTALL_NAME_DIR ON)
-
 if(PROJECT_IS_TOP_LEVEL)
   set(MLIR_TRT_EXTERNAL_PROJECT_BUILD OFF)
   message("Building MLIR-TensorRT as a standalone project.")
@@ -168,12 +133,14 @@ include(AddLLVM)
 include(AddMLIR)
 include(HandleLLVMOptions)
 
-find_package(CUDAToolkit)
-
 # -------------------------------------------------
 # Dependencies
 # -------------------------------------------------
 
+if(MLIR_TRT_REQUIRES_CUDA)
+  find_package(CUDAToolkit REQUIRED)
+endif()
+
 # Download Stablehlo if it isn't provided by a parent project.
 if(MLIR_TRT_ENABLE_HLO)
   find_package(Stablehlo REQUIRED)
@@ -192,8 +159,6 @@ if(MLIR_TRT_ENABLE_PYTHON)
   mlir_tensorrt_find_dlpack()
 endif()
 
-find_package(MLIRTensorRTCommon REQUIRED)
-
 #--------------------------------------------------
 # Diagnostics
 #--------------------------------------------------
@@ -222,6 +187,9 @@ include_directories(${MLIR_INCLUDE_DIRS})
 link_directories(${LLVM_BUILD_LIBRARY_DIR})
 add_definitions(${LLVM_DEFINITIONS} ${MLIR_TRT_DEFINITIONS})
 
+mtrt_write_feature_flags_header()
+include_directories("${CMAKE_CURRENT_BINARY_DIR}/include")
+
 add_compile_options(-Wno-deprecated-declarations)
 
 include(AddMLIRPython)
@@ -244,13 +212,7 @@ add_custom_target("mlir-tensorrt-doc")
 add_custom_target(check-all-mlir-tensorrt)
 add_custom_target(check-all-mlir-tensorrt-build-only)
 
-if(MLIR_TRT_ENABLE_EXECUTOR)
-  add_subdirectory(executor EXCLUDE_FROM_ALL)
-endif()
-
-add_subdirectory(tensorrt EXCLUDE_FROM_ALL)
-include_directories(${CMAKE_CURRENT_LIST_DIR}/tensorrt/include)
-include_directories(${CMAKE_CURRENT_BINARY_DIR}/tensorrt/include)
-
-add_subdirectory(compiler)
-add_subdirectory(integrations)
+add_subdirectory(common)
+foreach(project_name IN LISTS MLIR_TRT_ENABLE_PROJECTS)
+  add_subdirectory(${project_name})
+endforeach()

mlir-tensorrt/CMakeOptions.cmake

@@ -0,0 +1,40 @@
+mtrt_option(MLIR_TRT_ENABLE_HLO "Whether to include stablehlo features" ON)
+mtrt_option(MLIR_TRT_ENABLE_ASSERTIONS "Enables assertions" ON)
+mtrt_option(MLIR_TRT_TARGET_TENSORRT "Enable exporting TensorRT dialect IR to a TensorRT engine" ON)
+mtrt_option(MLIR_TRT_WITH_ASAN "Enable AddressSanitizer use" OFF)
+mtrt_option(MLIR_TRT_ENABLE_PYTHON "Enable building the mlir_tensorrt_compiler python package." ON)
+mtrt_option(MLIR_TRT_ENABLE_TESTING "Enable building optional tests" ON)
+mtrt_option(MLIR_TRT_ENABLE_TORCH "Whether to include torch-mlir features" OFF)
+mtrt_option(MLIR_TRT_ENABLE_NCCL "Enable the NCCL runtime module" OFF)
+mtrt_option(MLIR_TRT_ENABLE_CUBLAS "Enable CUBLAS in the executor" ON)
+mtrt_option(MLIR_TRT_ENABLE_CUDA "Enable the use of CUDA runtime" ON)
+mtrt_option(MLIR_TRT_ENABLE_MPI "Enable use of MPI in the runtime" ${MLIR_TRT_ENABLE_NCCL})
+
+# The 'python' directory must come last because it actually instantiates
+# the compiler and runtime python packages in '<build-dir>/python_packages'
+# whereas the definition of the source groups belonging to those packages
+# are distributed across multiple sub-directories.
+set(MLIR_TRT_ENABLE_PROJECTS_DEFAULT
+  executor tensorrt compiler integrations/python
+)
+
+if(NOT MLIR_TRT_ENABLE_PYTHON)
+  list(REMOVE_ITEM MLIR_TRT_ENABLE_PROJECTS_DEFAULT "integrations/python")
+endif()
+
+set(MLIR_TRT_ENABLE_PROJECTS ${MLIR_TRT_ENABLE_PROJECTS_DEFAULT} CACHE INTERNAL "Projects to enable")
+
+set(MLIR_TRT_TENSORRT_DIR "" CACHE STRING "Path to TensorRT install directory")
+set(MLIR_TRT_DOWNLOAD_TENSORRT_VERSION "10.12" CACHE STRING
+   "Version of TensorRT to download and use. It overrides MLIR_TRT_TENSORRT_DIR.")
+set(MLIR_TRT_PACKAGE_CACHE_DIR "" CACHE STRING "Directory where to cache downloaded C++ packages")
+# We need this option since CMAKE_CUDA_ARCHITECTURES=native has some issues.
+set(MLIR_TRT_CUDA_ARCHITECTURES "detect" CACHE STRING
+  "Equivalent to CMAKE_CUDA_ARCHITECTURES, but use \"detect\" to "
+  "specify that the compute capability should be taken by querying the driver.")
+
+if(MLIR_TRT_ENABLE_CUDA)
+  set(MLIR_TRT_CUDA_TARGET "CUDA::cudart" CACHE INTERNAL "")
+else()
+  set(MLIR_TRT_CUDA_TARGET "" CACHE INTERNAL "")
+endif()