Merge topic 'support_nvcc_native_flag'

14d8a2768d CUDA: Support nvcc 11.5 new -arch=all|all-major flags

Acked-by: Kitware Robot <[email protected]>
Reviewed-by: Raul Tambre <[email protected]>
Merge-request: !6652

Help/prop_tgt/CUDA_ARCHITECTURES.rst

@@ -20,6 +20,18 @@ variable if it is set when a target is created.
 The ``CUDA_ARCHITECTURES`` target property must be set to a non-empty value on targets
 that compile CUDA sources, or it is an error.  See policy :policy:`CMP0104`.
 
+.. versionadded:: 3.23
+
+  The ``CUDA_ARCHITECTURES`` may be set to the following special keywords:
+
+  ``all``
+    Requires NVIDIA 11.5+. Will compile for all supported major and minor real
+    architectures, and the highest major virtual architecture.
+
+  ``all-major``
+    Requires NVIDIA 11.5+. Will compile for all supported major real
+    architectures, and the highest major virtual architecture.
+
 Examples
 ^^^^^^^^
 

Help/release/dev/cuda-new-arch-modes.rst

@@ -0,0 +1,10 @@
+cuda-new-arch-modes
+-------------------
+
+* The :prop_tgt:`CUDA_ARCHITECTURES` target property now supports the
+  `all`, and `all-major` values when the CUDA compiler id is ``NVIDIA``,
+  and version is 11.5+.
+
+* The :variable:`CMAKE_CUDA_ARCHITECTURES` variable now supports the
+  `all`, and `all-major` values when the `CUDA` compiler id is ``NVIDIA``,
+  and version is 11.5+.

Modules/CMakeDetermineCUDACompiler.cmake

@@ -258,13 +258,22 @@ if(NOT CMAKE_CUDA_COMPILER_ID_RUN)
 
   # Append user-specified architectures.
   if(CMAKE_CUDA_ARCHITECTURES)
-    foreach(arch ${CMAKE_CUDA_ARCHITECTURES})
-      # Strip specifiers as PTX vs binary doesn't matter.
-      string(REGEX MATCH "[0-9]+" arch_name "${arch}")
-      string(APPEND clang_test_flags " --cuda-gpu-arch=sm_${arch_name}")
-      string(APPEND nvcc_test_flags " -gencode=arch=compute_${arch_name},code=sm_${arch_name}")
-      list(APPEND tested_architectures "${arch_name}")
-    endforeach()
+    if("x${CMAKE_CUDA_ARCHITECTURES}" STREQUAL "xall")
+      string(APPEND nvcc_test_flags " -arch=all")
+      set(architectures_mode all)
+    elseif("x${CMAKE_CUDA_ARCHITECTURES}" STREQUAL "xall-major")
+      string(APPEND nvcc_test_flags " -arch=all-major")
+      set(architectures_mode all-major)
+    else()
+      set(architectures_mode explicit)
+      foreach(arch ${CMAKE_CUDA_ARCHITECTURES})
+        # Strip specifiers as PTX vs binary doesn't matter.
+        string(REGEX MATCH "[0-9]+" arch_name "${arch}")
+        string(APPEND clang_test_flags " --cuda-gpu-arch=sm_${arch_name}")
+        string(APPEND nvcc_test_flags " -gencode=arch=compute_${arch_name},code=sm_${arch_name}")
+        list(APPEND tested_architectures "${arch_name}")
+      endforeach()
+    endif()
 
     # If the user has specified architectures we'll want to fail during compiler detection if they don't work.
     set(CMAKE_CUDA_COMPILER_ID_REQUIRE_SUCCESS ON)
@@ -597,7 +606,18 @@ if(DEFINED detected_architecture AND "${CMAKE_CUDA_ARCHITECTURES}" STREQUAL "")
   if(NOT CMAKE_CUDA_ARCHITECTURES)
     message(FATAL_ERROR "Failed to find a working CUDA architecture.")
   endif()
-elseif(architectures)
+elseif(architectures AND (architectures_mode STREQUAL "xall" OR
+                          architectures_mode STREQUAL "xall-major"))
+  if(NOT CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
+    message(FATAL_ERROR
+      "The CMAKE_CUDA_ARCHITECTURES:\n"
+      "  ${CMAKE_CUDA_ARCHITECTURES}\n"
+      "is not supported with the ${CMAKE_CUDA_COMPILER_ID} compiler.  Try:\n"
+      "  ${architectures}\n"
+      "instead.")
+  endif()
+
+elseif(architectures AND architectures_mode STREQUAL "xexplicit")
   # Sort since order mustn't matter.
   list(SORT architectures)
   list(SORT tested_architectures)
@@ -630,5 +650,7 @@ unset(_CUDA_LIBRARY_DIR)
 unset(_CUDA_TARGET_DIR)
 unset(_CUDA_TARGET_NAME)
 
+unset(architectures_mode)
+
 set(CMAKE_CUDA_COMPILER_ENV_VAR "CUDACXX")
 set(CMAKE_CUDA_HOST_COMPILER_ENV_VAR "CUDAHOSTCXX")

Source/cmGeneratorTarget.cxx

@@ -3465,6 +3465,22 @@ void cmGeneratorTarget::AddCUDAArchitectureFlags(std::string& flags) const
     return;
   }
 
+  std::string const& compiler =
+    this->Makefile->GetSafeDefinition("CMAKE_CUDA_COMPILER_ID");
+
+  // Check for special modes: `all`, `all-major`.
+  if (property == "all") {
+    if (compiler == "NVIDIA") {
+      flags += " -arch=all";
+      return;
+    }
+  } else if (property == "all-major") {
+    if (compiler == "NVIDIA") {
+      flags += " -arch=all-major";
+      return;
+    }
+  }
+
   struct CudaArchitecture
   {
     std::string name;
@@ -3506,9 +3522,6 @@ void cmGeneratorTarget::AddCUDAArchitectureFlags(std::string& flags) const
     }
   }
 
-  std::string const& compiler =
-    this->Makefile->GetSafeDefinition("CMAKE_CUDA_COMPILER_ID");
-
   if (compiler == "NVIDIA") {
     for (CudaArchitecture& architecture : architectures) {
       flags +=

Tests/CudaOnly/All/CMakeLists.txt

@@ -0,0 +1,44 @@
+cmake_minimum_required(VERSION 3.20)
+project(CudaOnlyAll CUDA)
+
+if(CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA" AND
+   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 11.5.0)
+
+  set(compile_options -Wno-deprecated-gpu-targets)
+  function(verify_output flag output_var)
+    string(REGEX MATCHALL "-arch compute_([0-9]+)" target_archs "${${output_var}}")
+    list(LENGTH target_archs count)
+    if(count LESS 2)
+      message(FATAL_ERROR "${flag} failed to map to multiple architectures")
+    endif()
+  endfunction()
+endif()
+
+if(COMMAND verify_output)
+  set(try_compile_flags -v ${compile_options})
+
+  set(CMAKE_CUDA_ARCHITECTURES all)
+  try_compile(all_archs_compiles
+    ${CMAKE_CURRENT_BINARY_DIR}/try_compile/all_archs_compiles
+    ${CMAKE_CURRENT_SOURCE_DIR}/main.cu
+    COMPILE_DEFINITIONS ${try_compile_flags}
+    OUTPUT_VARIABLE output
+    )
+  verify_output(all output)
+
+  set(CMAKE_CUDA_ARCHITECTURES all-major)
+  try_compile(all_major_archs_compiles
+    ${CMAKE_CURRENT_BINARY_DIR}/try_compile/all_major_archs_compiles
+    ${CMAKE_CURRENT_SOURCE_DIR}/main.cu
+    COMPILE_DEFINITIONS ${try_compile_flags}
+    OUTPUT_VARIABLE output
+    )
+  verify_output(all-major output)
+
+  if(all_archs_compiles AND all_major_archs_compiles)
+    add_executable(CudaOnlyAll main.cu)
+    target_compile_options(CudaOnlyAll PRIVATE ${compile_options})
+  endif()
+else()
+  add_executable(CudaOnlyAll main.cu)
+endif()

Tests/CudaOnly/All/main.cu

@@ -0,0 +1,3 @@
+int main()
+{
+}

Tests/CudaOnly/CMakeLists.txt

@@ -4,8 +4,10 @@ macro (add_cuda_test_macro name)
     PROPERTY LABELS "CUDA")
 endmacro ()
 
+add_cuda_test_macro(CudaOnly.All CudaOnlyAll)
 add_cuda_test_macro(CudaOnly.Architecture Architecture)
 add_cuda_test_macro(CudaOnly.CompileFlags CudaOnlyCompileFlags)
+
 add_cuda_test_macro(CudaOnly.EnableStandard CudaOnlyEnableStandard)
 add_cuda_test_macro(CudaOnly.ExportPTX CudaOnlyExportPTX)
 add_cuda_test_macro(CudaOnly.SharedRuntimePlusToolkit CudaOnlySharedRuntimePlusToolkit)