Make libcudacxx respect CMake options for CUDA archs. (#235)

* Allow compute_archs to be parsed with several sane delimiters * Remove compute_future stuff, just parse CMake options
NVIDIA · Jul 17, 2023 · ca0f635 · ca0f635
1 parent c44b5b7
commit ca0f635
Show file tree

Hide file tree

Showing 3 changed files with 12 additions and 15 deletions.
diff --git a/libcudacxx/.upstream-tests/test/CMakeLists.txt b/libcudacxx/.upstream-tests/test/CMakeLists.txt
@@ -13,11 +13,6 @@ foreach (COMPUTE_ARCH ${LIBCUDACXX_COMPUTE_ARCHS})
   set(LIBCUDACXX_COMPUTE_ARCHS_STRING "${LIBCUDACXX_COMPUTE_ARCHS_STRING} ${COMPUTE_ARCH}")
 endforeach ()
 
-option(LIBCUDACXX_ENABLE_COMPUTE_FUTURE "Enable code generation for tests for compute_${LIBCUDACXX_HIGHEST_COMPUTE_ARCH}" OFF)
-if (LIBCUDACXX_ENABLE_COMPUTE_FUTURE)
-  set(_compute_message "${_compute_message} compute_${LIBCUDACXX_HIGHEST_COMPUTE_ARCH}")
-endif ()
-
 message(STATUS "Enabled CUDA architectures:${_compute_message}")
 
 option(LIBCUDACXX_TEST_WITH_NVRTC
@@ -67,6 +62,9 @@ if (${CMAKE_CUDA_COMPILER_ID} STREQUAL "NVHPC")
     -stdpar")
 endif()
 
+set(LIBCUDACXX_COMPUTE_ARCHS_STRING
+  "${CMAKE_CUDA_ARCHITECTURES}")
+
 include(AddLLVM)
 
 set(LIBCUDACXX_TARGET_INFO "libcudacxx.test.target_info.LocalTI" CACHE STRING
@@ -83,8 +81,6 @@ endif()
 
 set(AUTO_GEN_COMMENT "## Autogenerated by libcudacxx configuration.\n# Do not edit!")
 
-pythonize_bool(LIBCUDACXX_ENABLE_COMPUTE_FUTURE)
-
 configure_lit_site_cfg(
   ${CMAKE_CURRENT_SOURCE_DIR}/lit.site.cfg.in
   ${CMAKE_CURRENT_BINARY_DIR}/lit.site.cfg)

diff --git a/libcudacxx/.upstream-tests/test/lit.site.cfg.in b/libcudacxx/.upstream-tests/test/lit.site.cfg.in
@@ -23,7 +23,6 @@ config.test_linker_flags        = "@LIBCUDACXX_TEST_LINKER_FLAGS@"
 config.test_compiler_flags      = "@LIBCUDACXX_TEST_COMPILER_FLAGS@"
 
 config.compute_archs            = "@LIBCUDACXX_COMPUTE_ARCHS_STRING@"
-config.enable_compute_future    = @LIBCUDACXX_ENABLE_COMPUTE_FUTURE@
 config.nvcc_host_compiler       = "@CMAKE_CUDA_HOST_COMPILER@"
 
 config.executor                 = "@LIBCUDACXX_EXECUTOR@"

diff --git a/libcudacxx/.upstream-tests/utils/libcudacxx/test/config.py b/libcudacxx/.upstream-tests/utils/libcudacxx/test/config.py
@@ -588,18 +588,20 @@ def configure_compile_flags(self):
             pre_sm_70 = False
             pre_sm_80 = False
             pre_sm_90 = False
-            compute_archs = [int(a) for a in sorted(shlex.split(compute_archs))]
-            for arch in compute_archs:
+            compute_archs = set(sorted(re.split('\s|;|,', compute_archs)))
+            for s in compute_archs:
+                # Split arch and mode i.e. 80-virtual -> 80, virtual
+                arch, *mode = re.split('-', s)
+                arch = int(arch)
                 if arch < 32: pre_sm_32 = True
                 if arch < 60: pre_sm_60 = True
                 if arch < 70: pre_sm_70 = True
                 if arch < 80: pre_sm_80 = True
                 if arch < 90: pre_sm_90 = True
-                arch_flag = '-gencode=arch=compute_{0},code=sm_{0}'.format(arch)
-                self.cxx.compile_flags += [arch_flag]
-            enable_compute_future = self.get_lit_conf('enable_compute_future')
-            if enable_compute_future:
-                arch_flag = '-gencode=arch=compute_{0},code=compute_{0}'.format(arch)
+                if mode.count("virtual"):
+                    arch_flag = '-gencode=arch=compute_{0},code=compute_{0}'.format(arch)
+                else:
+                    arch_flag = '-gencode=arch=compute_{0},code=sm_{0}'.format(arch)
                 self.cxx.compile_flags += [arch_flag]
         if pre_sm_32:
             self.config.available_features.add("pre-sm-32")