Fix default dbg and fastbuild optimization level for nvcc

bazel-contrib · Jul 10, 2023 · 0872806 · 0872806
1 parent b9cf8ec
commit 0872806
Show file tree

Hide file tree

Showing 2 changed files with 16 additions and 3 deletions.
diff --git a/cuda/private/toolchain_configs/nvcc.bzl b/cuda/private/toolchain_configs/nvcc.bzl
@@ -298,7 +298,7 @@ def _impl(ctx):
         flag_sets = [
             flag_set(
                 actions = [ACTION_NAMES.cuda_compile],
-                flag_groups = [flag_group(flags = ["-g"])],
+                flag_groups = [flag_group(flags = ["-O0", "-g"])],
             ),
         ],
         provides = ["compilation_mode"],
@@ -331,7 +331,7 @@ def _impl(ctx):
         flag_sets = [
             flag_set(
                 actions = [ACTION_NAMES.cuda_compile],
-                flag_groups = [flag_group(flags = ["--generate-line-info", "-Xcompiler", "-g1"])],
+                flag_groups = [flag_group(flags = ["-O0", "--generate-line-info", "-Xcompiler", "-g1"])],
             ),
         ],
         provides = ["compilation_mode"],

diff --git a/tests/flag/BUILD.bazel b/tests/flag/BUILD.bazel
@@ -36,10 +36,17 @@ cuda_library_flag_test(
 )
 
 # https://docs.bazel.build/versions/main/user-manual.html#flag--compilation_mode
+#
+# NOTE: -O specify optimization level for **host** code only, see
+# https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#optimize-level-o
 cuda_library_c_dbg_flag_test(
     name = "cuda_library_c_dbg_flag_test",
     action_mnemonic = "CudaCompile",
-    contain_flags = ["-g"],
+    contain_flags = [
+        # gcc default to -O0 so bazel does not explicitly set the flag. Whereas nvcc defaults to -O3.
+        "-O0",
+        "-g",
+    ],
     target_compatible_with = ["@platforms//os:linux"],
     target_under_test = "@rules_cuda_examples//basic:kernel",
 )
@@ -48,8 +55,14 @@ cuda_library_c_fastbuild_flag_test(
     name = "cuda_library_c_fastbuild_flag_test",
     action_mnemonic = "CudaCompile",
     contain_flags = [
+        # gcc default to -O0 so bazel does not explicitly set the flag. Whereas nvcc defaults to -O3.
+        "-O0",
         "--generate-line-info",
         "-g1",
+        # There is no -gmlt option (not merged) as claimed in bazel user manual
+        # https://codereview.appspot.com/4440072
+        # https://gcc.gnu.org/legacy-ml/gcc-patches/2011-04/msg02075.html
+        # "-Xcompiler -gmlt",
     ],
     not_contain_flags = ["-DNDEBUG"],
     target_compatible_with = ["@platforms//os:linux"],