Fix compiler error with cuda-clang #17171
Conversation
segment_reduction_ops.h requires cuda_kernel_helper.h to be included in clang because it uses some of the helpers directly in the header (e.g. CudaAtomicMax). It works with nvcc, because the usage is in a template context and nvcc checks that function is available only at template instantiation. However, clang does more strict erorr-checking for functions found during template instantiation and requires them to be found either by ADL or at the point of template declaration.
| @@ -16,6 +16,14 @@ limitations under the License. | |||
| #ifndef THIRD_PARTY_TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_ | |||
| #define THIRD_PARTY_TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_ | |||
|
|
|||
|
|
|||
| // This file requires the following include because it uses CudaAtomicMax: | |||
| // #include "tensorflow/core/util/cuda_kernel_helper.h" | |||
There was a problem hiding this comment.
What if we guard the include with "ifdef GOOGLE_CUDA" ?
There was a problem hiding this comment.
I tried doing that, but discovered that CPU builds also set GOOGLE_CUDA (see tensorflow.bzl#1 and tensorflow.bzl#2, both seem to propagate to '.cc' files as well as '.cu.cc' files)
Probably tensorflow.bzl shouldn't set GOOGLE_CUDA, as this define is handled by the [GPU crosstool] (
However when I tried removing '-DGOOGLE_CUDA=1' from tensorflow.bzl, but I got a bunch of link errors and didn't investigate any further. Even if we want to dig deeper into this, it would be nice to have a workaround to unbreak cuda_clang while we investigate.
There was a problem hiding this comment.
You are right, that seems like a bug.
If this works as is now, I am OK with merging it.
gunan
added
awaiting testing (then merge)
kokoro:force-run
segment_reduction_ops.h requires cuda_kernel_helper.h to be
included in clang because it uses some of the helpers directly in the
header (e.g. CudaAtomicMax). It works with nvcc, because the usage is
in a template context and nvcc checks that function is available only later
at template instantiation.
However, clang does more strict erorr-checking for functions found
during template instantiation and requires them to also be found either by
ADL or at the point of template declaration.