[REVIEW] Build only compute for the newest arch in CMAKE_CUDA_ARCHITECTURES
#706
Conversation
kkraus14
added
improvement
robertmaynard
force-pushed
the
build_only_compute_for_newest_arch
branch
from
84a6a7b to
468a8e5
Compare
|
@harrism any chance you could review before we merge? I'm not super familiar with nvcc gencode options 😄 |
| # newest arch only to build that way while the rest built only for sm. | ||
| list(SORT CMAKE_CUDA_ARCHITECTURES ORDER ASCENDING) | ||
| list(POP_BACK CMAKE_CUDA_ARCHITECTURES latest_arch) | ||
| list(TRANSFORM CMAKE_CUDA_ARCHITECTURES APPEND "-real") |
There was a problem hiding this comment.
Is -real a cmake command? What does this do?
There was a problem hiding this comment.
There was a problem hiding this comment.
-real and -virtual are special keywords that can be used with CMAKE_CUDA_ARCHITECTURES to provide abstractions around different CUDA compilers code generation API.
For nvcc:
| input | compiler invocation |
|---|---|
| 80 | --generate-code=arch=compute_80,code=[sm_80,compute_80] |
| 80-virtual | --generate-code=arch=compute_80,code=compute_80 |
| 80-real | --generate-code=arch=compute_80,code=sm_80 |
There was a problem hiding this comment.
Can I see the output of this command when I see above now.CMAKE_CUDA_ARCHITECTURES is unset?
We want SASS for all architectures we support, right? If we only include SASS ("-real"/) for 80, then users with anything but Ampere GPUs will experience looooong load/import times due to PTX-JIT to their present architecture. We do need to include PTX, but only for those who have GPUs we don't officially support (e.g. forward compatibility).
There was a problem hiding this comment.
I see, I had it backwards. The -real is appended to all but the last entry. I thought it was only being appended to the last entry. All good.
There was a problem hiding this comment.
We want SASS for all architectures we support, right? If we only include SASS ("-real"/) for 80, then users with anything but Ampere GPUs will experience looooong load/import times
You are correct. The code above is sneaky, as what we do is remove the 'newest' and only apply -real to any existing values. So input 70,80 becomes 70-real, 80 and input 80 becomes 80
harrism
added
the
5 - Ready to Merge
|
@gpucibot merge |
Based on #706 as they both modify `cmake/Modules/SetGPUArchs.cmake` This brings rmm's handling of `CMAKE_CUDA_ARCHITECTURES` to match that is proposed for cudf in rapidsai/cudf#7391 Authors: - Robert Maynard (@robertmaynard) Approvers: - Mark Harris (@harrism) - Keith Kraus (@kkraus14) URL: #709
RMM doesn't need to build
smandcomputefor each architecture, but only for the newest arch.