Add Pascal support to JCUDF transcode (row_conversion) #10329
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
rapidsai#8444 modified JCUDF transcoding logic (in Java/JNI) to use cudaMemcpyAsync() and cuda::barrier to allow for asynchronous memcpy on GPUs that support it. While this works for __CUDA_ARCH__ >= 700, for older GPUs (e.g. Pascal), JCUDF conversions cause CUDA errors and failures. E.g. ``` ai.rapids.cudf.CudfException: after reduction step 2: cudaErrorInvalidDeviceFunction: invalid device function ``` For older GPUs, rather than fail spectacularly, it would be good to provide a more stable (if less efficient) fallback implementation, via `memcpy()`. This commit adds code to conditionally use `cudaMemcpyAsync()` or `memcpy()`, depending on the GPU in play.
|
Note that this is based on @hyperbolic2346's work, over on his branch. This was tested with a replica of the JCUDF logic made in |
Codecov Report
@@ Coverage Diff @@
## branch-22.04 #10329 +/- ##
=============================================
Coverage 10.63% 10.63%
=============================================
Files 122 122
Lines 20940 20940
=============================================
Hits 2228 2228
Misses 18712 18712 Continue to review full report at Codecov.
|
|
🚀 |
|
Ah, shoot. It appears that the compilation introduces warnings to the build. Best not merge this until the warnings are sorted out. |
revans2
approved these changes
Feb 18, 2022
ttnghia
reviewed
Feb 18, 2022
ttnghia
approved these changes
Feb 19, 2022
|
I have managed to run tests, at last, against Pascal hardware: Some test logs: This should be safe to merge now. |
|
@gpucibot merge |
|
This change has been merged now. Thanks for the reviews, chaps. |
|
😍 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
#8444 modified JCUDF transcoding logic (in Java/JNI) to use
cudaMemcpyAsync()and
cuda::barrierto allow for asynchronous memcpy on GPUs that support it.While this works for
__CUDA_ARCH__ >= 700, for older GPUs (e.g. Pascal),JCUDF conversions cause CUDA errors and failures. E.g.
cudaMemcpyAsync()is not supported on Pascal GPUs or prior. (They lack the hardware,apparently.)
For older GPUs, rather than fail spectacularly, it would be good to provide
a more stable (if less efficient) fallback implementation, via
memcpy().This commit adds code to conditionally use
cudaMemcpyAsync()ormemcpy(),depending on the GPU in play.