Add patch for thrust-cub 1.16 to fix sort compile times #10577
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
davidwendt
added
3 - Ready for Review
Codecov Report
@@ Coverage Diff @@
## branch-22.06 #10577 +/- ##
=============================================
Coverage 86.33% 86.33%
=============================================
Files 140 140
Lines 22300 22300
=============================================
Hits 19253 19253
Misses 3047 3047 Continue to review full report at Codecov.
|
bdice
approved these changes
Apr 2, 2022
There was a problem hiding this comment.
This looks like it fixes the build time regressions and the diff-of-diffs looks fine. Thanks @davidwendt! I was mostly focused on the performance of benchmarks and hadn't checked the build time report on #10489.
ttnghia
approved these changes
Apr 4, 2022
vyasr
approved these changes
Apr 4, 2022
|
@gpucibot merge |
bdice
added a commit
to bdice/cudf
that referenced
this pull request
Apr 4, 2022
…dsai#10577)" This reverts commit ff1ff80.
rapids-bot bot
pushed a commit
that referenced
this pull request
Apr 5, 2022
PR #10489 updated from Thrust 1.15 to Thrust 1.16. However, this appears to be causing conflicts with other repositories -- [cuSpatial](rapidsai/cuspatial#511 (comment)) and cuGraph have reported issues where their builds are finding Thrust 1.16 from libcudf instead of Thrust 1.15 which is [currently pinned by rapids-cmake](https://github.com/rapidsai/rapids-cmake/blob/06a657281cdd83781e49afcdbb39abc491eeab17/rapids-cmake/cpm/versions.json#L26). This PR is intended to unblock local builds and CI builds for other RAPIDS packages until we are able to identify the root cause (which may be due to CMake include path orderingsrapids-cmake). Last time Thrust was updated, [rapids-cmake was updated](rapidsai/rapids-cmake#138) one day before [libcudf was updated](#9912). That may explain why we didn't notice this problem with the 1.15 update. The plan I currently have in mind is: 1. Merge this PR to roll back libcudf to Thrust 1.15 (and revert the patch for Thrust 1.16 [10577](#10577)). This will hopefully unblock CI for cugraph and cuspatial. 2. Try to work out whatever issues with CMake / include paths may exist. 3. Prepare all rapids-cmake repos for Thrust 1.16 compatibility. I've [done this for RMM already](rapidsai/rmm#1011), and I am working on [PR 4675](rapidsai/cuml#4675) to cuML now. I am planning to make the same fixes for `#include`s in cuCollections, raft, cuSpatial, and cuGraph so they will be compatible with Thrust 1.16. 4. Try to upgrade libcudf to Thrust 1.16 again (and re-apply the updated patch). If (2) has been resolved, I hope we won't see any issues in other RAPIDS libraries 5. Upgrade rapids-cmake to Thrust 1.16. Authors: - Bradley Dice (https://github.com/bdice) Approvers: - Vyas Ramasubramani (https://github.com/vyasr) - Mark Harris (https://github.com/harrism) URL: #10586
abellina
pushed a commit
to abellina/cudf
that referenced
this pull request
Apr 11, 2022
Fixes `thrust.patch` to patch the CUB source for `sort` to minimize the inlining of the comparator functor. The build was updated in rapidsai#10489 to thrust-1.16 which includes change to thrust sort using CUB's `DeviceMergeSort`. This means the previous patch does not apply to the new thrust/cub source. This dramatically increased the build for `sort.cu` and other related source files as can be seen in this Build Metrics Report from rapidsai#10489: https://gpuci.gpuopenanalytics.com/job/rapidsai/job/gpuci/job/cudf/job/prb/job/cudf-cpu-cuda-build/CUDA=11.5/8633/Build_20Metrics_20Report/ This PR moves the `pragma unroll` changes into the appropriate CUB source files reducing the build time back to the previous levels (or close to it I hope). Authors: - David Wendt (https://github.com/davidwendt) Approvers: - Bradley Dice (https://github.com/bdice) - Nghia Truong (https://github.com/ttnghia) - Vyas Ramasubramani (https://github.com/vyasr) URL: rapidsai#10577
abellina
pushed a commit
to abellina/cudf
that referenced
this pull request
Apr 11, 2022
PR rapidsai#10489 updated from Thrust 1.15 to Thrust 1.16. However, this appears to be causing conflicts with other repositories -- [cuSpatial](rapidsai/cuspatial#511 (comment)) and cuGraph have reported issues where their builds are finding Thrust 1.16 from libcudf instead of Thrust 1.15 which is [currently pinned by rapids-cmake](https://github.com/rapidsai/rapids-cmake/blob/06a657281cdd83781e49afcdbb39abc491eeab17/rapids-cmake/cpm/versions.json#L26). This PR is intended to unblock local builds and CI builds for other RAPIDS packages until we are able to identify the root cause (which may be due to CMake include path orderingsrapids-cmake). Last time Thrust was updated, [rapids-cmake was updated](rapidsai/rapids-cmake#138) one day before [libcudf was updated](rapidsai#9912). That may explain why we didn't notice this problem with the 1.15 update. The plan I currently have in mind is: 1. Merge this PR to roll back libcudf to Thrust 1.15 (and revert the patch for Thrust 1.16 [10577](rapidsai#10577)). This will hopefully unblock CI for cugraph and cuspatial. 2. Try to work out whatever issues with CMake / include paths may exist. 3. Prepare all rapids-cmake repos for Thrust 1.16 compatibility. I've [done this for RMM already](rapidsai/rmm#1011), and I am working on [PR 4675](rapidsai/cuml#4675) to cuML now. I am planning to make the same fixes for `#include`s in cuCollections, raft, cuSpatial, and cuGraph so they will be compatible with Thrust 1.16. 4. Try to upgrade libcudf to Thrust 1.16 again (and re-apply the updated patch). If (2) has been resolved, I hope we won't see any issues in other RAPIDS libraries 5. Upgrade rapids-cmake to Thrust 1.16. Authors: - Bradley Dice (https://github.com/bdice) Approvers: - Vyas Ramasubramani (https://github.com/vyasr) - Mark Harris (https://github.com/harrism) URL: rapidsai#10586
rapids-bot bot
pushed a commit
that referenced
this pull request
Aug 11, 2022
Updates the bundled version of Thrust to 1.17.0. I will run benchmarks and include results in a comment below. Depends on #11457. Supersedes #10489, #10577, #10586. Closes #10841. **This should be merged concurrently with rapidsai/rapids-cmake#231 Authors: - Bradley Dice (https://github.com/bdice) Approvers: - David Wendt (https://github.com/davidwendt) - Nghia Truong (https://github.com/ttnghia) - Robert Maynard (https://github.com/robertmaynard) URL: #11437
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.
Fixes
thrust.patchto patch the CUB source forsortto minimize the inlining of the comparator functor. The build was updated in #10489 to thrust-1.16 which includes change to thrust sort using CUB'sDeviceMergeSort. This means the previous patch does not apply to the new thrust/cub source. This dramatically increased the build forsort.cuand other related source files as can be seen in this Build Metrics Report from #10489: https://gpuci.gpuopenanalytics.com/job/rapidsai/job/gpuci/job/cudf/job/prb/job/cudf-cpu-cuda-build/CUDA=11.5/8633/Build_20Metrics_20Report/This PR moves the
pragma unrollchanges into the appropriate CUB source files reducing the build time back to the previous levels (or close to it I hope).