Fix skipped workspace initialization for parallel split-K in GemmUniversalBase - #3578
Open
AyaanFaisal21 wants to merge 1 commit into
Open
Fix skipped workspace initialization for parallel split-K in GemmUniversalBase#3578AyaanFaisal21 wants to merge 1 commit into
AyaanFaisal21 wants to merge 1 commit into
Conversation
…ersalBase Fixes NVIDIA#3538. GemmUniversalBase::initialize() only calls Params::init_workspace when mode == kGemm, while GemmWithKReduction's init_workspace only acts when mode == kGemmSplitKParallel, so its ptr_D / ptr_gemm_k_reduction redirect can never run: partials go through the user's D buffer (out of bounds for grid.k > 1) and the reduction reads uninitialized workspace. Before 3.2 the call was unconditional (v2.11, v3.1), and the Python cppgen backend still calls it unconditionally. Restore the unconditional call and move the mode condition onto the one side effect it plausibly guarded: the whole-workspace memset in ParamsBase, which only serial split-K needs for semaphore zeroing. Tested with example 23 on SM86 (m=1024 n=1024 k=8192, 8 slices): before, parallel split-k miscompares and compute-sanitizer reports OOB 16-byte global writes past D; after, plain, serial, and parallel all pass and compute-sanitizer reports 0 errors. Serial split-K exercises the re-gated memset path directly. This change was developed with AI assistance; I have reviewed and verified it.
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Problem
Fixes #3538.
Since 3.2, GemmUniversalBase::initialize() only calls Params::init_workspace() for kGemm, but GemmWithKReduction::Params::init_workspace() performs the parallel split-K pointer setup only for kGemmSplitKParallel. That makes the setup unreachable: partials are written through the user's ptr_D, causing OOB writes when grid.k > 1, while the reduction reads workspace that was never written.
The gate only appeared in the 3.2 squash commit; v2.11 and v3.1 called
init_workspace()in every mode. Every kernel that customizesinit_workspace()checks the mode inside its own implementation, so the unconditional call is the contract those implementations were written against.Fix
Restore the unconditional init_workspace() call and move the mode check to ParamsBase's workspace memset. Only serial split-K needs that memset to zero its semaphore; parallel split-K fully overwrites its workspace before reduction.
This restores the required parallel setup while still skipping the pointless whole-buffer memset for parallel split-K, since it fully overwrites its workspace before reduction.
Testing
Example 23 (ampere_gemm_operand_reduction_fusion), SM86, CUDA 12.8, m=1024 n=1024 k=8192:
Before: plain and serial split-K pass; 8-slice parallel split-K miscompares, with compute-sanitizer reporting invalid 16-byte writes past D.
After: all three pass; compute-sanitizer reports 0 errors.
Serial split-K also exercises the re-gated semaphore memset and remains unchanged.
AI assistance was used during investigation and drafting. The proposed logic, code changes, and reported test results were reviewed and verified by the author.