Optimize Indexing Backward Kernel with Sub-group Aggregation and Tailored Stride Dispatch #2749
Conversation
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
yucai-intel
changed the title
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Outdated suggestions cannot be applied.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.
Description
This PR optimizes the indexing_backward kernel on XPU by implementing a specialized aggregation strategy for sorted indices. The primary goal is to minimize global memory contention during gradient accumulation.
Key Optimizations
Duplicate Aggregation & Lookahead: Instead of performing individual atomic updates, the kernel identifies contiguous identical indices using an optimized lookahead mechanism (SKIP_SORTED_INDICES). This collapses multiple redundant updates into a single localized accumulation, significantly reducing atomic contention on grad_weight.
Sub-group Parallel Reduction: For clusters with high duplicate counts, the kernel utilizes sub-group shuffle primitives (shift_group_left) to perform parallel reductions. This ensures that large index blocks are processed across all lanes within a sub-group simultaneously, maximizing compute throughput.
Tiled Stride Optimization: Three specialized kernel variants are introduced to handle different data layouts:
stride_1: Optimized for scalar-like indexing with maximum throughput.
small_stride: Parallelizes across the feature dimension using local work-items.
generic_stride: Handles high-dimensional feature vectors with optimized memory tiling.
SLM-backed Duplicate Cache: Implemented a Shared Local Memory (SLM) cache (smem_dups_cache) to coordinate duplicate counts within a sub-group, reducing redundant global memory fetches for index metadata.