[ET-VK][conv2d] Re-implement pointwise conv2d with tiled compute and blocked weight packing#18300
Merged
SS-JIA merged 2 commits intogh/SS-JIA/491/origfrom Mar 18, 2026
Merged
Conversation
…blocked weight packing Pull Request resolved: #18292 Profiling EdgeTAM on Adreno shows pointwise 1×1 convolutions are a dominant bottleneck. This diff re-implements the stride=1, padding=0 pointwise path using the same tiled matmul approach as the recently landed linear shader rewrite. The new `conv2d_pw_tiled` shader reuses the shared linear tiled infrastructure (FPInputTile, FPWeightTile, FPOutTile, fp_accumulate_with_fp_weight, packed weight tile loading) with custom input/output tile load/store functions that map flat spatial indices to channels-packed texture3d coordinates. Weight packing uses the same 4OC×4IC blocked format as linear via the `pack_fp_linear_weight` shader. Dispatch uses DynamicDispatchNode for correct workgroup size updates during graph resizing. Only the stride=1, padding=0 pointwise path is changed; the general conv2d_pw shader for arbitrary stride/padding is left unchanged. EdgeTAM first frame on Samsung S25 (Adreno 830): 208 ms → 196 ms (~6%). Authored with Claude. ghstack-source-id: 353941147 @exported-using-ghexport Differential Revision: [D96756792](https://our.internmc.facebook.com/intern/diff/D96756792/)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18300
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
…device-based tile selection Pull Request resolved: #18293 Profiling showed depthwise conv2d is 5-15x slower on Mali GPUs vs Adreno due to register pressure from the 4x2 output tile (17 vec4 registers per thread). Benchmarking confirmed that reducing the tile to 1x1 (7 vec4 registers) gives 4-15x speedup on Mali with no regression on Adreno. This change extracts depthwise conv2d dispatch logic from Convolution.cpp into a new Conv2dDW.cpp (following the Conv2dPW.cpp pattern), and adds device-based tile size selection: b1x1 on Mali, b4x2 (current default) on Adreno. ghstack-source-id: 353940602 @exported-using-ghexport Differential Revision: [D97058158](https://our.internmc.facebook.com/intern/diff/D97058158/)
SS-JIA
approved these changes
Mar 18, 2026
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
2 participants
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.
This PR was created by the merge bot to help merge the original PR into the main branch.
ghstack PR number: #18292 by @SS-JIA
^ Please use this as the source of truth for the PR details, comments, and reviews
ghstack PR base: https://github.com/pytorch/executorch/tree/gh/SS-JIA/492/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/492/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/gh/SS-JIA/491/orig
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/492/orig
Differential Revision: D96756792
@diff-train-skip-merge