[ET-VK][conv2d] Cap im2col scratch memory via output-height tiling#20657
Merged
Conversation
Pull Request resolved: #20653 The im2col + GEMM conv2d path materializes an im2col scratch tensor of size M * K_total * elem (M = N * H_out * W_out, K_total = K_h * K_w * align_up_4(C_in)) as a single shared tensor, allocated during graph build and resident for the model's lifetime. For full-resolution convolutions this scratch is very large -- a 64-channel 3x3 conv at 256x256 in FP32 materializes ~144 MB, and at 512x512 in FP16 ~288 MB. On memory-constrained mobile GPUs, where GPU allocations come from unified, non-reclaimable system memory, this can nearly double peak process memory and trigger the OS low-memory killer. This change tiles the im2col + GEMM over output-height rows to a fixed scratch budget (kIm2colScratchBudgetBytes, 16 MB). A single scratch tensor sized to oh_tile output rows is reused across tiles, with an oh_offset selecting the live row window per tile. The GEMM inner loop is byte-identical, so the GEMM-based speedup is preserved; scratch becomes O(budget) instead of O(M * K_total), making it resolution-independent. Tiling along output-height (rather than flattened M) keeps the row -> (oh, ow) decode exact for all three storage variants (buffer, texture2d, texture3d). The fixed per-build tile count is safe because tensors are built at the dynamic upper bound, so runtime shapes only shrink and trailing tiles no-op via the shader's oh < H_out guard. oh_tile reaches the resize callbacks as a raw int packed into the resize_args slot (read via static_cast, not get_int) to avoid materializing a graph Value for a build-time constant. The direct-conv fallback for small shapes is unchanged. ghstack-source-id: 398747242 @exported-using-ghexport Differential Revision: [D110231992](https://our.internmc.facebook.com/intern/diff/D110231992/)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20657
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 1 Cancelled Job, 2 Unrelated FailuresAs of commit 8c545a4 with merge base 6d0b933 ( NEW FAILURE - The following job has failed:
CANCELLED JOB - The following job was cancelled. Please retry:
FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
SS-JIA
approved these changes
Jul 1, 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: #20653 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/566/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/566/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/main
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/566/orig
Differential Revision: D110231992
@diff-train-skip-merge