[ET-VK][ez] Use tree reduction in q8ta_linear_gemv shader#17808
Merged
Conversation
Replace the serial O(WGS) reduction loop with a tree reduction pattern (O(log2(WGS))). Previously, only thread 0 summed all 64 partial accumulators sequentially. Now all threads participate in a classic halving reduction, matching the pattern already used in linear_q4gsw_coop.glsl. Authored by Claude. Differential Revision: [D94949137](https://our.internmc.facebook.com/intern/diff/D94949137/) ghstack-source-id: 346524552 Pull Request resolved: #17792
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17808
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
This PR needs a
|
added 3 commits
March 3, 2026 09:58
Previously, the im2col + pointwise GEMM path (`q8ta_conv2d_im2col`) only supported non-grouped convolutions (groups=1). This diff extends it to handle grouped convolutions as well, providing significant speedups on Mali GPUs. The key changes are: **PW GEMM shader (`q8ta_conv2d_pw.glsl`)**: Added `K4_per_group` and `OC4_per_group` as push constants. The shader now computes a group index from the output channel block (`group_idx = oc_block_idx / OC4_per_group`) and offsets the im2col input read by `group_idx * K4_per_group`. For non-grouped cases (groups=1), `group_idx` is always 0, so behavior is unchanged. **PW node (`Q8taConv2dPW.cpp`)**: `add_q8ta_conv2d_pw_node` now accepts a `groups` parameter (default=1) and computes `K4_per_group` and `OC4_per_group` internally from the input/output tensor dimensions. `K4_per_group` and `OC4_per_group` were previously specialization constants; they are now push constants to avoid shader variant explosion when groups varies. **Im2col node (`Q8taConv2dIm2Col.cpp`)**: Removed the `groups == 1` assertion from `add_q8ta_im2col_node`. The im2col shader already handles groups correctly (each group's K range is contiguous in the output buffer). The `q8ta_conv2d_im2col` operator now passes the groups value through to the PW node. **Dispatch heuristic (`Q8taConv2d.cpp`)**: Updated `q8ta_conv2d` with device-aware dispatch. On Mali, im2col is used for all eligible cases (grouped and ungrouped) since it provides 1.2-3.6x speedups. On Adreno, im2col is only used for ungrouped convolutions (groups=1) where in_channels_per_group >= 32 or spatial_out <= 4096, since grouped convolutions show 0.7-0.95x regression with im2col. The heuristic uses `graph.device_is_mali()` to select the path. **Tests (`test_q8ta_conv2d.cpp`)**: Updated im2col test eligibility from `groups == 1 && channels.in % 4 == 0` to `in_channels_per_group % 4 == 0`, enabling im2col testing for grouped cases. Added SceneX v9 256x256 grouped convolution configs. Differential Revision: [D94949480](https://our.internmc.facebook.com/intern/diff/D94949480/) ghstack-source-id: 346525921 Pull Request resolved: #17793
…e-adaptive conv2d Performance testing of quantized int8 convolutions reveals that different algorithms perform better on different GPU architectures: im2col is faster on Mali while direct convolution is faster on Adreno. The optimal memory layout differs per algorithm (4C for im2col, 4C1W for direct convolution). This introduces a new "dynamic" memory layout PACKED_INT8_CONV2D that is serialized at export time and resolved to a concrete layout at runtime based on the device's GPU architecture. The resolution logic in ResolveLayouts.cpp mirrors the im2col vs direct convolution decision in Q8taConv2d.cpp. Differential Revision: [D94949134](https://our.internmc.facebook.com/intern/diff/D94949134/) ghstack-source-id: 346525918 Pull Request resolved: #17794
Add the ability to override the Vulkan device name at runtime so that device-adaptive code paths (e.g. memory layout selection) can be tested on hardware that doesn't match the overridden device type. PhysicalDevice::override_device_name() and Adapter::override_device_name() are added behind VULKAN_DEBUG. The device type detection logic is refactored into a reusable determine_device_type() helper to avoid duplication between the constructor and the override function. All test binaries in fb/test/models/ (classification, greenscreen, scenex, skin_seg) now accept --gpu_name to invoke the override before loading the model. The Skycastle CI workflows are updated to re-run classification and greenscreen tests with --gpu_name Mali-G715 in addition to the default run. Differential Revision: [D94949136](https://our.internmc.facebook.com/intern/diff/D94949136/) ghstack-source-id: 346525920 Pull Request resolved: #17795
SS-JIA
approved these changes
Mar 3, 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: #17792 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/453/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/453/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/453/orig
Differential Revision: D94949137
@diff-train-skip-merge