[ET-VK] Fix missing memory barrier for first-use writes on aliased tensors #17309
+18
−10
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
…nsors Tensors sharing physical memory via SharedObject each track their own `last_access_` independently. When a tensor's first access is a write, `prev_stage` is `NO_STAGE`, causing `transition()` to use `TOP_OF_PIPE_BIT` as `srcStageMask` with no `srcAccessMask` — effectively a no-op barrier. If the same physical memory was previously written through a different aliased tensor handle, this creates a WAW hazard where the new write may execute before or concurrently with the prior write, producing non-deterministic results. This was observed as non-deterministic q8ta_conv2d output in ResNet50: running the model twice with the same input produced slightly different quantized int8 values. Adding a debug print shader after each conv2d dispatch masked the issue because the print node's read-after-write barrier serialized GPU work. The fix: when `prev_stage` is `NO_STAGE` and the current access is a write, use `COMPUTE_SHADER_BIT` with `SHADER_WRITE_BIT` instead of `TOP_OF_PIPE_BIT` with no access flags. This ensures all prior compute shader work completes and its writes are made visible before the new write begins. Authored with Claude. Differential Revision: [D92715369](https://our.internmc.facebook.com/intern/diff/D92715369/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17309
Note: Links to docs will display an error until the docs builds have been completed. ❌ 7 New Failures, 3 Unrelated FailuresAs of commit 0c048bb with merge base ba89c69 ( NEW FAILURES - The following jobs have failed:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
BROKEN TRUNK - The following jobs failed but was present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
SS-JIA
pushed a commit
that referenced
this pull request
Feb 9, 2026
…nsors Tensors sharing physical memory via SharedObject each track their own `last_access_` independently. When a tensor's first access is a write, `prev_stage` is `NO_STAGE`, causing `transition()` to use `TOP_OF_PIPE_BIT` as `srcStageMask` with no `srcAccessMask` — effectively a no-op barrier. If the same physical memory was previously written through a different aliased tensor handle, this creates a WAW hazard where the new write may execute before or concurrently with the prior write, producing non-deterministic results. This was observed as non-deterministic q8ta_conv2d output in ResNet50: running the model twice with the same input produced slightly different quantized int8 values. Adding a debug print shader after each conv2d dispatch masked the issue because the print node's read-after-write barrier serialized GPU work. The fix: when `prev_stage` is `NO_STAGE` and the current access is a write, use `COMPUTE_SHADER_BIT` with `SHADER_WRITE_BIT` instead of `TOP_OF_PIPE_BIT` with no access flags. This ensures all prior compute shader work completes and its writes are made visible before the new write begins. Authored with Claude. Differential Revision: [D92715369](https://our.internmc.facebook.com/intern/diff/D92715369/) ghstack-source-id: 339490477 Pull Request resolved: #17309
This PR needs a
|
… aliased tensors" Tensors sharing physical memory via SharedObject each track their own `last_access_` independently. When a tensor's first access is a write, `prev_stage` is `NO_STAGE`, causing `transition()` to use `TOP_OF_PIPE_BIT` as `srcStageMask` with no `srcAccessMask` — effectively a no-op barrier. If the same physical memory was previously written through a different aliased tensor handle, this creates a WAW hazard where the new write may execute before or concurrently with the prior write, producing non-deterministic results. This was observed as non-deterministic q8ta_conv2d output in ResNet50: running the model twice with the same input produced slightly different quantized int8 values. Adding a debug print shader after each conv2d dispatch masked the issue because the print node's read-after-write barrier serialized GPU work. The fix: when `prev_stage` is `NO_STAGE` and the current access is a write, use `COMPUTE_SHADER_BIT` with `SHADER_WRITE_BIT` instead of `TOP_OF_PIPE_BIT` with no access flags. This ensures all prior compute shader work completes and its writes are made visible before the new write begins. Authored with Claude. Differential Revision: [D92715369](https://our.internmc.facebook.com/intern/diff/D92715369/) [ghstack-poisoned]
SS-JIA
pushed a commit
that referenced
this pull request
Feb 9, 2026
…nsors Pull Request resolved: #17309 Tensors sharing physical memory via SharedObject each track their own `last_access_` independently. When a tensor's first access is a write, `prev_stage` is `NO_STAGE`, causing `transition()` to use `TOP_OF_PIPE_BIT` as `srcStageMask` with no `srcAccessMask` — effectively a no-op barrier. If the same physical memory was previously written through a different aliased tensor handle, this creates a WAW hazard where the new write may execute before or concurrently with the prior write, producing non-deterministic results. This was observed as non-deterministic q8ta_conv2d output in ResNet50: running the model twice with the same input produced slightly different quantized int8 values. Adding a debug print shader after each conv2d dispatch masked the issue because the print node's read-after-write barrier serialized GPU work. The fix: when `prev_stage` is `NO_STAGE` and the current access is a write, use `COMPUTE_SHADER_BIT` with `SHADER_WRITE_BIT` instead of `TOP_OF_PIPE_BIT` with no access flags. This ensures all prior compute shader work completes and its writes are made visible before the new write begins. Authored with Claude. ghstack-source-id: 339541957 @exported-using-ghexport Differential Revision: [D92715369](https://our.internmc.facebook.com/intern/diff/D92715369/)
… aliased tensors" Tensors sharing physical memory via SharedObject each track their own `last_access_` independently. When a tensor's first access is a write, `prev_stage` is `NO_STAGE`, causing `transition()` to use `TOP_OF_PIPE_BIT` as `srcStageMask` with no `srcAccessMask` — effectively a no-op barrier. If the same physical memory was previously written through a different aliased tensor handle, this creates a WAW hazard where the new write may execute before or concurrently with the prior write, producing non-deterministic results. This was observed as non-deterministic q8ta_conv2d output in ResNet50: running the model twice with the same input produced slightly different quantized int8 values. Adding a debug print shader after each conv2d dispatch masked the issue because the print node's read-after-write barrier serialized GPU work. The fix: when `prev_stage` is `NO_STAGE` and the current access is a write, use `COMPUTE_SHADER_BIT` with `SHADER_WRITE_BIT` instead of `TOP_OF_PIPE_BIT` with no access flags. This ensures all prior compute shader work completes and its writes are made visible before the new write begins. Authored with Claude. Differential Revision: [D92715369](https://our.internmc.facebook.com/intern/diff/D92715369/) [ghstack-poisoned]
SS-JIA
pushed a commit
that referenced
this pull request
Feb 9, 2026
…nsors Pull Request resolved: #17309 Tensors sharing physical memory via SharedObject each track their own `last_access_` independently. When a tensor's first access is a write, `prev_stage` is `NO_STAGE`, causing `transition()` to use `TOP_OF_PIPE_BIT` as `srcStageMask` with no `srcAccessMask` — effectively a no-op barrier. If the same physical memory was previously written through a different aliased tensor handle, this creates a WAW hazard where the new write may execute before or concurrently with the prior write, producing non-deterministic results. This was observed as non-deterministic q8ta_conv2d output in ResNet50: running the model twice with the same input produced slightly different quantized int8 values. Adding a debug print shader after each conv2d dispatch masked the issue because the print node's read-after-write barrier serialized GPU work. The fix: when `prev_stage` is `NO_STAGE` and the current access is a write, use `COMPUTE_SHADER_BIT` with `SHADER_WRITE_BIT` instead of `TOP_OF_PIPE_BIT` with no access flags. This ensures all prior compute shader work completes and its writes are made visible before the new write begins. Authored with Claude. ghstack-source-id: 339541957 @exported-using-ghexport Differential Revision: [D92715369](https://our.internmc.facebook.com/intern/diff/D92715369/)
manuelcandales
approved these changes
Feb 9, 2026
… aliased tensors" Tensors sharing physical memory via SharedObject each track their own `last_access_` independently. When a tensor's first access is a write, `prev_stage` is `NO_STAGE`, causing `transition()` to use `TOP_OF_PIPE_BIT` as `srcStageMask` with no `srcAccessMask` — effectively a no-op barrier. If the same physical memory was previously written through a different aliased tensor handle, this creates a WAW hazard where the new write may execute before or concurrently with the prior write, producing non-deterministic results. This was observed as non-deterministic q8ta_conv2d output in ResNet50: running the model twice with the same input produced slightly different quantized int8 values. Adding a debug print shader after each conv2d dispatch masked the issue because the print node's read-after-write barrier serialized GPU work. The fix: when `prev_stage` is `NO_STAGE` and the current access is a write, use `COMPUTE_SHADER_BIT` with `SHADER_WRITE_BIT` instead of `TOP_OF_PIPE_BIT` with no access flags. This ensures all prior compute shader work completes and its writes are made visible before the new write begins. Authored with Claude. Differential Revision: [D92715369](https://our.internmc.facebook.com/intern/diff/D92715369/) [ghstack-poisoned]
SS-JIA
pushed a commit
that referenced
this pull request
Feb 10, 2026
…nsors Pull Request resolved: #17309 Tensors sharing physical memory via SharedObject each track their own `last_access_` independently. When a tensor's first access is a write, `prev_stage` is `NO_STAGE`, causing `transition()` to use `TOP_OF_PIPE_BIT` as `srcStageMask` with no `srcAccessMask` — effectively a no-op barrier. If the same physical memory was previously written through a different aliased tensor handle, this creates a WAW hazard where the new write may execute before or concurrently with the prior write, producing non-deterministic results. This was observed as non-deterministic q8ta_conv2d output in ResNet50: running the model twice with the same input produced slightly different quantized int8 values. Adding a debug print shader after each conv2d dispatch masked the issue because the print node's read-after-write barrier serialized GPU work. The fix: when `prev_stage` is `NO_STAGE` and the current access is a write, use `COMPUTE_SHADER_BIT` with `SHADER_WRITE_BIT` instead of `TOP_OF_PIPE_BIT` with no access flags. This ensures all prior compute shader work completes and its writes are made visible before the new write begins. Authored with Claude. ghstack-source-id: 339884030 @exported-using-ghexport Differential Revision: [D92715369](https://our.internmc.facebook.com/intern/diff/D92715369/)
meta-codesync
bot
merged commit d78828d
into
gh/SS-JIA/412/base
179 of 189 checks passed
SS-JIA
pushed a commit
that referenced
this pull request
Feb 10, 2026
…nsors Pull Request resolved: #17309 Tensors sharing physical memory via SharedObject each track their own `last_access_` independently. When a tensor's first access is a write, `prev_stage` is `NO_STAGE`, causing `transition()` to use `TOP_OF_PIPE_BIT` as `srcStageMask` with no `srcAccessMask` — effectively a no-op barrier. If the same physical memory was previously written through a different aliased tensor handle, this creates a WAW hazard where the new write may execute before or concurrently with the prior write, producing non-deterministic results. This was observed as non-deterministic q8ta_conv2d output in ResNet50: running the model twice with the same input produced slightly different quantized int8 values. Adding a debug print shader after each conv2d dispatch masked the issue because the print node's read-after-write barrier serialized GPU work. The fix: when `prev_stage` is `NO_STAGE` and the current access is a write, use `COMPUTE_SHADER_BIT` with `SHADER_WRITE_BIT` instead of `TOP_OF_PIPE_BIT` with no access flags. This ensures all prior compute shader work completes and its writes are made visible before the new write begins. Authored with Claude. ghstack-source-id: 339884030 @exported-using-ghexport Differential Revision: [D92715369](https://our.internmc.facebook.com/intern/diff/D92715369/)
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.
Stack from ghstack (oldest at bottom):
Tensors sharing physical memory via SharedObject each track their own
last_access_independently. When a tensor's first access is a write,prev_stageisNO_STAGE, causingtransition()to useTOP_OF_PIPE_BITassrcStageMaskwith nosrcAccessMask— effectivelya no-op barrier. If the same physical memory was previously written
through a different aliased tensor handle, this creates a WAW hazard
where the new write may execute before or concurrently with the prior
write, producing non-deterministic results.
This was observed as non-deterministic q8ta_conv2d output in ResNet50:
running the model twice with the same input produced slightly different
quantized int8 values. Adding a debug print shader after each conv2d
dispatch masked the issue because the print node's read-after-write
barrier serialized GPU work.
The fix: when
prev_stageisNO_STAGEand the current access is awrite, use
COMPUTE_SHADER_BITwithSHADER_WRITE_BITinstead ofTOP_OF_PIPE_BITwith no access flags. This ensures all prior computeshader work completes and its writes are made visible before the new
write begins.
Authored with Claude.
Differential Revision: D92715369