Arm backend: Phase 3 - Vulkan backend uses shared vulkan context. by wwwind · Pull Request #23151 · pytorch/executorch

wwwind · 2026-09-25T13:05:49Z

Implement Phase 3 of the shared Vulkan runtime RFC by integrating SharedVulkanContext into the ExecuTorch Vulkan backend.
RFC: #19298

The Vulkan delegate can now resolve, create, or reuse a shared Vulkan context during initialization and construct an Adapter around the shared VkInstance, VkDevice, and queue.

Key changes:

Add shared-context resolution to Vulkan backend initialization.
Extend Adapter to support borrowed shared devices and queues.
Route shared queue submit/wait operations through with_locked_queue() for cross-delegate synchronization.
Preserve correct shared-device lifetime and ownership semantics.
Keep the legacy Vulkan path unchanged when shared-context options are not provided.
Reuse the existing Vulkan VkInstance and avoid device-level volkLoadDevice() dispatch so multiple shared VkDevices can coexist safely.
Add validation and tests for shared-context creation, reuse, compatibility, lifetime, and queue handling.

This implements the Vulkan-side work described in the RFC and prepares the shared runtime for Phase 4 VGF integration.

Implementation has been assisted with Codex.

cc @SS-JIA @manuelcandales @digantdesai @cbilgin @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani

Signed-off-by: Elena Zhelezina <elena.zhelezina@arm.com> Change-Id: I6bf8e362c4a19c46e4cf2e01335d118d54b85635

pytorch-bot · 2026-09-25T13:05:53Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23151

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (2 Unrelated Failures)

As of commit 9983026 with merge base ad3b6d2 ():

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

Test Vulkan Backend (specialized runners) / test-model-vulkan-e2e / linux-job (gh) (trunk failure)
RuntimeError: Command bash /__w/_temp/exec_script failed with exit code 1
Test Vulkan Backend (specialized runners) / test-vulkan-nvidia / linux-job (gh) (trunk failure)
RuntimeError: Command bash /__w/_temp/exec_script failed with exit code 1

This comment was automatically generated by Dr. CI and updates every 15 minutes.

zingo · 2026-09-25T13:41:36Z

Hi @SS-JIA and @digantdesai is it you or someone else who what to know about this?

SS-JIA

AI review (Codex), static analysis. Findings rechecked against the current PR head.

Change-Id: I0092c8e5eabbe8e24aa7dd8a109602634ff5d26e

Change-Id: Ie723be86551ad70d36c0cd8b436ab593f61f966e

wwwind · 2026-10-27T13:09:52Z

@SS-JIA Could you please take a look ? I addressed the comments

Change-Id: Ibb53efbe3da9849ad58c77f8342cb7ffd7148bf9

Change-Id: I546b2199f8b64b70f84463e133984623c78e35c2

wwwind · 2026-11-03T13:42:37Z

@SS-JIA Thank you for the review. Comments are addressed.
Could you please re-review again ? Thanks

SS-JIA

LGTM, but there are a couple of build failures. Once those are resolved, it should be good to merge. Stamping for now to indicate that changes LGTM.

Change-Id: I6b219abb4abe86189397ab4b77e58d6181061534

Change-Id: I8b514d6ca9419798af6a40250dafa33162da5660

wwwind · 2026-11-05T13:24:17Z

@SS-JIA Please take a look again. I resolve an issue with this file.

meta-codesync · 2026-11-05T23:42:22Z

@SS-JIA has imported this pull request. If you are a Meta employee, you can view this in D129152314.

SS-JIA · 2026-11-06T01:21:11Z

Hi @wwwind — the internal import of this PR (D129152314) fails to build. All 8 failing signals trace to one stale path; no test code ran, so other compile errors may still surface after this fix.

Fix: in backends/vulkan/targets.bzl, the new vulkan_shared_device_test target lists a header that does not exist:

headers = ["runtime/SharedDevice.h"],

There is no runtime/SharedDevice.h in this PR. The header that was added is runtime/ResolveSharedDevice.h (included by ResolveSharedDevice.cpp and both new tests). Change the line to:

headers = ["runtime/ResolveSharedDevice.h"],

Evidence from D129152314:

File not found: `fbsource//xplat/executorch/backends/vulkan/runtime/SharedDevice.h`.
     Included in `xplat/executorch/backends/vulkan/BUCK` but does not exist
Unknown target `vulkan_shared_device_test_no_volk` ...

This blocks vulkan_shared_device_test and vulkan_shared_device_test_no_volk (xplat and fbcode), and the citadel-*-for-executorch-linux-sanity-check jobs that build them. The github-export-checks failure is consequential — it was cancelled because its dependency failed — and should clear once the build passes.

Please update that one line and re-run CI. Happy to re-check the import once updated.

Change-Id: Ie1822d9558b1cbe77f30f7b351fc331567140b24

Summary: #23151 broke several Buck builds: `vulkan_shared_device_test` compares Vulkan queue handles with `VK_NULL_HANDLE`, which some Vulkan headers define as `0`, so gtest compares a pointer with an int and the test does not compile. Compare with `nullptr`. The `_no_volk` variants of that test link `libvulkan`, which the Linux, Android and Apple test hosts do not have. Keep only the volk variant. `vulkan_compute_api` exported `vulkan_shared:runtime`, which pulls in the ExecuTorch runtime. A Windows binary that links the compute API with `link_style = "shared"` then builds `portable_type` as a DLL, which fails to link. The compute API only needs `SharedVulkanDevice`, so move it into its own `vulkan_shared:device` target with no runtime deps and export that instead. On Windows, callers declared the shared API `dllimport`, so a binary that links the static build of the library cannot resolve it. Drop `dllimport` for callers; a plain declaration links against both the import library and the static library. Authored with Metamate (agentcloud). Reviewed By: shoumikhin Differential Revision: D130716215

Arm backend: Phase 3 - Vulkan backend uses shared vulkan context.

6e24c7c

Signed-off-by: Elena Zhelezina <elena.zhelezina@arm.com> Change-Id: I6bf8e362c4a19c46e4cf2e01335d118d54b85635

wwwind requested review from SS-JIA, kirklandsign and larryliu0820 as code owners September 25, 2026 13:05

meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 25, 2026

wwwind mentioned this pull request Sep 25, 2026

RFC: Unified Vulkan and VGF GPU backend path in ExecuTorch #19298

Open

wwwind added the help wanted Extra attention is needed label Oct 5, 2026

SS-JIA reviewed Oct 19, 2026

View reviewed changes

Comment thread backends/vulkan/CMakeLists.txt Resolved

Comment thread backends/vulkan/runtime/SharedContext.cpp Resolved

Comment thread backends/vulkan/runtime/test/SharedGpuContextAdapterTest.cpp Resolved

wwwind added 3 commits October 27, 2026 03:29

Merge branch 'main' into phase3_vulkan

567e3ff

Merge branch 'main' into phase3_vulkan

912d859

Addressed comments.

14c2e84

Change-Id: I0092c8e5eabbe8e24aa7dd8a109602634ff5d26e

wwwind requested a review from mergennachin as a code owner October 27, 2026 04:57

wwwind added 2 commits October 27, 2026 06:28

Fixed wheels.

d82fbe9

Change-Id: Ie723be86551ad70d36c0cd8b436ab593f61f966e

Merge branch 'main' into phase3_vulkan

e42305b

SS-JIA requested changes Nov 2, 2026

View reviewed changes

wwwind added 3 commits November 3, 2026 10:09

Address Vulkan shared device review comments

f539cdc

Change-Id: Ibb53efbe3da9849ad58c77f8342cb7ffd7148bf9

Merge branch 'main' into phase3_vulkan

702d107

Small fix after updating

f47da1f

Change-Id: I546b2199f8b64b70f84463e133984623c78e35c2

SS-JIA approved these changes Nov 3, 2026

View reviewed changes

SS-JIA reviewed Nov 3, 2026

View reviewed changes

Comment thread backends/vulkan/CMakeLists.txt Resolved

wwwind added 2 commits November 3, 2026 21:55

Fix for builds

d987226

Change-Id: I6b219abb4abe86189397ab4b77e58d6181061534

Merge branch 'main' into phase3_vulkan

455c0ef

wwwind added 2 commits November 5, 2026 09:53

Remove obsolete Vulkan benchmark script

3b3e9a7

Change-Id: I8b514d6ca9419798af6a40250dafa33162da5660

Merge branch 'main' into phase3_vulkan

f81ceba

wwwind added 2 commits November 6, 2026 09:45

Fix for the CI failure

4b5350d

Change-Id: Ie1822d9558b1cbe77f30f7b351fc331567140b24

Merge branch 'main' into phase3_vulkan

9983026

SS-JIA merged commit 2e15ded into pytorch:main Nov 6, 2026

SS-JIA mentioned this pull request Nov 10, 2026

Vulkan: fix the shared-device test and Windows builds after #23151 #24843

Merged

Conversation

wwwind commented Sep 25, 2026 • edited by pytorch-bot Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot Bot commented Sep 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23151

✅ You can merge normally! (2 Unrelated Failures)

Uh oh!

zingo commented Sep 25, 2026

Uh oh!

SS-JIA left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

wwwind commented Oct 27, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

wwwind commented Nov 3, 2026

Uh oh!

SS-JIA left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

wwwind commented Nov 5, 2026

Uh oh!

meta-codesync Bot commented Nov 5, 2026

Uh oh!

SS-JIA commented Nov 6, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

wwwind commented Sep 25, 2026 •

edited by pytorch-bot Bot

Loading

pytorch-bot Bot commented Sep 25, 2026 •

edited

Loading