Split off the cooperative queue from the xfer queue by b-sumner · Pull Request #13665 · ROCm/rocm-systems

b-sumner · 2026-10-16T19:17:27Z

Motivation

Split the ROCm internal device queue into a general transfer queue and a cooperative queue

The single internal VirtualGPU (xferQueue_) served both as the blit/transfer queue (xferMgr() and friends) and as the device's cooperative-launch queue, so it was created as an HSA_QUEUE_TYPE_COOPERATIVE queue whenever HIP coop groups were enabled — even though only the coop-launch path needs cooperative semantics.

This separates the two roles:

xferQueue is now non-cooperative (new createInternalXferQueue() helper; createVirtualDevice unchanged). xferMgr() and all transfer callers are unchanged.
New coopQueue, created lazily on the first cooperative launch, holds the single per-device cooperative queue. submitKernel's coop path uses dev().coopQueue() instead of dev().xferQueue().

A cooperative HW queue is now allocated only when a cooperative launch actually occurs, rather than for every HIP process. The two queues each own their blit lock; transfers and coop launches are still ordered by the existing CPU-synchronous blit path and barrier-signal handshakes, not by queue sharing.

No functional change for transfers; cooperative-launch semantics (GWS init, barriers, exclusive access) are unchanged beyond the accessor name.

Test Plan

All existing tests should run unchanged

Test Result

Tests I tried ran unchanged

Submission Checklist

[ x] Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.

@skip-pr-bot

therock-pr-bot · 2026-10-16T19:18:33Z

❌ PR Check — Action Required

Check	Status	Details
📝 PR Description	❌ Fail	Error: PR description must reference a JIRA ID, ISSUE ID, or a GitHub closing keyword. Expected: include a `JIRA ID` / `ISSUE ID` line (separator `:` or `-`, or omitted; value may be a JIRA key, a number with/without `#`, or a link), OR a closing keyword + issue reference. Accepted examples: • `JIRA ID : TESTAUTO-6039` • `JIRA ID - #330` • `JIRA ID #330` • `JIRA ID` (on separate line) `ROCM-25757` • `ISSUE ID : TESTUTO-3334` • `ISSUE ID #3334` • `ISSUE ID - TESTAUTO-3433` • `ISSUE ID` (on separate line) `AIRUNTIME-2352` • `ISSUE ID : https://github.com/<org_name>/<repo_name>/issues/1234` • `Closes #10` • `Fixes octo-org/octo-repo#100` • `Resolves: #123` • `#123` • `https://github.com/<org_name>/<repo_name>/issues/123` Current: no valid JIRA/ISSUE/closing-keyword reference found
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	⚠️ Warning	Error: Source/code files changed without an accompanying unit test. Expected: add at least one test file named like `test_<name>.py` / `test_<name>.cpp` (or `<name>_test.`). Current:* code file(s) changed: `projects/clr/rocclr/device/device.hpp`, `projects/clr/rocclr/device/pal/paldevice.cpp`, `projects/clr/rocclr/device/pal/paldevice.hpp`, `projects/clr/rocclr/device/rocm/rocdevice.cpp`, `projects/clr/rocclr/device/rocm/rocdevice.hpp` (+1 more); no test file found
🔎 pre-commit	⏳ Pending	⏳ Still running…
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

⚠️ 1 policy check(s) failed. Please address the issues above before this PR can be Reviewed.

🚫 Please fix the failed policies

❌ PR Description

The Not ready to Review label was added to this PR. Once all policies pass, the label is removed automatically.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

therock-pr-bot · 2026-10-16T19:18:35Z

🚫 Please fix the failed policies before requesting reviews.

The following policy checks failed:

❌ PR Description

The Not ready to Review label has been added to this PR.
Once all policies pass, the label will be removed automatically.

b-sumner · 2026-10-16T19:20:37Z

This will reduce the exposure of the problem rocgdb has with cooperative queues.

Copilot

🟡 Changes recommended

Dynamic queue reclamation can leave the transfer queue without a hardware queue, causing subsequent kernel-backed transfers to crash.

1 open finding

Pin internal queue before creation to prevent hardware queue reclamation · New

What changed in this PR

Separates ROCm’s internal transfer queue from cooperative launches so cooperative hardware queues are allocated only when needed.

Changes:

Creates a non-cooperative internal transfer queue.
Adds lazy cooperative-queue creation and cleanup.
Routes cooperative kernel launches through the new queue.

File	Description
projects/clr/rocclr/device/rocm/rocvirtual.cpp	Uses the cooperative queue for cooperative launches.
projects/clr/rocclr/device/rocm/rocdevice.hpp	Declares separate queue state and accessors.
projects/clr/rocclr/device/rocm/rocdevice.cpp	Implements separate queue creation and cleanup.

🧠 Review effort: Balanced

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

mangupta

🤖 AI-assisted pre-review (for @mangupta)

mangupta · 2026-10-20T01:14:08Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (f9c4ffa36a76) — no new findings at or above the posting threshold.

mangupta · 2026-10-21T01:12:18Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (4dbf4d38f603) — no new findings at or above the posting threshold.

mangupta · 2026-10-22T01:17:32Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (ca96d33c801c) — no new findings at or above the posting threshold.

mangupta · 2026-10-22T15:32:42Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (68f37aa7d803) — no new findings at or above the posting threshold.

mangupta · 2026-10-23T01:12:41Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (1a6e53cc899f) — no new findings at or above the posting threshold.

mangupta · 2026-10-23T13:43:46Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (c6b913f80933) — no new findings at or above the posting threshold.

mangupta · 2026-10-25T15:37:56Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (cc30958ec225) — no new findings at or above the posting threshold.

Make the internal xfer queue non-cooperative and add a separate coopQueue, created lazily on the first cooperative launch. Route both through createVirtualDevice() via a new cooperative argument, and pin the non-cooperative xfer queue before create() so its HW queue is not reclaimed.

mangupta · 2026-10-26T15:19:10Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (02825b347c78) — no new findings at or above the posting threshold.

geomin12 · 2026-10-26T16:36:23Z

hi @b-sumner ! thanks for this PR, i notice you enabled ci:run-all-archs which is great! thanks for using that, we accidentially included some test machines that weren't meant to run. we are fixing it in #14272, so please re-base once this lands!

i noticed those jobs (particularly gfx117X) were massively queued. i cancelled the job!

b-sumner · 2026-10-26T18:18:09Z

I have created PR 14396 to eliminate some of the false failures that are showing up here.

mangupta · 2026-10-28T14:28:18Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (fbc351579f2c) — no new findings at or above the posting threshold.

mangupta · 2026-10-30T01:37:20Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (2876ae46e774) — no new findings at or above the posting threshold.

mangupta · 2026-10-30T15:57:10Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (b2c108e09037) — no new findings at or above the posting threshold.

therock-pr-bot · 2026-11-02T18:09:36Z

✅ Author chose to skip pr bot run hence removing label (@skip-pr-bot found in the PR description).

mangupta · 2026-11-03T01:38:22Z

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (84fb48fcb693) — no new findings at or above the posting threshold.

litvaOo · 2026-11-04T11:56:10Z

Errors are unrelated to PR's content, overriding.

systems-assistant · 2026-11-05T15:44:24Z

TheRock Submodule Bump Activity

Newest first

2026-11-05 — Included in TheRock via ROCm/TheRock#9739.

b-sumner requested a review from a team as a code owner October 16, 2026 19:17

Copilot AI balanced review requested due to automatic review settings October 16, 2026 19:17

github-actions Bot added the project: clr label Oct 16, 2026

Copilot started reviewing on behalf of b-sumner October 16, 2026 19:18

View session

therock-pr-bot Bot added the Not ready to Review label Oct 16, 2026

Copilot AI reviewed Oct 16, 2026

View reviewed changes

Comment thread projects/clr/rocclr/device/rocm/rocdevice.cpp Resolved

systems-assistant Bot added the organization: ROCm label Oct 16, 2026

b-sumner force-pushed the users/b-sumner/split-xfer-queue branch from f7d21af to 33f6eaa Compare October 16, 2026 22:50

gandryey approved these changes Oct 16, 2026

View reviewed changes

mangupta reviewed Oct 19, 2026

View reviewed changes

Comment thread projects/clr/rocclr/device/rocm/rocdevice.cpp Resolved

mangupta added the lrt-bot: review label Oct 19, 2026

mangupta added lrt-bot: ready and removed lrt-bot: review labels Oct 20, 2026

lumachad added the ci:run-all-archs Opt-in to building for all architectures on a pull request label Oct 20, 2026

lancesix mentioned this pull request Oct 23, 2026

"feat(hostcall) : Reland : enable the new hostcall implementation by default"" ROCm/TheRock#9432

Merged

litvaOo force-pushed the users/b-sumner/split-xfer-queue branch from cc30958 to 0ff9414 Compare October 26, 2026 08:27

Merge branch 'develop' into users/b-sumner/split-xfer-queue

02825b3

Merge branch 'develop' into users/b-sumner/split-xfer-queue

fbc3515

Merge branch 'develop' into users/b-sumner/split-xfer-queue

2876ae4

Merge branch 'develop' into users/b-sumner/split-xfer-queue

b2c108e

b-sumner added 2 commits November 2, 2026 06:51

Merge branch 'develop' into users/b-sumner/split-xfer-queue

8f073e5

Merge branch 'develop' into users/b-sumner/split-xfer-queue

84fb48f

therock-pr-bot Bot removed the Not ready to Review label Nov 2, 2026

litvaOo merged commit fe3963a into develop Nov 4, 2026
1543 of 1614 checks passed

litvaOo deleted the users/b-sumner/split-xfer-queue branch November 4, 2026 11:56

mangupta removed the lrt-bot: ready label Nov 4, 2026

Conversation

b-sumner commented Oct 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Test Plan

Test Result

Submission Checklist

Uh oh!

therock-pr-bot Bot commented Oct 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

❌ PR Check — Action Required

Uh oh!

therock-pr-bot Bot commented Oct 16, 2026

Uh oh!

b-sumner commented Oct 16, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

mangupta left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

mangupta commented Oct 20, 2026

Uh oh!

mangupta commented Oct 21, 2026

Uh oh!

mangupta commented Oct 22, 2026

Uh oh!

mangupta commented Oct 22, 2026

Uh oh!

mangupta commented Oct 23, 2026

Uh oh!

mangupta commented Oct 23, 2026

Uh oh!

mangupta commented Oct 25, 2026

Uh oh!

mangupta commented Oct 26, 2026

Uh oh!

geomin12 commented Oct 26, 2026

Uh oh!

b-sumner commented Oct 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mangupta commented Oct 28, 2026

Uh oh!

mangupta commented Oct 30, 2026

Uh oh!

mangupta commented Oct 30, 2026

Uh oh!

therock-pr-bot Bot commented Nov 2, 2026

Uh oh!

mangupta commented Nov 3, 2026

Uh oh!

litvaOo commented Nov 4, 2026

Uh oh!

Uh oh!

systems-assistant Bot commented Nov 5, 2026

TheRock Submodule Bump Activity

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

b-sumner commented Oct 16, 2026 •

edited

Loading

therock-pr-bot Bot commented Oct 16, 2026 •

edited

Loading

b-sumner commented Oct 26, 2026 •

edited

Loading