Codestin Search App

AaronStGeorge · 2025-10-24T18:11:15Z

Motivation

As part of the effort to create IREE/Fusilli based hipDNN plugin in TheRock, this PR enables github action based CI for fusilli-plugin, adds a CODEOWNERS section, and updates the label mapping for the PR labeling bot.

Technical Details

This PR is part of a series that will move the initial efforts toward bringing and IREE/Fusilli based hipDNN plugin into TheRock (see RFC0004-Fusilli-IREE-Kernel-Provider-hipDNN.md):

PR 1 Directly move existing code from nod-ai/shark-ai without any edits. At this point, the plugin will not build, and will not be running in rocm-libraries CI through TheRock or an individual GitHub actions runner.
- [FusilliPlugin] Lift and shift fusilli-plugin: shark-ai -> rocm-libraries #2252
PR 2 Follow up with changes to update and re-enable the build, enable CI, and set up fusilli-plugin as a proper rocm-libraries project. At this point fusilli-plugin CI should run in individual GitHub actions runner CI. Note: TheRock CI will still not be enabled. TheRock CI is the main CI platform for rocm-libraries, but a project must build as part of TheRock to participate.
- [FusilliPlugin] Re-Enable build and CI and add CODEOWNERS etc. #2269 (this PR)
PR 3 Remove fusilli-plugin from shark-ai.
PR 4/N Enable building fusilli plugin optionally in TheRock and enable TheRock CI.
PR N+1 Remove CI added in PR 2 as things will now be tested through TheRock CI.

Test Plan

CI now runs fusilli-plugin CI.

Test Result

CI looks green!

Submission Checklist

Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

jayhawk-commits

The files outside projects/fusilli-plugin directory look fine to me.

jayhawk-commits · 2025-10-27T21:46:52Z

@amd-jnovotny Could somebody from docs approve this PR? This adds a new project to rocm-libraries, which includes some .md files so it pulls in the docs team.

sjain-stanford · 2025-10-27T23:04:04Z

+
+#===----------------------------------------------------------------------===#
+#
+# This script is copied from fusilli:


Since you're downloading fusilli (at a given hash), is it possible to reuse this script from there instead of cloning it?

sjain-stanford

fusilli-plugin/ changes LGTM apart from a non-blocking comment.

ScottTodd · 2025-10-28T02:47:41Z

PR 2 ... Note: TheRock CI will still not be enabled. TheRock CI is the main CI platform for rocm-libraries, but a project must build as part of TheRock to participate.

PR N+1 Remove CI added in PR 2 as things will now be tested through TheRock CI.

I'm okay with the end state, but this intermediate state has some details that concern me a bit.

For example, none of the other subprojects have their own dedicated workflows and I don't see this fitting into the existing gardener rotations (https://github.com/ROCm/rocm-libraries/blob/develop/docs/gardening.md) until this uses the same workflow style as other subprojects.

ScottTodd · 2025-10-28T02:48:28Z

+    defaults:
+      run:
+        # --noprofile skips loading ~/.bash_profile
+        # --norc      skips loading ~/.bashrc
+        # -exo pipefail:
+        #   -e exit immediately if any command fails
+        #   -x print each command before executing
+        #   -o pipefail ensure pipeline fails if any command in it fails
+        # {0} github actions placeholder for the command to run
+        shell: bash --noprofile --norc -exo pipefail {0}


Comments should also explain why this is customized. GitHub Actions has good defaults for bash as the shell anyways, so I'm not sure what this is changing.

ScottTodd · 2025-10-28T02:48:52Z

+    strategy:
+      fail-fast: false
+      matrix:
+        name: ["gfx942_clang20_debug"]
+        include:
+          - name: gfx942_clang20_debug


Why is this using a matrix with only one element?

ScottTodd · 2025-10-28T02:51:06Z

+    - name: build fusilli-plugin
+      # The docker container mounts the working directory as a volume, so we
+      # must run from github.workspace to ensure fusilli-plugin can reach into
+      # ../sharkfuser for its fusilli dependency.
+      working-directory: ${{ github.workspace }}
+      run: |
+        ${{ env.FUSILLI_PLUGIN_DIR }}/build_tools/docker/exec_docker_ci.sh \
+        bash -c "cd projects/fusilli-plugin && \
+                 cmake -GNinja -S. -Bbuild \
+                   ${{ matrix.fusilli-plugin-cmake-options }} && \
+                 cmake --build build --target all"


Requiring Docker specific code with bundled source/dependencies for building is a huge red flag to me.

Does this project have plans for Windows support?

What is in the dockerfile?

Can the entire workflow run in a container instead of just this specific step via a script?

Can the build run on a CPU machine and tests run on a GPU machine?

ScottTodd · 2025-10-28T02:53:59Z

+# ROCm requires accesses to the host's /dev/kfd and /dev/dri/* device nodes, typically
+# owned by the `render` and `video` groups. The groups' GIDs in the container must
+# match the host's to access the resources. Sometimes the device nodes may be owned by
+# dynamic GIDs (that don't belong to the `render` or `video` groups). So instead of
+# adding user to the GIDs of named groups (obtained from `getent group render` or
+# `getent group video`), we simply check the owning GID of the device nodes on the host
+# and pass it to `docker run` with `--group-add=<GID>`.
+for DEV in /dev/kfd /dev/dri/*; do


Relating to my prior comment, there should be a separation between build and test containers and requirements. Mixing the two risks release builds generating artifacts that are not portable.

ScottTodd · 2025-10-28T02:54:37Z

+           -v "${PWD}":/workspace \
+           ${DOCKER_RUN_DEVICE_OPTS} \
+           --security-opt seccomp=unconfined \
+           ghcr.io/sjain-stanford/compiler-dev-ubuntu-24.04:main@sha256:d52a5eb21ce21509f5fd1064074ba34f7ad8810c5d5c6caff9790149c8e05b3c \


No using dockerfiles from user repositories. This needs to be moved to a shared location if used at all.

To be stronger on this, this is a red flag with prevents merging this as is.

ScottTodd · 2025-10-28T02:55:56Z

-# Options
-option(FUSILLI_PLUGIN_USE_LOCAL_FUSILLI "Use local Fusilli build from ../sharkfuser" ON)
+set(FUSILLI_HASH "5970f13835942213c7d4fd6171b936ee9fe9be73" CACHE STRING "Git hash for Fusilli")
+set(HIP_DNN_HASH "4e0a0452cfcb8fdb86e9c40a6e43debab4d4ecbc" CACHE STRING "Git hash for hipDNN")


While in this superrepo, this must depend on the source in https://github.com/ROCm/rocm-libraries/tree/develop/projects/hipdnn, not fetch from a fixed commit hash.

I strongly agree with Scott here. We cannot have another projects that pulls it's own version from source sitting next to it. Both need to work at HEAD.

Is a

FetchContent_Declare( hipdnn SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../hipdnn )

acceptable? Once fusilli-plugin is fully integrated into TheRock the plan was to use the find_package. IIUC find_package + TheRock dependency providers ensure that the build is using hipDNN built as part of TheRock. This was intermediate state.

ScottTodd · 2025-10-28T02:57:31Z

Most (all?) other licenses in rocm-libraries subprojects are MIT. Should we pursue a re-licensing of this project from "Apache v2 with LLVM Exceptions" to MIT?

This indeed needs to be clarified. Same applies to the workflow file added as part of the PR and licensed under Apache 2.0

Apache 2.0 W LLVM exceptions is marked "pre-approved for use in AMD projects" on the OSS License List so I think it's fine. That being said, with only three contributors it would be quite easy to re-license so we might want to move towards MIT for consistency.

marbre

I agree with Scott and see (the same) red flags which need to be addressed before this can be merged.

marbre · 2025-10-28T03:01:47Z

+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception


Not sure this should be the default license for workflows.

marbre · 2025-10-28T03:04:17Z

-# Options
-option(FUSILLI_PLUGIN_USE_LOCAL_FUSILLI "Use local Fusilli build from ../sharkfuser" ON)
+set(FUSILLI_HASH "5970f13835942213c7d4fd6171b936ee9fe9be73" CACHE STRING "Git hash for Fusilli")
+set(HIP_DNN_HASH "4e0a0452cfcb8fdb86e9c40a6e43debab4d4ecbc" CACHE STRING "Git hash for hipDNN")


I strongly agree with Scott here. We cannot have another projects that pulls it's own version from source sitting next to it. Both need to work at HEAD.

marbre · 2025-10-28T03:05:07Z

This indeed needs to be clarified. Same applies to the workflow file added as part of the PR and licensed under Apache 2.0

marbre · 2025-10-28T03:06:53Z

+           -v "${PWD}":/workspace \
+           ${DOCKER_RUN_DEVICE_OPTS} \
+           --security-opt seccomp=unconfined \
+           ghcr.io/sjain-stanford/compiler-dev-ubuntu-24.04:main@sha256:d52a5eb21ce21509f5fd1064074ba34f7ad8810c5d5c6caff9790149c8e05b3c \


To be stronger on this, this is a red flag with prevents merging this as is.

amd-jnovotny · 2025-10-28T12:36:17Z

@jayhawk-commits Done

AaronStGeorge · 2025-10-28T21:05:45Z

@marbre @ScottTodd thanks for the review!

Most of the blocking comments seem to be on temporary pieces intended to keep previous (non-TheRock-based) CI during the transition into TheRock. I think the simplest path forward is simply to go without CI while in a transitionary state. Effort that can be put into fixing intermediate state could be better allocated towards getting to the desired end state where fusilli-plugin is part of TheRock build and enabled through TheRock CI.

sjain-stanford · 2025-10-29T17:49:47Z

@AaronStGeorge let's drop the intermediate CI entirely and skip to working towards the end state. No point trying to get to a better intermediate state that is transient / throw-away anyway. Please remove the CI bits from here and let's try to land this soon.

AaronStGeorge · 2025-11-19T22:22:05Z

RFC0004-Fusilli-IREE-Kernel-Provider-hipDNN decided that fusilli plugin will be hosted in fusilli repo rather than rocm-libraries.

andrei-kochin · 2025-11-27T06:05:40Z

Fusilli removal from rocm-libraries will happen in #2777

github-actions Bot added github actions project: none Does not target any component labels Oct 24, 2025

AaronStGeorge force-pushed the fusilli-plugin-to-the-rock-2-of-n branch from c65442e to 376f0dc Compare October 24, 2025 18:16

assistant-librarian Bot added the organization: ROCm label Oct 24, 2025

AaronStGeorge force-pushed the fusilli-plugin-to-the-rock-2-of-n branch 2 times, most recently from 6228026 to 2a1f4cb Compare October 25, 2025 00:08

AaronStGeorge closed this Oct 25, 2025

AaronStGeorge deleted the fusilli-plugin-to-the-rock-2-of-n branch October 25, 2025 00:10

AaronStGeorge restored the fusilli-plugin-to-the-rock-2-of-n branch October 25, 2025 00:11

AaronStGeorge reopened this Oct 25, 2025

AaronStGeorge force-pushed the fusilli-plugin-to-the-rock-2-of-n branch 2 times, most recently from a2561ba to 301a775 Compare October 25, 2025 03:05

AaronStGeorge changed the title ~~[DO NOT MERGE] CI testing for fusilli-plugin~~ [FusilliPlugin] Re-Enable build and CI, add CODEOWNERS etc. Oct 25, 2025

AaronStGeorge changed the title ~~[FusilliPlugin] Re-Enable build and CI, add CODEOWNERS etc.~~ [FusilliPlugin] Re-Enable build and CI and add CODEOWNERS etc. Oct 25, 2025

AaronStGeorge force-pushed the fusilli-plugin-to-the-rock-2-of-n branch 2 times, most recently from 272a3b7 to 470079d Compare October 25, 2025 03:23

AaronStGeorge marked this pull request as ready for review October 25, 2025 03:24

AaronStGeorge requested review from a team as code owners October 25, 2025 03:24

AaronStGeorge requested a review from jayhawk-commits October 25, 2025 03:26

sjain-stanford reviewed Oct 25, 2025

View reviewed changes

Comment thread projects/fusilli-plugin/build_tools/docker/exec_docker_ci.sh Outdated

sjain-stanford added project: fusilli-provider and removed project: none Does not target any component labels Oct 25, 2025

jayhawk-commits approved these changes Oct 27, 2025

View reviewed changes

sjain-stanford reviewed Oct 27, 2025

View reviewed changes

sjain-stanford approved these changes Oct 27, 2025

View reviewed changes

ScottTodd requested changes Oct 28, 2025

View reviewed changes

marbre requested changes Oct 28, 2025

View reviewed changes

amd-jnovotny approved these changes Oct 28, 2025

View reviewed changes

AaronStGeorge added 6 commits November 16, 2025 03:12

Fix build, add CI

5819ea8

Address comment

80103e2

Use find_pakage

c0cae93

Fixes

15f66d6

Add new API method, update cmake

d804165

Debug stuff that should be removed

a4cc9b1

AaronStGeorge force-pushed the fusilli-plugin-to-the-rock-2-of-n branch from 393764f to a4cc9b1 Compare November 16, 2025 03:38

Fixes for new hipDNN APIs

11f0df4

AaronStGeorge closed this Nov 19, 2025

Conversation

AaronStGeorge commented Oct 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

Uh oh!

Uh oh!

jayhawk-commits left a comment

Choose a reason for hiding this comment

Uh oh!

jayhawk-commits commented Oct 27, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sjain-stanford left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

AaronStGeorge Oct 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

marbre left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

amd-jnovotny commented Oct 28, 2025

Uh oh!

AaronStGeorge commented Oct 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

sjain-stanford commented Oct 29, 2025

Uh oh!

AaronStGeorge commented Nov 19, 2025

Uh oh!

andrei-kochin commented Nov 27, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

AaronStGeorge commented Oct 24, 2025 •

edited

Loading

AaronStGeorge Oct 28, 2025 •

edited

Loading

AaronStGeorge commented Oct 28, 2025 •

edited

Loading