feat(rocke): grouped conv dispatcher

[bartekxk]

Motivation

Add possibility to use dispatcher for grouped conv and bench them

Technical Details

-dispatcher
-benchmark

Test Plan

CI

Test Result

Pass

Submission Checklist

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

JIRA ID: AICK-1750

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

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

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[Copilot]

Pull request overview

Adds a new grouped-convolution dispatcher in rocke (forward + wgrad routing) and a corresponding benchmark harness + JSON case suites to exercise/compare dispatcher-selected kernels across common model shapes.

Changes:

• Introduces dispatch/grouped_convolution.py with request/spec types, candidate registries, and dispatch entrypoints for grouped conv fwd + wgrad.
• Adds benchmark_grouped_conv.py to compile/time dispatcher sweep-space candidates (with optional verification + ckProfiler comparison).
• Adds forward/backward JSON case files to drive benchmark runs.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 5 comments.

File	Description
dnn-providers/hip-kernel-provider/rocke/library/dispatch/grouped_convolution.py	New dispatcher and candidate registry for grouped conv fwd/wgrad.
dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/grouped_conv/benchmark_grouped_conv.py	New benchmark CLI for dispatcher-driven grouped conv timing/verification.
dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/grouped_conv/bench_cases_conv.json	Forward benchmark case suite for grouped conv benchmarking.
dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/grouped_conv/bench_cases_conv_bwd.json	Backward (wgrad-focused) benchmark case suite.

Suppressed comments (1)

dnn-providers/hip-kernel-provider/rocke/library/dispatch/grouped_convolution.py:948

• Same split_k=-1 issue as the gfx942 wgrad candidate: selection returns split_k=-1 and the dispatcher uses it in the launch grid/spec hash. Resolve split_k=-1 at select time so the returned spec/grid are concrete.

    def select(req: OperatorRequest) -> ConvGroupedSpec:
        ok, why = candidate.admits(req)
        if not ok:
            raise ValueError(f"{name} does not support request: {why}")
        assert isinstance(req, ConvGroupedRequest)

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated no new comments.

Suppressed comments (10)

dnn-providers/hip-kernel-provider/rocke/library/dispatch/grouped_convolution.py:353

• _request_errors() calls _problem(req) unguarded; ConvProblem.post_init raises ValueError for invalid group divisibility or partial 3D fields, which would crash candidate.supported()/admits() instead of returning a clean rejection message.

    p = _problem(req)
    if p.Ho <= 0 or p.Wo <= 0:

dnn-providers/hip-kernel-provider/rocke/library/dispatch/grouped_convolution.py:1005

• This docstring also claims split-K is "forwarded from request", but there is no split_k field on ConvGroupedRequest and select_spec currently always uses split_k=-1. Keeping docs aligned avoids confusion for callers trying to force/disable split-K.

    """Backward-weight conv for gfx950: 64×64×64, 2×2, 32×32×16 MFMA.

    Epilogue derived from vec_size_c (cshuffle when >1, default otherwise).
    Split-K forwarded from request (1=disabled, -1=auto CK formula, >1=fixed).
    gfx1250 wgrad is not yet supported; use CDNA only.

dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/grouped_conv/benchmark_grouped_conv.py:258

• parse_json_case() treats pads_before as symmetric padding and ignores pads_after. This silently mis-parses asymmetric padding (e.g. causal conv3d cases with pads_before != pads_after in the provided JSON), producing an incorrect ConvProblem.

        pads_before = entry.get("pads_before", entry.get("pad", 0))
        if isinstance(pads_before, (list, tuple)):
            pD, pH, pW = int(pads_before[0]), int(pads_before[1]), int(pads_before[2])
        else:
            pD = pH = pW = int(pads_before)

dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/grouped_conv/benchmark_grouped_conv.py:317

• The conv2d pads_before/pads_after path also ignores pads_after. If a JSON case provides asymmetric pads, the benchmark will run a different problem than specified. Since ConvProblem only supports symmetric padding, it should reject cases where pads_after != pads_before.

    # Support both symmetric pad and pads_before/pads_after
    if "pads_before" in entry:
        pads_before = entry["pads_before"]
        pH = (
            int(pads_before[0])
            if isinstance(pads_before, (list, tuple))
            else int(pads_before)
        )
        pW = (
            int(pads_before[1])
            if isinstance(pads_before, (list, tuple))
            else int(pads_before)
        )
    else:

dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/grouped_conv/benchmark_grouped_conv.py:540

• The dispatcher (dispatch/grouped_convolution.py) rejects dtype=fp32, but this benchmark exposes --dtype fp32. That will deterministically produce "No dispatcher candidates" failures; better to remove fp32 here or explicitly skip/report it as unsupported by the grouped-conv dispatcher.

        "--dtype",
        default="fp16",
        choices=["fp16", "bf16", "fp32"],
        help="data type (default: fp16)",
    )

dnn-providers/hip-kernel-provider/rocke/library/dispatch/grouped_convolution.py:1196

• conv_grouped_sweep_space() deduplicates by spec.kernel_name(), but wgrad select_spec currently returns split_k=-1, so the kernel_name becomes "spk-1" and does not match the actually built kernel (which resolves split_k from the problem). This can also collapse distinct configs and makes benchmark output misleading.

    for candidate in registry.supported(req):
        spec = candidate.select_spec(req)
        h = spec.kernel_name()
        if h not in seen:

dnn-providers/hip-kernel-provider/rocke/library/dispatch/grouped_convolution.py:1210

• dispatch_conv_grouped() returns a wgrad spec with split_k=-1, but grid() resolves split_k on the fly. That means DispatchResult.spec / KernelId.spec_hash may not describe the actual kernel specialization used, which can break compile-key based caches and makes logs non-reproducible.

    registry = _registry_for(req)
    candidate = registry.select(req, ranker=ranker)
    spec = candidate.select_spec(req)
    kid = _kernel_id(req, candidate, spec)
    return DispatchResult(

dnn-providers/hip-kernel-provider/rocke/library/dispatch/grouped_convolution.py:897

• The docstring says "Split-K forwarded from request", but ConvGroupedRequest has no split_k field and this candidate hard-codes split_k=-1. Either add a request field or update the docstring to describe the current auto-resolution behavior.

This issue also appears on line 1001 of the same file.

    """Backward-weight conv for gfx942: 64×64×64, 2×2, 16×16×16 MFMA.

    Epilogue derived from vec_size_c (cshuffle when >1, default otherwise).
    Split-K forwarded from request (1=disabled, -1=auto CK formula, >1=fixed).
    """

dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/grouped_conv/benchmark_grouped_conv.py:202

• parse_json_case() currently accepts fp32 cases, but the grouped-conv dispatcher only supports fp16/bf16. With the provided JSON files containing fp32 entries, this leads to benchmark runs that always fail at dispatch time instead of being skipped with a clear warning.

This issue also appears in the following locations of the same file:

• line 254
• line 304

    dtype = entry.get("dtype", "fp16")
    if dtype not in ("fp16", "bf16", "fp32"):
        raise ValueError(f"dtype={dtype!r} is not supported (only fp16, bf16, fp32)")

dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/grouped_conv/benchmark_grouped_conv.py:1079

• The help text for --dump-fail says it implies --verify, and the code does run verification when dump_fail is set. However the summary table only shows the verify column when args.verify is true, so a --dump-fail run hides PASS/FAIL in the final results table.

    show_verify = args.verify
    width = 84 if show_verify else 72
    print(f"\n{'='*width}")

[yraparti]

can you move this to platform/dispatch?