feat(rocke): switch conv kernels to AOT by bartekxk · Pull Request #12783 · ROCm/rocm-libraries

bartekxk · 2026-09-30T11:18:50Z

Motivation

Make conv kernels more generic.

Technical Details

Switch from JIT compilation to AOT

Test Plan

run_all.py

Test Result

pass

Submission Checklist

Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.

JIRA ID: AICK-2366

therock-pr-bot · 2026-09-30T11:23:31Z

✅ 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?

Copilot

Copilot review overview

🟡 Changes recommended

Unresolved ABI construction, manifest packing, and split-K pipeline defects can prevent launches or produce incorrect results.

Review effort: Balanced
Findings: 5 · 1

Open (6)

Fix incorrect dtype argument in two-stage wgrad signature · New
Select direct ABI arguments for direct manifests · New
Recognize direct convolution layout variants · New
Accept documented 3-D convolution manifest fields · New
Mask split-K loads against supplied reduction extent · New
Enforce max_sub_gemms during compatibility filtering · New

What changed in this PR

Migrates rocKE convolution kernels from shape-specialized JIT compilation to shape-generic AOT kernels using runtime kernargs.

Changes:

Adds dynamic transforms, descriptors, pipelines, and mirrored C++ emitters.
Centralizes convolution ABI, manifests, caching, dispatch, and launch packing.
Updates convolution correctness and cache tests.

File	Description
`platform/python/rocke/portable_ir/drivers/roll_nd_coverage.py`	Removes obsolete JIT sampling probe.
`platform/python/rocke/helpers/transforms.py`	Adds dynamic transforms and descriptors.
`platform/python/rocke/helpers/pipeline.py`	Adds runtime-extent ping-pong pipeline.
`platform/python/rocke/helpers/manifest.py`	Emits AOT convolution signatures.
`platform/python/rocke/helpers/conv_abi.py`	Defines centralized convolution ABI.
`platform/python/rocke/examples/gfx950/conv_dgrad/run_one_dgrad.py`	Uses AOT dgrad arguments.
`platform/cpp/instances/common/conv_implicit_gemm_dgrad.cpp`	Converts dgrad emission to AOT.
`platform/cpp/instances/common/conv_implicit_gemm_conv_epilogue.cpp`	Uses runtime epilogue descriptors and bounds.
`platform/cpp/instances/common/conv_implicit_gemm_conv_build_glue.cpp`	Declares and wires runtime parameters.
`platform/cpp/instances/common/conv_direct_grouped_spec_and_validate.cpp`	Adds direct-convolution validation.
`platform/cpp/instances/common/conv_direct_grouped_common.cpp`	Adds shared direct AOT plumbing.
`platform/cpp/instances/common/conv_direct_grouped_build_dgrad.cpp`	Converts direct dgrad to AOT.
`platform/cpp/instances/common/conv_direct_grouped_build_depthwise.cpp`	Converts depthwise forward to AOT.
`platform/cpp/instances/common/conv_direct_grouped_build_depthwise_spatial.cpp`	Converts spatial depthwise loops to AOT.
`platform/cpp/instances/common/conv_direct_grouped_build_4c_phases_and_loop.cpp`	Adds runtime 4c row loops.
`platform/cpp/instances/common/conv_direct_grouped_build_16c_phases.cpp`	Adds runtime 16c descriptors.
`platform/cpp/include/rocke/instance_conv_implicit_gemm.h`	Declares dynamic descriptor APIs.
`platform/cpp/include/rocke/instance_conv_implicit_gemm_internal.h`	Extends build context with AOT values.
`platform/cpp/include/rocke/instance_conv_implicit_gemm_dgrad.h`	Adds dgrad search capability bound.
`platform/cpp/include/rocke/instance_conv_direct_grouped_internal.h`	Defines direct AOT parameter context.
`platform/cpp/include/rocke/helper_rocke.helpers.transforms.h`	Declares dynamic transform infrastructure.
`library/tests/test_direct_kernel_cache.py`	Tests direct AOT cache behavior.
`library/tests/test_direct_conv_correctness.py`	Updates direct launch correctness tests.
`library/tests/test_conv_wgrad_correctness.py`	Updates wgrad AOT launches.
`library/tests/test_conv_fwd_correctness.py`	Updates forward AOT launches.
`library/tests/test_conv_dgrad_correctness.py`	Updates dgrad AOT launches.
`library/kernels/common/manifest_runner/conv.py`	Packs runtime convolution arguments.
`library/kernels/common/conv_implicit_gemm_wgrad_two_stage.py`	Adds two-stage AOT launch helpers.
`library/dispatch/grouped_convolution.py`	Returns full AOT signatures and values.
`library/benchmarks/common/kernel_cache.py`	Adds AOT kernel identity and cache.
`library/benchmarks/common/grouped_conv/benchmark_grouped_conv.py`	Updates grouped benchmarks for AOT.

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

bartekxk · 2026-09-30T11:37:44Z

Review: #12783 — feat(rocke): switch conv kernels to AOT

Reviewed commit: ac0b2f517c3

I found five bugs: two break things outright, and three produce wrong results or wrong launches. I verified each one by reading the code, without running anything on a GPU. For the first one, pyflakes on the changed files also reports the undefined name.

1. Grouped-conv benchmark crashes with `NameError`

Location: library/benchmarks/common/grouped_conv/benchmark_grouped_conv.py:1001 (also lines 1261–1271 and 1318–1328)

_run_fwd and _run_wgrad read spec.tile_m, spec.tile_n, spec.tile_k and spec.split_k, but neither function has a variable named spec. The loop variables are dspec / instance_spec, so every fwd and wgrad candidate raises NameError.

Fix: use the correct loop variable.

2. Wgrad `unroll_k` loop runs the MFMA on the wrong buffer

Location: library/kernels/common/conv_implicit_gemm_wgrad.py:2329

Each iteration loads the next tile into the second buffer, but the MFMA always reads the first buffer, which only ever holds the first tile. With more than one K tile, dW ends up as the first tile's contribution multiplied by the tile count. The previous version alternated the buffers correctly.

Fix: switch the compute buffer on each iteration, as the previous version did.

3. Split-K wgrad can add a tile to dW twice

Location: platform/python/rocke/helpers/pipeline.py:382 (run_ping_pong_dynamic)

Each step of the loop computes a second tile and assumes anything past the loop bound reads as zero. That is only true at the real end of the tensor. The async wgrad path passes the end of its split-K slice as the bound. With an odd number of tiles per slice, it therefore also computes the first tile of the next slice.

Example: 5 K tiles split 2 ways gives 3 tiles per slice. Slice 0 also computes tile 3, which slice 1 computes too, so tile 3 is added to dW twice.

Fix: guard the second (tail) tile against the slice bound, not only the tensor bound.

4. Dispatcher launch values and signature disagree with the kernel and grid

Location: library/dispatch/grouped_convolution.py:1747 (also :1700)

launch_values_for and _signature_for read the raw spec.split_k and getattr(spec, "two_stage", False).
ConvGroupedSpec has no two_stage field, so that value is always False.
The kernel and the grid instead use _resolve_wgrad_split_k, which resolves auto split-K, clamps it, and can switch the kernel to two-stage.

Consequences:

Two-stage kernel: the signature leaves out ws_ptr/ws_bytes, so ks/ks_count get packed into the workspace-pointer slot.
Auto split-K resolving to, e.g., 4: the grid launches groups × 4 blocks in z, but the kernel is told there is 1 slice, so it decodes group and slice wrongly.

Fix: compute launch values and the signature from _resolve_wgrad_split_k, the same helper the kernel and grid use.

5. Grouped dgrad from the kernel cache only computes group 0

Location: library/benchmarks/common/benchmark_implicit_gemm_conv.py:3689

The launch grid is (blocks, 1, split_k), but the dgrad kernel reads the conv group from blockIdx.y. The dispatcher correctly puts the group count in grid y; this path does not. The cache sweep does build grouped dgrad kernels, so this path is reachable, and the benchmark still reports timings for the incomplete result.

Fix: launch with (blocks, groups, split_k).

Note on byte-identity

Bugs 2 and 3 are in kernel emission. The fixes have to be mirrored in the C++ engine and the golden re-blessed in the same change. Then run tools/check_byte_identity.py for both llvm20 and llvm22.

therock-pr-bot · 2026-09-30T11:38:43Z

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

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Copilot

Copilot review overview

🟡 Changes recommended

Odd-tail pipeline state and incomplete workspace validation can produce incorrect runtime results.

Review effort: Balanced
Findings: 2 · 1

Open (3)

Validate workspace pointer and size fields together · New
Mask Phase B state updates for out-of-range tiles · New
Reject buffer sequences with unsupported lengths · New

Resolved since last review (6)

Mask split-K loads against supplied reduction extent
Accept documented 3-D convolution manifest fields
Recognize direct convolution layout variants
Select direct ABI arguments for direct manifests
Fix incorrect dtype argument in two-stage wgrad signature
Enforce max_sub_gemms during compatibility filtering

aledudek

Could you expand the test plan? This changes the ABI and IR of every conv kernel, could you please list:

check_byte_identity.py at llvm20 and llvm22
a run with ROCKE_CPP_STRICT=1 (or the engine build_id), so we know the C++ engine actually ran
the on-device conv correctness tests (fwd/wgrad/dgrad/direct) and which GPU they ran on, with platform/tests and library/tests run separately

Copilot

Copilot review overview

🟡 Changes recommended

The dynamic magic transform lacks full Python/C++ API parity, and several documented launch workflows are currently invalid.

Review effort: Balanced
Findings: 1 · 1

Open (2)

Add mixed constant/value support and dim-1 parity · New
Fix module example for new AOT signature · New

Resolved since last review (3)

Mask Phase B state updates for out-of-range tiles
Validate workspace pointer and size fields together
Reject buffer sequences with unsupported lengths

aledudek

Looks good besides the group merge problem, left a few minor comments, good to merge after resolving that

…arkocot/conv-aot

…ol-cpp-port Adapts the column-streamed depthwise kernel (DirectDepthwiseColSpec) to the AOT conv ABI that develop switched the direct-conv kernels to (#12783). The dwcol kernel unrolled its input rows over the whole image, keeping a Ho x block_w accumulator band, so its trip count baked the image height -- exactly what AOT forbids. Following the DirectConvSpec precedent, the band is now a build-time row tile: - new block_h knob (default 16): each block covers block_h output rows; block_id_z encodes n * ceil(Ho / block_h) + h_tile, with the tile count derived from the p_Ho kernarg; stores are bounded against p_Ho. - the kernel declares the direct-conv kernarg block (emit_direct_params / rocke_dconv_emit_params) and uses the dynamic A/D descriptors, so N, H, W, groups and the tensor strides are runtime. - live_f32 is now block_h * block_w + KH; n_iters is (block_h - 1) * stride + KH. - ch_tile_exact / w_tile_exact guard elision is removed: the group count and Wo are kernargs, so the guards are always emitted. - direct_launch_geometry handles DirectDepthwiseColSpec. Both engines are mirrored; byte-identity is GREEN for every family at llvm20 and llvm22. The dwcol representative-IR goldens are re-blessed (only the dwcol entries moved). AOT cache: direct_depthwise_col is a new forward variant in DIRECT_CAPABILITIES. It is padding-generic (rows are bounded by p_Ho), so it takes every PAD in [0, KH-1] except stride-1 over-padding. The branch's per-direction depthwise block_w grids (DW_BLOCK_W_FWD / DW_BLOCK_W_DGRAD) and the col block_h/block_w grids move into direct_kernel_sweep, which develop made the single source for the JIT sweep and the cache. Conflict resolution: the preload depthwise builders take develop's loop-only form, which supersedes this branch's unroll threshold of 0. Tests: the dwcol GPU correctness suite gains row-tile shapes and a test that runs one compiled binary across several image shapes; the shape- invariance test covers dwcol; the cache tests cover the new variant; the preload tap-ceiling boundary test is re-pinned, since forward depthwise now only accepts odd filters with "same" padding. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ke-conv-bwd-merged-groups-heuristic Adapts the depthwise wgrad merged-groups heuristic to the AOT switch (#12783). Conflict was in test_grouped_conv_wgrad_dispatch.py, where both sides appended a test class; both are kept. AOT's TestLaunchContractMatchesKernel asserted the pre-merge invariant gridDim.z == groups * split_k, which a merged launch breaks (z is groups/Gm * split_k). It now asks WgradConvSpec.grid_groups, the same property _wgrad_grid divides by, so it reduces to the old assertion at Gm == 1. grouped_convolution.py merged cleanly and is correct as merged: both AOT entry points go through _resolve_wgrad_split_k, so a merged split_k > 1 resolves to two-stage and the ws_ptr/ws_bytes pair joins the kernarg ABI on its own. The launch values stay built from the true (unmerged) problem, which the kernel expects -- it scales by Gm in-kernel -- and p_num_pid_m/n still match the merged grid because the merge gate admits only single-tile merges. That invariant is now documented and pinned by a test rather than left implicit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ke-conv-fwd-merged-groups Brings in the switch of conv kernels from JIT to AOT compilation (#12783) and adapts the depthwise forward group-merge feature to it. Under AOT the problem shape is passed as kernel arguments, so a binary is shared by every compatible shape. Following the wgrad merge precedent, the kernargs keep describing the TRUE problem and the merged geometry is folded in-kernel from the build-time degree: merge is depthwise-only, so the merged cpg/kpg are exactly Gm (constants), and only the merged reduction K_gemm*Gm is a runtime product. - Builder (Python and C++, byte-identical): merged k decode in the split address form (shift/mask for the merged group, magic division for X), merged B load with the diagonal mask, merged K-loop extent for the plain, unroll_k and async drivers, merged group base and N bound in the epilogue. The grouped decode stays engaged at Gm == groups because the binary serves every group count Gm divides. 3-D is now rejected by the merge gate. - Default load/store widths come from the merged problem in both the builder and the validator (_sync_load_vecs, the cshuffle default). - Python's open-width cshuffle store now uses the kpg-derived width, as the C++ engine and is_valid_spec already did; the fixed 8 dropped every output when kpg was not a multiple of 8 (e.g. unmerged depthwise). - AOT cache: the sweep builds merged depthwise fwd kernels (cshuffle, plain and unroll_k); the cache serves them to depthwise problems Gm divides with Gm <= tile_n; run-from-cache launches groups/Gm along z. - Parity configs renumbered after develop's new ones (merged now 20-25, plus a merged unroll_k case); golden IR hashes re-blessed for the merged cases. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

assistant-librarian · 2026-10-07T12:35:25Z

TheRock Submodule Bump Activity

Newest first

2026-10-07 — Included in TheRock via ROCm/TheRock#8777.

## Motivation After #12783 (JIT → AOT), many conv shapes got slower. In some cases the AOT cache had no kernel at all. The causes: 1. The cache left out kernels the JIT sweep used to pick. 2. The shape-generic kernels compute index math with runtime divides inside the K loop. ## Technical Details **Cache coverage** - **Direct-conv wgrad** is now cached. `DirectConvWgradSpec.runtime_channels` makes one binary serve any channel count. - **Direct MFMA dgrad for dense problems** is now cached, through runtime-channel modes of `DirectConvSpec` and its transpose/reorganize helpers. The C++ engine gets ports of these. - **Implicit-GEMM wgrad and dgrad** entries now use the K-outer LDS layout, the same as the JIT sweep picks. - **Non-split wgrad** (split_k=1) is cached, and split-K degrees 256/512 are tried at launch. - **13x13 and 31x31 depthwise forward** filters are cached. **New kernel loop options** (off by default, so existing IR is unchanged) - `WgradConvSpec.prefetch_addr`: computes the next tile's addresses while the current loads are in flight. - `DgradConvSpec.uniform_k_decode`: moves the per-element divides out of the K loop. - Both are turned on by deterministic, per-geometry selection rules. **Benchmark tooling** - Fixed a device-memory leak in `--run-from-cache` (two-stage workspaces and dgrad sub-GEMM tables). - `--split-k-prune` now works on the cache path. - Early stop drops hopeless kernels after a single launch. - The compare script no longer skips direct conv for wgrad. The follow-up commit addresses self-review findings: C++ bounds and validation, dgrad cache coverage, early-stop robustness, and extra tests. Cache size grows about 16% per dtype. Most of it is the non-split wgrad entries. ## Test Plan - `check_byte_identity.py`: llvm20 and llvm22. New parity configs cover each new mode. - `platform/tests/run_all.py` with ctest. - CPU: `test_conv_abi`, `test_direct_kernel_cache`, `test_early_stop`, `test_conv_wgrad_prefetch_addr`, `test_conv_dgrad_uniform_k_decode`, layering/signature tests. - GPU (gfx950): `test_conv_wgrad_correctness`, `test_conv_dgrad_correctness`, `test_direct_conv_correctness`, `test_conv_fwd_correctness`. - `--run-from-cache --verify` on sampled wgrad, dgrad and depthwise shapes. ## Test Result - Byte identity is GREEN for all 73 families on both LLVM flavors. - ctest passes 14/14. - All the listed tests pass. - Every new kernel mode passes `--verify` against the torch reference. - `run_all.py` reports two failures that also occur on the base commit, unrelated to this change: the relative-path guard on `platform/CMakeLists.txt`, and `test_parity_matrix` for `tf32` on gfx942. **Performance** (relative only). How it was measured: - **Setup:** gfx950, NHWC, bf16 and fp16. Shapes are 297 forward, wgrad and dgrad problems from three model shape lists. - **Baseline:** best JIT result before #12783. - **New side:** best kernel from this branch's AOT cache, measured with `--run-from-cache` at 3 warmup / 10 iterations. - **Correctness:** the new kernel modes passed `--verify` against a reference. The other cached kernels were not re-verified in the timing runs. Results: - wgrad shapes went from typically 2–3× slower than the JIT baseline (worst case about 6×) to parity on most measured configurations. - dgrad shapes that were 1.1–1.3× slower are about level with the baseline. - The shape-generic kernel loops alone got 1.05–1.3× faster on the measured configurations. Caveats: - At this iteration count some rows varied by up to 1.6× between runs. - A few remaining gaps are outside this PR: - Direct depthwise kernels are slower in the JIT path too, which is being addressed separately. - dgrad with stride > 2, and direct dgrad with an odd number of 16-channel dY tiles (80, 112, …), are not yet cached. JIRA-ID : AICK-2519 --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

## Motivation After ROCm#12783 (JIT → AOT), many conv shapes got slower. In some cases the AOT cache had no kernel at all. The causes: 1. The cache left out kernels the JIT sweep used to pick. 2. The shape-generic kernels compute index math with runtime divides inside the K loop. ## Technical Details **Cache coverage** - **Direct-conv wgrad** is now cached. `DirectConvWgradSpec.runtime_channels` makes one binary serve any channel count. - **Direct MFMA dgrad for dense problems** is now cached, through runtime-channel modes of `DirectConvSpec` and its transpose/reorganize helpers. The C++ engine gets ports of these. - **Implicit-GEMM wgrad and dgrad** entries now use the K-outer LDS layout, the same as the JIT sweep picks. - **Non-split wgrad** (split_k=1) is cached, and split-K degrees 256/512 are tried at launch. - **13x13 and 31x31 depthwise forward** filters are cached. **New kernel loop options** (off by default, so existing IR is unchanged) - `WgradConvSpec.prefetch_addr`: computes the next tile's addresses while the current loads are in flight. - `DgradConvSpec.uniform_k_decode`: moves the per-element divides out of the K loop. - Both are turned on by deterministic, per-geometry selection rules. **Benchmark tooling** - Fixed a device-memory leak in `--run-from-cache` (two-stage workspaces and dgrad sub-GEMM tables). - `--split-k-prune` now works on the cache path. - Early stop drops hopeless kernels after a single launch. - The compare script no longer skips direct conv for wgrad. The follow-up commit addresses self-review findings: C++ bounds and validation, dgrad cache coverage, early-stop robustness, and extra tests. Cache size grows about 16% per dtype. Most of it is the non-split wgrad entries. ## Test Plan - `check_byte_identity.py`: llvm20 and llvm22. New parity configs cover each new mode. - `platform/tests/run_all.py` with ctest. - CPU: `test_conv_abi`, `test_direct_kernel_cache`, `test_early_stop`, `test_conv_wgrad_prefetch_addr`, `test_conv_dgrad_uniform_k_decode`, layering/signature tests. - GPU (gfx950): `test_conv_wgrad_correctness`, `test_conv_dgrad_correctness`, `test_direct_conv_correctness`, `test_conv_fwd_correctness`. - `--run-from-cache --verify` on sampled wgrad, dgrad and depthwise shapes. ## Test Result - Byte identity is GREEN for all 73 families on both LLVM flavors. - ctest passes 14/14. - All the listed tests pass. - Every new kernel mode passes `--verify` against the torch reference. - `run_all.py` reports two failures that also occur on the base commit, unrelated to this change: the relative-path guard on `platform/CMakeLists.txt`, and `test_parity_matrix` for `tf32` on gfx942. **Performance** (relative only). How it was measured: - **Setup:** gfx950, NHWC, bf16 and fp16. Shapes are 297 forward, wgrad and dgrad problems from three model shape lists. - **Baseline:** best JIT result before ROCm#12783. - **New side:** best kernel from this branch's AOT cache, measured with `--run-from-cache` at 3 warmup / 10 iterations. - **Correctness:** the new kernel modes passed `--verify` against a reference. The other cached kernels were not re-verified in the timing runs. Results: - wgrad shapes went from typically 2–3× slower than the JIT baseline (worst case about 6×) to parity on most measured configurations. - dgrad shapes that were 1.1–1.3× slower are about level with the baseline. - The shape-generic kernel loops alone got 1.05–1.3× faster on the measured configurations. Caveats: - At this iteration count some rows varied by up to 1.6× between runs. - A few remaining gaps are outside this PR: - Direct depthwise kernels are slower in the JIT path too, which is being addressed separately. - dgrad with stride > 2, and direct dgrad with an odd number of 16-channel dY tiles (80, 112, …), are not yet cached. JIRA-ID : AICK-2519 --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

bartekxk requested review from a team as code owners September 30, 2026 11:18

bartekxk requested a balanced review from Copilot September 30, 2026 11:19

github-actions Bot added project: hip-kernel-provider rocke: platform Touches rocKE platform labels Sep 30, 2026

Copilot started reviewing on behalf of bartekxk September 30, 2026 11:19

View session

Copilot AI reviewed Sep 30, 2026

View reviewed changes

assistant-librarian Bot added the organization: ROCm label Sep 30, 2026

aledudek requested changes Sep 30, 2026

View reviewed changes

github-actions Bot added the documentation label Oct 1, 2026

feat(rocke): switch conv kernels to AOT

f4cb7e1

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

bartekxk force-pushed the users/barkocot/conv-aot branch from 1419d3a to f4cb7e1 Compare October 1, 2026 11:52

bartekxk requested a balanced review from Copilot October 1, 2026 11:53

Copilot started reviewing on behalf of bartekxk October 1, 2026 11:54

View session

Copilot AI reviewed Oct 1, 2026

View reviewed changes

aledudek requested changes Oct 1, 2026

View reviewed changes

bartekxk added 2 commits October 1, 2026 23:00

fixes

0d323d8

review fixes

999cdd3

bartekxk requested a balanced review from Copilot October 2, 2026 09:43

Copilot started reviewing on behalf of bartekxk October 2, 2026 09:44

View session

Copilot AI reviewed Oct 2, 2026

View reviewed changes

Comment thread ...iders/hip-kernel-provider/rocke/platform/cpp/include/rocke/helper_rocke.helpers.transforms.h Resolved

Comment thread ...iders/hip-kernel-provider/rocke/library/kernels/common/conv_implicit_gemm_wgrad_two_stage.py Resolved

aledudek reviewed Oct 2, 2026

View reviewed changes

aledudek approved these changes Oct 2, 2026

View reviewed changes

bartekxk added 2 commits October 4, 2026 22:19

fxes

3223fe6

fxes

0065472

bartekxk added 2 commits October 5, 2026 09:28

Merge branch 'develop' of github.com:ROCm/rocm-libraries into users/b…

7716eb2

…arkocot/conv-aot

fxes

2e84e47

github-actions Bot added the rocke-lib: attention Touches rocKE library attention/FMHA kernels label Oct 5, 2026

bartekxk enabled auto-merge (squash) October 5, 2026 12:45

bartekxk merged commit 3c7094a into develop Oct 5, 2026
96 checks passed

bartekxk deleted the users/barkocot/conv-aot branch October 5, 2026 15:03

aledudek mentioned this pull request Oct 6, 2026

feat(hip-kernel-provider): runtime-dispatch packaged gfx950 rocKE conv forward through hipDNN #12780

Closed

1 task done

jakpiase mentioned this pull request Oct 8, 2026

fix(rocke): close conv perf gaps after the switch to AOT #13320

Merged

cgmillette mentioned this pull request Oct 9, 2026

feat(rocke): prototype device-side printing #10841

Draft

9 tasks done

cgmillette mentioned this pull request Oct 29, 2026

[rocke] test_conv_abi.py::test_cache_applies_runtime_launch_limits fails on develop since #13782 (kernel admission still refuses the depthwise dgrad that #13782 enabled) #14643

Open

Conversation

bartekxk commented Sep 30, 2026

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

Uh oh!

therock-pr-bot Bot commented Sep 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✅ All Checks Passed — Ready for Review

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

bartekxk commented Sep 30, 2026

Review: #12783 — feat(rocke): switch conv kernels to AOT

1. Grouped-conv benchmark crashes with NameError

2. Wgrad unroll_k loop runs the MFMA on the wrong buffer

3. Split-K wgrad can add a tile to dW twice

4. Dispatcher launch values and signature disagree with the kernel and grid

5. Grouped dgrad from the kernel cache only computes group 0

Note on byte-identity

Uh oh!

therock-pr-bot Bot commented Sep 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Uh oh!

aledudek left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

aledudek left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

assistant-librarian Bot commented Oct 7, 2026

therock-pr-bot Bot commented Sep 30, 2026 •

edited

Loading

1. Grouped-conv benchmark crashes with `NameError`

2. Wgrad `unroll_k` loop runs the MFMA on the wrong buffer

therock-pr-bot Bot commented Sep 30, 2026 •

edited

Loading