Two new template parameters are introduced:
* CShuffleDataType allows to apply multiple Ds before downcasting to
ODataType (prevents unexpected precision loss and/or overflow);
* CComputeDataType to use scales with int32 AccDataType (int8 gemms);
[CK][CI] Retry builds on node failure with automatic
rerouting (#8332)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
When a Jenkins node enters a bad state (missing GPU driver, dead Docker
daemon, full disk), every PR scheduled onto it fails the same way until
a human manually takes it offline. Some failures are also transient and
would pass on a simple retry. Today the pipeline does neither — every
failure goes straight to red on the same node.
## Technical Details
Two new retry behaviors based on failure type:
- **Different node** for persistent node faults (driver missing, daemon
down, disk full, container won't start)
- **Retry in place** for transient glitches (registry pull, DNS), then a
different node if retries are exhausted
Real build/compile failures and aborted builds are never retried.
**New:** `src/org/ck/NodeFault.groovy`, `TransientFault.groovy` — typed
exceptions in the shared library `src/` for stable classloader identity
under dynamic library loading.
**`vars/ck.groovy`:** adds `preflight()` (host health checks before
build), `pullImage()` (classifying pull failures at the call site,
replacing `getDockerImage()`), `runOnHealthyNode()` (outer reroute loop,
up to 3 nodes), `runInPlace()` (same-node transient retries). GitHub
failure status is only set once all retries are exhausted.
**`Jenkinsfile`:** all active `Build CK and run Tests` stages converted
to `agent none` + `ck.runOnHealthyNode(…)`.
## Test Plan
Tested on `users/brockhargreaves-amd/ck/node-failure-retry-logic` with
`USE_CURRENT_BRANCH_FOR_CK_GROOVY=true`. Verified preflight logging,
reroute on node fault, attempt counter in logs, no retry on aborts, and
single failure status report after budget exhausted.
## Test Result
Retry logic working as expected. Three bugs found and fixed during
testing: false `NodeFault` from host-level sccache probe (sccache is
in-container), `null` node name in catch logging, and `sh` calls outside
`node()` context in status reporting.
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[GFX1250][MX GEMM] Unified FLATMM GroupedGemm Implementation
for MX Data Types (#8325)
## Motivation
Design and test a unified FLATMM GroupedGemm interface so that it
supports all MX FP8, FP6, and FP4 data types on both the gfx950 and
gfx1250 architectures and works seamlessly across these platforms.
## Technical Details
Implementation exposes Grouped Gemm interface for MX FLATMM and MX TDM
FLATMM pipelines.
## Test Plan
Add the following tests:
- ck_tile/grouped_gemm_mx/test_grouped_gemm_mx_flatmm_non_tdm.cpp
- ck_tile/grouped_gemm_mx/test_grouped_gemm_mx_flatmm_tdm.cpp
- ck_tile/flatmm/test_mx_flatmm_persistent.cpp
Verify on the gfx950 and gfx1250 architectures.
## Test Result
All tests pass. Verified on A0 hardware with rocm-7.14.0a20260517
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CKTile] Fix MX GEMM: num_loop==3 dispatch, split-K,
unsupported-shape guard (#6663)
Three independent MX GEMM correctness bugs reported against
example/ck_tile/42_mx_gemm (fp8xfp8, A=Row/B=Col) on MI350X, plus one
host-side atomic-add accumulation bug in the example's repeat loop.
- Pipeline (gemm_pipeline_ag_bg_cr_comp_async.hpp): BlockHasHotloop
required num_loop > PrefetchStages, which let num_loop == 3 enter a hot
loop that produced 5 gemm accumulations instead of 3 (K == 3*K_Tile,
e.g. K=768, deterministically wrong). Require num_loop >= 4 instead:
pre-pipeline + TailNumber::Three already totals exactly 3.
- Kernel (gemm_mx_kernel.hpp): split-K was silently broken because
GridSize did not thread k_batch into blockIdx.z and the scale tile
windows were anchored at K=0 for every k_id. Every k_id >= 1 therefore
read the wrong packed scales. Fix:
* GridSize returns dim3(grid_x, 1, k_batch) (persistent and
non-persistent).
* MakeScaleA/BBlockWindows accept a k_elem_offset and translate it to a
packed-scale K offset (also apply pad_tensor_view so OOB scale loads
return zero, matching A/B padding).
* operator() derives k_id from blockIdx.z, uses GetSplitKElemOffset
(matches Underlying::SplitKBatchOffset's K1-aligned formula), and
dispatches the epilogue with memory_operation_enum::atomic_add for
k_batch > 1, set for k_batch == 1. Same fp16/bf16 even-vector-size guard
as UniversalGemmKernel.
* MakeCBlockWindows templated on DstInMemOp; unconditionally applies
pad_tensor_view using kPadM/kPadN so partial trailing M/N tiles are
handled correctly.
- Compile- and runtime unsupported-shape guards (gemm_mx_kernel.hpp):
add IsSupportedArgument and a static_assert for configurations that
produce silent wrong results:
* static_assert(!kPadK) -- the MX comp-async pipeline uses
async_load_tile whose OOB check is per-vector-start, so a vector
straddling the K pad boundary reads garbage. Until the async path learns
per-element pad masking, reject kPadK at compile time.
* Runtime: k_batch >= 1; M/N multiples of MPerBlock/NPerBlock when
kPadM/kPadN are false; M >= MPerBlock and N >= NPerBlock always
(CShuffleEpilogue cannot safely run with a single partial tile); K %
(KPerBlock * k_batch) == 0; and for k_batch > 1, K must be a multiple of
WarpTile_K * k_batch so every split lands on a packed-scale boundary.
* All error paths log under CK_TILE_LOGGING with actionable messages.
- Example (example/ck_tile/42_mx_gemm/mx_gemm_instance.hpp):
* Call Kernel::IsSupportedArgument up front and throw a clear
runtime_error for rejected shapes (was silently launching an unsupported
kernel).
* Switch to launch_kernel_time_mask with a clear_gemm_output preprocess
that zeroes C between iterations when k_batch > 1 (mirrors
universal_gemm_invoker). Without this the default -warmup=50 -repeat=100
accumulated 150 atomic_adds into C after the kernel-side split-K fix.
Tests (test/ck_tile/gemm_mx/):
- Add MXfp8_GemmConfig16_PadMN (kPadM = kPadN = true).
- test_mx_gemm_fp8.cpp: HotLoopTailNumLoopThree (K=768 regression),
SplitK (k_batch=2,4 across full_k/partial_k paths),
TestMxGemmFp8PadMN::{MNPaddingAligned, MPadding, NPadding, MNPadding}
covering trailing partial tiles along M, N, or both.
- Run(...) now takes k_batch.
- packScalesMNxK: guard against OOB (mn, k) reads from src and
initialise e8m0 bytes to the zero exponent (0x00) instead of the
default-constructed NaN (0xFF), so padded lanes don't poison the packed
int32_t shared with in-range lanes.
- test_mx_gemm_instance.hpp: call IsSupportedArgument before launch.
Verification on gfx950, ROCm 7.2.0:
- ctest -R test_ck_tile_mx_gemm -> 100% (2/2).
- Example sweep over the original bug-report shapes: all K-aligned
shapes now validate correct (including 4096^3 sk=2 and the K=768 cases);
all K=128 shapes cleanly rejected with the new error message instead of
producing silent wrong results.
Made-with: Cursor
## Motivation
<!-- Explain the purpose of this PR and the goals it aims to achieve.
-->
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Add tile shape for FMHA batch prefill on MI308X (on fp8,
hdim=256) (#8350)
## Motivation
Add a tile size appropriate for FMHA batch prefill fp8/hdim256 on MI308X
## Technical Details
Appending the tile shape to the existing factory such that it can be
picked up by Aiter
## Test Plan
Ran the performance test on both MI300X and MI308X
## Test Result
MI300X performance seems unaffected by this change. MI308X does improve.
## Submission Checklist
- [X] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Co-authored-by: Damien Lejeune <damien.lejeune@amd.com>
[CK Tile] Add transposed tile load implementation, and tests
for load_and_convert_tile (#5510)
## Motivation
Mixed precision b/fp16 x fp8 requires a transposed tile load
implementation that supports mixed precision using these types.
Implement this, use it in `load_and_convert_tile`, and add a unit test
for `load_and_convert_tile` which covers this functionality.
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Add graph capture support for FMHA backward(new
branch) (#8262)
## Motivation
Add HIP graph capture support for FMHA backward operations. The original
implementation only supported normal execution mode and would cause
use-after-free crashes when used with graph capture replay.
When FMHA backward is captured into a HIP graph:
- First replay: host callback executes and deletes the closure (as
designed for normal mode)
- Subsequent replays: use-after-free crash because the closure was
already freed
This PR enables `fmha_bwd_launcher::prepare_workspace_async()` to work
correctly in both normal execution and graph capture modes.
[CK Tile] WAVELET pipeline for backward-data grouped
convolution (#8220)
## Motivation
On the RetinaNet shapes (gfx950, fp16) CK Tile backward-data conv was
~18% behind classic
CK, with the gap concentrated in the K=2376 3x3 detection-head family
where bwd_data spends
most of its time. The WAVELET GEMM pipeline already gives uplift for
forward and
backward-weight conv; this ports it to backward-data and consolidates
the now-shared
machinery across all three directions.
## Technical Details
- Backward-data wavelet support in the tile kernel: launch extra load
waves when the
pipeline exposes `LaunchBlockSize`, and split the epilogue into math
waves (run the
CShuffle epilogue) and load waves (`RunBarrierStub`).
- Register 7 WAVELET instances (fp16 and bf16), tuned for
backward-data's tall-skinny GEMM
rather than the forward tile shapes: a big-M `256/128/64` workhorse, a
`VecA=4` variant for
the `K % 8 != 0` shapes, and a `NumGroupsToMerge=32` variant for grouped
(depthwise-style)
shapes.
- Implement the native backward-data instance parser in
`generate_instances.py`.
- Deduplicate the wavelet machinery shared by forward, backward-data,
and backward-weight:
`GroupedConvLaunchBlockSize`, `is_wavelet_pipeline`, and
`RunWaveletAwareEpilogue` in
`grouped_convolution_utils.hpp`; the three native instance parsers
collapse to one
parameterized parser. The three kernels now call the shared helpers.
## Test Plan
- Rebuild the full profiler instance pools for all three directions
(fp16/bf16/fp32,
nhwgc/ndhwgc) to exercise the shared helpers across every instantiation.
- Tile GTests on gfx950: `test_grouped_convnd_fwd_tile`,
`test_grouped_convnd_bwd_data_tile`,
`test_grouped_convnd_bwd_weight_tile`.
- Per-shape sweep of the 35 RetinaNet backward-data shapes vs classic CK
and the
non-wavelet tile pool (`profile_wavelet_bwd_data.py`); correctness
spot-checked with
GPU-reference verification on the new big-M and NumGroupsToMerge
instances.
## Test Result
- GTests pass: forward 9/9, backward-data 6/6, backward-weight 6/6.
- Backward-data perf (3x3 g=1 region, geomean classic/tile): 0.88 ->
1.11, i.e. the tile
path goes from ~12% slower than classic to ~8% faster. The largest
single backward-data
shape (256x100x100->2376) moves from 11% slower than classic to 12.5%
faster.
- The dedup refactor preserves behavior (net -174 lines across the
kernels/generator),
confirmed by the full rebuild and the GTests above.
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[ck] Enforce LF-only line endings in C/C++ sources
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Summary
Several CK source files carry Windows **CRLF** line endings (a trailing
carriage return on each line), introduced by editors configured for
Windows endings or copy/paste from Windows tooling. These are purely
cosmetic but they pollute diffs (whole-file churn the first time someone
makes an LF edit), confuse `clang-format`, and are inconsistent with the
LF-only convention used across the rest of the tree.
This PR (a) normalizes every existing CRLF file (6 files) to LF and (b)
adds a pre-checkin gate so new CRLF leaks are rejected before merge.
## File extensions covered
Both the cleanup scan and the new Jenkins enforcement stage use the same
predicate as the adjacent `ASCII Only Check` stage:
```
*.h *.hpp *.cpp *.h.in *.hpp.in *.cpp.in *.inc *.cl
```
(excluding `*/build/*` and `*/include/rapidjson/*`). The local
pre-commit hook's `c++/inc` type filter covers the same set.
## Why no enforcement today
CK is opted out of the rocm-libraries root `.pre-commit-config.yaml`, so
the existing `pre-commit` workflow doesn't touch CK. The local CK
`.pre-commit-config.yaml` only runs for developers who installed hooks.
The **authoritative gate is therefore the new Jenkins stage** in this
PR; the local hook is convenience.
## Commit layout (bisect-friendly)
1. `[ck] Normalize CRLF line endings to LF in C/C++ sources`
Mechanical line-ending cleanup across 6 files. No content change: every
edit is purely CRLF -> LF, verified with `git diff --ignore-cr-at-eol`
reporting an empty diff.
2. `[ck] Enforce LF-only line endings in C/C++ sources`
- New `projects/composablekernel/script/check_no_crlf.sh` (modeled on
`check_ascii_only.sh`).
- New `crlf-checker` entry in
`projects/composablekernel/.pre-commit-config.yaml` under the
local-hooks block (`types_or: [c++, inc]`).
- New `CRLF Check` parallel stage in
`projects/composablekernel/Jenkinsfile`'s `Static checks` block,
mirroring the adjacent `ASCII Only Check` stage. Always-on, no
`RUN_CPPCHECK` gate.
The tree is buildable at every commit boundary. Commit 1 leaves 0 CRLF
violations; commit 2 wires the gate.
## Demo
Script output on a synthesized violation:
```
$ printf 'int main() {}\r\n' > /tmp/bad.cpp
$ projects/composablekernel/script/check_no_crlf.sh /tmp/bad.cpp
ERROR: /tmp/bad.cpp contains CRLF (Windows) line endings:
1:int main() {}<CR>
Fix: convert to LF, e.g. 'sed -i 's/\r$//' /tmp/bad.cpp' or 'dos2unix /tmp/bad.cpp'
$ echo $?
1
```
Full repo scan after the cleanup commit:
```
$ cd projects/composablekernel && find . -type f \( -name '*.h' -o -name '*.hpp' -o -name '*.cpp' \
-o -name '*.h.in' -o -name '*.hpp.in' -o -name '*.cpp.in' -o -name '*.inc' -o -name '*.cl' \) \
-not -path '*/build/*' -not -path '*/include/rapidjson/*' -print0 \
| xargs -0 -P 8 -n 64 script/check_no_crlf.sh
$ echo $?
0
```
## Test plan
- [ ] Jenkins PR build: confirm new `Static checks -> CRLF Check` stage
runs green over the full predicate and the existing `ASCII Only Check` /
`Clang Format` stages are unaffected.
- [ ] Local: `pre-commit run crlf-checker --all-files` runs cleanly
after installing CK pre-commit hooks.
- [ ] Manually inject a CRLF line ending in any `.cpp/.hpp/.inc` file,
push: confirm Jenkins fails the new stage with a clear error.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
[CK] Pre-emptively add groovy/ folder and skip TheRock CI
filter (#8378)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
The CK Groovy library is growing and will be reorganized into a
self-describing `groovy/` folder rather than living under `src/` and
`vars/`. This PR creates that folder pre-emptively and adds it to the
TheRock CI skip-list so that future Groovy additions do not
unnecessarily trigger TheRock builds.
## Technical Details
- Added `projects/composablekernel/groovy/` with a `.gitkeep` to
establish the directory in the repo.
- Added `"projects/composablekernel/groovy/*"` to
`SKIPPABLE_PATH_PATTERNS` in `.github/scripts/therock_configure_ci.py`
alongside the existing `vars/*` entry, ensuring changes confined to
Groovy pipeline code are recognized as non-therock-relevant and skip the
TheRock CI pipeline.
## Test Plan
No code logic was changed. Verified that `therock_configure_ci.py`
pattern list is consistent with the existing `vars/*` skip entry and
that the new pattern follows the same glob convention.
## Test Result
N/A — directory scaffolding and CI filter only; no functional code
affected.
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] Re-enable HIPRTC codegen tests for all CK PRs.
## Motivation
At the time when we introduced the smart test filter to only build and
run tests affected by the PR changes, we disabled the client examples,
which required full CK build, and also the hiprtc tests that were
grouped with the client examples. This caused a few PRs to sneak through
that caused the hiprtc compilation to fail.
By restoring the hiprtc tests in all PRs, we should close this gap.
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Revert "[CK_TILE] Implement RTC API for a subset of FMHA
functionality for MGX" (#8333)
Reverts ROCm/rocm-libraries#6086
Need to revert as the codegen test for fmha is failing due to including
std header:
2026-06-11T22:36:03.673Z] In file included from
/tmp/comgr-953928-0-473822/include/ck/host/device_fmha_fwd/fmha_fwd_wrapper.hpp:8:
[2026-06-11T22:36:03.673Z] In file included from
/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/cmath:49:
[2026-06-11T22:36:03.673Z] In file included from
/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/std_abs.h:38:
[2026-06-11T22:36:03.673Z] /usr/include/stdlib.h:32:10: fatal error:
'stddef.h' file not found
[2026-06-11T22:36:03.673Z] 32 | #include <stddef.h>
[2026-06-11T22:36:03.673Z] | ^~~~~~~~~~
The ck_tile headers were never prepped for hiprtc compilation.
Users/tlakshma/ck/tile engine develop
## Motivation
This PR adds multiple new GPU kernel benchmarking operations to the CK
Tile Engine, expanding its coverage of GEMM-family operations:
- **gemm_multi_abd**: GEMM with multiple A, B, and D tensors, enabling
epilogue patterns such as scale/bias fusion.
- **batched_contraction**: Batched tensor contraction supporting
multi-dimensional batch (G), M, N, and K dimensions, targeting workloads
where the contraction indices span more than one logical axis.
- **mx_gemm**: MX-format GEMM with microscaling (e8m0) scale tensors.
- **gemm_rowcolquant**: Block-scale GEMM with row/column quantization.
- **gemm_tensor_quant**: Block-scale GEMM with tensor quantization.
- **grouped_gemm_rowcolquant**: Grouped GEMM with row/column
quantization.
- **grouped_gemm_tensorquant**: Grouped GEMM with tensor quantization.
- **batched_gemm**: Batched GEMM benchmarking support.
## Technical Details
### gemm_multi_abd
- New subdirectory: tile_engine/ops/gemm/gemm_multi_abd/
- CMakeLists.txt follows the same individual-target pattern as
gemm_universal / gemm_multi_d.
- gemm_multi_abd_instance_builder.py subclasses GemmKernelBuilder from
the shared gemm_instance_builder.py.
- gemm_multi_abd_benchmark.py delegates to the shared GemmBenchmark
parent class.
- Configs: default_config.json, default_ci_config.json,
user_provided_config.json.
- Supported GPU targets: gfx90a, gfx942, gfx950, gfx1201.
### batched_contraction
- New subdirectory: tile_engine/ops/gemm/batched_contraction/
- Extends GemmKernelBuilder via BatchedContractionKernelBuilder, adding
num_dim_g, num_dim_m, num_dim_n, num_dim_k, num_d_tensors, and
elementwise_function parameters.
- Layout string uses 3-character encoding (A+B+E), e.g. rcr.
- Self-contained benchmark sweep driver
(batched_contraction_benchmark.py) with JSON/CSV export and best-kernel
selection.
- Supported GPU targets: gfx90a, gfx942, gfx950.
### mx_gemm
- New subdirectory: tile_engine/ops/gemm/mx_gemm/
- Supports MX-format (e8m0) microscaling for A and B scale tensors.
### block_scale_gemm (gemm_rowcolquant, gemm_tensor_quant)
- New subdirectory: tile_engine/ops/gemm/block_scale_gemm/
- gemm_rowcolquant: row/column quantization epilogue.
- gemm_tensor_quant: tensor-level quantization epilogue.
### grouped_gemm_quant (grouped_gemm_rowcolquant,
grouped_gemm_tensorquant)
- New subdirectory: tile_engine/ops/gemm/grouped_gemm_quant/
- grouped_gemm_rowcolquant: grouped GEMM with row/column quantization.
- grouped_gemm_tensorquant: grouped GEMM with tensor quantization.
### batched_gemm
- New subdirectory: tile_engine/ops/gemm/batched_gemm/
- Batched GEMM benchmark support wired into the sampling/active-op
lists.
All new ops are registered in op_weights.json for budget allocation and
wired into the active-op sampling lists in CMakeLists.txt.
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Add cluster launch in test ck_tile mx gemm tdm wmma
## Motivation
Add cluster launch test in test_ck_tile_mx_gemm_pipeline_tdm_wmma on
gfx1250, so that we can check the performance on gfx1250 hardware.
## Technical Details
Added Out-of-bounds guard in RunGemm of MxGemmKernel to skip blocks
padded by cluster alignment.
Add ClusterEnable/ClusterDisable aliases and extend the tuple in
test_mx_gemm_pipeline_kernel_types.hpp by adding two kernel types with
ClusterEnable for F8 CompTDMV1 and CompTDMV2 respectively. The existing
F4 non-ClusterLaunch kernel types have issue to be fixed, so this PR
does not include F4 cases.
Read ClusterLaunch from the tuple in test_mx_gemm_pipeline_util.hpp.
Update invoke_mx_gemm to branch on ClusterLaunch, including Add cluster
size constants, Switch GemmShape type, TilePartitioner type, and the
kernel launch call.
## Test Plan
Tested the changes on gfx1250 FFM.
## Test Result
The added kernel types (instances) passed the tests on gfx1250 FFM.
## Submission Checklist
- [x ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Implement RTC API for a subset of FMHA
functionality for MGX (#6086)
## Motivation
Introduce a wrapper for the FmhaFwdKernel, for use in real time
compilation in MIGraphX.
## Technical Details
The intent of the API is to provide multiple instances of the
FmhaFwdKernelWrapper, suitable for a particular problem definition.
At the moment the wrapper only supports bias and causal masking, feature
expansion will come in a future pr.
The usage pattern is, in short:
1. Define fmha_fwd::Problem (input dimensions, data type, etc)
2. Fetch Solutions for target architecture (currently only gfx942) based
on Problem.
The solutions contain a map of template -> template parameter and can be
converted to a string representing the full instantiation of
FmhFwdKernelWrapper e.g. `ck_tile::FmhaFwdWrapper<ck_tile::fp16_t, 128,
64, 16, 32, 32, 32, 4, 1, 1, 4, 1, 1, 32, 32, 16, 32, 32, 16, false,
true, false, true, true, true, true, ck_tile::FmhaPipelineTag::QR>`
3. The instance can then be used in an RTC kernel. The kernel needs to:
* Construct a Descriptor (containing descriptions of all input tensors)
* Call IsValid() on the descriptor to check if the instance is
applicable. Note that this is constexpr by design so that it can fail
the kernel compilation as a signal that the kernel is not applicable.
* Pass the descriptor and input pointers to the wrapper Run method.
A more detailed example of usage can be found in
codegen/test/fmh_fwd.cpp
Beside work on creating the wrapper and the supporting API, the PR also
contains some changes necessary to enable compilation with HIPRTC.
The contents of the CK tile headers are embedded in a binary file which
is used to pass the header files as strings to HIPRTC.
Many of the ck tile headers contain host only code which leads to
compilation failures.
ck_tile_headers_preprocessor goes through the embedded headers and
removes the bodies of host only functions, thereby eliminating the
compilation failures.
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] Padding on K for global load for grouped conv bwd data
(#8272)
## Motivation
Fix incorrect results caused by lack of padding during global load in
grouped convolution backward data kernel. It is needed since there is no
OOB check for global load.
## Technical Details
Add padding needed for global load which not use OOB check.
## Test Plan
test_grouped_convnd_bwd_data*
## Test Result
Passed locally
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK dispatcher] - LGBM predict data_type FLOAT32->FLOAT64 in
ml_heuristic (#8132)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Summary
`ml_heuristic.hpp` calls `LGBM_BoosterPredictForMat(...,
/*data_type=*/0, ...)` (`C_API_DTYPE_FLOAT32`) against a
`std::array<double, NUM_FEATURES>` feature buffer. LightGBM reinterprets
the 8-byte doubles as 4-byte floats → invalid predictions → the
heuristic's argmax always tie-breaks to the first/smallest enumerated
config.
**Fix:** `data_type 0 → 1` (`C_API_DTYPE_FLOAT64`), matching the
`double` buffer. After the fix, predictions vary and track real TFLOPS
(the model correctly prefers larger tiles).
## Verification
- The feature buffer `f` is `std::array<double, NUM_FEATURES>`
(NUM_FEATURES = 72) → `f.data()` is a `double*`.
- The changed `0` is the 3rd positional `data_type` argument (not
`nrow`/`ncol`/`is_row_major`).
One-line functional change.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
[CK] increase time limit for fmha_bwd tests to prevent
timeouts (#8241)
## Motivation
Observed a CI failure due to fmha_bwd test timeout which never happened
before. Going to increase the time limit for the test to prevent any
further CI failures.
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK TILE] Fix performance regression caused by Dispatcher
codegen compiler flag. (#8019)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
Currently CK Tile two codegen paths: CK Builder and CK Tile Dispatcher.
The CK Tile Dispatcher codegen uses an additional compiler flag that is
not present in the CK Builder codegen workflow. The additional compiler
flag can cause performance regression for so instances as it disables
relevant compiler optimizations.
## Technical Details
Removed compiler flag `-mllvm -enable-noalias-to-md-conversion=0` from
the CMakeLists.txt that creates instance library from Dispatcher
codegen.
## Test Plan
Required testing is contained in the CI/CD pipeline.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] Fix scale init in profile_grouped_conv_fwd_outelementop
(#8208)
## Motivation
Wrong scale initialization caused random errors on CI.
## Technical Details
InvScale was initialized by 0 what caused nans during division. At now
zero are excluded from randing.
## Test Plan
TestGroupedConvndFwdConvInvscale3d
## Test Result
Passed in 100 runs
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
AICK-1400
[CK] Magic division for long_index_t
## Motivation
Improve performance for long_index_t kernels
## Technical Details
Support magic division for long_index_t
## Test Plan
test_grouped_convnd*
## Test Result
Passed locally
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
AICK-1386
[CK_Tile] Add wmma_bf16f32_16x16x32_bf16 warp-gemm test
(#8035)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Summary
Adds the warp-gemm unit test for `wmma_bf16f32_16x16x32_bf16`. **Stacked
on #8028** (the API change) and based on its branch, so #8028 shows the
isolated API diff and this PR shows just the test.
## Test
gfx125-guarded `WmmaBf16f32.ResidualPrecisionContrast`: computes `Y_bf16
= X_bf16·W_bf16 + R_fp32` via `WarpGemm::mac_downconvert`, compares
against an fp32 reference (within bf16 tolerance), and asserts it is at
least as accurate as the bf16-accumulate path — i.e. it demonstrates the
precision benefit of the fp32 accumulator (`C`) carried into the fused
bf16 down-convert.
Passes on gfx1250.
[ck] Unify Build_CK and buildHipClangJob into buildAndTest
(#8108)
## Motivation
`projects/composablekernel/vars/ck.groovy` had two near-identical build
functions, `buildHipClangJob` (lean: static checks, FMHA, tile-engine,
conv) and `Build_CK` (main per-arch matrix). This removes the
duplication and fixes a latent GitHub-status bug that lived in both.
## Technical Details
- Merged both into one `buildAndTest(Map conf)` gated by an explicit
`is_main_build` flag (default `false` = lean path; `true` adds the GPU
check + arch-gated inductor/perf/hipTensor; only `runBuildCKAndTests`
sets it).
- Deleted the `Build_CK_and_Reboot` / `buildHipClangJobAndReboot`
wrappers (they only logged and re-threw); all 13 call sites now call
`buildAndTest` directly.
- Widened the shared `catch` to `Exception` so build / image-pull / "GPU
not found" failures report **failure** instead of leaving the check
stuck **pending** (failing stages now go red).
- Removed the dead `no_reboot` key. No change to what is built or
tested.
## Test Plan
- Jenkins linter on the `Jenkinsfile`.
- One branch run covering both paths (per-arch matrix + lean stages),
spot-checking gfx1250 and a nogpu stage.
## Test Result
- Verified statically: no `buildHipClangJob*` / `Build_CK*` references
remain; `buildAndTest` defined once, all call sites wired.
- Pending: linter + branch run before merge.
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK Tile] Stream-K RDNA Support
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
Currently, CK Tile Stream-K only supports CDNA architectures. This
change adds Stream-K support on RDNA3/3.5 and RDNA4 architectures.
## Technical Details
Stream-K currently has 3 reduction strategies: 1) atomics, 2) linear,
and 3) tree. The linear and tree reductions require inter-workgroup
communication to a global flags buffer and a global partials buffer. To
ensure cache coherency, we use cache modifiers to skip cache levels that
are not visible to all workgroups. On CDNA architectures, scalar load
and scalar store instructions are available, which we use to read and
write to the flags buffer with appropriate cache skipping modifiers.
However, RDNA architectures do not support scalar store instructions, so
workgroups must use a buffer store instruction to write to flags.
Additionally, cache modifiers differ between CDNA and RDNA; they also
differ between RDNA3 and RDNA4. Given this information, the main changes
are as follows:
- Added RDNA flag signaling: Use buffer store instructions for writing
to global flags buffer
- Add appropriate cache modifiers for reading and writing to flags and
partials:
- RDNA3 (gfx11): Use `glc | dlc` coherence flags
- RDNA4 (gfx12): Use `DEVICE` coherence scope
- SFINAE-guarded overloads: Added compile-time dispatch for
`SignalStorePartialDone()` and `WaitStorePartialDone()` based on target
architecture
- RDNA alignment requirements: Increased flags buffer alignment from
128B to 256B due to RDNA cache line size
**A note about the `amd_buffer_coherence_enum`:**
- **Problem:** The `amd_buffer_coherence_enum` uses preprocessor
conditionals (`#if defined(__gfx12__)`) to define architecture-specific
values. Template specializations reference enum values from different
architectures (e.g., `glc_dlc` for GFX11). Due to C++ two-phase name
lookup, non-dependent names are resolved during template parsing
regardless of which architecture is being compiled, causing compilation
failures when referenced values do not exist in the active preprocessor
branch.
- **Temporary Solution**: Added compatibility enum values to each
architecture block. For example, I added `glc_dlc` in the `__gfx12__`
block. I will create a ticket to refactor this enum with a design that
has better scalability and tries to avoid the use of preprocessor
conditionals.
## Test Plan
### Summary
gtests were added to test wmma variants of Stream-K. These tests were
stressed tested locally on gfx11 and gfx12.
### More details
This PR makes the following changes/additions to the Stream-K gtests:
- Split tests into MFMA (CDNA) and WMMA (RDNA) variants
- Added 16 WMMA kernel types: FP16/BF16/FP8/BF8 × Linear/Tree reduction
- WMMA uses 16×16×16 wave tiles for RDNA (this is the only tile size
supported on RDNA)
- Fixed RDNA WGP mode: multiply multiProcessorCount by 2 for actual CU
count
- As described in [HIP
documentation](https://rocm.docs.amd.com/projects/HIP/en/docs-7.2.0/doxygen/html/group___global_defs.html#ggacc0acd7b9bda126c6bb3dfd6e2796d7ca3ac50041beb59111a5c76edf03da0898),
when in Workgroup Processor (WGP) mode, the value of
`hipDeviceAttributeMultiprocessorCount` is half of CUs, because a single
WGP contains two CUs. The default mode on RDNA is WGP mode, so when
creating (M, N, K) instances for gtests using the CU count, we need to
multiply the CU count by 2 to get the correct value. This is not needed
in the kernel host code, because the occupancy ensures that overall
`max_active_wgs` is correct.
## Test Result
All tests pass locally.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Using named functors instead of lambdas
## Motivation
Currently, in block-level GEMM pipelines, there is significant code
repetition for prefetching and tail handling, where lambda functions
create a unique instantiations at each call. This includes repeated
static_for instantiations and large loops such as MRepeat. Each
repetition results in additional instantiations, which increases
compilation time and binary bloat.
## Technical Details
Refactor repeated code blocks into named functors so the compiler can
reuse already instantiated code instead of generating multiple copies.
Scope of changes:
1. WMMAOPS pipeline internals:
projects/composablekernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp,
projects/composablekernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_v1.hpp,
projects/composablekernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_v3.hpp
2. XDLOPS and preshuffle pipeline variants across
projects/composablekernel/include/ck/tensor_operation/gpu/block
(v1/v2/v3/v4/v5, scale, dequant, gufusion, moe, mx, blockscale,
skip-b-lds, dpp, xdlops)
Shared functor file:
projects/composablekernel/include/ck/utility/vector_load_functor.hpp
## Test Plan
Note that the provided compilation traces by -ftime-trace do not report
unnamed lambda instantiations, so a clear baseline for instantiation
counts cannot be established. As a result, the impact of this change
will be evaluated based on runtime performance rather than direct
instantiation-count comparisons.
## Test Result
The effects of this were timed by the compilation of a single HIP object
through an example (grouped_gemm_wmma_splitk_fp16.cpp). The average user
time and speedup of this using the average of 100 compilations is:
- Mean compile time before the changes: 37.734 s
- Mean compile time after: 32.087 s
- Speedup: 17.6%
Ran a full CK compilation on Alola with the following results:
| Metric | Before (min) | After (min) | Absolute Reduction (min) | %
Reduction |
| ------ | ------------ | ----------- | ------------------------ |
[CK] Remove Stream-K from old CK
## Motivation
Since Stream-K has a CK Tile implementation, we no longer need Stream-K
in old CK. Hence, this PR removes Stream-K from old CK.
## Technical Details
All Stream-K artifacts in old CK have been removed including examples,
tests, kernels, and CK profiler artifacts.
## Test Plan
Ran a CI run on the branch before publishing PR.
## Test Result
All tests passed.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Co-authored-by: Claude Sonnet 4 <noreply@anthropic.com>
[CK Tile] Add WAVELET pipeline for forward grouped
convolution (#8009)
## Motivation
CK Tile forward grouped convolution trails classic CK on 3x3
convolutions whose
output-channel count is not divisible by 8, where the narrow output
store limits
the compute CShuffle epilogue. This ports the WAVELET pipeline (added
for
backward-weight in #7937) to the forward kernel to close that gap.
## Technical Details
- Kernel (`grouped_convolution_forward_kernel.hpp`): WAVELET
load/math-wave wiring,
mirroring the backward-weight implementation; the non-WAVELET path is
unchanged.
- Generator: implement `parse_native_fwd_instance`, the forward
native-instance parser.
- Registered WAVELET instances: profiler bf16 3 / fp16 5, tests 1 each.
WAVELET requires input channels divisible by 8 (it does not apply to
depthwise).
The bf16/fp16 instance asymmetry is intentional and measured: the VecC=8
tiles
never beat the compute pool in bf16 but win about 20% of divisible-by-8
3x3 shapes
in fp16, so VecC=8 is registered for fp16 only.
## Test Plan
- Correctness (CPU reference) for every registered profiler instance,
across VecC variants.
- Per-shape best-instance performance sweep over the 34 RetinaNet shapes
(bf16) and
a 200-shape cross-model sweep (bf16 and fp16), compared against classic
CK.
## Test Result
- Correctness: PASS for all instances.
- RetinaNet (bf16, vs classic CK): faster on 28 of 34 shapes, geomean
+19.5%; the
not-divisible-by-8 shapes up to 3.7x. One 1x1 stride-2 shape stays ~20%
behind
classic CK, unrelated to WAVELET.
- Cross-model (200 shapes): WAVELET wins 3x3 not-divisible-by-8 in both
dtypes
(up to 61% over the next-best compute instance); for divisible-by-8 3x3
it wins
about 20% of shapes in fp16 (3-11%) and none in bf16.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
[CK Tile] Fix Stream-K flag store: wave-uniform SGPR address
for scalar s_store/s_load (#8099)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
Stream-K grouped-conv (and GEMM) kernels fail to assemble for some
instances: the inline scalar flag store/load gets a VGPR address
operand, which scalar-memory instructions reject (`invalid operand for
instruction`). This blocks Stream-K instances from building.
## Technical Details
- `StreamKReductionOps::{Signal,Wait}StorePartialDone` (shared by GEMM
and conv, added in #5393) take `kargs` by `const&` and feed
`kargs.workspace_ptr` / `cta_idx` into inline
`s_store_dword`/`s_load_dword` with `"s"` constraints. For some
instantiations the compiler can't keep the pointer wave-uniform and
emits a VGPR address.
- Fix: route the pointer and offset through `amd_wave_read_first_lane`
so the scalar-memory address is a wave-uniform SGPR before the asm. Same
instructions, no algorithm change.
- Not arch-specific: the affected instance fails on
gfx908/gfx90a/gfx942/gfx950 without the fix; whether the compiler spills
to a VGPR depends on the instantiation (tile/warp/pipeline), not the
target.
## Test Plan
- Compile the previously-failing dispatcher instance for
gfx908/gfx90a/gfx942/gfx950.
- `test_ck_tile_grouped_conv_bwd_weight_streamk` on gfx942, gfx90a,
gfx950 hardware.
- gfx950 perf A/B (example, bf16/tree, 10 runs each) with vs without the
change.
## Test Result
- Failing instance now assembles on all four archs; previously failed on
every one.
- 30/30 conv Stream-K tests pass on gfx942, gfx90a, gfx950.
- gfx950 perf delta -0.13% (within run-to-run noise) — no regression
from the added readfirstlane on the cold flag path.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
[CK] Grouped Convolution Global Load/Store instances
## Motivation
Support global load and store in grouped convolutions using instance
factory.
## Technical Details
- add new instances for each direction
- add new tests for large cases
## Test Plan
New test for large cases
## Test Result
pending
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
AICK-1255
[Ck][CK Tile] Global Load/Store for Large Tensors support
(#6744)
## Motivation
Create solution to support large tensors in the entire ck tile.
## Technical Details
- add possiblity to use global load
- int64 indexing
## Test Plan
conv fwd tests
## Test Result
passed locally
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
AICK-913
[CK Tile] Extend type support EightWave pipeline
## Motivation
EightWave pipeline was designed for 8 bit types. This PR extend support
for any FP type
## Technical Details
- Generalize policy to support any FP type
- Change LDS layout to fix bank conflicts. This removes all bank
conflicts in the pipeline (checked for all supported types). Remaining
bank conflicts are related to Cshuffle epilogue.
## Test Plan
Added GEMM tests with new supported types. Note that FP6 is also
supported for MX GEMM but the PR was reverted so no tests were added for
it.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[tile_engine] Integrate gemm_streamk into budget-based
sampling system (#8079)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
`gemm_streamk` was the only GEMM op not participating in the tile
engine's budget-based sampling system. Without a budget cap, it would
always generate its full feasible set, making build times unpredictable
and inconsistent with the other ops.
## Technical Details
- **CMake budget propagation** (`ops/gemm/CMakeLists.txt`): Added
`gemm_streamk` to the active-ops detection loop so it receives a share
of the sampling budget. Because `gemm_streamk` lives in a sibling
subdirectory (`ops/gemm_streamk/`), its allocation is written via `CACHE
STRING "" FORCE` to make the variable visible across the CMake directory
boundary.
- **Per-combo budget division** (`ops/gemm_streamk/CMakeLists.txt`,
`ops/gemm/grouped_gemm/CMakeLists.txt`): Added the same per-combo
`MAX_INSTANCES` division that exists in `gemm_universal` and
`gemm_preshuffle`. The total budget is divided by `n_datatypes ×
n_layouts` before the inner `foreach` loop so that sampling fires
independently per `(dtype, layout)` combo rather than acting as a single
global cap.
- **Sampling integration** (`gemm_streamk_instance_builder.py`): Added
`_apply_sampling()` method to `GemmKernelBuilder`, mirroring the
Sobol+LHS+maximin sampling used by other ops. New constructor
parameters: `gpu_target`, `max_instances`, `seed`, `tier`,
`manifest_path`. New CLI arguments: `--gpu_target`, `--max-instances`,
`--seed`, `--tier`, `--manifest-path`. The `--gpu_target` argument is
now also forwarded on the `--list_kernels` invocation.
- **`GEMM_STREAMK_AXES`** (`sampling/feasible_set.py`): Defined as
`GEMM_AXES + ["reduction_strategy"]` to account for the extra axis
unique to stream-K. Added `reduction_strategy` to `CATEGORICAL_AXES`.
- **Weight rebalancing** (`sampling/op_weights.json`): Allocated 10%
weight to `gemm_streamk` by proportionally reducing `gemm_universal`
(0.35 → 0.30) and `gemm_preshuffle` (0.30 → 0.25). Total remains 1.00.
## Test Plan
- Configure with `TILE_ENGINE_SAMPLING_TIER=daily` and verify that
`gemm_streamk` receives a non-zero budget allocation and that
`GEMM_STREAMK_MAX_INSTANCES` is set correctly.
- Configure with `TILE_ENGINE_SAMPLING_TIER=daily` across multiple
`(dtype, layout)` combos and confirm per-combo budget = total /
n_combos.
- Configure with `-DGEMM_STREAMK_MAX_INSTANCES=50` explicit override and
verify the override is respected (budget allocation skipped).
- Verify `chosen_instances.json` manifest is written to the working path
when tier is active.
- Confirm `op_weights.json` weights still sum to 1.00.
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] [MIOPEN] Split convolution library by layout
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
# Split Composable Kernel convolution operations by data layout
TLDR:
1. This is a reorganization of files, folders, and CMakeLists for
convolution kernels and facilitates a splitting of the convolution
library into layouts.
2. The speedup can range anywhere between 15-40% depending on the target
architecture for miopen only builds of CK. For TheRock nightly builds of
CK, which includes both miopen and hip tensor kernel instances, this
constituted in a 10% decrease in compile time for gfx1100.
## Overview
Based on https://github.com/ROCm/composable_kernel/pull/3010/ (except
keeping 1 static library)
## What MIOpen Actually Uses
MIOpen **exclusively uses:
- **NHWGC** for all 2D convolutions
- **NDHWGC** for all 3D convolutions
This is because MIOpen's tensor descriptors natively use channel-last,
group-aware formats.
## Key Changes
### 1. Layout-Based Directory Structure
Reorganized convolution instance files from flat per-operation to
hierarchical layout-based structure. For example:
**Before:**
grouped_conv2d_fwd/
├── device_grouped_conv2d_fwd_xdl_nhwgc_*.cpp (MIOpen-required)
├── device_grouped_conv2d_fwd_xdl_gnhwc_*.cpp (optional)
└── device_grouped_conv2d_fwd_xdl_ngchw_*.cpp (optional)
**After:**
grouped_conv2d_fwd/
├── nhwgc/ ← MIOpen-required
│ ├── xdl/device_grouped_conv2d_fwd_xdl_*.cpp
│ └── wmma/device_grouped_conv2d_fwd_wmma_*.cpp
├── gnhwc/ ← Optional (excluded with MIOPEN_REQ_LIBS_ONLY)
└── ngchw/ ← Optional (excluded with MIOPEN_REQ_LIBS_ONLY)
### 2. Preserved Umbrella Library
As before, all convolution operations are consolidated into a single
static `device_conv_operations` library:
- Aggregates layout-specific instance object files via
`ADD_CONV_LAYOUT_INSTANCES` macro
- **Default build:** Includes all layouts (NHWGC + GNHWC + NGCHW +
NDHWGC + GNDHWC + NGCDHW)
- **MIOpen build (`MIOPEN_REQ_LIBS_ONLY=ON`):** Includes only NHWGC and
NDHWGC layouts
### 3. Binary Size Reduction
When building with `MIOPEN_REQ_LIBS_ONLY=ON`:
**Layouts Included (26 targets):**
- 7× NHWGC instances (2D operations + variants)
- 19× NDHWGC instances (3D operations + variants)
**Layouts Excluded (16 targets):**
- 3× GNHWC instances (2D operations)
- 3× NGCHW instances (2D operations)
- 3× GNDHWC instances (3D operations)
- 3× NGCDHW instances (3D operations)
- 2× GNWC instances (1D operations)
- 1× NWGC instance (1D operations)
- 1× additional NHWGC instance (grouped_conv1d_fwd, not needed by
MIOpen)
This represents a **~38% reduction in instance targets** (16 excluded
out of 42 total
layout-specific targets).
### Testing
- ✅ All existing CK tests link against the umbrella library
- ✅ MIOpen links successfully with the reduced umbrella library
- ✅ Profiler builds with all layout-specific targets explicitly listed
Notes from the Author:
Since this refactor moved most of the convolution files further into
subdirectories, I concentrated on ensuring that no source files were
excluded, including sharded sources: Targets are correctly migrated — no
missing targets, no shard count mismatches.
=?UTF-8?q?[CK=20TILE]=20Unification=20Work=20=E2=80=93=20?=
=?UTF-8?q?Add=20MFMA=20specialisations=20for=20`tf32=5Ft`=20(#6768)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
This PR adds two specialisations related to `tf32_t`.
## Technical Details
This change treats `tf32_t` as a concrete type rather than an empty
`struct`. It also adds two new specialisations for MFMA dense builtins
and resolves existing circular include issues.
## Test Plan
All the new wrappers were added to the test suite in
test_amdgcn_mma_layout.inc.
## Test Result
Test should pass.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] [QuantGEMM] Fix SplitK tail handling and other
improvements (#7199)
This pull request introduces improved and more robust split-K support
for quantized GEMM. The main changes add runtime validation, utility
functions for split-K batch calculations, pointer offset handling for
split-K in grouped kernels, and enhanced support for various tensor
layouts. The changes also improve error handling and provide more
flexibility for runtime tail handling in split-K pipelines.
**Split-K Support and Validation Enhancements:**
* Added runtime validation to ensure `k_batch` is a positive integer and
that split-K configurations do not produce empty final batches or
mismatched pipeline tails, with detailed error messages and logging for
misconfiguration.
[[1]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871R1184-R1211)
[[2]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871L1161-R1250)
* Introduced utility functions `get_splitk_batch_k_read` and
`get_splitk_last_batch_k` to compute per-batch K read sizes and handle
split rounding, ensuring correct and consistent split-K batch
partitioning.
[[1]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871R206-R234)
[[2]](diffhunk://#diff-635b89bdffa96b2b42f1632520cde36701d7d631e864185591f6b32f7645cf47L104-R107)
[[3]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871L388-R417)
[[4]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871L1161-R1250)
* Changed the default value of `k_batch` in `QuantGemmHostArgs` to 1 (no
split-K) for safer default behavior.
**Pointer Offsets and Grouped Kernel Handling:**
* Updated `QuantGroupedGemmKernel` to apply split-K per-batch offsets to
all input pointers, mirroring the behavior of non-grouped kernels and
ensuring correctness for split-K launches.
* Modified AQ tensor view handling to correctly reflect the remaining
K-groups from the split-K batch's offset position, improving accuracy
for split-K in grouped kernels.
**Pipeline and Layout Flexibility:**
* Added support for runtime selection of split-K tail handling via a new
template parameter `RuntimeSplitKTail_`, with new helper methods to
dispatch GEMM pipelines accordingly.
[[1]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871R273)
[[2]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871R1496-R1567)
[[3]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871L1427)
[[4]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871L1447-R1629)
[[5]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871L1459-R1641)
* Improved handling for tensor layout cases, including preshuffled B and
both row-major and column-major AQ layouts, ensuring correct pointer
arithmetic and compatibility checks.
[[1]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871R438-R454)
[[2]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871L464-R516)
[[3]](diffhunk://#diff-d000149a681cd42bfb9947872c603e556cea26cbd7fd4f8f60afc6595d975871R1184-R1211)
[CK Tile] Async support preshuffle GEMM
## Motivation
Add async support to existing preshuffle GEMM pipeline
## Technical Details
Notes:
the implementation avoids previous strategy of duplicating pipelines for
async support and instead add a switch `Async` to the ops Problem to
enable async pipeline. Then, integrate the async pipeline in the
existing one. This allows to avoid code duplication and facilitate the
integration of buffer load to lds in existing pipelines. In my opinion,
it should be used also for other pipelines which don't support buffer
load to lds yet and it would also be a good idea to refactor the
existing async GEMM pipelines with the same approach.
Summary:
- integrate buffer load to lds in existing pipeline
- add optimal tensor descriptors for vmem loading and lds reading. They
are currently optimized for 16x16 wave tiles but they also work for
32x32 wave tiles. Optimizations for 32x32 wave tile requires different
lds layout and it will be done in a follow-up issue
- Add async config to examples
- Add test (gfx950 only)
## Test Plan
New test for gfx950 `test_ck_tile_gemm_pipeline_wp_async`
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_Tile] Add wmma_bf16f32_16x16x32_bf16 via
fused-downconvert override (#8028)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Summary
Adds `__builtin_amdgcn_wmma_bf16f32_16x16x32_bf16` (fp32 accumulate →
bf16 output) to the CK Tile WMMA warp-gemm path. **API only** — the unit
test is split into a stacked PR (#8035) so this API change can be
reviewed in isolation.
## Changes (4 files)
- **16-bit trait:** `wmma_intrinsic_downconvert` (calls the bf16f32
builtin — fp32 C in, bf16 C out) plus `COutDataType = bf16_t` /
`COutVecType`.
- **`WarpGemmAttributeWmmaImpl` / `WarpGemmAttributeWmma`:**
`mac_downconvert(c_fp32, a, b)` (kTransC-aware) returning the bf16
C-output vector.
- **`WarpGemmImpl`:** `mac_downconvert` tail handler producing a bf16
C-output tile from the fp32 accumulator tile, reusing
`CWarpDstrEncoding` (output layout identical to the f32 C tile).
Verified on gfx1250 (via the stacked test PR #8035): the test passes;
the existing WMMA warp-gemm test is unaffected (additive change only).
[CK Tile] PermuteN support MX GEMM
## Motivation
Add PermuteN support to preshuffle MX GEMM
## Technical Details
- Modify `shuffle_b_permuteN` to support MX preshuffled layout
- Add `preShuffleScalePermuteN` with same functionality of
`preShuffleScale` but layout consistent with PermuteN
- Include MX pre-processing functions in the library
## Test Plan
Add test configuration for permuteN with preshuffle (both FP4 and FP8)
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Co-authored-by: Cong Ma <congma13@amd.com>
[ck] Updated CK Tile documentation to use mermaid diagrams
(#7955)
## Motivation
There were mermaid diagrams in the CK Tile doc that were converted to
svg. However, there is an extension for mermaid diagrams. The conf.py
and requirements.in have been updated to use that extension instead of
the svg files.
[CK] Load ck.groovy via Jenkins Shared Library
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
This allows the CI service to have a configuration source-of-truth
outside the PR under test, allowing rapid system changes. Bug fixes on
the develop branch propagate immediately to all pipelines that don't
override the parameter -- no rebase required.
A new `USE_CURRENT_BRANCH_FOR_CK_GROOVY` parameter lets contributors
test pipeline changes on their own branch without any extra
configuration.
## Technical Details
- `loadCk()` in the Jenkinsfile is updated to call
`library("ck@${branch}").ck.get()` instead of `checkout scm` + `load
"vars/ck.groovy"`. The `checkout scm` inside `loadCk()` is removed since
Jenkins now handles the library fetch internally.
- A `USE_CURRENT_BRANCH_FOR_CK_GROOVY` boolean parameter (default: off)
is added. When off, `ck.groovy` is always loaded from `develop` — all
normal PR builds are unaffected. When on, `ck.groovy` is loaded from the
current branch automatically via `env.CHANGE_BRANCH`, so contributors
testing pipeline changes just tick the box.
- `return this` is removed from the end of `ck.groovy`. This was
required by the `load` convention but is not needed (and can cause
errors) in a shared library context.
- `loadCk()` is kept at every call site rather than called once at the
top, preserving restart-from-stage safety — if a build is restarted from
a mid-pipeline stage, `ck` is still initialized correctly.
- The Jenkins Shared Library named `"ck"` must be registered in Jenkins
Global Pipeline Libraries
## Test Plan
1. Trigger "Build with Parameters" on the PR branch with
`USE_CURRENT_BRANCH_FOR_CK_GROOVY=true`
2. Verify "Determine CI Execution" stage completes and the library()
calls indicates the current branch
3. Verify "Static checks" stage completes.
4. Trigger a second build with `USE_CURRENT_BRANCH_FOR_CK_GROOVY=false`
(default) to confirm normal builds still load from `develop`.
## Test Result
Verified both paths. The develop library is loaded by default, the
branch library is loaded when the parameter is enabled.
## Submission Checklist
- [ X ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK] Fix latest build issues with staging compiler.
## Motivation
Fixing new warnings with staging compiler.
## Technical Details
<!-- Explain the changes along with any relevant GitHub links. -->
## Test Plan
<!-- Explain any relevant testing done to verify this PR. -->
## Test Result
<!-- Briefly summarize test outcomes. -->
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
