[CK Tile] Async support pipeline V3
## Motivation
Optimize pipeline V3 for gfx950 by enabling buffer load to lds (async
pipeline)
## Technical Details
- Add `Async` bool to `Problem` struct to enable async pipeline in
existing one
- Add `static_move_ys` to load transpose. This generates offset in
assembly instructions saving registers
- Add `is_valid` to `async_get_vectorized_elements`. Before hard coded
to true. It allows to support padding
- Remove unnecessary restrictions to `is_a_load_tr` and `is_b_load_tr`
(wider use of lds load transpose on gfx950)
- Integrate async support in existing V3 pipeline (avoid pipelines
duplication)
- Create policy to support both async and default cases. This could be
used by any async pipeline (next steps)
- Define `wg_attr_num_access` separately for A and B. This allows to
optimize ds_read instruction width for cases when one matrix is
transposed and the other is not. Before in such cases, `ds_read_b64` was
used instead of `ds_read_b128`
- Add test for V3 async. Currently only supporting cases with A and B
having the same type
## Test Plan
New test `test_ck_tile_gemm_pipeline_compv3_async`
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK Tile] Rule-based configuration generation in CK
Dispatcher codegen (#8157)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
The CK Tile Dispatcher code generation for CK Tile Profiler relies on
flat JSON files to list the generated configurations. This approach has
the following problems
- The JSON files are verbose
- The JSON files get easily out of sync with the CK Builder .config
files from which they were generated from.
- The JSON file based configuration make it hard to list explicitly the
rules that govern the instance generation.
## Technical Details
Replaced the JSON files with a rule based configuration. To preserve the
existing functionality, the `profiler` and the `tests` instance sets are
generated directly from the CK Builder config files. The JSON config
files are removed from source control, and the "on-the-fly" generation
guarantees that the Dispatcher codegen uses up to date configurations.
This is PR introduces six different rule sets for the CK Tile Dispatcher
code generation
1. `profiler`: matches with the old JSON set of profiler configurations.
2. `tests`: matches with the old JSON set of tests configurations.
3. `full`: full configuration set created from a rule-based config
selection
4. `full-tests`: a subset of `full` for generating configurations for
convolution integration tests.
5. `tiny`: a subset of `full-tests` to produce the minimal set of
configurations to test the Dispatcher codegen.
6. `default`: the default rules, which corresponds to the existing
heuristic rules for configuration selection. This ensures that ML based
kernel selection doesn't get broken.
The main use of the `full` rule set is to define a reasonable solution
space for the possible implicit GEMM configurations. We start from the
configurations that allowed by the device architecture. The `full` rule
set defines the relevant tile sizes for each convolution direction. From
the tile size we have a curated mapping to the number of waves over the
different GEMM axes, i.e., we describe how many waves each GEMM
dimensions corresponds to. The GEMM-K wave tile dimension can be
computed from the other parameters and does not need to be listed
explicitly.
An orthogonal axis to the tiling strategy is the vectorization strategy.
This mainly defined by the data type and hardware as in general, we want
to use the maximum possible load widths. The maximum sizes for each
convolution direction variant are defined by the implicit GEMM matrix
dimensions. For cases where have a low number of channels per
convolution group, we need smaller vector load sizes. These are captured
by the `VecStrategy` enumeration in the codegen rules.
The problem with the rule based configuration selection is that we "over
generate" configurations. The old JSON configurations compose
approximately 25% of all configuration that the `full` rule set creates.
The additional configurations are valid, but they many not provide any
performance benefits. Hence, we keep the `profiler` and `tests` rule set
for now to avoid building an excessive amount configurations by default.
The `full` rule set can be taken into use by specifying CMake
configuration flag `-D DISPATCHER_RULE_SET=full`. By default, the
`tests` rule set is used, i.e., we don't change the existing bahaviour.
## Test Plan
Added a new stage in the CI/CD pipeline that ensures the Dispatcher
codegen rules are up to date. Otherwise the functionality is covered by
the existing CI/CD tests. There are no functional changes to the
convolution kernels. Only how the different instances are generated.
## Test Result
If the CK Tile conv instances build without errors, the Dispatcher
codegen is generating valid code. If all tests in CI/CD pipeline are
passing, the Dispatcher codegen generates valid instances.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Add tile size for FMHA batch prefill bf16 for MI308X
## Motivation
Adding a tile size adapted to MI308X, for the FMHA Batch Prefill BF16
input type case
## Technical Details
N/A
## Test Plan
Benchmarking from the Aiter side with:
```
python3 op_tests/test_batch_prefill.py -s 8000 -p 1 -q 4 -k 1 --head_dim 256 -c true -d bf16 --input_dtype bf16 --quant_method none --kv_layout linear -t sglang -l 0.0 --return_lse false --profile
```
## Test Result
We see an improvement with the new tile size on MI308X (both with PLT
mode OFF and ON)
## 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>
[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 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] 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][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
=?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] 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] Enforce ASCII-only C/C++ sources for hipRTC
compatibility (#7829)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Summary
CK source files must be compilable via **hipRTC (HIP runtime
compilation)**, whose preprocessor does not accept non-ASCII bytes
anywhere in a translation unit — **including in comments**. Bytes that
are harmless under `hipcc` (em-dashes, smart quotes, multiplication
signs, Greek letters, box-drawing glyphs, etc.) cause hipRTC to fail at
preprocessing time. These regularly leak in via LLM-assisted authoring
or copy/paste from formatted documents and silently break hipRTC paths
that are not exercised by the default `hipcc`-based build matrix.
This PR (a) cleans every existing violation (53 files) and (b) adds a
pre-checkin gate so new violations are rejected before merge.
## File extensions covered
Both the cleanup scan and the new Jenkins enforcement stage use the same
predicate:
```
*.h *.hpp *.cpp *.h.in *.hpp.in *.cpp.in *.inc *.cl
```
(excluding `*/build/*` and `*/include/rapidjson/*`). This is a strict
superset of the existing `Clang Format` stage's predicate — `*.inc` is
added so test-fixture include files are also gated. 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. `79798aa6261` — **`[ck] Convert reflect/ rendering to ASCII for
hipRTC compatibility`**
Behavior change, isolated. `TreeFormatter` swaps `├─ / └─ / │ ` for `|-
/ +- / | ` (3-col width preserved so alignment is unchanged).
`conv_description.hpp` swaps `×` for `x` as the dimension separator.
`test_conv_description.cpp` expected strings updated in lockstep so the
snapshot test stays green. This is the only commit in the series with
observable runtime impact.
2. `738fdb0d81c` — **`[ck] Strip non-ASCII bytes from C++ sources for
hipRTC compatibility`**
Mechanical text cleanup across 53 files. Replacements happen in comments
or in `std::cout` strings that are not asserted on by any test. None of
the 174 `.inc` files in the tree required edits, but they were in the
scan's predicate so the enforcement stage's predicate is a superset of
what was scanned. Full replacement table in the commit message.
3. `1d7cd8ba235` — **`[ck] Enforce ASCII-only C/C++ sources for hipRTC
compatibility`**
- New `projects/composablekernel/script/check_ascii_only.sh` (modeled on
`check_copyright_year.sh`).
- New entry in `projects/composablekernel/.pre-commit-config.yaml` under
the local-hooks block (`types_or: [c++, inc]`).
- New `ASCII Only Check` parallel stage in
`projects/composablekernel/Jenkinsfile`'s `Static checks` block,
mirroring the existing `Clang Format` stage but with `*.inc` added to
the find predicate. Always-on, no `RUN_CPPCHECK` gate.
The tree is buildable at every commit boundary. Commit 1 leaves 50 known
violations; commit 2 leaves 0; commit 3 wires the gate.
## Demo
Script output on a synthesized violation:
```
$ printf '// em-dash test \xe2\x80\x94 here\n' > /tmp/bad.cpp
$ projects/composablekernel/script/check_ascii_only.sh /tmp/bad.cpp
ERROR: /tmp/bad.cpp contains non-ASCII bytes:
1:// em-dash test — here
Fix: replace with ASCII (em-dash -> --, smart quotes -> ", arrows -> ->, etc.)
$ echo $?
1
```
Full repo scan after the cleanup commits (note the `-name '*.inc'`
clause):
```
$ 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_ascii_only.sh
$ echo $?
0
```
## Test plan
- [ ] Jenkins PR build: confirm new `Static checks -> ASCII Only Check`
stage runs green over the full predicate (incl. `*.inc`) and existing
`Clang Format` stage is unaffected.
- [ ] `test_conv_description` passes against the ASCII tree-formatter
output (touched in commit 1).
- [ ] Local: `pre-commit run ascii-only-checker --all-files` runs
cleanly after installing CK pre-commit hooks via
`script/install_precommit.sh`.
- [ ] Manually inject a non-ASCII byte in any `.cpp/.hpp/.inc` file,
push: confirm Jenkins fails the new stage with a clear error.
- [ ] Spot-check a representative subset of touched files under hipRTC
compilation to confirm no remaining hipRTC-blocking content (optional,
since the static byte check is a sufficient condition for hipRTC
preprocessor acceptance on this dimension).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
=?UTF-8?q?[CK=20TILE]=20Unification=20Work=20=E2=80=93=20?=
=?UTF-8?q?More=20accurate=20tests=20for=20MmaPipelines=20(#6212)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Motivation
This PR solves several issues:
#### More accurate tests for MmaPipelines
The current tests for the MmaPipelines (test_amdgcn_sparse_mma,
test_amdgcn_wavewise_mma) use explicit input fragment vectors filled
with 1s, and only check the output of a single lane. We should have
tests that actually use the MmaPipelines with non-trivial input matrices
and verify the complete output.
Some other aspects of the current MmaPipelines tests that I noticed and
deserve some attention:
1. There is sometimes iteration over K outside of the pipeline, which is
then included in WaveTileK or FragK, which is not correct. We should
remove it, move K iteration inside of the pipeline, or be more clear
about this outer-K loop size and how it propagates downwards.
2. There is very tight coupling between the kernel, gtest code, and
test_pipeline helper, requiring a lot of information and functions to be
passed back and forth.
3. The test_pipeline helper is doing a bunch of register-related logic
on the host (related to point 1)
4. Without this register logic the only thing it does is check the
device, call the kernel, and check the output, but with a lot of
boilerplate.
#### Test helper for detecting target arch at HOST runtime
There is a really apparent issue we faced while writing tests:
Scenario:
1. Compile a test that supports both gfx950 and gfx1201 for gfx950
2. Run the test on a server that only has gfx1201 GPU
Actual:
Segmentation fault
Expected:
The test can correctly detect from HOST runtime that the DEVICE
target_id was different and skips the test.
Notes:
The only way of detecting the COMPILER_TARGET_ID in the existing "arch"
framework is launching a kernel and calling `get_compiler_target()` (so,
from a DEVICE code). This will create a segmentation fault if the
current arch differs from the target arch. To cope with this issue, we
propose to export the compiler target(s) (note they can be many) through
`projects/composablekernel/test/ck_tile/core/arch/CMakeLists.txt` and
define a test helper to deal with such cases.
#### Add composition support to Transforms
We have a small number of Transforms which act on MmaOp input and output
data, before and after the MmaOp call respectively. These are currently
implemented to work on an MmaTile level, but in theory they are also
supposed to work at a WaveTile level, i.e. after composition of multiple
MmaTiles to create larger effective MNK dimensions. Currently the
composed MmaTiles look like 2D C-style arrays of the individual MmaTile
level register vectors (see WaveWiseMmaPipeline). The transforms should
be able to take these and perform the proper transforms to the whole
WaveTile at once. This might allow for better performing
transformations.
Note: This PR handles the SparseTransform case and if we don't end up
doing scale as a transformation, there isn't really much left to do. If
we end up having only the sparse transform as a non-trivial transform,
then we could also consider removing the Transform framework.
[CK_TILE][FMHA] Optimize long-context decoding on gfx11/12
(#7500)
## Motivation
Relevant issue: ROCM-22065
FMHA has less-than-optimal performance of long-context decoding (i.e.
when seqlen_q = 1) on gfx11/12.
This PR optimizes the splitkv pipeline and configs for such scenarios.
## Technical Details
Optimizations applied in this PR:
1. use tiles with smaller M0 (16 vs 64), these tiles are used when
seqlen_q <= 16
2. adapt qr_nwarp_sshuffle pipeline for gfx11, it allows to use more
warps even for M0 = 16 (the qr pipeline parallelizes work between warps
in M dim so with M0 = 16 it allows to use only 1 warp)
3. enable kMergeNumHeadGroupsSeqLenQ (an optimization that merges one
group of heads in GQA) for all hdim values, not only 128
4. increase the number of splits (multiply by the number of head groups)
if (3) is used
5. increase the number of splits for RDNAs (`multiProcessorCount` is the
number of WGPs on RDNAs, not CUs, so it should be doubled to have
meaning similar to CDNAs)
Performance on gfx1151:
| Case | develop (GB/s) | This PR (GB/s) |
|:-------|-------:|-------:|
| [fp16\|group\|bshd] b:1, h:32/32, s:1/45056, d:64/64 | 127.58 | 183.11
|
| [fp16\|group\|bhsd] b:1, h:32/32, s:1/45056, d:64/64 | 153.64 | 215.02
|
| [fp16\|group\|bshd] b:1, h:16/8, s:1/77184, d:128/128 | 120.51 |
225.76 |
| [fp16\|group\|bhsd] b:1, h:16/8, s:1/77184, d:128/128 | 130.62 |
223.84 |
| [fp16\|group\|bshd] b:1, h:32/32, s:1/9600, d:128/128 | 82.65 | 138.44
|
| [fp16\|group\|bhsd] b:1, h:32/32, s:1/9600, d:128/128 | 105.75 |
220.45 |
| [fp16\|group\|bshd] b:1, h:8/1, s:1/401024, d:256/256 | 16.27 | 187.89
|
| [fp16\|group\|bhsd] b:1, h:8/1, s:1/401024, d:256/256 | 16.28 | 188.19
|
## Test Plan
An additional test case is added to the exiting test. It uses seqlen_q =
1, GQA, no mask to trigger the changes
```
ninja test_ck_tile_fmha_fwd_fp16 && bin/test_ck_tile_fmha_fwd_fp16 --gtest_filter="*SplitKV*
ninja test_ck_tile_fmha_fwd_bf16 && bin/test_ck_tile_fmha_fwd_bf16 --gtest_filter="*SplitKV*
```
Manual testing can be done with these commands:
```
bin/tile_example_fmha_fwd -prec=fp16 -mode=1 -page_block_size=128 -b=1 -h=32 -h_k=32 -d=64 -s=1 -s_k=$((352 * 128)) -lse=1 -mask=0 -num_splits=0 -kname=1 -v=1
bin/tile_example_fmha_fwd -prec=fp16 -mode=1 -page_block_size=128 -b=1 -h=16 -h_k=8 -d=128 -s=1 -s_k=$((603 * 128)) -lse=1 -mask=0 -num_splits=0 -kname=1 -v=1
bin/tile_example_fmha_fwd -prec=fp16 -mode=1 -page_block_size=128 -b=1 -h=32 -h_k=32 -d=128 -s=1 -s_k=$((75 * 128)) -lse=1 -mask=0 -num_splits=0 -kname=1 -v=1
bin/tile_example_fmha_fwd -prec=fp16 -mode=1 -page_block_size=128 -b=1 -h=8 -h_k=1 -d=256 -s=1 -s_k=$((3133 * 128)) -lse=1 -mask=0 -num_splits=0 -kname=1 -v=1
```
## Test Result
All the tests must pass.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
[CK_TILE] Use gfx11 float buffer atomics in FMHA Bwd
## Motivation
FlashAttention CK backward on gfx11 can hit out-of-bounds/tail writes in
the dQ accumulator atomic-add path when sequence rows are padded at the
tile level but not marked invalid in the DQDKDV main tensor view.
With the generic global atomic fallback, an incorrectly-valid tail
element can issue an actual pointer-based `atomicAdd`. With the buffer
atomic path, the write is issued through a buffer resource with bounds
information and follows the same backend already used by gfx9/gfx12.
This fixes the gfx11 FMHA BWD failure without changing the gfx11 default
for unrelated CK Tile kernels.
## Technical Details
This PR enables the existing CK Tile AMD buffer float atomic-add path
only for generated FMHA BWD gfx11 translation units.
gfx11 normally uses the generic global atomic fallback for
floating-point `buffer_view::atomic_add`. That fallback performs the
atomic through a raw computed pointer and depends on the software
validity predicate to avoid invalid elements. In FMHA BWD dQ
accumulation, padded tail rows can reach this path, so using the buffer
atomic backend is safer: it uses a buffer resource with base pointer,
bounds information, and an element offset, matching the backend already
used by gfx9/gfx12.
Enabling `CK_TILE_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT` globally for gfx11 is
too broad and can break unrelated gfx11 CK builds such as GEMM. Instead,
`config.hpp` now preserves an explicitly pre-defined
`CK_TILE_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT`, while keeping the existing
default disabled for gfx11.
## Test Plan
Validated the change with the FlashAttention CK full test suite with
backward pass enabled on gfx11.
pytest -q -s tests/test_flash_attn_ck.py
## Test Result
FlashAttention CK gfx11 test result:
260680 passed, 152076 skipped
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
[CK_TILE][GFX1250] Enable MX GEMM FLATMM with ASYNC
## Motivation
Enables MX GEMM FLATMM pipeline on gfx1250. The pipeline uses an async
load instruction for tensor A, which complements the existing MX GEMM
FLATMM pipeline with TDM load. At this time, only FLATMM MX pipelines
are enabled on gfx1250.
## Technical Details
The existing gfx950 implementation was extended to support gfx1250
architecture. All three MX FP data types are supported across the two
ASICs.
It should be noted that while the TDM pipeline uses an emulated
32x32x128 warp-tile instruction, the present submission relies on the
built-in 16x16x128 instruction, called 4 times per warp.
## Test Plan
Existing `test/ck_tile/flatmm` tests were extended to cover new gfx1250
functionality.
To help facilitate the testing in development,
`example/ck_tile/18_flatmm/script/smoke_test_mx.sh` script was
introduced to verify various combinations of supported data types and
pipeline versions.
## Test Result
The present submission is expected to work on both gfx950 and gfx1250
hardware for all reasonable sizes and all MX FP8/FP6/FP4 data types.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
- [x] Relies on #6978 and should only be merged after the changes are
merged to the `develop`.
[CK] Replace deprecated load_module function in python
(#7833)
## Motivation
Recent pytorch builds with python 3.15 failed in CK due to deprecation
of load_module function. This should fix the issue.
## 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.
