[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] suppress compiler warnings while building pytorch. (#7760)
## Motivation
Recently added compiler flags that are required to suppress false
warnings by latest staging compiler are not recognized by older compiler
versions and are triggering an avalanche of warnings. Previous attempt
to suppress them by using -Wno-unknown-warning-option flag didn't help,
because that flag wasn't recognized either and just added more warnings.
I've verified that current approach by checking the clang version
actually works as intended and makes the warnings go away.
## 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 Engine] Daily tier sampling for tile engine GEMM (#7311)
Summary
- Replace uniform random instance sampling (random.shuffle) with
scrambled Sobol + Latin Hypercube + maximin space-filling
sampling, per the Tile Engine Benchmark Sampling RFC
- Add op-weighted budget allocation via new
TILE_ENGINE_SAMPLING_TIER=daily CMake knob that auto-distributes 8,000
instances across
ops proportional to registered weights in op_weights.json
- Emit chosen_instances.json manifests for reproducibility tracking
- Consolidate 5 copies of sampling logic into single _apply_sampling()
method on the base class
Jenkinsfile changes
Replace per-op -D *_MAX_INSTANCES=250 with single -D
TILE_ENGINE_SAMPLING_TIER=daily in gfx942/gfx950/gfx1201 stages. Budget
auto-distributes (8000 total per GPU target).
---------
Co-authored-by: Claude Sonnet 4 <noreply@anthropic.com>
[CK] Suppress new staging compiler errors (#7384)
## Motivation
This should make new builds with staging compiler pass.
## 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] Add pooling in tile_engine (#4469)
## Motivation
<!-- Explain the purpose of this PR and the goals it aims to achieve.
-->
Add pooling in ck tile engine
## 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.
---------
Co-authored-by: Adam Osewski <19374865+aosewski@users.noreply.github.com>
[CK TILE ENGINE] Add grouped_gemm operator to Tile Engine (gfx942/gfx950) (#4996)
## Motivation
The grouped_gemm CK Tile kernel exists (e.g.,
`example/17_grouped_gemm/`) but has no Tile Engine wrapper. Grouped GEMM
handles multiple independent GEMM problems with varying M/N/K dimensions
in a single kernel launch. This PR adds the Tile Engine infrastructure
for automated kernel generation, benchmarking, and profiling of grouped
GEMM kernels.
Jira: AICK-809
## Technical Details
- Created Tile Engine wrapper under `tile_engine/ops/gemm/grouped_gemm/`
following the `gemm_universal` template
- Files added: `CMakeLists.txt`, `grouped_gemm_common.hpp`,
`grouped_gemm_benchmark.hpp`, `grouped_gemm_profiler.hpp`,
`grouped_gemm_benchmark.py`, `grouped_gemm_benchmark_single.cpp`,
`grouped_gemm_instance_builder.py`, `configs/`
- Supported datatypes: fp16, fp8, bf16, bf8
- Supported layouts: rcr, rrr, ccr, crr
- Target GPUs: gfx942, gfx950
- CK Tile kernel: `ck_tile::GroupedGemmKernel` from
`include/ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp`
- Instance builder extends `GemmKernelBuilder` base class
- Registered in `tile_engine/ops/gemm/CMakeLists.txt`
- Updated Jenkinsfile to build and benchmark grouped_gemm targets in CI
- Benchmark infrastructure includes JSON output, CSV export, and
verification support
## Test Plan
- CMake configure succeeds for grouped_gemm targets
- Kernel instance builder generates valid kernel headers for all
(datatype, layout) combinations
- At least one kernel binary compiles and runs per datatype/layout
combination
- Correctness passes with `--verify 1` on gfx942/gfx950
## Test Result
## Submission Checklist
- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
