ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-13 17:55:48 +00:00
Code Issues Packages Projects Releases Wiki Activity
3,266 Commits 742 Branches 38 Tags
develop
Commit Graph

2 Commits

Author SHA1 Message Date
Yaswanth Raparti
6989cf800c [rocm-libraries] ROCm/rocm-libraries#6327 (commit 1e7a12e) 
[CK][CK TILE] Dispatcher kernel selection heuristic for
 grouped conv (#6327)

## Motivation
The ML heuristic in dispatcher does not support grouped-conv operator
yet. In this PR, the support for fwd, bdw-data, and bwd-weight
grouped-conv kernels have been added. A tile_engine utility has also
been added to compile and run any selected kernel configuration through
dispatcher infrastructure.

## Technical Details

1. Tile engine utility is added to benchmark each shape with all the
possible kernel+tile_size combinations here -
[https://github.com/ROCm/rocm-libraries/blob/users/yraparti/ck/dispatcher-grouped-conv-heuristics/projects/composablekernel/tile_engine/ops/grouped_conv/grouped_conv_full_benchmark.py](url)
2. New LGBM regressor models for grouped conv are added to models
directory. We have 3 separate models for fwd, bwd-data, and bwd-weights
[https://github.com/ROCm/rocm-libraries/tree/users/yraparti/ck/dispatcher-grouped-conv-heuristics/projects/composablekernel/dispatcher/heuristics/models](url)
3. Implemented lazy GPU initialization (dispatcher/python)
- **Issue**: ProcessPoolExecutor fork() + GPU context caused memory
access faults
- **Solution**: Mirror FMHA pattern - defer GPU initialization until
first run()
  - **Changes**:
- setup_multiple_grouped_conv_dispatchers() returns List[Path], not
loaded libs
    - GpuGroupedConvRunner.__init__() no longer calls ctypes.CDLL
    - Added _ensure_initialized() method for lazy GPU loading
    - GPU context created only on first run() call
  - **Benefit**: Parallel compilation now works without GPU conflicts
4. Addressed few miscellaneous issues such as:
  - Fixed BF16->FP16 naming bug in the dispatcher wrapper
- Added new tile sizes, and comp_v5 pipeline to the arch spec to expand
the kernel selection
- Added automatic padding support for unsupported shapes in dispatcher
runner
- Created a single source of truth between tile_engine and dispatcher
about the architecture and tile_size details
- Build a validation scripts to compare oracle_best vs ml_heuristic
comparison

## Test Plan

1. Validated fwd, bwd-data, and bwd-weight kernels with both known and
unseen data sets with up to 300 problems.
2. Ensured that test cases are added in both dispatcher and tile_engine
to validate the heuristic.

## Test Result
Results on Unseen shapes validated on gfx950
#### Forward Pass Model
- **Training Data**: 48,845 measurements across 1,372 unique problem
shapes
- **Validation Set**: 300 unseen problems from model crawler
- **Validation Performance** (vs. oracle):
  - Mean Efficiency: **93.05%**
  - Median Efficiency: **96.8%**
  - P10 Efficiency: **79.9%**

#### Backward Data Gradient (bwd_data) Model
- **Training Data**: 18,773 measurements across 891 unique problem
shapes
- **Validation Set**: 300 unseen problems from model crawler
- **Validation Performance** (vs. oracle):
  - Mean Efficiency: **93.8%**
  - Median Efficiency: **96.5%**
  - P10 Efficiency: **82.9%**

#### Backward Weight Gradient (bwd_weight) Model
- **Training Data**: 34,900 measurements across 1,508 unique problem
shapes
- **Validation Set**: 300 unseen problems from model crawler
- **Validation Performance** (vs. oracle):
  - Mean Efficiency: **96.1%**
  - Median Efficiency: **99.2%**
  - P10 Efficiency: **89.4%**

## Submission Checklist

- [ x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
 2026-05-08 20:48:42 +00:00
Yaswanth Raparti
c1127a36f5 [rocm-libraries] ROCm/rocm-libraries#5676 (commit 1d18339) 
[CK][CK TILE]Autotuning heuristics infra for universal GEMM
 kernel selection (#5676)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

## Motivation

This PR adds ML-based kernel selection heuristics to the CK Tile
dispatcher, enabling fast and accurate automatic kernel selection for
Universal Gemm kernels. Instead of requiring exhaustive search through
4600+ kernel configurations (taking ~46 seconds per problem shape), the
ML heuristic predicts optimal kernels in microseconds while achieving
>98% of oracle-best performance.

## Technical Details

**ML infrastructure**

https://github.com/ROCm/rocm-libraries/tree/users/vanantha/ck/dispatcher-heuristics/projects/composablekernel/dispatcher/heuristics
* Feature Engine
([feature_engine.py](https://github.com/ROCm/rocm-libraries/blob/users/vanantha/ck/dispatcher-heuristics/projects/composablekernel/dispatcher/heuristics/feature_engine.py)):
55-feature extraction including problem dimensions, kernel
configuration, tile efficiency, and hardware profile
* Training Pipeline
([train.py](https://github.com/ROCm/rocm-libraries/blob/users/vanantha/ck/dispatcher-heuristics/projects/composablekernel/dispatcher/heuristics/train.py)):
LightGBM regression with log-transform, GroupKFold cross-validation,
warm-start support
* Predictor
([predict.py](https://github.com/ROCm/rocm-libraries/blob/users/vanantha/ck/dispatcher-heuristics/projects/composablekernel/dispatcher/heuristics/predict.py)):
Kernel ranking and TFLOPS prediction for problem shapes
* Evaluation
([evaluate.py](https://github.com/ROCm/rocm-libraries/blob/users/vanantha/ck/dispatcher-heuristics/projects/composablekernel/dispatcher/heuristics/evaluate.py)):
Comprehensive metrics including efficiency, NDCG@k, shape family
analysis

**Data Generation Tools:**

*
[generate_benchmark_data.py](https://github.com/ROCm/rocm-libraries/blob/users/vanantha/ck/dispatcher-heuristics/projects/composablekernel/dispatcher/heuristics/generate_benchmark_data.py):
Build and benchmark kernels across diverse problem shapes
*
[convert_json_to_parquet.py](https://github.com/ROCm/rocm-libraries/blob/users/vanantha/ck/dispatcher-heuristics/projects/composablekernel/dispatcher/heuristics/convert_json_to_parquet.py):
Convert benchmark JSON to training-ready parquet format
*
[data_pipeline.py](https://github.com/ROCm/rocm-libraries/blob/users/vanantha/ck/dispatcher-heuristics/projects/composablekernel/dispatcher/heuristics/data_pipeline.py):
Parse streaming benchmark logs into canonical datasets

**Examples**
*
[09_ml_heuristic.cpp](https://github.com/ROCm/rocm-libraries/blob/users/vanantha/ck/dispatcher-heuristics/projects/composablekernel/dispatcher/examples/gemm/cpp/09_ml_heuristic.cpp):
C++ example demonstrating ML-based kernel selection
*
[09_ml_heuristic.py](https://github.com/ROCm/rocm-libraries/blob/users/vanantha/ck/dispatcher-heuristics/projects/composablekernel/dispatcher/examples/gemm/python/09_ml_heuristic.py):
Python example with validation

**Pre-trained Models
(projects/composablekernel/dispatcher/heuristics/models/):**
* gemm_universal_fp8_gfx950/: fp8 RCR model (42K trees, 97.51% mean
efficiency)
* gemm_universal_fp16_gfx950/: fp16 RCR model (20K trees, 99.36% mean
efficiency)

## Test Plan

* Evaluated on 25 diverse shapes for fp16, 168 shapes for fp8
* All shape families tested: tiny M (M<8), small M, medium M, large M
(M≥1024)
* All pipeline types: compv3, compv4, mem

## Test Result

**fp16 Model (gfx950, RCR layout)**
* Mean Efficiency: 99.36%
* P10 Efficiency: 98.05% (90th percentile of shapes achieve ≥98% of
oracle best)
* Min Efficiency: 95.45%

**fp8 Model (gfx950, RCR layout)**
* Mean Efficiency: 98.28% (original), 97.51% (wide coverage)
* P10 Efficiency: 94.64% (original), 93.89% (wide coverage)
* Min Efficiency: 84.5%

## Submission Checklist

- [x ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
 2026-04-02 02:26:32 +00:00