composable_kernel

ROCm/composable_kernel

Fork 0

mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 09:16:52 +00:00

Commit Graph

Author	SHA1	Message	Date
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
Vidyasagar Ananthan	9e049a32a1	Adding dispatcher architecture (#3300 ) * WIP POC of dispatcher * Dispatcher python workflow setup. * Dispatcher cleanup and updates. Further dispatcher cleanup and updates. Build fixes Improvements and python to CK example Improvements to readme * Fixes to python paths * Cleaning up code * Improving dispatcher support for different arch Fixing typos * Fix formatting errors * Cleaning up examples * Improving codegeneration * Improving and fixing C++ examples * Adding conv functionality (fwd,bwd,bwdw) and examples. * Fixes based on feedback. * Further fixes based on feedback. * Adding stress test for autogeneration and autocorrection, and fixing preshuffle bug. * Another round of improvements based on feedback. * Trimming out unnecessary code. * Fixing the multi-D implementation. * Using gpu verification for gemms and fixing convolutions tflops calculation. * Fix counter usage issue and arch filtering per ops. * Adding changelog and other fixes. * Improve examples and resolve critical bugs. * Reduce build time for python examples. * Fixing minor bug. * Fix compilation error. * Improve installation instructions for dispatcher. * Add docker based installation instructions for dispatcher. * Fixing arch-based filtering to match tile engine. * Remove dead code and fix arch filtering. * Minor bugfix. * Updates after rebase. * Trimming code. * Fix copyright headers. * Consolidate examples, cut down code. * Minor fixes. * Improving python examples. * Update readmes. * Remove conv functionality. * Cleanup following conv removable.	2026-01-22 09:34:33 -08:00

Author

SHA1

Message

Date

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

Vidyasagar Ananthan

9e049a32a1

Adding dispatcher architecture (#3300 )

* WIP POC of dispatcher

* Dispatcher python workflow setup.

* Dispatcher cleanup and updates.

Further dispatcher cleanup and updates.

Build fixes

Improvements and python to CK example

Improvements to readme

* Fixes to python paths

* Cleaning up code

* Improving dispatcher support for different arch

Fixing typos

* Fix formatting errors

* Cleaning up examples

* Improving codegeneration

* Improving and fixing C++ examples

* Adding conv functionality (fwd,bwd,bwdw) and examples.

* Fixes based on feedback.

* Further fixes based on feedback.

* Adding stress test for autogeneration and autocorrection, and fixing preshuffle bug.

* Another round of improvements  based on feedback.

* Trimming out unnecessary code.

* Fixing the multi-D implementation.

* Using gpu verification for gemms and fixing convolutions tflops calculation.

* Fix counter usage issue and arch filtering per ops.

* Adding changelog and other fixes.

* Improve examples and resolve critical bugs.

* Reduce build time for python examples.

* Fixing minor bug.

* Fix compilation error.

* Improve installation instructions for dispatcher.

* Add docker based  installation instructions for dispatcher.

* Fixing arch-based filtering to match tile engine.

* Remove dead code and fix arch filtering.

* Minor bugfix.

* Updates after rebase.

* Trimming code.

* Fix copyright headers.

* Consolidate examples, cut down code.

* Minor fixes.

* Improving python examples.

* Update readmes.

* Remove conv functionality.

* Cleanup following conv removable.

2026-01-22 09:34:33 -08:00

2 Commits