mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-07-02 13:17:36 +00:00
2a71304bbb03dbfc7c83df7c6702d5faad065708
4 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Ville Pietilä
|
60b276647b |
[rocm-libraries] ROCm/rocm-libraries#8157 (commit b0d9d39)
[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. |
||
|
Ville Pietilä
|
78d657c4f7 |
[rocm-libraries] ROCm/rocm-libraries#7284 (commit e7d25b2)
[CK_TILE] Integrate CK Tile Dispatcher code generation into CK Tile Profiler (#7284) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Motivation CK Tile is going to be delivered to hipDNN via CK Dispatcher. Currently the CK Tile Profiler using CK Builder for generating the profiled instances from the configuration files that identify the instances that old CK exposes. We need to replace this instance generation with the CK Tile Dispatcher codegen. ## Technical Details The old CK Profiler config files are converted to JSON files that the CK Tile Dispatcher can digest. The conversion script for configurations is stored to source control in case we need to update the JSON configurations later. The dispatcher generates instance libraries per conv direction (fwd, bwd data, and bwd weight) that are linked to the CK Profiler executable. I also implemented codegne for the stream-K and depthwise conv instances. The proposed solution replaces the CK Builder codegen with the CK Tile Dispatcher codegen. There are two new methods that are exposed via the dispatcher backend - `is_supported` - required to enabled the profiler workflow where we check the applicability of the kernel instance before running it. - `get_instance_string` - this mainly for verification. This provide the CK Builder instance string for verifying that the old CK Builder based profiler and the new CK Tile Dispatcher based profiler have the same instances. The rules that limit the generated instances are now collected to a single location under the dispacther. The CK Builder codegen uses these, which ensures that the two codegen pipelines are in sync. The next step (different PR) is to remove the CK Builder codegen pipeline altogether. ## Test Plan Verified that the old CK Builder based profiler and the new CK Tile Dispatcher based profiler have the same instances, that is, the Dispatcher based codgen can generate the same instances as the old CK Builder. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Yaswanth Raparti
|
017dca1b9d |
[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. --------- Co-authored-by: Vidyasagar Ananthan <vidyasagar.ananthan@amd.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Jan Patrick Lehr <JanPatrick.Lehr@amd.com> |
||
|
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. |