mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-29 19:28:33 +00:00
60b276647b
[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.
Grouped Convolution — Python Examples
Examples and test harnesses for the grouped convolution dispatcher (forward, bwd_data, bwd_weight) using the Python JIT codegen + hipcc workflow.
Run scripts from this directory:
cd dispatcher/examples/grouped_conv/python
python3 -u <script.py> # use -u for unbuffered logs
GPU arch is auto-detected (detect_gpu_arch()); pass --arch gfx950 to override.
Examples
| Script | Purpose |
|---|---|
01_basic_grouped_conv.py |
End-to-end smoke test: build + run forward kernel, verify output. |
02_forward.py |
Forward variant (NHWGC / GKYXC), small 2D problem. |
03_bwd_data.py |
Backward-data variant. Runner contract: run(dY, W, prob). |
04_bwd_weight.py |
Backward-weight variant. Runner contract: run(X, dY, prob). |
05_benchmark.py |
Multi-kernel sweep + timing (slow; runs many configs). |
06_registry_json.py |
Build a registry from a JSON config file. |
09_ml_heuristic.py |
Demo of LightGBM heuristic (requires lightgbm); see ML heuristic below. |
10_test_all_pipelines.py |
For each variant, test all 8 pipelines with intrawave. |
11_test_schedulers.py |
For each variant, test all 8 pipelines × {intrawave, interwave}. |
12_test_config_options.py |
Test the 5 config options (see Config-options harness below). |
Runner argument contract
runner.run(input_np, weight_np, prob) — order matters per variant:
| Variant | input_np |
weight_np |
|---|---|---|
forward |
X (NHWGC) |
W (GKYXC) |
bwd_data |
dY |
W |
bwd_weight |
X |
dY |
Pipelines & schedulers
All 8 pipelines: basic_v1, mem, compv3, compv4, compv5, compv6, comp_async, basic_async_v1.
compv4andcomp_asyncrequiredouble_smem_buffer=True(loudstatic_assertotherwise).- Not every pipeline supports both
intrawaveandinterwave.11_test_schedulers.pytreats a pipeline as supported if at least one scheduler runs successfully.
Config-options harness (12_test_config_options.py)
Verifies the 5 GroupedConvKernelConfig options:
double_smem_buffer— LDS ping-pong (required for compv4 / comp_async).num_groups_to_merge— fuse groups into one tile.split_image— split spatial dims for large tensors.explicit_gemm— explicit GEMM path (experimental).two_stage— two-stage bwd_weight with fp32 workspace.
Each test is run in its own subprocess (--single-test '<json>' mode) so a
single GPU page fault doesn’t take down the whole sweep — failing combinations
are reported as CRASH and the run continues.
Test problem sizes are kept small (e.g. 2D: N=1, G=2, C=K=64, Hi=Wi=8, 3×3)
to avoid OOM / aperture violations on the test GPU.
ML heuristic (09_ml_heuristic.py)
LightGBM regression model that predicts kernel TFLOPS and selects a kernel for
a given problem. Requires the lightgbm Python package.
- Models live in
dispatcher/heuristics/models/grouped_conv_<variant>_bf16_<arch>/(forward, bwd_data, bwd_weight all available). - Feature engine:
dispatcher/heuristics/feature_engine_grouped_conv.py. - Training entry point:
dispatcher/heuristics/train.py. - Prediction:
dispatcher/heuristics/predict.py(usePredictorwithGroupedConvFeatureEngine; build the candidate kernel pool from a training/holdout parquet viadf["kernel_name"].unique()).
Typical training flow:
# 1. Benchmark to CSV (slow)
cd tile_engine/ops/grouped_conv
python3 -u grouped_conv_full_benchmark.py configs/forward_bf16.json \
--arch gfx950 --problems forward_training \
--csv benchmark_forward_bf16_gfx950.csv --workers 8
# 2. CSV → Parquet
cd ../../../dispatcher/heuristics
python3 convert_csv_to_parquet.py \
--input ../../tile_engine/ops/grouped_conv/benchmark_forward_bf16_gfx950.csv \
--output data/grouped_conv_forward_bf16_gfx950.parquet --arch gfx950
# 3. Train
python3 train.py --data_dir data \
--out_dir models/grouped_conv_forward_bf16_gfx950 \
--op grouped_conv --dtype bf16 --arch gfx950 --targets tflops --n_splits 5
To add a new pipeline (e.g. compv6) update:
dispatcher/codegen/grouped_config_rules.py (VARIANT_PIPELINES),
dispatcher/heuristics/feature_engine_grouped_conv.py (add the is_<name>
flag), and the relevant tile_engine/ops/grouped_conv/configs/*.json. Then
re-run the benchmark + train flow above.
Notes
- Use
python3 -ufor any long-running script so logs aren’t buffered. - Kernels are compiled once and cached under
/tmp/dispatcher/; subsequent runs reuse the cached.so. - This repo has 1 GPU — do not run benchmarks in parallel.