mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-28 18:56:59 +00:00
276863ca874bedeee72fa8f46094de085c258aa6
15 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
BrianHarrisonAMD
|
f0545b5c15 |
[rocm-libraries] ROCm/rocm-libraries#8132 (commit 57d21a1)
[CK dispatcher] - LGBM predict data_type FLOAT32->FLOAT64 in ml_heuristic (#8132) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Summary `ml_heuristic.hpp` calls `LGBM_BoosterPredictForMat(..., /*data_type=*/0, ...)` (`C_API_DTYPE_FLOAT32`) against a `std::array<double, NUM_FEATURES>` feature buffer. LightGBM reinterprets the 8-byte doubles as 4-byte floats → invalid predictions → the heuristic's argmax always tie-breaks to the first/smallest enumerated config. **Fix:** `data_type 0 → 1` (`C_API_DTYPE_FLOAT64`), matching the `double` buffer. After the fix, predictions vary and track real TFLOPS (the model correctly prefers larger tiles). ## Verification - The feature buffer `f` is `std::array<double, NUM_FEATURES>` (NUM_FEATURES = 72) → `f.data()` is a `double*`. - The changed `0` is the 3rd positional `data_type` argument (not `nrow`/`ncol`/`is_row_major`). One-line functional change. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
Ville Pietilä
|
c6c55db757 |
[rocm-libraries] ROCm/rocm-libraries#8019 (commit 6472935)
[CK TILE] Fix performance regression caused by Dispatcher codegen compiler flag. (#8019) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Motivation Currently CK Tile two codegen paths: CK Builder and CK Tile Dispatcher. The CK Tile Dispatcher codegen uses an additional compiler flag that is not present in the CK Builder codegen workflow. The additional compiler flag can cause performance regression for so instances as it disables relevant compiler optimizations. ## Technical Details Removed compiler flag `-mllvm -enable-noalias-to-md-conversion=0` from the CMakeLists.txt that creates instance library from Dispatcher codegen. ## Test Plan Required testing is contained in the CI/CD pipeline. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
John Afaganis
|
96c39b331e |
[rocm-libraries] ROCm/rocm-libraries#7829 (commit 13af7da)
[ck] Enforce ASCII-only C/C++ sources for hipRTC compatibility (#7829) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Summary CK source files must be compilable via **hipRTC (HIP runtime compilation)**, whose preprocessor does not accept non-ASCII bytes anywhere in a translation unit — **including in comments**. Bytes that are harmless under `hipcc` (em-dashes, smart quotes, multiplication signs, Greek letters, box-drawing glyphs, etc.) cause hipRTC to fail at preprocessing time. These regularly leak in via LLM-assisted authoring or copy/paste from formatted documents and silently break hipRTC paths that are not exercised by the default `hipcc`-based build matrix. This PR (a) cleans every existing violation (53 files) and (b) adds a pre-checkin gate so new violations are rejected before merge. ## File extensions covered Both the cleanup scan and the new Jenkins enforcement stage use the same predicate: ``` *.h *.hpp *.cpp *.h.in *.hpp.in *.cpp.in *.inc *.cl ``` (excluding `*/build/*` and `*/include/rapidjson/*`). This is a strict superset of the existing `Clang Format` stage's predicate — `*.inc` is added so test-fixture include files are also gated. The local pre-commit hook's `c++/inc` type filter covers the same set. ## Why no enforcement today CK is opted out of the rocm-libraries root `.pre-commit-config.yaml`, so the existing `pre-commit` workflow doesn't touch CK. The local CK `.pre-commit-config.yaml` only runs for developers who installed hooks. The **authoritative gate is therefore the new Jenkins stage** in this PR; the local hook is convenience. ## Commit layout (bisect-friendly) 1. `79798aa6261` — **`[ck] Convert reflect/ rendering to ASCII for hipRTC compatibility`** Behavior change, isolated. `TreeFormatter` swaps `├─ / └─ / │ ` for `|- / +- / | ` (3-col width preserved so alignment is unchanged). `conv_description.hpp` swaps `×` for `x` as the dimension separator. `test_conv_description.cpp` expected strings updated in lockstep so the snapshot test stays green. This is the only commit in the series with observable runtime impact. 2. `738fdb0d81c` — **`[ck] Strip non-ASCII bytes from C++ sources for hipRTC compatibility`** Mechanical text cleanup across 53 files. Replacements happen in comments or in `std::cout` strings that are not asserted on by any test. None of the 174 `.inc` files in the tree required edits, but they were in the scan's predicate so the enforcement stage's predicate is a superset of what was scanned. Full replacement table in the commit message. 3. `1d7cd8ba235` — **`[ck] Enforce ASCII-only C/C++ sources for hipRTC compatibility`** - New `projects/composablekernel/script/check_ascii_only.sh` (modeled on `check_copyright_year.sh`). - New entry in `projects/composablekernel/.pre-commit-config.yaml` under the local-hooks block (`types_or: [c++, inc]`). - New `ASCII Only Check` parallel stage in `projects/composablekernel/Jenkinsfile`'s `Static checks` block, mirroring the existing `Clang Format` stage but with `*.inc` added to the find predicate. Always-on, no `RUN_CPPCHECK` gate. The tree is buildable at every commit boundary. Commit 1 leaves 50 known violations; commit 2 leaves 0; commit 3 wires the gate. ## Demo Script output on a synthesized violation: ``` $ printf '// em-dash test \xe2\x80\x94 here\n' > /tmp/bad.cpp $ projects/composablekernel/script/check_ascii_only.sh /tmp/bad.cpp ERROR: /tmp/bad.cpp contains non-ASCII bytes: 1:// em-dash test — here Fix: replace with ASCII (em-dash -> --, smart quotes -> ", arrows -> ->, etc.) $ echo $? 1 ``` Full repo scan after the cleanup commits (note the `-name '*.inc'` clause): ``` $ cd projects/composablekernel && find . -type f \( -name '*.h' -o -name '*.hpp' -o -name '*.cpp' \ -o -name '*.h.in' -o -name '*.hpp.in' -o -name '*.cpp.in' -o -name '*.inc' -o -name '*.cl' \) \ -not -path '*/build/*' -not -path '*/include/rapidjson/*' -print0 \ | xargs -0 -P 8 -n 64 script/check_ascii_only.sh $ echo $? 0 ``` ## Test plan - [ ] Jenkins PR build: confirm new `Static checks -> ASCII Only Check` stage runs green over the full predicate (incl. `*.inc`) and existing `Clang Format` stage is unaffected. - [ ] `test_conv_description` passes against the ASCII tree-formatter output (touched in commit 1). - [ ] Local: `pre-commit run ascii-only-checker --all-files` runs cleanly after installing CK pre-commit hooks via `script/install_precommit.sh`. - [ ] Manually inject a non-ASCII byte in any `.cpp/.hpp/.inc` file, push: confirm Jenkins fails the new stage with a clear error. - [ ] Spot-check a representative subset of touched files under hipRTC compilation to confirm no remaining hipRTC-blocking content (optional, since the static byte check is a sufficient condition for hipRTC preprocessor acceptance on this dimension). 🤖 Generated with [Claude Code](https://claude.com/claude-code) |
||
|
Illia Silin
|
8d97265896 |
[rocm-libraries] ROCm/rocm-libraries#7863 (commit 0845ce7)
[CK] apply the compiler warning suppression flags in cmake files (#7863) ## Motivation Apply the blanket suppression flags for latest clang warnings in staging compiler such as: lifetime-safety-lifetimebound-violation lifetime-safety-intra-tu-suggestions lifetime-safety-cross-tu-suggestions unknown-warning-option ## 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. |
||
|
Yaswanth Raparti
|
fe085f8a69 |
[rocm-libraries] ROCm/rocm-libraries#7761 (commit 237b766)
[CK][CK TILE] Clean up tile_engine grouped_conv harness (#7761) ## Motivation Tile_engine grouped_conv contains ML heuristic validation scripts which cause confusion to new developers. So, this PR is intended to relocate the scripts into dispatcher/heuristic directory to maintain separation of concern. ## Technical Details The grouped_conv tile_engine directory is a benchmarking harness for grouped convolution kernels; ML-heuristic content does not belong there. - Move compare_ml_vs_oracle.py and validate_ml_vs_oracle.py from tile_engine/ops/grouped_conv/ to dispatcher/heuristics/validation/grouped_conv/, and rebase their sys.path / oracle CSV / model dir lookups for the new location (CSV path is now an --oracle-csv flag instead of a hard-coded sibling). - Move GROUPED_CONV_HEURISTIC_REPORT.md (system-level ML report) into dispatcher/heuristics/ where the rest of the heuristic docs live. - Rewrite tile_engine/ops/grouped_conv/README.md as a pure benchmarking / dispatcher-sweep doc (kernel enumeration, JIT pipeline, CSV schema, problem registry), in the style of tile_engine/ops/fmha/README.md. All ML training / model-efficiency content is removed and replaced with a pointer to dispatcher/heuristics/. ## Test Plan Validation scripts are re-wired and tested locally ## Test Result Tests passed on local machine. ## 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. |
||
|
Illia Silin
|
e02c566795 |
[rocm-libraries] ROCm/rocm-libraries#7612 (commit 5427d24)
[CK] upgrade CI to rocm7.13 as default compiler (#7612) ## Motivation Upgrade the default docker and compiler version in CI to rocm7.13. In order to pass all the checks I had to also clean up a lot of non-ascii characters in the source code comments and modify a couple of tests that were affected by a new compiler logic. ## 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: Aviral Goel <aviral.goel@amd.com> |
||
|
Vidyasagar Ananthan
|
b20458e19e |
[rocm-libraries] ROCm/rocm-libraries#5260 (commit a1834d2)
[CK] [CK_Tile] Add FMHA scaffolding to CK kernel dispatcher (#5260) ## Motivation The CK Tile dispatcher currently supports GEMM and Grouped Convolution but has no support for Fused Multi-Head Attention (FMHA). The example/ck_tile/01_fmha folder contains a comprehensive FMHA implementation with forward, backward, split-KV, paged-KV, append-KV, and batch-prefill kernels across multiple GPU architectures — but there is no unified dispatch layer for it. This PR ports the FMHA stack into the dispatcher, following the same architectural patterns established by GEMM and Grouped Convolution, enabling runtime kernel selection, JIT compilation from Python, and a declarative C++ example flow. Autotuning heuristics to follow. ## Technical Details This PR adds FMHA scaffolding to the CK dispatcher framework, mirroring GEMM's layered architecture. Seven new C++ runtime headers provide type definitions (coexisting with upstream headers via __has_include, requiring zero modifications to example/ck_tile/01_fmha/), a problem builder with 18+ setters, Signature + Algorithm kernel key matching, a virtual kernel instance, a DECL_FMHA_KERNEL_SET macro with wildcard support and named tile/wave/warp setters, arch-aware registry with JSON export, and a dispatcher with seqtune-aware selection, configurable timing, and multi-stage execution plans for split-KV (two-stage) and backward (three-stage). The codegen pipeline is driven by a fmha_arch_specs.json capturing per-arch tile tables and pipeline constraints for five architectures (gfx90a/942/950/1100/1201), migrated from hardcoded logic in 01_fmha/codegen/, with supporting modules for C++ symbol mappings, validation rules, and named receipt profiles (ck_default, flash, pytorch, aiter, fp32, fp8). Python integration (fmha_utils.py) mirrors the C++ layer with JIT compilation, parallel multi-kernel builds, HIP memory management via ctypes, tolerance-based validation, and a NumPy CPU reference with GQA support. Twenty-seven C++ and thirty-two Python examples cover the full feature surface — forward, split-KV, masks, bias, dropout, GQA, backward, append-KV, batch prefill, fp8, logits soft cap, sink tokens, and parameter sweeps — all JIT-compiled on the fly. ## Test Plan Seven test files cover the runtime types, codegen, and end-to-end correctness. C++ unit tests validate the problem builder, dispatcher planning (single-stage for forward/paged-KV/append-KV; multi-stage for split-KV and backward), registry operations, and the kernel-set declaration macro. Python unit tests verify codegen emission, profile filtering, and 15 validation rules for masks, hdim constraints, and pipeline requirements. GPU execution validation in 01_basic_fmha --validate reports zero errors across 65,536 elements with max absolute error of 7.29e-05. A gold-standard parity suite (test_fmha_parity.py) runs 14 configurations through both the upstream tile_example_fmha_fwd and the dispatcher, comparing exit codes to confirm behavioral parity — all 14 match. ## Test Result The C++ smoke test builds and passes all 9 compiled examples, and a Python JIT sweep (29_sweep_seqlen.py) passes 7/7 configurations reaching up to 375 TFLOPS at seqlen 2048. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Yaswanth Raparti <113389104+yraparti@users.noreply.github.com> Co-authored-by: Mohsen Saffari <mohsen.saffari@amd.com> Co-authored-by: Maksim (Max) Podkorytov <Maksim.Podkorytov@amd.com> Co-authored-by: yashagar <yashagar@amd.com> |
||
|
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> |
||
|
Yaswanth Raparti
|
eb7cb6302d |
[rocm-libraries] ROCm/rocm-libraries#6528 (commit aa81df5)
[CK][CK_TILE] Fix dispatcher cpp tests - registry key mismatch and string assertions (#6528) ## Motivation CPP tests in dispatcher were failing due to a mismatch in registry key and string representation. ## Technical Details Bug 1 - Registry key mismatch: The registry stored kernels using get_name() but lookups used encode_identifier(), causing all registry lookups to fail. Fixed by changing registry.cpp:58 to use encode_identifier() for storage. Bug 2 - String representation changes: Tests checked for "persist"/"nopers" substrings, but the code emits "True"/"False". Fixed by replacing brittle substring checks with comparison-based assertions in test_kernel_key.cpp and test_kernel_key_extended.cpp. ## Test Plan Tested with CPP tests in dispatcher ## Test Result Validation: All three core cpp tests now pass: - test_kernel_key - 6/6 tests passing - test_kernel_key_extended - 25/25 tests passing - test_registry - 8/8 tests passing ## Submission Checklist - [ x] 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> |
||
|
Yaswanth Raparti
|
8fc06f80f9 |
[rocm-libraries] ROCm/rocm-libraries#6445 (commit 2225e10)
[CK][CK_TILE] Fix library caching bug in gemm dispatcher (#6445) ## Motivation setup_gemm_dispatcher() was rebuilding libraries on every call instead of reusing cached libraries. **Root Cause**: 1. Library names only included dtype+layout, causing different tile/wave/warp configs to overwrite each other 2. No cache checking - always loaded default library, detected mismatch, then rebuilt ## Technical Details **Solution**: 1. Complete library naming with all distinguishing parameters: libdispatcher_gemm_{dtype}_{layout}_{tile}_{wave}_{warp}_{pipeline}_{epilogue}_{scheduler}.so 2. Cache checking before rebuild: - Check if library for exact config already exists - Reuse if found (500x faster: 0.02s vs 10s) - Only rebuild when no cached library exists 3. Better error handling for kernel generation failures Files Changed: - dispatcher/python/ctypes_utils.py - dispatcher/tests/test_library_caching.py (new unit test) ## Test Plan Use `dispatcher/tests/test_library_caching.py ` to ensure that libraries are cached and only rebuilt if they are not present in build directory 1. **test_01_unique_library_naming** - Library names include all parameters (dtype, layout, tile, wave, warp, pipeline, epilogue, scheduler) 2. **test_02_library_build_and_cache** - Libraries are built once and then cached for reuse 3. **test_03_different_configs_different_libraries** - Different configs create different library files 4. **test_04_cache_message_verification** - Cache hit messages are logged correctly 5. **test_05_code_fix_verification** - Code changes are present in ctypes_utils.py ## Test Result All the test above passed. ## Submission Checklist - [ x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Vidyasagar Ananthan
|
ca28efac88 |
[rocm-libraries] ROCm/rocm-libraries#5168 (commit 8b5afcb)
[CK] [CK_Tile] Add GroupConv to Kernel Dispatcher (#5168) ## Motivation This PR adds CK Tile group convolution (forward, backward-data, backward-weight) support to the kernel dispatcher, matching and unifying with the existing dispatcher GEMM infrastructure in architecture and usability. The dispatcher provides a unified kernel dispatch system with both C++ and Python frontends, and until now only supported GEMM operations. This PR enables framework integrators to use the same declarative kernel workflow for convolutions as they do for GEMM: declare kernels, build a registry JIT, select kernels within the registry at runtime, and dispatch to GPU. Future PRs will include runtime kernel selection heuristics for autotuning of kernel parameters based on (problem, hardware arch). ## Technical Details Grouped convolution support has been added to the CK Tile Dispatcher with generated_conv_backend.hpp enabling dispatcher.run(in, wei, out, problem) for all 6 conv variants (fwd/bwdd/bwdw x 2D/3D), runtime heuristic kernel selection, and GroupedConvKernelKey with full ConvConfigBase fields. Python side adds parallel JIT via registry.build(max_workers) and heuristic registry.select(). Includes 7 C++ and 6 Python examples covering all directions with CPU reference validation, and shared infrastructure improvements (BaseRegistry CRTP, structured exceptions). As a sanity check, JIT compile times for a single kernel remains the same and for multiple kernels there is better parallelism: Kernels | 1 worker | 8 workers 1 | 7.7 s | 7.7 s 2 | 15.9 s | 8.2 s 4 | 33.4 s | 9.7 s 6 | 52.3 s | 10.2 s ## Test Plan 145 ephemeral unit tests have been added to test basic functionality. All 30 examples/integration tests run end-to-end on gfx950 (MI350): 7 C++ conv, 7 C++ GEMM, 6 Python conv, 10 Python GEMM. CPU reference validation for forward, backward-data, and backward-weight (2D) in both C++ and Python examples pass. ## Test Result 30 examples pass. Peak performance: 132 TFLOPS (Batch-32 forward 56x56), 53 TFLOPS (pointwise 1x1). CPU reference accuracy: max_abs_diff < 0.002 for all directions (fp16 vs fp32 reference). ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Yaswanth Raparti <113389104+yraparti@users.noreply.github.com> |
||
|
Yaswanth Raparti
|
644fc05a87 |
[rocm-libraries] ROCm/rocm-libraries#5676 (commit 1d18339)
[CK][CK TILE]Autotuning heuristics infra for universal GEMM kernel selection (#5676) ## 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. --------- Co-authored-by: Vidyasagar Ananthan <vidyasagar.ananthan@amd.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
John Afaganis
|
91014bc2d9 |
[rocm-libraries] ROCm/rocm-libraries#5071 (commit 444cd13)
Add Operation Support Matrix to Dispatcher README (#5071) Added an Operation Support Matrix to the Dispatcher README, detailing CK Tile operations with support status for various data types, layouts, and GPU targets. ## Motivation Provide a clear understanding of which operators (and variants) are supported by dispatcher. ## Technical Details Entirely generated by a skill. ## Test Plan N/A. This is a documentation-only change. ## Test Result N/A. This is a documentation-only change. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
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. |