mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-10 16:28:38 +00:00
[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)
CK Tile Distribution Encoding Tutorial
Overview
Every load_tile and store_tile in CK needs to know which thread reads which data element.
This mapping is defined by a tile_distribution_encoding — a compile-time struct with 6 template
parameters:
tile_distribution_encoding<Rs, Hs, Ps_major, Ps_minor, Ys_major, Ys_minor>
Every level of Hs (hierarchical dimensions) is assigned to exactly one role:
| Role | Meaning |
|---|---|
| P (parallel) | Thread ID selects which slice — different threads get different data |
| Y (yield) | Each thread owns the entire range in its buffer |
| R (replicate) | Identical data broadcast to multiple thread groups |
Tutorials
These tutorials use the exact tile sizes from the naive GEMM tutorial
(01_naive_gemm/): MPerBlock=256, NPerBlock=128, KPerBlock=32, BlockSize=256, fp16.
| # | File | Matrix | Tile | Key Concept |
|---|---|---|---|---|
| 1 | tile_distribution_1.cpp |
A (DRAM load) | 256×32 | NDimP=2, warp_id→M1, lane_id→M2×K0 (coalesced) |
| 2 | tile_distribution_2.cpp |
B (DRAM load) | 128×32 | Same pattern as A, but N0=2 iterations (vs A's M0=4) due to smaller N |
| 3 | tile_distribution_3.cpp |
C (registers) | 256×128 | Warp-level MFMA output + block-level composition, standard vs transposed |
Tutorial 3 responds to CK_TILE_ENABLE_TRANSPOSED_C_DISTRIBUTION — rebuild with =0 or =1
to see both C register layouts.
Architecture note: All comments and concrete values assume CDNA (warp_size=64). On RDNA (warp_size=32), the thread-to-data mapping will differ.
Building
cd <repo-root>/projects/composablekernel/build
# Build all tutorials:
make tutorials -j
# or: ninja tutorials
# Or build individually:
make tile_tutorial_tile_distribution_1 -j
make tile_tutorial_tile_distribution_2 -j
make tile_tutorial_tile_distribution_3 -j
# Tutorial 3 with standard (non-transposed) C:
cmake -DCMAKE_CXX_FLAGS="-DCK_TILE_ENABLE_TRANSPOSED_C_DISTRIBUTION=0" ..
make tile_tutorial_tile_distribution_3 -j
Reference
- Encoding definition:
include/ck_tile/core/tensor/tile_distribution_encoding.hpp - Thread identity (NDimP):
include/ck_tile/core/tensor/tile_distribution.hpp - MFMA warp output layout:
include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma.hpp - Production A/B distributions:
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp - Naive GEMM tutorial:
tutorial/ck_tile/gemm/01_naive_gemm/