tirx-kernels
Guides TVM TIRx kernel authoring, execution scopes, layouts, tile primitives, lowering, GPU testing, and backend limits.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- tirx-kernels
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Guides TVM TIRx kernel authoring, execution scopes, layouts, tile primitives, lowering, GPU testing, and backend limits.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/gpu-testing.md2.5 KB
- 参考文档references/layout-and-primitives.md2.6 KB
- 参考文档references/native-workflows.md3.1 KB
- 参考文档references/troubleshooting.md1.9 KB
- 脚本scripts/monitor_gpu.sh3.1 KB
- 脚本scripts/tirx_layout_probe.py1.6 KB
03 SKILL.md 正文
TIRx Kernel Workflows
Use this route for TVM's native TIRx kernel DSL: from tvm.script import tirx as T, tvm.tirx.build, execution scopes, tensor layouts, tile primitives,
lowering, CUDA codegen decisions, and TIRx-specific tests.
Route
- Confirm the package imports with TIRx enabled. If not, use
../install-build/SKILL.md. - Classify the task: parser/printer, layout math, well-formedness verifier, tile primitive dispatch, lowering/codegen, or GPU execution.
- For layout or verifier questions, start with
scripts/tirx_layout_probe.pyandreferences/layout-and-primitives.md. - For author/compile/run loops, read
references/native-workflows.md. Usetvm.tirx.build(..., pipeline="default")ortvm.compile(..., tir_pipeline="tirx")only after the PrimFunc/IRModule is well formed. - For CUDA, Blackwell, or GPU CI tasks, read
references/gpu-testing.mdbefore running anything. Do not claim GPU or Blackwell coverage from CPU parser/layout tests. - Use
references/troubleshooting.mdfor parser, layout, scope, dispatch, target, or runtime failures.
API anchors
from tvm.script import tirx as T
import tvm.tirx
tvm.tirx.build(mod, target=None, pipeline="default")
tvm.compile(mod, target=None, tir_pipeline="tirx")
The verified layout/verifier anchors are TileLayout(spec),
ComposeLayout(per_element, swizzle_len, atom_len, tile_layout, swizzle_inner=True), and
verify_tirx_well_formed(obj, assert_mode=True, device_func=False).
Boundaries
- General install/build issues: install-build.
- Relax model/pipeline compilation: relax-compile.
- S-TIR schedules and meta-schedule: s-tir-tuning.
- RPC deployment of compiled modules: rpc-deployment.
- External
tirx-kernelsworkload registry: optional external dependency, not required by this generated skill.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).