CANN/cannbot-skills工具目录

Tool Catalog

【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体，本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skills

Use this file to choose a repository tool before opening or modifying the script itself. This is a selection index, not an implementation reference.

Index schema

This file is also the machine-readable metadata source foragent/index/tools.json. The index builder reads:

• each###entry heading as one tool record
• the surrounding##section as the tool category
• top-level entry bullets such aspurpose,use_for,not_for,important_outputs, andpair_with
• nested bullet items under those fields as ordered list values

If you edit this catalog, keep that field structure stable.

Current tools

agent/scripts/estimate_matmul_datamove.py

• purpose:
  • estimate matmul data movement for candidate tile/core-split strategies
  • reject illegal tile-space combinations before kernel authoring goes too far
• use_for:
  • choosingTILE_M,TILE_N,TILE_K
  • comparingsplit_m,split_n, andmix
  • checkingdbuf_left,dbuf_right, and fixeddbuf_l0ccapacity assumptions
  • resolving large-matmul strategy questions before touching kernel code
• not_for:
  • launching kernels
  • validating numerical correctness
  • replacing simulator-based output checks
• important outputs:
  • per-core datamove estimate
  • best strategy candidate set
  • PrettyTable strategy report
  • expansion ratio and DBUF-aware capacity decisions
• pair_with:
  • agent/references/constraints/tiling.md
  • agent/scripts/tools_summary.md
  • large tiled examples such asagent/example/kernels/a5/matmul_mknk_2dgrid_splitn.py,agent/example/kernels/a5/matmul_mknk_2dgrid_splitk.py, andagent/example/kernels/a5/matmul_mknk_2dgrid_splitk_add1.py

agent/scripts/select_kernel_example.py

• purpose:
  • rank existing kernel examples for a new task using the generated kernel index
  • reduce manual catalog and source-file scanning before kernel authoring
• use_for:
  • selecting a first kernel to study by topology, query text, tags, or lightweight features
  • narrowing candidate examples before openingagent/example/kernels/source files
  • surfacingstudy_foranddo_not_copy_whenguidance together with the example path
• not_for:
  • generating kernel logic
  • proving that one example is the uniquely correct template
  • replacing constraint/reference reading for tiling, autosync, counters, or precision
• important outputs:
  • ranked kernel paths
  • short match reasons
  • study_foranddo_not_copy_whensummaries
  • optional JSON output for tool chaining
• pair_with:
  • agent/references/examples/kernel-catalog.md
  • agent/index/kernels.json
  • agent/playbooks/kernel-authoring.md
  • agent/references/constraints/

agent/scripts/check_counter_lifetimes.py

• purpose:
  • flag suspicious counter sharing and stage-lifetime mixing with static AST checks
  • give a fast pre-review pass before deeper manual reasoning or simulator work
• use_for:
  • spotting counters incremented at multiple loop-owned stages
  • spotting suspicious delayed-stage sharing across sibling conditional branches
  • summarizing which buffers and positions each counter indexes
  • adding a lightweight CI gate with--fail-on-warning
• not_for:
  • proving full kernel correctness
  • replacing autosync reasoning or simulator validation
  • deciding automatically whether every multi-buffer counter is semantically wrong
• important_outputs:
  • warning codes with line numbers
  • per-counter buffer and position summary
  • optional JSON output for tool chaining or CI
• pair_with:
  • agent/references/constraints/counters.md
  • agent/example/kernels/a5/matmul_mknk_2dgrid_splitn.py
  • agent/example/kernels/a5/matmul_mknk_2dgrid_splitk_add1.py
  • agent/example/kernels/a5/test_mla_entire.py

agent/scripts/check_kernel_catalog.py

• purpose:
  • verify that the human-readable kernel catalog, generated kernel index, and actualagent/example/kernels/*.pyfiles stay consistent
  • catch source-of-truth drift before selectors or routing layers start returning stale metadata
• use_for:
  • checking that every catalog entry points to a real kernel file
  • checking that required metadata fields are present and non-empty
  • checking that newly added kernel files are not missing from the catalog
  • checking whetheragent/index/kernels.jsonis stale relative to the current catalog
• not_for:
  • rewriting catalog entries automatically
  • ranking or selecting kernels for study
  • validating kernel numerical correctness
• important_outputs:
  • consistency warning codes such as missing file, missing field, uncataloged kernel, or stale index
  • optional JSON output for CI or tool chaining
• pair_with:
  • agent/references/examples/kernel-catalog.md
  • agent/index/kernels.json
  • agent/scripts/build_agent_index.py
  • agent/scripts/select_kernel_example.py

agent/scripts/gen_kernel_skeleton.py

• purpose:
  • generate repository-style kernel scaffolds so new kernel work can start from a controlled skeleton instead of peeling an old file by hand
  • encode the current pure-cube, cube->vec, vec->cube, vec->cube->vec, cube->vec->cube->vec, and vec->cube->vec->cube template structure directly into a reusable CLI tool
• use_for:
  • creating a newcube-only,cube->vec,vec->cube,vec->cube->vec,cube->vec->cube->vec, orvec->cube->vec->cubekernel scaffold underagent/example/kernels/
  • previewing grid-aware cube skeletons withtile-m,tile-n, ormix
  • choosing whether the scaffold should include a tiledKloop or a one-shot cube call
  • selecting a topology-specific profile such ascube-only --profile splitk,cube-only --profile kmkn,cube->vec --profile dual-output,cube->vec --profile normalize-two-pass,cube->vec --profile half-row-post,vec->cube --profile nz-publish,vec->cube --profile half-row-pre,cube->vec->cube->vec --profile lookahead-basic, orvec->cube->vec->cube --profile lookahead-basic
  • getting explicit counter ownership and tiled loop structure before filling real matmul logic
• not_for:
  • generating a semantically complete kernel body
  • deciding exact tile/core split values automatically
  • replacing the need to choose the correct topology first
• important_outputs:
  • generated kernel file or printed scaffold text
  • tiled pure-cube, cube->vec, vec->cube, vec->cube->vec, cube->vec->cube->vec, or vec->cube->vec->cube loop skeletons
  • CvMutexhandoff skeleton and vec postprocess placeholder forcube->vec, includingsimple-post, explicit split-Mhalf-row-post,normalize-simple,normalize-two-pass, anddual-outputvariants
  • VcMutexpublish skeleton and vec preprocess placeholder forvec->cube, includingnd-publish,nz-publish, and explicit split-Mhalf-row-prevariants
  • overlapped stage1/stage2 delayed-drain skeleton withVcMutex+CvMutexforvec->cube->vec, including explicitoverlap-basic,half-row-post, and deeper-queuedelayed-postvariants
  • lookahead-stylecube->vec->cube->vecstreaming skeleton withCvMutex+VcMutex+CvMutex, separate delayed-stage counters, and a loop-external finalize stage
  • mirroredvec->cube->vec->cubestreaming skeleton withVcMutex+CvMutex+VcMutex, streamed vec preprocess, delayed second-stage cube consume, and direct final cube writeback
  • __main__validation stub unless disabled
• pair_with:
  • agent/playbooks/kernel-authoring.md
  • agent/references/patterns/cube-only.md
  • agent/scripts/select_kernel_example.py
  • agent/scripts/estimate_matmul_datamove.py

Fast selection hint

If the question is "which tile/core split should I use before writing the kernel body?", start here:

• agent/scripts/estimate_matmul_datamove.py

If the question is "which existing kernel should I study before opening source files?", start here:

• agent/scripts/select_kernel_example.py

If the question is "does this kernel have obvious counter-lifetime smells before I go deeper?", start here:

• agent/scripts/check_counter_lifetimes.py

If the question is "did the kernel catalog or generated kernel index drift out of sync?", start here:

• agent/scripts/check_kernel_catalog.py

If the question is "give me a new pure-cube or mixed-pipeline kernel scaffold that follows the repository's tiled template", start here:

• agent/scripts/gen_kernel_skeleton.py

【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体，本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skills