AI agent skills for Ascend C

CANNBot repository provides skills for AI agents with specialized knowledge about Ascend C API, NPU architecture, Tiling design, and Matmul/GEMM optimization. These skills give the AI coding assistant context-aware help with operator development, API usage, and explaining Ascend C specifics. The following instructions are recommended to use with OpenCode client.

Recommended Skills

Skill	Purpose
`ascendc-api-best-practices`	Comprehensive reference for all Ascend C API categories: arithmetic, reduction, sort, DataCopy, Transpose, Buffer management, precision conversion, pipeline operations, and API restrictions
`ascendc-blaze-best-practice`	Matmul/Cube/GEMM/BMM development via Blaze/tensor_api path. Covers AIC/StreamK/FixpOpti modes, dispatch modes, Tiling strategies, and workspace configuration
`ascendc-code-review`	Code review pipeline for Ascend C: file review, PR review, large PR handling, design consistency checks. Includes 11 reference rule sets covering API usage, performance, security, SIMT constraints, and C++ coding standards
`ascendc-crash-debug`	Debugging methodology for hang/crash/deadlock scenarios: symptom-to-cause mapping, Buffer deadlock diagnosis, cross-core sync issues, memory error detection via mssanitizer. Includes crash workflows and memcheck references
`ascendc-docs-search`	Local search across API documents and code examples from `asc-devkit`. Falls back to online Huawei Ascend Community search when local resources are insufficient
`ascendc-performance-best-practices`	Performance optimization patterns by operator family: MatMul, RadixSort, Scalar, Reduction, Elementwise. Covers pingpong, swap, streamk, fullload, and 8 optimization rules
`ascendc-perf-optimize`	4-step performance optimization strategy: Tiling modeling → inter-card pipeline → inter-core pipeline → intra-core pipeline. Bound diagnosis (compute/memory/communication) with Tiling parameter refinement. Self-contained tiling models for all operator categories
`ascendc-precision-debug`	Systematic precision issue diagnosis: pipeline sync (EnQue/DeQue), DataCopy alignment, Cast RoundMode, FP16/BF16 cross-validation, DumpTensor 7-step method. Includes error analysis scripts and boundary test generation
`ascendc-runtime-debug`	Runtime error debugging: comprehensive error code reference (161xxx/361xxx/561xxx), kernel binary debugging (compilation, cache, SEL matching, opParaSize), plog analysis. Includes debug workflows and error code tables
`ascendc-tiling-design`	Tiling methodology for all operator categories: Reduction, Sort, Elementwise, Broadcast, MatMul, Conv. Covers UB budgeting, Buffer strategies, and chunk sizing
`npu-arch`	NPU architecture knowledge: DAV_2201 vs DAV_3510, buffer sizes (UB/L0C/BT), Cube/Vector/MTE units, SIMD vs SIMD-RegBase, NDDMA, CCU
`ascendc-regbase-best-practice`	DAV_3510 RegBase (SIMD-RegBase) development: API constraints, implementation structure, common pitfalls, and real reference operators for the `RegTensor` / `MaskReg` / `asc_vf_call` / `__simd_vf__` paradigm
`ascendc-sync-audit`	Ascend C signal synchronization verification and fix: detects missing/mismatched sync between pipelines (MTE2/MTE3/Vector/Cube) and across cores (CrossCoreSetFlag/SyncAll), flag reuse conflicts, and buffer-index inconsistencies that cause hangs, deadlocks, or half-finished data; ships static analyzers (`sync_audit.py`, `ascendc_flow_analyzer.py`) plus a 333-PR case retriever for fix patterns
`ops-profiling`	On-board performance collection and analysis via `msprof`: standard / compare / quick / batch modes (`msprof_profile_run.sh`), bottleneck diagnosis (CUBE/VECTOR/MTE2/MTE1/FIXPIPE/SCALAR); MC2/multi-rank (fork) operators must use `msprof` (not `msprof op`)
`ops-simulator`	CANN Simulator (`npusim`, formerly `cannsim`) for functional and performance simulation without NPU hardware: Ascend 950 only, single-card, AI Core compute only (no MC2/HCCL); `summary.json` quick diagnosis and trace bubble analysis with PC-to-source mapping
`aiss-tiling-solver`	`TilingSolver` CLI for automatic optimal tiling of MatMul / Vector operators (Z3-based): `platform_info` collects hardware params (L0A/L0B/L0C/L1/UB/L2/HBM sizes, bandwidths, CUBE_OPS_PER_CYCLE) and the solver returns base/single-core tile sizes and double-buffer flags
`ops-precision-standard`	Operator precision standards (mixed tolerance atol/rtol) per dtype and operator category: random / non-compute / integer / quantization / float; used for ST verification and FP16/BF16/FP32 acceptance criteria

Directory Structure

The cannbot-skills repository should be cloned as a sibling directory to pyasc (not inside the project). Skills are symlinked into .opencode/skills/, the ops-code-reviewer agents into .opencode/agents/, and asc-devkit is symlinked in the project root. All paths use the $CANNBOT_SKILLS environment variable.

Installation

# Clone cannbot-skills as sibling to pyasc
export CANNBOT_SKILLS=$PWD/../cannbot-skills
git clone https://gitcode.com/cann/cannbot-skills.git $CANNBOT_SKILLS

# Clone asc-devkit (API docs + examples)
git clone https://gitcode.com/cann/asc-devkit.git $CANNBOT_SKILLS/plugins-official/ops-direct-invoke/asc-devkit

# Create symlinks
skills=(
    ascendc-api-best-practices
    ascendc-blaze-best-practice
    ascendc-code-review
    ascendc-crash-debug
    ascendc-docs-search
    ascendc-performance-best-practices
    ascendc-perf-optimize
    ascendc-precision-debug
    ascendc-runtime-debug
    ascendc-tiling-design
    npu-arch
    ascendc-regbase-best-practice
    ascendc-sync-audit
    ops-profiling
    ops-simulator
    aiss-tiling-solver
    ops-precision-standard
)
for skill in "${skills[@]}"; do
    ln -sfn $CANNBOT_SKILLS/ops/$skill .opencode/skills/$skill
done
ln -sfn $CANNBOT_SKILLS/plugins-official/ops-direct-invoke/asc-devkit ./asc-devkit

# ops-code-reviewer plugin: agent symlinks only (do not overwrite AGENTS.md)
mkdir -p .opencode/agents
for agent in ascendc-code-summarizer ascendc-ops-reviewer; do
    ln -sfn $CANNBOT_SKILLS/plugins-official/ops-code-reviewer/agents/$agent.md .opencode/agents/$agent.md
done

# Verify skills installation
for s in .opencode/skills/*/; do
  [ -f "$s/SKILL.md" ] && echo "OK: $s" || echo "MISSING: $s"
done

Restart OpenCode to load the skills.

Hiding Symlinks from Git

The symlinks created above (asc-devkit, skills under .opencode/skills/, and agents under .opencode/agents/) are external dependencies and should not be tracked by git. It is recommended to add them to the local .git/info/exclude file:

echo "
# CANNBot skills symlinks
asc-devkit
.opencode/skills/ascendc-*
.opencode/skills/npu-arch
.opencode/skills/ops-*
.opencode/skills/aiss-*
.opencode/agents/ascendc-*" >> .git/info/exclude

Updating

cd $CANNBOT_SKILLS && git pull
cd plugins-official/ops-direct-invoke/asc-devkit && git pull

Restart OpenCode to apply updates.

Alternative: Automated Installation

Use init.sh from a plugin directory to install all skills from that plugin:

cd $CANNBOT_SKILLS/plugins-official/<plugin-name>
bash init.sh project opencode /path/to/pyasc

See the cannbot-skills README for available plugins.