CANN Operator Generation

Generate optimized NPU kernels by writing solution.json.

Workflow

Step 1: Research APIs

cann_get_knowledge()              # List all API categories
cann_search_api("DataCopyPad")    # Get API signature → returns header_file path
cann_search_operator("avg_pool")  # Find similar implementations → returns file paths

IMPORTANT: These tools return file paths, not code. Use the Read tool to view the actual content.

Example workflow:

1. cann_search_operator("avg_pool") → {"primary": {"kernel_files": ["/path/to/kernel.h"]}}
2. Read("/path/to/kernel.h") → see actual implementation

Step 2: Read Input Files

1. •constraints.md - Code template structure & JSON format requirements
2. •vector.md or cube.md - Hardware specs & critical rules
3. •signature.json - Operator interface (inputs, outputs, params)
4. •python_reference.py - Reference implementation
5. •solution_template.json - Example JSON format

Step 3: Pre-flight Checks

Before writing code, verify:

Output dimensions alignment:

• •If output width/height is NOT a multiple of 8, you MUST use DataCopyPad instead of DataCopy
• •Example: outW=46 → use DataCopyPad to write exactly 46 elements

Unfamiliar APIs:

• •If using an API you're unsure about, call cann_search_api("ApiName") first
• •Then Read the returned header_file to see the exact signature

Step 4: Write solution.json

CRITICAL: Use the Write tool to create solution.json

All 6 fields required:

• •kernel_impl - Kernel class definition
• •kernel_entry_body - Instantiate and call kernel
• •tiling_fields - JSON array [{"type": "T", "name": "N"}, ...]
• •tiling_func_body - Host-side tiling calculation
• •infer_shape_body - Output shape inference
• •output_alloc_code - C++ code at::Tensor result = ...;

Quick Reference

Operator Naming

For operator foo_bar:

• •Kernel class: KernelFooBar
• •Tiling class: FooBarCustomTilingData

Buffer Initialization

// TQue: 3 arguments (que, BUFFER_NUM, size)
TQue<QuePosition::VECIN, 2> inQue;
pipe.InitBuffer(inQue, 2, bufferSize);

// TBuf: 2 arguments only (buf, size) - NO BUFFER_NUM!
TBuf<QuePosition::VECCALC> tmpBuf;
pipe.InitBuffer(tmpBuf, scratchSize);

Data Transfer

// ✅ Always use DataCopy for GM↔UB (uses DMA)
DataCopy(localTensor, xGm[offset], count);

// ⚠️ Avoid GetPhyAddr() for data transfer - very slow
// float val = *((__gm__ float*)xGm.GetPhyAddr() + idx);  // ~100x slower!

Error Troubleshooting