Question 1

How do I optimize a CATLASS matmul kernel on Ascend NPU?

Accepted Answer

Edit the DispatchPolicy, L1TileShape, L0TileShape, and BlockScheduler swizzle in catlass_kernel.asc. Choose the kernel family based on M/N/K shape and alignment, then tune tile sizes so block counts approach a multiple of the AIC core count.

Question 2

How to choose between Pingpong and Preload DispatchPolicy in CATLASS?

Accepted Answer

Use MmadAtlasA2Pingpong as the baseline for regular aligned GEMM. Try MmadAtlasA2Preload for large matrices to reduce data-movement bubbles, but note Preload adds pipeline overhead and can slow down small matrices, so confirm with profiling.

Question 3

When should I use Split-K kernels in CATLASS?

Accepted Answer

Use Split-K only when K is very large, a single core cannot saturate the K dimension, and multi-stage K reduction is actually enabled. If the configuration yields a split factor of 1, Split-K is equivalent to plain GEMM and gives no benefit.

Question 4

Why does tile tuning not improve my CATLASS kernel latency?

Accepted Answer

Latency is often measured end-to-end including dtype conversion and npu_format_cast, so if profiling shows time in Transdata rather than MMAD, tile changes have little effect. Instead edit catlass_torch.cpp to remove redundant format casts or copies.

Question 5

What happens if CATLASS tile shapes exceed L1 memory?

Accepted Answer

Exceeding L1 capacity causes compilation failure or silently selects a wrong configuration. Always compute L1/L0 usage against hardware constraints before submitting a configuration for evaluation.

catlass-matmul-optimization

SYSTEM DOCUMENTATION & REQUIREMENTS

Frequently Asked Questions about catlass-matmul-optimization

Related Skills

triton-cuda-matmul

rust-build-optimization

triton-ascend-matmul

triton-cuda-matmul