mhc_res Performance

Optimization Strategy

• Multi-core parallel: Each AI Core handles one or more (batch, target_stream) pairs
• Dynamic tiling: Tile size computed based on UB size and dtype
• Double buffering: Overlapped data movement and computation

Tile Size Calculation

// 5 buffers: inQue(2) + outQue(2) + tmpBuf(1)
tile = UB_SIZE / (5 * sizeof(T))

For bf16: 6 buffers (extra fp32 accumulator)

Benchmark vs einsum

Config	einsum	AscendC	Speedup
batch=8, seq=1024, dim=512, ns=4	0.44ms	0.18ms	2.5x
batch=16, seq=256, dim=1024, ns=8	1.11ms	0.36ms	3.1x

Memory Bandwidth

• Read: batch × num_streams × seq × dim × sizeof(T)
• Read: num_streams × num_streams × sizeof(T) (weights)
• Write: batch × num_streams × seq × dim × sizeof(T)