Linear Attention & SSM

2 ops, 10 workloads.

One table per op, one row per workload. Ratio is the baseline device time divided by ours in the same measurement regime, so green is faster than it, plain is level with it, red is slower. Times are in ms. How these numbers are taken.

Scan

CumprodFwd

• W1x: [2048, 4096]dtype=bf16hidden-state-scan
• W2x: [2048, 4096]dtype=f16hidden-state-scan
• W3x: [64, 32768]dtype=bf16long-seq-scan

dtype=f16

• W4input: [2, 5, 7]dim=1non-last-axis-host-permute
• W5input: [4097]nondivisible-4097

Workload	Ratio	Device time	Alternatives		Throughput	SOL	Bound
	alt / ours	ms	name · tier	ms	TFLOP/s	of ceiling	by
W1	16.39×	8.1635	torch_compile_aclgraph torch_compile basis torch_npu eager vendor library	133.9033 134.0408	·	·	·
W2	16.44×	8.1628	torch_compile_aclgraph torch_compile basis torch_npu eager vendor library	134.0186 134.1221	·	·	·
W3	13.70×	2.0857	torch_compile_aclgraph torch_compile basis torch_npu eager vendor library	28.5701 28.5780	·	·	·
W4	1.37×	0.0232	torch_compile_aclgraph torch_compile torch_npu eager vendor library basis	0.0333 0.0309	·	·	·
W5	3.36×	0.0267	torch_compile_aclgraph torch_compile torch_npu eager vendor library basis	0.0899 0.0882	·	·	·

CumsumFwd

• W1x: [2048, 4096]dtype=bf16hidden-state-scan
• W2x: [2048, 4096]dtype=f16hidden-state-scan
• W3x: [64, 32768]dtype=bf16long-seq-scan

dtype=f16

• W4input: [2, 5, 7]dim=1non-last-axis-host-permute
• W5input: [4097]nondivisible-4097

Workload	Ratio	Device time	Alternatives		Throughput	SOL	Bound
	alt / ours	ms	name · tier	ms	TFLOP/s	of ceiling	by
W1	0.05×	8.1691	torch_compile_aclgraph torch_compile torch_compile_ge torch_compile torch_npu eager vendor library basis	0.4340 0.3981 0.3949	·	·	·
W2	0.05×	8.1645	torch_compile_aclgraph torch_compile torch_compile_ge torch_compile torch_npu eager vendor library basis	0.4339 0.3956 0.3940	·	·	·
W3	0.07×	2.0892	torch_compile_aclgraph torch_compile torch_compile_ge torch_compile torch_npu eager vendor library basis	0.1686 0.1424 0.1446	·	·	·
W4	0.86×	0.0238	torch_compile_aclgraph torch_compile torch_compile_ge torch_compile torch_npu eager vendor library basis	0.0120 0.0107 0.0107	·	·	·
W5	0.96×	0.0283	torch_compile_aclgraph torch_compile torch_compile_ge torch_compile torch_npu eager vendor library basis	0.0272 0.0125 0.0185	·	·	·