Performance regression in causal_conv1d

[ChristosMatzoros]

PR #592 replaced the merged causal_conv1d kernel (#572) without benchmarking against it, causing degradation.

Timeline

Date	PR	Name	Description
2026-06-15	#555		"add gdn custom conv1d" (@zhaozx-cn) opened — still open, never merged
2026-06-25	#572	Variant PTO (#572)	"causal_conv1d: PTO-ISA rewrite" (@ChristosMatzoros) opened
2026-07-05	#572	Variant PTO (#572)	Variant PTO merged into main by @ping1jing2
2026-07-09	#592	Variant ASC (#592)	"Custom causal conv1d" (@zhaozx-cn) opened in a new PR (same code as #555)
2026-07-10	#592	Variant ASC (#592)	Variant ASC merged into main by @iforgetmyname, replacing Variant PTO

Variant PTO (the PTO-ISA causal_conv1d kernel) was merged into main on 2026-07-05. Four days later, Variant ASC was merged (2026-07-10) and replaced it with the implementation from the still-open #555, reusing that PR's description and benchmark tables unchanged, and removing the merged kernel entirely (main today contains none of it). The creators of Variant ASC never benchmarked against Variant PTO, the kernel actually on main at the time. Its performance tables compare against an older, superseded baseline, not the incumbent it replaced.

Note that Variant PTO had already been benchmarked head-to-head against #555 as part of its own review, where that older kernel measured slower. So Variant ASC re-submitted an implementation that had already been compared against Variant PTO and measured slower, without re-running that comparison. Re-measuring today, the kernel now on main is slower than Variant PTO, so this merge was a performance regression.

Head-to-head: Variant ASC vs VariantPTO

We benchmarked the Variant ASC (kernel now on main) against VariantPTO for identical shapes and hardware, same input x=[B, seq, dim], weight, bias, output, and conv_states write-back, so both kernels do identical total work (only the internal tiling differs).

Method: Ascend 910B2, fp16. e2e latency via torch.npu.Event, min over 5 windows, 100 iters/window, 5 warmups.

Result for 69 shapes: VariantPTO is faster on all 69, median 1.49× (range 1.03×–1.90×). Both kernels are correct on every shape, so this is a pure performance regression.

Full per-shape sweep (69 shapes)

B	dim	seq	B·seq·dim	Variant PTO #572 (µs)	Variant ASC #592 (µs)	Variant ASC / Variant PTO
64	1024	512	33554432	358.6	408.6	1.14×
64	1024	1024	67108864	707.0	785.3	1.11×
64	1024	2048	134217728	1400.7	1541.6	1.10×
64	2048	512	67108864	602.4	628.1	1.04×
64	2048	1024	134217728	1191.7	1232.0	1.03×
64	2048	2048	268435456	2370.3	2444.8	1.03×
64	4096	512	134217728	898.9	1084.7	1.21×
64	4096	1024	268435456	1782.6	2146.8	1.20×
64	4096	2048	536870912	3550.8	4269.3	1.20×
64	6144	512	201326592	1261.6	2144.6	1.70×
64	6144	1024	402653184	2508.8	4265.8	1.70×
64	6144	2048	805306368	5001.8	8513.3	1.70×
128	1024	512	67108864	535.3	787.6	1.47×
128	1024	1024	134217728	1055.7	1540.7	1.46×
128	1024	2048	268435456	2096.7	3051.8	1.46×
128	2048	512	134217728	899.0	1234.5	1.37×
128	2048	1024	268435456	1783.4	2444.6	1.37×
128	2048	2048	536870912	3551.0	4867.7	1.37×
128	4096	512	268435456	1788.7	2147.7	1.20×
128	4096	1024	536870912	3558.0	4271.9	1.20×
128	4096	2048	1073741824	7095.0	8517.7	1.20×
128	6144	512	402653184	2514.9	4268.9	1.70×
128	6144	1024	805306368	5009.3	8515.6	1.70×
128	6144	2048	1610612736	9997.5	17006.4	1.70×
256	1024	512	134217728	1061.4	1545.0	1.46×
256	1024	1024	268435456	2102.2	3056.3	1.45×
256	1024	2048	536870912	4183.5	6079.2	1.45×
256	2048	512	268435456	1789.0	2447.6	1.37×
256	2048	1024	536870912	3557.0	4871.7	1.37×
256	2048	2048	1073741824	7093.8	9717.6	1.37×
256	4096	512	536870912	3272.4	4274.0	1.31×
256	4096	1024	1073741824	6515.2	8520.7	1.31×
256	4096	2048	2147483648	12998.8	17013.1	1.31×
256	6144	512	805306368	4605.6	8524.3	1.85×
256	6144	1024	1610612736	9177.6	17015.0	1.85×
256	6144	2048	3221225472	18321.6	34002.0	1.86×
512	1024	512	268435456	1936.8	3062.3	1.58×
512	1024	1024	536870912	3846.9	6083.9	1.58×
512	1024	2048	1073741824	7660.7	12128.2	1.58×
512	2048	512	536870912	3271.5	4879.3	1.49×
512	2048	1024	1073741824	6516.3	9724.1	1.49×
512	2048	2048	2147483648	12997.9	19416.1	1.49×
512	4096	512	1073741824	6538.4	8529.9	1.30×
512	4096	1024	2147483648	13021.3	17023.3	1.31×
512	4096	2048	4294967296	25991.8	34008.8	1.31×
512	6144	512	1610612736	9201.1	17030.7	1.85×
512	6144	1024	3221225472	18344.8	34017.8	1.85×
512	6144	2048	6442450944	36628.2	67990.1	1.86×
1024	1024	512	536870912	3860.1	6098.9	1.58×
1024	1024	1024	1073741824	7676.3	12145.4	1.58×
1024	1024	2048	2147483648	15314.1	24232.3	1.58×
1024	2048	512	1073741824	6535.0	9739.5	1.49×
1024	2048	1024	2147483648	13019.8	19431.2	1.49×
1024	2048	2048	4294967296	25985.3	38813.8	1.49×
1024	4096	512	2147483648	12772.7	17040.8	1.33×
1024	4096	1024	4294967296	25438.1	34029.8	1.34×
1024	6144	512	3221225472	17976.0	34049.2	1.89×
1024	6144	1024	6442450944	35847.5	68024.1	1.90×
2048	1024	512	1073741824	7536.5	12170.5	1.61×
2048	1024	1024	2147483648	14992.4	24260.6	1.62×
2048	1024	2048	4294967296	29905.8	48437.2	1.62×
2048	2048	512	2147483648	12761.5	19463.1	1.53×
2048	2048	1024	4294967296	25437.4	38846.1	1.53×
2048	4096	512	4294967296	25520.0	34068.8	1.33×
2048	6144	512	6442450944	35936.0	68083.4	1.89×
4096	1024	512	2147483648	15053.3	24315.3	1.62×
4096	1024	1024	4294967296	29971.4	48496.5	1.62×
4096	2048	512	4294967296	25509.2	38909.7	1.53×
8192	1024	512	4294967296	29921.6	48602.3	1.62×

Request

1. Please confirm our benchmark findings.
2. After confirming our results, we request to revert main back to the Variant PTO implementation.

Cc

@ping1jing2 @iforgetmyname @zhaozx-cn @raphael-s-steiner

Links: #572 (merged) · #555 (open) · #592 (merged)