GroupNorm Vulkan subgroup reduce optimization

[futz12]

Add single-dispatch subgroup arithmetic fast-path for GroupNorm Vulkan.

• groupnorm_reduce_subgroup.comp: pack1, two-pass (sum->mean, sqsum->var)
• groupnorm_reduce_subgroup_pack4.comp: pack4 with component-boundary checks
• groupnorm_vulkan.cpp/h: conditional subgroup path with fallback
• tests/perf/perf_groupnorm.cpp: benchmark for representative shapes

Performance (NVIDIA RTX 4060 Laptop, gpu-1, median ms):

shape	precision	baseline	optimized	speedup
[64,64,128]	fp32	37.94	15.12	2.5x
[64,64,128]	fp16	36.18	11.85	3.1x
[64,64,128]	fp16psa	36.11	11.67	3.1x
[64,64,128]	bf16	36.62	12.28	3.0x
[32,32,256]	fp32	30.37	5.56	5.5x
[32,32,256]	fp16	29.66	0.92	32.2x
[32,32,256]	fp16psa	29.71	0.92	32.3x
[32,32,256]	bf16	29.80	0.95	31.4x
[16,16,512]	fp32	29.66	4.46	6.7x
[16,16,512]	fp16	29.11	2.89	10.1x
[16,16,512]	fp16psa	29.20	2.58	11.3x
[16,16,512]	bf16	29.40	2.37	12.4x
[8,8,512]	fp32	24.54	6.38	3.8x
[8,8,512]	fp16	24.38	5.67	4.3x
[8,8,512]	fp16psa	24.34	5.55	4.4x
[8,8,512]	bf16	24.39	5.68	4.3x
[224,224,3]	fp32	38.49	35.63	1.1x
[224,224,3]	fp16	38.30	31.06	1.2x
[224,224,3]	fp16psa	38.35	31.78	1.2x
[224,224,3]	bf16	38.48	30.66	1.3x
[224,224,64]	fp32	13.22	9.95	1.3x
[224,224,64]	fp16	8.53	8.93	1.0x
[224,224,64]	fp16psa	8.55	8.96	1.0x
[224,224,64]	bf16	8.88	9.24	1.0x
[4096,1,1]	fp32	25.18	8.53	3.0x
[4096,1,1]	fp16	25.07	7.58	3.3x
[4096,1,1]	fp16psa	25.12	7.49	3.4x
[4096,1,1]	bf16	25.13	7.63	3.3x
[512,1,1]	fp32	6.96	6.25	1.1x
[512,1,1]	fp16	4.40	5.34	0.8x
[512,1,1]	fp16psa	2.87	5.45	0.5x
[512,1,1]	bf16	2.29	5.41	0.4x

SD-like shapes (32x32x256, 16x16x512) benefit most from eliminating ~10-15 dispatches down to 3 dispatches (subgroup reduce -> coeffs -> norm). Small shapes (512x1x1) see regression because subgroup shader fixed-cost overhead exceeds savings for tiny group sizes.

test_groupnorm passes on NVIDIA RTX 4060

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 621533af32

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Respect disabled subgroup ops

When a caller sets opt.use_subgroup_ops = false, this branch still creates the subgroup GroupNorm pipeline on any device advertising arithmetic support, and forward_inplace later selects any non-null subgroup pipeline instead of the fallback path. This re-enables subgroup shaders for users/drivers that explicitly disabled them, so the condition should also gate on opt.use_subgroup_ops.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Accumulate pack1 reductions in float

For pack1 tensors with fp16 arithmetic enabled, afp is float16_t, so the new subgroup path accumulates the mean reduction in half precision (and the variance reduction does the same below) even though the existing fallback uses the groupnorm_reduce_sum4_fp16_to_fp32 path. On large groups such as a single group over a 224x224x3 tensor, the half-precision running sum/sqsum can overflow or lose enough precision to produce incorrect normalization; use float accumulators for these reductions like the pack4 shader does.

Useful? React with 👍 / 👎.