vulkan: optimize mul_mat_id loading row ids into shared memory

[jeffbolznv]

• Spread the work across the whole workgroup. Using more threads seems to far outweigh the synchronization overhead.
• Specialize the code for when the division is by a power of two.

5090 before

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench.exe -fa 1 -n 0 -p 512 -r 100 --prio 1 -m c:\models\Qwen_Qwen3-30B-A3B-Q2_K.gguf -m c:\models\\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf -m c:\models\gpt-oss-20b-mxfp4.gguf
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| qwen3moe 30B.A3B Q2_K - Medium |  10.15 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |      4530.43 ± 31.16 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |     7549.93 ± 116.08 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           pp512 |     7275.45 ± 115.27 |

5090 after

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench.exe -fa 1 -n 0 -p 512 -r 100 --prio 1 -m c:\models\Qwen_Qwen3-30B-A3B-Q2_K.gguf -m c:\models\\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf -m c:\models\gpt-oss-20b-mxfp4.gguf
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| qwen3moe 30B.A3B Q2_K - Medium |  10.15 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |      6169.11 ± 71.60 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |     8449.56 ± 125.48 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           pp512 |     7554.83 ± 156.67 |

4070 before

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench.exe -fa 1 -n 0 -p 512 -r 100 --prio 1 -m c:\models\Qwen_Qwen3-30B-A3B-Q2_K.gguf -m c:\models\\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf -m c:\models\gpt-oss-20b-mxfp4.gguf
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 4070 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| qwen3moe 30B.A3B Q2_K - Medium |  10.15 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |       1547.55 ± 9.91 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |       2188.88 ± 9.85 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           pp512 |      2757.82 ± 25.81 |

4070 after

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench.exe -fa 1 -n 0 -p 512 -r 100 --prio 1 -m c:\models\Qwen_Qwen3-30B-A3B-Q2_K.gguf -m c:\models\\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf -m c:\models\gpt-oss-20b-mxfp4.gguf
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 4070 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| qwen3moe 30B.A3B Q2_K - Medium |  10.15 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |      2230.97 ± 21.65 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |      2722.84 ± 13.74 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           pp512 |      2995.57 ± 28.40 |

[0cc4m]

Wow, great improvement. This closes the gap between CUDA and Vulkan MMID significantly. In some cases Vulkan even beats CUDA in pp512 now on my RTX 3090.

Example result, on Master:

model	size	params	backend	ngl	fa	test	t/s
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	0	pp512	1254.06 ± 7.48
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	0	tg128	140.24 ± 0.75
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	1	pp512	1284.92 ± 6.66
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	1	tg128	143.79 ± 0.53

PR:

model	size	params	backend	ngl	fa	test	t/s
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	0	pp512	1986.57 ± 16.09
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	0	tg128	137.40 ± 1.86
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	1	pp512	2055.43 ± 15.58
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	Vulkan	99	1	tg128	139.03 ± 0.14

CUDA:

model	size	params	backend	ngl	fa	test	t/s
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	CUDA	99	0	pp512	1972.23 ± 16.20
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	CUDA	99	0	tg128	139.29 ± 0.31
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	CUDA	99	1	pp512	2106.30 ± 9.63
qwen3moe 30B.A3B Q4_K - Medium	17.28 GiB	30.53 B	CUDA	99	1	tg128	141.42 ± 0.42

No change on (non-coopmat) AMD and Intel, of course.