Vulkan: Optimize Matmul parameters for AMD GPUs with Coopmat support

[0cc4m]

I tuned this on AMD Radeon 8060S, but a brief test also showed improvements on AMD RX 9060 XT.

ggml_vulkan: 0 = Radeon 8060S Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat

model	size	params	ngl	test	t/s (ROCm)	t/s (before)	t/s (after)	diff
llama 8B Q4_0	4.33 GiB	8.03 B	99	pp512	815.22 ± 16.88	644.84 ± 4.49	875.24 ± 17.99	+35.7%
llama 8B Q4_0	4.33 GiB	8.03 B	99	pp512 @ d8192	321.66 ± 3.02	384.27 ± 0.29	463.72 ± 1.32	+20.7%
llama 8B Q4_K - Small	4.36 GiB	8.03 B	99	pp512	995.81 ± 32.32	529.50 ± 13.07	793.51 ± 9.67	+49.9%
llama 8B Q4_K - Small	4.36 GiB	8.03 B	99	pp512 @ d8192	352.08 ± 3.17	341.86 ± 1.96	435.55 ± 2.63	+27.4%
llama 8B Q8_0	7.95 GiB	8.03 B	99	pp512	755.86 ± 21.11	422.37 ± 21.61	742.46 ± 15.49	+75.8%
llama 8B Q8_0	7.95 GiB	8.03 B	99	pp512 @ d8192	317.44 ± 2.31	306.97 ± 0.36	419.24 ± 4.59	+36.6%

ggml_vulkan: 0 = AMD Radeon RX 9060 XT (RADV GFX1200) (radv) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat

model	size	params	ngl	test	t/s (ROCm)	t/s (before)	t/s (after)	diff
llama 8B Q4_K - Small	4.36 GiB	8.03 B	99	pp512	648.41 ± 14.37	1437.13 ± 1.77	1902.23 ± 1.65	+32.4%
llama 8B Q4_K - Small	4.36 GiB	8.03 B	99	pp512 @ d8192	410.01 ± 6.22	757.59 ± 2.71	841.76 ± 3.02	+11.1%

[daniandtheweb]

There are some small but consistent improvements on RDNA3 as well with this.

model	size	params	backend	ngl	fa	test	t/s (ROCm)	t/s (before)	t/s (after)	diff
llama 7B Q4_0	3.56 GiB	6.74 B	Vulkan	100	0	pp512	2213.37 ± 25.38	1964.06 ± 17.79	1992.17 ± 24.98	+1.43%
llama 7B Q4_0	3.56 GiB	6.74 B	Vulkan	100	1	pp512	2480.36 ± 2.56	2057.06 ± 6.28	2120.18 ± 2.38	+3.06%

[0cc4m]

Thank you for testing it! I was hoping for more, I guess it is too different from 8060S. There's probably some more tuning for RDNA3/4 dGPUs that can be done, but I don't have the hardware for that.

[netrunnereve]

You should try this for mul_mat_id as well!

[characharm]

9070xt

model	test	master t/s (±)	PR t/s (±)	diff t/s	diff %
gpt-oss 20B MXFP4 MoE	pp512	4646.61 ± 74.94	3464.81 ± 61.21	-1181.80	-25.4%
gpt-oss 20B MXFP4 MoE	tg128	175.24 ± 0.50	178.45 ± 1.65	+3.21	+1.8%
gpt-oss 20B MXFP4 MoE	pp512 @ d8192	2077.60 ± 19.08	1688.42 ± 15.70	-389.18	-18.7%
gpt-oss 20B MXFP4 MoE	tg128 @ d8192	149.41 ± 0.73	149.97 ± 0.50	+0.56	+0.37%

[0cc4m]

You should try this for mul_mat_id as well!

I tried, but didn't find a good parameter set yet. I'll keep trying.

@characharm Is that Windows or Linux?

[characharm]

@0cc4m Windows. I rebooted between tests for accuracy. The numbers are stable.

[0cc4m]

Thank you for testing, I wish the drivers would behave more similarly. I'll disable the change on Windows.

[0cc4m]

Can you also test a dense model, though? Those should be more affected than MoE.

[characharm]

model	test	master t/s (±)	PR t/s (±)	diff t/s	diff %
llama 8B Q4_0	pp512 @ d8192	1659.33 ± 20.44	588.10 ± 1.59	-1071.23	-64.56%
llama 8B Q4_0	tg128 @ d8192	87.20 ± 0.48	87.20 ± 0.92	0.00	0.0%

[0cc4m]

@characharm Please check if #18763 restores your performance.

[characharm]

No, the performance is the same. I compared it with the CI build, so the problem is not in my build.
clarify:
The driver exclusion fix didn't work. I thought it might be a local build issue, but I verified with b7707 and got the same results.

[jeffbolznv]

I think these tile sizes are also used for matmul id in some cases, so that could explain the effect on gpt-oss.

[0cc4m]

@characharm Sorry, I missed disabling the large tile size. Try again, please.

I think these tile sizes are also used for matmul id in some cases, so that could explain the effect on gpt-oss.

No, I didn't enable the large tile for mul_mat_id, so unless the check somewhere is wrong, it should not be used at all.

[characharm]

@0cc4m Yes, now dens and moe show the same numbers as before 18749.

[acbits]

Not sure whether this helps, but Vulkan performance on RX 7600 has been going down. I don't use such a big model on this GPU, but interesting that it has degraded.

model	size	params	backend	ngl	threads	type_k	type_v	fa	test	t/s
qwen3 14B Q4_K - Medium	8.53 GiB	14.77 B	Vulkan	99	8	q8_0	q8_0	1	pp512	12.76 ± 0.00
qwen3 14B Q4_K - Medium	8.53 GiB	14.77 B	Vulkan	99	8	q8_0	q8_0	1	tg128	1.35 ± 0.00

build: 7d77f07 (7108)

model	size	params	backend	ngl	type_k	type_v	fa	test	t/s
qwen3 14B Q4_K - Medium	8.53 GiB	14.77 B	Vulkan	99	q8_0	q8_0	1	pp512	4.67 ± 0.00
qwen3 14B Q4_K - Medium	8.53 GiB	14.77 B	Vulkan	99	q8_0	q8_0	1	tg128	1.35 ± 0.00

build: c9ced49 (7710)

[0cc4m]

Can you test a model that actually fits into your GPU? That would likely give more usable data.

[acbits]

Can you test a model that actually fits into your GPU? That would likely give more usable data.

Luckily, I had copied the results from an old build. Yeah, even for smaller models, there has been degradation. Not sure whether kernel upgrade played a role.

Kernel: 6.6 (don't remember the exact patch)

model	size	params	backend	ngl	threads	type_k	type_v	fa	test	t/s
qwen3 8B Q5_K - Medium	5.44 GiB	8.19 B	Vulkan	99	8	q8_0	q8_0	1	pp512	498.47 ± 0.00
qwen3 8B Q5_K - Medium	5.44 GiB	8.19 B	Vulkan	99	8	q8_0	q8_0	1	tg128	34.83 ± 0.00

build: dd5e8ca (6916)

Kernel: 6.17.9-200

model	size	params	backend	ngl	type_k	type_v	fa	test	t/s
qwen3 8B Q5_K - Medium	5.44 GiB	8.19 B	Vulkan	99	q8_0	q8_0	1	pp512	240.94 ± 0.00
qwen3 8B Q5_K - Medium	5.44 GiB	8.19 B	Vulkan	99	q8_0	q8_0	1	tg128	12.13 ± 0.00

build: c9ced49 (7710)

[0cc4m]

Can you add more information about your setup? What OS, what driver, what does your device info string say, etc?

[acbits]

Can you add more information about your setup? What OS, what driver, what does your device info string say, etc?

OS: Fedora 42
Kernel: 6.17.9-200
MESA: 25.1.9
RX7600-vulkaninfo.json

[0cc4m]

My guess would be that your driver is too old, for good Mesa coopmat performance you usually want 25.3 or higher. But I didn't want to cause an issue for older versions.

[acbits]

My guess would be that your driver is too old, for good Mesa coopmat performance you usually want 25.3 or higher. But I didn't want to cause an issue for older versions.

25.1.9 is the latest. No updates are available for Fedora 42.

[Nindaleth]

25.1.9, despite being the newest for Fedora 42, is not good enough. If upgrade to Fedora 43 is not an option for you at the moment, you could try a newer mesa build from the che/mesa COPR repo.

For example a merge request providing a significant PP speed improvement was merged into Mesa repo in August and is available since release 25.2.x or 25.3.x (not sure here).