Add DMMV ESIMD Q3_K kernel

[malsbat]

Overview

Continuing the work in #25942 to add the DMMV ESIMD Q3_K kernel.

Additional information

Note: all measurements were captured with #25741 applied locally as the parent PR does not include the newer upstream fix. This only affects pp numbers which are not the important part of this PR.

Performance on B70

export GGML_SYCL_PRIORITIZE_DMMV=0
export GGML_SYCL_ENABLE_ESIMD=0
./build/bin/llama-bench -m /models/Qwen3-14B-Q3_K_M.gguf -ngl 100

model	size	params	backend	ngl	test	t/s
qwen3 14B Q3_K - Medium	6.81 GiB	14.77 B	SYCL	100	pp512	1305.08 ± 4.29
qwen3 14B Q3_K - Medium	6.81 GiB	14.77 B	SYCL	100	tg128	28.62 ± 0.19

export GGML_SYCL_PRIORITIZE_DMMV=1
export GGML_SYCL_ENABLE_ESIMD=0
./build/bin/llama-bench -m /models/Qwen3-14B-Q3_K_M.gguf -ngl 100

model	size	params	backend	ngl	test	t/s
qwen3 14B Q3_K - Medium	6.81 GiB	14.77 B	SYCL	100	pp512	1296.01 ± 5.29
qwen3 14B Q3_K - Medium	6.81 GiB	14.77 B	SYCL	100	tg128	31.37 ± 0.27

export GGML_SYCL_PRIORITIZE_DMMV=0
export GGML_SYCL_ENABLE_ESIMD=1
./build/bin/llama-bench -m /models/Qwen3-14B-Q3_K_M.gguf -ngl 100

model	size	params	backend	ngl	test	t/s
qwen3 14B Q3_K - Medium	6.81 GiB	14.77 B	SYCL	100	pp512	1296.22 ± 4.78
qwen3 14B Q3_K - Medium	6.81 GiB	14.77 B	SYCL	100	tg128	53.61 ± 0.27

Perplexity

model=/models/Qwen3-14B-Q3_K_M.gguf

export GGML_SYCL_PRIORITIZE_DMMV=0
export GGML_SYCL_ENABLE_ESIMD=0
./build/bin/llama-perplexity -m ${model} -f wikitext-2-raw/wiki.test.raw
4.19.245.222 I Final estimate: PPL = 9.2916 +/- 0.07317
./build/bin/llama-perplexity --hellaswag -f hellaswag_val_full.txt -m ${model}
400     78.25000000%    [73.9487%, 82.0138%]
./build/bin/llama-perplexity -f winogrande-debiased-eval.csv --winogrande -m ${model}
2.39.277.933 I Final Winogrande score(1267 tasks): 72.8493 +/- 1.2499

export GGML_SYCL_PRIORITIZE_DMMV=0
export GGML_SYCL_ENABLE_ESIMD=1
./build/bin/llama-perplexity -m ${model} -f wikitext-2-raw/wiki.test.raw
4.19.304.028 I Final estimate: PPL = 9.2917 +/- 0.07317
./build/bin/llama-perplexity --hellaswag -f hellaswag_val_full.txt -m ${model}
400     78.25000000%    [73.9487%, 82.0138%]
./build/bin/llama-perplexity -f winogrande-debiased-eval.csv --winogrande -m ${model}
2.39.320.785 I Final Winogrande score(1267 tasks): 72.7703 +/- 1.2511

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, used Claude to help author kernel based on existing patterns.

[malsbat]

This is currently a draft built upon #25942, I will rebase and mark ready for review when that PR is merged.

[arthw]

Qwen3-14B-Q3_K_M.gguf on B60:
export GGML_SYCL_PRIORITIZE_DMMV=0
export GGML_SYCL_ENABLE_ESIMD=1

Test	fa	Base t/s	Primary t/s	Increase Rate (Primary vs Base)
pp512	0	92.19	99.59	8.03%
pp512	1	75.33	81.65	8.39%
tg128	0	4.43	8.57	93.45%
tg128	1	4.56	9.08	99.12%

[arthw]

It's good job!

Thank you!

[arthw]

@malsbat
Please change all ready PR status from "Draft" to "Open".

[malsbat]

@malsbat Please change all ready PR status from "Draft" to "Open".

I have reached the open pull request limit, I will mark the others as ready for review once these get merged.

[arthw]

@malsbat
OK, I see.
You have 2 PRs to be ready to merge and 2 PRs as Draft.

We will merge them as soon!

Thank you!