Add Q5_K ESIMD kernel

[malsbat]

Overview

This completes the usage of ESIMD for DMMV QK kernels begun in #25942.

Measured tg improvement of 1.04x to 1.35x (depending on hardware) using Qwen35-27B-Q5_K_M.

Additional information

Note: all measurements were captured with aicss-genai/fix-row-calculation (already merged in master) and aicss-genai/fattn-single-device-sync (better solution already merged in master, will only improve pp results below) merged as the parent branch predates those necessary patches.

PVC performance

Baseline

export GGML_SYCL_PRIORITIZE_DMMV=0
export GGML_SYCL_ENABLE_ESIMD=0

model	size	params	backend	ngl	test	t/s
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	pp512	908.19 ± 3.04
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	tg128	17.25 ± 0.00

DMMV

export GGML_SYCL_PRIORITIZE_DMMV=1
export GGML_SYCL_ENABLE_ESIMD=0

model	size	params	backend	ngl	test	t/s
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	pp512	904.31 ± 7.70
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	tg128	11.84 ± 0.00

ESIMD

export GGML_SYCL_PRIORITIZE_DMMV=0
export GGML_SYCL_ENABLE_ESIMD=1

model	size	params	backend	ngl	test	t/s
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	pp512	901.58 ± 5.68
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	tg128	23.29 ± 0.03

B70 performance

Baseline

model	size	params	backend	ngl	test	t/s
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	pp512	826.10 ± 2.80
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	tg128	21.88 ± 0.04

DMMV

model	size	params	backend	ngl	test	t/s
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	pp512	829.38 ± 6.64
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	tg128	12.87 ± 0.05

ESIMD

model	size	params	backend	ngl	test	t/s
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	pp512	833.75 ± 3.07
qwen35 27B Q5_K - Medium	18.25 GiB	26.90 B	SYCL	100	tg128	22.69 ± 0.05

Perplexity tests across wikitext, hellaswag, and winogrande showed no significant changes between baseline and ESIMD.

Requirements

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

[malsbat]

Same as the others, I will rebase this once #25942 is merged.

[arthw]

With gemma-4-E2B-it-Q5_K_M.gguf on B60

Test	fa	Base t/s	Primary t/s	Increase Rate (Primary vs Base)
pp512	0	2154.07	2156.09	0.09%
pp512	1	2143.85	2133.25	-0.49%
tg128	0	99.42	104.56	5.17%
tg128	1	99.12	103.82	4.74%

[arthw]

It's good job!

Thank you!