hexagon: flatten matmul into 2d to use HMX in multi-sequence

[jhen0409]

Overview

Cont #29344 (mentioned about ssm_out MUL_MAT). Flatten 3d src1/dst (original use HVX MATMUL_4D_REPACKED_IMPL) into 2d to use HMX, which is faster on n_seqs > 1. If the rows of dims 1..3 can be walked with one stride and src0 not batched, the flattening is possible.

Based on my investigation, models like lfm2.cpp, mamba-base.cpp (mamba, jamba build_mamba_layer), qwen35|qwen3next|qwen35moe|qwen4exp.cpp (ssm_out), plamo2.cpp will cover the path. (or maybe minimax-01.cpp and more)

Additional information

Test on IQ-9075, run two models (that cover the path) with llama-batched-bench --device HTP0:0 -ngl 99 -fa on -t 6 --cpu-mask 0xfc -c 4096 -b 512 -ub 256 -npp 128 -ntg 32 -npl 1,2,4:

Model	npl	test	master (t/s)	patch (t/s)	change
Qwen3.5-2B Q4_0	1	pp128	673.17	673.00	-0.0%
Qwen3.5-2B Q4_0	1	tg32	20.68	20.66	-0.1%
Qwen3.5-2B Q4_0	2	pp128 x 2	452.70	719.49	+58.9%
Qwen3.5-2B Q4_0	2	tg32 x 2	29.53	30.19	+2.2%
Qwen3.5-2B Q4_0	4	pp128 x 4	445.93	703.51	+57.8%
Qwen3.5-2B Q4_0	4	tg32 x 4	33.58	34.18	+1.8%
LFM2-2.6B Q4_0	1	pp128	685.60	683.94	-0.2%
LFM2-2.6B Q4_0	1	tg32	22.32	22.64	+1.4%
LFM2-2.6B Q4_0	2	pp128 x 2	209.92	687.22	+227.4%
LFM2-2.6B Q4_0	2	tg32 x 2	27.62	28.13	+1.9%
LFM2-2.6B Q4_0	4	pp128 x 4	209.26	679.47	+224.7%
LFM2-2.6B Q4_0	4	tg32 x 4	30.74	31.47	+2.4%

Per-call MUL_MAT time with a batched src1 (GGML_HEXAGON_PROFILE=1, npl 2, one run)

Model	npl	weight	MUL_MAT shape (src0 x src1 -> dst)	calls	master (ms)	patch (ms)	change
Qwen3.5-2B Q4_0	2	ssm_out	2048:2048 x 2048:128:2:1 -> 2048:128:2:1	18	12.203	0.545	-95.5%
Qwen3.5-2B Q4_0	2	ssm_out	2048:2048 x 2048:1:2:1 -> 2048:1:2:1	72	0.174	0.112	-35.9%
LFM2-2.6B Q4_0	2	shortconv.in_proj	2048:6144 x 2048:128:2:1 -> 6144:128:2:1	22	30.077	1.276	-95.8%
LFM2-2.6B Q4_0	2	shortconv.out_proj	2048:2048 x 2048:128:2:1 -> 2048:128:2:1	22	10.254	0.543	-94.7%
LFM2-2.6B Q4_0	2	shortconv.in_proj	2048:6144 x 2048:1:2:1 -> 6144:1:2:1	88	0.303	0.252	-16.9%
LFM2-2.6B Q4_0	2	shortconv.out_proj	2048:2048 x 2048:1:2:1 -> 2048:1:2:1	88	0.109	0.092	-15.6%

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, Claude Code, I did the validation and review.