ggml-vulkan : tune L-tile warp micro-dimension for RDNA3 iGPUs

[Troncooooo]

Overview

Based on suggestion of Google AI search/Gemini, my LLM (qwen3.8, ollama/pi dev) improved preprocessing speed on my Minisforum um790 pro. This has resulted into an improved pp of about 61%. As to give exact details, the writer (A.I) give the following hint:

On RADV PHOENIX (gfx1151, UMA, KHR cooperative matrix) the 64x64 warp micro-tile is LDS/register limited. Halving the M micro-dimension (WM 64 -> 32 on l_warptile and the mmq/mmq_int variants) gives a large prompt-processing speedup with no effect on token generation (the mul_mat_vec path is untouched):

llama-bench -p 512 -n 64 -r 3, Mesa/radv:
Qwen3.8-9B-Q4_K_M: pp512 250.5 -> 404.4 t/s (+61%); tg64 14.42 -> 14.43
Qwen3.8-27B-UD-Q4_K_M: pp512 72.4 -> 117.5 t/s (+62%); tg64 4.73 -> 4.73

Scope: AMD + coopmat + non-proprietary driver (existing branch), further restricted to RDNA3 + UMA so discrete RDNA3 and other chips are unaffected.

Additional information

I've tested the improvement and kept the temperature at 0 (--temp 0) to validate output stayed the same.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, Gemini for looking for possible performance gains for my local machine. Local Qwen3.8 27B (Ollama) + Pi.dev for writing tests and code improvements. I've run the tests and verified the performance boost.

[0cc4m]

This is not a valid tile configuration. You're seeing performance improvements because it skips half of the work, output will be corrupt with a larger prompt.