Markus Fritsche
107ccddbb1
Final deliverables (mesa-panvk-shared-tls-wls + llama.cpp-mali-mmq-warptile) at top of patches/; iteration trail moved to patches/iterations/. Public-facing README with the 4.8x result and the two-fix summary. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Patches
Apply these two (the deliverables)
| patch | target | what |
|---|---|---|
mesa-panvk-shared-tls-wls.patch |
Mesa 26.1.3, src/panfrost/vulkan/ |
Fix the panvk priv-VA-heap exhaustion (OOM at ub>64). Device-shared TLS scratch + WLS reuse. Also contains a small PANVK_AS_LOG priv-heap instrument (env-gated). |
llama.cpp-mali-mmq-warptile.patch |
llama.cpp 389ff61, ggml/src/ggml-vulkan/ |
Mali/panvk mul_mm/mul_mmq warptile branch + BK_STEP spec constant (L-scoped). The 4.8× prefill. Also contains an env-gated [panvk-tls]/PANVK_WLS_LOG shader-spill trace. |
Env flags used at runtime (both patches): PANVK_SHARED_TLS=1 PANVK_WLS_REUSE=1. Trace flags
(optional): PANVK_AS_LOG=1, PANVK_WLS_LOG=1.
Build panvk minimally: meson setup build -Dvulkan-drivers=panfrost -Dgallium-drivers= ...
(mirrors a bifrost-style co-install). Point VK_ICD_FILENAMES at the built libvulkan_panfrost.so.
Run on a panthor kernel with the GPU bound to panthor, not the vendor mali blob.
iterations/ — the tuning trail (not for applying)
The single-variable shader experiments, in order. See docs/shader-tuning.md for the numbers.
llama.cpp-0001-vk-tunable-submit-thresholds.patch— exploratory, no effect (kept for the record).llama.cpp-0002-vk-cmd-cleanup-freq-env.patch— falsified hypothesis (cmdbufs are in-flight).llama.cpp-0003-mali-warptile-iter1.patch— 16-acc tile, spill 5328→96 B, ub128 4×.llama.cpp-0003-mali-warptile-iter3-champion.patch— bigger workgroup, same registers, +10%.llama.cpp-0004-mali-shader-iter5-champion.patch— scoped BK_STEP=2 (superseded by the final patch, which is iter5's content).