llama.cpp turns on an Adreno matmul path worth 25% prefill on gpt-oss-20b after finding one kernel eating 40.8% of GPU time

GitHub·medium signalXBlueskyLinkedInCopy link

Build b10687 (2026-08-29) defaults the xmem F16xF32 GEMM on for Adreno X2E after measuring kernel_mul_mm_f16_f32_l4_lm running gpt-oss-20b attention projections at roughly a quarter the speed of the tuned dense q4_0 GEMM, with that single kernel consuming 40.8% of prefill GPU time. The route already existed but was left opt-in, so nobody hit it; it is gated to X2E because the Adreno 840 measures neutral. The same commit bypasses the tiled f32 GEMM on the Adreno A7X, whose compiler allocates 488 B/WI of private memory against 304 on the next generation and spills in the K-loop.