Repository navigation

Gfx11 - #29

Closed

liangliangchang wants to merge 22 commits into

AMD-Ecosystem:gfx11from

liangliangchang:gfx11

Closed

Gfx11#29
liangliangchang wants to merge 22 commits into
AMD-Ecosystem:gfx11from
liangliangchang:gfx11

Commits on Jun 23, 2026
- Add kernel dimension logging for profiling analysis
  
  liangliangchang
  and
  claude
  committed
  a70d5dd
- Fix M/N/K dimension logging for MMQ and MMVQ profiling.
  
  liangliangchang
  and
  cursoragent
  committed
  3c63ba2
- Improve RDNA3.5 MMQ occupancy by reducing mmq_x when LDS limits WGs/CU.
  
  liangliangchang
  and
  cursoragent
  committed
  6051c1f
- Improve RDNA3.5 MMQ prefill via LDS layout and activation load pipelining.
  
  liangliangchang
  and
  cursoragent
  committed
  36249b0
- Add opt-in HIP MMQ phase profiling to split dequant vs MMA time.
  
  liangliangchang
  and
  cursoragent
  committed
  7248ec9
- Improve RDNA3.5 Q4_K MMQ dequant/vec_dot and fix phase profiling overhead.
  
  liangliangchang
  and
  cursoragent
  committed
  7ed8d1e
Commits on Jun 24, 2026
- Refactor RDNA3.5 Q4_K qs load for WMMA ldmatrix layout.
  
  liangliangchang
  and
  cursoragent
  committed
  fb63447
- Gate MMQ kernel dimension logs behind GGML_CUDA_MM_LOG.
  
  liangliangchang
  and
  cursoragent
  committed
  5924b0a
- cuda: precompute Q4_K vec_dot scale factors before WMMA accumulate
  
  liangliangchang
  and
  cursoragent
  committed
  afa1455
- cuda: fix MMQ phase profiler clock64 deltas on HIP
  
  liangliangchang
  and
  cursoragent
  committed
  46b38f2
Commits on Jun 25, 2026
- cuda: optimize RDNA3.5 tiny-M MMQ for gate Q8_0 prefill
  
  liangliangchang
  and
  cursoragent
  committed
  bee2b6d
- cuda: hoist Q6_K vec_dot scales on RDNA3.5 WMMA
  
  liangliangchang
  and
  cursoragent
  committed
  6e777ce
- cuda: use mmq_x=128 for Q6_K narrow-N prefill on RDNA3.5
  
  liangliangchang
  and
  cursoragent
  committed
  4d9b940
Commits on Jun 30, 2026
- cuda: Q6_K MMQ B-prefetch, y-tile pipeline, and vec_dot profiling on RDNA3.5
  
  liangliangchang
  and
  cursoragent
  committed
  03927f1
- cuda: drop RDNA3.5 activation y-pipeline from MMQ
  
  liangliangchang
  and
  cursoragent
  committed
  cf83d46
- cuda: batch Q6_K B loads with load_generic on RDNA3.5 WMMA
  
  liangliangchang
  and
  cursoragent
  committed
  d4511bd
- cuda: restore Q6_K mmq_x=128 for narrow-N FFN down on RDNA3.5
  
  liangliangchang
  and
  cursoragent
  committed
  eb2a7de
- cuda: restore Q6_K vec_dot load_ldmatrix path on RDNA3.5
  
  liangliangchang
  and
  cursoragent
  committed
  49b036d
- cuda: remove always-on MMQ profiling overhead on RDNA3.5 hot path
  
  liangliangchang
  and
  cursoragent
  committed
  94e9d1d
- cuda: pipeline Q6_K B-tile loads in RDNA3.5 WMMA vec_dot
  
  liangliangchang
  and
  cursoragent
  committed
  272c12e
- cuda: use dual-WG mmq_x=64 for Q6_K narrow-N on RDNA3.5
  
  liangliangchang
  and
  cursoragent
  committed
  4648486
- cuda: limit RDNA3.5 dual-WG mmq_x downgrade to K-quants
  
  liangliangchang
  and
  cursoragent
  committed
  35bd178

Commits

Commits on Jun 23, 2026

Commits on Jun 24, 2026

Commits on Jun 25, 2026

Commits on Jun 30, 2026