Repository navigation

Feature/2.0 - #23

Merged

sdcb merged 67 commits into

mainfrom

feature/2.0

Sep 30, 2026

Merged

Feature/2.0#23
sdcb merged 67 commits into
mainfrom
feature/2.0

Commits on Sep 26, 2026
Commits on Sep 27, 2026
Commits on Sep 28, 2026
- tools: check in the GPU tuning harness as GpuBench
  sdcb
  committed
  9fdc2bb
- gitignore: split the .wrangler/ entry back out
  sdcb
  committed
  faeadf2
- tools(GpuBench): ship the --sgtest subgroup-size probe
  sdcb
  committed
  9318ea8
- tools(GpuBench): check in the Windows bench scripts
  sdcb
  committed
  94183a8
- Merge pull request #19 from sdcb/vulkan-gpu-backend
  sdcb
  authored
  1c87543
- vulkan: accept UMA host-cached memory for CPU readback
  sdcb
  committed
  7bff285
- vulkan: add a no-coopmat GEMM and leave it off the default route
  sdcb
  committed
  67852e0
- perf(vulkan): 64x64 fp32-staged no-coopmat GEMM pinned to 8 lanes
  sdcb
  committed
  84665cd
- perf(vulkan): route large K<=128 pointwise convs to the no-coopmat GEMM
  sdcb
  committed
  ad14bcf
- vulkan: run the no-coopmat tier only when Vulkan is chosen explicitly
  sdcb
  committed
  2696073
- vulkan: drop unused SE split shaders and the old GEMM probe
  sdcb
  committed
  f1ad00e
- vulkan: gate the no-coopmat tier on the cm set's own MMA shape (8x16x16 on sg16)
  
  sdcb
  and
  cursoragent
  committed
  7b2c5bc
- Merge pull request #20 from sdcb/vulkan-no-coopmat
  sdcb
  authored
  6e0dc36
- vulkan: load libvulkan.so on Android
  sdcb
  committed
  c0473cf
- test: add an Android bench host for the Vulkan backend
  sdcb
  committed
  98da6e2
- vulkan: direct-load gemm_nc where no narrow subgroup can be pinned
  sdcb
  committed
  981b6fe
- vulkan: pad fp16 constant uploads to whole 16 bytes
  sdcb
  committed
  04f38bc
- vulkan: keep the im2col matrix within one binding's maxStorageBufferRange
  sdcb
  committed
  afb0f38
- test(android): GEMM variant sweep, FMA peak probe, pure-GPU profile script
  sdcb
  committed
  b865f97
- perf(vulkan): fp16 shared tiles with BK 8 in the direct-load gemm_nc
  sdcb
  committed
  e228638
- perf(vulkan): two-phase spatial mean at any size on wide no-coopmat parts
  sdcb
  committed
  8961dc2
- perf(vulkan): direct kxk conv at every K on wide no-coopmat parts
  sdcb
  committed
  9156482
- test(android): alternating CPU/Vulkan e2e script; stop the old process before a launch
  sdcb
  committed
  b294c2a
- vulkan: Auto takes the no-coopmat tier where subgroups cannot go below 64
  sdcb
  committed
  a318575
- docs: Snapdragon 8 Gen 3 / Adreno 750 Vulkan report
  sdcb
  committed
  23e53e2
Commits on Sep 29, 2026

Commits

Commits on Sep 26, 2026

Commits on Sep 27, 2026

Commits on Sep 28, 2026

Commits on Sep 29, 2026