Repository navigation

feat: add support for cuda-oxide (without breaking rust-cuda) - #13

Merged

sebcrozet merged 33 commits into

mainfrom

cuda-oxide-build-path

Sep 30, 2026

Merged

feat: add support for cuda-oxide (without breaking rust-cuda)#13
sebcrozet merged 33 commits into
mainfrom
cuda-oxide-build-path

Commits on Sep 30, 2026
- fix(webgpu): declare SPIR-V entry points for passthrough shader modules
  sebcrozet
  committed
  eed3b44
- khal-derive + khal-std: cuda-oxide (native-CUDA) device backend
  
  haixuanTao
  authored and
  sebcrozet
  committed
  8a2cb1f
- khal/cuda: print entry name on load_function failure (debug aid for cubin/host hash mismatches)
  
  haixuanTao
  authored and
  sebcrozet
  committed
  d94f6ab
- khal-std: cap glam <0.33 for the spirv-std shader build
  
  haixuanTao
  authored and
  sebcrozet
  committed
  7ac12db
- fix(cuda): push element count, not byte length, for slice kernel args
  
  haixuanTao
  authored and
  sebcrozet
  committed
  1029ddc
- khal-std: drop obsolete glam<0.33 cap (defer to glamx 0.3, matches upstream); khal-builder: KHAL_SKIP_SPIRV skip for cuda-only builds
  
  haixuanTao
  authored and
  sebcrozet
  committed
  79c0ae2
- Port CUDA runtime APIs from fork branches: compute_capability, graph capture (begin/end_capture, CapturedGraph), KHAL_CUDA_PROFILE kernel profiling, prebuilt-PTX escape hatch (CUDA_OXIDE_SHADERS_PT…
  
  haixuanTao
  authored and
  sebcrozet
  committed
  784d8d8
- khal-std: trig Float methods, generic SmemBuf, panic-free atomics for cuda-oxide
  
  haixuanTao
  authored and
  sebcrozet
  committed
  033d51b
- khal-std/derive: unified-mode cuda-oxide path (host-target, no --device)
  
  haixuanTao
  authored and
  sebcrozet
  committed
  63af1e9
- khal: never degrade the backend selection silently
  
  haixuanTao
  authored and
  sebcrozet
  committed
  d72fd4b
- fix(cuda): make graph capture actually work — owned stream + as_cuda on the backend enum
  
  haixuanTao
  authored and
  sebcrozet
  committed
  0991e11
- fix(khal-std): resolve cuda-device by git rev, not an absolute local path (#5)
  
  authored and
  sebcrozet
  committed
  96e3d5f
- khal-std: Index/IndexMut for the cuda-oxide SmemBuf glue
  
  haixuanTao
  authored and
  sebcrozet
  committed
  b0a3392
- khal-std: cuda-oxide arms for the RMW atomics (fetch_add/max/min/or/and/swap)
  
  haixuanTao
  authored and
  sebcrozet
  committed
  4b45508
- Revert "khal-std: cuda-oxide arms for the RMW atomics (fetch_add/max/min/or/and/swap)"
  
  haixuanTao
  authored and
  sebcrozet
  committed
  0aa6779
- khal-std: legacy-backend-compatible barrier (llvm.nvvm.barrier0 extern)
  
  haixuanTao
  authored and
  sebcrozet
  committed
  876afad
- fix(builder): rerun-if-changed on prebuilt PTX/cubin path
  
  haixuanTao
  authored and
  sebcrozet
  committed
  12b7f7d
- debug(cuda): env-gated capture diagnostics
  
  haixuanTao
  authored and
  sebcrozet
  committed
  cc7682c
- khal-std: pin cuda-device to upstream NVlabs cuda-oxide
  
  haixuanTao
  authored and
  sebcrozet
  committed
  19a58ac
- docs: the unified host-target path is production; cuda-device comes from NVlabs
  
  haixuanTao
  authored and
  sebcrozet
  committed
  3187c3f
- khal-std: bump cuda-device to NVlabs main with our merged fixes (#518, #520)
  
  haixuanTao
  authored and
  sebcrozet
  committed
  a70c8f7
- khal-builder: create the shader output dir on the prebuilt-cubin path
  
  haixuanTao
  authored and
  sebcrozet
  committed
  bf7ba72
- webgpu: wasm perf counters at the browser boundary (+ pub webgpu module)
  
  haixuanTao
  authored and
  sebcrozet
  committed
  523cea5
- metal: fix write_buffer/GPU race + WebGPU-parity shader semantics
  
  haixuanTao
  authored and
  sebcrozet
  committed
  c42e2ab
- metal: set threadGroupSizeIsMultipleOfThreadExecutionWidth on PSOs
  
  haixuanTao
  authored and
  sebcrozet
  committed
  4ecf417
- webgpu: name the oversize dispatch in the 65535-cap panic
  
  haixuanTao
  authored and
  sebcrozet
  committed
  930da99
- cuda-oxide + rust-cuda: end-to-end PTX/cubin build path and device fixes
  sebcrozet
  committed
  1cffa24
- Fix MPM hang (rust-cuda) and NaNs (cuda-oxide) in multi-dim workgroups
  sebcrozet
  committed
  e5cc8b2
- GpuReadback: implement the CUDA arm (async device->pinned copy + event)
  sebcrozet
  committed
  2f2c1c7
- CUDA: indirect dispatch without host sync (persistent grid + entry loop)
  sebcrozet
  committed
  6347dcd
- khal-builder: fall back to __nvcc_device_query and a cached arch when nvidia-smi fails
  sebcrozet
  committed
  173cea5
- Backend-level CUDA graph capture (GpuGraph, begin/end_capture)
  sebcrozet
  committed
  f130c14
- chore: CI fixes
  sebcrozet
  committed
  78724f3

Commits

Commits on Sep 30, 2026