Release v0.4.0

[sebcrozet]

v0.4.0

Added

• cuda-oxide support as an alternative to rust-cuda for building CUDA kernels. Enable the cuda-oxide feature on khal-builder and on khal-std (with default-features = false; the new default rust-cuda feature keeps the previous behavior). #[spirv_bindgen] emits a matching kernel entry. (feat: add support for cuda-oxide (without breaking rust-cuda) #13)
• Compute graphs: GpuBackend::begin_capture/end_capture record submitted work into a GpuGraph that is replayed with GpuGraph::launch. CUDA only for now (GpuBackend::supports_graphs); other backends return the new GpuBackendError::Unsupported.
• GpuBackend::auto: picks CUDA on sm_120+ devices when compiled in, WebGPU otherwise. Overridable with KHAL_BACKEND=cuda|webgpu|metal.
• GpuReadback is now non-blocking on CUDA (async copy into pinned host memory polled with an event).
• Cuda::compute_capability.
• khal-builder: CUDA PTX is assembled into a cubin with ptxas when available (avoids driver/toolkit PTX version mismatches; opt out with KHAL_CUDA_KEEP_PTX=1; target arch from KHAL_CUDA_ARCH or auto-detected). KHAL_SKIP_SPIRV=1 skips the SPIR-V build, and CUDA_OXIDE_SHADERS_PTX_<CRATE> embeds a prebuilt PTX/cubin.
• atomic_sub_u32 in khal-std.
• Debugging env vars for the CUDA backend: KHAL_CUDA_TRACE, KHAL_CUDA_PROFILE (with cuda::dump_kernel_profile), KHAL_CUDA_ALLOC_TRACE, KHAL_CUDA_GRAPH_DOT.

Changed

• Breaking (CUDA): indirect dispatches no longer read the workgroup count back to the host. Kernels are launched with a fixed number of resident blocks (KHAL_CUDA_PERSISTENT_BLOCKS, default 16×SM count) that loop over the virtual workgroups. This changes the generated kernel ABI (slice bindings now receive an element count instead of a byte length, plus a trailing indirect-args parameter), so CUDA shaders must be rebuilt. KHAL_CUDA_INDIRECT_SYNC=1 restores the old readback path.
• The CUDA backend now runs on its own stream instead of the legacy default stream (required for graph capture).
• load_function errors on CUDA name the missing entry point. (cuda: name the missing entry point in load_function errors #10)
• Metal now matches WebGPU semantics: buffers are zero-initialized, out-of-bounds reads return zero and writes are skipped, workgroup memory is zero-initialized, loops are bounded, fast math is disabled, and MSL targets Metal 3.0.
• WebGPU: slow_read_buffer reuses pooled staging buffers. On wasm, consecutive compute passes are merged into one to cut per-pass overhead.
• WebGPU: dispatches exceeding the 65535 per-dimension workgroup limit panic with an explicit message.
• GpuBackend::is_cuda is available without the cuda feature.
• cargo-cuda disables MIR jump-threading, which could duplicate workgroup barriers and deadlock kernels.

Fixed

• Metal: write_buffer while GPU work is in flight no longer races with running kernels; it is now ordered after previously submitted work.
• WebGPU SPIR-V passthrough modules now declare their entry points to wgpu, which otherwise failed to load them on Vulkan with Unable to find entry point.