scir-gpu
| Crates.io | scir-gpu |
| lib.rs | scir-gpu |
| version | 0.3.2 |
| created_at | 2025-09-10 21:10:13.55221+00 |
| updated_at | 2025-09-11 11:29:12.510651+00 |
| description | SciR GPU foundations: device arrays and CUDA (feature-gated) elementwise/FIR kernels with CPU parity. |
| homepage | https://github.com/SoftOboros/scir |
| repository | https://github.com/SoftOboros/scir |
| max_upload_size | |
| id | 1833011 |
| size | 44,634 |
Ira Abbott (iraabbott)
documentation
https://docs.rs/scir-gpu
README
scir-gpu
Overview
- GPU foundations for SciR. Provides a minimal
DeviceArray<T>abstraction, CPU baselines, and optional CUDA paths behind thecudafeature.
Features
- cuda: enables CUDA Driver API usage with embedded PTX kernels for f32 elementwise add/mul and batched FIR.
Requirements (CUDA)
- NVIDIA GPU with recent driver installed (libcuda present on host).
- On Linux, ensure
libcuda.sois visible to the container or process; on Windows,nvcuda.dllis required.
Quick start
- CPU tests:
cargo test -p scir-gpu - CUDA tests:
cargo test -p scir-gpu --features cuda
APIs
DeviceArray<T>: shaped arrays withdeviceanddtype. CPU-backed storage today.- Elementwise:
add_scalar_auto,mul_scalar_auto,add_auto(f32) dispatch to CUDA when available. - FIR:
fir1d_batched_f32_auto(x, taps, device)chooses CUDA or CPU and falls back to CPU if CUDA is unavailable.