A machine learning framework that sets tensors ablaze.
🚧 WORK IN PROGRESS 🚧
- singe-cuda: the CUDA driver and runtime APIs, including runtime compilation of CUDA C++ and NVTX annotations.
- singe-cuda-macros: derive macros for types that can be copied to the GPU, and
cuda_module!for typed kernel launches. - singe-cuda-find: finds the CUDA Toolkit and NVIDIA libraries for build scripts.
Each wrapper crate has a -sys crate of raw bindings beside it, such as singe-cuda-sys.
- singe-cublas: cuBLAS dense linear algebra, cuBLASLt matrix multiplication, and cuBLASXt BLAS3 across several GPUs.
- singe-cudnn: cuDNN deep neural network primitives, including a Rust port of NVIDIA's cuDNN Frontend.
- singe-cudss: cuDSS direct sparse solvers.
- singe-cufft: cuFFT Fourier transforms.
- singe-cufile: cuFile, which moves data between files and GPU memory with GPUDirect Storage.
- singe-cupti: CUPTI tracing and profiling.
- singe-curand: cuRAND random number generation on the GPU or CPU.
- singe-cusolver: cuSOLVER dense matrix factorizations and solvers.
- singe-cusparse: cuSPARSE sparse linear algebra.
- singe-cutensor: cuTENSOR tensor contractions, reductions, and elementwise operations.
- singe-nccl: NCCL collective communication between GPUs.
- singe-npp: NPP image and signal processing.
- singe-nvcomp: nvCOMP compression and decompression on the GPU.
- singe-nvimagecodec: nvImageCodec image decoding and encoding.
- singe-nvjpeg: nvJPEG and nvJPEG 2000 image decoding and encoding.
- singe-nvml: NVML GPU monitoring and management.
- singe-nvtiff: nvTIFF TIFF decoding and encoding.
- singe-onnx: decodes ONNX models from their protobuf encoding into typed Rust values.
- singe-ptx: a PTX parser, and a CPU emulator that runs PTX kernels without an NVIDIA GPU.