Compute Library Reference

Documentation · Tensors and frameworks · 中文 | 日本語 | Deutsch | Русский

Compute libraries implement operations, the runtime executes them on devices, and the tensor framework composes them. Library names describe responsibilities, not identical APIs or complete feature parity with similarly named CUDA libraries.

Choose a library

Library	Cargo package	Rust crate	Main operations
ruBLAS	`rublas`	`rublas`	Matrix multiplication, vector operations, grouped matrix multiplication, and INT4
ruDNN	`ruDNN`	`rudnn`	Attention, convolution, pooling, and MoE
ruTENSOR	`ruTENSOR`	`rutensor`	General tensor contractions, einsum, reductions, permutations, and elementwise operations
ruPRIM	`ruPRIM`	`ruprim`	Reductions, scans, elementwise operations, and indexing
ruFFT	`ruda-fft`	`rufft`	Real FFT and inverse transforms
ruRAND	`ruRAND`	`rurand`	Uniform, normal, and Bernoulli distributions
ruSPARSE	`ruSPARSE`	`rusparse`	Sparse matrix formats and operations
ruCCL	`ruCCL`	`ruccl`	Collective communication and orchestration
ruSOLVER	`ruda-solver`	`rusolver`	Experimental FP64 solvers/CG; opt-in small FP32 device solve
ruINTEGRATE	`ruintegrate`	`ruintegrate`	Experimental host quadrature and non-stiff ODEs

Interface layers

Kernel/launch interfaces accept device bindings and execution configuration for kernel and library development.
Tensor interfaces handle allocation, layouts, and calls, commonly using RudaTensor<R>.
Framework interfaces dispatch to libraries through components such as ruda-tensor-device.

Operations with the same name at different layers may have different arguments, return types, and error handling. Check the package, module path, and feature together; kernel entry point signatures are not tensor entry point signatures.

Using the guides

Each library guide covers its purpose, features, interfaces, data contracts, and constraints. Explicitly select the features needed by a general-purpose path rather than assuming defaults suit every device.