coeus-ops
Numerical tensor operators for Coeus.
This is the kernel library and the BackendOps dispatch surface: the point
where a tensor operation is resolved to a concrete backend implementation.
Operations are generic over <T: Scalar, B: ComputeBackend> and monomorphize
to direct specializations.
What is here
- Elementwise unary and binary operators.
matmul, batched matmul, and outer product.- Reductions, scans,
topk, and norms. - Convolution:
conv1d/conv2d/conv3dand their transposed forms. - Pooling, including adaptive pooling.
- Embedding, interpolation, and
unfold/fold. - Scaled dot-product attention.
- A lazy fused-expression DAG for combining elementwise operations.
- Provider-dispatched optimizer steps used by
coeus-optim.
CPU execution routes through coeus-leto into the
Leto array kernels.
Documentation
API docs: https://docs.rs/coeus-ops
License
Licensed under either of Apache License, Version 2.0 or MIT license at your option.