baracuda-cudnn

Safe Rust wrappers for NVIDIA cuDNN — deep-learning primitive ops: convolution, pooling, activation, normalization, RNN, attention.

Coverage

Workhorse-level coverage. Covers the classic op API comprehensively plus scaffolding for the modern Backend (Graph) API:

• Activation: forward + backward; ReLU, sigmoid, tanh, ELU, etc.
• Convolution: forward + backward-data + backward-filter + backward-bias, all algorithm variants. N-D Filter / Convolution descriptors.
• Pooling: max, avg, deterministic; forward + backward.
• Softmax + LogSoftmax: forward + backward.
• Batch normalization: inference, training, backward.
• LRN, dropout, op-tensor, reduce-tensor, transform, add, scale.
• Spatial transformer.
• CTC loss.
• RNN: forward + backward; LSTM / GRU / vanilla cell types.
• N-D tensor / filter / pooling / convolution descriptors.
• Backend (Graph) API: BackendDescriptor with set_attribute_raw, finalize, execute. Lower-level than the classic ops but supports operations the classic API doesn't.

Half-precision interop

Enable the half-crate feature to use half::f16 and half::bf16 as CudnnDataType directly, without going through baracuda's Half / BFloat16 conversion types:

baracuda-cudnn = { version = "0.0.1-alpha.7", features = ["half-crate"] }

Stream binding

Set the active stream on a handle with Handle::set_stream. All ops on that handle dispatch to that stream until you change it again.

Pairs with baracuda-cudnn-sys for the raw FFI surface.

Part of the baracuda workspace.

License

Dual MIT / Apache-2.0.