|
MayaFlux 1.0.0
Digital-First Multimedia Processing Framework
|
|
inlinenodiscard |
One nn_lstm_sequence dispatch: an LSTM cell over a whole sequence in one workgroup.
The input width equals the hidden width. Weights are in the order of PyTorch: gate rows input, forget, cell, output. w_ih and w_hh are [4 * hidden, hidden], b_ih and b_hh are [4 * hidden]. state holds h then c, 2 * hidden floats, and is read at the start and written at the end so the recurrence continues across dispatches. hs receives the hidden vector of every step.
| x | Device address of the input, [steps, hidden]. |
| hs | Device address of the output, [steps, hidden]. |
| hidden | Hidden width, 1 to lstm_max_hidden. |
| std::invalid_argument | When steps or hidden is out of range. |
Definition at line 311 of file Stages.hpp.
References lstm_max_hidden, lstm_sequence_shader, lstm_workgroup, MayaFlux::Yantra::TensorKernels::detail::make_stage(), and MayaFlux::Yantra::TensorKernels::detail::refuse().
Referenced by MayaFlux::Yantra::SileroVadModel::Impl::run().