LSTM Layer Functions

riscv_nmsis_nn_status riscv_lstm_unidirectional_s16_s8(nmsis_nn_lstm_context *scratch_buffers, const int8_t *input_data, const nmsis_nn_lstm_dims *lstm_dims, const int8_t *in_to_in_weights, const int8_t *in_to_forget_weights, const int8_t *in_to_cell_weights, const int8_t *in_to_out_weights, const int8_t *recurrent_to_in_weights, const int8_t *recurrent_to_forget_weights, const int8_t *recurrent_to_cell_weights, const int8_t *recurrent_to_out_weights, const int16_t *cell_to_in_weights, const int16_t *cell_to_forget_weights, const int16_t *cell_to_out_weights, const int8_t *projection_weights, const nmsis_nn_lstm_params *lstm, int8_t *output_state, int16_t *cell_state, int8_t *output_data)
group LSTM

Functions

riscv_nmsis_nn_status riscv_lstm_unidirectional_s16_s8(nmsis_nn_lstm_context *scratch_buffers, const int8_t *input_data, const nmsis_nn_lstm_dims *lstm_dims, const int8_t *in_to_in_weights, const int8_t *in_to_forget_weights, const int8_t *in_to_cell_weights, const int8_t *in_to_out_weights, const int8_t *recurrent_to_in_weights, const int8_t *recurrent_to_forget_weights, const int8_t *recurrent_to_cell_weights, const int8_t *recurrent_to_out_weights, const int16_t *cell_to_in_weights, const int16_t *cell_to_forget_weights, const int16_t *cell_to_out_weights, const int8_t *projection_weights, const nmsis_nn_lstm_params *lstm, int8_t *output_state, int16_t *cell_state, int8_t *output_data)

LSTM unidirectional function with 8 bit input and output and 16 bit gate output Peephole connections, projection, clipping, combined input/forget gate and layer normalization are not supported.

1 Input to input weight can not be nullptr. Otherwise nullptr for combined input/forgat gate. 2 Cell weights are not used and should be nullptr. Otherwise needed for peephole connections. 3 Projection weight is not used and should be nullpr. Otherwise needed for projection.

  1. Supported framework: TensorFlow Lite micro

Note

Following assumptions are done based on LSTM functionality as supported by Keras version 2.9.0 at the time of development. As stated here, https://github.com/tensorflow/community/blob/master/rfcs/20180920-unify-rnn-interface.md Keras’s LSTMCell is equivalent to TensorFlow’s BasicLSTMCell, which does not support peephole, clipping or projection. Layer normalization and combined input/forget gate are not supported either.

Parameters

  • scratch_buffers – [in] Struct containing scratch buffers Expected size for each scratch buffer is lstm_dims->num_batches * lstm_dims->num_outputs.

  • input_data – [in] Pointer to input data

  • lstm_dims – [in] LSTM input parameters related to dimensions

  • input_to_input_weights – [in] Input to input weights

  • input_to_forget_weights – [in] Input to forget weights

  • input_to_cell_weights – [in] Input to cell weights

  • input_to_output_weights – [in] Input to output weights

  • recurrent_to_input_weights – [in] Recurrent to input weights

  • recurrent_to_forget_weights – [in] Recurrent to forget weights

  • recurrent_to_cell_weights – [in] Recurrent to cell weights

  • recurrent_to_output_weights – [in] Recurrent to output weights

  • cell_to_input_weights – [in] Cell to input weights. Not used.

  • cell_to_forget_weights – [in] Cell to forget weights. Not used.

  • cell_to_output_weights – [in] Cell to output weights. Not used.

  • projection_weights – [in] Projection weights. Not used.

  • lstm – [in] LSTM parameters. See struct declaration

  • output_state – [in] Pointer to (recurrent) output state

  • cell_state – [in] Pointer to cell state

  • output_data – [in] Pointer to output state

Returns

The function returns RISCV_NMSIS_NN_SUCCESS