LSTM

Support functions for LSTM. More...

Functions
void	riscv_nn_lstm_calculate_gate_s8_s16 (const int8_t *input, const int8_t *input_to_gate_weights, const int32_t *input_to_gate_bias, const nmsis_nn_scaling input_to_gate_scaling, const int8_t *output_state, const int8_t *recurrent_to_gate_weights, const int32_t *recurrent_to_gate_bias, const nmsis_nn_scaling recurrent_to_gate, const int32_t n_batch, const int32_t n_input, const int32_t n_output, const int32_t n_cell, const riscv_nn_activation_type activation_type, int16_t *gate)
	Updates a LSTM gate for an iteration step of LSTM function, int8x8_16 version. More...
riscv_nmsis_nn_status	riscv_nn_lstm_step_s8_s16 (const int8_t *input, const int8_t *input_to_input_weight, const int8_t *input_to_forget_weight, const int8_t *input_to_cell_weight, const int8_t *input_to_output_weight, const int8_t *recurrent_to_input_weight, const int8_t *recurrent_to_forget_weight, const int8_t *recurrent_to_cell_weight, const int8_t *recurrent_to_output_weight, const nmsis_nn_lstm_params *lstm, const int n_batch, const int n_cell, const int n_input, const int n_output, int8_t *output_state, int16_t *cell_state, int8_t *output, nmsis_nn_lstm_context *scratch_buffers)
	Update LSTM function for an iteration step. More...
void	riscv_nn_lstm_update_cell_state_s16 (const int32_t n_block, const int32_t cell_state_scale, int16_t *cell_state, const int16_t *input_gate, const int16_t *forget_gate, const int16_t *cell_gate)
	Update cell state for a single LSTM iteration step, int8x8_16 version. More...
void	riscv_nn_lstm_update_output_s8_s16 (const int n_batch, const int n_cell, int16_t *cell_state, const int32_t cell_state_scale, const int16_t *output_gate, const nmsis_nn_scaling hidden_scaling, const int32_t hidden_offset, int8_t *output_state, int16_t *cell_gate_scratch)
	Calculate the output state tensor of an LSTM step, s8 input/output and s16 weight version. More...
void	riscv_nn_vec_mat_mul_result_acc_s8 (const int8_t *lhs_in, const int8_t *rhs_in, const int32_t *bias, int16_t *dst, const int32_t dst_offset, const int32_t dst_multiplier, const int32_t dst_shift, const int32_t rhs_cols, const int32_t rhs_rows, const int32_t batch)
	The result of the multiplication is accumulated to the passed result buffer. Multiplies a matrix by a "batched" vector (i.e. a matrix with a batch dimension composed by input vectors independent from each other). More...

Detailed Description

Support functions for LSTM.

Function Documentation

riscv_nn_lstm_calculate_gate_s8_s16()

void riscv_nn_lstm_calculate_gate_s8_s16	(	const int8_t *	input,
		const int8_t *	input_to_gate_weights,
		const int32_t *	input_to_gate_bias,
		const nmsis_nn_scaling	input_to_gate_scaling,
		const int8_t *	output_state,
		const int8_t *	recurrent_to_gate_weights,
		const int32_t *	recurrent_to_gate_bias,
		const nmsis_nn_scaling	recurrent_to_gate_scaling,
		const int32_t	n_batch,
		const int32_t	n_input,
		const int32_t	n_output,
		const int32_t	n_cell,
		const riscv_nn_activation_type	activation_type,
		int16_t *	gate
	)

Updates a LSTM gate for an iteration step of LSTM function, int8x8_16 version.

param[in] input Input data param[in] input_to_gate_weights Input to gate weights param[in] input_to_gate_bias Input to gate weights param[in] input_to_gate_scaling Input to gate scaling param[in] activation Actival min and max values param[in] output_state Output state param[in] recurrent_to_gate_weights Recurrent to gate weights param[in] recurrent_to_gate_bias Recurrent to gate bias param[in] recurrent_to_gate_scaling Recurrent to gate scaling param[in] n_batch Batch size param[in] n_input Input size param[out] n_output Output size param[in] activation_type Activation type (sigmoid or tanh) param[out] n_cell Cell size

riscv_nn_lstm_step_s8_s16()

riscv_nmsis_nn_status riscv_nn_lstm_step_s8_s16	(	const int8_t *	input,
		const int8_t *	input_to_input_weight,
		const int8_t *	input_to_forget_weight,
		const int8_t *	input_to_cell_weight,
		const int8_t *	input_to_output_weight,
		const int8_t *	recurrent_to_input_weight,
		const int8_t *	recurrent_to_forget_weight,
		const int8_t *	recurrent_to_cell_weight,
		const int8_t *	recurrent_to_output_weight,
		const nmsis_nn_lstm_params *	lstm,
		const int	n_batch,
		const int	n_cell,
		const int	n_input,
		const int	n_output,
		int8_t *	output_state,
		int16_t *	cell_state,
		int8_t *	output,
		nmsis_nn_lstm_context *	scratch_buffers
	)

Update LSTM function for an iteration step.

param[in] input Input data param[in] input_to_input_weight Input to input gate weights param[in] input_to_forget_weight Input to forget gate weights param[in] input_to_cell_weight Input to cell gate weights param[in] input_to_output_weight Input to output weights param[in] recurrent_to_input_weight Recurrent signal to input weights param[in] recurrent_to_forget_weight Recurrent signal to forget gate weights param[in] recurrent_to_cell_weight Recurrent signal to cell gate weighst param[in] recurrent_to_output_weight Recurrent signal to output weights param[in] lstm LSTM parameters param[in] n_batch Batch size param[in] n_cell Cell size param[in] n_input Input size param[in] n_output Output size param[out] output_state Output state param[out] cell_state Internal state param[out] output Output signal param[in] *scratch_buffers Struct containing scratch buffers

riscv_nn_lstm_update_cell_state_s16()

void riscv_nn_lstm_update_cell_state_s16	(	const int32_t	n_block,
		const int32_t	cell_state_scale,
		int16_t *	cell_state,
		const int16_t *	input_gate,
		const int16_t *	forget_gate,
		const int16_t *	cell_gate
	)

Update cell state for a single LSTM iteration step, int8x8_16 version.

Parameters

[in]	n_block	total number of cells for all batches
[in]	cell_state_scale	Scaling factor of cell state
[in]	cell_state	Input/output vector, size n_batch*n_cell
[in]	input_gate	Input vector of size n_block
[in]	forget_gate	Input/scratch vector of size n_block, always modified
[in]	cell_gate	Input vector of size, n_block

riscv_nn_lstm_update_output_s8_s16()

void riscv_nn_lstm_update_output_s8_s16	(	const int	n_batch,
		const int	n_cell,
		int16_t *	cell_state,
		const int32_t	cell_state_scale,
		const int16_t *	output_gate,
		const nmsis_nn_scaling	hidden_scale,
		const int32_t	hidden_offset,
		int8_t *	output_state,
		int16_t *	cell_gate_scratch
	)

Calculate the output state tensor of an LSTM step, s8 input/output and s16 weight version.

Parameters

[in]	n_batch	The number of distinct vectors in each array
[in]	n_cell	Number of cells
[in,out]	cell_state	Cell state, size n_batch*n_cell
[in]	cell_state_scale	Scaling of cell_state
[in]	output_gate	Output gate
[in]	hidden_scale	Effective scaling of cell_state .* output_gate
[in]	hidden_offset	Zero point for cell_state .* output_gate
[out]	output_state	Output state
[in]	cell_gate_scratch	Scratch buffer

riscv_nn_vec_mat_mul_result_acc_s8()

void riscv_nn_vec_mat_mul_result_acc_s8	(	const int8_t *	lhs_in,
		const int8_t *	rhs_in,
		const int32_t *	bias,
		int16_t *	dst,
		const int32_t	dst_offset,
		const int32_t	dst_multiplier,
		const int32_t	dst_shift,
		const int32_t	rhs_cols,
		const int32_t	rhs_rows,
		const int32_t	batch
	)

The result of the multiplication is accumulated to the passed result buffer. Multiplies a matrix by a "batched" vector (i.e. a matrix with a batch dimension composed by input vectors independent from each other).

Parameters

[in]	lhs_in	Batched vector
[in]	rhs_in	Weights - input matrix (H(Rows)xW(Columns))
[in]	bias	Bias vector
[out]	dst	Output
[in]	dst_offset	Output offset
[in]	dst_multiplier	Multiplier for quantization
[in]	dst_shift	Shift for quantization
[in]	rhs_cols	Vector/matarix column length
[in]	rhs_rows	Row count of matrix
[in]	batch	Batch size