Softmax Functions

Functions
void	riscv_softmax_q15 (const q15_t *vec_in, const uint16_t dim_vec, q15_t *p_out)
	Q15 softmax function. More...
void	riscv_softmax_q7 (const q7_t *vec_in, const uint16_t dim_vec, q7_t *p_out)
	Q7 softmax function. More...
riscv_nmsis_nn_status	riscv_softmax_s16 (const int16_t *input, const int32_t num_rows, const int32_t row_size, const int32_t mult, const int32_t shift, const nmsis_nn_softmax_lut_s16 *softmax_params, int16_t *output)
	S16 softmax function. More...
void	riscv_softmax_s8 (const int8_t *input, const int32_t num_rows, const int32_t row_size, const int32_t mult, const int32_t shift, const int32_t diff_min, int8_t *output)
	S8 softmax function. More...
void	riscv_softmax_s8_s16 (const int8_t *input, const int32_t num_rows, const int32_t row_size, const int32_t mult, const int32_t shift, const int32_t diff_min, int16_t *output)
	S8 to s16 softmax function. More...
void	riscv_softmax_u8 (const uint8_t *input, const int32_t num_rows, const int32_t row_size, const int32_t mult, const int32_t shift, const int32_t diff_min, uint8_t *output)
	U8 softmax function. More...
void	riscv_softmax_with_batch_q7 (const q7_t *vec_in, const uint16_t nb_batches, const uint16_t dim_vec, q7_t *p_out)
	Q7 softmax function with batch parameter. More...

Detailed Description

Function Documentation

riscv_softmax_q15()

void riscv_softmax_q15	(	const q15_t *	vec_in,
		const uint16_t	dim_vec,
		q15_t *	p_out
	)

Q15 softmax function.

Parameters

[in]	vec_in	pointer to input vector
[in]	dim_vec	input vector dimention
[out]	p_out	pointer to output vector

Here, instead of typical e based softmax, we use 2-based softmax, i.e.,:

y_i = 2^(x_i) / sum(2^x_j)

The relative output will be different here. But mathematically, the gradient will be the same with a log(2) scaling factor.

riscv_softmax_q7()

void riscv_softmax_q7	(	const q7_t *	vec_in,
		const uint16_t	dim_vec,
		q7_t *	p_out
	)

Q7 softmax function.

Parameters

[in]	vec_in	pointer to input vector
[in]	dim_vec	input vector dimention
[out]	p_out	pointer to output vector

Here, instead of typical natural logarithm e based softmax, we use 2-based softmax here, i.e.,:

y_i = 2^(x_i) / sum(2^x_j)

The relative output will be different here. But mathematically, the gradient will be the same with a log(2) scaling factor.

riscv_softmax_s16()

riscv_nmsis_nn_status riscv_softmax_s16	(	const int16_t *	input,
		const int32_t	num_rows,
		const int32_t	row_size,
		const int32_t	mult,
		const int32_t	shift,
		const nmsis_nn_softmax_lut_s16 *	softmax_params,
		int16_t *	output
	)

S16 softmax function.

Parameters

[in]	input	Pointer to the input tensor
[in]	num_rows	Number of rows in the input tensor
[in]	row_size	Number of elements in each input row
[in]	mult	Input quantization multiplier
[in]	shift	Input quantization shift within the range [0, 31]
[in]	softmax_params	Softmax s16 layer parameters with two pointers to LUTs speficied below. For indexing the high 9 bits are used and 7 remaining for interpolation. That means 512 entries for the 9-bit indexing and 1 extra for interpolation, i.e. 513 values for each LUT. Lookup table for exp(x), where x uniform distributed between [-10.0 , 0.0] Lookup table for 1 / (1 + x), where x uniform distributed between [0.0 , 1.0]
[out]	output	Pointer to the output tensor

Returns

The function returns RISCV_NMSIS_NN_ARG_ERROR Argument error check failed RISCV_NMSIS_NN_SUCCESS - Successful operation

Note

Supported framework: TensorFlow Lite micro (bit-accurate)

riscv_softmax_s8()

void riscv_softmax_s8	(	const int8_t *	input,
		const int32_t	num_rows,
		const int32_t	row_size,
		const int32_t	mult,
		const int32_t	shift,
		const int32_t	diff_min,
		int8_t *	output
	)

S8 softmax function.

Parameters

[in]	input	Pointer to the input tensor
[in]	num_rows	Number of rows in the input tensor
[in]	row_size	Number of elements in each input row
[in]	mult	Input quantization multiplier
[in]	shift	Input quantization shift within the range [0, 31]
[in]	diff_min	Minimum difference with max in row. Used to check if the quantized exponential operation can be performed
[out]	output	Pointer to the output tensor

Note

Supported framework: TensorFlow Lite micro (bit-accurate)

riscv_softmax_s8_s16()

void riscv_softmax_s8_s16	(	const int8_t *	input,
		const int32_t	num_rows,
		const int32_t	row_size,
		const int32_t	mult,
		const int32_t	shift,
		const int32_t	diff_min,
		int16_t *	output
	)

S8 to s16 softmax function.

Parameters

[in]	input	Pointer to the input tensor
[in]	num_rows	Number of rows in the input tensor
[in]	row_size	Number of elements in each input row
[in]	mult	Input quantization multiplier
[in]	shift	Input quantization shift within the range [0, 31]
[in]	diff_min	Minimum difference with max in row. Used to check if the quantized exponential operation can be performed
[out]	output	Pointer to the output tensor

Note

Supported framework: TensorFlow Lite micro (bit-accurate)

riscv_softmax_u8()

void riscv_softmax_u8	(	const uint8_t *	input,
		const int32_t	num_rows,
		const int32_t	row_size,
		const int32_t	mult,
		const int32_t	shift,
		const int32_t	diff_min,
		uint8_t *	output
	)

U8 softmax function.

Parameters

[in]	input	Pointer to the input tensor
[in]	num_rows	Number of rows in the input tensor
[in]	row_size	Number of elements in each input row
[in]	mult	Input quantization multiplier
[in]	shift	Input quantization shift within the range [0, 31]
[in]	diff_min	Minimum difference with max in row. Used to check if the quantized exponential operation can be performed
[out]	output	Pointer to the output tensor

Note

Supported framework: TensorFlow Lite micro (bit-accurate)

riscv_softmax_with_batch_q7()

void riscv_softmax_with_batch_q7	(	const q7_t *	vec_in,
		const uint16_t	nb_batches,
		const uint16_t	dim_vec,
		q7_t *	p_out
	)

Q7 softmax function with batch parameter.

Parameters

[in]	vec_in	pointer to input vector
[in]	nb_batches	number of batches
[in]	dim_vec	input vector dimention
[out]	p_out	pointer to output vector

Here, instead of typical natural logarithm e based softmax, we use 2-based softmax here, i.e.,:

y_i = 2^(x_i) / sum(2^x_j)

The relative output will be different here. But mathematically, the gradient will be the same with a log(2) scaling factor.