NN Activation Functions

The activation functions are used to filter out some input data. They include sigmoid, tanh and ReLU (Rectified Linear Unit) functions. More...

Functions
static void	hpm_nn_activate_s8 (q7_t *in_out, uint32_t size, uint16_t int_bits, riscv_nn_activation_fun act_fun)
	This function uses the sigmoid or tanh function to perform activation for signed 8-bit integer input vectors. More...
static void	hpm_nn_activate_s16 (q15_t *in_out, uint32_t size, uint16_t int_bits, riscv_nn_activation_fun act_fun)
	This function uses sigmoid or tanh function to perform activation for signed 16-bit integer input vectors. More...
static void	hpm_nn_leaky_relu_s8 (q7_t *in_out, uint32_t size, q15_t slope) riscv_nn_leaky_relu_s8(in_out
	This function uses the leaky ReLU function to perform activation for signed 8-bit integer input vectors. More...
static void	hpm_nn_relu_any_s8 (q7_t *data, uint16_t size, q7_t max_val)
	This function uses the ReLU function to perform activation for signed 8-bit integer input vectors. More...
static void	hpm_nn_relu_s8 (q7_t *in_out, uint32_t size)
	This function uses the ReLU function to perform activation for signed 8-bit integer input vectors. More...
static void	hpm_nn_relu_s16 (q15_t *in_out, uint32_t size)
	This function uses the ReLU function to perform activation for signed 16-bit integer input vectors. More...

Variables
static void	size
static void	slope

Detailed Description

The activation functions are used to filter out some input data. They include sigmoid, tanh and ReLU (Rectified Linear Unit) functions.

Function Documentation

hpm_nn_activate_s16()

static void hpm_nn_activate_s16 ( q15_t *  in_out, uint32_t  size, uint16_t  int_bits, riscv_nn_activation_fun  act_fun  )

inlinestatic

#include <middleware/hpm_math/hpm_math.h>

This function uses sigmoid or tanh function to perform activation for signed 16-bit integer input vectors.

Parameters

[in,out]	in_out	pointer of the input/output vector
[in]	size	number of elements in the input/output vector
[in]	int_bits	number of the bits in the integer part, which is supposed to be smaller than 4
[in]	act_fun	selection of activation functions. See the Note below for details.

Note

The availbale activation functions for selection include:

• NN_SIGMOID: Use the sigmoid activation function
• NN_TANH: Use the tanh activation function

hpm_nn_activate_s8()

static void hpm_nn_activate_s8 ( q7_t *  in_out, uint32_t  size, uint16_t  int_bits, riscv_nn_activation_fun  act_fun  )

inlinestatic

#include <middleware/hpm_math/hpm_math.h>

This function uses the sigmoid or tanh function to perform activation for signed 8-bit integer input vectors.

Parameters

[in,out]	in_out	pointer of the input/output vector
[in]	size	number of elements in the input/output vector
[in]	int_bits	number of the bits in the integer part, which is supposed to be smaller than 4
[in]	act_fun	selection of activation functions. See the Note below for details.

Note

The available activation functions for selection include:

• NN_SIGMOID: Use the sigmoid activation function
• NN_TANH: Use the tanh activation function

Example:

#define SIZE 32
q7_t in_out[SIZE] = {...};
hpm_nn_activate_s8(in_out, SIZE, 0, NN_SIGMOID);

hpm_nn_leaky_relu_s8()

static void hpm_nn_leaky_relu_s8 ( q7_t *  in_out, uint32_t  size, q15_t  slope  )

inlinestatic

#include <middleware/hpm_math/hpm_math.h>

This function uses the leaky ReLU function to perform activation for signed 8-bit integer input vectors.

Parameters

[in,out]	in_out	pointer of the input/output vector
[in]	size	number of elements in the input/output vector
[in]	slope	slope value to be multiplied with the negative inputs. The result will be right shifted 15 bits to scale back to signed 8-bit integer.

Example:

#define SIZE 1024
q15_t slope = 16384;
q7_t in_out[SIZE] = {...};
hpm_nn_leaky_relu_s8(in_out, SIZE, slope);

hpm_nn_relu_any_s8()

static void hpm_nn_relu_any_s8 ( q7_t *  data, uint16_t  size, q7_t  max_val  )

inlinestatic

#include <middleware/hpm_math/hpm_math.h>

This function uses the ReLU function to perform activation for signed 8-bit integer input vectors.

Parameters

[in,out]	data	pointer of the input/output vector
[in]	size	number of elements in the input/output vector
[in]	max_val	maximum value to limit the output vector

hpm_nn_relu_s16()

static void hpm_nn_relu_s16 ( q15_t *  in_out, uint32_t  size  )

inlinestatic

#include <middleware/hpm_math/hpm_math.h>

This function uses the ReLU function to perform activation for signed 16-bit integer input vectors.

Parameters

[in,out]	in_out	pointer of the input/output vector
[in]	size	number of elements in the input/output vector

hpm_nn_relu_s8()

static void hpm_nn_relu_s8 ( q7_t *  in_out, uint32_t  size  )

inlinestatic

#include <middleware/hpm_math/hpm_math.h>

This function uses the ReLU function to perform activation for signed 8-bit integer input vectors.

Parameters

[in,out]	in_out	pointer of the input/output vector
[in]	size	number of elements in the input/output vector

Example:

#define H 16
#define W 16
#define CH 5
#define NUM (H * W *CH)
q7_t in_out[NUM] = {...};
hpm_nn_relu_s8(in_out, NUM);

Variable Documentation

size

static int32_t size

#include <middleware/hpm_math/hpm_math.h>

slope

void slope

#include <middleware/hpm_math/hpm_math.h>