780 lines
33 KiB
C
780 lines
33 KiB
C
#pragma once
|
|
#include "dl_lib_matrix3d.h"
|
|
|
|
typedef int16_t qtp_t;
|
|
|
|
/**
|
|
* Matrix for input, filter, and output
|
|
* @Warning: the sequence of variables is fixed, cannot be modified, otherwise there will be errors in
|
|
* some handwrite xtensa instruction functions
|
|
*/
|
|
typedef struct
|
|
{
|
|
/******* fix start *******/
|
|
int w; /*!< Width */
|
|
int h; /*!< Height */
|
|
int c; /*!< Channel */
|
|
int n; /*!< Number of filter, input and output must be 1 */
|
|
int stride; /*!< Step between lines */
|
|
int exponent; /*!< Exponent for quantization */
|
|
qtp_t *item; /*!< Data */
|
|
/******* fix end *******/
|
|
} dl_matrix3dq_t;
|
|
|
|
#ifndef DL_QTP_SHIFT
|
|
#define DL_QTP_SHIFT 15
|
|
#define DL_ITMQ(m, x, y) m->itemq[(y) + (x)*m->stride]
|
|
#define DL_QTP_RANGE ((1 << DL_QTP_SHIFT) - 1)
|
|
#define DL_QTP_MAX 32767
|
|
#define DL_QTP_MIN -32768
|
|
|
|
#define DL_QTP_EXP_NA 255 //non-applicable exponent because matrix is null
|
|
|
|
#define DL_SHIFT_AUTO 32
|
|
#endif
|
|
|
|
/**
|
|
* Implementation of matrix relative operations
|
|
*/
|
|
typedef enum
|
|
{
|
|
DL_C_IMPL = 0, /*!< ANSI C */
|
|
DL_XTENSA_IMPL = 1 /*!< Handwrite xtensa instruction */
|
|
} dl_conv_mode;
|
|
|
|
|
|
/**
|
|
* Configuration of mobilenet operation
|
|
*/
|
|
typedef struct
|
|
{
|
|
int stride_x; /*!< Strides of width */
|
|
int stride_y; /*!< Strides of height */
|
|
dl_padding_type padding; /*!< Padding type */
|
|
dl_conv_mode mode; /*!< Implementation mode */
|
|
int dilate_exponent; /*!< Exponent of dilation filter */
|
|
int depthwise_exponent; /*!< Exponent of depthwise filter */
|
|
int compress_exponent; /*!< Exponent of compress filter */
|
|
} dl_matrix3dq_mobilenet_config_t;
|
|
|
|
//
|
|
// Utility
|
|
//
|
|
|
|
/*
|
|
* @brief Allocate a 3d quantised matrix
|
|
*
|
|
* @param n Number of filters, for input and output, should be 1
|
|
* @param w Width of matrix
|
|
* @param h Height of matrix
|
|
* @param c Channel of matrix
|
|
* @param e Exponent of matrix data
|
|
* @return 3d quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dq_alloc(int n, int w, int h, int c, int e);
|
|
|
|
/*
|
|
* @brief Free a 3d quantized matrix
|
|
*
|
|
* @param m 3d quantised matrix
|
|
*/
|
|
void dl_matrix3dq_free(dl_matrix3dq_t *m);
|
|
|
|
/**
|
|
* @brief Copy a range of items from an existing matrix to a preallocated matrix
|
|
*
|
|
* @param dst The resulting slice matrix
|
|
* @param src Old matrix to slice.
|
|
* @param x X-offset of the origin of the returned matrix within the sliced matrix
|
|
* @param y Y-offset of the origin of the returned matrix within the sliced matrix
|
|
* @param w Width of the resulting matrix
|
|
* @param h Height of the resulting matrix
|
|
*/
|
|
void dl_matrix3dq_slice_copy(dl_matrix3dq_t *dst, dl_matrix3dq_t *src, int x, int y, int w, int h);
|
|
|
|
/**
|
|
* @brief Transform a fixed point matrix to a float point matrix
|
|
*
|
|
* @param m Quantized matrix
|
|
* @return Float point matrix
|
|
*/
|
|
dl_matrix3d_t *dl_matrix3d_from_matrixq(dl_matrix3dq_t *m);
|
|
|
|
/**
|
|
* @brief Transform a float point matrix to a fixed point matrix with pre-defined exponent
|
|
*
|
|
* @param m Float point matrix
|
|
* @param exponent Exponent for resulting matrix
|
|
* @return Fixed point matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrixq_from_matrix3d_qmf(dl_matrix3d_t *m, int exponent);
|
|
|
|
/**
|
|
* @brief Transform a float point matrix to a fixed point matrix. The exponent is defined by the distribution of the input matrix.
|
|
*
|
|
* @param m Float point matrix
|
|
* @return Fixed point matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrixq_from_matrix3d(dl_matrix3d_t *m);
|
|
|
|
qtp_t dl_matrix3dq_quant_range_exceeded_checking(int64_t value, char *location);
|
|
|
|
/**
|
|
* @brief Reform a quantized matrix with exponent
|
|
*
|
|
* @param out Preallocated resulting matrix
|
|
* @param in Input matrix
|
|
* @param exponent Exponent for resulting matrix
|
|
*/
|
|
void dl_matrix3dq_shift_exponent(dl_matrix3dq_t *out, dl_matrix3dq_t *in, int exponent);
|
|
|
|
/**
|
|
* @brief Do batch normalization for a quantized matrix
|
|
*
|
|
* @param m Input and output quantized matrix, data will be updated
|
|
* @param scale Scale of batch-norm
|
|
* @param offset Offset of batch-norm
|
|
*/
|
|
void dl_matrix3dq_batch_normalize(dl_matrix3dq_t *m, dl_matrix3dq_t *scale, dl_matrix3dq_t *offset);
|
|
|
|
/**
|
|
* @brief Add two quantized matrix with a pre-defined exponent
|
|
*
|
|
* @param in_1 Adder 1
|
|
* @param in_2 Adder 2
|
|
* @param exponent Exponent for resulting matrix
|
|
* @return Result of accumulation of two matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dq_add(dl_matrix3dq_t *in_1, dl_matrix3dq_t *in_2, int exponent);
|
|
|
|
//
|
|
// Activation
|
|
//
|
|
/**
|
|
* @brief Do relu for a quantized matrix
|
|
*
|
|
* @param in Input and output quantized matrix, data will be updated
|
|
*/
|
|
void dl_matrix3dq_relu(dl_matrix3dq_t *in);
|
|
|
|
/**
|
|
* @brief Do relu with clips for a quantized matrix
|
|
*
|
|
* @param in Input and output quantized matrix, data will be updated
|
|
* @param clip Float point value to limit the maximum data
|
|
*/
|
|
void dl_matrix3dq_relu_clip(dl_matrix3dq_t *in, fptp_t clip);
|
|
|
|
/**
|
|
* @brief Do leaky relu for a quantized matrix
|
|
*
|
|
* @param in Input and output quantized matrix, data will be updated
|
|
* @param alpha Float point value to multiply for those less than zero
|
|
* @param clip Float point value to limit the maximum data
|
|
*/
|
|
void dl_matrix3dq_leaky_relu(dl_matrix3dq_t *in, fptp_t alpha, fptp_t clip);
|
|
|
|
/**
|
|
* @brief Do prelu for a quantized matrix
|
|
*
|
|
* @param in Input and output quantized matrix, data will be updated
|
|
* @param alpha Quantized matrix to multiply for those less than zero
|
|
*/
|
|
void dl_matrix3dq_p_relu(dl_matrix3dq_t *in, dl_matrix3dq_t *alpha);
|
|
|
|
//
|
|
// Concat
|
|
//
|
|
/**
|
|
* @brief Concatenate two quantized matrix in channel
|
|
*
|
|
* @param in_1 Quantized matrix to be concatenated
|
|
* @param in_2 Quantized matrix to be concatenated
|
|
* @return Quantized matrix with the same width and height of in_1 and in_2, and with the sum of channel number of in_1 and in_2
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dq_concat(dl_matrix3dq_t *in_1,
|
|
dl_matrix3dq_t *in_2);
|
|
|
|
/**
|
|
* @brief Concatenate four quantized matrix in channel
|
|
*
|
|
* @param in_1 Quantized matrix to be concatenated
|
|
* @param in_2 Quantized matrix to be concatenated
|
|
* @param in_3 Quantized matrix to be concatenated
|
|
* @param in_4 Quantized matrix to be concatenated
|
|
* @return Quantized matrix with the same width and height of all inputs, and with the sum of channel number of all inputs
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dq_concat_4(dl_matrix3dq_t *in_1,
|
|
dl_matrix3dq_t *in_2,
|
|
dl_matrix3dq_t *in_3,
|
|
dl_matrix3dq_t *in_4);
|
|
|
|
/**
|
|
* @brief Concatenate four quantized matrix in channel
|
|
*
|
|
* @param in_1 Quantized matrix to be concatenated
|
|
* @param in_2 Quantized matrix to be concatenated
|
|
* @param in_3 Quantized matrix to be concatenated
|
|
* @param in_4 Quantized matrix to be concatenated
|
|
* @param in_5 Quantized matrix to be concatenated
|
|
* @param in_6 Quantized matrix to be concatenated
|
|
* @param in_7 Quantized matrix to be concatenated
|
|
* @param in_8 Quantized matrix to be concatenated
|
|
* @return Quantized matrix with the same width and height of all inputs, and with the sum of channel number of all inputs
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dq_concat_8(dl_matrix3dq_t *in_1,
|
|
dl_matrix3dq_t *in_2,
|
|
dl_matrix3dq_t *in_3,
|
|
dl_matrix3dq_t *in_4,
|
|
dl_matrix3dq_t *in_5,
|
|
dl_matrix3dq_t *in_6,
|
|
dl_matrix3dq_t *in_7,
|
|
dl_matrix3dq_t *in_8);
|
|
|
|
//
|
|
// Conv 1x1
|
|
//
|
|
/**
|
|
* @brief Do 1x1 convolution with a quantized matrix
|
|
*
|
|
* @param out Preallocated quantized matrix, size (1, w, h, n)
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 1x1 filter, size (n, 1, 1, c)
|
|
* @param mode Implementation mode
|
|
*/
|
|
void dl_matrix3dqq_conv_1x1(dl_matrix3dq_t *out,
|
|
dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_conv_mode mode);
|
|
|
|
/**
|
|
* @brief Do 1x1 convolution with a quantized matrix, with relu activation
|
|
*
|
|
* @param out Preallocated quantized matrix, size (1, w, h, n)
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 1x1 filter, size (n, 1, 1, c)
|
|
* @param mode Implementation mode
|
|
*/
|
|
void dl_matrix3dqq_conv_1x1_with_relu(dl_matrix3dq_t *out,
|
|
dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_conv_mode mode);
|
|
|
|
/**
|
|
* @brief Do 1x1 convolution with a quantized matrix, with bias adding
|
|
*
|
|
* @param out Preallocated quantized matrix, size (1, w, h, n)
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 1x1 filter, size (n, 1, 1, c)
|
|
* @param bias Bias, size (1, 1, 1, n)
|
|
* @param mode Implementation mode
|
|
* @param name Layer name to debug
|
|
*/
|
|
void dl_matrix3dqq_conv_1x1_with_bias(dl_matrix3dq_t *out,
|
|
dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_matrix3dq_t *bias,
|
|
dl_conv_mode mode,
|
|
char *name);
|
|
|
|
/**
|
|
* @brief Do 1x1 convolution with a quantized matrix, with bias adding and relu activation
|
|
*
|
|
* @param out Preallocated quantized matrix, size (1, w, h, n)
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 1x1 filter, size (n, 1, 1, c)
|
|
* @param bias Bias, size (1, 1, 1, n)
|
|
* @param mode Implementation mode
|
|
*/
|
|
void dl_matrix3dqq_conv_1x1_with_bias_relu(dl_matrix3dq_t *out,
|
|
dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_matrix3dq_t *bias,
|
|
dl_conv_mode mode);
|
|
|
|
void dl_matrix3dqq_conv_1x1_with_prelu(dl_matrix3dq_t *out,
|
|
dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_matrix3dq_t *prelu,
|
|
dl_conv_mode mode);
|
|
|
|
/**
|
|
* @brief Do 1x1 convolution with an 8-bit fixed point matrix
|
|
*
|
|
* @param out Preallocated quantized matrix, size (1, w, h, n)
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 1x1 filter, size (n, 1, 1, c)
|
|
* @param mode Implementation mode
|
|
*/
|
|
void dl_matrix3duq_conv_1x1(dl_matrix3dq_t *out,
|
|
dl_matrix3du_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_conv_mode mode);
|
|
|
|
/**
|
|
* @brief Do 1x1 convolution with an 8-bit fixed point matrix, with bias adding
|
|
*
|
|
* @param out Preallocated quantized matrix, size (1, w, h, n)
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 1x1 filter, size (n, 1, 1, c)
|
|
* @param bias Bias, size (1, 1, 1, n)
|
|
* @param mode Implementation mode
|
|
*/
|
|
void dl_matrix3duq_conv_1x1_with_bias(dl_matrix3dq_t *out,
|
|
dl_matrix3du_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_matrix3dq_t *bias,
|
|
dl_conv_mode mode);
|
|
|
|
//
|
|
// Conv 3x3
|
|
//
|
|
/**
|
|
* @brief Do 3x3 convolution basic operation with a quantized matrix
|
|
*
|
|
* @param out Preallocated quantized matrix
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 3x3 filter, size (n, 3, 3, c)
|
|
* @param stride_x Stride of width
|
|
* @param stride_y Stride of height
|
|
*/
|
|
void dl_matrix3dqq_conv_3x3_op(dl_matrix3dq_t *out,
|
|
dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
int stride_x,
|
|
int stride_y);
|
|
|
|
/**
|
|
* @brief Do 3x3 convolution with a quantized matrix
|
|
*
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 3x3 filter, size (n, 3, 3, c)
|
|
* @param stride_x Stride of width
|
|
* @param stride_y Stride of height
|
|
* @param padding Padding type, 0: valid, 1: same
|
|
* @param exponent Exponent for resulting matrix
|
|
* @return Resulting quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dqq_conv_3x3(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent);
|
|
|
|
/**
|
|
* @brief Do 3x3 convolution with a quantized matrix, with bias adding
|
|
*
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 3x3 filter, size (n, 3, 3, c)
|
|
* @param bias Bias, size (1, 1, 1, n)
|
|
* @param stride_x Stride of width
|
|
* @param stride_y Stride of height
|
|
* @param padding Padding type, 0: valid, 1: same
|
|
* @param exponent Exponent for resulting matrix
|
|
* @return Resulting quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dqq_conv_3x3_with_bias(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *f,
|
|
dl_matrix3dq_t *bias,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent,
|
|
int relu);
|
|
|
|
dl_matrix3dq_t *dl_matrix3duq_conv_3x3_with_bias(dl_matrix3du_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_matrix3dq_t *bias,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent,
|
|
char *name);
|
|
|
|
dl_matrix3dq_t *dl_matrix3duq_conv_3x3_with_bias_prelu(dl_matrix3du_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_matrix3dq_t *bias,
|
|
dl_matrix3dq_t *prelu,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent,
|
|
char *name);
|
|
|
|
//
|
|
// Conv common
|
|
//
|
|
|
|
/**
|
|
* @brief Do a general convolution layer pass, size is (number, width, height, channel)
|
|
*
|
|
* @param in Input image
|
|
* @param filter Weights of the neurons
|
|
* @param bias Bias for the CNN layer.
|
|
* @param stride_x The step length of the convolution window in x(width) direction
|
|
* @param stride_y The step length of the convolution window in y(height) direction
|
|
* @param padding One of VALID or SAME
|
|
* @param mode Do convolution using C implement or xtensa implement, 0 or 1, with respect.
|
|
* If ESP_PLATFORM is not defined, this value is not used.
|
|
* @return The result of CNN layer.
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dqq_conv_common(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_matrix3dq_t *bias,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent,
|
|
dl_conv_mode mode);
|
|
|
|
/**
|
|
* @brief Do a general convolution layer pass for an 8-bit fixed point matrix, size is (number, width, height, channel)
|
|
*
|
|
* @param in Input image
|
|
* @param filter Weights of the neurons
|
|
* @param bias Bias for the CNN layer.
|
|
* @param stride_x The step length of the convolution window in x(width) direction
|
|
* @param stride_y The step length of the convolution window in y(height) direction
|
|
* @param padding One of VALID or SAME
|
|
* @param mode Do convolution using C implement or xtensa implement, 0 or 1, with respect.
|
|
* If ESP_PLATFORM is not defined, this value is not used.
|
|
* @return The result of CNN layer.
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3duq_conv_common(dl_matrix3du_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_matrix3dq_t *bias,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent,
|
|
dl_conv_mode mode);
|
|
|
|
//
|
|
// Depthwise 3x3
|
|
//
|
|
/**
|
|
* @brief Do 3x3 depthwise convolution with an 8-bit fixed point matrix
|
|
*
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 3x3 filter, size (1, 3, 3, c)
|
|
* @param stride_x Stride of width
|
|
* @param stride_y Stride of height
|
|
* @param padding Padding type, 0: valid, 1: same
|
|
* @param exponent Exponent for resulting matrix
|
|
* @return Resulting quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3duq_depthwise_conv_3x3(dl_matrix3du_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent);
|
|
|
|
/**
|
|
* @brief Do 3x3 depthwise convolution with a quantized matrix
|
|
*
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 3x3 filter, size (1, 3, 3, c)
|
|
* @param stride_x Stride of width
|
|
* @param stride_y Stride of height
|
|
* @param padding Padding type, 0: valid, 1: same
|
|
* @param exponent Exponent for resulting matrix
|
|
* @return Resulting quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dqq_depthwise_conv_3x3(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent);
|
|
|
|
#if CONFIG_DEVELOPING_CODE
|
|
dl_matrix3dq_t *dl_matrix3dqq_depthwise_conv_3x3_2(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent);
|
|
|
|
dl_matrix3dq_t *dl_matrix3dqq_depthwise_conv_3x3_3(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent);
|
|
#endif
|
|
|
|
/**
|
|
* @brief Do 3x3 depthwise convolution with a quantized matrix, with bias adding
|
|
*
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 3x3 filter, size (1, 3, 3, c)
|
|
* @param bias Bias, size (1, 1, 1, c)
|
|
* @param stride_x Stride of width
|
|
* @param stride_y Stride of height
|
|
* @param padding Padding type, 0: valid, 1: same
|
|
* @param exponent Exponent for resulting matrix
|
|
* @param relu Whether to use relu activation
|
|
* @return Resulting quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dqq_depthwise_conv_3x3_with_bias(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *f,
|
|
dl_matrix3dq_t *bias,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent,
|
|
int relu);
|
|
|
|
/**
|
|
* @brief Do 3x3 depthwise convolution with a quantized matrix, with bias adding and stride 1
|
|
*
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param filter 3x3 filter, size (1, 3, 3, c)
|
|
* @param bias Bias, size (1, 1, 1, n)
|
|
* @param padding Padding type, 0: valid, 1: same
|
|
* @param exponent Exponent for resulting matrix
|
|
* @param relu Whether to use relu activation
|
|
* @return Resulting quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dqq_depthwise_conv_3x3s1_with_bias(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *f,
|
|
dl_matrix3dq_t *bias,
|
|
dl_padding_type padding,
|
|
int exponent,
|
|
int relu);
|
|
|
|
dl_matrix3dq_t *dl_matrix3dqq_depthwise_conv_3x3_with_prelu(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_matrix3dq_t *prelu,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent);
|
|
|
|
|
|
//
|
|
// Depthwise Common
|
|
//
|
|
#if CONFIG_DEVELOPING_CODE
|
|
dl_matrix3dq_t *dl_matrix3dqq_depthwise_conv_common(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent,
|
|
dl_conv_mode mode);
|
|
|
|
dl_matrix3dq_t *dl_matrix3duq_depthwise_conv_common(dl_matrix3du_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
int exponent,
|
|
dl_conv_mode mode);
|
|
#endif
|
|
|
|
//
|
|
// Dot Product
|
|
//
|
|
|
|
void dl_matrix3dqq_dot_product(dl_matrix3dq_t *out,
|
|
dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_conv_mode mode);
|
|
|
|
//
|
|
// FC
|
|
//
|
|
/**
|
|
* @brief Do fully connected layer forward.
|
|
*
|
|
* @param out Preallocated resulting matrix, size (1, 1, 1, h)
|
|
* @param in Input matrix, size (1, 1, 1, w)
|
|
* @param filter Filter matrix, size (1, w, h, 1)
|
|
* @param mode Implementation mode
|
|
*/
|
|
void dl_matrix3dqq_fc(dl_matrix3dq_t *out,
|
|
dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_conv_mode mode);
|
|
|
|
/**
|
|
* @brief Do fully connected layer forward, with bias adding
|
|
*
|
|
* @param out Preallocated resulting matrix, size (1, 1, 1, h)
|
|
* @param in Input matrix, size (1, 1, 1, w)
|
|
* @param filter Filter matrix, size (1, w, h, 1)
|
|
* @param bias Bias matrix, size (1, 1, 1, h)
|
|
* @param mode Implementation mode
|
|
*/
|
|
void dl_matrix3dqq_fc_with_bias(dl_matrix3dq_t *out,
|
|
dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *filter,
|
|
dl_matrix3dq_t *bias,
|
|
dl_conv_mode mode,
|
|
char *name);
|
|
|
|
//
|
|
// Mobilefaceblock
|
|
//
|
|
/**
|
|
* @brief Do mobilefacenet process with splited pointwise 1x1 convolution, the process sequence is 1x1 pointwise->bn->relu->3x3 depthwise->bn->relu->1x1 pointwise->bn
|
|
*
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param pw_1 Pointwise 1x1 filter, size (n1/2, 1, 1, c)
|
|
* @param pw_2 Pointwise 1x1 filter, size (n1/2, 1, 1, c)
|
|
* @param pw_bias Pointwise bias, size (1, 1, 1, n1)
|
|
* @param dw Depthwise 3x3 filter, size (1, 3, 3, n1)
|
|
* @param dw_bias Depthwise bias, size (1, 1, 1, n1)
|
|
* @param pw_linear_1 Pointwise 1x1 filter, size (n2/2, 1, 1, n1)
|
|
* @param pw_linear_2 Pointwise 1x1 filter, size (n2/2, 1, 1, n1)
|
|
* @param pw_linear_bias Pointwise bias, size (1, 1, 1, n2)
|
|
* @param pw_exponent Exponent for pointwise resulting matrix
|
|
* @param dw_exponent Exponent for depthwise resulting matrix
|
|
* @param pw_linear_exponent Exponent for pointwise resulting matrix
|
|
* @param stride_x Stride of width
|
|
* @param stride_y Stride of height
|
|
* @param padding Padding type, 0: valid, 1: same
|
|
* @param mode Implementation mode
|
|
* @param shortcut Whether has a shortcut at pointwise linear
|
|
* @return Resulting quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dqq_mobilefaceblock_split(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *pw_1,
|
|
dl_matrix3dq_t *pw_2,
|
|
dl_matrix3dq_t *pw_bias,
|
|
dl_matrix3dq_t *dw,
|
|
dl_matrix3dq_t *dw_bias,
|
|
dl_matrix3dq_t *pw_linear_1,
|
|
dl_matrix3dq_t *pw_linear_2,
|
|
dl_matrix3dq_t *pw_linear_bias,
|
|
int pw_exponent,
|
|
int dw_exponent,
|
|
int pw_linear_exponent,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
dl_conv_mode mode,
|
|
int shortcut);
|
|
|
|
/**
|
|
* @brief Do mobilefacenet process, the process sequence is 1x1 pointwise->bn->relu->3x3 depthwise->bn->relu->1x1 pointwise->bn
|
|
*
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param pw Pointwise 1x1 filter, size (n1, 1, 1, c)
|
|
* @param pw_bias Pointwise bias, size (1, 1, 1, n1)
|
|
* @param dw Depthwise 3x3 filter, size (1, 3, 3, n1)
|
|
* @param dw_bias Depthwise bias, size (1, 1, 1, n1)
|
|
* @param pw_linear Pointwise 1x1 filter, size (n2, 1, 1, n1)
|
|
* @param pw_linear_bias Pointwise bias, size (1, 1, 1, n2)
|
|
* @param pw_exponent Exponent for pointwise resulting matrix
|
|
* @param dw_exponent Exponent for depthwise resulting matrix
|
|
* @param pw_linear_exponent Exponent for pointwise resulting matrix
|
|
* @param stride_x Stride of width
|
|
* @param stride_y Stride of height
|
|
* @param padding Padding type, 0: valid, 1: same
|
|
* @param mode Implementation mode
|
|
* @param shortcut Whether has a shortcut at pointwise linear
|
|
* @return Resulting quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dqq_mobilefaceblock(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *pw,
|
|
dl_matrix3dq_t *pw_bias,
|
|
dl_matrix3dq_t *dw,
|
|
dl_matrix3dq_t *dw_bias,
|
|
dl_matrix3dq_t *pw_linear,
|
|
dl_matrix3dq_t *pw_linear_bias,
|
|
int pw_exponent,
|
|
int dw_exponent,
|
|
int pw_linear_exponent,
|
|
int stride_x,
|
|
int stride_y,
|
|
dl_padding_type padding,
|
|
dl_conv_mode mode,
|
|
int shortcut);
|
|
|
|
//
|
|
// Mobilenet
|
|
//
|
|
|
|
/**
|
|
* @brief Do mobilenet process, the process sequence is 1x1 dilated->prelu->3x3 depthwise->prelu->1x1 compress->bias
|
|
*
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @param dilate Pointwise 1x1 filter, size (n1, 1, 1, c)
|
|
* @param dilate_prelu Pointwise prelu, size (1, 1, 1, n1)
|
|
* @param depthwise Depthwise 3x3 filter, size (1, 3, 3, n1)
|
|
* @param depthwise_prelu Depthwise prelu, size (1, 1, 1, n1)
|
|
* @param compress Pointwise 1x1 filter, size (n2, 1, 1, n1)
|
|
* @param bias Pointwise bias, size (1, 1, 1, n2)
|
|
* @param config Mobilenet configuration
|
|
* @return Resulting quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dqq_mobilenet(dl_matrix3dq_t *in,
|
|
dl_matrix3dq_t *dilate,
|
|
dl_matrix3dq_t *dilate_prelu,
|
|
dl_matrix3dq_t *depthwise,
|
|
dl_matrix3dq_t *depth_prelu,
|
|
dl_matrix3dq_t *compress,
|
|
dl_matrix3dq_t *bias,
|
|
dl_matrix3dq_mobilenet_config_t config,
|
|
char *name);
|
|
|
|
/**
|
|
* @brief Do mobilenet process, the process sequence is 1x1 dilated->prelu->3x3 depthwise->prelu->1x1 compress->bias
|
|
*
|
|
* @param in Input matrix, 8-bit fixed point, size (1, w, h, c)
|
|
* @param dilate Pointwise 1x1 filter, size (n1, 1, 1, c)
|
|
* @param dilate_prelu Pointwise prelu, size (1, 1, 1, n1)
|
|
* @param depthwise Depthwise 3x3 filter, size (1, 3, 3, n1)
|
|
* @param depthwise_prelu Depthwise prelu, size (1, 1, 1, n1)
|
|
* @param compress Pointwise 1x1 filter, size (n2, 1, 1, n1)
|
|
* @param bias Pointwise bias, size (1, 1, 1, n2)
|
|
* @param config Mobilenet configuration
|
|
* @return Resulting quantized matrix
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3duq_mobilenet(dl_matrix3du_t *in,
|
|
dl_matrix3dq_t *dilate,
|
|
dl_matrix3dq_t *dilate_prelu,
|
|
dl_matrix3dq_t *depthwise,
|
|
dl_matrix3dq_t *depth_prelu,
|
|
dl_matrix3dq_t *compress,
|
|
dl_matrix3dq_t *bias,
|
|
dl_matrix3dq_mobilenet_config_t config,
|
|
char *name);
|
|
|
|
//
|
|
// Padding
|
|
//
|
|
|
|
dl_error_type dl_matrix3dqq_padding(dl_matrix3dq_t **padded_in,
|
|
dl_matrix3dq_t **out,
|
|
dl_matrix3dq_t *in,
|
|
int out_c,
|
|
int stride_x,
|
|
int stride_y,
|
|
int padding,
|
|
int exponent);
|
|
|
|
dl_error_type dl_matrix3duq_padding(dl_matrix3du_t **padded_in,
|
|
dl_matrix3dq_t **out,
|
|
dl_matrix3du_t *in,
|
|
int out_c,
|
|
int stride_x,
|
|
int stride_y,
|
|
int padding,
|
|
int exponent);
|
|
|
|
//
|
|
// Pooling
|
|
//
|
|
/**
|
|
* @brief Calculate average value of a feature map
|
|
*
|
|
* @param in Input matrix, size (1, w, h, c)
|
|
* @return Resulting matrix, size (1, 1, 1, c)
|
|
*/
|
|
dl_matrix3dq_t *dl_matrix3dq_global_pool(dl_matrix3dq_t *in);
|