SpaceToBatch

将输入在空间维度按照 block_size 分块并重新排列到 batch 维度,同时根据 paddings 执行必要的零填充。

输入:
  • input - 输入数据地址。

  • block_size - 分块因子,格式为 [block_h, block_w]

  • paddings - 填充参数,格式为 [top, bottom, left, right]

  • input_shape - 输入形状,格式为 [batch, height, width, channel]

  • data_size - 单个元素字节数(例如 sizeof(float))。

  • core_mask(int, 可选) - 核掩码(仅适用于共享存储版本)。

输出:
  • output - 输出数据地址。

支持平台:

FT78NE MT7004

备注

  • FT78NE 支持的数据类型: fp32、fp64、cplx64、cplx128、int16、int8、int32。

  • MT7004 支持的数据类型: fp32、fp16、cplx64、int16、int32。

共享存储版本:

void i8_spacetobatch_s(int8_t *input, int8_t *output, const int *block_size, const int *paddings, const int *input_shape, int data_size, int core_mask)
void i16_spacetobatch_s(int16_t *input, int16_t *output, const int *block_size, const int *paddings, const int *input_shape, int data_size, int core_mask)
void i32_spacetobatch_s(int32_t *input, int32_t *output, const int *block_size, const int *paddings, const int *input_shape, int data_size, int core_mask)
void hp_spacetobatch_s(half *input, half *output, const int *block_size, const int *paddings, const int *input_shape, int data_size, int core_mask)
void fp_spacetobatch_s(float *input, float *output, const int *block_size, const int *paddings, const int *input_shape, int data_size, int core_mask)
void dp_spacetobatch_s(double *input, double *output, const int *block_size, const int *paddings, const int *input_shape, int data_size, int core_mask)
void c64_spacetobatch_s(float *input, float *output, const int *block_size, const int *paddings, const int *input_shape, int data_size, int core_mask)
void c128_spacetobatch_s(double *input, double *output, const int *block_size, const int *paddings, const int *input_shape, int data_size, int core_mask)

C 调用示例:

 1// FT78NE 多核示例
 2#include <stdio.h>
 3
 4int main(int argc, char *argv[]) {
 5    float *input = (float *)0xA0000000;    // 输入在 DDR 地址 0xA0000000
 6    float *output = (float *)0xB0000000;   // 输出在 DDR 地址 0xB0000000
 7    int block_size[2] = {2, 2};
 8    int paddings[4] = {0, 0, 0, 0};
 9    int input_shape[4] = {1, 100, 10, 10};
10    int core_mask = 0xff;
11    fp_spacetobatch_s(input, output, block_size, paddings, input_shape, sizeof(float), core_mask);
12    return 0;
13}

私有存储版本:

void i8_spacetobatch_p(int8_t *input, int8_t *output, const int *block_size, const int *paddings, const int *input_shape, int data_size)
void i16_spacetobatch_p(int16_t *input, int16_t *output, const int *block_size, const int *paddings, const int *input_shape, int data_size)
void i32_spacetobatch_p(int32_t *input, int32_t *output, const int *block_size, const int *paddings, const int *input_shape, int data_size)
void hp_spacetobatch_p(half *input, half *output, const int *block_size, const int *paddings, const int *input_shape, int data_size)
void fp_spacetobatch_p(float *input, float *output, const int *block_size, const int *paddings, const int *input_shape, int data_size)
void dp_spacetobatch_p(double *input, double *output, const int *block_size, const int *paddings, const int *input_shape, int data_size)
void c64_spacetobatch_p(float *input, float *output, const int *block_size, const int *paddings, const int *input_shape, int data_size)
void c128_spacetobatch_p(double *input, double *output, const int *block_size, const int *paddings, const int *input_shape, int data_size)

C 调用示例:

 1// FT78NE 单核示例
 2#include <stdio.h>
 3
 4int main(int argc, char *argv[]) {
 5    float *input = (float *)0x10000000;   // 单核版本:输入放在 L2 地址 0x10000000
 6    float *output = (float *)0x10040000;  // 单核版本:输出放在 L2 地址 0x10040000
 7    int block_size[2] = {2, 2};
 8    int paddings[4] = {0, 0, 0, 0};
 9    int input_shape[4] = {1, 100, 10, 10};
10    fp_spacetobatch_p(input, output, block_size, paddings, input_shape, sizeof(float));
11    return 0;
12}