97 lines
4.8 KiB
ReStructuredText
97 lines
4.8 KiB
ReStructuredText
BatchToSpace
|
||
=================
|
||
将输入张量在批维度上分块并重新分布到空间维度，同时按照 ``crops`` 对输出空间范围进行裁剪。
|
||
|
||
.. math::
|
||
|
||
\begin{aligned}
|
||
N_{\text{out}} &= \frac{N}{b_h \times b_w}, \\
|
||
H_{\text{out}} &= b_h \times H - c_{\text{top}} - c_{\text{bottom}}, \\
|
||
W_{\text{out}} &= b_w \times W - c_{\text{left}} - c_{\text{right}}, \\
|
||
ext{output}[n, h, w, c] &= \text{input}[n', h', w', c]
|
||
\end{aligned}
|
||
|
||
其中 :math:`N, H, W, C` 分别表示输入的 batch、高度、宽度和通道数；:math:`b_h, b_w` 为 ``block_size``；:math:`c_{*}` 来源于 ``crops``；:math:`n', h', w'` 由 ``BatchToSpace`` 映射关系确定。
|
||
|
||
输入：
|
||
- **input** - 输入数据地址。
|
||
- **input_shape** - 输入形状，格式为 ``[batch, height, width, channel]``。
|
||
- **block_size** - 分块因子，格式为 ``[block_h, block_w]``。
|
||
- **crops** - 裁剪参数，格式为 ``[top, bottom, left, right]``。
|
||
- **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``
|
||
|
||
.. note::
|
||
- FT78NE 支持 fp32、fp64、cplx64、cplx128、int16、int8、int32 数据类型。
|
||
- MT7004 支持 fp32、fp16、cplx64、int16、int32 数据类型。
|
||
|
||
|
||
**共享存储版本:**
|
||
|
||
.. c:function:: void i8_batchtospace_s(int8_t *input, int8_t *output, const int *input_shape, const int *block_size, const int *crops, int data_size, int core_mask)
|
||
.. c:function:: void i16_batchtospace_s(int16_t *input, int16_t *output, const int *input_shape, const int *block_size, const int *crops, int data_size, int core_mask)
|
||
.. c:function:: void i32_batchtospace_s(int32_t *input, int32_t *output, const int *input_shape, const int *block_size, const int *crops, int data_size, int core_mask)
|
||
.. c:function:: void hp_batchtospace_s(half *input, half *output, const int *input_shape, const int *block_size, const int *crops, int data_size, int core_mask)
|
||
.. c:function:: void fp_batchtospace_s(float *input, float *output, const int *input_shape, const int *block_size, const int *crops, int data_size, int core_mask)
|
||
.. c:function:: void dp_batchtospace_s(double *input, double *output, const int *input_shape, const int *block_size, const int *crops, int data_size, int core_mask)
|
||
.. c:function:: void c64_batchtospace_s(float *input, float *output, const int *input_shape, const int *block_size, const int *crops, int data_size, int core_mask)
|
||
.. c:function:: void c128_batchtospace_s(double *input, double *output, const int *input_shape, const int *block_size, const int *crops, int data_size, int core_mask)
|
||
|
||
|
||
**C 调用示例：**
|
||
|
||
.. code-block:: c
|
||
:linenos:
|
||
:emphasize-lines: 11
|
||
|
||
// 多核（共享存储）示例
|
||
#include <stdio.h>
|
||
|
||
int main(int argc, char *argv[]) {
|
||
float *input = (float *)0xA0000000; // 输入在 DDR 空间
|
||
float *output = (float *)0xB0000000;
|
||
int input_shape[4] = {400, 2, 2, 3};
|
||
int block_size[2] = {2, 2};
|
||
int crops[4] = {0, 0, 0, 0};
|
||
int core_mask = 0xff;
|
||
fp_batchtospace_s(input, output, input_shape, block_size, crops, sizeof(float), core_mask);
|
||
return 0;
|
||
}
|
||
|
||
|
||
**私有存储版本:**
|
||
|
||
.. c:function:: void i8_batchtospace_p(int8_t *input, int8_t *output, const int *input_shape, const int *block_size, const int *crops, int data_size)
|
||
.. c:function:: void i16_batchtospace_p(int16_t *input, int16_t *output, const int *input_shape, const int *block_size, const int *crops, int data_size)
|
||
.. c:function:: void i32_batchtospace_p(int32_t *input, int32_t *output, const int *input_shape, const int *block_size, const int *crops, int data_size)
|
||
.. c:function:: void hp_batchtospace_p(half *input, half *output, const int *input_shape, const int *block_size, const int *crops, int data_size)
|
||
.. c:function:: void fp_batchtospace_p(float *input, float *output, const int *input_shape, const int *block_size, const int *crops, int data_size)
|
||
.. c:function:: void dp_batchtospace_p(double *input, double *output, const int *input_shape, const int *block_size, const int *crops, int data_size)
|
||
.. c:function:: void c64_batchtospace_p(float *input, float *output, const int *input_shape, const int *block_size, const int *crops, int data_size)
|
||
.. c:function:: void c128_batchtospace_p(double *input, double *output, const int *input_shape, const int *block_size, const int *crops, int data_size)
|
||
|
||
**C 调用示例：**
|
||
|
||
.. code-block:: c
|
||
:linenos:
|
||
:emphasize-lines: 10
|
||
|
||
// 单核（私有存储）示例
|
||
#include <stdio.h>
|
||
|
||
int main(int argc, char *argv[]) {
|
||
float *input = (float *)0x10000000; // 输入在 L2 空间
|
||
float *output = (float *)0x10010000;
|
||
int input_shape[4] = {400, 2, 2, 3};
|
||
int block_size[2] = {2, 2};
|
||
int crops[4] = {0, 0, 0, 0};
|
||
fp_batchtospace_p(input, output, input_shape, block_size, crops, sizeof(float));
|
||
return 0;
|
||
}
|