109 lines
4.7 KiB
ReStructuredText
109 lines
4.7 KiB
ReStructuredText
Stack
|
||
=================
|
||
|
||
沿一个新轴对输入张量序列进行堆叠。所有输入张量必须具有相同的形状。
|
||
|
||
.. math::
|
||
|
||
\text{若输入 } num\_inputs \text{ 个形状为 } (d_0, d_1, \dots, d_{n-1}) \text{ 的张量，并在 } axis \text{ 处堆叠：}
|
||
|
||
.. math::
|
||
|
||
output\_shape = (d_0, \dots, d_{axis-1}, num\_inputs, d_{axis}, \dots, d_{n-1})
|
||
|
||
输入：
|
||
- **inputs** - 指针数组，包含所有输入张量起始地址的数组。
|
||
- **input_shape** - 所有输入张量共用的形状（shape）数组。
|
||
- **num_inputs** - 输入张量的数量。
|
||
- **axis** - 堆叠操作插入新维度的位置。
|
||
- **input_ndim** - 输入张量的维度数（秩）。
|
||
- **core_mask(int, 可选)** - 核掩码（仅适用于共享存储版本）。
|
||
|
||
输出：
|
||
- **output** - 计算结果存储地址。
|
||
|
||
支持平台：
|
||
``FT78NE``
|
||
``MT7004``
|
||
|
||
.. note::
|
||
- FT78NE 支持 int8, int16, int32, fp32, fp64, cplx64, cplx128
|
||
- MT7004 支持 fp16, fp32, int16, int32, cplx64
|
||
- 所有输入张量必须具有完全一致的 `input_shape`。
|
||
- 输出张量的维度将比输入张量多 1。
|
||
|
||
**共享存储版本:**
|
||
|
||
.. c:function:: void i8_stack_s(int8_t* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, int8_t* output, int core_mask)
|
||
.. c:function:: void i16_stack_s(int16_t* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, int16_t* output, int core_mask)
|
||
.. c:function:: void i32_stack_s(int32_t* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, int32_t* output, int core_mask)
|
||
.. c:function:: void hp_stack_s(half* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, half* output, int core_mask)
|
||
.. c:function:: void fp_stack_s(float* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, float* output, int core_mask)
|
||
.. c:function:: void dp_stack_s(double* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, double* output, int core_mask)
|
||
.. c:function:: void c64_stack_s(float* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, float* output, int core_mask)
|
||
.. c:function:: void c128_stack_s(double* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, double* output, int core_mask)
|
||
|
||
**C调用示例：**
|
||
|
||
.. code-block:: c
|
||
:linenos:
|
||
:emphasize-lines: 20
|
||
|
||
// FT78NE 示例（共享存储）
|
||
#include <stdio.h>
|
||
#include "78NE/utils.h"
|
||
|
||
int main() {
|
||
// 输入形状为 {4, 10, 8, 12}
|
||
int input_shape[] = { 4, 10, 8, 12 };
|
||
int input_ndim = 4;
|
||
int num_inputs = 3;
|
||
int axis = 1;
|
||
int core_mask = 0b1011;
|
||
|
||
float *in0 = (float *)0xA0000000;
|
||
float *in1 = (float *)0xA2000000;
|
||
float *in2 = (float *)0xA4000000;
|
||
float *inputs[] = { in0, in1, in2 };
|
||
float *output = (float *)0xB0000000;
|
||
// 输出形状将变为 {4, 3, 10, 8, 12}
|
||
|
||
fp_stack_s(inputs, input_shape, num_inputs, axis, input_ndim, output, core_mask);
|
||
return 0;
|
||
}
|
||
|
||
|
||
**私有存储版本:**
|
||
|
||
.. c:function:: void i8_stack_p(int8_t* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, int8_t* output)
|
||
.. c:function:: void i16_stack_p(int16_t* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, int16_t* output)
|
||
.. c:function:: void i32_stack_p(int32_t* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, int32_t* output)
|
||
.. c:function:: void hp_stack_p(half* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, half* output)
|
||
.. c:function:: void fp_stack_p(float* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, float* output)
|
||
.. c:function:: void dp_stack_p(double* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, double* output)
|
||
.. c:function:: void c64_stack_p(float* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, float* output)
|
||
.. c:function:: void c128_stack_p(double* inputs[], int* input_shape, int num_inputs, int axis, int input_ndim, double* output)
|
||
|
||
**C调用示例：**
|
||
|
||
.. code-block:: c
|
||
:linenos:
|
||
:emphasize-lines: 15
|
||
|
||
// MT7004 示例（私有存储）
|
||
#include <stdio.h>
|
||
|
||
int main() {
|
||
int input_shape[] = { 2, 3, 4, 5 };
|
||
int input_ndim = 4;
|
||
int num_inputs = 2;
|
||
int axis = 0;
|
||
|
||
float *in0 = (float *)0x10810000;
|
||
float *in1 = (float *)0x10812000;
|
||
float *inputs[] = { in0, in1 };
|
||
float *output = (float *)0x10820000;
|
||
|
||
fp_stack_p(inputs, input_shape, num_inputs, axis, input_ndim, output);
|
||
return 0;
|
||
} |