97 lines
4.3 KiB
ReStructuredText
97 lines
4.3 KiB
ReStructuredText
SliceFusion
|
||
=================
|
||
|
||
从输入张量中提取一个子集(切片)。根据给定的起始索引(begin)和切片大小(size),在每个维度上截取数据。
|
||
|
||
.. math::
|
||
|
||
output[i_0, i_1, \dots, i_{n-1}] = input[begin_0 + i_0, begin_1 + i_1, \dots, begin_{n-1} + i_{n-1}]
|
||
|
||
其中 $n$ 为维度数(ndim),且满足 $0 \le i_k < size_k$。
|
||
|
||
输入:
|
||
- **input** - 输入张量数据地址。
|
||
- **input_shape** - 输入张量的形状数组地址。
|
||
- **ndim** - 输入张量的维度数。
|
||
- **begin** - 切片起始索引数组地址(长度为 ndim)。
|
||
- **size** - 切片大小数组地址(长度为 ndim)。
|
||
- **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(输出张量形状即为 size)。
|
||
|
||
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
|
||
- 切片操作不改变数据数值,仅改变数据的空间排布,常用于特征提取或张量分解。
|
||
|
||
**共享存储版本:**
|
||
|
||
.. c:function:: void i8_slice_s(int8_t* input, int8_t* output, int* input_shape, int ndim, int* begin, int* size, int core_mask)
|
||
.. c:function:: void i16_slice_s(int16_t* input, int16_t* output, int* input_shape, int ndim, int* begin, int* size, int core_mask)
|
||
.. c:function:: void i32_slice_s(int32_t* input, int32_t* output, int* input_shape, int ndim, int* begin, int* size, int core_mask)
|
||
.. c:function:: void hp_slice_s(half* input, half* output, int* input_shape, int ndim, int* begin, int* size, int core_mask)
|
||
.. c:function:: void fp_slice_s(float* input, float* output, int* input_shape, int ndim, int* begin, int* size, int core_mask)
|
||
.. c:function:: void dp_slice_s(double* input, double* output, int* input_shape, int ndim, int* begin, int* size, int core_mask)
|
||
.. c:function:: void c64_slice_s(float* input, float* output, int* input_shape, int ndim, int* begin, int* size, int core_mask)
|
||
.. c:function:: void c128_slice_s(double* input, double* output, int* input_shape, int ndim, int* begin, int* size, int core_mask)
|
||
|
||
**C调用示例:**
|
||
|
||
.. code-block:: c
|
||
:linenos:
|
||
:emphasize-lines: 14
|
||
|
||
// FT78NE 示例(多核并行切片)
|
||
#include <stdio.h>
|
||
#include "78NE/utils.h"
|
||
|
||
int main() {
|
||
float *input = (float *)0xA0000000;
|
||
float *output = (float *)0xB0000000;
|
||
int input_shape[] = {16, 32, 64, 128};
|
||
int begin[] = {2, 4, 8, 16};
|
||
int size[] = {8, 12, 20, 32};
|
||
int ndim = 4;
|
||
int core_mask = 0xFF; // 使用8核并行
|
||
|
||
fp_slice_s(input, output, input_shape, ndim, begin, size, core_mask);
|
||
return 0;
|
||
}
|
||
|
||
|
||
**私有存储版本:**
|
||
|
||
.. c:function:: void i8_slice_p(int8_t* input, int8_t* output, int* input_shape, int ndim, int* begin, int* size)
|
||
.. c:function:: void i16_slice_p(int16_t* input, int16_t* output, int* input_shape, int ndim, int* begin, int* size)
|
||
.. c:function:: void i32_slice_p(int32_t* input, int32_t* output, int* input_shape, int ndim, int* begin, int* size)
|
||
.. c:function:: void hp_slice_p(half* input, half* output, int* input_shape, int ndim, int* begin, int* size)
|
||
.. c:function:: void fp_slice_p(float* input, float* output, int* input_shape, int ndim, int* begin, int* size)
|
||
.. c:function:: void dp_slice_p(double* input, double* output, int* input_shape, int ndim, int* begin, int* size)
|
||
.. c:function:: void c64_slice_p(float* input, float* output, int* input_shape, int ndim, int* begin, int* size)
|
||
.. c:function:: void c128_slice_p(double* input, double* output, int* input_shape, int ndim, int* begin, int* size)
|
||
|
||
**C调用示例:**
|
||
|
||
.. code-block:: c
|
||
:linenos:
|
||
:emphasize-lines: 12
|
||
|
||
// MT7004 示例(单核私有存储切片)
|
||
#include <stdio.h>
|
||
|
||
int main() {
|
||
float *input = (float *)0x10000000;
|
||
float *output = (float *)0x10010000;
|
||
int input_shape[] = {4, 8, 16, 32};
|
||
int begin[] = {1, 2, 3, 4};
|
||
int size[] = {2, 3, 6, 8};
|
||
int ndim = 4;
|
||
|
||
fp_slice_p(input, output, input_shape, ndim, begin, size);
|
||
return 0;
|
||
} |