forked from FT_DSP/FT_DSP.gitlink.net
92 lines
3.7 KiB
ReStructuredText
92 lines
3.7 KiB
ReStructuredText
FusedBatchNorm
|
||
=================
|
||
对输入张量执行融合批归一化（Fused Batch Normalization），在多核间拆分批量单元并行完成归一化与仿射变换。
|
||
|
||
.. math::
|
||
|
||
\hat{x}_{b,c} = \frac{x_{b,c} - mean_c}{\sqrt{variance_c + \epsilon}}, \quad y_{b,c} = scale_c \cdot \hat{x}_{b,c} + offset_c
|
||
|
||
|
||
输入：
|
||
- **input** - 输入张量首地址，形状为 ``[unit, channel]``。
|
||
- **scale** - 缩放系数数组首地址，长度为 ``channel``。
|
||
- **offset** - 平移系数数组首地址，长度为 ``channel``。
|
||
- **mean** - 归一化均值数组首地址，长度为 ``channel``。
|
||
- **variance** - 归一化方差数组首地址，长度为 ``channel``。
|
||
- **epsilon** - 数值稳定项。
|
||
- **channel** - 通道数。
|
||
- **unit** - 归一化单元数量（批量大小 × 高 × 宽）。
|
||
- **core_mask(int, 可选)** - 核掩码（仅适用于共享存储版本）。
|
||
|
||
输出：
|
||
- **output** - 写回融合批归一化计算结果的张量首地址。
|
||
|
||
支持平台：
|
||
``FT78NE``
|
||
``MT7004``
|
||
|
||
.. note::
|
||
- FT78NE 支持 fp32 数据类型。
|
||
- MT7004 支持 fp16、fp32 数据类型。
|
||
|
||
**共享存储版本:**
|
||
|
||
.. c:function:: void hp_fusedbatchnorm_s(const half *input, const half *scale, const half *offset, const half *mean, const half *variance, float epsilon, int channel, int unit, int core_mask, half *output)
|
||
.. c:function:: void fp_fusedbatchnorm_s(const float *input, const float *scale, const float *offset, const float *mean, const float *variance, float epsilon, int channel, int unit, int core_mask, float *output)
|
||
|
||
**C调用示例：**
|
||
|
||
.. code-block:: c
|
||
:linenos:
|
||
:emphasize-lines: 15
|
||
|
||
// FT78NE 多核示例
|
||
#include <stdio.h>
|
||
|
||
int main(void) {
|
||
const float *input = (const float *)0xA0000000; // DDR 存储
|
||
const float *scale = (const float *)0xB0000000;
|
||
const float *offset = (const float *)0xB0001000;
|
||
const float *mean = (const float *)0xB0002000;
|
||
const float *variance = (const float *)0xB0003000;
|
||
float *output = (float *)0xC0000000;
|
||
int channel = 64;
|
||
int unit = 1024;
|
||
float epsilon = 1e-5f;
|
||
int core_mask = 0xff;
|
||
fp_fusedbatchnorm_s(input, scale, offset, mean, variance,
|
||
epsilon, channel, unit, core_mask,
|
||
output);
|
||
return 0;
|
||
}
|
||
|
||
|
||
**私有存储版本:**
|
||
|
||
.. c:function:: void hp_fusedbatchnorm_p(const half *input, const half *scale, const half *offset, const half *mean, const half *variance, float epsilon, int channel, int unit, half *output)
|
||
.. c:function:: void fp_fusedbatchnorm_p(const float *input, const float *scale, const float *offset, const float *mean, const float *variance, float epsilon, int channel, int unit, float *output)
|
||
|
||
**C调用示例：**
|
||
|
||
.. code-block:: c
|
||
:linenos:
|
||
:emphasize-lines: 14
|
||
|
||
// MT7004 单核示例
|
||
#include <stdio.h>
|
||
|
||
int main(void) {
|
||
const half *input = (const half *)0x10000000; // L2 存储
|
||
const half *scale = (const half *)0x10004000;
|
||
const half *offset = (const half *)0x10008000;
|
||
const half *mean = (const half *)0x1000C000;
|
||
const half *variance = (const half *)0x10010000;
|
||
half *output = (half *)0x10014000;
|
||
int channel = 32;
|
||
int unit = 512;
|
||
float epsilon = 1e-4f;
|
||
hp_fusedbatchnorm_p(input, scale, offset, mean, variance,
|
||
epsilon, channel, unit, output);
|
||
return 0;
|
||
}
|