Qiea/c-cnn
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

优化convolution函数

Browse Source 
降低malloc使用的内存
This commit is contained in:
Qiea
2024-11-11 19:44:37 +08:00
parent a6adf831b3
commit dfc83243df
1 changed files with 20 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
cnn.c
View File

@@ -43,8 +43,8 @@ float* convolution(Model model_w, Model model_b, const float* input_matrix, int
float conv_temp; // 临时变量，用于存储卷积计算的中间结果 float conv_temp; // 临时变量，用于存储卷积计算的中间结果
//用于合并前的数组，具有32*64*50*50(第二层)的大小 //用于合并前的数组，具有32*64*50*50(第二层)的大小
float* _conv_rlst = (float *) malloc(sizeof (float) * model_w.channel * model_w.num_kernels * (cr_l * cr_l)); float* _conv_rlst = (float *) malloc(sizeof (float) * model_w.num_kernels * (cr_l * cr_l));
memset(_conv_rlst, 0, sizeof (float) * model_w.channel * model_w.num_kernels * (cr_l * cr_l)); memset(_conv_rlst, 0, sizeof (float) * model_w.num_kernels * (cr_l * cr_l));
//子图合并后的数组 //子图合并后的数组
float* conv_rlst = (float *) malloc(sizeof (float) * model_w.num_kernels * (cr_l * cr_l)); float* conv_rlst = (float *) malloc(sizeof (float) * model_w.num_kernels * (cr_l * cr_l));
memset(conv_rlst, 0, sizeof (float) * model_w.num_kernels * (cr_l * cr_l)); memset(conv_rlst, 0, sizeof (float) * model_w.num_kernels * (cr_l * cr_l));
@@ -63,32 +63,35 @@ float* convolution(Model model_w, Model model_b, const float* input_matrix, int
* model_w.array[((c+k*model_w.channel)*3*3) + (x*3+y)]; * model_w.array[((c+k*model_w.channel)*3*3) + (x*3+y)];
} }
} }
_conv_rlst[(c*model_w.num_kernels*cr_l*cr_l) + (k*cr_l*cr_l) + (row*cr_l+col)] = conv_temp; _conv_rlst[(k*cr_l*cr_l) + (row*cr_l+col)] = conv_temp;
} }
} }
} }
} //合并子图
//合并子图
{
for(int k=0; k < model_w.num_kernels; k++) { for(int k=0; k < model_w.num_kernels; k++) {
for (int row = 0; row < cr_l; row++) { for (int row = 0; row < cr_l; row++) {
for (int col = 0; col < cr_l; col++) { for (int col = 0; col < cr_l; col++) {
conv_temp = 0; // 每个输出像素初始化为0 conv_rlst[(k*cr_l*cr_l) + (row*cr_l+col)] += _conv_rlst[(k*cr_l*cr_l) + (row*cr_l+col)];
for (int c = 0; c < model_w.channel; c++) {
conv_temp += _conv_rlst[(c * model_w.num_kernels * cr_l * cr_l) + (k * cr_l * cr_l) + (row * cr_l + col)];
}
// 加上对应卷积核的偏置
conv_temp += model_b.array[k];
// 激活函数：ReLU（将小于0的值设为0）
if (conv_temp > 0)
conv_rlst[(k * (cr_l * cr_l)) + (row * cr_l) + (col)] = conv_temp; // 如果卷积结果大于0，存入结果数组
else
conv_rlst[(k * (cr_l * cr_l)) + (row * cr_l) + (col)] = 0; // 否则存入0
} }
} }
} }
} }
for(int k=0; k < model_w.num_kernels; k++) {
for (int row = 0; row < cr_l; row++) {
for (int col = 0; col < cr_l; col++) {
conv_temp = conv_rlst[(k * cr_l * cr_l) + (row * cr_l + col)];
// 加上对应卷积核的偏置
conv_temp += model_b.array[k];
// 激活函数：ReLU（将小于0的值设为0）
if (conv_temp > 0)
conv_rlst[(k * (cr_l * cr_l)) + (row * cr_l) + (col)] = conv_temp; // 如果卷积结果大于0，存入结果数组
else
conv_rlst[(k * (cr_l * cr_l)) + (row * cr_l) + (col)] = 0; // 否则存入0
}
}
}
free(_conv_rlst); free(_conv_rlst);
_conv_rlst = NULL;
return conv_rlst; return conv_rlst;
} }