Ncnn OP forward code learning

OpenMP Supported programming languages :C、C++ and Fortran; Support OpenMp The compiler includes Visual studio,Sun Compiler,GNU Compiler and Intel Compiler, Clang. Specific view :https://www.openmp.org/resources/openmp-compilers-tools/

OpenMP One of the most powerful functions ： Based on the source code of serial program , Just make a few changes , You can parallelize many serial for loop , Achieve the effect of significantly improving performance .

1. AbsVal: There are two interfaces ,forward and forward_inplace, The difference is inplace Replace in place , No additional memory space will be requested , In the specific operation , stay channel Dimension do pragma omp parallel for Parallel acceleration , if bottom value Being positive , be top value = bottom value, Otherwise, reverse .

2. ArgMax: There are two output modes ,1).  Output only max index  2). Output max index At the same time , It will also output the index Corresponding val. Which output method is used , adopt load_param(FILE*  paramfp) Interface out_max_val Parameters to determine , if out_max_val It's true , Then for 2) Mode output , Before output topk The value of is max index, after topk The values are max val; Otherwise, it is the way 1 Output .

int nscan = fscanf(paramfp, "%d %d", &out_max_val, &topk);

It is worth mentioning that internal targeting topk Sort ,ncnn Interface std::partial_sort Interface .

std::partial_sort(vec.begin(), vec.begin() + topk, vec.end(), std::greater< std::pair<float, int> >());

3. BatchNorm: BatchNorm There are also two interfaces ,forward and forward_inplace, Don't over elaborate . The core optimization is to calculate all the preset calculations during initialization , In practice forward Reduce the amount of calculation .

#pragma omp parallel for

for(int q = 0; q < channels; q++){

const float* ptr = bottom_blob.channel(q);

float* outptr = top_blob.channel(q);

float a = a_data_ptr[q];

float b = b_data_ptr[q];

for(int i=0; i<size; i++){

outptr[i] = b*ptr[i] + a;
}