Implement 2D and 1D std::array in opencl kernel

Question

I am asked to implement the following part of code into kernel code. Actually, I have tried but not sure about the std::array.

This is the original code for the function.

int KNN_classifier(array<array<int, 20>, 4344>& array_X_train, array<int, 4344>& array_Y_train,
 int k, array<int, 20>& data_point) {
 // Calculate the distance between data_point and all points.
 int index = 0;
 static array<pair<int, float>, 4344> array_dist{};
 pair<int, float> pair_tmp;
 //Apply uroll:__attribute__((opencl_unroll_hint(n)))
 for (size_t i = 0; i < array_X_train.size(); ++i) {
 pair_tmp = make_pair(i, Euclidean_distance(array_X_train[i], data_point));
 // dist_vec.push_back(pair_tmp);
 array_dist[index] = pair_tmp;
 index++; 
 }

This is the kernel code:

 #define N1 20 //num of cols
#define N2 4344 //num of rows
inline float Euclidean_distance(array<int, 20>& array_point_A, array<int, 20>& array_point_B) {
 float sum = 0.0;
 static array<float, 20> w = { 0.0847282, 0.0408621, 0.105036, 0.0619821, 0.0595455, 0.0416739, 0.0181147, 0.00592921,
 0.040049, 0.0766054, 0.0441091, 0.0376111, 0.0124285, 0.0733558, 0.0587338, 0.0303001, 0.0579207, 0.0449221,
 0.0530462, 0.0530462 };
 for (size_t i = 0; i < array_point_A.size(); ++i) {
 int a = array_point_A[i] - array_point_B[i];
 float wieghted_distance = w[i] * (static_cast<float> (a) * static_cast<float> (a)); 
 sum += wieghted_distance;
 }
 return sqrt(sum);
}
__kernel int KNN_classifier(__global const int* restrict array_X_train, __global const int* restrict array_Y_train,
 __global const int* restrict data_point, int k)
{
 int index = 0;
 static array<pair<int, float>, 4344> array_dist{};
 pair<int, float> pair_tmp;
 for (size_t i = 0; i < array_X_train.size(); ++i) {
 pair_tmp = make_pair(i, Euclidean_distance(array_X_train[i], data_point));
 array_dist[index] = pair_tmp;
 index++;
 }

Answer 1

Review

Welcome to Code Review. There are some suggestions as follows.

Const Definition

You already define the value N1 and N2, why not use that in the array template parameter directly (e.g. array<float, N1>)? Actually, not sure what kind of problem you want to solve with KNN algorithm, you specified the dimension of KNN algorithm in 20 here. How about the case of the dimension which is different from 20?

The weight w in Euclidean_distance function

Do you have any reason to do this? In general, this w array seems to be a kind of data, which could be passed as a parameter like Euclidean_distance(array<int, 20>& array_point_A, array<int, 20>& array_point_B, array<int, 20>& w) instead and you provide the value of w parameter when you call it. It is not necessary to define the const data in a function.