Geting wrong results in matrix multiplication

OpenCL
1

Hi, i wrote this kernel to multiply matrices (without caring about the types), but i’m getting wrong results. It seems like there’s a problem in the way i read the values from memory, because just some values of the result matrix are wrong.

Here’s the code:

#include "name_func.h"
#define FUNCNAME NAME(dot_matrix, Type1, Type2) 


__kernel void FUNCNAME(const int M, const int N, 
   const int K, const int Srows, const int Scols, 
		 const __global Type1* A,
		 const __global Type2* B,
		 __global Out_Type* C,
		 __local Type1* Asub,
		 __local Type2* Bsub) {

    const int row = get_local_id(0);
    const int col = get_local_id(1);
    const int globalRow = Srows*get_group_id(0) + row;
    const int globalCol = Scols*get_group_id(1) + col;
    Out_Type acc = 0;

    const int num_tiles = ceil((float) N / Scols);
    for(int t = 0; t < num_tiles; t++) {
        const int tiledCol = Scols*t + col;
        const int tiledRow = Srows*t + row;
        Asub[row*Scols + col] = A[(globalRow*N + tiledCol)]; 
        Bsub[row*Scols + col] = B[(tiledRow*K + globalCol)];
        barrier(CLK_LOCAL_MEM_FENCE);

        const bool change_size = (t+1) == num_tiles;
        int new_size = N - (num_tiles - 1)*Scols;

        //local size = new_size if change_size == True, local_size otherwise
        new_size = change_size * new_size + (1 - change_size) * Scols;

        for(int i = 0; i < new_size; i++) {
            acc += Asub[row*Scols + i] * Bsub[i*Scols + col];
        }
    }
    C[globalRow*K + globalCol] = acc;
}

#undef FUNCNAME

I have an AMD APU A10-8700p with GPU AMD Radeon R6