OpenCL course homeworks
Directory: matrix_mul_ocl Device code: mat_mp.cl Host code: main.cpp
Input for device code is 2 randomly generated matrices, result buffer and all sizes (N x M x K).
Device code uses local group memory, splites matrix on tiles, transposes input matrix.
Tiles have fixed size LOC_WORK^2, all of N, M, K must devide on LOC_WORK (LOC_WORK is compile defined in common/utils.h).
Devision condition is automatically fixed in code.
GFLOPS is ~10 times lower than Intel UHD 620 specification.
For OpenMP version look at: matrix_mul_omp
Directory: prefsum_ocl Device code: prefsum.cl Host code: main.cpp
Counts prefix sum of float array. Uses tree-building algorithm. (splits array on power of 2 chunks and aggregate them)
Whole array is splitted on local groups of fixed size.
Actually is straightforward, stupid implementation.