A Strategy for Automatic Performance Tuning of Stencil Computations on GPUs

Algorithm 1

Generic 3D stencil computation.

void stencil_computation (

float (array1) [Y_SIZE] [X_SIZE],

float (array2) [Y_SIZE] [X_SIZE])

float (in) [Y_SIZE] [X_SIZE] = array1;

float (out) [Y_SIZE] [X_SIZE] = array2;

for (int t = 0; t < T_MAX; ++t)

for (int z = 0; z < Z_SIZE; ++z)

for (int y = 0; y < Y_SIZE; ++y)

for (int x = 0; x < X_SIZE; ++x)

float temp0 = in [z+] [y+] [x+];

float temp1 = in [z+] [y+] [x+];

⋮

float tempN = in [z+] [y+] [x+];

out [z] [y] [x]= f();

// Swap in and out pointers