Greetings!
Is it possible to somehow vary the tile size of the tile clause during runtime?
E.g.:
for(time = 0; time < timeMax; time++){ #pragma acc parallel loop tile(tl1, tl2) present(arrA, arrB) for(int x = 0; x < SIZE; x++){ for(int y = 0; y < SIZE; y++){ //Some work } } //Compute new tl1 and tl2 values for the next timestep tl1 = ... tl2 = ...
}
I initially thought this would throw an error since tl1 and tl2 are not constants, but instead the code compiled and ran smoothly. OpenACC seems to have ignored the tile clause and executed as if the tile was not being used (using Nvidia HPC 21.5).