Question about the parameters in the kernel about gpa-minimod-artifacts HOT 3 CLOSED

N_RADIUS does not depend on N_THREADS_PER_BLOCK_DIM. N_RADIUS = 4 refers to the case which the order of a stencil is high. Figure 1 in this paper (https://arxiv.org/ftp/arxiv/papers/2002/2002.05983.pdf) shows the comparison between a low order stencil and a high order stencil. Usually we use the same number for all three dimensions. @rsrice can provide more information.

from gpa-minimod-artifacts.

When N_THREADS_PER_BLOCK_DIM is equal to 4, the min and max values are 0. If you could reproduce that, please suggest the code changes. I assume min and max values are not supposed to be 0.

./main --grid 100
ndamp = 27 27 27
grid = 100 100 100
time step 100 / 1000
time step 200 / 1000
time step 300 / 1000
time step 400 / 1000
time step 500 / 1000
time step 600 / 1000
time step 700 / 1000
time step 800 / 1000
time step 900 / 1000
time step 1000 / 1000
FINAL min_u,  max_u = 0.000000, 0.000000

from gpa-minimod-artifacts.

N_THREADS_PER_BLOCK_DIM and N_RADIUS in this implementation are not parameters. The optimizations of shared memory fetching, such as L40-45, assume the grid size of 8x8x8 as we have discussed in this paper (https://arxiv.org/abs/2009.04619). They are not parameters but are extracted out as macros for better readability instead of magic numbers being all over the place.

With that said, we are working towards a semi-automated code generation for minimod that allows the arbitrary grid sizes. It also implies one can specify different sizes across three dimensions. If you are interested in minimod and plan to experimenting different grid sizes, such as 8x8x4 or 7x5x6, please stay tuned.

from gpa-minimod-artifacts.