Fused coefficient in symmetric-isotropic case

[christiealappatt]

Right now in the case of symmetric, isotropic stencils the coefficients are not fused i.e. for example python stempel.py -D 2 -r 1 -s -k star -C variable

output:

for(int j=2; j < M-2; j++){
for(int i=2; i < N-2; i++){
b[j][i] = W[j][i][0] * a[j][i]
+ W[j][i][1] * (a[j][i-1] + a[j][i+1])
+ W[j][i][1] * (a[j-1][i] + a[j+1][i])
+ W[j][i][2] * (a[j][i-2] + a[j][i+2])
+ W[j][i][2] * (a[j-2][i] + a[j+2][i])
;
}
}

I think it would be better if the output is:

for(int j=2; j < M-2; j++){
for(int i=2; i < N-2; i++){
b[j][i] = W[j][i][0] * a[j][i]
+ W[j][i][1] * (a[j][i-1] + a[j][i+1])
                    + a[j-1][i] + a[j+1][i])
+ W[j][i][2] * (a[j][i-2] + a[j][i+2])
                   + a[j-2][i] + a[j+2][i])
;
}
}

Since sometimes compiler might generate additional multiplies for the current case,

[sguera]

I see what you mean. Maybe @cod3monk can tell us how would kerncraft deal with it.

[cod3monk]

Kerncraft does not deal with it. It does not do any code transformations and leaves that part to the compiler. I have seen issues with this particular problem in a 3D boxed stencil. For kerncraft the solution is to fuse the coefficients manually, but since you are generating the code for kerncraft it might be useful to do that automatically.

[sguera]

Ok thanks for clarifying. Then I will fuse them while generating the code.

[sguera]

fixed in 3e57c7dcdae6422891c4054eb8ef6560155c4343 and 856049e55ca8e2e4a7857aee27a74215bd8879f0