Laplace potential kernel - rinv evaluation

[dmalhotra]

Laplace potential kernel - rinv evaluation

[dmalhotra]

1. rsqrt_approx_intrin is less accurate (_mm_rsqrt_ps).
2. You need to manually delete the precomputed files after making changes to the kernel function.

How these relate to your observations:

• I believe in the first case, the errors are so large because you did not delete the precomputed files.
• You get convergence warnings (not errors) during precomputation because the kernel function is less accurate. The warnings here are irrelevant since that part of the code is only used for periodic boundary conditions. The warning goes away if the volume-potential term in kernel construction is replace by NULL. The double-layer kernel and scale-invariance condition don't make any difference to the warnings.
• In the second case you get much more reasonable relative errors (0.05 for a kernel function only accurate to 4-digits) because changing the scale-invariance condition required that the precomputed data had to be computed again for tree levels > 0. You could have achieved the same effect by deleting the precomputed files.

[thomasgillis]

Thank you very much for the detailed information. Helped a lot!
Is it possible to disable the Precomp_XX writing using some flags?

[dmalhotra]

There is no option to disable writing the precomputed file, but you can hack it. You can use the configure option --with-precomp-dir="some-dir-path" to specify the directory where the file is stored. If you give a path which doesn't exist, then the file can not be written.