| commit | d1a07e21a9e92e374ebd40da2c39ff223a9138e1 | |
| author | Berk Hess <hess@kth.se> | |
| Thu, 25 Feb 2016 19:27:36 +0000 (25 20:27 +0100) | ||
| committer | Mark Abraham <mark.j.abraham@gmail.com> | |
| Sat, 27 Feb 2016 17:49:14 +0000 (27 18:49 +0100) | ||
| tree | 03569069a701c463a63a4b9da5056d2af9df9c0b | treesnapshot (tar.gz zip) |
| parent | cb313c7e66b36b25e4c54db9934cad613418ecd1 | commitdiff |
Remove epsfac from GPU kernel inner-loops
Multiplying the i-charge instead of the j-charge with epsfac in
the GPU kernels removes a flop form the inner-loop. On Maxwell with
CUDA the gain is 2-3% in the force kernels and 3-4% in the energy
kernels (which is more than a flop, probably a register less is used).
There is a single additional division in the energy kernel, but the
gain more than compensates for this.
Change-Id: I6924b32f3f61d3b7bbe532f95361b7fefb577609
Multiplying the i-charge instead of the j-charge with epsfac in
the GPU kernels removes a flop form the inner-loop. On Maxwell with
CUDA the gain is 2-3% in the force kernels and 3-4% in the energy
kernels (which is more than a flop, probably a register less is used).
There is a single additional division in the energy kernel, but the
gain more than compensates for this.
Change-Id: I6924b32f3f61d3b7bbe532f95361b7fefb577609