Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Help compiler exploit hanging-node symmetries by reduced reg loads #13000

Merged
merged 1 commit into from
Nov 28, 2021
Merged

Help compiler exploit hanging-node symmetries by reduced reg loads #13000

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
51 changes: 43 additions & 8 deletions include/deal.II/matrix_free/evaluation_kernels_hanging_nodes.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,18 +93,53 @@ namespace internal
temp[k] = values[my_offset + k * stride];

// perform interpolation point by point and write back
for (unsigned int k = 0; k < points; ++k)
for (unsigned int k = 0; k < points / 2; ++k)
{
Number sum0 = Number(), sum1 = Number();
const unsigned int kmirror = points - 1 - k;
Number sum0 = Number(), sum1 = Number(), sum2 = Number(),
sum3 = Number();
for (unsigned int h = 0; h < points; ++h)
{
// sum0 belongs to the other interpolation matrix, but we
// can use the symmetry here to reduce the number of loads
sum0 += temp[h] *
weights[(transpose ? 1 : points) * (points - 1 - k) +
(transpose ? points : 1) * (points - 1 - h)];
sum1 += temp[h] * weights[(transpose ? 1 : points) * k +
const unsigned int hmirror = points - 1 - h;
// load from both sides of the interpolation matrix to
// reflect symmetry between the two subfaces along that
// direction
const Number w0 = weights[(transpose ? 1 : points) * kmirror +
(transpose ? points : 1) * hmirror];
const Number w1 = weights[(transpose ? 1 : points) * k +
(transpose ? points : 1) * h];
sum0 += temp[h] * w0;
sum1 += temp[h] * w1;
sum2 += temp[hmirror] * w1;
sum3 += temp[hmirror] * w0;
}
values[my_offset + k * stride] =
temp[k] +
mask_write * (sum0 + mask_weight * (sum1 - sum0) - temp[k]);
values[my_offset + kmirror * stride] =
temp[kmirror] +
mask_write *
(sum2 + mask_weight * (sum3 - sum2) - temp[kmirror]);
}

// cleanup case
if (points % 2)
{
const unsigned int k = points / 2;
Number sum0 = temp[k] * weights[(transpose ? 1 : points) * k +
(transpose ? points : 1) * k],
sum1 = sum0;
for (unsigned int h = 0; h < points / 2; ++h)
{
const unsigned int hmirror = points - 1 - h;
const Number w0 = weights[(transpose ? 1 : points) * k +
(transpose ? points : 1) * hmirror];
const Number w1 = weights[(transpose ? 1 : points) * k +
(transpose ? points : 1) * h];
sum0 += temp[h] * w0;
sum0 += temp[hmirror] * w1;
sum1 += temp[h] * w1;
sum1 += temp[hmirror] * w0;
}
values[my_offset + k * stride] =
temp[k] +
Expand Down