mulimoen
/
SummationByParts
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

simdify blocks in diff_op_col

This commit is contained in:
Magnus Ulimoen 2020-05-01 19:08:55 +02:00
parent cbf6042055
commit 52bd4f3f8f
1 changed files with 92 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

103
sbp/src/operators.rs
View File

@ -293,20 +293,6 @@ fn diff_op_col_simd(
// How many elements that can be simdified
let simdified = SimdT::lanes() * (ny / SimdT::lanes());
// First block
{
for (bl, mut fut) in block.iter().zip(fut.axis_iter_mut(ndarray::Axis(1))) {
fut.fill(0.0);
debug_assert_eq!(fut.len(), prev.shape()[0]);
for (&bl, prev) in bl.iter().zip(prev.axis_iter(ndarray::Axis(1))) {
debug_assert_eq!(prev.len(), fut.len());
fut.scaled_add(idx * bl, &prev);
}
}
}
// Diagonal elements
{
let half_diag_width = (diag.len() - 1) / 2;
assert!(half_diag_width <= block.len());
@ -324,11 +310,51 @@ fn diff_op_col_simd(
unsafe { prev_base_ptr.offset(prev_stride * i as isize + j as isize) }
};
// Not algo necessary, but gives performance increase
assert_eq!(fut_stride, prev_stride);
// First block
{
for (ifut, &bl) in block.iter().enumerate() {
for j in (0..simdified).step_by(SimdT::lanes()) {
let index_to_simd = |i| unsafe {
// j never moves past end of slice due to step_by and
// rounding down
SimdT::from_slice_unaligned(std::slice::from_raw_parts(
prev_ptr(j, i),
SimdT::lanes(),
))
};
let mut f = SimdT::splat(0.0);
for (iprev, &bl) in bl.iter().enumerate() {
f = index_to_simd(iprev).mul_adde(SimdT::splat(bl), f);
}
f = f * idx;
unsafe {
f.write_to_slice_unaligned(std::slice::from_raw_parts_mut(
fut_ptr(j, ifut),
SimdT::lanes(),
));
}
}
for j in simdified..ny {
unsafe {
let mut f = 0.0;
for (iprev, bl) in bl.iter().enumerate() {
f += bl * *prev_ptr(j, iprev);
}
*fut_ptr(j, ifut) = f * idx;
}
}
}
}
// Diagonal elements
{
for ifut in block.len()..nx - block.len() {
for j in (0..simdified).step_by(SimdT::lanes()) {
let index_to_simd = |(j, i)| unsafe {
let index_to_simd = |i| unsafe {
// j never moves past end of slice due to step_by and
// rounding down
SimdT::from_slice_unaligned(std::slice::from_raw_parts(
@ -339,7 +365,7 @@ fn diff_op_col_simd(
let mut f = SimdT::splat(0.0);
for (id, &d) in diag.iter().enumerate() {
let offset = ifut - half_diag_width + id;
f = index_to_simd((j, offset)).mul_adde(SimdT::splat(d), f);
f = index_to_simd(offset).mul_adde(SimdT::splat(d), f);
}
f = f * idx;
unsafe {
@ -367,13 +393,46 @@ fn diff_op_col_simd(
// End block
{
for (bl, mut fut) in block.iter().zip(fut.axis_iter_mut(ndarray::Axis(1)).rev()) {
fut.fill(0.0);
for (&bl, prev) in bl.iter().zip(prev.axis_iter(ndarray::Axis(1)).rev()) {
if symmetry == Symmetry::Symmetric {
fut.scaled_add(idx * bl, &prev);
// Get blocks and corresponding offsets
// (rev to iterate in ifut increasing order)
for (bl, ifut) in block.iter().zip((0..nx).rev()) {
for j in (0..simdified).step_by(SimdT::lanes()) {
let index_to_simd = |i| unsafe {
// j never moves past end of slice due to step_by and
// rounding down
SimdT::from_slice_unaligned(std::slice::from_raw_parts(
prev_ptr(j, i),
SimdT::lanes(),
))
};
let mut f = SimdT::splat(0.0);
for (&bl, iprev) in bl.iter().zip((0..nx).rev()) {
f = index_to_simd(iprev).mul_adde(SimdT::splat(bl), f);
}
f = if symmetry == Symmetry::Symmetric {
f * idx
} else {
fut.scaled_add(-idx * bl, &prev);
-f * idx
};
unsafe {
f.write_to_slice_unaligned(std::slice::from_raw_parts_mut(
fut_ptr(j, ifut),
SimdT::lanes(),
));
}
}
for j in simdified..ny {
unsafe {
let mut f = 0.0;
for (&bl, iprev) in bl.iter().zip((0..nx).rev()).rev() {
f += bl * *prev_ptr(j, iprev);
}
*fut_ptr(j, ifut) = if symmetry == Symmetry::Symmetric {
f * idx
} else {
-f * idx
};
}
}
}