public inbox for gcc-cvs@sourceware.org
help / color / mirror / Atom feed
* [gcc(refs/vendors/vrull/heads/slp-improvements)] Fix (perf regression): revert to previous ordering
@ 2024-01-23 20:59 Philipp Tomsich
0 siblings, 0 replies; only message in thread
From: Philipp Tomsich @ 2024-01-23 20:59 UTC (permalink / raw)
To: gcc-cvs
https://gcc.gnu.org/g:d8a63369bf3cb2a8f6450162683b51a388021312
commit d8a63369bf3cb2a8f6450162683b51a388021312
Author: Manolis Tsamis <manolis.tsamis@vrull.eu>
Date: Tue Jan 23 17:58:05 2024 +0100
Fix (perf regression): revert to previous ordering
The 'more natural' ABCD ordering (as opposed to the ACBD order) of all
nodes from the refactored implementation resulted in a x264
regression, where the number of tbl instructions has increased from 26
to 286 in total.
This changes the order back to ACBD, reducing the number of
tbl-instructions back to 26 and resolves the observed performance
regression from the fix for #356.
Ref #356
Signed-off-by: Philipp Tomsich <philipp.tomsich@vrull.eu>
Diff:
---
gcc/tree-vect-slp.cc | 20 ++++++++++----------
1 file changed, 10 insertions(+), 10 deletions(-)
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index f7970bafd45..6b5edf1cebc 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -1920,14 +1920,14 @@ try_rearrange_oprnd_info (vec<slp_oprnd_info> &oprnds_info, unsigned group_size)
{
/* Given oprnd[0] -> A1, A1, B1, B1, A2, A2, B2, B2, ...
Given oprnd[1] -> C1, C1, D1, D1, C2, C2, D2, D2, ...
- Create a single node -> A1, B1, C1, D1, A2, B2, C2, D2, ... */
+ Create a single node -> A1, C1, B1, D1, A2, C2, B2, D2, ... */
/* The ordering here is at least to some extent arbitrary.
A generilized version needs to use some explicit ordering. */
- oprnds_info[0]->def_stmts[j+1] = oprnds_info[0]->def_stmts[j+2];
- oprnds_info[0]->ops[j+1] = oprnds_info[0]->ops[j+2];
- oprnds_info[0]->def_stmts[j+2] = oprnds_info[1]->def_stmts[j];
- oprnds_info[0]->ops[j+2] = oprnds_info[1]->ops[j];
+ oprnds_info[0]->def_stmts[j+1] = oprnds_info[1]->def_stmts[j];
+ oprnds_info[0]->ops[j+1] = oprnds_info[1]->ops[j];
+ oprnds_info[0]->def_stmts[j+2] = oprnds_info[0]->def_stmts[j+2];
+ oprnds_info[0]->ops[j+2] = oprnds_info[0]->ops[j+2];
oprnds_info[0]->def_stmts[j+3] = oprnds_info[1]->def_stmts[j+2];
oprnds_info[0]->ops[j+3] = oprnds_info[1]->ops[j+2];
}
@@ -2817,7 +2817,7 @@ fail:
}
else if (rearrange_pattern == SLP_OPRND_PATTERN_AABB)
{
- /* Given a single node -> A1, B1, C1, D1, A2, B2, C2, D2, ...
+ /* Given a single node -> A1, C1, B1, D1, A2, C2, B2, D2, ...
Create node "one" -> A1, A1, B1, B1, A2, A2, B2, B2, ...
Create node "two" -> C1, C1, D1, D1, C2, C2, D2, D2, ... */
@@ -2825,11 +2825,11 @@ fail:
{
SLP_TREE_LANE_PERMUTATION(one).safe_push (std::make_pair (0, j + 0));
SLP_TREE_LANE_PERMUTATION(one).safe_push (std::make_pair (0, j + 0));
- SLP_TREE_LANE_PERMUTATION(one).safe_push (std::make_pair (0, j + 1));
- SLP_TREE_LANE_PERMUTATION(one).safe_push (std::make_pair (0, j + 1));
+ SLP_TREE_LANE_PERMUTATION(one).safe_push (std::make_pair (0, j + 2));
+ SLP_TREE_LANE_PERMUTATION(one).safe_push (std::make_pair (0, j + 2));
- SLP_TREE_LANE_PERMUTATION(two).safe_push (std::make_pair (0, j + 2));
- SLP_TREE_LANE_PERMUTATION(two).safe_push (std::make_pair (0, j + 2));
+ SLP_TREE_LANE_PERMUTATION(two).safe_push (std::make_pair (0, j + 1));
+ SLP_TREE_LANE_PERMUTATION(two).safe_push (std::make_pair (0, j + 1));
SLP_TREE_LANE_PERMUTATION(two).safe_push (std::make_pair (0, j + 3));
SLP_TREE_LANE_PERMUTATION(two).safe_push (std::make_pair (0, j + 3));
}
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2024-01-23 20:59 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2024-01-23 20:59 [gcc(refs/vendors/vrull/heads/slp-improvements)] Fix (perf regression): revert to previous ordering Philipp Tomsich
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).