Differential D44659

[libcxx] Optimize <experimental/simd> -O0 performance for operators
Needs ReviewPublic
Actions

Authored by timshen on Mar 19 2018, 5:44 PM.

Tags

None

Subscribers

jlebar, bixia, miyuki and 4 others

Details

Reviewers

mclow.lists
EricWF

Summary

When vector extension (attribute((vector_size(...)))) is available
use its operators, instead of generating loops of scalar operations.

Also use memcpy() when possible.

Diff Detail

Build Status

Buildable 16783
Build 16783: arc lint + arc unit

Event Timeline

timshen created this revision.Mar 19 2018, 5:44 PM

Herald added a reviewer: EricWF. · View Herald TranscriptMar 19 2018, 5:44 PM

Herald added subscribers: christof, sanjoy. · View Herald Transcript

Harbormaster completed remote builds in B16241: Diff 139051.Mar 19 2018, 5:46 PM

timshen added a parent revision: D44658: [libcxx] Implement <experimental/simd> aligned load and store when compiled with Clang..Mar 19 2018, 5:49 PM

timshen added a child revision: D44660: [libcxx] unroll the loops in <experimental/simd> for Clang, until LLVM bugs are fixed.

Also optimize -O0 load and stores by using memcpy.

timshen edited the summary of this revision. (Show Details)Apr 5 2018, 11:42 AM

miyuki added a subscriber: miyuki.Jul 4 2018, 7:34 AM

Herald added subscribers: bixia, jlebar. · View Herald TranscriptJul 4 2018, 7:34 AM

Revision Contents
Changeset List

			Path	Packages
M			libcxx/include/experimental/simd (434 lines)
M			libcxx/test/std/experimental/simd/simd.elementwise/operators.pass.cpp (190 lines)

Diff 141189

libcxx/include/experimental/simd

Loading...

libcxx/test/std/experimental/simd/simd.elementwise/operators.pass.cpp

Loading...