coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2026-10-28 18:07:12 +08:00
Code Issues Packages Projects Releases Wiki Activity

Update patch set 7

Browse Source 
Patch Set 7:

(5 comments)

Addressed all review comments in the latest patchset.

Patch-set: 7
This commit is contained in:
Gerrit User 1115898 2026-10-27 15:52:53 -07:00 committed by Gerrit Code Review
parent 848eb00428
commit 905ff3ad4f
1 changed files with 90 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

90
1b1216abb649754576866b54f03a5229a02d9e52
View File

@ -34,6 +34,24 @@
"revId": "1b1216abb649754576866b54f03a5229a02d9e52",
"serverId": "3ce6091f-6c88-37e8-8c75-72f92ae8dfba"
},
{
"unresolved": false,
"key": {
"uuid": "837a75a9_ebeae93e",
"filename": "source/planar_functions.cc",
"patchSetId": 7
},
"lineNbr": 2936,
"author": {
"id": 1115898
},
"writtenOn": "2026-10-27T22:52:53Z",
"side": 1,
"message": "Done. Removed the unreachable #elif defined(HAS_MIRRORUVROW_NEON) branch. Keeping dedicated 16-bit element row functions (MirrorRow_16_* and CopyRow_16_*) allows tail masking to be done with 16-bit predicates or rep movsw/stosw.",
"parentUuid": "ef108627_f67f97d7",
"revId": "1b1216abb649754576866b54f03a5229a02d9e52",
"serverId": "3ce6091f-6c88-37e8-8c75-72f92ae8dfba"
},
{
"unresolved": true,
"key": {
@ -51,6 +69,24 @@
"revId": "1b1216abb649754576866b54f03a5229a02d9e52",
"serverId": "3ce6091f-6c88-37e8-8c75-72f92ae8dfba"
},
{
"unresolved": false,
"key": {
"uuid": "ff8f5d5e_04637c87",
"filename": "source/row_neon.cc",
"patchSetId": 7
},
"lineNbr": 3995,
"author": {
"id": 1115898
},
"writtenOn": "2026-10-27T22:52:53Z",
"side": 1,
"message": "Done. Updated CopyPlane_16 in source/planar_functions.cc to dispatch CopyRow_16 (including CopyRow_16_NEON / CopyRow_16_Any_NEON and CopyRow_16_HVX) and added CopyPlane_16 unit tests in unit_test/planar_test.cc. Keeping dedicated 16-bit element functions (CopyRow_16_* and MirrorRow_16_*) allows tail masking to be done with 16-bit predicates or rep movsw/stosw.",
"parentUuid": "393b10e1_4fa315af",
"revId": "1b1216abb649754576866b54f03a5229a02d9e52",
"serverId": "3ce6091f-6c88-37e8-8c75-72f92ae8dfba"
},
{
"unresolved": true,
"key": {
@ -68,6 +104,24 @@
"revId": "1b1216abb649754576866b54f03a5229a02d9e52",
"serverId": "3ce6091f-6c88-37e8-8c75-72f92ae8dfba"
},
{
"unresolved": false,
"key": {
"uuid": "9d802295_65c69e9a",
"filename": "source/row_neon.cc",
"patchSetId": 7
},
"lineNbr": 4188,
"author": {
"id": 1115898
},
"writtenOn": "2026-10-27T22:52:53Z",
"side": 1,
"message": "Done. Replaced vmov + vmla.f32 with vmul.f32 + vadd.f32 to avoid the two vmov instructions per iteration.",
"parentUuid": "cb6aa027_8a0f308c",
"revId": "1b1216abb649754576866b54f03a5229a02d9e52",
"serverId": "3ce6091f-6c88-37e8-8c75-72f92ae8dfba"
},
{
"unresolved": true,
"key": {
@ -85,6 +139,24 @@
"revId": "1b1216abb649754576866b54f03a5229a02d9e52",
"serverId": "3ce6091f-6c88-37e8-8c75-72f92ae8dfba"
},
{
"unresolved": false,
"key": {
"uuid": "337fea1f_c5575252",
"filename": "source/row_neon64.cc",
"patchSetId": 7
},
"lineNbr": 5927,
"author": {
"id": 1115898
},
"writtenOn": "2026-10-27T22:52:53Z",
"side": 1,
"message": "We benchmarked both on Apple M6: packing all 4 bytes with 3 bfi instructions and a single post-indexed str %w5, [%0], #4 took 8.24 ms (vs 8.17 ms for C) due to the serial bfi dependency chain waiting on all 4 table loads, whereas strh + strb + strb + add took 7.43 ms (1.10x faster than C) since each store depends on at most one ALU op after its load.",
"parentUuid": "f2aed825_a3572d3a",
"revId": "1b1216abb649754576866b54f03a5229a02d9e52",
"serverId": "3ce6091f-6c88-37e8-8c75-72f92ae8dfba"
},
{
"unresolved": true,
"key": {
@ -101,6 +173,24 @@
"message": "Two small notes here:\n1. Like the SSE2 block above, this calls `ARGBAffineRow_NEON` unconditionally before checking `TestCpuFlag(kCpuHasNEON)` on line 1583.\n2. This only tests `width \u003d 1280` (a multiple of 4) and `src_argb_stride \u003d 0`, so the 1-3 pixel tail loop and non-zero `src_argb_stride` path in `ARGBAffineRow_NEON` aren\u0027t exercised.",
"revId": "1b1216abb649754576866b54f03a5229a02d9e52",
"serverId": "3ce6091f-6c88-37e8-8c75-72f92ae8dfba"
},
{
"unresolved": false,
"key": {
"uuid": "46373fbe_f81b14e2",
"filename": "unit_test/planar_test.cc",
"patchSetId": 7
},
"lineNbr": 1579,
"author": {
"id": 1115898
},
"writtenOn": "2026-10-27T22:52:53Z",
"side": 1,
"message": "Done. Guarded both SSE2 and NEON calls with TestCpuFlag checks and added a width \u003d 1279 test case with non-zero src_argb_stride to exercise both the 1-3 pixel tail loop and stride arithmetic.",
"parentUuid": "32ca45d5_15db4360",
"revId": "1b1216abb649754576866b54f03a5229a02d9e52",
"serverId": "3ce6091f-6c88-37e8-8c75-72f92ae8dfba"
}
]
}