[x265] [PATCH] Avoid pointer copies in ReferencePlanes::{lowresMC, lowresQPelCost}

[George Steed]

These copies are reasonably hot in perf traces and there is no need to
copy the pointers here. Just set a pointer to the array to read from and
use that directly.

Benchmarks show this gives a nice improvement in FPS reported for an
encoding using --preset=medium:

Cortex-A715: +0.2%
  Cortex-X3: +0.7%
Neoverse V2: +0.9%

Change-Id: I33ba6854e7fbe6ecb1778b8201030a717f60f34e
---
 source/common/lowres.h | 12 ++----------
 1 file changed, 2 insertions(+), 10 deletions(-)

diff --git a/source/common/lowres.h b/source/common/lowres.h
index c9c2e3712..7e6baa844 100644
--- a/source/common/lowres.h
+++ b/source/common/lowres.h
@@ -71,11 +71,7 @@ struct ReferencePlanes
     inline pixel *lowresMC(intptr_t blockOffset, const MV& qmv, pixel *buf, 
intptr_t& outstride, bool hme)
     {
         intptr_t YStride = hme ? lumaStride / 2 : lumaStride;
-        pixel *plane[4];
-        for (int i = 0; i < 4; i++)
-        {
-            plane[i] = hme ? lowerResPlane[i] : lowresPlane[i];
-        }
+        pixel **plane = hme ? lowerResPlane : lowresPlane;
         if ((qmv.x | qmv.y) & 1)
         {
             int hpelA = (qmv.y & 2) | ((qmv.x & 2) >> 1);
@@ -98,11 +94,7 @@ struct ReferencePlanes
     inline int lowresQPelCost(pixel *fenc, intptr_t blockOffset, const MV& 
qmv, pixelcmp_t comp, bool hme)
     {
         intptr_t YStride = hme ? lumaStride / 2 : lumaStride;
-        pixel *plane[4];
-        for (int i = 0; i < 4; i++)
-        {
-            plane[i] = hme ? lowerResPlane[i] : lowresPlane[i];
-        }
+        pixel **plane = hme ? lowerResPlane : lowresPlane;
         if ((qmv.x | qmv.y) & 1)
         {
             ALIGN_VAR_16(pixel, subpelbuf[8 * 8]);
-- 
2.34.1


>From 1e6f0252b7d47d153d2fe2f13977f0e70b9bf4a2 Mon Sep 17 00:00:00 2001
From: George Steed <george.st...@arm.com>
Date: Thu, 10 Oct 2024 14:26:12 +0100
Subject: [PATCH] Avoid pointer copies in
 ReferencePlanes::{lowresMC,lowresQPelCost}

These copies are reasonably hot in perf traces and there is no need to
copy the pointers here. Just set a pointer to the array to read from and
use that directly.

Benchmarks show this gives a nice improvement in FPS reported for an
encoding using --preset=medium:

Cortex-A715: +0.2%
  Cortex-X3: +0.7%
Neoverse V2: +0.9%

Change-Id: I33ba6854e7fbe6ecb1778b8201030a717f60f34e
---
 source/common/lowres.h | 12 ++----------
 1 file changed, 2 insertions(+), 10 deletions(-)

diff --git a/source/common/lowres.h b/source/common/lowres.h
index c9c2e3712..7e6baa844 100644
--- a/source/common/lowres.h
+++ b/source/common/lowres.h
@@ -71,11 +71,7 @@ struct ReferencePlanes
     inline pixel *lowresMC(intptr_t blockOffset, const MV& qmv, pixel *buf, intptr_t& outstride, bool hme)
     {
         intptr_t YStride = hme ? lumaStride / 2 : lumaStride;
-        pixel *plane[4];
-        for (int i = 0; i < 4; i++)
-        {
-            plane[i] = hme ? lowerResPlane[i] : lowresPlane[i];
-        }
+        pixel **plane = hme ? lowerResPlane : lowresPlane;
         if ((qmv.x | qmv.y) & 1)
         {
             int hpelA = (qmv.y & 2) | ((qmv.x & 2) >> 1);
@@ -98,11 +94,7 @@ struct ReferencePlanes
     inline int lowresQPelCost(pixel *fenc, intptr_t blockOffset, const MV& qmv, pixelcmp_t comp, bool hme)
     {
         intptr_t YStride = hme ? lumaStride / 2 : lumaStride;
-        pixel *plane[4];
-        for (int i = 0; i < 4; i++)
-        {
-            plane[i] = hme ? lowerResPlane[i] : lowresPlane[i];
-        }
+        pixel **plane = hme ? lowerResPlane : lowresPlane;
         if ((qmv.x | qmv.y) & 1)
         {
             ALIGN_VAR_16(pixel, subpelbuf[8 * 8]);
-- 
2.34.1


_______________________________________________
x265-devel mailing list
x265-devel@videolan.org
https://mailman.videolan.org/listinfo/x265-devel