From: Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
Date: Tue, 15 Feb 2022 19:54:18 +0000 (+0100)
Subject: gdk: Specialise RGBA8 → premultiplied BGRA8 conversion
X-Git-Tag: archive/raspbian/4.6.5+ds-1+rpi1~1^2~19^2~3^2~79^2~1
X-Git-Url: https://dgit.raspbian.org/?a=commitdiff_plain;h=0e3ed7a7380a377c650dfe665db741498e183933;p=gtk4.git

gdk: Specialise RGBA8 → premultiplied BGRA8 conversion

On x86 on a Kaby Lake CPU, this makes it go from 6.63% of the total
execution time (loading some PNGs using the cairo backend) down to
3.20%.

On ARM on a Cortex-A7, on the same workload, this makes it go from 57%
to 8.36%.
---

diff --git a/gdk/gdkmemoryformat.c b/gdk/gdkmemoryformat.c
index 60c3317bc8..21667c1dcb 100644
--- a/gdk/gdkmemoryformat.c
+++ b/gdk/gdkmemoryformat.c
@@ -166,6 +166,29 @@ r32g32b32a32_float_from_float (guchar      *dest,
   memcpy (dest, src, sizeof (float) * n * 4);
 }
 
+// This one conversion is quite important, it converts from RGBA with straight
+// alpha (as found in PNG for instance) to BGRA with premultiplied alpha (the
+// sole cairo format available).
+static void
+r8g8b8a8_to_b8g8r8a8_premultiplied (guchar *dest,
+                                    const guchar *src,
+                                    gsize n)
+{
+  for (; n > 0; n--)
+    {
+      guchar a = src[3];
+      guint16 r = (guint16)src[0] * a + 127;
+      guint16 g = (guint16)src[1] * a + 127;
+      guint16 b = (guint16)src[2] * a + 127;
+      dest[0] = (b + (b >> 8) + 1) >> 8;
+      dest[1] = (g + (g >> 8) + 1) >> 8;
+      dest[2] = (r + (r >> 8) + 1) >> 8;
+      dest[3] = a;
+      dest += 4;
+      src += 4;
+    }
+}
+
 struct _GdkMemoryFormatDescription
 {
   GdkMemoryAlpha alpha;
@@ -479,6 +502,17 @@ gdk_memory_convert (guchar              *dest_data,
   g_assert (dest_format < GDK_MEMORY_N_FORMATS);
   g_assert (src_format < GDK_MEMORY_N_FORMATS);
 
+  if (src_format == GDK_MEMORY_R8G8B8A8 && dest_format == GDK_MEMORY_B8G8R8A8_PREMULTIPLIED)
+    {
+      for (y = 0; y < height; y++)
+        {
+          r8g8b8a8_to_b8g8r8a8_premultiplied (dest_data, src_data, width);
+          src_data += src_stride;
+          dest_data += dest_stride;
+        }
+      return;
+    }
+
   tmp = g_new (float, width * 4);
 
   for (y = 0; y < height; y++)