mewin/libpng
1
0
Fork 0
mirror of https://git.code.sf.net/p/libpng/code.git synced 2025-07-10 18:04:09 +02:00
Code Issues Projects Releases Wiki Activity

Move sse opts into contrib/intel

Browse Source
This commit is contained in:
Matt Sarett
2016-02-18 12:43:50 -05:00
parent fb375429b5
commit 342c4eab2a
7 changed files with 73 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
contrib/intel/intel_init.c Normal file
View File

@@ -0,0 +1,45 @@
/* intel_init.c - SSE2 optimized filter functions
*
* Copyright (c) 2016 Google, Inc.
*
* This code is released under the libpng license.
* For conditions of distribution and use, see the disclaimer
* and license in png.h
*/
#include "../pngpriv.h"
#ifdef PNG_READ_SUPPORTED
#if PNG_INTEL_SSE_IMPLEMENTATION > 0
void
png_init_filter_functions_sse2(png_structp pp, unsigned int bpp)
{
// The techniques used to implement each of these filters in SSE operate on
// one pixel at a time.
// So they generally speed up 3bpp images about 3x, 4bpp images about 4x.
// They can scale up to 6 and 8 bpp images and down to 2 bpp images,
// but they'd not likely have any benefit for 1bpp images.
// Most of these can be implemented using only MMX and 64-bit registers,
// but they end up a bit slower than using the equally-ubiquitous SSE2.
if (bpp == 3)
{
pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_sse2;
pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_sse2;
pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
png_read_filter_row_paeth3_sse2;
}
else if (bpp == 4)
{
pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub4_sse2;
pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_sse2;
pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
png_read_filter_row_paeth4_sse2;
}
// No need optimize PNG_FILTER_VALUE_UP. The compiler should autovectorize.
}
#endif /* PNG_INTEL_SSE_IMPLEMENTATION > 0 */
#endif /* PNG_READ_SUPPORTED */