1
 2/* intel_init.c - SSE2 optimized filter functions
 3 *
 4 * Copyright (c) 2016-2017 Glenn Randers-Pehrson
 5 * Written by Mike Klein and Matt Sarett, Google, Inc.
 6 * Derived from arm/arm_init.c
 7 *
 8 * Last changed in libpng 1.6.29 [March 16, 2017]
 9 *
10 * This code is released under the libpng license.
11 * For conditions of distribution and use, see the disclaimer
12 * and license in png.h
13 */
14
15#include "../pngpriv.h"
16
17#ifdef PNG_READ_SUPPORTED
18#if PNG_INTEL_SSE_IMPLEMENTATION > 0
19
20void
21png_init_filter_functions_sse2(png_structp pp, unsigned int bpp)
22{
23   /* The techniques used to implement each of these filters in SSE operate on
24    * one pixel at a time.
25    * So they generally speed up 3bpp images about 3x, 4bpp images about 4x.
26    * They can scale up to 6 and 8 bpp images and down to 2 bpp images,
27    * but they'd not likely have any benefit for 1bpp images.
28    * Most of these can be implemented using only MMX and 64-bit registers,
29    * but they end up a bit slower than using the equally-ubiquitous SSE2.
30   */
31   png_debug(1, "in png_init_filter_functions_sse2");
32   if (bpp == 3)
33   {
34      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_sse2;
35      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_sse2;
36      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
37         png_read_filter_row_paeth3_sse2;
38   }
39   else if (bpp == 4)
40   {
41      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub4_sse2;
42      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_sse2;
43      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
44          png_read_filter_row_paeth4_sse2;
45   }
46
47   /* No need optimize PNG_FILTER_VALUE_UP.  The compiler should
48    * autovectorize.
49    */
50}
51
52#endif /* PNG_INTEL_SSE_IMPLEMENTATION > 0 */
53#endif /* PNG_READ_SUPPORTED */