3f23bf67994383655c1283e0a07fac35aff0c9e5
[ffmpeg.git] / libavfilter / x86 / vf_gradfun_init.c
1 /*
2  * Copyright (C) 2009 Loren Merritt <lorenm@u.washington.edu>
3  *
4  * This file is part of Libav.
5  *
6  * Libav is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * Libav is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with Libav; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 #include "config.h"
22 #include "libavutil/attributes.h"
23 #include "libavutil/cpu.h"
24 #include "libavutil/mem.h"
25 #include "libavutil/x86/asm.h"
26 #include "libavutil/x86/cpu.h"
27 #include "libavfilter/gradfun.h"
28
29 void ff_gradfun_filter_line_mmxext(intptr_t x, uint8_t *dst, uint8_t *src,
30                                    uint16_t *dc, int thresh,
31                                    const uint16_t *dithers);
32
33 void ff_gradfun_filter_line_ssse3(intptr_t x, uint8_t *dst, uint8_t *src,
34                                   uint16_t *dc, int thresh,
35                                   const uint16_t *dithers);
36
37 void ff_gradfun_blur_line_movdqa_sse2(intptr_t x, uint16_t *buf,
38                                       uint16_t *buf1, uint16_t *dc,
39                                       uint8_t *src1, uint8_t *src2);
40 void ff_gradfun_blur_line_movdqu_sse2(intptr_t x, uint16_t *buf,
41                                       uint16_t *buf1, uint16_t *dc,
42                                       uint8_t *src1, uint8_t *src2);
43
44 #if HAVE_YASM
45 static void gradfun_filter_line(uint8_t *dst, uint8_t *src, uint16_t *dc,
46                                 int width, int thresh, const uint16_t *dithers,
47                                 int alignment)
48 {
49     intptr_t x;
50     if (width & alignment) {
51         x = width & ~alignment;
52         ff_gradfun_filter_line_c(dst + x, src + x, dc + x / 2,
53                                  width - x, thresh, dithers);
54         width = x;
55     }
56     x = -width;
57     ff_gradfun_filter_line_mmxext(x, dst + width, src + width, dc + width / 2,
58                                   thresh, dithers);
59 }
60
61 static void gradfun_filter_line_mmxext(uint8_t *dst, uint8_t *src, uint16_t *dc,
62                                        int width, int thresh,
63                                        const uint16_t *dithers)
64 {
65     gradfun_filter_line(dst, src, dc, width, thresh, dithers, 3);
66 }
67
68 static void gradfun_filter_line_ssse3(uint8_t *dst, uint8_t *src, uint16_t *dc,
69                                       int width, int thresh,
70                                       const uint16_t *dithers)
71 {
72     gradfun_filter_line(dst, src, dc, width, thresh, dithers, 7);
73 }
74
75 static void gradfun_blur_line_sse2(uint16_t *dc, uint16_t *buf, uint16_t *buf1,
76                                    uint8_t *src, int src_linesize, int width)
77 {
78     intptr_t x = -2 * width;
79     if (((intptr_t) src | src_linesize) & 15)
80         ff_gradfun_blur_line_movdqu_sse2(x, buf + width, buf1 + width,
81                                          dc + width, src + width * 2,
82                                          src + width * 2 + src_linesize);
83     else
84         ff_gradfun_blur_line_movdqa_sse2(x, buf + width, buf1 + width,
85                                          dc + width, src + width * 2,
86                                          src + width * 2 + src_linesize);
87 }
88 #endif /* HAVE_YASM */
89
90 av_cold void ff_gradfun_init_x86(GradFunContext *gf)
91 {
92 #if HAVE_YASM
93     int cpu_flags = av_get_cpu_flags();
94
95     if (EXTERNAL_MMXEXT(cpu_flags))
96         gf->filter_line = gradfun_filter_line_mmxext;
97     if (EXTERNAL_SSSE3(cpu_flags))
98         gf->filter_line = gradfun_filter_line_ssse3;
99
100     if (EXTERNAL_SSE2(cpu_flags))
101         gf->blur_line = gradfun_blur_line_sse2;
102 #endif /* HAVE_YASM */
103 }