5bd293fb6b3de3b4f16d080b103b0be349daa310
[ffmpeg.git] / libavcodec / x86 / idctdsp_init.c
1 /*
2  * This file is part of Libav.
3  *
4  * Libav is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU Lesser General Public
6  * License as published by the Free Software Foundation; either
7  * version 2.1 of the License, or (at your option) any later version.
8  *
9  * Libav is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
12  * Lesser General Public License for more details.
13  *
14  * You should have received a copy of the GNU Lesser General Public
15  * License along with Libav; if not, write to the Free Software
16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17  */
18
19 #include "config.h"
20 #include "libavutil/attributes.h"
21 #include "libavutil/cpu.h"
22 #include "libavutil/x86/cpu.h"
23 #include "libavcodec/avcodec.h"
24 #include "libavcodec/idctdsp.h"
25 #include "libavcodec/simple_idct.h"
26 #include "idct_xvid.h"
27 #include "idctdsp.h"
28
29 /* Input permutation for the simple_idct_mmx */
30 static const uint8_t simple_mmx_permutation[64] = {
31     0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
32     0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
33     0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
34     0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
35     0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
36     0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
37     0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
38     0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
39 };
40
41 static const uint8_t idct_sse2_row_perm[8] = { 0, 4, 1, 5, 2, 6, 3, 7 };
42
43 av_cold int ff_init_scantable_permutation_x86(uint8_t *idct_permutation,
44                                               enum idct_permutation_type perm_type)
45 {
46     int i;
47
48     switch (perm_type) {
49     case FF_IDCT_PERM_SIMPLE:
50         for (i = 0; i < 64; i++)
51             idct_permutation[i] = simple_mmx_permutation[i];
52         return 1;
53     case FF_IDCT_PERM_SSE2:
54         for (i = 0; i < 64; i++)
55             idct_permutation[i] = (i & 0x38) | idct_sse2_row_perm[i & 7];
56         return 1;
57     }
58
59     return 0;
60 }
61
62 av_cold void ff_idctdsp_init_x86(IDCTDSPContext *c, AVCodecContext *avctx,
63                                  unsigned high_bit_depth)
64 {
65     int cpu_flags = av_get_cpu_flags();
66
67     if (INLINE_MMX(cpu_flags)) {
68         c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
69         c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
70         c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
71
72         if (!high_bit_depth) {
73             switch (avctx->idct_algo) {
74             case FF_IDCT_AUTO:
75             case FF_IDCT_SIMPLEMMX:
76                 c->idct_put              = ff_simple_idct_put_mmx;
77                 c->idct_add              = ff_simple_idct_add_mmx;
78                 c->idct                  = ff_simple_idct_mmx;
79                 c->perm_type             = FF_IDCT_PERM_SIMPLE;
80                 break;
81             case FF_IDCT_XVIDMMX:
82                 c->idct_put              = ff_idct_xvid_mmx_put;
83                 c->idct_add              = ff_idct_xvid_mmx_add;
84                 c->idct                  = ff_idct_xvid_mmx;
85                 break;
86             }
87         }
88     }
89
90     if (INLINE_MMXEXT(cpu_flags)) {
91         if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
92             c->idct_put = ff_idct_xvid_mmxext_put;
93             c->idct_add = ff_idct_xvid_mmxext_add;
94             c->idct     = ff_idct_xvid_mmxext;
95         }
96     }
97
98     if (INLINE_SSE2(cpu_flags)) {
99         if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
100             c->idct_put              = ff_idct_xvid_sse2_put;
101             c->idct_add              = ff_idct_xvid_sse2_add;
102             c->idct                  = ff_idct_xvid_sse2;
103             c->perm_type             = FF_IDCT_PERM_SSE2;
104         }
105     }
106 }