avcodec/x86: move simple_idct to external assembly
[ffmpeg.git] / libavcodec / x86 / idctdsp_init.c
1 /*
2  * This file is part of FFmpeg.
3  *
4  * FFmpeg 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  * FFmpeg 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 FFmpeg; 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 "idctdsp.h"
26 #include "simple_idct.h"
27
28 /* Input permutation for the simple_idct_mmx */
29 static const uint8_t simple_mmx_permutation[64] = {
30     0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
31     0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
32     0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
33     0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
34     0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
35     0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
36     0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
37     0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
38 };
39
40 static const uint8_t idct_sse2_row_perm[8] = { 0, 4, 1, 5, 2, 6, 3, 7 };
41
42 av_cold int ff_init_scantable_permutation_x86(uint8_t *idct_permutation,
43                                               enum idct_permutation_type perm_type)
44 {
45     int i;
46
47     switch (perm_type) {
48     case FF_IDCT_PERM_SIMPLE:
49         for (i = 0; i < 64; i++)
50             idct_permutation[i] = simple_mmx_permutation[i];
51         return 1;
52     case FF_IDCT_PERM_SSE2:
53         for (i = 0; i < 64; i++)
54             idct_permutation[i] = (i & 0x38) | idct_sse2_row_perm[i & 7];
55         return 1;
56     }
57
58     return 0;
59 }
60
61 av_cold void ff_idctdsp_init_x86(IDCTDSPContext *c, AVCodecContext *avctx,
62                                  unsigned high_bit_depth)
63 {
64     int cpu_flags = av_get_cpu_flags();
65
66     if (EXTERNAL_MMX(cpu_flags)) {
67         c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
68         c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
69         c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
70
71             if (!high_bit_depth &&
72                 avctx->lowres == 0 &&
73                 (avctx->idct_algo == FF_IDCT_AUTO ||
74                  avctx->idct_algo == FF_IDCT_SIMPLEAUTO ||
75                  avctx->idct_algo == 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             }
81     }
82
83     if (EXTERNAL_SSE2(cpu_flags)) {
84         c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_sse2;
85         c->put_pixels_clamped        = ff_put_pixels_clamped_sse2;
86         c->add_pixels_clamped        = ff_add_pixels_clamped_sse2;
87
88             if (!high_bit_depth &&
89                 avctx->lowres == 0 &&
90                 (avctx->idct_algo == FF_IDCT_AUTO ||
91                  avctx->idct_algo == FF_IDCT_SIMPLEAUTO ||
92                  avctx->idct_algo == FF_IDCT_SIMPLEMMX)) {
93                     c->idct_put  = ff_simple_idct_put_sse2;
94                     c->idct_add  = ff_simple_idct_add_sse2;
95                     c->perm_type = FF_IDCT_PERM_SIMPLE;
96             }
97     }
98
99     if (ARCH_X86_64 && avctx->lowres == 0) {
100         if (avctx->bits_per_raw_sample == 10 &&
101             (avctx->idct_algo == FF_IDCT_AUTO ||
102              avctx->idct_algo == FF_IDCT_SIMPLEAUTO ||
103              avctx->idct_algo == FF_IDCT_SIMPLE)) {
104             if (EXTERNAL_SSE2(cpu_flags)) {
105                 c->idct_put  = ff_simple_idct10_put_sse2;
106                 c->idct_add  = NULL;
107                 c->idct      = ff_simple_idct10_sse2;
108                 c->perm_type = FF_IDCT_PERM_TRANSPOSE;
109
110             }
111             if (EXTERNAL_AVX(cpu_flags)) {
112                 c->idct_put  = ff_simple_idct10_put_avx;
113                 c->idct_add  = NULL;
114                 c->idct      = ff_simple_idct10_avx;
115                 c->perm_type = FF_IDCT_PERM_TRANSPOSE;
116             }
117         }
118
119         if (avctx->bits_per_raw_sample == 12 &&
120             (avctx->idct_algo == FF_IDCT_AUTO ||
121              avctx->idct_algo == FF_IDCT_SIMPLEMMX)) {
122             if (EXTERNAL_SSE2(cpu_flags)) {
123                 c->idct_put  = ff_simple_idct12_put_sse2;
124                 c->idct_add  = NULL;
125                 c->idct      = ff_simple_idct12_sse2;
126                 c->perm_type = FF_IDCT_PERM_TRANSPOSE;
127             }
128             if (EXTERNAL_AVX(cpu_flags)) {
129                 c->idct_put  = ff_simple_idct12_put_avx;
130                 c->idct_add  = NULL;
131                 c->idct      = ff_simple_idct12_avx;
132                 c->perm_type = FF_IDCT_PERM_TRANSPOSE;
133             }
134         }
135     }
136 }