9de79859a9b0e8c2e95424f859276e49af42b4d2
[x262.git] / common / bitstream.c
1 /*****************************************************************************
2  * bitstream.c: bitstream writing
3  *****************************************************************************
4  * Copyright (C) 2003-2013 x264 project
5  *
6  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
7  *          Jason Garrett-Glaser <darkshikari@gmail.com>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  *
19  * You should have received a copy of the GNU General Public License
20  * along with this program; if not, write to the Free Software
21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
22  *
23  * This program is also available under a commercial proprietary license.
24  * For more information, contact us at licensing@x264.com.
25  *****************************************************************************/
26
27 #include "common.h"
28
29 static uint8_t *x264_nal_escape_c( uint8_t *dst, uint8_t *src, uint8_t *end )
30 {
31     if( src < end ) *dst++ = *src++;
32     if( src < end ) *dst++ = *src++;
33     while( src < end )
34     {
35         if( src[0] <= 0x03 && !dst[-2] && !dst[-1] )
36             *dst++ = 0x03;
37         *dst++ = *src++;
38     }
39     return dst;
40 }
41
42 uint8_t *x264_nal_escape_mmx2( uint8_t *dst, uint8_t *src, uint8_t *end );
43 uint8_t *x264_nal_escape_sse2( uint8_t *dst, uint8_t *src, uint8_t *end );
44 uint8_t *x264_nal_escape_avx2( uint8_t *dst, uint8_t *src, uint8_t *end );
45 void x264_cabac_block_residual_rd_internal_sse2       ( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
46 void x264_cabac_block_residual_rd_internal_sse2_lzcnt ( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
47 void x264_cabac_block_residual_rd_internal_ssse3      ( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
48 void x264_cabac_block_residual_rd_internal_ssse3_lzcnt( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
49 void x264_cabac_block_residual_8x8_rd_internal_sse2       ( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
50 void x264_cabac_block_residual_8x8_rd_internal_sse2_lzcnt ( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
51 void x264_cabac_block_residual_8x8_rd_internal_ssse3      ( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
52 void x264_cabac_block_residual_8x8_rd_internal_ssse3_lzcnt( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
53 void x264_cabac_block_residual_internal_sse2       ( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
54 void x264_cabac_block_residual_internal_sse2_lzcnt ( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
55 void x264_cabac_block_residual_internal_avx2_bmi2 ( dctcoef *l, int b_interlaced, intptr_t ctx_block_cat, x264_cabac_t *cb );
56
57 /****************************************************************************
58  * x264_nal_encode:
59  ****************************************************************************/
60 void x264_nal_encode( x264_t *h, uint8_t *dst, x264_nal_t *nal )
61 {
62     uint8_t *src = nal->p_payload;
63     uint8_t *end = nal->p_payload + nal->i_payload;
64     uint8_t *orig_dst = dst;
65
66     if( h->param.b_annexb )
67     {
68         if( nal->b_long_startcode )
69             *dst++ = 0x00;
70         *dst++ = 0x00;
71         *dst++ = 0x00;
72         *dst++ = 0x01;
73     }
74     else /* save room for size later */
75         dst += 4;
76
77     /* nal header */
78     *dst++ = ( 0x00 << 7 ) | ( nal->i_ref_idc << 5 ) | nal->i_type;
79
80     dst = h->bsf.nal_escape( dst, src, end );
81     int size = (dst - orig_dst) - 4;
82
83     /* Write the size header for mp4/etc */
84     if( !h->param.b_annexb )
85     {
86         /* Size doesn't include the size of the header we're writing now. */
87         orig_dst[0] = size>>24;
88         orig_dst[1] = size>>16;
89         orig_dst[2] = size>> 8;
90         orig_dst[3] = size>> 0;
91     }
92
93     nal->i_payload = size+4;
94     nal->p_payload = orig_dst;
95     x264_emms();
96 }
97
98 void x264_bitstream_init( int cpu, x264_bitstream_function_t *pf )
99 {
100     memset( pf, 0, sizeof(*pf) );
101
102     pf->nal_escape = x264_nal_escape_c;
103 #if HAVE_MMX
104 #if ARCH_X86_64
105     pf->cabac_block_residual_internal = x264_cabac_block_residual_internal_sse2;
106     pf->cabac_block_residual_rd_internal = x264_cabac_block_residual_rd_internal_sse2;
107     pf->cabac_block_residual_8x8_rd_internal = x264_cabac_block_residual_8x8_rd_internal_sse2;
108 #endif
109
110     if( cpu&X264_CPU_MMX2 )
111         pf->nal_escape = x264_nal_escape_mmx2;
112     if( cpu&X264_CPU_SSE2 )
113     {
114 #if ARCH_X86_64
115         if( cpu&X264_CPU_LZCNT )
116         {
117             pf->cabac_block_residual_internal = x264_cabac_block_residual_internal_sse2_lzcnt;
118             pf->cabac_block_residual_rd_internal = x264_cabac_block_residual_rd_internal_sse2_lzcnt;
119             pf->cabac_block_residual_8x8_rd_internal = x264_cabac_block_residual_8x8_rd_internal_sse2_lzcnt;
120         }
121 #endif
122         if( cpu&X264_CPU_SSE2_IS_FAST )
123             pf->nal_escape = x264_nal_escape_sse2;
124     }
125 #if ARCH_X86_64
126     if( cpu&X264_CPU_SSSE3 )
127     {
128         pf->cabac_block_residual_rd_internal = x264_cabac_block_residual_rd_internal_ssse3;
129         pf->cabac_block_residual_8x8_rd_internal = x264_cabac_block_residual_8x8_rd_internal_ssse3;
130         if( cpu&X264_CPU_LZCNT )
131         {
132             pf->cabac_block_residual_rd_internal = x264_cabac_block_residual_rd_internal_ssse3_lzcnt;
133             pf->cabac_block_residual_8x8_rd_internal = x264_cabac_block_residual_8x8_rd_internal_ssse3_lzcnt;
134         }
135     }
136
137     if( cpu&X264_CPU_AVX2 )
138     {
139         pf->nal_escape = x264_nal_escape_avx2;
140         if( cpu&X264_CPU_BMI2 )
141             pf->cabac_block_residual_internal = x264_cabac_block_residual_internal_avx2_bmi2;
142     }
143 #endif
144 #endif
145 }