H.264: Use 64-/128-bit write-combining macros for copies
[ffmpeg.git] / libavcodec / h264.h
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file libavcodec/h264.h
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #ifndef AVCODEC_H264_H
29 #define AVCODEC_H264_H
30
31 #include "dsputil.h"
32 #include "cabac.h"
33 #include "mpegvideo.h"
34 #include "h264pred.h"
35 #include "rectangle.h"
36
37 #define interlaced_dct interlaced_dct_is_a_bad_name
38 #define mb_intra mb_intra_is_not_initialized_see_mb_type
39
40 #define LUMA_DC_BLOCK_INDEX   25
41 #define CHROMA_DC_BLOCK_INDEX 26
42
43 #define CHROMA_DC_COEFF_TOKEN_VLC_BITS 8
44 #define COEFF_TOKEN_VLC_BITS           8
45 #define TOTAL_ZEROS_VLC_BITS           9
46 #define CHROMA_DC_TOTAL_ZEROS_VLC_BITS 3
47 #define RUN_VLC_BITS                   3
48 #define RUN7_VLC_BITS                  6
49
50 #define MAX_SPS_COUNT 32
51 #define MAX_PPS_COUNT 256
52
53 #define MAX_MMCO_COUNT 66
54
55 #define MAX_DELAYED_PIC_COUNT 16
56
57 /* Compiling in interlaced support reduces the speed
58  * of progressive decoding by about 2%. */
59 #define ALLOW_INTERLACE
60
61 #define ALLOW_NOCHROMA
62
63 /**
64  * The maximum number of slices supported by the decoder.
65  * must be a power of 2
66  */
67 #define MAX_SLICES 16
68
69 #ifdef ALLOW_INTERLACE
70 #define MB_MBAFF h->mb_mbaff
71 #define MB_FIELD h->mb_field_decoding_flag
72 #define FRAME_MBAFF h->mb_aff_frame
73 #define FIELD_PICTURE (s->picture_structure != PICT_FRAME)
74 #else
75 #define MB_MBAFF 0
76 #define MB_FIELD 0
77 #define FRAME_MBAFF 0
78 #define FIELD_PICTURE 0
79 #undef  IS_INTERLACED
80 #define IS_INTERLACED(mb_type) 0
81 #endif
82 #define FIELD_OR_MBAFF_PICTURE (FRAME_MBAFF || FIELD_PICTURE)
83
84 #ifdef ALLOW_NOCHROMA
85 #define CHROMA h->sps.chroma_format_idc
86 #else
87 #define CHROMA 1
88 #endif
89
90 #ifndef CABAC
91 #define CABAC h->pps.cabac
92 #endif
93
94 #define EXTENDED_SAR          255
95
96 #define MB_TYPE_REF0       MB_TYPE_ACPRED //dirty but it fits in 16 bit
97 #define MB_TYPE_8x8DCT     0x01000000
98 #define IS_REF0(a)         ((a) & MB_TYPE_REF0)
99 #define IS_8x8DCT(a)       ((a) & MB_TYPE_8x8DCT)
100
101 /**
102  * Value of Picture.reference when Picture is not a reference picture, but
103  * is held for delayed output.
104  */
105 #define DELAYED_PIC_REF 4
106
107
108 /* NAL unit types */
109 enum {
110     NAL_SLICE=1,
111     NAL_DPA,
112     NAL_DPB,
113     NAL_DPC,
114     NAL_IDR_SLICE,
115     NAL_SEI,
116     NAL_SPS,
117     NAL_PPS,
118     NAL_AUD,
119     NAL_END_SEQUENCE,
120     NAL_END_STREAM,
121     NAL_FILLER_DATA,
122     NAL_SPS_EXT,
123     NAL_AUXILIARY_SLICE=19
124 };
125
126 /**
127  * SEI message types
128  */
129 typedef enum {
130     SEI_BUFFERING_PERIOD             =  0, ///< buffering period (H.264, D.1.1)
131     SEI_TYPE_PIC_TIMING              =  1, ///< picture timing
132     SEI_TYPE_USER_DATA_UNREGISTERED  =  5, ///< unregistered user data
133     SEI_TYPE_RECOVERY_POINT          =  6  ///< recovery point (frame # to decoder sync)
134 } SEI_Type;
135
136 /**
137  * pic_struct in picture timing SEI message
138  */
139 typedef enum {
140     SEI_PIC_STRUCT_FRAME             = 0, ///<  0: %frame
141     SEI_PIC_STRUCT_TOP_FIELD         = 1, ///<  1: top field
142     SEI_PIC_STRUCT_BOTTOM_FIELD      = 2, ///<  2: bottom field
143     SEI_PIC_STRUCT_TOP_BOTTOM        = 3, ///<  3: top field, bottom field, in that order
144     SEI_PIC_STRUCT_BOTTOM_TOP        = 4, ///<  4: bottom field, top field, in that order
145     SEI_PIC_STRUCT_TOP_BOTTOM_TOP    = 5, ///<  5: top field, bottom field, top field repeated, in that order
146     SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM = 6, ///<  6: bottom field, top field, bottom field repeated, in that order
147     SEI_PIC_STRUCT_FRAME_DOUBLING    = 7, ///<  7: %frame doubling
148     SEI_PIC_STRUCT_FRAME_TRIPLING    = 8  ///<  8: %frame tripling
149 } SEI_PicStructType;
150
151 /**
152  * Sequence parameter set
153  */
154 typedef struct SPS{
155
156     int profile_idc;
157     int level_idc;
158     int chroma_format_idc;
159     int transform_bypass;              ///< qpprime_y_zero_transform_bypass_flag
160     int log2_max_frame_num;            ///< log2_max_frame_num_minus4 + 4
161     int poc_type;                      ///< pic_order_cnt_type
162     int log2_max_poc_lsb;              ///< log2_max_pic_order_cnt_lsb_minus4
163     int delta_pic_order_always_zero_flag;
164     int offset_for_non_ref_pic;
165     int offset_for_top_to_bottom_field;
166     int poc_cycle_length;              ///< num_ref_frames_in_pic_order_cnt_cycle
167     int ref_frame_count;               ///< num_ref_frames
168     int gaps_in_frame_num_allowed_flag;
169     int mb_width;                      ///< pic_width_in_mbs_minus1 + 1
170     int mb_height;                     ///< pic_height_in_map_units_minus1 + 1
171     int frame_mbs_only_flag;
172     int mb_aff;                        ///<mb_adaptive_frame_field_flag
173     int direct_8x8_inference_flag;
174     int crop;                   ///< frame_cropping_flag
175     unsigned int crop_left;            ///< frame_cropping_rect_left_offset
176     unsigned int crop_right;           ///< frame_cropping_rect_right_offset
177     unsigned int crop_top;             ///< frame_cropping_rect_top_offset
178     unsigned int crop_bottom;          ///< frame_cropping_rect_bottom_offset
179     int vui_parameters_present_flag;
180     AVRational sar;
181     int video_signal_type_present_flag;
182     int full_range;
183     int colour_description_present_flag;
184     enum AVColorPrimaries color_primaries;
185     enum AVColorTransferCharacteristic color_trc;
186     enum AVColorSpace colorspace;
187     int timing_info_present_flag;
188     uint32_t num_units_in_tick;
189     uint32_t time_scale;
190     int fixed_frame_rate_flag;
191     short offset_for_ref_frame[256]; //FIXME dyn aloc?
192     int bitstream_restriction_flag;
193     int num_reorder_frames;
194     int scaling_matrix_present;
195     uint8_t scaling_matrix4[6][16];
196     uint8_t scaling_matrix8[2][64];
197     int nal_hrd_parameters_present_flag;
198     int vcl_hrd_parameters_present_flag;
199     int pic_struct_present_flag;
200     int time_offset_length;
201     int cpb_cnt;                       ///< See H.264 E.1.2
202     int initial_cpb_removal_delay_length; ///< initial_cpb_removal_delay_length_minus1 +1
203     int cpb_removal_delay_length;      ///< cpb_removal_delay_length_minus1 + 1
204     int dpb_output_delay_length;       ///< dpb_output_delay_length_minus1 + 1
205     int bit_depth_luma;                ///< bit_depth_luma_minus8 + 8
206     int bit_depth_chroma;              ///< bit_depth_chroma_minus8 + 8
207     int residual_color_transform_flag; ///< residual_colour_transform_flag
208 }SPS;
209
210 /**
211  * Picture parameter set
212  */
213 typedef struct PPS{
214     unsigned int sps_id;
215     int cabac;                  ///< entropy_coding_mode_flag
216     int pic_order_present;      ///< pic_order_present_flag
217     int slice_group_count;      ///< num_slice_groups_minus1 + 1
218     int mb_slice_group_map_type;
219     unsigned int ref_count[2];  ///< num_ref_idx_l0/1_active_minus1 + 1
220     int weighted_pred;          ///< weighted_pred_flag
221     int weighted_bipred_idc;
222     int init_qp;                ///< pic_init_qp_minus26 + 26
223     int init_qs;                ///< pic_init_qs_minus26 + 26
224     int chroma_qp_index_offset[2];
225     int deblocking_filter_parameters_present; ///< deblocking_filter_parameters_present_flag
226     int constrained_intra_pred; ///< constrained_intra_pred_flag
227     int redundant_pic_cnt_present; ///< redundant_pic_cnt_present_flag
228     int transform_8x8_mode;     ///< transform_8x8_mode_flag
229     uint8_t scaling_matrix4[6][16];
230     uint8_t scaling_matrix8[2][64];
231     uint8_t chroma_qp_table[2][64];  ///< pre-scaled (with chroma_qp_index_offset) version of qp_table
232     int chroma_qp_diff;
233 }PPS;
234
235 /**
236  * Memory management control operation opcode.
237  */
238 typedef enum MMCOOpcode{
239     MMCO_END=0,
240     MMCO_SHORT2UNUSED,
241     MMCO_LONG2UNUSED,
242     MMCO_SHORT2LONG,
243     MMCO_SET_MAX_LONG,
244     MMCO_RESET,
245     MMCO_LONG,
246 } MMCOOpcode;
247
248 /**
249  * Memory management control operation.
250  */
251 typedef struct MMCO{
252     MMCOOpcode opcode;
253     int short_pic_num;  ///< pic_num without wrapping (pic_num & max_pic_num)
254     int long_arg;       ///< index, pic_num, or num long refs depending on opcode
255 } MMCO;
256
257 /**
258  * H264Context
259  */
260 typedef struct H264Context{
261     MpegEncContext s;
262     int nal_ref_idc;
263     int nal_unit_type;
264     uint8_t *rbsp_buffer[2];
265     unsigned int rbsp_buffer_size[2];
266
267     /**
268       * Used to parse AVC variant of h264
269       */
270     int is_avc; ///< this flag is != 0 if codec is avc1
271     int got_avcC; ///< flag used to parse avcC data only once
272     int nal_length_size; ///< Number of bytes used for nal length (1, 2 or 4)
273
274     int chroma_qp[2]; //QPc
275
276     int qp_thresh;      ///< QP threshold to skip loopfilter
277
278     int prev_mb_skipped;
279     int next_mb_skipped;
280
281     //prediction stuff
282     int chroma_pred_mode;
283     int intra16x16_pred_mode;
284
285     int top_mb_xy;
286     int left_mb_xy[2];
287
288     int8_t intra4x4_pred_mode_cache[5*8];
289     int8_t (*intra4x4_pred_mode)[8];
290     H264PredContext hpc;
291     unsigned int topleft_samples_available;
292     unsigned int top_samples_available;
293     unsigned int topright_samples_available;
294     unsigned int left_samples_available;
295     uint8_t (*top_borders[2])[16+2*8];
296     uint8_t left_border[2*(17+2*9)];
297
298     /**
299      * non zero coeff count cache.
300      * is 64 if not available.
301      */
302     DECLARE_ALIGNED_8(uint8_t, non_zero_count_cache)[6*8];
303
304     /*
305     .UU.YYYY
306     .UU.YYYY
307     .vv.YYYY
308     .VV.YYYY
309     */
310     uint8_t (*non_zero_count)[32];
311
312     /**
313      * Motion vector cache.
314      */
315     DECLARE_ALIGNED_16(int16_t, mv_cache)[2][5*8][2];
316     DECLARE_ALIGNED_8(int8_t, ref_cache)[2][5*8];
317 #define LIST_NOT_USED -1 //FIXME rename?
318 #define PART_NOT_AVAILABLE -2
319
320     /**
321      * is 1 if the specific list MV&references are set to 0,0,-2.
322      */
323     int mv_cache_clean[2];
324
325     /**
326      * number of neighbors (top and/or left) that used 8x8 dct
327      */
328     int neighbor_transform_size;
329
330     /**
331      * block_offset[ 0..23] for frame macroblocks
332      * block_offset[24..47] for field macroblocks
333      */
334     int block_offset[2*(16+8)];
335
336     uint32_t *mb2b_xy; //FIXME are these 4 a good idea?
337     uint32_t *mb2b8_xy;
338     int b_stride; //FIXME use s->b4_stride
339     int b8_stride;
340
341     int mb_linesize;   ///< may be equal to s->linesize or s->linesize*2, for mbaff
342     int mb_uvlinesize;
343
344     int emu_edge_width;
345     int emu_edge_height;
346
347     int halfpel_flag;
348     int thirdpel_flag;
349
350     int unknown_svq3_flag;
351     int next_slice_index;
352
353     SPS *sps_buffers[MAX_SPS_COUNT];
354     SPS sps; ///< current sps
355
356     PPS *pps_buffers[MAX_PPS_COUNT];
357     /**
358      * current pps
359      */
360     PPS pps; //FIXME move to Picture perhaps? (->no) do we need that?
361
362     uint32_t dequant4_buffer[6][52][16];
363     uint32_t dequant8_buffer[2][52][64];
364     uint32_t (*dequant4_coeff[6])[16];
365     uint32_t (*dequant8_coeff[2])[64];
366     int dequant_coeff_pps;     ///< reinit tables when pps changes
367
368     int slice_num;
369     uint16_t *slice_table_base;
370     uint16_t *slice_table;     ///< slice_table_base + 2*mb_stride + 1
371     int slice_type;
372     int slice_type_nos;        ///< S free slice type (SI/SP are remapped to I/P)
373     int slice_type_fixed;
374
375     //interlacing specific flags
376     int mb_aff_frame;
377     int mb_field_decoding_flag;
378     int mb_mbaff;              ///< mb_aff_frame && mb_field_decoding_flag
379
380     DECLARE_ALIGNED_8(uint16_t, sub_mb_type)[4];
381
382     //POC stuff
383     int poc_lsb;
384     int poc_msb;
385     int delta_poc_bottom;
386     int delta_poc[2];
387     int frame_num;
388     int prev_poc_msb;             ///< poc_msb of the last reference pic for POC type 0
389     int prev_poc_lsb;             ///< poc_lsb of the last reference pic for POC type 0
390     int frame_num_offset;         ///< for POC type 2
391     int prev_frame_num_offset;    ///< for POC type 2
392     int prev_frame_num;           ///< frame_num of the last pic for POC type 1/2
393
394     /**
395      * frame_num for frames or 2*frame_num+1 for field pics.
396      */
397     int curr_pic_num;
398
399     /**
400      * max_frame_num or 2*max_frame_num for field pics.
401      */
402     int max_pic_num;
403
404     //Weighted pred stuff
405     int use_weight;
406     int use_weight_chroma;
407     int luma_log2_weight_denom;
408     int chroma_log2_weight_denom;
409     int luma_weight[2][48];
410     int luma_offset[2][48];
411     int chroma_weight[2][48][2];
412     int chroma_offset[2][48][2];
413     int implicit_weight[48][48];
414
415     //deblock
416     int deblocking_filter;         ///< disable_deblocking_filter_idc with 1<->0
417     int slice_alpha_c0_offset;
418     int slice_beta_offset;
419
420     int redundant_pic_count;
421
422     int direct_spatial_mv_pred;
423     int dist_scale_factor[16];
424     int dist_scale_factor_field[2][32];
425     int map_col_to_list0[2][16+32];
426     int map_col_to_list0_field[2][2][16+32];
427
428     /**
429      * num_ref_idx_l0/1_active_minus1 + 1
430      */
431     unsigned int ref_count[2];   ///< counts frames or fields, depending on current mb mode
432     unsigned int list_count;
433     uint8_t *list_counts;            ///< Array of list_count per MB specifying the slice type
434     Picture *short_ref[32];
435     Picture *long_ref[32];
436     Picture default_ref_list[2][32]; ///< base reference list for all slices of a coded picture
437     Picture ref_list[2][48];         /**< 0..15: frame refs, 16..47: mbaff field refs.
438                                           Reordered version of default_ref_list
439                                           according to picture reordering in slice header */
440     int ref2frm[MAX_SLICES][2][64];  ///< reference to frame number lists, used in the loop filter, the first 2 are for -2,-1
441     Picture *delayed_pic[MAX_DELAYED_PIC_COUNT+2]; //FIXME size?
442     int outputed_poc;
443
444     /**
445      * memory management control operations buffer.
446      */
447     MMCO mmco[MAX_MMCO_COUNT];
448     int mmco_index;
449
450     int long_ref_count;  ///< number of actual long term references
451     int short_ref_count; ///< number of actual short term references
452
453     //data partitioning
454     GetBitContext intra_gb;
455     GetBitContext inter_gb;
456     GetBitContext *intra_gb_ptr;
457     GetBitContext *inter_gb_ptr;
458
459     DECLARE_ALIGNED_16(DCTELEM, mb)[16*24];
460     DCTELEM mb_padding[256];        ///< as mb is addressed by scantable[i] and scantable is uint8_t we can either check that i is not too large or ensure that there is some unused stuff after mb
461
462     /**
463      * Cabac
464      */
465     CABACContext cabac;
466     uint8_t      cabac_state[460];
467     int          cabac_init_idc;
468
469     /* 0x100 -> non null luma_dc, 0x80/0x40 -> non null chroma_dc (cb/cr), 0x?0 -> chroma_cbp(0,1,2), 0x0? luma_cbp */
470     uint16_t     *cbp_table;
471     int cbp;
472     int top_cbp;
473     int left_cbp;
474     /* chroma_pred_mode for i4x4 or i16x16, else 0 */
475     uint8_t     *chroma_pred_mode_table;
476     int         last_qscale_diff;
477     int16_t     (*mvd_table[2])[2];
478     DECLARE_ALIGNED_16(int16_t, mvd_cache)[2][5*8][2];
479     uint8_t     *direct_table;
480     uint8_t     direct_cache[5*8];
481
482     uint8_t zigzag_scan[16];
483     uint8_t zigzag_scan8x8[64];
484     uint8_t zigzag_scan8x8_cavlc[64];
485     uint8_t field_scan[16];
486     uint8_t field_scan8x8[64];
487     uint8_t field_scan8x8_cavlc[64];
488     const uint8_t *zigzag_scan_q0;
489     const uint8_t *zigzag_scan8x8_q0;
490     const uint8_t *zigzag_scan8x8_cavlc_q0;
491     const uint8_t *field_scan_q0;
492     const uint8_t *field_scan8x8_q0;
493     const uint8_t *field_scan8x8_cavlc_q0;
494
495     int x264_build;
496
497     /**
498      * @defgroup multithreading Members for slice based multithreading
499      * @{
500      */
501     struct H264Context *thread_context[MAX_THREADS];
502
503     /**
504      * current slice number, used to initalize slice_num of each thread/context
505      */
506     int current_slice;
507
508     /**
509      * Max number of threads / contexts.
510      * This is equal to AVCodecContext.thread_count unless
511      * multithreaded decoding is impossible, in which case it is
512      * reduced to 1.
513      */
514     int max_contexts;
515
516     /**
517      *  1 if the single thread fallback warning has already been
518      *  displayed, 0 otherwise.
519      */
520     int single_decode_warning;
521
522     int last_slice_type;
523     /** @} */
524
525     int mb_xy;
526
527     uint32_t svq3_watermark_key;
528
529     /**
530      * pic_struct in picture timing SEI message
531      */
532     SEI_PicStructType sei_pic_struct;
533
534     /**
535      * Complement sei_pic_struct
536      * SEI_PIC_STRUCT_TOP_BOTTOM and SEI_PIC_STRUCT_BOTTOM_TOP indicate interlaced frames.
537      * However, soft telecined frames may have these values.
538      * This is used in an attempt to flag soft telecine progressive.
539      */
540     int prev_interlaced_frame;
541
542     /**
543      * Bit set of clock types for fields/frames in picture timing SEI message.
544      * For each found ct_type, appropriate bit is set (e.g., bit 1 for
545      * interlaced).
546      */
547     int sei_ct_type;
548
549     /**
550      * dpb_output_delay in picture timing SEI message, see H.264 C.2.2
551      */
552     int sei_dpb_output_delay;
553
554     /**
555      * cpb_removal_delay in picture timing SEI message, see H.264 C.1.2
556      */
557     int sei_cpb_removal_delay;
558
559     /**
560      * recovery_frame_cnt from SEI message
561      *
562      * Set to -1 if no recovery point SEI message found or to number of frames
563      * before playback synchronizes. Frames having recovery point are key
564      * frames.
565      */
566     int sei_recovery_frame_cnt;
567
568     int is_complex;
569
570     int luma_weight_flag[2];   ///< 7.4.3.2 luma_weight_lX_flag
571     int chroma_weight_flag[2]; ///< 7.4.3.2 chroma_weight_lX_flag
572
573     // Timestamp stuff
574     int sei_buffering_period_present;  ///< Buffering period SEI flag
575     int initial_cpb_removal_delay[32]; ///< Initial timestamps for CPBs
576 }H264Context;
577
578
579 extern const uint8_t ff_h264_chroma_qp[52];
580
581
582 /**
583  * Decode SEI
584  */
585 int ff_h264_decode_sei(H264Context *h);
586
587 /**
588  * Decode SPS
589  */
590 int ff_h264_decode_seq_parameter_set(H264Context *h);
591
592 /**
593  * Decode PPS
594  */
595 int ff_h264_decode_picture_parameter_set(H264Context *h, int bit_length);
596
597 /**
598  * Decodes a network abstraction layer unit.
599  * @param consumed is the number of bytes used as input
600  * @param length is the length of the array
601  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
602  * @returns decoded bytes, might be src+1 if no escapes
603  */
604 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length);
605
606 /**
607  * identifies the exact end of the bitstream
608  * @return the length of the trailing, or 0 if damaged
609  */
610 int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src);
611
612 /**
613  * frees any data that may have been allocated in the H264 context like SPS, PPS etc.
614  */
615 av_cold void ff_h264_free_context(H264Context *h);
616
617 /**
618  * reconstructs bitstream slice_type.
619  */
620 int ff_h264_get_slice_type(const H264Context *h);
621
622 /**
623  * allocates tables.
624  * needs width/height
625  */
626 int ff_h264_alloc_tables(H264Context *h);
627
628 /**
629  * fills the default_ref_list.
630  */
631 int ff_h264_fill_default_ref_list(H264Context *h);
632
633 int ff_h264_decode_ref_pic_list_reordering(H264Context *h);
634 void ff_h264_fill_mbaff_ref_list(H264Context *h);
635 void ff_h264_remove_all_refs(H264Context *h);
636
637 /**
638  * Executes the reference picture marking (memory management control operations).
639  */
640 int ff_h264_execute_ref_pic_marking(H264Context *h, MMCO *mmco, int mmco_count);
641
642 int ff_h264_decode_ref_pic_marking(H264Context *h, GetBitContext *gb);
643
644
645 /**
646  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
647  */
648 int ff_h264_check_intra4x4_pred_mode(H264Context *h);
649
650 /**
651  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
652  */
653 int ff_h264_check_intra_pred_mode(H264Context *h, int mode);
654
655 void ff_h264_write_back_intra_pred_mode(H264Context *h);
656 void ff_h264_hl_decode_mb(H264Context *h);
657 int ff_h264_frame_start(H264Context *h);
658 av_cold int ff_h264_decode_init(AVCodecContext *avctx);
659 av_cold int ff_h264_decode_end(AVCodecContext *avctx);
660 av_cold void ff_h264_decode_init_vlc(void);
661
662 /**
663  * decodes a macroblock
664  * @returns 0 if OK, AC_ERROR / DC_ERROR / MV_ERROR if an error is noticed
665  */
666 int ff_h264_decode_mb_cavlc(H264Context *h);
667
668 /**
669  * decodes a CABAC coded macroblock
670  * @returns 0 if OK, AC_ERROR / DC_ERROR / MV_ERROR if an error is noticed
671  */
672 int ff_h264_decode_mb_cabac(H264Context *h);
673
674 void ff_h264_init_cabac_states(H264Context *h);
675
676 void ff_h264_direct_dist_scale_factor(H264Context * const h);
677 void ff_h264_direct_ref_list_init(H264Context * const h);
678 void ff_h264_pred_direct_motion(H264Context * const h, int *mb_type);
679
680 void ff_h264_filter_mb_fast( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
681 void ff_h264_filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
682
683 /**
684  * Reset SEI values at the beginning of the frame.
685  *
686  * @param h H.264 context.
687  */
688 void ff_h264_reset_sei(H264Context *h);
689
690
691 /*
692 o-o o-o
693  / / /
694 o-o o-o
695  ,---'
696 o-o o-o
697  / / /
698 o-o o-o
699 */
700 //This table must be here because scan8[constant] must be known at compiletime
701 static const uint8_t scan8[16 + 2*4]={
702  4+1*8, 5+1*8, 4+2*8, 5+2*8,
703  6+1*8, 7+1*8, 6+2*8, 7+2*8,
704  4+3*8, 5+3*8, 4+4*8, 5+4*8,
705  6+3*8, 7+3*8, 6+4*8, 7+4*8,
706  1+1*8, 2+1*8,
707  1+2*8, 2+2*8,
708  1+4*8, 2+4*8,
709  1+5*8, 2+5*8,
710 };
711
712 static av_always_inline uint32_t pack16to32(int a, int b){
713 #if HAVE_BIGENDIAN
714    return (b&0xFFFF) + (a<<16);
715 #else
716    return (a&0xFFFF) + (b<<16);
717 #endif
718 }
719
720 /**
721  * gets the chroma qp.
722  */
723 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
724     return h->pps.chroma_qp_table[t][qscale];
725 }
726
727 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my);
728
729 static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deblock){
730     MpegEncContext * const s = &h->s;
731     const int mb_xy= h->mb_xy;
732     int topleft_xy, top_xy, topright_xy, left_xy[2];
733     int topleft_type, top_type, topright_type, left_type[2];
734     const uint8_t * left_block;
735     int topleft_partition= -1;
736     int i;
737     static const uint8_t left_block_options[4][16]={
738         {0,1,2,3,7,10,8,11,7+0*8, 7+1*8, 7+2*8, 7+3*8, 2+0*8, 2+3*8, 2+1*8, 2+2*8},
739         {2,2,3,3,8,11,8,11,7+2*8, 7+2*8, 7+3*8, 7+3*8, 2+1*8, 2+2*8, 2+1*8, 2+2*8},
740         {0,0,1,1,7,10,7,10,7+0*8, 7+0*8, 7+1*8, 7+1*8, 2+0*8, 2+3*8, 2+0*8, 2+3*8},
741         {0,2,0,2,7,10,7,10,7+0*8, 7+2*8, 7+0*8, 7+2*8, 2+0*8, 2+3*8, 2+0*8, 2+3*8}
742     };
743
744     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
745
746     //FIXME deblocking could skip the intra and nnz parts.
747 //     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
748 //         return;
749
750     /* Wow, what a mess, why didn't they simplify the interlacing & intra
751      * stuff, I can't imagine that these complex rules are worth it. */
752
753     topleft_xy = top_xy - 1;
754     topright_xy= top_xy + 1;
755     left_xy[1] = left_xy[0] = mb_xy-1;
756     left_block = left_block_options[0];
757     if(FRAME_MBAFF){
758         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
759         const int top_pair_xy      = pair_xy     - s->mb_stride;
760         const int topleft_pair_xy  = top_pair_xy - 1;
761         const int topright_pair_xy = top_pair_xy + 1;
762         const int topleft_mb_field_flag  = IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
763         const int top_mb_field_flag      = IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
764         const int topright_mb_field_flag = IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
765         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
766         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
767         const int bottom = (s->mb_y & 1);
768         tprintf(s->avctx, "fill_caches: curr_mb_field_flag:%d, left_mb_field_flag:%d, topleft_mb_field_flag:%d, top_mb_field_flag:%d, topright_mb_field_flag:%d\n", curr_mb_field_flag, left_mb_field_flag, topleft_mb_field_flag, top_mb_field_flag, topright_mb_field_flag);
769
770         if (curr_mb_field_flag && (bottom || top_mb_field_flag)){
771             top_xy -= s->mb_stride;
772         }
773         if (curr_mb_field_flag && (bottom || topleft_mb_field_flag)){
774             topleft_xy -= s->mb_stride;
775         } else if(bottom && !curr_mb_field_flag && left_mb_field_flag) {
776             topleft_xy += s->mb_stride;
777             // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
778             topleft_partition = 0;
779         }
780         if (curr_mb_field_flag && (bottom || topright_mb_field_flag)){
781             topright_xy -= s->mb_stride;
782         }
783         if (left_mb_field_flag != curr_mb_field_flag) {
784             left_xy[1] = left_xy[0] = pair_xy - 1;
785             if (curr_mb_field_flag) {
786                 left_xy[1] += s->mb_stride;
787                 left_block = left_block_options[3];
788             } else {
789                 left_block= left_block_options[2 - bottom];
790             }
791         }
792     }
793
794     h->top_mb_xy = top_xy;
795     h->left_mb_xy[0] = left_xy[0];
796     h->left_mb_xy[1] = left_xy[1];
797     if(for_deblock){
798
799         //for sufficiently low qp, filtering wouldn't do anything
800         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
801         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
802         int qp = s->current_picture.qscale_table[mb_xy];
803         if(qp <= qp_thresh
804            && (left_xy[0]<0 || ((qp + s->current_picture.qscale_table[left_xy[0]] + 1)>>1) <= qp_thresh)
805            && (left_xy[1]<0 || ((qp + s->current_picture.qscale_table[left_xy[1]] + 1)>>1) <= qp_thresh)
806            && (top_xy   < 0 || ((qp + s->current_picture.qscale_table[top_xy ] + 1)>>1) <= qp_thresh)){
807             return 1;
808         }
809         if(IS_INTRA(mb_type))
810             return 0;
811
812         AV_COPY64(&h->non_zero_count_cache[0+8*1], &h->non_zero_count[mb_xy][ 0]);
813         AV_COPY64(&h->non_zero_count_cache[0+8*2], &h->non_zero_count[mb_xy][ 8]);
814         *((uint32_t*)&h->non_zero_count_cache[0+8*5])= *((uint32_t*)&h->non_zero_count[mb_xy][16]);
815         *((uint32_t*)&h->non_zero_count_cache[4+8*3])= *((uint32_t*)&h->non_zero_count[mb_xy][20]);
816         AV_COPY64(&h->non_zero_count_cache[0+8*4], &h->non_zero_count[mb_xy][24]);
817
818         h->cbp= h->cbp_table[mb_xy];
819
820         top_type     = h->slice_table[top_xy     ] < 0xFFFF ? s->current_picture.mb_type[top_xy]     : 0;
821         left_type[0] = h->slice_table[left_xy[0] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[0]] : 0;
822         left_type[1] = h->slice_table[left_xy[1] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[1]] : 0;
823
824         {
825             int list;
826             for(list=0; list<h->list_count; list++){
827                 int8_t *ref;
828                 int y, b_stride;
829                 int16_t (*mv_dst)[2];
830                 int16_t (*mv_src)[2];
831
832                 if(!USES_LIST(mb_type, list)){
833                     fill_rectangle(  h->mv_cache[list][scan8[0]], 4, 4, 8, pack16to32(0,0), 4);
834                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
835                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] =
836                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
837                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = ((LIST_NOT_USED)&0xFF)*0x01010101;
838                     continue;
839                 }
840
841                 ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
842                 {
843                     int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
844                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
845                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
846                     ref += h->b8_stride;
847                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
848                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
849                 }
850
851                 b_stride = h->b_stride;
852                 mv_dst   = &h->mv_cache[list][scan8[0]];
853                 mv_src   = &s->current_picture.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
854                 for(y=0; y<4; y++){
855                     AV_COPY128(mv_dst + 8*y, mv_src + y*b_stride);
856                 }
857
858             }
859         }
860     }else{
861         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
862         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
863         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
864         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
865         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
866
867         if(IS_INTRA(mb_type)){
868             int type_mask= h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
869             h->topleft_samples_available=
870             h->top_samples_available=
871             h->left_samples_available= 0xFFFF;
872             h->topright_samples_available= 0xEEEA;
873
874             if(!(top_type & type_mask)){
875                 h->topleft_samples_available= 0xB3FF;
876                 h->top_samples_available= 0x33FF;
877                 h->topright_samples_available= 0x26EA;
878             }
879             if(IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[0])){
880                 if(IS_INTERLACED(mb_type)){
881                     if(!(left_type[0] & type_mask)){
882                         h->topleft_samples_available&= 0xDFFF;
883                         h->left_samples_available&= 0x5FFF;
884                     }
885                     if(!(left_type[1] & type_mask)){
886                         h->topleft_samples_available&= 0xFF5F;
887                         h->left_samples_available&= 0xFF5F;
888                     }
889                 }else{
890                     int left_typei = h->slice_table[left_xy[0] + s->mb_stride ] == h->slice_num
891                                     ? s->current_picture.mb_type[left_xy[0] + s->mb_stride] : 0;
892                     assert(left_xy[0] == left_xy[1]);
893                     if(!((left_typei & type_mask) && (left_type[0] & type_mask))){
894                         h->topleft_samples_available&= 0xDF5F;
895                         h->left_samples_available&= 0x5F5F;
896                     }
897                 }
898             }else{
899                 if(!(left_type[0] & type_mask)){
900                     h->topleft_samples_available&= 0xDF5F;
901                     h->left_samples_available&= 0x5F5F;
902                 }
903             }
904
905             if(!(topleft_type & type_mask))
906                 h->topleft_samples_available&= 0x7FFF;
907
908             if(!(topright_type & type_mask))
909                 h->topright_samples_available&= 0xFBFF;
910
911             if(IS_INTRA4x4(mb_type)){
912                 if(IS_INTRA4x4(top_type)){
913                     h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
914                     h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
915                     h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
916                     h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
917                 }else{
918                     int pred;
919                     if(!(top_type & type_mask))
920                         pred= -1;
921                     else{
922                         pred= 2;
923                     }
924                     h->intra4x4_pred_mode_cache[4+8*0]=
925                     h->intra4x4_pred_mode_cache[5+8*0]=
926                     h->intra4x4_pred_mode_cache[6+8*0]=
927                     h->intra4x4_pred_mode_cache[7+8*0]= pred;
928                 }
929                 for(i=0; i<2; i++){
930                     if(IS_INTRA4x4(left_type[i])){
931                         h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
932                         h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
933                     }else{
934                         int pred;
935                         if(!(left_type[i] & type_mask))
936                             pred= -1;
937                         else{
938                             pred= 2;
939                         }
940                         h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
941                         h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
942                     }
943                 }
944             }
945         }
946     }
947
948
949 /*
950 0 . T T. T T T T
951 1 L . .L . . . .
952 2 L . .L . . . .
953 3 . T TL . . . .
954 4 L . .L . . . .
955 5 L . .. . . . .
956 */
957 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
958     if(top_type){
959         *(uint32_t*)&h->non_zero_count_cache[4+8*0]= *(uint32_t*)&h->non_zero_count[top_xy][4+3*8];
960         if(!for_deblock){
961             h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][1+1*8];
962             h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][2+1*8];
963
964             h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][1+2*8];
965             h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][2+2*8];
966         }
967     }else if(!for_deblock){
968             h->non_zero_count_cache[1+8*0]=
969             h->non_zero_count_cache[2+8*0]=
970
971             h->non_zero_count_cache[1+8*3]=
972             h->non_zero_count_cache[2+8*3]=
973             *(uint32_t*)&h->non_zero_count_cache[4+8*0]= CABAC && !IS_INTRA(mb_type) ? 0 : 0x40404040;
974     }
975
976     for (i=0; i<2; i++) {
977         if(left_type[i]){
978             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[8+0+2*i]];
979             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[8+1+2*i]];
980             if(!for_deblock){
981                 h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[8+4+2*i]];
982                 h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[8+5+2*i]];
983             }
984         }else if(!for_deblock){
985                 h->non_zero_count_cache[3+8*1 + 2*8*i]=
986                 h->non_zero_count_cache[3+8*2 + 2*8*i]=
987                 h->non_zero_count_cache[0+8*1 +   8*i]=
988                 h->non_zero_count_cache[0+8*4 +   8*i]= CABAC && !IS_INTRA(mb_type) ? 0 : 64;
989         }
990     }
991
992     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
993     if(for_deblock && !CABAC && h->pps.transform_8x8_mode){
994         if(IS_8x8DCT(top_type)){
995             h->non_zero_count_cache[4+8*0]=
996             h->non_zero_count_cache[5+8*0]= h->cbp_table[top_xy] & 4;
997             h->non_zero_count_cache[6+8*0]=
998             h->non_zero_count_cache[7+8*0]= h->cbp_table[top_xy] & 8;
999         }
1000         if(IS_8x8DCT(left_type[0])){
1001             h->non_zero_count_cache[3+8*1]=
1002             h->non_zero_count_cache[3+8*2]= h->cbp_table[left_xy[0]]&2; //FIXME check MBAFF
1003         }
1004         if(IS_8x8DCT(left_type[1])){
1005             h->non_zero_count_cache[3+8*3]=
1006             h->non_zero_count_cache[3+8*4]= h->cbp_table[left_xy[1]]&8; //FIXME check MBAFF
1007         }
1008
1009         if(IS_8x8DCT(mb_type)){
1010             h->non_zero_count_cache[scan8[0   ]]= h->non_zero_count_cache[scan8[1   ]]=
1011             h->non_zero_count_cache[scan8[2   ]]= h->non_zero_count_cache[scan8[3   ]]= h->cbp & 1;
1012
1013             h->non_zero_count_cache[scan8[0+ 4]]= h->non_zero_count_cache[scan8[1+ 4]]=
1014             h->non_zero_count_cache[scan8[2+ 4]]= h->non_zero_count_cache[scan8[3+ 4]]= h->cbp & 2;
1015
1016             h->non_zero_count_cache[scan8[0+ 8]]= h->non_zero_count_cache[scan8[1+ 8]]=
1017             h->non_zero_count_cache[scan8[2+ 8]]= h->non_zero_count_cache[scan8[3+ 8]]= h->cbp & 4;
1018
1019             h->non_zero_count_cache[scan8[0+12]]= h->non_zero_count_cache[scan8[1+12]]=
1020             h->non_zero_count_cache[scan8[2+12]]= h->non_zero_count_cache[scan8[3+12]]= h->cbp & 8;
1021         }
1022     }
1023
1024     if( CABAC && !for_deblock) {
1025         // top_cbp
1026         if(top_type) {
1027             h->top_cbp = h->cbp_table[top_xy];
1028         } else if(IS_INTRA(mb_type)) {
1029             h->top_cbp = 0x1C0;
1030         } else {
1031             h->top_cbp = 0;
1032         }
1033         // left_cbp
1034         if (left_type[0]) {
1035             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
1036         } else if(IS_INTRA(mb_type)) {
1037             h->left_cbp = 0x1C0;
1038         } else {
1039             h->left_cbp = 0;
1040         }
1041         if (left_type[0]) {
1042             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
1043         }
1044         if (left_type[1]) {
1045             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
1046         }
1047     }
1048
1049 #if 1
1050     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
1051         int list;
1052         for(list=0; list<h->list_count; list++){
1053             if(!for_deblock && !USES_LIST(mb_type, list) && !IS_DIRECT(mb_type)){
1054                 /*if(!h->mv_cache_clean[list]){
1055                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
1056                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
1057                     h->mv_cache_clean[list]= 1;
1058                 }*/
1059                 continue;
1060             }
1061             h->mv_cache_clean[list]= 0;
1062
1063             if(USES_LIST(top_type, list)){
1064                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
1065                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
1066                 AV_COPY128(h->mv_cache[list][scan8[0] + 0 - 1*8], s->current_picture.motion_val[list][b_xy + 0]);
1067                 if(for_deblock){
1068                     int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
1069                     h->ref_cache[list][scan8[0] + 0 - 1*8]=
1070                     h->ref_cache[list][scan8[0] + 1 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 0]];
1071                     h->ref_cache[list][scan8[0] + 2 - 1*8]=
1072                     h->ref_cache[list][scan8[0] + 3 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 1]];
1073                 }else{
1074                     h->ref_cache[list][scan8[0] + 0 - 1*8]=
1075                     h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
1076                     h->ref_cache[list][scan8[0] + 2 - 1*8]=
1077                     h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
1078                 }
1079             }else{
1080                 AV_ZERO128(h->mv_cache[list][scan8[0] + 0 - 1*8]);
1081                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= (((for_deblock||top_type) ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
1082             }
1083
1084             for(i=0; i<2; i++){
1085                 int cache_idx = scan8[0] - 1 + i*2*8;
1086                 if(USES_LIST(left_type[i], list)){
1087                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
1088                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
1089                     *(uint32_t*)h->mv_cache[list][cache_idx  ]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[0+i*2]];
1090                     *(uint32_t*)h->mv_cache[list][cache_idx+8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[1+i*2]];
1091                     if(for_deblock){
1092                         int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[i]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
1093                         h->ref_cache[list][cache_idx  ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)]];
1094                         h->ref_cache[list][cache_idx+8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)]];
1095                     }else{
1096                         h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
1097                         h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
1098                     }
1099                 }else{
1100                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
1101                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
1102                     h->ref_cache[list][cache_idx  ]=
1103                     h->ref_cache[list][cache_idx+8]= (for_deblock||left_type[i]) ? LIST_NOT_USED : PART_NOT_AVAILABLE;
1104                 }
1105             }
1106
1107             if(for_deblock || ((IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred) && !FRAME_MBAFF))
1108                 continue;
1109
1110             if(USES_LIST(topleft_type, list)){
1111                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
1112                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
1113                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
1114                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
1115             }else{
1116                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
1117                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
1118             }
1119
1120             if(USES_LIST(topright_type, list)){
1121                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
1122                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
1123                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
1124                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
1125             }else{
1126                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
1127                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
1128             }
1129
1130             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
1131                 continue;
1132
1133             h->ref_cache[list][scan8[5 ]+1] =
1134             h->ref_cache[list][scan8[7 ]+1] =
1135             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
1136             h->ref_cache[list][scan8[4 ]] =
1137             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
1138             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
1139             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
1140             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
1141             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
1142             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
1143
1144             if( CABAC ) {
1145                 /* XXX beurk, Load mvd */
1146                 if(USES_LIST(top_type, list)){
1147                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
1148                     AV_COPY128(h->mvd_cache[list][scan8[0] + 0 - 1*8], h->mvd_table[list][b_xy + 0]);
1149                 }else{
1150                     AV_ZERO128(h->mvd_cache[list][scan8[0] + 0 - 1*8]);
1151                 }
1152                 if(USES_LIST(left_type[0], list)){
1153                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
1154                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 0*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[0]];
1155                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[1]];
1156                 }else{
1157                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
1158                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
1159                 }
1160                 if(USES_LIST(left_type[1], list)){
1161                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
1162                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 2*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[2]];
1163                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 3*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[3]];
1164                 }else{
1165                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
1166                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
1167                 }
1168                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
1169                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
1170                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
1171                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
1172                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
1173
1174                 if(h->slice_type_nos == FF_B_TYPE){
1175                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
1176
1177                     if(IS_DIRECT(top_type)){
1178                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
1179                     }else if(IS_8X8(top_type)){
1180                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
1181                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
1182                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
1183                     }else{
1184                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
1185                     }
1186
1187                     if(IS_DIRECT(left_type[0]))
1188                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
1189                     else if(IS_8X8(left_type[0]))
1190                         h->direct_cache[scan8[0] - 1 + 0*8]= h->direct_table[h->mb2b8_xy[left_xy[0]] + 1 + h->b8_stride*(left_block[0]>>1)];
1191                     else
1192                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
1193
1194                     if(IS_DIRECT(left_type[1]))
1195                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
1196                     else if(IS_8X8(left_type[1]))
1197                         h->direct_cache[scan8[0] - 1 + 2*8]= h->direct_table[h->mb2b8_xy[left_xy[1]] + 1 + h->b8_stride*(left_block[2]>>1)];
1198                     else
1199                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
1200                 }
1201             }
1202
1203             if(FRAME_MBAFF){
1204 #define MAP_MVS\
1205                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
1206                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
1207                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
1208                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
1209                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
1210                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
1211                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
1212                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
1213                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
1214                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
1215                 if(MB_FIELD){
1216 #define MAP_F2F(idx, mb_type)\
1217                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
1218                         h->ref_cache[list][idx] <<= 1;\
1219                         h->mv_cache[list][idx][1] /= 2;\
1220                         h->mvd_cache[list][idx][1] /= 2;\
1221                     }
1222                     MAP_MVS
1223 #undef MAP_F2F
1224                 }else{
1225 #define MAP_F2F(idx, mb_type)\
1226                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
1227                         h->ref_cache[list][idx] >>= 1;\
1228                         h->mv_cache[list][idx][1] <<= 1;\
1229                         h->mvd_cache[list][idx][1] <<= 1;\
1230                     }
1231                     MAP_MVS
1232 #undef MAP_F2F
1233                 }
1234             }
1235         }
1236     }
1237 #endif
1238
1239     if(!for_deblock)
1240         h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
1241     return 0;
1242 }
1243
1244 static void fill_decode_caches(H264Context *h, int mb_type){
1245     fill_caches(h, mb_type, 0);
1246 }
1247
1248 /**
1249  *
1250  * @returns non zero if the loop filter can be skiped
1251  */
1252 static int fill_filter_caches(H264Context *h, int mb_type){
1253     return fill_caches(h, mb_type, 1);
1254 }
1255
1256 /**
1257  * gets the predicted intra4x4 prediction mode.
1258  */
1259 static inline int pred_intra_mode(H264Context *h, int n){
1260     const int index8= scan8[n];
1261     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
1262     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
1263     const int min= FFMIN(left, top);
1264
1265     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
1266
1267     if(min<0) return DC_PRED;
1268     else      return min;
1269 }
1270
1271 static inline void write_back_non_zero_count(H264Context *h){
1272     const int mb_xy= h->mb_xy;
1273
1274     AV_COPY64(&h->non_zero_count[mb_xy][ 0], &h->non_zero_count_cache[0+8*1]);
1275     AV_COPY64(&h->non_zero_count[mb_xy][ 8], &h->non_zero_count_cache[0+8*2]);
1276     *((uint32_t*)&h->non_zero_count[mb_xy][16]) = *((uint32_t*)&h->non_zero_count_cache[0+8*5]);
1277     *((uint32_t*)&h->non_zero_count[mb_xy][20]) = *((uint32_t*)&h->non_zero_count_cache[4+8*3]);
1278     AV_COPY64(&h->non_zero_count[mb_xy][24], &h->non_zero_count_cache[0+8*4]);
1279 }
1280
1281 static inline void write_back_motion(H264Context *h, int mb_type){
1282     MpegEncContext * const s = &h->s;
1283     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1284     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1285     int list;
1286
1287     if(!USES_LIST(mb_type, 0))
1288         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1289
1290     for(list=0; list<h->list_count; list++){
1291         int y, b_stride;
1292         int16_t (*mv_dst)[2];
1293         int16_t (*mv_src)[2];
1294
1295         if(!USES_LIST(mb_type, list))
1296             continue;
1297
1298         b_stride = h->b_stride;
1299         mv_dst   = &s->current_picture.motion_val[list][b_xy];
1300         mv_src   = &h->mv_cache[list][scan8[0]];
1301         for(y=0; y<4; y++){
1302             AV_COPY128(mv_dst + y*b_stride, mv_src + 8*y);
1303         }
1304         if( CABAC ) {
1305             int16_t (*mvd_dst)[2] = &h->mvd_table[list][b_xy];
1306             int16_t (*mvd_src)[2] = &h->mvd_cache[list][scan8[0]];
1307             if(IS_SKIP(mb_type))
1308                 fill_rectangle(mvd_dst, 4, 4, h->b_stride, 0, 4);
1309             else
1310             for(y=0; y<4; y++){
1311                 AV_COPY128(mvd_dst + y*b_stride, mvd_src + 8*y);
1312             }
1313         }
1314
1315         {
1316             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1317             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1318             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1319             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1320             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1321         }
1322     }
1323
1324     if(h->slice_type_nos == FF_B_TYPE && CABAC){
1325         if(IS_8X8(mb_type)){
1326             uint8_t *direct_table = &h->direct_table[b8_xy];
1327             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1328             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1329             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1330         }
1331     }
1332 }
1333
1334 static inline int get_dct8x8_allowed(H264Context *h){
1335     if(h->sps.direct_8x8_inference_flag)
1336         return !(*(uint64_t*)h->sub_mb_type & ((MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_8x8                )*0x0001000100010001ULL));
1337     else
1338         return !(*(uint64_t*)h->sub_mb_type & ((MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_8x8|MB_TYPE_DIRECT2)*0x0001000100010001ULL));
1339 }
1340
1341 static void predict_field_decoding_flag(H264Context *h){
1342     MpegEncContext * const s = &h->s;
1343     const int mb_xy= h->mb_xy;
1344     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
1345                 ? s->current_picture.mb_type[mb_xy-1]
1346                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
1347                 ? s->current_picture.mb_type[mb_xy-s->mb_stride]
1348                 : 0;
1349     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
1350 }
1351
1352 /**
1353  * decodes a P_SKIP or B_SKIP macroblock
1354  */
1355 static void decode_mb_skip(H264Context *h){
1356     MpegEncContext * const s = &h->s;
1357     const int mb_xy= h->mb_xy;
1358     int mb_type=0;
1359
1360     memset(h->non_zero_count[mb_xy], 0, 32);
1361     memset(h->non_zero_count_cache + 8, 0, 8*5); //FIXME ugly, remove pfui
1362
1363     if(MB_FIELD)
1364         mb_type|= MB_TYPE_INTERLACED;
1365
1366     if( h->slice_type_nos == FF_B_TYPE )
1367     {
1368         // just for fill_caches. pred_direct_motion will set the real mb_type
1369         mb_type|= MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2|MB_TYPE_SKIP;
1370
1371         fill_decode_caches(h, mb_type); //FIXME check what is needed and what not ...
1372         ff_h264_pred_direct_motion(h, &mb_type);
1373         mb_type|= MB_TYPE_SKIP;
1374     }
1375     else
1376     {
1377         int mx, my;
1378         mb_type|= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P1L0|MB_TYPE_SKIP;
1379
1380         fill_decode_caches(h, mb_type); //FIXME check what is needed and what not ...
1381         pred_pskip_motion(h, &mx, &my);
1382         fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, 0, 1);
1383         fill_rectangle(  h->mv_cache[0][scan8[0]], 4, 4, 8, pack16to32(mx,my), 4);
1384     }
1385
1386     write_back_motion(h, mb_type);
1387     s->current_picture.mb_type[mb_xy]= mb_type;
1388     s->current_picture.qscale_table[mb_xy]= s->qscale;
1389     h->slice_table[ mb_xy ]= h->slice_num;
1390     h->prev_mb_skipped= 1;
1391 }
1392
1393 #include "h264_mvpred.h" //For pred_pskip_motion()
1394
1395 #endif /* AVCODEC_H264_H */