theora-exp/lib/internal.h

   1 #if !defined(_internal_H)
   2 # define _internal_H (1)
   3 # include <stdlib.h>
   4 # if defined(HAVE_CONFIG_H)
   5 #  include <config.h>
   6 # endif
   7 # include "theora/codec.h"
   8 # include "ocintrin.h"
   9 # include "huffman.h"
  10 # include "quant.h"
  11
  12 /*Thank you Microsoft, I know the order of operations.*/
  13 # if defined(_MSCVER)
  14 #  pragma warning(disable:4554)
  15 # endif
  16
  17 /*This library's version.*/
  18 # define OC_VENDOR_STRING "derf's experimental encoder library " __DATE__
  19
  20 /*Theora bitstream version.*/
  21 # define TH_VERSION_MAJOR (3)
  22 # define TH_VERSION_MINOR (2)
  23 # define TH_VERSION_SUB   (0)
  24
  25 /*A keyframe.*/
  26 #define OC_INTRA_FRAME (0)
  27 /*A predicted frame.*/
  28 #define OC_INTER_FRAME (1)
  29 /*A frame of unknown type (frame type decision has not yet been made).*/
  30 #define OC_UNKWN_FRAME (-1)
  31
  32 /*The amount of padding to add to the reconstructed frame buffers on all
  33    sides.
  34   This is used to allow unrestricted motion vectors without special casing.
  35   This must be a multiple of 2.*/
  36 #define OC_UMV_PADDING (16)
  37
  38 /*Frame classification indices.*/
  39 /*The previous golden frame.*/
  40 #define OC_FRAME_GOLD (0)
  41 /*The previous frame.*/
  42 #define OC_FRAME_PREV (1)
  43 /*The current frame.*/
  44 #define OC_FRAME_SELF (2)
  45
  46 /*The input or output buffer.*/
  47 #define OC_FRAME_IO   (3)
  48
  49 /*Macroblock modes.*/
  50 /*Macro block is invalid: It is never coded.*/
  51 #define OC_MODE_INVALID        (-1)
  52
  53 /*Encoded with no motion compensated prediction.*/
  54 #define OC_MODE_INTRA          (0)
  55 /*Encoded difference from the same macro block in the previous frame.*/
  56 #define OC_MODE_INTER_NOMV     (1)
  57 /*Encoded difference from the previous frame offset by the given motion
  58    vector.*/
  59 #define OC_MODE_INTER_MV       (2)
  60 /*Encoded difference from the previous frame offset by the last coded motion
  61    vector.*/
  62 #define OC_MODE_INTER_MV_LAST  (3)
  63 /*Encoded difference from the previous frame offset by the second to last
  64    coded motion vector.*/
  65 #define OC_MODE_INTER_MV_LAST2 (4)
  66 /*Encoded difference from the previous frame offset by the individual motion
  67    vectors given for each block.*/
  68 #define OC_MODE_INTER_MV_FOUR  (5)
  69 /*Encoded difference from the same macro block in the previous golden frame.*/
  70 #define OC_MODE_GOLDEN_NOMV    (6)
  71 /*Encoded difference from the previous golden frame offset by the given motion
  72    vector.*/
  73 #define OC_MODE_GOLDEN_MV      (7)
  74 /*The number of (coded) modes.*/
  75 #define OC_NMODES              (8)
  76
  77 /*Macro block is not coded.*/
  78 #define OC_MODE_NOT_CODED      (8)
  79
  80 /*Predictor bit flags.*/
  81 /*Left.*/
  82 #define OC_PL  (1)
  83 /*Upper-left.*/
  84 #define OC_PUL (2)
  85 /*Up.*/
  86 #define OC_PU  (4)
  87 /*Upper-right.*/
  88 #define OC_PUR (8)
  89
  90 /*Constants for the packet state machine common between encoder and decoder.*/
  91
  92 /*Next packet to emit/read: Codec info header.*/
  93 #define OC_PACKET_INFO_HDR    (-3)
  94 /*Next packet to emit/read: Comment header.*/
  95 #define OC_PACKET_COMMENT_HDR (-2)
  96 /*Next packet to emit/read: Codec setup header.*/
  97 #define OC_PACKET_SETUP_HDR   (-1)
  98 /*No more packets to emit/read.*/
  99 #define OC_PACKET_DONE        (INT_MAX)
 100
 101
 102
 103 typedef struct oc_theora_state oc_theora_state;
 104
 105
 106
 107 /*A map from a super block to fragment numbers.*/
 108 typedef int oc_sb_map[4][4];
 109 /*A map from a macro block to fragment numbers.*/
 110 typedef int oc_mb_map[3][4];
 111 /*A motion vector.*/
 112 typedef char oc_mv[2];
 113
 114
 115
 116 /*Super block information.
 117   Super blocks are 32x32 segments of pixels in a single color plane indexed
 118    in image order.
 119   Internally, super blocks are broken up into four quadrants, each of which
 120    contains a 2x2 pattern of blocks, each of which is an 8x8 block of pixels.
 121   Quadrants, and the blocks within them, are indexed in a special order called
 122    a "Hilbert curve" within the super block.
 123
 124   In order to differentiate between the Hilbert-curve indexing strategy and
 125    the regular image order indexing strategy, blocks indexed in image order
 126    are called "fragments".
 127   Fragments are indexed in image order, left to right, then bottom to top,
 128    from Y plane to Cb plane to Cr plane.*/
 129 typedef struct{
 130   unsigned  coded_fully:1;
 131   unsigned  coded_partially:1;
 132   unsigned  quad_valid:4;
 133   oc_sb_map map;
 134 }oc_sb;
 135
 136
 137
 138 /*Macro block information.
 139   The co-located fragments in all image planes corresponding to the location of
 140    a single luma plane super block quadrant forms a macro block.
 141   Thus there is only a single set of macro blocks for all planes, which
 142    contains between 6 and 12 fragments, depending on the pixel format.
 143   Therefore macro block information is kept in a separate array from super
 144    blocks, to avoid unused space in the other planes.*/
 145 typedef struct{
 146   /*The current macro block mode.
 147     A negative number indicates the macro block lies entirely outside the
 148      coded frame.*/
 149   int           mode;
 150   /*The X location of the macro block's upper-left hand pixel.*/
 151   int           x;
 152   /*The Y location of the macro block's upper-right hand pixel.*/
 153   int           y;
 154   /*The fragments that belong to this macro block in each color plane.
 155     Fragments are stored in image order (left to right then top to bottom).
 156     When chroma components are decimated, the extra fragments have an index of
 157      -1.*/
 158   oc_mb_map     map;
 159 }oc_mb;
 160
 161
 162
 163 /*Information about a fragment which intersects the border of the displayable
 164    region.
 165   This marks which pixels belong to the displayable region, and is used to
 166    ensure that pixels outside of this region are never referenced.
 167   This allows applications to pass in buffers that are really the size of the
 168    displayable region without causing a seg fault.*/
 169 typedef struct{
 170   /*A bit mask marking which pixels are in the displayable region.
 171     Pixel (x,y) corresponds to bit (y<<3|x).*/
 172   ogg_int64_t mask;
 173   /*The number of pixels in the displayable region.
 174     This is always positive, and always less than 64.*/
 175   int         npixels;
 176 }oc_border_info;
 177
 178
 179
 180 /*Fragment information.*/
 181 typedef struct{
 182   /*A flag indicating whether or not this fragment is coded.*/
 183   unsigned        coded:1;
 184   /*A flag indicating that all of this fragment lies outside the displayable
 185      region of the frame.
 186     Note the contrast with an invalid macro block, which is outside the coded
 187      frame, not just the displayable one.*/
 188   unsigned        invalid:1;
 189   /*The quality index used for this fragment's AC coefficients.*/
 190   unsigned        qi:6;
 191   /*The mode of the macroblock this fragment belongs to.*/
 192   int             mbmode:8;
 193   /*The prediction-corrected DC component.*/
 194   int             dc:16;
 195   /*A pointer to the portion of an image covered by this fragment in several
 196      images.
 197     The first three are reconstructed frame buffers, while the last is the
 198      input image buffer.
 199     The appropriate stride value is determined by the color plane the fragment
 200      belongs in.*/
 201   unsigned char  *buffer[4];
 202   /*Information for fragments which lie partially outside the displayable
 203      region.
 204     For fragments completely inside or outside this region, this is NULL.*/
 205   oc_border_info *border;
 206   /*The motion vector used for this fragment.*/
 207   oc_mv           mv;
 208 }oc_fragment;
 209
 210
 211
 212 /*A description of each fragment plane.*/
 213 typedef struct{
 214   /*The number of fragments in the horizontal direction.*/
 215   int nhfrags;
 216   /*The number of fragments in the vertical direction.*/
 217   int nvfrags;
 218   /*The offset of the first fragment in the plane.*/
 219   int froffset;
 220   /*The total number of fragments in the plane.*/
 221   int nfrags;
 222   /*The number of super blocks in the horizontal direction.*/
 223   int nhsbs;
 224   /*The number of super blocks in the vertical direction.*/
 225   int nvsbs;
 226   /*The offset of the first super block in the plane.*/
 227   int sboffset;
 228   /*The total number of super blocks in the plane.*/
 229   int nsbs;
 230 }oc_fragment_plane;
 231
 232
 233
 234 /*The shared (encoder and decoder) functions that have accelerated variants.*/
 235 typedef struct{
 236   void (*frag_recon_intra)(unsigned char *_dst,int _dst_ystride,
 237    const ogg_int16_t *_residue);
 238   void (*frag_recon_inter)(unsigned char *_dst,int _dst_ystride,
 239    const unsigned char *_src,int _src_ystride,const ogg_int16_t *_residue);
 240   void (*frag_recon_inter2)(unsigned char *_dst,int _dst_ystride,
 241    const unsigned char *_src1,int _src1_ystride,const unsigned char *_src2,
 242    int _src2_ystride,const ogg_int16_t *_residue);
 243   void (*state_frag_copy)(const oc_theora_state *_state,
 244    const int *_fragis,int _nfragis,int _dst_frame,int _src_frame,int _pli);
 245   void (*state_frag_recon)(oc_theora_state *_state,const oc_fragment *_frag,
 246    int _pli,ogg_int16_t _dct_coeffs[128],int _last_zzi,int _ncoefs,
 247    ogg_uint16_t _dc_iquant,const ogg_uint16_t _ac_iquant[64]);
 248   void (*state_loop_filter_frag_rows)(oc_theora_state *_state,int *_bv,
 249    int _refi,int _pli,int _fragy0,int _fragy_end);
 250   void (*restore_fpu)(void);
 251 }oc_base_opt_vtable;
 252
 253
 254
 255 /*Common state information between the encoder and decoder.*/
 256 struct oc_theora_state{
 257   /*The stream information.*/
 258   th_info               info;
 259   /*Table for shared accelerated functions.*/
 260   oc_base_opt_vtable    opt_vtable;
 261   /*CPU flags to detect the presence of extended instruction sets.*/
 262   ogg_uint32_t          cpu_flags;
 263   /*The fragment plane descriptions.*/
 264   oc_fragment_plane     fplanes[3];
 265   /*The total number of fragments in a single frame.*/
 266   int                   nfrags;
 267   /*The list of fragments, indexed in image order.*/
 268   oc_fragment          *frags;
 269   /*The total number of super blocks in a single frame.*/
 270   int                   nsbs;
 271   /*The list of super blocks, indexed in image order.*/
 272   oc_sb                *sbs;
 273   /*The number of macro blocks in the X direction.*/
 274   int                   nhmbs;
 275   /*The number of macro blocks in the Y direction.*/
 276   int                   nvmbs;
 277   /*The total number of macro blocks.*/
 278   int                   nmbs;
 279   /*The list of macro blocks, indexed in super block order.
 280     That is, the macro block corresponding to the macro block mbi in (luma
 281      plane) super block sbi is (sbi<<2|mbi).*/
 282   oc_mb                *mbs;
 283   /*The list of coded fragments, in coded order.*/
 284   int                  *coded_fragis;
 285   /*The number of coded fragments in each plane.*/
 286   int                   ncoded_fragis[3];
 287   /*The list of uncoded fragments.
 288     This just past the end of the list, which is in reverse order, and
 289      uses the same block of allocated storage as the coded_fragis list.*/
 290   int                  *uncoded_fragis;
 291   /*The number of uncoded fragments in each plane.*/
 292   int                   nuncoded_fragis[3];
 293   /*The list of coded macro blocks in the Y plane, in coded order.*/
 294   int                  *coded_mbis;
 295   /*The number of coded macro blocks in the Y plane.*/
 296   int                   ncoded_mbis;
 297   /*A copy of the image data used to fill the input pointers in each fragment.
 298     If the data pointers or strides change, these input pointers must be
 299      re-populated.*/
 300   th_ycbcr_buffer       input;
 301   /*The number of unique border patterns.*/
 302   int                   nborders;
 303   /*The storage for the border info for all border fragments.
 304     This data is pointed to from the appropriate fragments.*/
 305   oc_border_info        borders[16];
 306   /*The index of the buffers being used for each OC_FRAME_* reference frame.*/
 307   int                   ref_frame_idx[3];
 308   /*The actual buffers used for the previously decoded frames.*/
 309   th_ycbcr_buffer       ref_frame_bufs[3];
 310   /*The storage for the reference frame buffers.*/
 311   unsigned char        *ref_frame_data;
 312   /*The frame number of the last keyframe.*/
 313   ogg_int64_t           keyframe_num;
 314   /*The frame number of the current frame.*/
 315   ogg_int64_t           curframe_num;
 316   /*The granpos of the current frame.*/
 317   ogg_int64_t           granpos;
 318   /*The type of the current frame.*/
 319   int                   frame_type;
 320   /*The quality indices of the current frame.*/
 321   int                   qis[3];
 322   /*The number of quality indices used in the current frame.*/
 323   int                   nqis;
 324   /*The dequantization tables.*/
 325   oc_quant_table       *dequant_tables[2][3];
 326   oc_quant_tables       dequant_table_data[2][3];
 327   /*Loop filter strength parameters.*/
 328   unsigned char         loop_filter_limits[64];
 329 };
 330
 331
 332
 333 /*The function type used to fill in the chroma plane motion vectors for a
 334    macro block when 4 different motion vectors are specified in the luma
 335    plane.
 336   _cbmvs: The chroma block-level motion vectors to fill in.
 337   _lmbmv: The luma macro-block level motion vector to fill in for use in
 338            prediction.
 339   _lbmvs: The luma block-level motion vectors.*/
 340 typedef void (*oc_set_chroma_mvs_func)(oc_mv _cbmvs[4],
 341  const oc_mv _lbmvs[4]);
 342
 343
 344
 345 /*A map from the index in the zig zag scan to the coefficient number in a
 346    block.
 347   The extra 64 entries send out of bounds indexes to index 64.
 348   This is used to safely ignore invalid zero runs when decoding
 349    coefficients.*/
 350 extern const int OC_FZIG_ZAG[128];
 351 /*A map from the coefficient number in a block to its index in the zig zag
 352    scan.*/
 353 extern const int OC_IZIG_ZAG[64];
 354 /*The predictor frame to use for each macro block mode.*/
 355 extern const int OC_FRAME_FOR_MODE[OC_NMODES];
 356 /*A map from physical macro block ordering to bitstream macro block
 357    ordering within a super block.*/
 358 extern const int OC_MB_MAP[2][2];
 359 /*A list of the indices in the oc_mb.map array that can be valid for each of
 360    the various chroma decimation types.*/
 361 extern const int OC_MB_MAP_IDXS[TH_PF_NFORMATS][12];
 362 /*The number of indices in the oc_mb.map array that can be valid for each of
 363    the various chroma decimation types.*/
 364 extern const int OC_MB_MAP_NIDXS[TH_PF_NFORMATS];
 365 /*A table of functions used to fill in the Cb,Cr plane motion vectors for a
 366    macro block when 4 different motion vectors are specified in the luma
 367    plane.*/
 368 extern const oc_set_chroma_mvs_func OC_SET_CHROMA_MVS_TABLE[TH_PF_NFORMATS];
 369
 370
 371
 372 int oc_ilog(unsigned _v);
 373 void **oc_malloc_2d(size_t _height,size_t _width,size_t _sz);
 374 void **oc_calloc_2d(size_t _height,size_t _width,size_t _sz);
 375 void oc_free_2d(void *_ptr);
 376
 377 void oc_ycbcr_buffer_flip(th_ycbcr_buffer _dst,const th_ycbcr_buffer _src);
 378
 379 int oc_dct_token_skip(int _token,int _extra_bits);
 380
 381 int oc_frag_pred_dc(const oc_fragment *_frag,
 382  const oc_fragment_plane *_fplane,int _x,int _y,int _pred_last[3]);
 383
 384 int oc_state_init(oc_theora_state *_state,const th_info *_info);
 385 void oc_state_clear(oc_theora_state *_state);
 386 void oc_state_vtable_init_c(oc_theora_state *_state);
 387 void oc_state_borders_fill_rows(oc_theora_state *_state,int _refi,int _pli,
 388  int _y0,int _yend);
 389 void oc_state_borders_fill_caps(oc_theora_state *_state,int _refi,int _pli);
 390 void oc_state_borders_fill(oc_theora_state *_state,int _refi);
 391 void oc_state_fill_buffer_ptrs(oc_theora_state *_state,int _buf_idx,
 392  th_ycbcr_buffer _img);
 393 int oc_state_mbi_for_pos(oc_theora_state *_state,int _mbx,int _mby);
 394 int oc_state_get_mv_offsets(oc_theora_state *_state,int *_offset0,
 395  int *_offset1,int _dx,int _dy,int _ystride,int _pli);
 396
 397 int oc_state_loop_filter_init(oc_theora_state *_state,int *_bv);
 398 void oc_state_loop_filter(oc_theora_state *_state,int _frame);
 399 #if defined(OC_DUMP_IMAGES)
 400 int oc_state_dump_frame(const oc_theora_state *_state,int _frame,
 401  const char *_suf);
 402 #endif
 403
 404 /*Shared accelerated functions.*/
 405 void oc_frag_recon_intra(const oc_theora_state *_state,unsigned char *_dst,int _dst_ystride,
 406  const ogg_int16_t *_residue);
 407 void oc_frag_recon_inter(const oc_theora_state *_state,unsigned char *_dst,int _dst_ystride,
 408  const unsigned char *_src,int _src_ystride,const ogg_int16_t *_residue);
 409 void oc_frag_recon_inter2(const oc_theora_state *_state,unsigned char *_dst,int _dst_ystride,
 410  const unsigned char *_src1,int _src1_ystride,const unsigned char *_src2,
 411  int _src2_ystride,const ogg_int16_t *_residue);
 412 void oc_state_frag_copy(const oc_theora_state *_state,const int *_fragis,
 413  int _nfragis,int _dst_frame,int _src_frame,int _pli);
 414 void oc_state_frag_recon(oc_theora_state *_state,const oc_fragment *_frag,
 415  int _pli,ogg_int16_t _dct_coeffs[128],int _last_zzi,int _ncoefs,
 416  ogg_uint16_t _dc_iquant,const ogg_uint16_t _ac_iquant[64]);
 417 void oc_state_loop_filter_frag_rows(oc_theora_state *_state,int *_bv,
 418  int _refi,int _pli,int _fragy0,int _fragy_end);
 419 void oc_restore_fpu(const oc_theora_state *_state);
 420
 421 /*Default pure-C implementations.*/
 422 void oc_frag_recon_intra_c(unsigned char *_dst,int _dst_ystride,
 423  const ogg_int16_t *_residue);
 424 void oc_frag_recon_inter_c(unsigned char *_dst,int _dst_ystride,
 425  const unsigned char *_src,int _src_ystride,const ogg_int16_t *_residue);
 426 void oc_frag_recon_inter2_c(unsigned char *_dst,int _dst_ystride,
 427  const unsigned char *_src1,int _src1_ystride,const unsigned char *_src2,
 428  int _src2_ystride,const ogg_int16_t *_residue);
 429 void oc_state_frag_copy_c(const oc_theora_state *_state,const int *_fragis,
 430  int _nfragis,int _dst_frame,int _src_frame,int _pli);
 431 void oc_state_frag_recon_c(oc_theora_state *_state,const oc_fragment *_frag,
 432  int _pli,ogg_int16_t _dct_coeffs[128],int _last_zzi,int _ncoefs,
 433  ogg_uint16_t _dc_iquant,const ogg_uint16_t _ac_iquant[64]);
 434 void oc_state_loop_filter_frag_rows_c(oc_theora_state *_state,int *_bv,
 435  int _refi,int _pli,int _fragy0,int _fragy_end);
 436 void oc_restore_fpu_c(void);
 437
 438 /*We need a way to call a few enocder functions without introducing a link-time
 439    dependency into the decoder, while still allowing the old alpha API which
 440    does not distinguish between encoder and decoder objects to be used.
 441   We do this by placing a function table at the start of the encoder object
 442    which can dispatch into the encoder library.*/
 443 /*We don't ship theora.h, so we don't want to include it to force theora_state
 444    to be defined; thus it is replaced by void * below.*/
 445 typedef void (*oc_state_clear_func)(void *_th);
 446 typedef int (*oc_state_control_func)(void *th,int req,
 447  void *buf,size_t buf_sz);
 448 typedef ogg_int64_t (*oc_state_granule_frame_func)(void *_th,
 449  ogg_int64_t _granulepos);
 450 typedef double (*oc_state_granule_time_func)(void *_th,
 451  ogg_int64_t _granulepos);
 452
 453 typedef struct oc_state_dispatch_vtbl oc_state_dispatch_vtbl;
 454
 455 struct oc_state_dispatch_vtbl{
 456   oc_state_clear_func         clear;
 457   oc_state_control_func       control;
 458   oc_state_granule_frame_func granule_frame;
 459   oc_state_granule_time_func  granule_time;
 460 };
 461
 462 #endif