1 // Copyright 2010 Google Inc. All Rights Reserved. 2 // 3 // Use of this source code is governed by a BSD-style license 4 // that can be found in the COPYING file in the root of the source 5 // tree. An additional intellectual property rights grant can be found 6 // in the file PATENTS. All contributing project authors may 7 // be found in the AUTHORS file in the root of the source tree. 8 // ----------------------------------------------------------------------------- 9 // 10 // VP8 decoder: internal header. 11 // 12 // Author: Skal (pascal.massimino@gmail.com) 13 14 #ifndef WEBP_DEC_VP8I_H_ 15 #define WEBP_DEC_VP8I_H_ 16 17 #include <string.h> // for memcpy() 18 #include "./common_dec.h" 19 #include "./vp8li_dec.h" 20 #include "../utils/bit_reader_utils.h" 21 #include "../utils/random_utils.h" 22 #include "../utils/thread_utils.h" 23 #include "../dsp/dsp.h" 24 25 #ifdef __cplusplus 26 extern "C" { 27 #endif 28 29 //------------------------------------------------------------------------------ 30 // Various defines and enums 31 32 // version numbers 33 #define DEC_MAJ_VERSION 0 34 #define DEC_MIN_VERSION 6 35 #define DEC_REV_VERSION 0 36 37 // YUV-cache parameters. Cache is 32-bytes wide (= one cacheline). 38 // Constraints are: We need to store one 16x16 block of luma samples (y), 39 // and two 8x8 chroma blocks (u/v). These are better be 16-bytes aligned, 40 // in order to be SIMD-friendly. We also need to store the top, left and 41 // top-left samples (from previously decoded blocks), along with four 42 // extra top-right samples for luma (intra4x4 prediction only). 43 // One possible layout is, using 32 * (17 + 9) bytes: 44 // 45 // .+------ <- only 1 pixel high 46 // .|yyyyt. 47 // .|yyyyt. 48 // .|yyyyt. 49 // .|yyyy.. 50 // .+--.+-- <- only 1 pixel high 51 // .|uu.|vv 52 // .|uu.|vv 53 // 54 // Every character is a 4x4 block, with legend: 55 // '.' = unused 56 // 'y' = y-samples 'u' = u-samples 'v' = u-samples 57 // '|' = left sample, '-' = top sample, '+' = top-left sample 58 // 't' = extra top-right sample for 4x4 modes 59 #define YUV_SIZE (BPS * 17 + BPS * 9) 60 #define Y_SIZE (BPS * 17) 61 #define Y_OFF (BPS * 1 + 8) 62 #define U_OFF (Y_OFF + BPS * 16 + BPS) 63 #define V_OFF (U_OFF + 16) 64 65 // minimal width under which lossy multi-threading is always disabled 66 #define MIN_WIDTH_FOR_THREADS 512 67 68 //------------------------------------------------------------------------------ 69 // Headers 70 71 typedef struct { 72 uint8_t key_frame_; 73 uint8_t profile_; 74 uint8_t show_; 75 uint32_t partition_length_; 76 } VP8FrameHeader; 77 78 typedef struct { 79 uint16_t width_; 80 uint16_t height_; 81 uint8_t xscale_; 82 uint8_t yscale_; 83 uint8_t colorspace_; // 0 = YCbCr 84 uint8_t clamp_type_; 85 } VP8PictureHeader; 86 87 // segment features 88 typedef struct { 89 int use_segment_; 90 int update_map_; // whether to update the segment map or not 91 int absolute_delta_; // absolute or delta values for quantizer and filter 92 int8_t quantizer_[NUM_MB_SEGMENTS]; // quantization changes 93 int8_t filter_strength_[NUM_MB_SEGMENTS]; // filter strength for segments 94 } VP8SegmentHeader; 95 96 // probas associated to one of the contexts 97 typedef uint8_t VP8ProbaArray[NUM_PROBAS]; 98 99 typedef struct { // all the probas associated to one band 100 VP8ProbaArray probas_[NUM_CTX]; 101 } VP8BandProbas; 102 103 // Struct collecting all frame-persistent probabilities. 104 typedef struct { 105 uint8_t segments_[MB_FEATURE_TREE_PROBS]; 106 // Type: 0:Intra16-AC 1:Intra16-DC 2:Chroma 3:Intra4 107 VP8BandProbas bands_[NUM_TYPES][NUM_BANDS]; 108 const VP8BandProbas* bands_ptr_[NUM_TYPES][16 + 1]; 109 } VP8Proba; 110 111 // Filter parameters 112 typedef struct { 113 int simple_; // 0=complex, 1=simple 114 int level_; // [0..63] 115 int sharpness_; // [0..7] 116 int use_lf_delta_; 117 int ref_lf_delta_[NUM_REF_LF_DELTAS]; 118 int mode_lf_delta_[NUM_MODE_LF_DELTAS]; 119 } VP8FilterHeader; 120 121 //------------------------------------------------------------------------------ 122 // Informations about the macroblocks. 123 124 typedef struct { // filter specs 125 uint8_t f_limit_; // filter limit in [3..189], or 0 if no filtering 126 uint8_t f_ilevel_; // inner limit in [1..63] 127 uint8_t f_inner_; // do inner filtering? 128 uint8_t hev_thresh_; // high edge variance threshold in [0..2] 129 } VP8FInfo; 130 131 typedef struct { // Top/Left Contexts used for syntax-parsing 132 uint8_t nz_; // non-zero AC/DC coeffs (4bit for luma + 4bit for chroma) 133 uint8_t nz_dc_; // non-zero DC coeff (1bit) 134 } VP8MB; 135 136 // Dequantization matrices 137 typedef int quant_t[2]; // [DC / AC]. Can be 'uint16_t[2]' too (~slower). 138 typedef struct { 139 quant_t y1_mat_, y2_mat_, uv_mat_; 140 141 int uv_quant_; // U/V quantizer value 142 int dither_; // dithering amplitude (0 = off, max=255) 143 } VP8QuantMatrix; 144 145 // Data needed to reconstruct a macroblock 146 typedef struct { 147 int16_t coeffs_[384]; // 384 coeffs = (16+4+4) * 4*4 148 uint8_t is_i4x4_; // true if intra4x4 149 uint8_t imodes_[16]; // one 16x16 mode (#0) or sixteen 4x4 modes 150 uint8_t uvmode_; // chroma prediction mode 151 // bit-wise info about the content of each sub-4x4 blocks (in decoding order). 152 // Each of the 4x4 blocks for y/u/v is associated with a 2b code according to: 153 // code=0 -> no coefficient 154 // code=1 -> only DC 155 // code=2 -> first three coefficients are non-zero 156 // code=3 -> more than three coefficients are non-zero 157 // This allows to call specialized transform functions. 158 uint32_t non_zero_y_; 159 uint32_t non_zero_uv_; 160 uint8_t dither_; // local dithering strength (deduced from non_zero_*) 161 uint8_t skip_; 162 uint8_t segment_; 163 } VP8MBData; 164 165 // Persistent information needed by the parallel processing 166 typedef struct { 167 int id_; // cache row to process (in [0..2]) 168 int mb_y_; // macroblock position of the row 169 int filter_row_; // true if row-filtering is needed 170 VP8FInfo* f_info_; // filter strengths (swapped with dec->f_info_) 171 VP8MBData* mb_data_; // reconstruction data (swapped with dec->mb_data_) 172 VP8Io io_; // copy of the VP8Io to pass to put() 173 } VP8ThreadContext; 174 175 // Saved top samples, per macroblock. Fits into a cache-line. 176 typedef struct { 177 uint8_t y[16], u[8], v[8]; 178 } VP8TopSamples; 179 180 //------------------------------------------------------------------------------ 181 // VP8Decoder: the main opaque structure handed over to user 182 183 struct VP8Decoder { 184 VP8StatusCode status_; 185 int ready_; // true if ready to decode a picture with VP8Decode() 186 const char* error_msg_; // set when status_ is not OK. 187 188 // Main data source 189 VP8BitReader br_; 190 191 // headers 192 VP8FrameHeader frm_hdr_; 193 VP8PictureHeader pic_hdr_; 194 VP8FilterHeader filter_hdr_; 195 VP8SegmentHeader segment_hdr_; 196 197 // Worker 198 WebPWorker worker_; 199 int mt_method_; // multi-thread method: 0=off, 1=[parse+recon][filter] 200 // 2=[parse][recon+filter] 201 int cache_id_; // current cache row 202 int num_caches_; // number of cached rows of 16 pixels (1, 2 or 3) 203 VP8ThreadContext thread_ctx_; // Thread context 204 205 // dimension, in macroblock units. 206 int mb_w_, mb_h_; 207 208 // Macroblock to process/filter, depending on cropping and filter_type. 209 int tl_mb_x_, tl_mb_y_; // top-left MB that must be in-loop filtered 210 int br_mb_x_, br_mb_y_; // last bottom-right MB that must be decoded 211 212 // number of partitions minus one. 213 uint32_t num_parts_minus_one_; 214 // per-partition boolean decoders. 215 VP8BitReader parts_[MAX_NUM_PARTITIONS]; 216 217 // Dithering strength, deduced from decoding options 218 int dither_; // whether to use dithering or not 219 VP8Random dithering_rg_; // random generator for dithering 220 221 // dequantization (one set of DC/AC dequant factor per segment) 222 VP8QuantMatrix dqm_[NUM_MB_SEGMENTS]; 223 224 // probabilities 225 VP8Proba proba_; 226 int use_skip_proba_; 227 uint8_t skip_p_; 228 229 // Boundary data cache and persistent buffers. 230 uint8_t* intra_t_; // top intra modes values: 4 * mb_w_ 231 uint8_t intra_l_[4]; // left intra modes values 232 233 VP8TopSamples* yuv_t_; // top y/u/v samples 234 235 VP8MB* mb_info_; // contextual macroblock info (mb_w_ + 1) 236 VP8FInfo* f_info_; // filter strength info 237 uint8_t* yuv_b_; // main block for Y/U/V (size = YUV_SIZE) 238 239 uint8_t* cache_y_; // macroblock row for storing unfiltered samples 240 uint8_t* cache_u_; 241 uint8_t* cache_v_; 242 int cache_y_stride_; 243 int cache_uv_stride_; 244 245 // main memory chunk for the above data. Persistent. 246 void* mem_; 247 size_t mem_size_; 248 249 // Per macroblock non-persistent infos. 250 int mb_x_, mb_y_; // current position, in macroblock units 251 VP8MBData* mb_data_; // parsed reconstruction data 252 253 // Filtering side-info 254 int filter_type_; // 0=off, 1=simple, 2=complex 255 VP8FInfo fstrengths_[NUM_MB_SEGMENTS][2]; // precalculated per-segment/type 256 257 // Alpha 258 struct ALPHDecoder* alph_dec_; // alpha-plane decoder object 259 const uint8_t* alpha_data_; // compressed alpha data (if present) 260 size_t alpha_data_size_; 261 int is_alpha_decoded_; // true if alpha_data_ is decoded in alpha_plane_ 262 uint8_t* alpha_plane_mem_; // memory allocated for alpha_plane_ 263 uint8_t* alpha_plane_; // output. Persistent, contains the whole data. 264 const uint8_t* alpha_prev_line_; // last decoded alpha row (or NULL) 265 int alpha_dithering_; // derived from decoding options (0=off, 100=full) 266 }; 267 268 //------------------------------------------------------------------------------ 269 // internal functions. Not public. 270 271 // in vp8.c 272 int VP8SetError(VP8Decoder* const dec, 273 VP8StatusCode error, const char* const msg); 274 275 // in tree.c 276 void VP8ResetProba(VP8Proba* const proba); 277 void VP8ParseProba(VP8BitReader* const br, VP8Decoder* const dec); 278 // parses one row of intra mode data in partition 0, returns !eof 279 int VP8ParseIntraModeRow(VP8BitReader* const br, VP8Decoder* const dec); 280 281 // in quant.c 282 void VP8ParseQuant(VP8Decoder* const dec); 283 284 // in frame.c 285 int VP8InitFrame(VP8Decoder* const dec, VP8Io* const io); 286 // Call io->setup() and finish setting up scan parameters. 287 // After this call returns, one must always call VP8ExitCritical() with the 288 // same parameters. Both functions should be used in pair. Returns VP8_STATUS_OK 289 // if ok, otherwise sets and returns the error status on *dec. 290 VP8StatusCode VP8EnterCritical(VP8Decoder* const dec, VP8Io* const io); 291 // Must always be called in pair with VP8EnterCritical(). 292 // Returns false in case of error. 293 int VP8ExitCritical(VP8Decoder* const dec, VP8Io* const io); 294 // Return the multi-threading method to use (0=off), depending 295 // on options and bitstream size. Only for lossy decoding. 296 int VP8GetThreadMethod(const WebPDecoderOptions* const options, 297 const WebPHeaderStructure* const headers, 298 int width, int height); 299 // Initialize dithering post-process if needed. 300 void VP8InitDithering(const WebPDecoderOptions* const options, 301 VP8Decoder* const dec); 302 // Process the last decoded row (filtering + output). 303 int VP8ProcessRow(VP8Decoder* const dec, VP8Io* const io); 304 // To be called at the start of a new scanline, to initialize predictors. 305 void VP8InitScanline(VP8Decoder* const dec); 306 // Decode one macroblock. Returns false if there is not enough data. 307 int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br); 308 309 // in alpha.c 310 const uint8_t* VP8DecompressAlphaRows(VP8Decoder* const dec, 311 const VP8Io* const io, 312 int row, int num_rows); 313 314 //------------------------------------------------------------------------------ 315 316 #ifdef __cplusplus 317 } // extern "C" 318 #endif 319 320 #endif /* WEBP_DEC_VP8I_H_ */ 321