| 1 | /******************************************************************** |
| 2 | * * |
| 3 | * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE. * |
| 4 | * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS * |
| 5 | * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE * |
| 6 | * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. * |
| 7 | * * |
| 8 | * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009 * |
| 9 | * by the Xiph.Org Foundation and contributors http://www.xiph.org/ * |
| 10 | * * |
| 11 | ******************************************************************** |
| 12 | |
| 13 | function: |
| 14 | last mod: $Id$ |
| 15 | |
| 16 | ********************************************************************/ |
| 17 | |
| 18 | #include <stdlib.h> |
| 19 | #include <string.h> |
| 20 | #include <ogg/ogg.h> |
| 21 | #include "quant.h" |
| 22 | #include "decint.h" |
| 23 | |
| 24 | /*The maximum output of the DCT with +/- 255 inputs is +/- 8157. |
| 25 | These minimum quantizers ensure the result after quantization (and after |
| 26 | prediction for DC) will be no more than +/- 510. |
| 27 | The tokenization system can handle values up to +/- 580, so there is no need |
| 28 | to do any coefficient clamping. |
| 29 | I would rather have allowed smaller quantizers and had to clamp, but these |
| 30 | minimums were required when constructing the original VP3 matrices and have |
| 31 | been formalized in the spec.*/ |
| 32 | static const unsigned OC_DC_QUANT_MIN[2]={4<<2,8<<2}; |
| 33 | static const unsigned OC_AC_QUANT_MIN[2]={2<<2,4<<2}; |
| 34 | |
| 35 | /*Initializes the dequantization tables from a set of quantizer info. |
| 36 | Currently the dequantizer (and elsewhere enquantizer) tables are expected to |
| 37 | be initialized as pointing to the storage reserved for them in the |
| 38 | oc_theora_state (resp. oc_enc_ctx) structure. |
| 39 | If some tables are duplicates of others, the pointers will be adjusted to |
| 40 | point to a single copy of the tables, but the storage for them will not be |
| 41 | freed. |
| 42 | If you're concerned about the memory footprint, the obvious thing to do is |
| 43 | to move the storage out of its fixed place in the structures and allocate |
| 44 | it on demand. |
| 45 | However, a much, much better option is to only store the quantization |
| 46 | matrices being used for the current frame, and to recalculate these as the |
| 47 | qi values change between frames (this is what VP3 did).*/ |
| 48 | void oc_dequant_tables_init(ogg_uint16_t *_dequant[64][3][2], |
| 49 | int _pp_dc_scale[64],const th_quant_info *_qinfo){ |
| 50 | /*Coding mode: intra or inter.*/ |
| 51 | int qti; |
| 52 | /*Y', C_b, C_r*/ |
| 53 | int pli; |
| 54 | for(qti=0;qti<2;qti++)for(pli=0;pli<3;pli++){ |
| 55 | /*Quality index.*/ |
| 56 | int qi; |
| 57 | /*Range iterator.*/ |
| 58 | int qri; |
| 59 | for(qi=0,qri=0;qri<=_qinfo->qi_ranges[qti][pli].nranges;qri++){ |
| 60 | th_quant_base base; |
| 61 | ogg_uint32_t q; |
| 62 | int qi_start; |
| 63 | int qi_end; |
| 64 | memcpy(base,_qinfo->qi_ranges[qti][pli].base_matrices[qri], |
| 65 | sizeof(base)); |
| 66 | qi_start=qi; |
| 67 | if(qri==_qinfo->qi_ranges[qti][pli].nranges)qi_end=qi+1; |
| 68 | else qi_end=qi+_qinfo->qi_ranges[qti][pli].sizes[qri]; |
| 69 | /*Iterate over quality indicies in this range.*/ |
| 70 | for(;;){ |
| 71 | ogg_uint32_t qfac; |
| 72 | int zzi; |
| 73 | int ci; |
| 74 | /*In the original VP3.2 code, the rounding offset and the size of the |
| 75 | dead zone around 0 were controlled by a "sharpness" parameter. |
| 76 | The size of our dead zone is now controlled by the per-coefficient |
| 77 | quality thresholds returned by our HVS module. |
| 78 | We round down from a more accurate value when the quality of the |
| 79 | reconstruction does not fall below our threshold and it saves bits. |
| 80 | Hence, all of that VP3.2 code is gone from here, and the remaining |
| 81 | floating point code has been implemented as equivalent integer code |
| 82 | with exact precision.*/ |
| 83 | qfac=(ogg_uint32_t)_qinfo->dc_scale[qi]*base[0]; |
| 84 | /*For postprocessing, not dequantization.*/ |
| 85 | if(_pp_dc_scale!=NULL)_pp_dc_scale[qi]=(int)(qfac/160); |
| 86 | /*Scale DC the coefficient from the proper table.*/ |
| 87 | q=(qfac/100)<<2; |
| 88 | q=OC_CLAMPI(OC_DC_QUANT_MIN[qti],q,OC_QUANT_MAX); |
| 89 | _dequant[qi][pli][qti][0]=(ogg_uint16_t)q; |
| 90 | /*Now scale AC coefficients from the proper table.*/ |
| 91 | for(zzi=1;zzi<64;zzi++){ |
| 92 | q=((ogg_uint32_t)_qinfo->ac_scale[qi]*base[OC_FZIG_ZAG[zzi]]/100)<<2; |
| 93 | q=OC_CLAMPI(OC_AC_QUANT_MIN[qti],q,OC_QUANT_MAX); |
| 94 | _dequant[qi][pli][qti][zzi]=(ogg_uint16_t)q; |
| 95 | } |
| 96 | /*If this is a duplicate of a previous matrix, use that instead. |
| 97 | This simple check helps us improve cache coherency later.*/ |
| 98 | { |
| 99 | int dupe; |
| 100 | int qtj; |
| 101 | int plj; |
| 102 | dupe=0; |
| 103 | for(qtj=0;qtj<=qti;qtj++){ |
| 104 | for(plj=0;plj<(qtj<qti?3:pli);plj++){ |
| 105 | if(!memcmp(_dequant[qi][pli][qti],_dequant[qi][plj][qtj], |
| 106 | sizeof(oc_quant_table))){ |
| 107 | dupe=1; |
| 108 | break; |
| 109 | } |
| 110 | } |
| 111 | if(dupe)break; |
| 112 | } |
| 113 | if(dupe)_dequant[qi][pli][qti]=_dequant[qi][plj][qtj]; |
| 114 | } |
| 115 | if(++qi>=qi_end)break; |
| 116 | /*Interpolate the next base matrix.*/ |
| 117 | for(ci=0;ci<64;ci++){ |
| 118 | base[ci]=(unsigned char)( |
| 119 | (2*((qi_end-qi)*_qinfo->qi_ranges[qti][pli].base_matrices[qri][ci]+ |
| 120 | (qi-qi_start)*_qinfo->qi_ranges[qti][pli].base_matrices[qri+1][ci]) |
| 121 | +_qinfo->qi_ranges[qti][pli].sizes[qri])/ |
| 122 | (2*_qinfo->qi_ranges[qti][pli].sizes[qri])); |
| 123 | } |
| 124 | } |
| 125 | } |
| 126 | } |
| 127 | } |
| 128 | |