LCOV - code coverage report
Current view: directory - media/libjpeg - jcdctmgr.c (source / functions) Found Hit Coverage
Test: app.info Lines: 205 85 41.5 %
Date: 2012-06-02 Functions: 10 5 50.0 %

       1                 : /*
       2                 :  * jcdctmgr.c
       3                 :  *
       4                 :  * Copyright (C) 1994-1996, Thomas G. Lane.
       5                 :  * Copyright (C) 1999-2006, MIYASAKA Masaru.
       6                 :  * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
       7                 :  * Copyright (C) 2011 D. R. Commander
       8                 :  * This file is part of the Independent JPEG Group's software.
       9                 :  * For conditions of distribution and use, see the accompanying README file.
      10                 :  *
      11                 :  * This file contains the forward-DCT management logic.
      12                 :  * This code selects a particular DCT implementation to be used,
      13                 :  * and it performs related housekeeping chores including coefficient
      14                 :  * quantization.
      15                 :  */
      16                 : 
      17                 : #define JPEG_INTERNALS
      18                 : #include "jinclude.h"
      19                 : #include "jpeglib.h"
      20                 : #include "jdct.h"             /* Private declarations for DCT subsystem */
      21                 : #include "jsimddct.h"
      22                 : 
      23                 : 
      24                 : /* Private subobject for this module */
      25                 : 
      26                 : typedef JMETHOD(void, forward_DCT_method_ptr, (DCTELEM * data));
      27                 : typedef JMETHOD(void, float_DCT_method_ptr, (FAST_FLOAT * data));
      28                 : 
      29                 : typedef JMETHOD(void, convsamp_method_ptr,
      30                 :                 (JSAMPARRAY sample_data, JDIMENSION start_col,
      31                 :                  DCTELEM * workspace));
      32                 : typedef JMETHOD(void, float_convsamp_method_ptr,
      33                 :                 (JSAMPARRAY sample_data, JDIMENSION start_col,
      34                 :                  FAST_FLOAT *workspace));
      35                 : 
      36                 : typedef JMETHOD(void, quantize_method_ptr,
      37                 :                 (JCOEFPTR coef_block, DCTELEM * divisors,
      38                 :                  DCTELEM * workspace));
      39                 : typedef JMETHOD(void, float_quantize_method_ptr,
      40                 :                 (JCOEFPTR coef_block, FAST_FLOAT * divisors,
      41                 :                  FAST_FLOAT * workspace));
      42                 : 
      43                 : METHODDEF(void) quantize (JCOEFPTR, DCTELEM *, DCTELEM *);
      44                 : 
      45                 : typedef struct {
      46                 :   struct jpeg_forward_dct pub;  /* public fields */
      47                 : 
      48                 :   /* Pointer to the DCT routine actually in use */
      49                 :   forward_DCT_method_ptr dct;
      50                 :   convsamp_method_ptr convsamp;
      51                 :   quantize_method_ptr quantize;
      52                 : 
      53                 :   /* The actual post-DCT divisors --- not identical to the quant table
      54                 :    * entries, because of scaling (especially for an unnormalized DCT).
      55                 :    * Each table is given in normal array order.
      56                 :    */
      57                 :   DCTELEM * divisors[NUM_QUANT_TBLS];
      58                 : 
      59                 :   /* work area for FDCT subroutine */
      60                 :   DCTELEM * workspace;
      61                 : 
      62                 : #ifdef DCT_FLOAT_SUPPORTED
      63                 :   /* Same as above for the floating-point case. */
      64                 :   float_DCT_method_ptr float_dct;
      65                 :   float_convsamp_method_ptr float_convsamp;
      66                 :   float_quantize_method_ptr float_quantize;
      67                 :   FAST_FLOAT * float_divisors[NUM_QUANT_TBLS];
      68                 :   FAST_FLOAT * float_workspace;
      69                 : #endif
      70                 : } my_fdct_controller;
      71                 : 
      72                 : typedef my_fdct_controller * my_fdct_ptr;
      73                 : 
      74                 : 
      75                 : /*
      76                 :  * Find the highest bit in an integer through binary search.
      77                 :  */
      78                 : LOCAL(int)
      79             576 : flss (UINT16 val)
      80                 : {
      81                 :   int bit;
      82                 : 
      83             576 :   bit = 16;
      84                 : 
      85             576 :   if (!val)
      86               0 :     return 0;
      87                 : 
      88             576 :   if (!(val & 0xff00)) {
      89             576 :     bit -= 8;
      90             576 :     val <<= 8;
      91                 :   }
      92             576 :   if (!(val & 0xf000)) {
      93               0 :     bit -= 4;
      94               0 :     val <<= 4;
      95                 :   }
      96             576 :   if (!(val & 0xc000)) {
      97             123 :     bit -= 2;
      98             123 :     val <<= 2;
      99                 :   }
     100             576 :   if (!(val & 0x8000)) {
     101             174 :     bit -= 1;
     102             174 :     val <<= 1;
     103                 :   }
     104                 : 
     105             576 :   return bit;
     106                 : }
     107                 : 
     108                 : /*
     109                 :  * Compute values to do a division using reciprocal.
     110                 :  *
     111                 :  * This implementation is based on an algorithm described in
     112                 :  *   "How to optimize for the Pentium family of microprocessors"
     113                 :  *   (http://www.agner.org/assem/).
     114                 :  * More information about the basic algorithm can be found in
     115                 :  * the paper "Integer Division Using Reciprocals" by Robert Alverson.
     116                 :  *
     117                 :  * The basic idea is to replace x/d by x * d^-1. In order to store
     118                 :  * d^-1 with enough precision we shift it left a few places. It turns
     119                 :  * out that this algoright gives just enough precision, and also fits
     120                 :  * into DCTELEM:
     121                 :  *
     122                 :  *   b = (the number of significant bits in divisor) - 1
     123                 :  *   r = (word size) + b
     124                 :  *   f = 2^r / divisor
     125                 :  *
     126                 :  * f will not be an integer for most cases, so we need to compensate
     127                 :  * for the rounding error introduced:
     128                 :  *
     129                 :  *   no fractional part:
     130                 :  *
     131                 :  *       result = input >> r
     132                 :  *
     133                 :  *   fractional part of f < 0.5:
     134                 :  *
     135                 :  *       round f down to nearest integer
     136                 :  *       result = ((input + 1) * f) >> r
     137                 :  *
     138                 :  *   fractional part of f > 0.5:
     139                 :  *
     140                 :  *       round f up to nearest integer
     141                 :  *       result = (input * f) >> r
     142                 :  *
     143                 :  * This is the original algorithm that gives truncated results. But we
     144                 :  * want properly rounded results, so we replace "input" with
     145                 :  * "input + divisor/2".
     146                 :  *
     147                 :  * In order to allow SIMD implementations we also tweak the values to
     148                 :  * allow the same calculation to be made at all times:
     149                 :  * 
     150                 :  *   dctbl[0] = f rounded to nearest integer
     151                 :  *   dctbl[1] = divisor / 2 (+ 1 if fractional part of f < 0.5)
     152                 :  *   dctbl[2] = 1 << ((word size) * 2 - r)
     153                 :  *   dctbl[3] = r - (word size)
     154                 :  *
     155                 :  * dctbl[2] is for stupid instruction sets where the shift operation
     156                 :  * isn't member wise (e.g. MMX).
     157                 :  *
     158                 :  * The reason dctbl[2] and dctbl[3] reduce the shift with (word size)
     159                 :  * is that most SIMD implementations have a "multiply and store top
     160                 :  * half" operation.
     161                 :  *
     162                 :  * Lastly, we store each of the values in their own table instead
     163                 :  * of in a consecutive manner, yet again in order to allow SIMD
     164                 :  * routines.
     165                 :  */
     166                 : LOCAL(int)
     167             576 : compute_reciprocal (UINT16 divisor, DCTELEM * dtbl)
     168                 : {
     169                 :   UDCTELEM2 fq, fr;
     170                 :   UDCTELEM c;
     171                 :   int b, r;
     172                 : 
     173             576 :   b = flss(divisor) - 1;
     174             576 :   r  = sizeof(DCTELEM) * 8 + b;
     175                 : 
     176             576 :   fq = ((UDCTELEM2)1 << r) / divisor;
     177             576 :   fr = ((UDCTELEM2)1 << r) % divisor;
     178                 : 
     179             576 :   c = divisor / 2; /* for rounding */
     180                 : 
     181             576 :   if (fr == 0) { /* divisor is power of two */
     182                 :     /* fq will be one bit too large to fit in DCTELEM, so adjust */
     183             414 :     fq >>= 1;
     184             414 :     r--;
     185             162 :   } else if (fr <= (divisor / 2U)) { /* fractional part is < 0.5 */
     186              42 :     c++;
     187                 :   } else { /* fractional part is > 0.5 */
     188             120 :     fq++;
     189                 :   }
     190                 : 
     191             576 :   dtbl[DCTSIZE2 * 0] = (DCTELEM) fq;      /* reciprocal */
     192             576 :   dtbl[DCTSIZE2 * 1] = (DCTELEM) c;       /* correction + roundfactor */
     193             576 :   dtbl[DCTSIZE2 * 2] = (DCTELEM) (1 << (sizeof(DCTELEM)*8*2 - r));  /* scale */
     194             576 :   dtbl[DCTSIZE2 * 3] = (DCTELEM) r - sizeof(DCTELEM)*8; /* shift */
     195                 : 
     196             576 :   if(r <= 16) return 0;
     197             576 :   else return 1;
     198                 : }
     199                 : 
     200                 : /*
     201                 :  * Initialize for a processing pass.
     202                 :  * Verify that all referenced Q-tables are present, and set up
     203                 :  * the divisor table for each one.
     204                 :  * In the current implementation, DCT of all components is done during
     205                 :  * the first pass, even if only some components will be output in the
     206                 :  * first scan.  Hence all components should be examined here.
     207                 :  */
     208                 : 
     209                 : METHODDEF(void)
     210               3 : start_pass_fdctmgr (j_compress_ptr cinfo)
     211                 : {
     212               3 :   my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
     213                 :   int ci, qtblno, i;
     214                 :   jpeg_component_info *compptr;
     215                 :   JQUANT_TBL * qtbl;
     216                 :   DCTELEM * dtbl;
     217                 : 
     218              15 :   for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
     219               9 :        ci++, compptr++) {
     220               9 :     qtblno = compptr->quant_tbl_no;
     221                 :     /* Make sure specified quantization table is present */
     222              18 :     if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
     223               9 :         cinfo->quant_tbl_ptrs[qtblno] == NULL)
     224               0 :       ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
     225               9 :     qtbl = cinfo->quant_tbl_ptrs[qtblno];
     226                 :     /* Compute divisors for this quant table */
     227                 :     /* We may do this more than once for same table, but it's not a big deal */
     228               9 :     switch (cinfo->dct_method) {
     229                 : #ifdef DCT_ISLOW_SUPPORTED
     230                 :     case JDCT_ISLOW:
     231                 :       /* For LL&M IDCT method, divisors are equal to raw quantization
     232                 :        * coefficients multiplied by 8 (to counteract scaling).
     233                 :        */
     234               9 :       if (fdct->divisors[qtblno] == NULL) {
     235               6 :         fdct->divisors[qtblno] = (DCTELEM *)
     236               6 :           (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
     237                 :                                       (DCTSIZE2 * 4) * SIZEOF(DCTELEM));
     238                 :       }
     239               9 :       dtbl = fdct->divisors[qtblno];
     240             585 :       for (i = 0; i < DCTSIZE2; i++) {
     241             576 :         if(!compute_reciprocal(qtbl->quantval[i] << 3, &dtbl[i])
     242               0 :           && fdct->quantize == jsimd_quantize)
     243               0 :           fdct->quantize = quantize;
     244                 :       }
     245               9 :       break;
     246                 : #endif
     247                 : #ifdef DCT_IFAST_SUPPORTED
     248                 :     case JDCT_IFAST:
     249                 :       {
     250                 :         /* For AA&N IDCT method, divisors are equal to quantization
     251                 :          * coefficients scaled by scalefactor[row]*scalefactor[col], where
     252                 :          *   scalefactor[0] = 1
     253                 :          *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
     254                 :          * We apply a further scale factor of 8.
     255                 :          */
     256                 : #define CONST_BITS 14
     257                 :         static const INT16 aanscales[DCTSIZE2] = {
     258                 :           /* precomputed values scaled up by 14 bits */
     259                 :           16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
     260                 :           22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
     261                 :           21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
     262                 :           19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
     263                 :           16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
     264                 :           12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
     265                 :            8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
     266                 :            4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
     267                 :         };
     268                 :         SHIFT_TEMPS
     269                 : 
     270               0 :         if (fdct->divisors[qtblno] == NULL) {
     271               0 :           fdct->divisors[qtblno] = (DCTELEM *)
     272               0 :             (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
     273                 :                                         (DCTSIZE2 * 4) * SIZEOF(DCTELEM));
     274                 :         }
     275               0 :         dtbl = fdct->divisors[qtblno];
     276               0 :         for (i = 0; i < DCTSIZE2; i++) {
     277               0 :           if(!compute_reciprocal(
     278               0 :             DESCALE(MULTIPLY16V16((INT32) qtbl->quantval[i],
     279                 :                                   (INT32) aanscales[i]),
     280               0 :                     CONST_BITS-3), &dtbl[i])
     281               0 :             && fdct->quantize == jsimd_quantize)
     282               0 :             fdct->quantize = quantize;
     283                 :         }
     284                 :       }
     285               0 :       break;
     286                 : #endif
     287                 : #ifdef DCT_FLOAT_SUPPORTED
     288                 :     case JDCT_FLOAT:
     289                 :       {
     290                 :         /* For float AA&N IDCT method, divisors are equal to quantization
     291                 :          * coefficients scaled by scalefactor[row]*scalefactor[col], where
     292                 :          *   scalefactor[0] = 1
     293                 :          *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
     294                 :          * We apply a further scale factor of 8.
     295                 :          * What's actually stored is 1/divisor so that the inner loop can
     296                 :          * use a multiplication rather than a division.
     297                 :          */
     298                 :         FAST_FLOAT * fdtbl;
     299                 :         int row, col;
     300                 :         static const double aanscalefactor[DCTSIZE] = {
     301                 :           1.0, 1.387039845, 1.306562965, 1.175875602,
     302                 :           1.0, 0.785694958, 0.541196100, 0.275899379
     303                 :         };
     304                 : 
     305               0 :         if (fdct->float_divisors[qtblno] == NULL) {
     306               0 :           fdct->float_divisors[qtblno] = (FAST_FLOAT *)
     307               0 :             (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
     308                 :                                         DCTSIZE2 * SIZEOF(FAST_FLOAT));
     309                 :         }
     310               0 :         fdtbl = fdct->float_divisors[qtblno];
     311               0 :         i = 0;
     312               0 :         for (row = 0; row < DCTSIZE; row++) {
     313               0 :           for (col = 0; col < DCTSIZE; col++) {
     314               0 :             fdtbl[i] = (FAST_FLOAT)
     315               0 :               (1.0 / (((double) qtbl->quantval[i] *
     316               0 :                        aanscalefactor[row] * aanscalefactor[col] * 8.0)));
     317               0 :             i++;
     318                 :           }
     319                 :         }
     320                 :       }
     321               0 :       break;
     322                 : #endif
     323                 :     default:
     324               0 :       ERREXIT(cinfo, JERR_NOT_COMPILED);
     325               0 :       break;
     326                 :     }
     327                 :   }
     328               3 : }
     329                 : 
     330                 : 
     331                 : /*
     332                 :  * Load data into workspace, applying unsigned->signed conversion.
     333                 :  */
     334                 : 
     335                 : METHODDEF(void)
     336               0 : convsamp (JSAMPARRAY sample_data, JDIMENSION start_col, DCTELEM * workspace)
     337                 : {
     338                 :   register DCTELEM *workspaceptr;
     339                 :   register JSAMPROW elemptr;
     340                 :   register int elemr;
     341                 : 
     342               0 :   workspaceptr = workspace;
     343               0 :   for (elemr = 0; elemr < DCTSIZE; elemr++) {
     344               0 :     elemptr = sample_data[elemr] + start_col;
     345                 : 
     346                 : #if DCTSIZE == 8                /* unroll the inner loop */
     347               0 :     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
     348               0 :     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
     349               0 :     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
     350               0 :     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
     351               0 :     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
     352               0 :     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
     353               0 :     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
     354               0 :     *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
     355                 : #else
     356                 :     {
     357                 :       register int elemc;
     358                 :       for (elemc = DCTSIZE; elemc > 0; elemc--)
     359                 :         *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
     360                 :     }
     361                 : #endif
     362                 :   }
     363               0 : }
     364                 : 
     365                 : 
     366                 : /*
     367                 :  * Quantize/descale the coefficients, and store into coef_blocks[].
     368                 :  */
     369                 : 
     370                 : METHODDEF(void)
     371               0 : quantize (JCOEFPTR coef_block, DCTELEM * divisors, DCTELEM * workspace)
     372                 : {
     373                 :   int i;
     374                 :   DCTELEM temp;
     375                 :   UDCTELEM recip, corr, shift;
     376                 :   UDCTELEM2 product;
     377               0 :   JCOEFPTR output_ptr = coef_block;
     378                 : 
     379               0 :   for (i = 0; i < DCTSIZE2; i++) {
     380               0 :     temp = workspace[i];
     381               0 :     recip = divisors[i + DCTSIZE2 * 0];
     382               0 :     corr =  divisors[i + DCTSIZE2 * 1];
     383               0 :     shift = divisors[i + DCTSIZE2 * 3];
     384                 : 
     385               0 :     if (temp < 0) {
     386               0 :       temp = -temp;
     387               0 :       product = (UDCTELEM2)(temp + corr) * recip;
     388               0 :       product >>= shift + sizeof(DCTELEM)*8;
     389               0 :       temp = product;
     390               0 :       temp = -temp;
     391                 :     } else {
     392               0 :       product = (UDCTELEM2)(temp + corr) * recip;
     393               0 :       product >>= shift + sizeof(DCTELEM)*8;
     394               0 :       temp = product;
     395                 :     }
     396                 : 
     397               0 :     output_ptr[i] = (JCOEF) temp;
     398                 :   }
     399               0 : }
     400                 : 
     401                 : 
     402                 : /*
     403                 :  * Perform forward DCT on one or more blocks of a component.
     404                 :  *
     405                 :  * The input samples are taken from the sample_data[] array starting at
     406                 :  * position start_row/start_col, and moving to the right for any additional
     407                 :  * blocks. The quantized coefficients are returned in coef_blocks[].
     408                 :  */
     409                 : 
     410                 : METHODDEF(void)
     411             396 : forward_DCT (j_compress_ptr cinfo, jpeg_component_info * compptr,
     412                 :              JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
     413                 :              JDIMENSION start_row, JDIMENSION start_col,
     414                 :              JDIMENSION num_blocks)
     415                 : /* This version is used for integer DCT implementations. */
     416                 : {
     417                 :   /* This routine is heavily used, so it's worth coding it tightly. */
     418             396 :   my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
     419             396 :   DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
     420                 :   DCTELEM * workspace;
     421                 :   JDIMENSION bi;
     422                 : 
     423                 :   /* Make sure the compiler doesn't look up these every pass */
     424             396 :   forward_DCT_method_ptr do_dct = fdct->dct;
     425             396 :   convsamp_method_ptr do_convsamp = fdct->convsamp;
     426             396 :   quantize_method_ptr do_quantize = fdct->quantize;
     427             396 :   workspace = fdct->workspace;
     428                 : 
     429             396 :   sample_data += start_row;     /* fold in the vertical offset once */
     430                 : 
     431             792 :   for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
     432                 :     /* Load data into workspace, applying unsigned->signed conversion */
     433             396 :     (*do_convsamp) (sample_data, start_col, workspace);
     434                 : 
     435                 :     /* Perform the DCT */
     436             396 :     (*do_dct) (workspace);
     437                 : 
     438                 :     /* Quantize/descale the coefficients, and store into coef_blocks[] */
     439             396 :     (*do_quantize) (coef_blocks[bi], divisors, workspace);
     440                 :   }
     441             396 : }
     442                 : 
     443                 : 
     444                 : #ifdef DCT_FLOAT_SUPPORTED
     445                 : 
     446                 : 
     447                 : METHODDEF(void)
     448               0 : convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col, FAST_FLOAT * workspace)
     449                 : {
     450                 :   register FAST_FLOAT *workspaceptr;
     451                 :   register JSAMPROW elemptr;
     452                 :   register int elemr;
     453                 : 
     454               0 :   workspaceptr = workspace;
     455               0 :   for (elemr = 0; elemr < DCTSIZE; elemr++) {
     456               0 :     elemptr = sample_data[elemr] + start_col;
     457                 : #if DCTSIZE == 8                /* unroll the inner loop */
     458               0 :     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
     459               0 :     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
     460               0 :     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
     461               0 :     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
     462               0 :     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
     463               0 :     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
     464               0 :     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
     465               0 :     *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
     466                 : #else
     467                 :     {
     468                 :       register int elemc;
     469                 :       for (elemc = DCTSIZE; elemc > 0; elemc--)
     470                 :         *workspaceptr++ = (FAST_FLOAT)
     471                 :                           (GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
     472                 :     }
     473                 : #endif
     474                 :   }
     475               0 : }
     476                 : 
     477                 : 
     478                 : METHODDEF(void)
     479               0 : quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors, FAST_FLOAT * workspace)
     480                 : {
     481                 :   register FAST_FLOAT temp;
     482                 :   register int i;
     483               0 :   register JCOEFPTR output_ptr = coef_block;
     484                 : 
     485               0 :   for (i = 0; i < DCTSIZE2; i++) {
     486                 :     /* Apply the quantization and scaling factor */
     487               0 :     temp = workspace[i] * divisors[i];
     488                 : 
     489                 :     /* Round to nearest integer.
     490                 :      * Since C does not specify the direction of rounding for negative
     491                 :      * quotients, we have to force the dividend positive for portability.
     492                 :      * The maximum coefficient size is +-16K (for 12-bit data), so this
     493                 :      * code should work for either 16-bit or 32-bit ints.
     494                 :      */
     495               0 :     output_ptr[i] = (JCOEF) ((int) (temp + (FAST_FLOAT) 16384.5) - 16384);
     496                 :   }
     497               0 : }
     498                 : 
     499                 : 
     500                 : METHODDEF(void)
     501               0 : forward_DCT_float (j_compress_ptr cinfo, jpeg_component_info * compptr,
     502                 :                    JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
     503                 :                    JDIMENSION start_row, JDIMENSION start_col,
     504                 :                    JDIMENSION num_blocks)
     505                 : /* This version is used for floating-point DCT implementations. */
     506                 : {
     507                 :   /* This routine is heavily used, so it's worth coding it tightly. */
     508               0 :   my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
     509               0 :   FAST_FLOAT * divisors = fdct->float_divisors[compptr->quant_tbl_no];
     510                 :   FAST_FLOAT * workspace;
     511                 :   JDIMENSION bi;
     512                 : 
     513                 : 
     514                 :   /* Make sure the compiler doesn't look up these every pass */
     515               0 :   float_DCT_method_ptr do_dct = fdct->float_dct;
     516               0 :   float_convsamp_method_ptr do_convsamp = fdct->float_convsamp;
     517               0 :   float_quantize_method_ptr do_quantize = fdct->float_quantize;
     518               0 :   workspace = fdct->float_workspace;
     519                 : 
     520               0 :   sample_data += start_row;     /* fold in the vertical offset once */
     521                 : 
     522               0 :   for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
     523                 :     /* Load data into workspace, applying unsigned->signed conversion */
     524               0 :     (*do_convsamp) (sample_data, start_col, workspace);
     525                 : 
     526                 :     /* Perform the DCT */
     527               0 :     (*do_dct) (workspace);
     528                 : 
     529                 :     /* Quantize/descale the coefficients, and store into coef_blocks[] */
     530               0 :     (*do_quantize) (coef_blocks[bi], divisors, workspace);
     531                 :   }
     532               0 : }
     533                 : 
     534                 : #endif /* DCT_FLOAT_SUPPORTED */
     535                 : 
     536                 : 
     537                 : /*
     538                 :  * Initialize FDCT manager.
     539                 :  */
     540                 : 
     541                 : GLOBAL(void)
     542               3 : jinit_forward_dct (j_compress_ptr cinfo)
     543                 : {
     544                 :   my_fdct_ptr fdct;
     545                 :   int i;
     546                 : 
     547               3 :   fdct = (my_fdct_ptr)
     548               3 :     (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
     549                 :                                 SIZEOF(my_fdct_controller));
     550               3 :   cinfo->fdct = (struct jpeg_forward_dct *) fdct;
     551               3 :   fdct->pub.start_pass = start_pass_fdctmgr;
     552                 : 
     553                 :   /* First determine the DCT... */
     554               3 :   switch (cinfo->dct_method) {
     555                 : #ifdef DCT_ISLOW_SUPPORTED
     556                 :   case JDCT_ISLOW:
     557               3 :     fdct->pub.forward_DCT = forward_DCT;
     558               3 :     if (jsimd_can_fdct_islow())
     559               3 :       fdct->dct = jsimd_fdct_islow;
     560                 :     else
     561               0 :       fdct->dct = jpeg_fdct_islow;
     562               3 :     break;
     563                 : #endif
     564                 : #ifdef DCT_IFAST_SUPPORTED
     565                 :   case JDCT_IFAST:
     566               0 :     fdct->pub.forward_DCT = forward_DCT;
     567               0 :     if (jsimd_can_fdct_ifast())
     568               0 :       fdct->dct = jsimd_fdct_ifast;
     569                 :     else
     570               0 :       fdct->dct = jpeg_fdct_ifast;
     571               0 :     break;
     572                 : #endif
     573                 : #ifdef DCT_FLOAT_SUPPORTED
     574                 :   case JDCT_FLOAT:
     575               0 :     fdct->pub.forward_DCT = forward_DCT_float;
     576               0 :     if (jsimd_can_fdct_float())
     577               0 :       fdct->float_dct = jsimd_fdct_float;
     578                 :     else
     579               0 :       fdct->float_dct = jpeg_fdct_float;
     580               0 :     break;
     581                 : #endif
     582                 :   default:
     583               0 :     ERREXIT(cinfo, JERR_NOT_COMPILED);
     584               0 :     break;
     585                 :   }
     586                 : 
     587                 :   /* ...then the supporting stages. */
     588               3 :   switch (cinfo->dct_method) {
     589                 : #ifdef DCT_ISLOW_SUPPORTED
     590                 :   case JDCT_ISLOW:
     591                 : #endif
     592                 : #ifdef DCT_IFAST_SUPPORTED
     593                 :   case JDCT_IFAST:
     594                 : #endif
     595                 : #if defined(DCT_ISLOW_SUPPORTED) || defined(DCT_IFAST_SUPPORTED)
     596               3 :     if (jsimd_can_convsamp())
     597               3 :       fdct->convsamp = jsimd_convsamp;
     598                 :     else
     599               0 :       fdct->convsamp = convsamp;
     600               3 :     if (jsimd_can_quantize())
     601               3 :       fdct->quantize = jsimd_quantize;
     602                 :     else
     603               0 :       fdct->quantize = quantize;
     604               3 :     break;
     605                 : #endif
     606                 : #ifdef DCT_FLOAT_SUPPORTED
     607                 :   case JDCT_FLOAT:
     608               0 :     if (jsimd_can_convsamp_float())
     609               0 :       fdct->float_convsamp = jsimd_convsamp_float;
     610                 :     else
     611               0 :       fdct->float_convsamp = convsamp_float;
     612               0 :     if (jsimd_can_quantize_float())
     613               0 :       fdct->float_quantize = jsimd_quantize_float;
     614                 :     else
     615               0 :       fdct->float_quantize = quantize_float;
     616               0 :     break;
     617                 : #endif
     618                 :   default:
     619               0 :     ERREXIT(cinfo, JERR_NOT_COMPILED);
     620               0 :     break;
     621                 :   }
     622                 : 
     623                 :   /* Allocate workspace memory */
     624                 : #ifdef DCT_FLOAT_SUPPORTED
     625               3 :   if (cinfo->dct_method == JDCT_FLOAT)
     626               0 :     fdct->float_workspace = (FAST_FLOAT *)
     627               0 :       (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
     628                 :                                   SIZEOF(FAST_FLOAT) * DCTSIZE2);
     629                 :   else
     630                 : #endif
     631               3 :     fdct->workspace = (DCTELEM *)
     632               3 :       (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
     633                 :                                   SIZEOF(DCTELEM) * DCTSIZE2);
     634                 : 
     635                 :   /* Mark divisor tables unallocated */
     636              15 :   for (i = 0; i < NUM_QUANT_TBLS; i++) {
     637              12 :     fdct->divisors[i] = NULL;
     638                 : #ifdef DCT_FLOAT_SUPPORTED
     639              12 :     fdct->float_divisors[i] = NULL;
     640                 : #endif
     641                 :   }
     642               3 : }

Generated by: LCOV version 1.7