• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * jcdctmgr.c
3  *
4  * This file was part of the Independent JPEG Group's software:
5  * Copyright (C) 1994-1996, Thomas G. Lane.
6  * libjpeg-turbo Modifications:
7  * Copyright (C) 1999-2006, MIYASAKA Masaru.
8  * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
9  * Copyright (C) 2011, 2014-2015, D. R. Commander.
10  * For conditions of distribution and use, see the accompanying README.ijg
11  * file.
12  *
13  * This file contains the forward-DCT management logic.
14  * This code selects a particular DCT implementation to be used,
15  * and it performs related housekeeping chores including coefficient
16  * quantization.
17  */
18 
19 #define JPEG_INTERNALS
20 #include "jinclude.h"
21 #include "jpeglib.h"
22 #include "jdct.h"               /* Private declarations for DCT subsystem */
23 #include "jsimddct.h"
24 
25 
26 /* Private subobject for this module */
27 
28 typedef void (*forward_DCT_method_ptr) (DCTELEM *data);
29 typedef void (*float_DCT_method_ptr) (FAST_FLOAT *data);
30 
31 typedef void (*convsamp_method_ptr) (JSAMPARRAY sample_data,
32                                      JDIMENSION start_col,
33                                      DCTELEM *workspace);
34 typedef void (*float_convsamp_method_ptr) (JSAMPARRAY sample_data,
35                                            JDIMENSION start_col,
36                                            FAST_FLOAT *workspace);
37 
38 typedef void (*quantize_method_ptr) (JCOEFPTR coef_block, DCTELEM *divisors,
39                                      DCTELEM *workspace);
40 typedef void (*float_quantize_method_ptr) (JCOEFPTR coef_block,
41                                            FAST_FLOAT *divisors,
42                                            FAST_FLOAT *workspace);
43 
44 METHODDEF(void) quantize(JCOEFPTR, DCTELEM *, DCTELEM *);
45 
46 typedef struct {
47   struct jpeg_forward_dct pub;  /* public fields */
48 
49   /* Pointer to the DCT routine actually in use */
50   forward_DCT_method_ptr dct;
51   convsamp_method_ptr convsamp;
52   quantize_method_ptr quantize;
53 
54   /* The actual post-DCT divisors --- not identical to the quant table
55    * entries, because of scaling (especially for an unnormalized DCT).
56    * Each table is given in normal array order.
57    */
58   DCTELEM *divisors[NUM_QUANT_TBLS];
59 
60   /* work area for FDCT subroutine */
61   DCTELEM *workspace;
62 
63 #ifdef DCT_FLOAT_SUPPORTED
64   /* Same as above for the floating-point case. */
65   float_DCT_method_ptr float_dct;
66   float_convsamp_method_ptr float_convsamp;
67   float_quantize_method_ptr float_quantize;
68   FAST_FLOAT *float_divisors[NUM_QUANT_TBLS];
69   FAST_FLOAT *float_workspace;
70 #endif
71 } my_fdct_controller;
72 
73 typedef my_fdct_controller *my_fdct_ptr;
74 
75 
76 #if BITS_IN_JSAMPLE == 8
77 
78 /*
79  * Find the highest bit in an integer through binary search.
80  */
81 
82 LOCAL(int)
flss(UINT16 val)83 flss(UINT16 val)
84 {
85   int bit;
86 
87   bit = 16;
88 
89   if (!val)
90     return 0;
91 
92   if (!(val & 0xff00)) {
93     bit -= 8;
94     val <<= 8;
95   }
96   if (!(val & 0xf000)) {
97     bit -= 4;
98     val <<= 4;
99   }
100   if (!(val & 0xc000)) {
101     bit -= 2;
102     val <<= 2;
103   }
104   if (!(val & 0x8000)) {
105     bit -= 1;
106     val <<= 1;
107   }
108 
109   return bit;
110 }
111 
112 
113 /*
114  * Compute values to do a division using reciprocal.
115  *
116  * This implementation is based on an algorithm described in
117  *   "How to optimize for the Pentium family of microprocessors"
118  *   (http://www.agner.org/assem/).
119  * More information about the basic algorithm can be found in
120  * the paper "Integer Division Using Reciprocals" by Robert Alverson.
121  *
122  * The basic idea is to replace x/d by x * d^-1. In order to store
123  * d^-1 with enough precision we shift it left a few places. It turns
124  * out that this algoright gives just enough precision, and also fits
125  * into DCTELEM:
126  *
127  *   b = (the number of significant bits in divisor) - 1
128  *   r = (word size) + b
129  *   f = 2^r / divisor
130  *
131  * f will not be an integer for most cases, so we need to compensate
132  * for the rounding error introduced:
133  *
134  *   no fractional part:
135  *
136  *       result = input >> r
137  *
138  *   fractional part of f < 0.5:
139  *
140  *       round f down to nearest integer
141  *       result = ((input + 1) * f) >> r
142  *
143  *   fractional part of f > 0.5:
144  *
145  *       round f up to nearest integer
146  *       result = (input * f) >> r
147  *
148  * This is the original algorithm that gives truncated results. But we
149  * want properly rounded results, so we replace "input" with
150  * "input + divisor/2".
151  *
152  * In order to allow SIMD implementations we also tweak the values to
153  * allow the same calculation to be made at all times:
154  *
155  *   dctbl[0] = f rounded to nearest integer
156  *   dctbl[1] = divisor / 2 (+ 1 if fractional part of f < 0.5)
157  *   dctbl[2] = 1 << ((word size) * 2 - r)
158  *   dctbl[3] = r - (word size)
159  *
160  * dctbl[2] is for stupid instruction sets where the shift operation
161  * isn't member wise (e.g. MMX).
162  *
163  * The reason dctbl[2] and dctbl[3] reduce the shift with (word size)
164  * is that most SIMD implementations have a "multiply and store top
165  * half" operation.
166  *
167  * Lastly, we store each of the values in their own table instead
168  * of in a consecutive manner, yet again in order to allow SIMD
169  * routines.
170  */
171 
172 LOCAL(int)
compute_reciprocal(UINT16 divisor,DCTELEM * dtbl)173 compute_reciprocal(UINT16 divisor, DCTELEM *dtbl)
174 {
175   UDCTELEM2 fq, fr;
176   UDCTELEM c;
177   int b, r;
178 
179   if (divisor == 1) {
180     /* divisor == 1 means unquantized, so these reciprocal/correction/shift
181      * values will cause the C quantization algorithm to act like the
182      * identity function.  Since only the C quantization algorithm is used in
183      * these cases, the scale value is irrelevant.
184      */
185     dtbl[DCTSIZE2 * 0] = (DCTELEM)1;                        /* reciprocal */
186     dtbl[DCTSIZE2 * 1] = (DCTELEM)0;                        /* correction */
187     dtbl[DCTSIZE2 * 2] = (DCTELEM)1;                        /* scale */
188     dtbl[DCTSIZE2 * 3] = -(DCTELEM)(sizeof(DCTELEM) * 8);   /* shift */
189     return 0;
190   }
191 
192   b = flss(divisor) - 1;
193   r  = sizeof(DCTELEM) * 8 + b;
194 
195   fq = ((UDCTELEM2)1 << r) / divisor;
196   fr = ((UDCTELEM2)1 << r) % divisor;
197 
198   c = divisor / 2;                      /* for rounding */
199 
200   if (fr == 0) {                        /* divisor is power of two */
201     /* fq will be one bit too large to fit in DCTELEM, so adjust */
202     fq >>= 1;
203     r--;
204   } else if (fr <= (divisor / 2U)) {    /* fractional part is < 0.5 */
205     c++;
206   } else {                              /* fractional part is > 0.5 */
207     fq++;
208   }
209 
210   dtbl[DCTSIZE2 * 0] = (DCTELEM)fq;     /* reciprocal */
211   dtbl[DCTSIZE2 * 1] = (DCTELEM)c;      /* correction + roundfactor */
212 #ifdef WITH_SIMD
213   dtbl[DCTSIZE2 * 2] = (DCTELEM)(1 << (sizeof(DCTELEM) * 8 * 2 - r)); /* scale */
214 #else
215   dtbl[DCTSIZE2 * 2] = 1;
216 #endif
217   dtbl[DCTSIZE2 * 3] = (DCTELEM)r - sizeof(DCTELEM) * 8; /* shift */
218 
219   if (r <= 16) return 0;
220   else return 1;
221 }
222 
223 #endif
224 
225 
226 /*
227  * Initialize for a processing pass.
228  * Verify that all referenced Q-tables are present, and set up
229  * the divisor table for each one.
230  * In the current implementation, DCT of all components is done during
231  * the first pass, even if only some components will be output in the
232  * first scan.  Hence all components should be examined here.
233  */
234 
235 METHODDEF(void)
start_pass_fdctmgr(j_compress_ptr cinfo)236 start_pass_fdctmgr(j_compress_ptr cinfo)
237 {
238   my_fdct_ptr fdct = (my_fdct_ptr)cinfo->fdct;
239   int ci, qtblno, i;
240   jpeg_component_info *compptr;
241   JQUANT_TBL *qtbl;
242   DCTELEM *dtbl;
243 
244   for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
245        ci++, compptr++) {
246     qtblno = compptr->quant_tbl_no;
247     /* Make sure specified quantization table is present */
248     if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
249         cinfo->quant_tbl_ptrs[qtblno] == NULL)
250       ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
251     qtbl = cinfo->quant_tbl_ptrs[qtblno];
252     /* Compute divisors for this quant table */
253     /* We may do this more than once for same table, but it's not a big deal */
254     switch (cinfo->dct_method) {
255 #ifdef DCT_ISLOW_SUPPORTED
256     case JDCT_ISLOW:
257       /* For LL&M IDCT method, divisors are equal to raw quantization
258        * coefficients multiplied by 8 (to counteract scaling).
259        */
260       if (fdct->divisors[qtblno] == NULL) {
261         fdct->divisors[qtblno] = (DCTELEM *)
262           (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
263                                       (DCTSIZE2 * 4) * sizeof(DCTELEM));
264       }
265       dtbl = fdct->divisors[qtblno];
266       for (i = 0; i < DCTSIZE2; i++) {
267 #if BITS_IN_JSAMPLE == 8
268         if (!compute_reciprocal(qtbl->quantval[i] << 3, &dtbl[i]) &&
269             fdct->quantize == jsimd_quantize)
270           fdct->quantize = quantize;
271 #else
272         dtbl[i] = ((DCTELEM)qtbl->quantval[i]) << 3;
273 #endif
274       }
275       break;
276 #endif
277 #ifdef DCT_IFAST_SUPPORTED
278     case JDCT_IFAST:
279       {
280         /* For AA&N IDCT method, divisors are equal to quantization
281          * coefficients scaled by scalefactor[row]*scalefactor[col], where
282          *   scalefactor[0] = 1
283          *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
284          * We apply a further scale factor of 8.
285          */
286 #define CONST_BITS  14
287         static const INT16 aanscales[DCTSIZE2] = {
288           /* precomputed values scaled up by 14 bits */
289           16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
290           22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
291           21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
292           19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
293           16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
294           12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
295            8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
296            4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
297         };
298         SHIFT_TEMPS
299 
300         if (fdct->divisors[qtblno] == NULL) {
301           fdct->divisors[qtblno] = (DCTELEM *)
302             (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
303                                         (DCTSIZE2 * 4) * sizeof(DCTELEM));
304         }
305         dtbl = fdct->divisors[qtblno];
306         for (i = 0; i < DCTSIZE2; i++) {
307 #if BITS_IN_JSAMPLE == 8
308           if (!compute_reciprocal(
309                 DESCALE(MULTIPLY16V16((JLONG)qtbl->quantval[i],
310                                       (JLONG)aanscales[i]),
311                         CONST_BITS - 3), &dtbl[i]) &&
312               fdct->quantize == jsimd_quantize)
313             fdct->quantize = quantize;
314 #else
315           dtbl[i] = (DCTELEM)
316             DESCALE(MULTIPLY16V16((JLONG)qtbl->quantval[i],
317                                   (JLONG)aanscales[i]),
318                     CONST_BITS - 3);
319 #endif
320         }
321       }
322       break;
323 #endif
324 #ifdef DCT_FLOAT_SUPPORTED
325     case JDCT_FLOAT:
326       {
327         /* For float AA&N IDCT method, divisors are equal to quantization
328          * coefficients scaled by scalefactor[row]*scalefactor[col], where
329          *   scalefactor[0] = 1
330          *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
331          * We apply a further scale factor of 8.
332          * What's actually stored is 1/divisor so that the inner loop can
333          * use a multiplication rather than a division.
334          */
335         FAST_FLOAT *fdtbl;
336         int row, col;
337         static const double aanscalefactor[DCTSIZE] = {
338           1.0, 1.387039845, 1.306562965, 1.175875602,
339           1.0, 0.785694958, 0.541196100, 0.275899379
340         };
341 
342         if (fdct->float_divisors[qtblno] == NULL) {
343           fdct->float_divisors[qtblno] = (FAST_FLOAT *)
344             (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
345                                         DCTSIZE2 * sizeof(FAST_FLOAT));
346         }
347         fdtbl = fdct->float_divisors[qtblno];
348         i = 0;
349         for (row = 0; row < DCTSIZE; row++) {
350           for (col = 0; col < DCTSIZE; col++) {
351             fdtbl[i] = (FAST_FLOAT)
352               (1.0 / (((double)qtbl->quantval[i] *
353                        aanscalefactor[row] * aanscalefactor[col] * 8.0)));
354             i++;
355           }
356         }
357       }
358       break;
359 #endif
360     default:
361       ERREXIT(cinfo, JERR_NOT_COMPILED);
362       break;
363     }
364   }
365 }
366 
367 
368 /*
369  * Load data into workspace, applying unsigned->signed conversion.
370  */
371 
372 METHODDEF(void)
convsamp(JSAMPARRAY sample_data,JDIMENSION start_col,DCTELEM * workspace)373 convsamp(JSAMPARRAY sample_data, JDIMENSION start_col, DCTELEM *workspace)
374 {
375   register DCTELEM *workspaceptr;
376   register JSAMPROW elemptr;
377   register int elemr;
378 
379   workspaceptr = workspace;
380   for (elemr = 0; elemr < DCTSIZE; elemr++) {
381     elemptr = sample_data[elemr] + start_col;
382 
383 #if DCTSIZE == 8                /* unroll the inner loop */
384     *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
385     *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
386     *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
387     *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
388     *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
389     *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
390     *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
391     *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
392 #else
393     {
394       register int elemc;
395       for (elemc = DCTSIZE; elemc > 0; elemc--)
396         *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
397     }
398 #endif
399   }
400 }
401 
402 
403 /*
404  * Quantize/descale the coefficients, and store into coef_blocks[].
405  */
406 
407 METHODDEF(void)
quantize(JCOEFPTR coef_block,DCTELEM * divisors,DCTELEM * workspace)408 quantize(JCOEFPTR coef_block, DCTELEM *divisors, DCTELEM *workspace)
409 {
410   int i;
411   DCTELEM temp;
412   JCOEFPTR output_ptr = coef_block;
413 
414 #if BITS_IN_JSAMPLE == 8
415 
416   UDCTELEM recip, corr;
417   int shift;
418   UDCTELEM2 product;
419 
420   for (i = 0; i < DCTSIZE2; i++) {
421     temp = workspace[i];
422     recip = divisors[i + DCTSIZE2 * 0];
423     corr =  divisors[i + DCTSIZE2 * 1];
424     shift = divisors[i + DCTSIZE2 * 3];
425 
426     if (temp < 0) {
427       temp = -temp;
428       product = (UDCTELEM2)(temp + corr) * recip;
429       product >>= shift + sizeof(DCTELEM) * 8;
430       temp = (DCTELEM)product;
431       temp = -temp;
432     } else {
433       product = (UDCTELEM2)(temp + corr) * recip;
434       product >>= shift + sizeof(DCTELEM) * 8;
435       temp = (DCTELEM)product;
436     }
437     output_ptr[i] = (JCOEF)temp;
438   }
439 
440 #else
441 
442   register DCTELEM qval;
443 
444   for (i = 0; i < DCTSIZE2; i++) {
445     qval = divisors[i];
446     temp = workspace[i];
447     /* Divide the coefficient value by qval, ensuring proper rounding.
448      * Since C does not specify the direction of rounding for negative
449      * quotients, we have to force the dividend positive for portability.
450      *
451      * In most files, at least half of the output values will be zero
452      * (at default quantization settings, more like three-quarters...)
453      * so we should ensure that this case is fast.  On many machines,
454      * a comparison is enough cheaper than a divide to make a special test
455      * a win.  Since both inputs will be nonnegative, we need only test
456      * for a < b to discover whether a/b is 0.
457      * If your machine's division is fast enough, define FAST_DIVIDE.
458      */
459 #ifdef FAST_DIVIDE
460 #define DIVIDE_BY(a, b)  a /= b
461 #else
462 #define DIVIDE_BY(a, b)  if (a >= b) a /= b;  else a = 0
463 #endif
464     if (temp < 0) {
465       temp = -temp;
466       temp += qval >> 1;        /* for rounding */
467       DIVIDE_BY(temp, qval);
468       temp = -temp;
469     } else {
470       temp += qval >> 1;        /* for rounding */
471       DIVIDE_BY(temp, qval);
472     }
473     output_ptr[i] = (JCOEF)temp;
474   }
475 
476 #endif
477 
478 }
479 
480 
481 /*
482  * Perform forward DCT on one or more blocks of a component.
483  *
484  * The input samples are taken from the sample_data[] array starting at
485  * position start_row/start_col, and moving to the right for any additional
486  * blocks. The quantized coefficients are returned in coef_blocks[].
487  */
488 
489 METHODDEF(void)
forward_DCT(j_compress_ptr cinfo,jpeg_component_info * compptr,JSAMPARRAY sample_data,JBLOCKROW coef_blocks,JDIMENSION start_row,JDIMENSION start_col,JDIMENSION num_blocks)490 forward_DCT(j_compress_ptr cinfo, jpeg_component_info *compptr,
491             JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
492             JDIMENSION start_row, JDIMENSION start_col, JDIMENSION num_blocks)
493 /* This version is used for integer DCT implementations. */
494 {
495   /* This routine is heavily used, so it's worth coding it tightly. */
496   my_fdct_ptr fdct = (my_fdct_ptr)cinfo->fdct;
497   DCTELEM *divisors = fdct->divisors[compptr->quant_tbl_no];
498   DCTELEM *workspace;
499   JDIMENSION bi;
500 
501   /* Make sure the compiler doesn't look up these every pass */
502   forward_DCT_method_ptr do_dct = fdct->dct;
503   convsamp_method_ptr do_convsamp = fdct->convsamp;
504   quantize_method_ptr do_quantize = fdct->quantize;
505   workspace = fdct->workspace;
506 
507   sample_data += start_row;     /* fold in the vertical offset once */
508 
509   for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
510     /* Load data into workspace, applying unsigned->signed conversion */
511     (*do_convsamp) (sample_data, start_col, workspace);
512 
513     /* Perform the DCT */
514     (*do_dct) (workspace);
515 
516     /* Quantize/descale the coefficients, and store into coef_blocks[] */
517     (*do_quantize) (coef_blocks[bi], divisors, workspace);
518   }
519 }
520 
521 
522 #ifdef DCT_FLOAT_SUPPORTED
523 
524 METHODDEF(void)
convsamp_float(JSAMPARRAY sample_data,JDIMENSION start_col,FAST_FLOAT * workspace)525 convsamp_float(JSAMPARRAY sample_data, JDIMENSION start_col,
526                FAST_FLOAT *workspace)
527 {
528   register FAST_FLOAT *workspaceptr;
529   register JSAMPROW elemptr;
530   register int elemr;
531 
532   workspaceptr = workspace;
533   for (elemr = 0; elemr < DCTSIZE; elemr++) {
534     elemptr = sample_data[elemr] + start_col;
535 #if DCTSIZE == 8                /* unroll the inner loop */
536     *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
537     *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
538     *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
539     *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
540     *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
541     *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
542     *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
543     *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
544 #else
545     {
546       register int elemc;
547       for (elemc = DCTSIZE; elemc > 0; elemc--)
548         *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
549     }
550 #endif
551   }
552 }
553 
554 
555 METHODDEF(void)
quantize_float(JCOEFPTR coef_block,FAST_FLOAT * divisors,FAST_FLOAT * workspace)556 quantize_float(JCOEFPTR coef_block, FAST_FLOAT *divisors,
557                FAST_FLOAT *workspace)
558 {
559   register FAST_FLOAT temp;
560   register int i;
561   register JCOEFPTR output_ptr = coef_block;
562 
563   for (i = 0; i < DCTSIZE2; i++) {
564     /* Apply the quantization and scaling factor */
565     temp = workspace[i] * divisors[i];
566 
567     /* Round to nearest integer.
568      * Since C does not specify the direction of rounding for negative
569      * quotients, we have to force the dividend positive for portability.
570      * The maximum coefficient size is +-16K (for 12-bit data), so this
571      * code should work for either 16-bit or 32-bit ints.
572      */
573     output_ptr[i] = (JCOEF)((int)(temp + (FAST_FLOAT)16384.5) - 16384);
574   }
575 }
576 
577 
578 METHODDEF(void)
forward_DCT_float(j_compress_ptr cinfo,jpeg_component_info * compptr,JSAMPARRAY sample_data,JBLOCKROW coef_blocks,JDIMENSION start_row,JDIMENSION start_col,JDIMENSION num_blocks)579 forward_DCT_float(j_compress_ptr cinfo, jpeg_component_info *compptr,
580                   JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
581                   JDIMENSION start_row, JDIMENSION start_col,
582                   JDIMENSION num_blocks)
583 /* This version is used for floating-point DCT implementations. */
584 {
585   /* This routine is heavily used, so it's worth coding it tightly. */
586   my_fdct_ptr fdct = (my_fdct_ptr)cinfo->fdct;
587   FAST_FLOAT *divisors = fdct->float_divisors[compptr->quant_tbl_no];
588   FAST_FLOAT *workspace;
589   JDIMENSION bi;
590 
591 
592   /* Make sure the compiler doesn't look up these every pass */
593   float_DCT_method_ptr do_dct = fdct->float_dct;
594   float_convsamp_method_ptr do_convsamp = fdct->float_convsamp;
595   float_quantize_method_ptr do_quantize = fdct->float_quantize;
596   workspace = fdct->float_workspace;
597 
598   sample_data += start_row;     /* fold in the vertical offset once */
599 
600   for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
601     /* Load data into workspace, applying unsigned->signed conversion */
602     (*do_convsamp) (sample_data, start_col, workspace);
603 
604     /* Perform the DCT */
605     (*do_dct) (workspace);
606 
607     /* Quantize/descale the coefficients, and store into coef_blocks[] */
608     (*do_quantize) (coef_blocks[bi], divisors, workspace);
609   }
610 }
611 
612 #endif /* DCT_FLOAT_SUPPORTED */
613 
614 
615 /*
616  * Initialize FDCT manager.
617  */
618 
619 GLOBAL(void)
jinit_forward_dct(j_compress_ptr cinfo)620 jinit_forward_dct(j_compress_ptr cinfo)
621 {
622   my_fdct_ptr fdct;
623   int i;
624 
625   fdct = (my_fdct_ptr)
626     (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
627                                 sizeof(my_fdct_controller));
628   cinfo->fdct = (struct jpeg_forward_dct *)fdct;
629   fdct->pub.start_pass = start_pass_fdctmgr;
630 
631   /* First determine the DCT... */
632   switch (cinfo->dct_method) {
633 #ifdef DCT_ISLOW_SUPPORTED
634   case JDCT_ISLOW:
635     fdct->pub.forward_DCT = forward_DCT;
636     if (jsimd_can_fdct_islow())
637       fdct->dct = jsimd_fdct_islow;
638     else
639       fdct->dct = jpeg_fdct_islow;
640     break;
641 #endif
642 #ifdef DCT_IFAST_SUPPORTED
643   case JDCT_IFAST:
644     fdct->pub.forward_DCT = forward_DCT;
645     if (jsimd_can_fdct_ifast())
646       fdct->dct = jsimd_fdct_ifast;
647     else
648       fdct->dct = jpeg_fdct_ifast;
649     break;
650 #endif
651 #ifdef DCT_FLOAT_SUPPORTED
652   case JDCT_FLOAT:
653     fdct->pub.forward_DCT = forward_DCT_float;
654     if (jsimd_can_fdct_float())
655       fdct->float_dct = jsimd_fdct_float;
656     else
657       fdct->float_dct = jpeg_fdct_float;
658     break;
659 #endif
660   default:
661     ERREXIT(cinfo, JERR_NOT_COMPILED);
662     break;
663   }
664 
665   /* ...then the supporting stages. */
666   switch (cinfo->dct_method) {
667 #ifdef DCT_ISLOW_SUPPORTED
668   case JDCT_ISLOW:
669 #endif
670 #ifdef DCT_IFAST_SUPPORTED
671   case JDCT_IFAST:
672 #endif
673 #if defined(DCT_ISLOW_SUPPORTED) || defined(DCT_IFAST_SUPPORTED)
674     if (jsimd_can_convsamp())
675       fdct->convsamp = jsimd_convsamp;
676     else
677       fdct->convsamp = convsamp;
678     if (jsimd_can_quantize())
679       fdct->quantize = jsimd_quantize;
680     else
681       fdct->quantize = quantize;
682     break;
683 #endif
684 #ifdef DCT_FLOAT_SUPPORTED
685   case JDCT_FLOAT:
686     if (jsimd_can_convsamp_float())
687       fdct->float_convsamp = jsimd_convsamp_float;
688     else
689       fdct->float_convsamp = convsamp_float;
690     if (jsimd_can_quantize_float())
691       fdct->float_quantize = jsimd_quantize_float;
692     else
693       fdct->float_quantize = quantize_float;
694     break;
695 #endif
696   default:
697     ERREXIT(cinfo, JERR_NOT_COMPILED);
698     break;
699   }
700 
701   /* Allocate workspace memory */
702 #ifdef DCT_FLOAT_SUPPORTED
703   if (cinfo->dct_method == JDCT_FLOAT)
704     fdct->float_workspace = (FAST_FLOAT *)
705       (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
706                                   sizeof(FAST_FLOAT) * DCTSIZE2);
707   else
708 #endif
709     fdct->workspace = (DCTELEM *)
710       (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
711                                   sizeof(DCTELEM) * DCTSIZE2);
712 
713   /* Mark divisor tables unallocated */
714   for (i = 0; i < NUM_QUANT_TBLS; i++) {
715     fdct->divisors[i] = NULL;
716 #ifdef DCT_FLOAT_SUPPORTED
717     fdct->float_divisors[i] = NULL;
718 #endif
719   }
720 }
721