1/*
2 * jidctred.c
3 *
4 * Copyright (C) 1994-1998, Thomas G. Lane.
5 * This file is part of the Independent JPEG Group's software.
6 * For conditions of distribution and use, see the accompanying README file.
7 *
8 * This file contains inverse-DCT routines that produce reduced-size output:
9 * either 4x4, 2x2, or 1x1 pixels from an 8x8 DCT block.
10 *
11 * The implementation is based on the Loeffler, Ligtenberg and Moschytz (LL&M)
12 * algorithm used in jidctint.c.  We simply replace each 8-to-8 1-D IDCT step
13 * with an 8-to-4 step that produces the four averages of two adjacent outputs
14 * (or an 8-to-2 step producing two averages of four outputs, for 2x2 output).
15 * These steps were derived by computing the corresponding values at the end
16 * of the normal LL&M code, then simplifying as much as possible.
17 *
18 * 1x1 is trivial: just take the DC coefficient divided by 8.
19 *
20 * See jidctint.c for additional comments.
21 */
22
23#define JPEG_INTERNALS
24#include "jinclude.h"
25#include "jpeglib.h"
26#include "jdct.h"		/* Private declarations for DCT subsystem */
27
28#ifdef IDCT_SCALING_SUPPORTED
29
30
31/*
32 * This module is specialized to the case DCTSIZE = 8.
33 */
34
35#if DCTSIZE != 8
36  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
37#endif
38
39
40/* Scaling is the same as in jidctint.c. */
41
42#if BITS_IN_JSAMPLE == 8
43#define CONST_BITS  13
44#define PASS1_BITS  2
45#else
46#define CONST_BITS  13
47#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
48#endif
49
50/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
51 * causing a lot of useless floating-point operations at run time.
52 * To get around this we use the following pre-calculated constants.
53 * If you change CONST_BITS you may want to add appropriate values.
54 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
55 */
56
57#if CONST_BITS == 13
58#define FIX_0_211164243  ((INT32)  1730)	/* FIX(0.211164243) */
59#define FIX_0_509795579  ((INT32)  4176)	/* FIX(0.509795579) */
60#define FIX_0_601344887  ((INT32)  4926)	/* FIX(0.601344887) */
61#define FIX_0_720959822  ((INT32)  5906)	/* FIX(0.720959822) */
62#define FIX_0_765366865  ((INT32)  6270)	/* FIX(0.765366865) */
63#define FIX_0_850430095  ((INT32)  6967)	/* FIX(0.850430095) */
64#define FIX_0_899976223  ((INT32)  7373)	/* FIX(0.899976223) */
65#define FIX_1_061594337  ((INT32)  8697)	/* FIX(1.061594337) */
66#define FIX_1_272758580  ((INT32)  10426)	/* FIX(1.272758580) */
67#define FIX_1_451774981  ((INT32)  11893)	/* FIX(1.451774981) */
68#define FIX_1_847759065  ((INT32)  15137)	/* FIX(1.847759065) */
69#define FIX_2_172734803  ((INT32)  17799)	/* FIX(2.172734803) */
70#define FIX_2_562915447  ((INT32)  20995)	/* FIX(2.562915447) */
71#define FIX_3_624509785  ((INT32)  29692)	/* FIX(3.624509785) */
72#else
73#define FIX_0_211164243  FIX(0.211164243)
74#define FIX_0_509795579  FIX(0.509795579)
75#define FIX_0_601344887  FIX(0.601344887)
76#define FIX_0_720959822  FIX(0.720959822)
77#define FIX_0_765366865  FIX(0.765366865)
78#define FIX_0_850430095  FIX(0.850430095)
79#define FIX_0_899976223  FIX(0.899976223)
80#define FIX_1_061594337  FIX(1.061594337)
81#define FIX_1_272758580  FIX(1.272758580)
82#define FIX_1_451774981  FIX(1.451774981)
83#define FIX_1_847759065  FIX(1.847759065)
84#define FIX_2_172734803  FIX(2.172734803)
85#define FIX_2_562915447  FIX(2.562915447)
86#define FIX_3_624509785  FIX(3.624509785)
87#endif
88
89
90/* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
91 * For 8-bit samples with the recommended scaling, all the variable
92 * and constant values involved are no more than 16 bits wide, so a
93 * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
94 * For 12-bit samples, a full 32-bit multiplication will be needed.
95 */
96
97#if BITS_IN_JSAMPLE == 8
98#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
99#else
100#define MULTIPLY(var,const)  ((var) * (const))
101#endif
102
103
104/* Dequantize a coefficient by multiplying it by the multiplier-table
105 * entry; produce an int result.  In this module, both inputs and result
106 * are 16 bits or less, so either int or short multiply will work.
107 */
108
109#define DEQUANTIZE(coef,quantval)  (((ISLOW_MULT_TYPE) (coef)) * (quantval))
110
111
112/*
113 * Perform dequantization and inverse DCT on one block of coefficients,
114 * producing a reduced-size 4x4 output block.
115 */
116
117GLOBAL(void)
118jpeg_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
119	       JCOEFPTR coef_block,
120	       JSAMPARRAY output_buf, JDIMENSION output_col)
121{
122  INT32 tmp0, tmp2, tmp10, tmp12;
123  INT32 z1, z2, z3, z4;
124  JCOEFPTR inptr;
125  ISLOW_MULT_TYPE * quantptr;
126  int * wsptr;
127  JSAMPROW outptr;
128  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
129  int ctr;
130  int workspace[DCTSIZE*4];	/* buffers data between passes */
131  SHIFT_TEMPS
132
133  /* Pass 1: process columns from input, store into work array. */
134
135  inptr = coef_block;
136  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
137  wsptr = workspace;
138  for (ctr = DCTSIZE; ctr > 0; inptr++, quantptr++, wsptr++, ctr--) {
139    /* Don't bother to process column 4, because second pass won't use it */
140    if (ctr == DCTSIZE-4)
141      continue;
142    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
143	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*5] == 0 &&
144	inptr[DCTSIZE*6] == 0 && inptr[DCTSIZE*7] == 0) {
145      /* AC terms all zero; we need not examine term 4 for 4x4 output */
146      int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
147
148      wsptr[DCTSIZE*0] = dcval;
149      wsptr[DCTSIZE*1] = dcval;
150      wsptr[DCTSIZE*2] = dcval;
151      wsptr[DCTSIZE*3] = dcval;
152
153      continue;
154    }
155
156    /* Even part */
157
158    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
159    tmp0 <<= (CONST_BITS+1);
160
161    z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
162    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
163
164    tmp2 = MULTIPLY(z2, FIX_1_847759065) + MULTIPLY(z3, - FIX_0_765366865);
165
166    tmp10 = tmp0 + tmp2;
167    tmp12 = tmp0 - tmp2;
168
169    /* Odd part */
170
171    z1 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
172    z2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
173    z3 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
174    z4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
175
176    tmp0 = MULTIPLY(z1, - FIX_0_211164243) /* sqrt(2) * (c3-c1) */
177	 + MULTIPLY(z2, FIX_1_451774981) /* sqrt(2) * (c3+c7) */
178	 + MULTIPLY(z3, - FIX_2_172734803) /* sqrt(2) * (-c1-c5) */
179	 + MULTIPLY(z4, FIX_1_061594337); /* sqrt(2) * (c5+c7) */
180
181    tmp2 = MULTIPLY(z1, - FIX_0_509795579) /* sqrt(2) * (c7-c5) */
182	 + MULTIPLY(z2, - FIX_0_601344887) /* sqrt(2) * (c5-c1) */
183	 + MULTIPLY(z3, FIX_0_899976223) /* sqrt(2) * (c3-c7) */
184	 + MULTIPLY(z4, FIX_2_562915447); /* sqrt(2) * (c1+c3) */
185
186    /* Final output stage */
187
188    wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp2, CONST_BITS-PASS1_BITS+1);
189    wsptr[DCTSIZE*3] = (int) DESCALE(tmp10 - tmp2, CONST_BITS-PASS1_BITS+1);
190    wsptr[DCTSIZE*1] = (int) DESCALE(tmp12 + tmp0, CONST_BITS-PASS1_BITS+1);
191    wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 - tmp0, CONST_BITS-PASS1_BITS+1);
192  }
193
194  /* Pass 2: process 4 rows from work array, store into output array. */
195
196  wsptr = workspace;
197  for (ctr = 0; ctr < 4; ctr++) {
198    outptr = output_buf[ctr] + output_col;
199    /* It's not clear whether a zero row test is worthwhile here ... */
200
201#ifndef NO_ZERO_ROW_TEST
202    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 &&
203	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
204      /* AC terms all zero */
205      JSAMPLE dcval = range_limit[(int) DESCALE((INT32) wsptr[0], PASS1_BITS+3)
206				  & RANGE_MASK];
207
208      outptr[0] = dcval;
209      outptr[1] = dcval;
210      outptr[2] = dcval;
211      outptr[3] = dcval;
212
213      wsptr += DCTSIZE;		/* advance pointer to next row */
214      continue;
215    }
216#endif
217
218    /* Even part */
219
220    tmp0 = ((INT32) wsptr[0]) << (CONST_BITS+1);
221
222    tmp2 = MULTIPLY((INT32) wsptr[2], FIX_1_847759065)
223	 + MULTIPLY((INT32) wsptr[6], - FIX_0_765366865);
224
225    tmp10 = tmp0 + tmp2;
226    tmp12 = tmp0 - tmp2;
227
228    /* Odd part */
229
230    z1 = (INT32) wsptr[7];
231    z2 = (INT32) wsptr[5];
232    z3 = (INT32) wsptr[3];
233    z4 = (INT32) wsptr[1];
234
235    tmp0 = MULTIPLY(z1, - FIX_0_211164243) /* sqrt(2) * (c3-c1) */
236	 + MULTIPLY(z2, FIX_1_451774981) /* sqrt(2) * (c3+c7) */
237	 + MULTIPLY(z3, - FIX_2_172734803) /* sqrt(2) * (-c1-c5) */
238	 + MULTIPLY(z4, FIX_1_061594337); /* sqrt(2) * (c5+c7) */
239
240    tmp2 = MULTIPLY(z1, - FIX_0_509795579) /* sqrt(2) * (c7-c5) */
241	 + MULTIPLY(z2, - FIX_0_601344887) /* sqrt(2) * (c5-c1) */
242	 + MULTIPLY(z3, FIX_0_899976223) /* sqrt(2) * (c3-c7) */
243	 + MULTIPLY(z4, FIX_2_562915447); /* sqrt(2) * (c1+c3) */
244
245    /* Final output stage */
246
247    outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp2,
248					  CONST_BITS+PASS1_BITS+3+1)
249			    & RANGE_MASK];
250    outptr[3] = range_limit[(int) DESCALE(tmp10 - tmp2,
251					  CONST_BITS+PASS1_BITS+3+1)
252			    & RANGE_MASK];
253    outptr[1] = range_limit[(int) DESCALE(tmp12 + tmp0,
254					  CONST_BITS+PASS1_BITS+3+1)
255			    & RANGE_MASK];
256    outptr[2] = range_limit[(int) DESCALE(tmp12 - tmp0,
257					  CONST_BITS+PASS1_BITS+3+1)
258			    & RANGE_MASK];
259
260    wsptr += DCTSIZE;		/* advance pointer to next row */
261  }
262}
263
264
265/*
266 * Perform dequantization and inverse DCT on one block of coefficients,
267 * producing a reduced-size 2x2 output block.
268 */
269
270GLOBAL(void)
271jpeg_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
272	       JCOEFPTR coef_block,
273	       JSAMPARRAY output_buf, JDIMENSION output_col)
274{
275  INT32 tmp0, tmp10, z1;
276  JCOEFPTR inptr;
277  ISLOW_MULT_TYPE * quantptr;
278  int * wsptr;
279  JSAMPROW outptr;
280  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
281  int ctr;
282  int workspace[DCTSIZE*2];	/* buffers data between passes */
283  SHIFT_TEMPS
284
285  /* Pass 1: process columns from input, store into work array. */
286
287  inptr = coef_block;
288  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
289  wsptr = workspace;
290  for (ctr = DCTSIZE; ctr > 0; inptr++, quantptr++, wsptr++, ctr--) {
291    /* Don't bother to process columns 2,4,6 */
292    if (ctr == DCTSIZE-2 || ctr == DCTSIZE-4 || ctr == DCTSIZE-6)
293      continue;
294    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*3] == 0 &&
295	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*7] == 0) {
296      /* AC terms all zero; we need not examine terms 2,4,6 for 2x2 output */
297      int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
298
299      wsptr[DCTSIZE*0] = dcval;
300      wsptr[DCTSIZE*1] = dcval;
301
302      continue;
303    }
304
305    /* Even part */
306
307    z1 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
308    tmp10 = z1 << (CONST_BITS+2);
309
310    /* Odd part */
311
312    z1 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
313    tmp0 = MULTIPLY(z1, - FIX_0_720959822); /* sqrt(2) * (c7-c5+c3-c1) */
314    z1 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
315    tmp0 += MULTIPLY(z1, FIX_0_850430095); /* sqrt(2) * (-c1+c3+c5+c7) */
316    z1 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
317    tmp0 += MULTIPLY(z1, - FIX_1_272758580); /* sqrt(2) * (-c1+c3-c5-c7) */
318    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
319    tmp0 += MULTIPLY(z1, FIX_3_624509785); /* sqrt(2) * (c1+c3+c5+c7) */
320
321    /* Final output stage */
322
323    wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp0, CONST_BITS-PASS1_BITS+2);
324    wsptr[DCTSIZE*1] = (int) DESCALE(tmp10 - tmp0, CONST_BITS-PASS1_BITS+2);
325  }
326
327  /* Pass 2: process 2 rows from work array, store into output array. */
328
329  wsptr = workspace;
330  for (ctr = 0; ctr < 2; ctr++) {
331    outptr = output_buf[ctr] + output_col;
332    /* It's not clear whether a zero row test is worthwhile here ... */
333
334#ifndef NO_ZERO_ROW_TEST
335    if (wsptr[1] == 0 && wsptr[3] == 0 && wsptr[5] == 0 && wsptr[7] == 0) {
336      /* AC terms all zero */
337      JSAMPLE dcval = range_limit[(int) DESCALE((INT32) wsptr[0], PASS1_BITS+3)
338				  & RANGE_MASK];
339
340      outptr[0] = dcval;
341      outptr[1] = dcval;
342
343      wsptr += DCTSIZE;		/* advance pointer to next row */
344      continue;
345    }
346#endif
347
348    /* Even part */
349
350    tmp10 = ((INT32) wsptr[0]) << (CONST_BITS+2);
351
352    /* Odd part */
353
354    tmp0 = MULTIPLY((INT32) wsptr[7], - FIX_0_720959822) /* sqrt(2) * (c7-c5+c3-c1) */
355	 + MULTIPLY((INT32) wsptr[5], FIX_0_850430095) /* sqrt(2) * (-c1+c3+c5+c7) */
356	 + MULTIPLY((INT32) wsptr[3], - FIX_1_272758580) /* sqrt(2) * (-c1+c3-c5-c7) */
357	 + MULTIPLY((INT32) wsptr[1], FIX_3_624509785); /* sqrt(2) * (c1+c3+c5+c7) */
358
359    /* Final output stage */
360
361    outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp0,
362					  CONST_BITS+PASS1_BITS+3+2)
363			    & RANGE_MASK];
364    outptr[1] = range_limit[(int) DESCALE(tmp10 - tmp0,
365					  CONST_BITS+PASS1_BITS+3+2)
366			    & RANGE_MASK];
367
368    wsptr += DCTSIZE;		/* advance pointer to next row */
369  }
370}
371
372
373/*
374 * Perform dequantization and inverse DCT on one block of coefficients,
375 * producing a reduced-size 1x1 output block.
376 */
377
378GLOBAL(void)
379jpeg_idct_1x1 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
380	       JCOEFPTR coef_block,
381	       JSAMPARRAY output_buf, JDIMENSION output_col)
382{
383  int dcval;
384  ISLOW_MULT_TYPE * quantptr;
385  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
386  SHIFT_TEMPS
387
388  /* We hardly need an inverse DCT routine for this: just take the
389   * average pixel value, which is one-eighth of the DC coefficient.
390   */
391  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
392  dcval = DEQUANTIZE(coef_block[0], quantptr[0]);
393  dcval = (int) DESCALE((INT32) dcval, 3);
394
395  output_buf[0][output_col] = range_limit[dcval & RANGE_MASK];
396}
397
398#endif /* IDCT_SCALING_SUPPORTED */
399