1 /*
2  * jidctred.c
3  *
4  * This file was part of the Independent JPEG Group's software.
5  * Copyright (C) 1994-1998, Thomas G. Lane.
6  * libjpeg-turbo Modifications:
7  * Copyright (C) 2015, D. R. Commander
8  * For conditions of distribution and use, see the accompanying README file.
9  *
10  * This file contains inverse-DCT routines that produce reduced-size output:
11  * either 4x4, 2x2, or 1x1 pixels from an 8x8 DCT block.
12  *
13  * The implementation is based on the Loeffler, Ligtenberg and Moschytz (LL&M)
14  * algorithm used in jidctint.c.  We simply replace each 8-to-8 1-D IDCT step
15  * with an 8-to-4 step that produces the four averages of two adjacent outputs
16  * (or an 8-to-2 step producing two averages of four outputs, for 2x2 output).
17  * These steps were derived by computing the corresponding values at the end
18  * of the normal LL&M code, then simplifying as much as possible.
19  *
20  * 1x1 is trivial: just take the DC coefficient divided by 8.
21  *
22  * See jidctint.c for additional comments.
23  */
24 
25 #define JPEG_INTERNALS
26 #include "jinclude.h"
27 #include "jpeglib.h"
28 #include "jdct.h"               /* Private declarations for DCT subsystem */
29 
30 #ifdef IDCT_SCALING_SUPPORTED
31 
32 
33 /*
34  * This module is specialized to the case DCTSIZE = 8.
35  */
36 
37 #if DCTSIZE != 8
38   Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
39 #endif
40 
41 
42 /* Scaling is the same as in jidctint.c. */
43 
44 #if BITS_IN_JSAMPLE == 8
45 #define CONST_BITS  13
46 #define PASS1_BITS  2
47 #else
48 #define CONST_BITS  13
49 #define PASS1_BITS  1           /* lose a little precision to avoid overflow */
50 #endif
51 
52 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
53  * causing a lot of useless floating-point operations at run time.
54  * To get around this we use the following pre-calculated constants.
55  * If you change CONST_BITS you may want to add appropriate values.
56  * (With a reasonable C compiler, you can just rely on the FIX() macro...)
57  */
58 
59 #if CONST_BITS == 13
60 #define FIX_0_211164243  ((INT32)  1730)        /* FIX(0.211164243) */
61 #define FIX_0_509795579  ((INT32)  4176)        /* FIX(0.509795579) */
62 #define FIX_0_601344887  ((INT32)  4926)        /* FIX(0.601344887) */
63 #define FIX_0_720959822  ((INT32)  5906)        /* FIX(0.720959822) */
64 #define FIX_0_765366865  ((INT32)  6270)        /* FIX(0.765366865) */
65 #define FIX_0_850430095  ((INT32)  6967)        /* FIX(0.850430095) */
66 #define FIX_0_899976223  ((INT32)  7373)        /* FIX(0.899976223) */
67 #define FIX_1_061594337  ((INT32)  8697)        /* FIX(1.061594337) */
68 #define FIX_1_272758580  ((INT32)  10426)       /* FIX(1.272758580) */
69 #define FIX_1_451774981  ((INT32)  11893)       /* FIX(1.451774981) */
70 #define FIX_1_847759065  ((INT32)  15137)       /* FIX(1.847759065) */
71 #define FIX_2_172734803  ((INT32)  17799)       /* FIX(2.172734803) */
72 #define FIX_2_562915447  ((INT32)  20995)       /* FIX(2.562915447) */
73 #define FIX_3_624509785  ((INT32)  29692)       /* FIX(3.624509785) */
74 #else
75 #define FIX_0_211164243  FIX(0.211164243)
76 #define FIX_0_509795579  FIX(0.509795579)
77 #define FIX_0_601344887  FIX(0.601344887)
78 #define FIX_0_720959822  FIX(0.720959822)
79 #define FIX_0_765366865  FIX(0.765366865)
80 #define FIX_0_850430095  FIX(0.850430095)
81 #define FIX_0_899976223  FIX(0.899976223)
82 #define FIX_1_061594337  FIX(1.061594337)
83 #define FIX_1_272758580  FIX(1.272758580)
84 #define FIX_1_451774981  FIX(1.451774981)
85 #define FIX_1_847759065  FIX(1.847759065)
86 #define FIX_2_172734803  FIX(2.172734803)
87 #define FIX_2_562915447  FIX(2.562915447)
88 #define FIX_3_624509785  FIX(3.624509785)
89 #endif
90 
91 
92 /* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
93  * For 8-bit samples with the recommended scaling, all the variable
94  * and constant values involved are no more than 16 bits wide, so a
95  * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
96  * For 12-bit samples, a full 32-bit multiplication will be needed.
97  */
98 
99 #if BITS_IN_JSAMPLE == 8
100 #define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
101 #else
102 #define MULTIPLY(var,const)  ((var) * (const))
103 #endif
104 
105 
106 /* Dequantize a coefficient by multiplying it by the multiplier-table
107  * entry; produce an int result.  In this module, both inputs and result
108  * are 16 bits or less, so either int or short multiply will work.
109  */
110 
111 #define DEQUANTIZE(coef,quantval)  (((ISLOW_MULT_TYPE) (coef)) * (quantval))
112 
113 
114 /*
115  * Perform dequantization and inverse DCT on one block of coefficients,
116  * producing a reduced-size 4x4 output block.
117  */
118 
119 GLOBAL(void)
120 jpeg_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
121                JCOEFPTR coef_block,
122                JSAMPARRAY output_buf, JDIMENSION output_col)
123 {
124   INT32 tmp0, tmp2, tmp10, tmp12;
125   INT32 z1, z2, z3, z4;
126   JCOEFPTR inptr;
127   ISLOW_MULT_TYPE * quantptr;
128   int * wsptr;
129   JSAMPROW outptr;
130   JSAMPLE *range_limit = IDCT_range_limit(cinfo);
131   int ctr;
132   int workspace[DCTSIZE*4];     /* buffers data between passes */
133   SHIFT_TEMPS
134 
135   /* Pass 1: process columns from input, store into work array. */
136 
137   inptr = coef_block;
138   quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
139   wsptr = workspace;
140   for (ctr = DCTSIZE; ctr > 0; inptr++, quantptr++, wsptr++, ctr--) {
141     /* Don't bother to process column 4, because second pass won't use it */
142     if (ctr == DCTSIZE-4)
143       continue;
144     if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
145         inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*5] == 0 &&
146         inptr[DCTSIZE*6] == 0 && inptr[DCTSIZE*7] == 0) {
147       /* AC terms all zero; we need not examine term 4 for 4x4 output */
148       int dcval = LEFT_SHIFT(DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]),
149                              PASS1_BITS);
150 
151       wsptr[DCTSIZE*0] = dcval;
152       wsptr[DCTSIZE*1] = dcval;
153       wsptr[DCTSIZE*2] = dcval;
154       wsptr[DCTSIZE*3] = dcval;
155 
156       continue;
157     }
158 
159     /* Even part */
160 
161     tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
162     tmp0 = LEFT_SHIFT(tmp0, CONST_BITS+1);
163 
164     z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
165     z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
166 
167     tmp2 = MULTIPLY(z2, FIX_1_847759065) + MULTIPLY(z3, - FIX_0_765366865);
168 
169     tmp10 = tmp0 + tmp2;
170     tmp12 = tmp0 - tmp2;
171 
172     /* Odd part */
173 
174     z1 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
175     z2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
176     z3 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
177     z4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
178 
179     tmp0 = MULTIPLY(z1, - FIX_0_211164243) /* sqrt(2) * (c3-c1) */
180          + MULTIPLY(z2, FIX_1_451774981) /* sqrt(2) * (c3+c7) */
181          + MULTIPLY(z3, - FIX_2_172734803) /* sqrt(2) * (-c1-c5) */
182          + MULTIPLY(z4, FIX_1_061594337); /* sqrt(2) * (c5+c7) */
183 
184     tmp2 = MULTIPLY(z1, - FIX_0_509795579) /* sqrt(2) * (c7-c5) */
185          + MULTIPLY(z2, - FIX_0_601344887) /* sqrt(2) * (c5-c1) */
186          + MULTIPLY(z3, FIX_0_899976223) /* sqrt(2) * (c3-c7) */
187          + MULTIPLY(z4, FIX_2_562915447); /* sqrt(2) * (c1+c3) */
188 
189     /* Final output stage */
190 
191     wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp2, CONST_BITS-PASS1_BITS+1);
192     wsptr[DCTSIZE*3] = (int) DESCALE(tmp10 - tmp2, CONST_BITS-PASS1_BITS+1);
193     wsptr[DCTSIZE*1] = (int) DESCALE(tmp12 + tmp0, CONST_BITS-PASS1_BITS+1);
194     wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 - tmp0, CONST_BITS-PASS1_BITS+1);
195   }
196 
197   /* Pass 2: process 4 rows from work array, store into output array. */
198 
199   wsptr = workspace;
200   for (ctr = 0; ctr < 4; ctr++) {
201     outptr = output_buf[ctr] + output_col;
202     /* It's not clear whether a zero row test is worthwhile here ... */
203 
204 #ifndef NO_ZERO_ROW_TEST
205     if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 &&
206         wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
207       /* AC terms all zero */
208       JSAMPLE dcval = range_limit[(int) DESCALE((INT32) wsptr[0], PASS1_BITS+3)
209                                   & RANGE_MASK];
210 
211       outptr[0] = dcval;
212       outptr[1] = dcval;
213       outptr[2] = dcval;
214       outptr[3] = dcval;
215 
216       wsptr += DCTSIZE;         /* advance pointer to next row */
217       continue;
218     }
219 #endif
220 
221     /* Even part */
222 
223     tmp0 = LEFT_SHIFT((INT32) wsptr[0], CONST_BITS+1);
224 
225     tmp2 = MULTIPLY((INT32) wsptr[2], FIX_1_847759065)
226          + MULTIPLY((INT32) wsptr[6], - FIX_0_765366865);
227 
228     tmp10 = tmp0 + tmp2;
229     tmp12 = tmp0 - tmp2;
230 
231     /* Odd part */
232 
233     z1 = (INT32) wsptr[7];
234     z2 = (INT32) wsptr[5];
235     z3 = (INT32) wsptr[3];
236     z4 = (INT32) wsptr[1];
237 
238     tmp0 = MULTIPLY(z1, - FIX_0_211164243) /* sqrt(2) * (c3-c1) */
239          + MULTIPLY(z2, FIX_1_451774981) /* sqrt(2) * (c3+c7) */
240          + MULTIPLY(z3, - FIX_2_172734803) /* sqrt(2) * (-c1-c5) */
241          + MULTIPLY(z4, FIX_1_061594337); /* sqrt(2) * (c5+c7) */
242 
243     tmp2 = MULTIPLY(z1, - FIX_0_509795579) /* sqrt(2) * (c7-c5) */
244          + MULTIPLY(z2, - FIX_0_601344887) /* sqrt(2) * (c5-c1) */
245          + MULTIPLY(z3, FIX_0_899976223) /* sqrt(2) * (c3-c7) */
246          + MULTIPLY(z4, FIX_2_562915447); /* sqrt(2) * (c1+c3) */
247 
248     /* Final output stage */
249 
250     outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp2,
251                                           CONST_BITS+PASS1_BITS+3+1)
252                             & RANGE_MASK];
253     outptr[3] = range_limit[(int) DESCALE(tmp10 - tmp2,
254                                           CONST_BITS+PASS1_BITS+3+1)
255                             & RANGE_MASK];
256     outptr[1] = range_limit[(int) DESCALE(tmp12 + tmp0,
257                                           CONST_BITS+PASS1_BITS+3+1)
258                             & RANGE_MASK];
259     outptr[2] = range_limit[(int) DESCALE(tmp12 - tmp0,
260                                           CONST_BITS+PASS1_BITS+3+1)
261                             & RANGE_MASK];
262 
263     wsptr += DCTSIZE;           /* advance pointer to next row */
264   }
265 }
266 
267 
268 /*
269  * Perform dequantization and inverse DCT on one block of coefficients,
270  * producing a reduced-size 2x2 output block.
271  */
272 
273 GLOBAL(void)
jpeg_idct_2x2(j_decompress_ptr cinfo,jpeg_component_info * compptr,JCOEFPTR coef_block,JSAMPARRAY output_buf,JDIMENSION output_col)274 jpeg_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
275                JCOEFPTR coef_block,
276                JSAMPARRAY output_buf, JDIMENSION output_col)
277 {
278   INT32 tmp0, tmp10, z1;
279   JCOEFPTR inptr;
280   ISLOW_MULT_TYPE * quantptr;
281   int * wsptr;
282   JSAMPROW outptr;
283   JSAMPLE *range_limit = IDCT_range_limit(cinfo);
284   int ctr;
285   int workspace[DCTSIZE*2];     /* buffers data between passes */
286   SHIFT_TEMPS
287 
288   /* Pass 1: process columns from input, store into work array. */
289 
290   inptr = coef_block;
291   quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
292   wsptr = workspace;
293   for (ctr = DCTSIZE; ctr > 0; inptr++, quantptr++, wsptr++, ctr--) {
294     /* Don't bother to process columns 2,4,6 */
295     if (ctr == DCTSIZE-2 || ctr == DCTSIZE-4 || ctr == DCTSIZE-6)
296       continue;
297     if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*3] == 0 &&
298         inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*7] == 0) {
299       /* AC terms all zero; we need not examine terms 2,4,6 for 2x2 output */
300       int dcval = LEFT_SHIFT(DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]),
301                              PASS1_BITS);
302 
303       wsptr[DCTSIZE*0] = dcval;
304       wsptr[DCTSIZE*1] = dcval;
305 
306       continue;
307     }
308 
309     /* Even part */
310 
311     z1 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
312     tmp10 = LEFT_SHIFT(z1, CONST_BITS+2);
313 
314     /* Odd part */
315 
316     z1 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
317     tmp0 = MULTIPLY(z1, - FIX_0_720959822); /* sqrt(2) * (c7-c5+c3-c1) */
318     z1 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
319     tmp0 += MULTIPLY(z1, FIX_0_850430095); /* sqrt(2) * (-c1+c3+c5+c7) */
320     z1 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
321     tmp0 += MULTIPLY(z1, - FIX_1_272758580); /* sqrt(2) * (-c1+c3-c5-c7) */
322     z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
323     tmp0 += MULTIPLY(z1, FIX_3_624509785); /* sqrt(2) * (c1+c3+c5+c7) */
324 
325     /* Final output stage */
326 
327     wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp0, CONST_BITS-PASS1_BITS+2);
328     wsptr[DCTSIZE*1] = (int) DESCALE(tmp10 - tmp0, CONST_BITS-PASS1_BITS+2);
329   }
330 
331   /* Pass 2: process 2 rows from work array, store into output array. */
332 
333   wsptr = workspace;
334   for (ctr = 0; ctr < 2; ctr++) {
335     outptr = output_buf[ctr] + output_col;
336     /* It's not clear whether a zero row test is worthwhile here ... */
337 
338 #ifndef NO_ZERO_ROW_TEST
339     if (wsptr[1] == 0 && wsptr[3] == 0 && wsptr[5] == 0 && wsptr[7] == 0) {
340       /* AC terms all zero */
341       JSAMPLE dcval = range_limit[(int) DESCALE((INT32) wsptr[0], PASS1_BITS+3)
342                                   & RANGE_MASK];
343 
344       outptr[0] = dcval;
345       outptr[1] = dcval;
346 
347       wsptr += DCTSIZE;         /* advance pointer to next row */
348       continue;
349     }
350 #endif
351 
352     /* Even part */
353 
354     tmp10 = LEFT_SHIFT((INT32) wsptr[0], CONST_BITS+2);
355 
356     /* Odd part */
357 
358     tmp0 = MULTIPLY((INT32) wsptr[7], - FIX_0_720959822) /* sqrt(2) * (c7-c5+c3-c1) */
359          + MULTIPLY((INT32) wsptr[5], FIX_0_850430095) /* sqrt(2) * (-c1+c3+c5+c7) */
360          + MULTIPLY((INT32) wsptr[3], - FIX_1_272758580) /* sqrt(2) * (-c1+c3-c5-c7) */
361          + MULTIPLY((INT32) wsptr[1], FIX_3_624509785); /* sqrt(2) * (c1+c3+c5+c7) */
362 
363     /* Final output stage */
364 
365     outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp0,
366                                           CONST_BITS+PASS1_BITS+3+2)
367                             & RANGE_MASK];
368     outptr[1] = range_limit[(int) DESCALE(tmp10 - tmp0,
369                                           CONST_BITS+PASS1_BITS+3+2)
370                             & RANGE_MASK];
371 
372     wsptr += DCTSIZE;           /* advance pointer to next row */
373   }
374 }
375 
376 
377 /*
378  * Perform dequantization and inverse DCT on one block of coefficients,
379  * producing a reduced-size 1x1 output block.
380  */
381 
382 GLOBAL(void)
jpeg_idct_1x1(j_decompress_ptr cinfo,jpeg_component_info * compptr,JCOEFPTR coef_block,JSAMPARRAY output_buf,JDIMENSION output_col)383 jpeg_idct_1x1 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
384                JCOEFPTR coef_block,
385                JSAMPARRAY output_buf, JDIMENSION output_col)
386 {
387   int dcval;
388   ISLOW_MULT_TYPE * quantptr;
389   JSAMPLE *range_limit = IDCT_range_limit(cinfo);
390   SHIFT_TEMPS
391 
392   /* We hardly need an inverse DCT routine for this: just take the
393    * average pixel value, which is one-eighth of the DC coefficient.
394    */
395   quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
396   dcval = DEQUANTIZE(coef_block[0], quantptr[0]);
397   dcval = (int) DESCALE((INT32) dcval, 3);
398 
399   output_buf[0][output_col] = range_limit[dcval & RANGE_MASK];
400 }
401 
402 #endif /* IDCT_SCALING_SUPPORTED */
403