1 /*
2  * jidctfst.c
3  *
4  * Copyright (C) 1994-1998, Thomas G. Lane.
5  * Modified 2015 by Guido Vollbeding.
6  * This file is part of the Independent JPEG Group's software.
7  * For conditions of distribution and use, see the accompanying README file.
8  *
9  * This file contains a fast, not so accurate integer implementation of the
10  * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
11  * must also perform dequantization of the input coefficients.
12  *
13  * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
14  * on each row (or vice versa, but it's more convenient to emit a row at
15  * a time).  Direct algorithms are also available, but they are much more
16  * complex and seem not to be any faster when reduced to code.
17  *
18  * This implementation is based on Arai, Agui, and Nakajima's algorithm for
19  * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
20  * Japanese, but the algorithm is described in the Pennebaker & Mitchell
21  * JPEG textbook (see REFERENCES section in file README).  The following code
22  * is based directly on figure 4-8 in P&M.
23  * While an 8-point DCT cannot be done in less than 11 multiplies, it is
24  * possible to arrange the computation so that many of the multiplies are
25  * simple scalings of the final outputs.  These multiplies can then be
26  * folded into the multiplications or divisions by the JPEG quantization
27  * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
28  * to be done in the DCT itself.
29  * The primary disadvantage of this method is that with fixed-point math,
30  * accuracy is lost due to imprecise representation of the scaled
31  * quantization values.  The smaller the quantization table entry, the less
32  * precise the scaled value, so this implementation does worse with high-
33  * quality-setting files than with low-quality ones.
34  */
35 
36 #define JPEG_INTERNALS
37 #include "jinclude.h"
38 #include "jpeglib.h"
39 #include "jdct.h"		/* Private declarations for DCT subsystem */
40 
41 #ifdef DCT_IFAST_SUPPORTED
42 
43 
44 /*
45  * This module is specialized to the case DCTSIZE = 8.
46  */
47 
48 #if DCTSIZE != 8
49   Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
50 #endif
51 
52 
53 /* Scaling decisions are generally the same as in the LL&M algorithm;
54  * see jidctint.c for more details.  However, we choose to descale
55  * (right shift) multiplication products as soon as they are formed,
56  * rather than carrying additional fractional bits into subsequent additions.
57  * This compromises accuracy slightly, but it lets us save a few shifts.
58  * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
59  * everywhere except in the multiplications proper; this saves a good deal
60  * of work on 16-bit-int machines.
61  *
62  * The dequantized coefficients are not integers because the AA&N scaling
63  * factors have been incorporated.  We represent them scaled up by PASS1_BITS,
64  * so that the first and second IDCT rounds have the same input scaling.
65  * For 8-bit JSAMPLEs, we choose IFAST_SCALE_BITS = PASS1_BITS so as to
66  * avoid a descaling shift; this compromises accuracy rather drastically
67  * for small quantization table entries, but it saves a lot of shifts.
68  * For 12-bit JSAMPLEs, there's no hope of using 16x16 multiplies anyway,
69  * so we use a much larger scaling factor to preserve accuracy.
70  *
71  * A final compromise is to represent the multiplicative constants to only
72  * 8 fractional bits, rather than 13.  This saves some shifting work on some
73  * machines, and may also reduce the cost of multiplication (since there
74  * are fewer one-bits in the constants).
75  */
76 
77 #if BITS_IN_JSAMPLE == 8
78 #define CONST_BITS  8
79 #define PASS1_BITS  2
80 #else
81 #define CONST_BITS  8
82 #define PASS1_BITS  1		/* lose a little precision to avoid overflow */
83 #endif
84 
85 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
86  * causing a lot of useless floating-point operations at run time.
87  * To get around this we use the following pre-calculated constants.
88  * If you change CONST_BITS you may want to add appropriate values.
89  * (With a reasonable C compiler, you can just rely on the FIX() macro...)
90  */
91 
92 #if CONST_BITS == 8
93 #define FIX_1_082392200  ((INT32)  277)		/* FIX(1.082392200) */
94 #define FIX_1_414213562  ((INT32)  362)		/* FIX(1.414213562) */
95 #define FIX_1_847759065  ((INT32)  473)		/* FIX(1.847759065) */
96 #define FIX_2_613125930  ((INT32)  669)		/* FIX(2.613125930) */
97 #else
98 #define FIX_1_082392200  FIX(1.082392200)
99 #define FIX_1_414213562  FIX(1.414213562)
100 #define FIX_1_847759065  FIX(1.847759065)
101 #define FIX_2_613125930  FIX(2.613125930)
102 #endif
103 
104 
105 /* We can gain a little more speed, with a further compromise in accuracy,
106  * by omitting the addition in a descaling shift.  This yields an incorrectly
107  * rounded result half the time...
108  */
109 
110 #ifndef USE_ACCURATE_ROUNDING
111 #undef DESCALE
112 #define DESCALE(x,n)  RIGHT_SHIFT(x, n)
113 #endif
114 
115 
116 /* Multiply a DCTELEM variable by an INT32 constant, and immediately
117  * descale to yield a DCTELEM result.
118  */
119 
120 #define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
121 
122 
123 /* Dequantize a coefficient by multiplying it by the multiplier-table
124  * entry; produce a DCTELEM result.  For 8-bit data a 16x16->16
125  * multiplication will do.  For 12-bit data, the multiplier table is
126  * declared INT32, so a 32-bit multiply will be used.
127  */
128 
129 #if BITS_IN_JSAMPLE == 8
130 #define DEQUANTIZE(coef,quantval)  (((IFAST_MULT_TYPE) (coef)) * (quantval))
131 #else
132 #define DEQUANTIZE(coef,quantval)  \
133 	DESCALE((coef)*(quantval), IFAST_SCALE_BITS-PASS1_BITS)
134 #endif
135 
136 
137 /*
138  * Perform dequantization and inverse DCT on one block of coefficients.
139  *
140  * cK represents cos(K*pi/16).
141  */
142 
143 GLOBAL(void)
144 jpeg_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
145 		 JCOEFPTR coef_block,
146 		 JSAMPARRAY output_buf, JDIMENSION output_col)
147 {
148   DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
149   DCTELEM tmp10, tmp11, tmp12, tmp13;
150   DCTELEM z5, z10, z11, z12, z13;
151   JCOEFPTR inptr;
152   IFAST_MULT_TYPE * quantptr;
153   int * wsptr;
154   JSAMPROW outptr;
155   JSAMPLE *range_limit = IDCT_range_limit(cinfo);
156   int ctr;
157   int workspace[DCTSIZE2];	/* buffers data between passes */
158   SHIFT_TEMPS			/* for DESCALE */
159   ISHIFT_TEMPS			/* for IRIGHT_SHIFT */
160 
161   /* Pass 1: process columns from input, store into work array. */
162 
163   inptr = coef_block;
164   quantptr = (IFAST_MULT_TYPE *) compptr->dct_table;
165   wsptr = workspace;
166   for (ctr = DCTSIZE; ctr > 0; ctr--) {
167     /* Due to quantization, we will usually find that many of the input
168      * coefficients are zero, especially the AC terms.  We can exploit this
169      * by short-circuiting the IDCT calculation for any column in which all
170      * the AC terms are zero.  In that case each output is equal to the
171      * DC coefficient (with scale factor as needed).
172      * With typical images and quantization tables, half or more of the
173      * column DCT calculations can be simplified this way.
174      */
175 
176     if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
177 	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
178 	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
179 	inptr[DCTSIZE*7] == 0) {
180       /* AC terms all zero */
181       int dcval = (int) DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
182 
183       wsptr[DCTSIZE*0] = dcval;
184       wsptr[DCTSIZE*1] = dcval;
185       wsptr[DCTSIZE*2] = dcval;
186       wsptr[DCTSIZE*3] = dcval;
187       wsptr[DCTSIZE*4] = dcval;
188       wsptr[DCTSIZE*5] = dcval;
189       wsptr[DCTSIZE*6] = dcval;
190       wsptr[DCTSIZE*7] = dcval;
191 
192       inptr++;			/* advance pointers to next column */
193       quantptr++;
194       wsptr++;
195       continue;
196     }
197 
198     /* Even part */
199 
200     tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
201     tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
202     tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
203     tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
204 
205     tmp10 = tmp0 + tmp2;	/* phase 3 */
206     tmp11 = tmp0 - tmp2;
207 
208     tmp13 = tmp1 + tmp3;	/* phases 5-3 */
209     tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */
210 
211     tmp0 = tmp10 + tmp13;	/* phase 2 */
212     tmp3 = tmp10 - tmp13;
213     tmp1 = tmp11 + tmp12;
214     tmp2 = tmp11 - tmp12;
215 
216     /* Odd part */
217 
218     tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
219     tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
220     tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
221     tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
222 
223     z13 = tmp6 + tmp5;		/* phase 6 */
224     z10 = tmp6 - tmp5;
225     z11 = tmp4 + tmp7;
226     z12 = tmp4 - tmp7;
227 
228     tmp7 = z11 + z13;		/* phase 5 */
229     tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
230 
231     z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
232     tmp10 = z5 - MULTIPLY(z12, FIX_1_082392200); /* 2*(c2-c6) */
233     tmp12 = z5 - MULTIPLY(z10, FIX_2_613125930); /* 2*(c2+c6) */
234 
235     tmp6 = tmp12 - tmp7;	/* phase 2 */
236     tmp5 = tmp11 - tmp6;
237     tmp4 = tmp10 - tmp5;
238 
239     wsptr[DCTSIZE*0] = (int) (tmp0 + tmp7);
240     wsptr[DCTSIZE*7] = (int) (tmp0 - tmp7);
241     wsptr[DCTSIZE*1] = (int) (tmp1 + tmp6);
242     wsptr[DCTSIZE*6] = (int) (tmp1 - tmp6);
243     wsptr[DCTSIZE*2] = (int) (tmp2 + tmp5);
244     wsptr[DCTSIZE*5] = (int) (tmp2 - tmp5);
245     wsptr[DCTSIZE*3] = (int) (tmp3 + tmp4);
246     wsptr[DCTSIZE*4] = (int) (tmp3 - tmp4);
247 
248     inptr++;			/* advance pointers to next column */
249     quantptr++;
250     wsptr++;
251   }
252 
253   /* Pass 2: process rows from work array, store into output array.
254    * Note that we must descale the results by a factor of 8 == 2**3,
255    * and also undo the PASS1_BITS scaling.
256    */
257 
258   wsptr = workspace;
259   for (ctr = 0; ctr < DCTSIZE; ctr++) {
260     outptr = output_buf[ctr] + output_col;
261 
262     /* Add range center and fudge factor for final descale and range-limit. */
263     z5 = (DCTELEM) wsptr[0] +
264 	   ((((DCTELEM) RANGE_CENTER) << (PASS1_BITS+3)) +
265 	    (1 << (PASS1_BITS+2)));
266 
267     /* Rows of zeroes can be exploited in the same way as we did with columns.
268      * However, the column calculation has created many nonzero AC terms, so
269      * the simplification applies less often (typically 5% to 10% of the time).
270      * On machines with very fast multiplication, it's possible that the
271      * test takes more time than it's worth.  In that case this section
272      * may be commented out.
273      */
274 
275 #ifndef NO_ZERO_ROW_TEST
276     if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
277 	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
278       /* AC terms all zero */
279       JSAMPLE dcval = range_limit[(int) IRIGHT_SHIFT(z5, PASS1_BITS+3)
280 				  & RANGE_MASK];
281 
282       outptr[0] = dcval;
283       outptr[1] = dcval;
284       outptr[2] = dcval;
285       outptr[3] = dcval;
286       outptr[4] = dcval;
287       outptr[5] = dcval;
288       outptr[6] = dcval;
289       outptr[7] = dcval;
290 
291       wsptr += DCTSIZE;		/* advance pointer to next row */
292       continue;
293     }
294 #endif
295 
296     /* Even part */
297 
298     tmp10 = z5 + (DCTELEM) wsptr[4];
299     tmp11 = z5 - (DCTELEM) wsptr[4];
300 
301     tmp13 = (DCTELEM) wsptr[2] + (DCTELEM) wsptr[6];
302     tmp12 = MULTIPLY((DCTELEM) wsptr[2] - (DCTELEM) wsptr[6],
303 		     FIX_1_414213562) - tmp13; /* 2*c4 */
304 
305     tmp0 = tmp10 + tmp13;
306     tmp3 = tmp10 - tmp13;
307     tmp1 = tmp11 + tmp12;
308     tmp2 = tmp11 - tmp12;
309 
310     /* Odd part */
311 
312     z13 = (DCTELEM) wsptr[5] + (DCTELEM) wsptr[3];
313     z10 = (DCTELEM) wsptr[5] - (DCTELEM) wsptr[3];
314     z11 = (DCTELEM) wsptr[1] + (DCTELEM) wsptr[7];
315     z12 = (DCTELEM) wsptr[1] - (DCTELEM) wsptr[7];
316 
317     tmp7 = z11 + z13;		/* phase 5 */
318     tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
319 
320     z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
321     tmp10 = z5 - MULTIPLY(z12, FIX_1_082392200); /* 2*(c2-c6) */
322     tmp12 = z5 - MULTIPLY(z10, FIX_2_613125930); /* 2*(c2+c6) */
323 
324     tmp6 = tmp12 - tmp7;	/* phase 2 */
325     tmp5 = tmp11 - tmp6;
326     tmp4 = tmp10 - tmp5;
327 
328     /* Final output stage: scale down by a factor of 8 and range-limit */
329 
330     outptr[0] = range_limit[(int) IRIGHT_SHIFT(tmp0 + tmp7, PASS1_BITS+3)
331 			    & RANGE_MASK];
332     outptr[7] = range_limit[(int) IRIGHT_SHIFT(tmp0 - tmp7, PASS1_BITS+3)
333 			    & RANGE_MASK];
334     outptr[1] = range_limit[(int) IRIGHT_SHIFT(tmp1 + tmp6, PASS1_BITS+3)
335 			    & RANGE_MASK];
336     outptr[6] = range_limit[(int) IRIGHT_SHIFT(tmp1 - tmp6, PASS1_BITS+3)
337 			    & RANGE_MASK];
338     outptr[2] = range_limit[(int) IRIGHT_SHIFT(tmp2 + tmp5, PASS1_BITS+3)
339 			    & RANGE_MASK];
340     outptr[5] = range_limit[(int) IRIGHT_SHIFT(tmp2 - tmp5, PASS1_BITS+3)
341 			    & RANGE_MASK];
342     outptr[3] = range_limit[(int) IRIGHT_SHIFT(tmp3 + tmp4, PASS1_BITS+3)
343 			    & RANGE_MASK];
344     outptr[4] = range_limit[(int) IRIGHT_SHIFT(tmp3 - tmp4, PASS1_BITS+3)
345 			    & RANGE_MASK];
346 
347     wsptr += DCTSIZE;		/* advance pointer to next row */
348   }
349 }
350 
351 #endif /* DCT_IFAST_SUPPORTED */
352