1/*
2 * jidctflt.c
3 *
4 * Copyright (C) 1994-1998, Thomas G. Lane.
5 * This file is part of the Independent JPEG Group's software.
6 *
7 * The authors make NO WARRANTY or representation, either express or implied,
8 * with respect to this software, its quality, accuracy, merchantability, or
9 * fitness for a particular purpose.  This software is provided "AS IS", and you,
10 * its user, assume the entire risk as to its quality and accuracy.
11 *
12 * This software is copyright (C) 1991-1998, Thomas G. Lane.
13 * All Rights Reserved except as specified below.
14 *
15 * Permission is hereby granted to use, copy, modify, and distribute this
16 * software (or portions thereof) for any purpose, without fee, subject to these
17 * conditions:
18 * (1) If any part of the source code for this software is distributed, then this
19 * README file must be included, with this copyright and no-warranty notice
20 * unaltered; and any additions, deletions, or changes to the original files
21 * must be clearly indicated in accompanying documentation.
22 * (2) If only executable code is distributed, then the accompanying
23 * documentation must state that "this software is based in part on the work of
24 * the Independent JPEG Group".
25 * (3) Permission for use of this software is granted only if the user accepts
26 * full responsibility for any undesirable consequences; the authors accept
27 * NO LIABILITY for damages of any kind.
28 *
29 * These conditions apply to any software derived from or based on the IJG code,
30 * not just to the unmodified library.  If you use our work, you ought to
31 * acknowledge us.
32 *
33 * Permission is NOT granted for the use of any IJG author's name or company name
34 * in advertising or publicity relating to this software or products derived from
35 * it.  This software may be referred to only as "the Independent JPEG Group's
36 * software".
37 *
38 * We specifically permit and encourage the use of this software as the basis of
39 * commercial products, provided that all warranty or liability claims are
40 * assumed by the product vendor.
41 *
42 *
43 * This file contains a floating-point implementation of the
44 * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
45 * must also perform dequantization of the input coefficients.
46 *
47 * This implementation should be more accurate than either of the integer
48 * IDCT implementations.  However, it may not give the same results on all
49 * machines because of differences in roundoff behavior.  Speed will depend
50 * on the hardware's floating point capacity.
51 *
52 * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
53 * on each row (or vice versa, but it's more convenient to emit a row at
54 * a time).  Direct algorithms are also available, but they are much more
55 * complex and seem not to be any faster when reduced to code.
56 *
57 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
58 * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
59 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
60 * JPEG textbook (see REFERENCES section in file README).  The following code
61 * is based directly on figure 4-8 in P&M.
62 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
63 * possible to arrange the computation so that many of the multiplies are
64 * simple scalings of the final outputs.  These multiplies can then be
65 * folded into the multiplications or divisions by the JPEG quantization
66 * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
67 * to be done in the DCT itself.
68 * The primary disadvantage of this method is that with a fixed-point
69 * implementation, accuracy is lost due to imprecise representation of the
70 * scaled quantization values.  However, that problem does not arise if
71 * we use floating point arithmetic.
72 */
73
74#include <stdint.h>
75#include "tinyjpeg-internal.h"
76
77#define FAST_FLOAT float
78#define DCTSIZE	   8
79#define DCTSIZE2   (DCTSIZE*DCTSIZE)
80
81#define DEQUANTIZE(coef,quantval)  (((FAST_FLOAT) (coef)) * (quantval))
82
83#if 1 && defined(__GNUC__) && (defined(__i686__) || defined(__x86_64__))
84
85static inline unsigned char descale_and_clamp(int x, int shift)
86{
87  __asm__ (
88      "add %3,%1\n"
89      "\tsar %2,%1\n"
90      "\tsub $-128,%1\n"
91      "\tcmovl %5,%1\n"	/* Use the sub to compare to 0 */
92      "\tcmpl %4,%1\n"
93      "\tcmovg %4,%1\n"
94      : "=r"(x)
95      : "0"(x), "Ir"(shift), "ir"(1UL<<(shift-1)), "r" (0xff), "r" (0)
96      );
97  return x;
98}
99
100#else
101static inline unsigned char descale_and_clamp(int x, int shift)
102{
103  x += (1UL<<(shift-1));
104  if (x<0)
105    x = (x >> shift) | ((~(0UL)) << (32-(shift)));
106  else
107    x >>= shift;
108  x += 128;
109  if (x>255)
110    return 255;
111  else if (x<0)
112    return 0;
113  else
114    return x;
115}
116#endif
117
118/*
119 * Perform dequantization and inverse DCT on one block of coefficients.
120 */
121
122void
123tinyjpeg_idct_float (struct component *compptr, uint8_t *output_buf, int stride)
124{
125  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
126  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
127  FAST_FLOAT z5, z10, z11, z12, z13;
128  int16_t *inptr;
129  FAST_FLOAT *quantptr;
130  FAST_FLOAT *wsptr;
131  uint8_t *outptr;
132  int ctr;
133  FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
134
135  /* Pass 1: process columns from input, store into work array. */
136
137  inptr = compptr->DCT;
138  quantptr = compptr->Q_table;
139  wsptr = workspace;
140  for (ctr = DCTSIZE; ctr > 0; ctr--) {
141    /* Due to quantization, we will usually find that many of the input
142     * coefficients are zero, especially the AC terms.  We can exploit this
143     * by short-circuiting the IDCT calculation for any column in which all
144     * the AC terms are zero.  In that case each output is equal to the
145     * DC coefficient (with scale factor as needed).
146     * With typical images and quantization tables, half or more of the
147     * column DCT calculations can be simplified this way.
148     */
149
150    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
151	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
152	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
153	inptr[DCTSIZE*7] == 0) {
154      /* AC terms all zero */
155      FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
156
157      wsptr[DCTSIZE*0] = dcval;
158      wsptr[DCTSIZE*1] = dcval;
159      wsptr[DCTSIZE*2] = dcval;
160      wsptr[DCTSIZE*3] = dcval;
161      wsptr[DCTSIZE*4] = dcval;
162      wsptr[DCTSIZE*5] = dcval;
163      wsptr[DCTSIZE*6] = dcval;
164      wsptr[DCTSIZE*7] = dcval;
165
166      inptr++;			/* advance pointers to next column */
167      quantptr++;
168      wsptr++;
169      continue;
170    }
171
172    /* Even part */
173
174    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
175    tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
176    tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
177    tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
178
179    tmp10 = tmp0 + tmp2;	/* phase 3 */
180    tmp11 = tmp0 - tmp2;
181
182    tmp13 = tmp1 + tmp3;	/* phases 5-3 */
183    tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
184
185    tmp0 = tmp10 + tmp13;	/* phase 2 */
186    tmp3 = tmp10 - tmp13;
187    tmp1 = tmp11 + tmp12;
188    tmp2 = tmp11 - tmp12;
189
190    /* Odd part */
191
192    tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
193    tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
194    tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
195    tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
196
197    z13 = tmp6 + tmp5;		/* phase 6 */
198    z10 = tmp6 - tmp5;
199    z11 = tmp4 + tmp7;
200    z12 = tmp4 - tmp7;
201
202    tmp7 = z11 + z13;		/* phase 5 */
203    tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
204
205    z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
206    tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
207    tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
208
209    tmp6 = tmp12 - tmp7;	/* phase 2 */
210    tmp5 = tmp11 - tmp6;
211    tmp4 = tmp10 + tmp5;
212
213    wsptr[DCTSIZE*0] = tmp0 + tmp7;
214    wsptr[DCTSIZE*7] = tmp0 - tmp7;
215    wsptr[DCTSIZE*1] = tmp1 + tmp6;
216    wsptr[DCTSIZE*6] = tmp1 - tmp6;
217    wsptr[DCTSIZE*2] = tmp2 + tmp5;
218    wsptr[DCTSIZE*5] = tmp2 - tmp5;
219    wsptr[DCTSIZE*4] = tmp3 + tmp4;
220    wsptr[DCTSIZE*3] = tmp3 - tmp4;
221
222    inptr++;			/* advance pointers to next column */
223    quantptr++;
224    wsptr++;
225  }
226
227  /* Pass 2: process rows from work array, store into output array. */
228  /* Note that we must descale the results by a factor of 8 == 2**3. */
229
230  wsptr = workspace;
231  outptr = output_buf;
232  for (ctr = 0; ctr < DCTSIZE; ctr++) {
233    /* Rows of zeroes can be exploited in the same way as we did with columns.
234     * However, the column calculation has created many nonzero AC terms, so
235     * the simplification applies less often (typically 5% to 10% of the time).
236     * And testing floats for zero is relatively expensive, so we don't bother.
237     */
238
239    /* Even part */
240
241    tmp10 = wsptr[0] + wsptr[4];
242    tmp11 = wsptr[0] - wsptr[4];
243
244    tmp13 = wsptr[2] + wsptr[6];
245    tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT) 1.414213562) - tmp13;
246
247    tmp0 = tmp10 + tmp13;
248    tmp3 = tmp10 - tmp13;
249    tmp1 = tmp11 + tmp12;
250    tmp2 = tmp11 - tmp12;
251
252    /* Odd part */
253
254    z13 = wsptr[5] + wsptr[3];
255    z10 = wsptr[5] - wsptr[3];
256    z11 = wsptr[1] + wsptr[7];
257    z12 = wsptr[1] - wsptr[7];
258
259    tmp7 = z11 + z13;
260    tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562);
261
262    z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
263    tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
264    tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
265
266    tmp6 = tmp12 - tmp7;
267    tmp5 = tmp11 - tmp6;
268    tmp4 = tmp10 + tmp5;
269
270    /* Final output stage: scale down by a factor of 8 and range-limit */
271
272    outptr[0] = descale_and_clamp((int)(tmp0 + tmp7), 3);
273    outptr[7] = descale_and_clamp((int)(tmp0 - tmp7), 3);
274    outptr[1] = descale_and_clamp((int)(tmp1 + tmp6), 3);
275    outptr[6] = descale_and_clamp((int)(tmp1 - tmp6), 3);
276    outptr[2] = descale_and_clamp((int)(tmp2 + tmp5), 3);
277    outptr[5] = descale_and_clamp((int)(tmp2 - tmp5), 3);
278    outptr[4] = descale_and_clamp((int)(tmp3 + tmp4), 3);
279    outptr[3] = descale_and_clamp((int)(tmp3 - tmp4), 3);
280
281
282    wsptr += DCTSIZE;		/* advance pointer to next row */
283    outptr += stride;
284  }
285}
286