blob: adfb6e363f6199f6128f974aeda7f79c1527b3c7 [file] [log] [blame]
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +00001/*
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +00002 * jfdctint.c
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +00003 *
DRC8e9cef22015-09-21 12:57:41 -05004 * This file was part of the Independent JPEG Group's software.
Thomas G. Lane489583f1996-02-07 00:00:00 +00005 * Copyright (C) 1991-1996, Thomas G. Lane.
DRC8e9cef22015-09-21 12:57:41 -05006 * libjpeg-turbo Modifications:
7 * Copyright (C) 2015, D. R. Commander
DRC7e3acc02015-10-10 10:25:46 -05008 * For conditions of distribution and use, see the accompanying README.ijg
9 * file.
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +000010 *
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000011 * This file contains a slow-but-accurate integer implementation of the
12 * forward DCT (Discrete Cosine Transform).
13 *
14 * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
15 * on each column. Direct algorithms are also available, but they are
16 * much more complex and seem not to be any faster when reduced to code.
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +000017 *
Thomas G. Lane88aeed41992-12-10 00:00:00 +000018 * This implementation is based on an algorithm described in
19 * C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
20 * Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
21 * Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
22 * The primary algorithm described there uses 11 multiplies and 29 adds.
23 * We use their alternate method with 12 multiplies and 32 adds.
24 * The advantage of this method is that no data path contains more than one
25 * multiplication; this allows a very simple and accurate implementation in
26 * scaled fixed-point arithmetic, with a minimal number of shifts.
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +000027 */
28
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000029#define JPEG_INTERNALS
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +000030#include "jinclude.h"
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000031#include "jpeglib.h"
DRCe5eaf372014-05-09 18:00:32 +000032#include "jdct.h" /* Private declarations for DCT subsystem */
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000033
34#ifdef DCT_ISLOW_SUPPORTED
35
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +000036
Thomas G. Lane4a6b7301992-03-17 00:00:00 +000037/*
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000038 * This module is specialized to the case DCTSIZE = 8.
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +000039 */
40
Thomas G. Lane4a6b7301992-03-17 00:00:00 +000041#if DCTSIZE != 8
42 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +000043#endif
44
Thomas G. Lanebd543f01991-12-13 00:00:00 +000045
Thomas G. Lane88aeed41992-12-10 00:00:00 +000046/*
Thomas G. Lane88aeed41992-12-10 00:00:00 +000047 * The poop on this scaling stuff is as follows:
48 *
49 * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
50 * larger than the true DCT outputs. The final outputs are therefore
51 * a factor of N larger than desired; since N=8 this can be cured by
52 * a simple right shift at the end of the algorithm. The advantage of
53 * this arrangement is that we save two multiplications per 1-D DCT,
54 * because the y0 and y4 outputs need not be divided by sqrt(N).
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000055 * In the IJG code, this factor of 8 is removed by the quantization step
56 * (in jcdctmgr.c), NOT in this module.
Thomas G. Lanebd543f01991-12-13 00:00:00 +000057 *
58 * We have to do addition and subtraction of the integer inputs, which
59 * is no problem, and multiplication by fractional constants, which is
60 * a problem to do in integer arithmetic. We multiply all the constants
Thomas G. Lane88aeed41992-12-10 00:00:00 +000061 * by CONST_SCALE and convert them to integer constants (thus retaining
62 * CONST_BITS bits of precision in the constants). After doing a
63 * multiplication we have to divide the product by CONST_SCALE, with proper
64 * rounding, to produce the correct output. This division can be done
65 * cheaply as a right shift of CONST_BITS bits. We postpone shifting
66 * as long as possible so that partial sums can be added together with
67 * full fractional precision.
Thomas G. Lanebd543f01991-12-13 00:00:00 +000068 *
Thomas G. Lane88aeed41992-12-10 00:00:00 +000069 * The outputs of the first pass are scaled up by PASS1_BITS bits so that
70 * they are represented to better-than-integral precision. These outputs
71 * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000072 * with the recommended scaling. (For 12-bit sample data, the intermediate
DRC1e32fe32015-10-14 17:32:39 -050073 * array is JLONG anyway.)
Thomas G. Lane88aeed41992-12-10 00:00:00 +000074 *
75 * To avoid overflow of the 32-bit intermediate results in pass 2, we must
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000076 * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26. Error analysis
Thomas G. Lane88aeed41992-12-10 00:00:00 +000077 * shows that the values given below are the most effective.
Thomas G. Lanebd543f01991-12-13 00:00:00 +000078 */
79
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +000080#if BITS_IN_JSAMPLE == 8
Thomas G. Lane88aeed41992-12-10 00:00:00 +000081#define CONST_BITS 13
82#define PASS1_BITS 2
Thomas G. Lanebd543f01991-12-13 00:00:00 +000083#else
Thomas G. Lane88aeed41992-12-10 00:00:00 +000084#define CONST_BITS 13
DRCe5eaf372014-05-09 18:00:32 +000085#define PASS1_BITS 1 /* lose a little precision to avoid overflow */
Thomas G. Lanebd543f01991-12-13 00:00:00 +000086#endif
87
Thomas G. Lane88aeed41992-12-10 00:00:00 +000088/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
89 * causing a lot of useless floating-point operations at run time.
90 * To get around this we use the following pre-calculated constants.
91 * If you change CONST_BITS you may want to add appropriate values.
92 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
93 */
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +000094
Thomas G. Lane88aeed41992-12-10 00:00:00 +000095#if CONST_BITS == 13
DRC1e32fe32015-10-14 17:32:39 -050096#define FIX_0_298631336 ((JLONG) 2446) /* FIX(0.298631336) */
97#define FIX_0_390180644 ((JLONG) 3196) /* FIX(0.390180644) */
98#define FIX_0_541196100 ((JLONG) 4433) /* FIX(0.541196100) */
99#define FIX_0_765366865 ((JLONG) 6270) /* FIX(0.765366865) */
100#define FIX_0_899976223 ((JLONG) 7373) /* FIX(0.899976223) */
101#define FIX_1_175875602 ((JLONG) 9633) /* FIX(1.175875602) */
102#define FIX_1_501321110 ((JLONG) 12299) /* FIX(1.501321110) */
103#define FIX_1_847759065 ((JLONG) 15137) /* FIX(1.847759065) */
104#define FIX_1_961570560 ((JLONG) 16069) /* FIX(1.961570560) */
105#define FIX_2_053119869 ((JLONG) 16819) /* FIX(2.053119869) */
106#define FIX_2_562915447 ((JLONG) 20995) /* FIX(2.562915447) */
107#define FIX_3_072711026 ((JLONG) 25172) /* FIX(3.072711026) */
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000108#else
109#define FIX_0_298631336 FIX(0.298631336)
110#define FIX_0_390180644 FIX(0.390180644)
111#define FIX_0_541196100 FIX(0.541196100)
112#define FIX_0_765366865 FIX(0.765366865)
113#define FIX_0_899976223 FIX(0.899976223)
114#define FIX_1_175875602 FIX(1.175875602)
115#define FIX_1_501321110 FIX(1.501321110)
116#define FIX_1_847759065 FIX(1.847759065)
117#define FIX_1_961570560 FIX(1.961570560)
118#define FIX_2_053119869 FIX(2.053119869)
119#define FIX_2_562915447 FIX(2.562915447)
120#define FIX_3_072711026 FIX(3.072711026)
121#endif
Thomas G. Lanebd543f01991-12-13 00:00:00 +0000122
Thomas G. Lanebd543f01991-12-13 00:00:00 +0000123
DRC1e32fe32015-10-14 17:32:39 -0500124/* Multiply an JLONG variable by an JLONG constant to yield an JLONG result.
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000125 * For 8-bit samples with the recommended scaling, all the variable
126 * and constant values involved are no more than 16 bits wide, so a
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000127 * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
128 * For 12-bit samples, a full 32-bit multiplication will be needed.
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000129 */
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +0000130
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000131#if BITS_IN_JSAMPLE == 8
132#define MULTIPLY(var,const) MULTIPLY16C16(var,const)
133#else
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000134#define MULTIPLY(var,const) ((var) * (const))
135#endif
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +0000136
137
Thomas G. Lanebd543f01991-12-13 00:00:00 +0000138/*
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +0000139 * Perform the forward DCT on one block of samples.
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +0000140 */
141
Thomas G. Lane489583f1996-02-07 00:00:00 +0000142GLOBAL(void)
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000143jpeg_fdct_islow (DCTELEM * data)
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +0000144{
DRC1e32fe32015-10-14 17:32:39 -0500145 JLONG tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
146 JLONG tmp10, tmp11, tmp12, tmp13;
147 JLONG z1, z2, z3, z4, z5;
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000148 DCTELEM *dataptr;
149 int ctr;
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000150 SHIFT_TEMPS
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +0000151
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000152 /* Pass 1: process rows. */
153 /* Note results are scaled up by sqrt(8) compared to a true DCT; */
154 /* furthermore, we scale the results by 2**PASS1_BITS. */
Thomas G. Lane4a6b7301992-03-17 00:00:00 +0000155
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000156 dataptr = data;
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000157 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000158 tmp0 = dataptr[0] + dataptr[7];
159 tmp7 = dataptr[0] - dataptr[7];
160 tmp1 = dataptr[1] + dataptr[6];
161 tmp6 = dataptr[1] - dataptr[6];
162 tmp2 = dataptr[2] + dataptr[5];
163 tmp5 = dataptr[2] - dataptr[5];
164 tmp3 = dataptr[3] + dataptr[4];
165 tmp4 = dataptr[3] - dataptr[4];
DRCe5eaf372014-05-09 18:00:32 +0000166
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000167 /* Even part per LL&M figure 1 --- note that published figure is faulty;
168 * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
169 */
DRCe5eaf372014-05-09 18:00:32 +0000170
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000171 tmp10 = tmp0 + tmp3;
172 tmp13 = tmp0 - tmp3;
173 tmp11 = tmp1 + tmp2;
174 tmp12 = tmp1 - tmp2;
DRCe5eaf372014-05-09 18:00:32 +0000175
DRC8e9cef22015-09-21 12:57:41 -0500176 dataptr[0] = (DCTELEM) LEFT_SHIFT(tmp10 + tmp11, PASS1_BITS);
177 dataptr[4] = (DCTELEM) LEFT_SHIFT(tmp10 - tmp11, PASS1_BITS);
DRCe5eaf372014-05-09 18:00:32 +0000178
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000179 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
180 dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
DRCe5eaf372014-05-09 18:00:32 +0000181 CONST_BITS-PASS1_BITS);
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000182 dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
DRCe5eaf372014-05-09 18:00:32 +0000183 CONST_BITS-PASS1_BITS);
184
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000185 /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
186 * cK represents cos(K*pi/16).
187 * i0..i3 in the paper are tmp4..tmp7 here.
188 */
DRCe5eaf372014-05-09 18:00:32 +0000189
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000190 z1 = tmp4 + tmp7;
191 z2 = tmp5 + tmp6;
192 z3 = tmp4 + tmp6;
193 z4 = tmp5 + tmp7;
194 z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
DRCe5eaf372014-05-09 18:00:32 +0000195
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000196 tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
197 tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
198 tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
199 tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
200 z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
201 z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
202 z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
203 z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
DRCe5eaf372014-05-09 18:00:32 +0000204
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000205 z3 += z5;
206 z4 += z5;
DRCe5eaf372014-05-09 18:00:32 +0000207
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000208 dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
209 dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
210 dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
211 dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
DRCe5eaf372014-05-09 18:00:32 +0000212
213 dataptr += DCTSIZE; /* advance pointer to next row */
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000214 }
Thomas G. Lane4a6b7301992-03-17 00:00:00 +0000215
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000216 /* Pass 2: process columns.
217 * We remove the PASS1_BITS scaling, but leave the results scaled up
218 * by an overall factor of 8.
219 */
Thomas G. Lane4a6b7301992-03-17 00:00:00 +0000220
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000221 dataptr = data;
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000222 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000223 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
224 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
225 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
226 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
227 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
228 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
229 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
230 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
DRCe5eaf372014-05-09 18:00:32 +0000231
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000232 /* Even part per LL&M figure 1 --- note that published figure is faulty;
233 * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
234 */
DRCe5eaf372014-05-09 18:00:32 +0000235
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000236 tmp10 = tmp0 + tmp3;
237 tmp13 = tmp0 - tmp3;
238 tmp11 = tmp1 + tmp2;
239 tmp12 = tmp1 - tmp2;
DRCe5eaf372014-05-09 18:00:32 +0000240
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000241 dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
242 dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
DRCe5eaf372014-05-09 18:00:32 +0000243
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000244 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
245 dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
DRCe5eaf372014-05-09 18:00:32 +0000246 CONST_BITS+PASS1_BITS);
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000247 dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
DRCe5eaf372014-05-09 18:00:32 +0000248 CONST_BITS+PASS1_BITS);
249
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000250 /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
251 * cK represents cos(K*pi/16).
252 * i0..i3 in the paper are tmp4..tmp7 here.
253 */
DRCe5eaf372014-05-09 18:00:32 +0000254
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000255 z1 = tmp4 + tmp7;
256 z2 = tmp5 + tmp6;
257 z3 = tmp4 + tmp6;
258 z4 = tmp5 + tmp7;
259 z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
DRCe5eaf372014-05-09 18:00:32 +0000260
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000261 tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
262 tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
263 tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
264 tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
265 z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
266 z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
267 z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
268 z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
DRCe5eaf372014-05-09 18:00:32 +0000269
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000270 z3 += z5;
271 z4 += z5;
DRCe5eaf372014-05-09 18:00:32 +0000272
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000273 dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp4 + z1 + z3,
DRCe5eaf372014-05-09 18:00:32 +0000274 CONST_BITS+PASS1_BITS);
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000275 dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp5 + z2 + z4,
DRCe5eaf372014-05-09 18:00:32 +0000276 CONST_BITS+PASS1_BITS);
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000277 dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp6 + z2 + z3,
DRCe5eaf372014-05-09 18:00:32 +0000278 CONST_BITS+PASS1_BITS);
Thomas G. Lane88aeed41992-12-10 00:00:00 +0000279 dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp7 + z1 + z4,
DRCe5eaf372014-05-09 18:00:32 +0000280 CONST_BITS+PASS1_BITS);
281
282 dataptr++; /* advance pointer to next column */
Thomas G. Lane4a6b7301992-03-17 00:00:00 +0000283 }
Thomas G. Lane2cbeb8a1991-10-07 00:00:00 +0000284}
Thomas G. Lane36a4ccc1994-09-24 00:00:00 +0000285
286#endif /* DCT_ISLOW_SUPPORTED */