• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * jfdctint.c
3  *
4  * This file was part of the Independent JPEG Group's software:
5  * Copyright (C) 1991-1996, Thomas G. Lane.
6  * libjpeg-turbo Modifications:
7  * Copyright (C) 2015, 2020, D. R. Commander.
8  * For conditions of distribution and use, see the accompanying README.ijg
9  * file.
10  *
11  * This file contains a slower but more 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.
17  *
18  * 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.
27  */
28 
29 #define JPEG_INTERNALS
30 #include "jinclude.h"
31 #include "jpeglib.h"
32 #include "jdct.h"               /* Private declarations for DCT subsystem */
33 
34 #ifdef DCT_ISLOW_SUPPORTED
35 
36 
37 /*
38  * This module is specialized to the case DCTSIZE = 8.
39  */
40 
41 #if DCTSIZE != 8
42   Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
43 #endif
44 
45 
46 /*
47  * 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).
55  * In the IJG code, this factor of 8 is removed by the quantization step
56  * (in jcdctmgr.c), NOT in this module.
57  *
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
61  * 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.
68  *
69  * 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
72  * with the recommended scaling.  (For 12-bit sample data, the intermediate
73  * array is JLONG anyway.)
74  *
75  * To avoid overflow of the 32-bit intermediate results in pass 2, we must
76  * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
77  * shows that the values given below are the most effective.
78  */
79 
80 #if BITS_IN_JSAMPLE == 8
81 #define CONST_BITS  13
82 #define PASS1_BITS  2
83 #else
84 #define CONST_BITS  13
85 #define PASS1_BITS  1           /* lose a little precision to avoid overflow */
86 #endif
87 
88 /* 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  */
94 
95 #if CONST_BITS == 13
96 #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) */
108 #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
122 
123 
124 /* Multiply an JLONG variable by an JLONG constant to yield an JLONG result.
125  * 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
127  * 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.
129  */
130 
131 #if BITS_IN_JSAMPLE == 8
132 #define MULTIPLY(var, const)  MULTIPLY16C16(var, const)
133 #else
134 #define MULTIPLY(var, const)  ((var) * (const))
135 #endif
136 
137 
138 /*
139  * Perform the forward DCT on one block of samples.
140  */
141 
142 GLOBAL(void)
143 jpeg_fdct_islow(DCTELEM *data)
144 {
145   JLONG tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
146   JLONG tmp10, tmp11, tmp12, tmp13;
147   JLONG z1, z2, z3, z4, z5;
148   DCTELEM *dataptr;
149   int ctr;
150   SHIFT_TEMPS
151 
152   /* 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. */
155 
156   dataptr = data;
157   for (ctr = DCTSIZE - 1; ctr >= 0; ctr--) {
158     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];
166 
167     /* Even part per LL&M figure 1 --- note that published figure is faulty;
168      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
169      */
170 
171     tmp10 = tmp0 + tmp3;
172     tmp13 = tmp0 - tmp3;
173     tmp11 = tmp1 + tmp2;
174     tmp12 = tmp1 - tmp2;
175 
176     dataptr[0] = (DCTELEM)LEFT_SHIFT(tmp10 + tmp11, PASS1_BITS);
177     dataptr[4] = (DCTELEM)LEFT_SHIFT(tmp10 - tmp11, PASS1_BITS);
178 
179     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
180     dataptr[2] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
181                                   CONST_BITS - PASS1_BITS);
182     dataptr[6] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp12, -FIX_1_847759065),
183                                   CONST_BITS - PASS1_BITS);
184 
185     /* 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      */
189 
190     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 */
195 
196     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) */
204 
205     z3 += z5;
206     z4 += z5;
207 
208     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);
212 
213     dataptr += DCTSIZE;         /* advance pointer to next row */
214   }
215 
216   /* 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    */
220 
221   dataptr = data;
222   for (ctr = DCTSIZE - 1; ctr >= 0; ctr--) {
223     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];
231 
232     /* Even part per LL&M figure 1 --- note that published figure is faulty;
233      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
234      */
235 
236     tmp10 = tmp0 + tmp3;
237     tmp13 = tmp0 - tmp3;
238     tmp11 = tmp1 + tmp2;
239     tmp12 = tmp1 - tmp2;
240 
241     dataptr[DCTSIZE * 0] = (DCTELEM)DESCALE(tmp10 + tmp11, PASS1_BITS);
242     dataptr[DCTSIZE * 4] = (DCTELEM)DESCALE(tmp10 - tmp11, PASS1_BITS);
243 
244     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
245     dataptr[DCTSIZE * 2] =
246       (DCTELEM)DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
247                        CONST_BITS + PASS1_BITS);
248     dataptr[DCTSIZE * 6] =
249       (DCTELEM)DESCALE(z1 + MULTIPLY(tmp12, -FIX_1_847759065),
250                        CONST_BITS + PASS1_BITS);
251 
252     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
253      * cK represents cos(K*pi/16).
254      * i0..i3 in the paper are tmp4..tmp7 here.
255      */
256 
257     z1 = tmp4 + tmp7;
258     z2 = tmp5 + tmp6;
259     z3 = tmp4 + tmp6;
260     z4 = tmp5 + tmp7;
261     z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
262 
263     tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
264     tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
265     tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
266     tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
267     z1 = MULTIPLY(z1, -FIX_0_899976223); /* sqrt(2) * ( c7-c3) */
268     z2 = MULTIPLY(z2, -FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
269     z3 = MULTIPLY(z3, -FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
270     z4 = MULTIPLY(z4, -FIX_0_390180644); /* sqrt(2) * ( c5-c3) */
271 
272     z3 += z5;
273     z4 += z5;
274 
275     dataptr[DCTSIZE * 7] = (DCTELEM)DESCALE(tmp4 + z1 + z3,
276                                             CONST_BITS + PASS1_BITS);
277     dataptr[DCTSIZE * 5] = (DCTELEM)DESCALE(tmp5 + z2 + z4,
278                                             CONST_BITS + PASS1_BITS);
279     dataptr[DCTSIZE * 3] = (DCTELEM)DESCALE(tmp6 + z2 + z3,
280                                             CONST_BITS + PASS1_BITS);
281     dataptr[DCTSIZE * 1] = (DCTELEM)DESCALE(tmp7 + z1 + z4,
282                                             CONST_BITS + PASS1_BITS);
283 
284     dataptr++;                  /* advance pointer to next column */
285   }
286 }
287 
288 #endif /* DCT_ISLOW_SUPPORTED */
289