• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * jidctflt.c
3  *
4  * This file was part of the Independent JPEG Group's software:
5  * Copyright (C) 1994-1998, Thomas G. Lane.
6  * Modified 2010 by Guido Vollbeding.
7  * libjpeg-turbo Modifications:
8  * Copyright (C) 2014, D. R. Commander.
9  * For conditions of distribution and use, see the accompanying README.ijg
10  * file.
11  *
12  * This file contains a floating-point implementation of the
13  * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
14  * must also perform dequantization of the input coefficients.
15  *
16  * This implementation should be more accurate than either of the integer
17  * IDCT implementations.  However, it may not give the same results on all
18  * machines because of differences in roundoff behavior.  Speed will depend
19  * on the hardware's floating point capacity.
20  *
21  * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
22  * on each row (or vice versa, but it's more convenient to emit a row at
23  * a time).  Direct algorithms are also available, but they are much more
24  * complex and seem not to be any faster when reduced to code.
25  *
26  * This implementation is based on Arai, Agui, and Nakajima's algorithm for
27  * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
28  * Japanese, but the algorithm is described in the Pennebaker & Mitchell
29  * JPEG textbook (see REFERENCES section in file README.ijg).  The following
30  * code is based directly on figure 4-8 in P&M.
31  * While an 8-point DCT cannot be done in less than 11 multiplies, it is
32  * possible to arrange the computation so that many of the multiplies are
33  * simple scalings of the final outputs.  These multiplies can then be
34  * folded into the multiplications or divisions by the JPEG quantization
35  * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
36  * to be done in the DCT itself.
37  * The primary disadvantage of this method is that with a fixed-point
38  * implementation, accuracy is lost due to imprecise representation of the
39  * scaled quantization values.  However, that problem does not arise if
40  * we use floating point arithmetic.
41  */
42 
43 #define JPEG_INTERNALS
44 #include "jinclude.h"
45 #include "jpeglib.h"
46 #include "jdct.h"               /* Private declarations for DCT subsystem */
47 
48 #ifdef DCT_FLOAT_SUPPORTED
49 
50 
51 /*
52  * This module is specialized to the case DCTSIZE = 8.
53  */
54 
55 #if DCTSIZE != 8
56   Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
57 #endif
58 
59 
60 /* Dequantize a coefficient by multiplying it by the multiplier-table
61  * entry; produce a float result.
62  */
63 
64 #define DEQUANTIZE(coef, quantval)  (((FAST_FLOAT)(coef)) * (quantval))
65 
66 
67 /*
68  * Perform dequantization and inverse DCT on one block of coefficients.
69  */
70 
71 GLOBAL(void)
72 jpeg_idct_float(j_decompress_ptr cinfo, jpeg_component_info *compptr,
73                 JCOEFPTR coef_block, JSAMPARRAY output_buf,
74                 JDIMENSION output_col)
75 {
76   FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
77   FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
78   FAST_FLOAT z5, z10, z11, z12, z13;
79   JCOEFPTR inptr;
80   FLOAT_MULT_TYPE *quantptr;
81   FAST_FLOAT *wsptr;
82   JSAMPROW outptr;
83   JSAMPLE *range_limit = cinfo->sample_range_limit;
84   int ctr;
85   FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
86 #define _0_125  ((FLOAT_MULT_TYPE)0.125)
87 
88   /* Pass 1: process columns from input, store into work array. */
89 
90   inptr = coef_block;
91   quantptr = (FLOAT_MULT_TYPE *)compptr->dct_table;
92   wsptr = workspace;
93   for (ctr = DCTSIZE; ctr > 0; ctr--) {
94     /* Due to quantization, we will usually find that many of the input
95      * coefficients are zero, especially the AC terms.  We can exploit this
96      * by short-circuiting the IDCT calculation for any column in which all
97      * the AC terms are zero.  In that case each output is equal to the
98      * DC coefficient (with scale factor as needed).
99      * With typical images and quantization tables, half or more of the
100      * column DCT calculations can be simplified this way.
101      */
102 
103     if (inptr[DCTSIZE * 1] == 0 && inptr[DCTSIZE * 2] == 0 &&
104         inptr[DCTSIZE * 3] == 0 && inptr[DCTSIZE * 4] == 0 &&
105         inptr[DCTSIZE * 5] == 0 && inptr[DCTSIZE * 6] == 0 &&
106         inptr[DCTSIZE * 7] == 0) {
107       /* AC terms all zero */
108       FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE * 0],
109                                     quantptr[DCTSIZE * 0] * _0_125);
110 
111       wsptr[DCTSIZE * 0] = dcval;
112       wsptr[DCTSIZE * 1] = dcval;
113       wsptr[DCTSIZE * 2] = dcval;
114       wsptr[DCTSIZE * 3] = dcval;
115       wsptr[DCTSIZE * 4] = dcval;
116       wsptr[DCTSIZE * 5] = dcval;
117       wsptr[DCTSIZE * 6] = dcval;
118       wsptr[DCTSIZE * 7] = dcval;
119 
120       inptr++;                  /* advance pointers to next column */
121       quantptr++;
122       wsptr++;
123       continue;
124     }
125 
126     /* Even part */
127 
128     tmp0 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0] * _0_125);
129     tmp1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2] * _0_125);
130     tmp2 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4] * _0_125);
131     tmp3 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6] * _0_125);
132 
133     tmp10 = tmp0 + tmp2;        /* phase 3 */
134     tmp11 = tmp0 - tmp2;
135 
136     tmp13 = tmp1 + tmp3;        /* phases 5-3 */
137     tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT)1.414213562) - tmp13; /* 2*c4 */
138 
139     tmp0 = tmp10 + tmp13;       /* phase 2 */
140     tmp3 = tmp10 - tmp13;
141     tmp1 = tmp11 + tmp12;
142     tmp2 = tmp11 - tmp12;
143 
144     /* Odd part */
145 
146     tmp4 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1] * _0_125);
147     tmp5 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3] * _0_125);
148     tmp6 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5] * _0_125);
149     tmp7 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7] * _0_125);
150 
151     z13 = tmp6 + tmp5;          /* phase 6 */
152     z10 = tmp6 - tmp5;
153     z11 = tmp4 + tmp7;
154     z12 = tmp4 - tmp7;
155 
156     tmp7 = z11 + z13;           /* phase 5 */
157     tmp11 = (z11 - z13) * ((FAST_FLOAT)1.414213562); /* 2*c4 */
158 
159     z5 = (z10 + z12) * ((FAST_FLOAT)1.847759065); /* 2*c2 */
160     tmp10 = z5 - z12 * ((FAST_FLOAT)1.082392200); /* 2*(c2-c6) */
161     tmp12 = z5 - z10 * ((FAST_FLOAT)2.613125930); /* 2*(c2+c6) */
162 
163     tmp6 = tmp12 - tmp7;        /* phase 2 */
164     tmp5 = tmp11 - tmp6;
165     tmp4 = tmp10 - tmp5;
166 
167     wsptr[DCTSIZE * 0] = tmp0 + tmp7;
168     wsptr[DCTSIZE * 7] = tmp0 - tmp7;
169     wsptr[DCTSIZE * 1] = tmp1 + tmp6;
170     wsptr[DCTSIZE * 6] = tmp1 - tmp6;
171     wsptr[DCTSIZE * 2] = tmp2 + tmp5;
172     wsptr[DCTSIZE * 5] = tmp2 - tmp5;
173     wsptr[DCTSIZE * 3] = tmp3 + tmp4;
174     wsptr[DCTSIZE * 4] = tmp3 - tmp4;
175 
176     inptr++;                    /* advance pointers to next column */
177     quantptr++;
178     wsptr++;
179   }
180 
181   /* Pass 2: process rows from work array, store into output array. */
182 
183   wsptr = workspace;
184   for (ctr = 0; ctr < DCTSIZE; ctr++) {
185     outptr = output_buf[ctr] + output_col;
186     /* Rows of zeroes can be exploited in the same way as we did with columns.
187      * However, the column calculation has created many nonzero AC terms, so
188      * the simplification applies less often (typically 5% to 10% of the time).
189      * And testing floats for zero is relatively expensive, so we don't bother.
190      */
191 
192     /* Even part */
193 
194     /* Apply signed->unsigned and prepare float->int conversion */
195     z5 = wsptr[0] + ((FAST_FLOAT)CENTERJSAMPLE + (FAST_FLOAT)0.5);
196     tmp10 = z5 + wsptr[4];
197     tmp11 = z5 - wsptr[4];
198 
199     tmp13 = wsptr[2] + wsptr[6];
200     tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT)1.414213562) - tmp13;
201 
202     tmp0 = tmp10 + tmp13;
203     tmp3 = tmp10 - tmp13;
204     tmp1 = tmp11 + tmp12;
205     tmp2 = tmp11 - tmp12;
206 
207     /* Odd part */
208 
209     z13 = wsptr[5] + wsptr[3];
210     z10 = wsptr[5] - wsptr[3];
211     z11 = wsptr[1] + wsptr[7];
212     z12 = wsptr[1] - wsptr[7];
213 
214     tmp7 = z11 + z13;
215     tmp11 = (z11 - z13) * ((FAST_FLOAT)1.414213562);
216 
217     z5 = (z10 + z12) * ((FAST_FLOAT)1.847759065); /* 2*c2 */
218     tmp10 = z5 - z12 * ((FAST_FLOAT)1.082392200); /* 2*(c2-c6) */
219     tmp12 = z5 - z10 * ((FAST_FLOAT)2.613125930); /* 2*(c2+c6) */
220 
221     tmp6 = tmp12 - tmp7;
222     tmp5 = tmp11 - tmp6;
223     tmp4 = tmp10 - tmp5;
224 
225     /* Final output stage: float->int conversion and range-limit */
226 
227     outptr[0] = range_limit[((int)(tmp0 + tmp7)) & RANGE_MASK];
228     outptr[7] = range_limit[((int)(tmp0 - tmp7)) & RANGE_MASK];
229     outptr[1] = range_limit[((int)(tmp1 + tmp6)) & RANGE_MASK];
230     outptr[6] = range_limit[((int)(tmp1 - tmp6)) & RANGE_MASK];
231     outptr[2] = range_limit[((int)(tmp2 + tmp5)) & RANGE_MASK];
232     outptr[5] = range_limit[((int)(tmp2 - tmp5)) & RANGE_MASK];
233     outptr[3] = range_limit[((int)(tmp3 + tmp4)) & RANGE_MASK];
234     outptr[4] = range_limit[((int)(tmp3 - tmp4)) & RANGE_MASK];
235 
236     wsptr += DCTSIZE;           /* advance pointer to next row */
237   }
238 }
239 
240 #endif /* DCT_FLOAT_SUPPORTED */
241