1 #if !defined(_FX_JPEG_TURBO_)
2 /*
3  * jidctint.c
4  *
5  * Copyright (C) 1991-1998, Thomas G. Lane.
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 slow-but-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 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 IDCT step produces outputs which are a factor of sqrt(N)
50  * larger than the true IDCT 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 IDCT,
54  * because the y0 and y4 inputs need not be divided by sqrt(N).
55  *
56  * We have to do addition and subtraction of the integer inputs, which
57  * is no problem, and multiplication by fractional constants, which is
58  * a problem to do in integer arithmetic.  We multiply all the constants
59  * by CONST_SCALE and convert them to integer constants (thus retaining
60  * CONST_BITS bits of precision in the constants).  After doing a
61  * multiplication we have to divide the product by CONST_SCALE, with proper
62  * rounding, to produce the correct output.  This division can be done
63  * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
64  * as long as possible so that partial sums can be added together with
65  * full fractional precision.
66  *
67  * The outputs of the first pass are scaled up by PASS1_BITS bits so that
68  * they are represented to better-than-integral precision.  These outputs
69  * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
70  * with the recommended scaling.  (To scale up 12-bit sample data further, an
71  * intermediate INT32 array would be needed.)
72  *
73  * To avoid overflow of the 32-bit intermediate results in pass 2, we must
74  * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
75  * shows that the values given below are the most effective.
76  */
77 
78 #if BITS_IN_JSAMPLE == 8
79 #define CONST_BITS  13
80 #define PASS1_BITS  2
81 #else
82 #define CONST_BITS  13
83 #define PASS1_BITS  1		/* lose a little precision to avoid overflow */
84 #endif
85 
86 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
87  * causing a lot of useless floating-point operations at run time.
88  * To get around this we use the following pre-calculated constants.
89  * If you change CONST_BITS you may want to add appropriate values.
90  * (With a reasonable C compiler, you can just rely on the FIX() macro...)
91  */
92 
93 #if CONST_BITS == 13
94 #define FIX_0_298631336  ((INT32)  2446)	/* FIX(0.298631336) */
95 #define FIX_0_390180644  ((INT32)  3196)	/* FIX(0.390180644) */
96 #define FIX_0_541196100  ((INT32)  4433)	/* FIX(0.541196100) */
97 #define FIX_0_765366865  ((INT32)  6270)	/* FIX(0.765366865) */
98 #define FIX_0_899976223  ((INT32)  7373)	/* FIX(0.899976223) */
99 #define FIX_1_175875602  ((INT32)  9633)	/* FIX(1.175875602) */
100 #define FIX_1_501321110  ((INT32)  12299)	/* FIX(1.501321110) */
101 #define FIX_1_847759065  ((INT32)  15137)	/* FIX(1.847759065) */
102 #define FIX_1_961570560  ((INT32)  16069)	/* FIX(1.961570560) */
103 #define FIX_2_053119869  ((INT32)  16819)	/* FIX(2.053119869) */
104 #define FIX_2_562915447  ((INT32)  20995)	/* FIX(2.562915447) */
105 #define FIX_3_072711026  ((INT32)  25172)	/* FIX(3.072711026) */
106 #else
107 #define FIX_0_298631336  FIX(0.298631336)
108 #define FIX_0_390180644  FIX(0.390180644)
109 #define FIX_0_541196100  FIX(0.541196100)
110 #define FIX_0_765366865  FIX(0.765366865)
111 #define FIX_0_899976223  FIX(0.899976223)
112 #define FIX_1_175875602  FIX(1.175875602)
113 #define FIX_1_501321110  FIX(1.501321110)
114 #define FIX_1_847759065  FIX(1.847759065)
115 #define FIX_1_961570560  FIX(1.961570560)
116 #define FIX_2_053119869  FIX(2.053119869)
117 #define FIX_2_562915447  FIX(2.562915447)
118 #define FIX_3_072711026  FIX(3.072711026)
119 #endif
120 
121 
122 /* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
123  * For 8-bit samples with the recommended scaling, all the variable
124  * and constant values involved are no more than 16 bits wide, so a
125  * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
126  * For 12-bit samples, a full 32-bit multiplication will be needed.
127  */
128 
129 #if BITS_IN_JSAMPLE == 8
130 #define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
131 #else
132 #define MULTIPLY(var,const)  ((var) * (const))
133 #endif
134 
135 
136 /* Dequantize a coefficient by multiplying it by the multiplier-table
137  * entry; produce an int result.  In this module, both inputs and result
138  * are 16 bits or less, so either int or short multiply will work.
139  */
140 
141 #define DEQUANTIZE(coef,quantval)  (((ISLOW_MULT_TYPE) (coef)) * (quantval))
142 
143 
144 /*
145  * Perform dequantization and inverse DCT on one block of coefficients.
146  */
147 
148 GLOBAL(void)
149 jpeg_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
150 		 JCOEFPTR coef_block,
151 		 JSAMPARRAY output_buf, JDIMENSION output_col)
152 {
153   INT32 tmp0, tmp1, tmp2, tmp3;
154   INT32 tmp10, tmp11, tmp12, tmp13;
155   INT32 z1, z2, z3, z4, z5;
156   JCOEFPTR inptr;
157   ISLOW_MULT_TYPE * quantptr;
158   int * wsptr;
159   JSAMPROW outptr;
160   JSAMPLE *range_limit = IDCT_range_limit(cinfo);
161   int ctr;
162   int workspace[DCTSIZE2];	/* buffers data between passes */
163   SHIFT_TEMPS
164 
165   /* Pass 1: process columns from input, store into work array. */
166   /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
167   /* furthermore, we scale the results by 2**PASS1_BITS. */
168 
169   inptr = coef_block;
170   quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
171   wsptr = workspace;
172   for (ctr = DCTSIZE; ctr > 0; ctr--) {
173     /* Due to quantization, we will usually find that many of the input
174      * coefficients are zero, especially the AC terms.  We can exploit this
175      * by short-circuiting the IDCT calculation for any column in which all
176      * the AC terms are zero.  In that case each output is equal to the
177      * DC coefficient (with scale factor as needed).
178      * With typical images and quantization tables, half or more of the
179      * column DCT calculations can be simplified this way.
180      */
181 
182     if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
183 	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
184 	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
185 	inptr[DCTSIZE*7] == 0) {
186       /* AC terms all zero */
187       int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
188 
189       wsptr[DCTSIZE*0] = dcval;
190       wsptr[DCTSIZE*1] = dcval;
191       wsptr[DCTSIZE*2] = dcval;
192       wsptr[DCTSIZE*3] = dcval;
193       wsptr[DCTSIZE*4] = dcval;
194       wsptr[DCTSIZE*5] = dcval;
195       wsptr[DCTSIZE*6] = dcval;
196       wsptr[DCTSIZE*7] = dcval;
197 
198       inptr++;			/* advance pointers to next column */
199       quantptr++;
200       wsptr++;
201       continue;
202     }
203 
204     /* Even part: reverse the even part of the forward DCT. */
205     /* The rotator is sqrt(2)*c(-6). */
206 
207     z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
208     z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
209 
210     z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
211     tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065);
212     tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865);
213 
214     z2 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
215     z3 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
216 
217     tmp0 = (z2 + z3) << CONST_BITS;
218     tmp1 = (z2 - z3) << CONST_BITS;
219 
220     tmp10 = tmp0 + tmp3;
221     tmp13 = tmp0 - tmp3;
222     tmp11 = tmp1 + tmp2;
223     tmp12 = tmp1 - tmp2;
224 
225     /* Odd part per figure 8; the matrix is unitary and hence its
226      * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
227      */
228 
229     tmp0 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
230     tmp1 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
231     tmp2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
232     tmp3 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
233 
234     z1 = tmp0 + tmp3;
235     z2 = tmp1 + tmp2;
236     z3 = tmp0 + tmp2;
237     z4 = tmp1 + tmp3;
238     z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
239 
240     tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
241     tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
242     tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
243     tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
244     z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
245     z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
246     z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
247     z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
248 
249     z3 += z5;
250     z4 += z5;
251 
252     tmp0 += z1 + z3;
253     tmp1 += z2 + z4;
254     tmp2 += z2 + z3;
255     tmp3 += z1 + z4;
256 
257     /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
258 
259     wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS);
260     wsptr[DCTSIZE*7] = (int) DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS);
261     wsptr[DCTSIZE*1] = (int) DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS);
262     wsptr[DCTSIZE*6] = (int) DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS);
263     wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS);
264     wsptr[DCTSIZE*5] = (int) DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS);
265     wsptr[DCTSIZE*3] = (int) DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS);
266     wsptr[DCTSIZE*4] = (int) DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS);
267 
268     inptr++;			/* advance pointers to next column */
269     quantptr++;
270     wsptr++;
271   }
272 
273   /* Pass 2: process rows from work array, store into output array. */
274   /* Note that we must descale the results by a factor of 8 == 2**3, */
275   /* and also undo the PASS1_BITS scaling. */
276 
277   wsptr = workspace;
278   for (ctr = 0; ctr < DCTSIZE; ctr++) {
279     outptr = output_buf[ctr] + output_col;
280     /* Rows of zeroes can be exploited in the same way as we did with columns.
281      * However, the column calculation has created many nonzero AC terms, so
282      * the simplification applies less often (typically 5% to 10% of the time).
283      * On machines with very fast multiplication, it's possible that the
284      * test takes more time than it's worth.  In that case this section
285      * may be commented out.
286      */
287 
288 #ifndef NO_ZERO_ROW_TEST
289     if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
290 	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
291       /* AC terms all zero */
292       JSAMPLE dcval = range_limit[(int) DESCALE((INT32) wsptr[0], PASS1_BITS+3)
293 				  & RANGE_MASK];
294 
295       outptr[0] = dcval;
296       outptr[1] = dcval;
297       outptr[2] = dcval;
298       outptr[3] = dcval;
299       outptr[4] = dcval;
300       outptr[5] = dcval;
301       outptr[6] = dcval;
302       outptr[7] = dcval;
303 
304       wsptr += DCTSIZE;		/* advance pointer to next row */
305       continue;
306     }
307 #endif
308 
309     /* Even part: reverse the even part of the forward DCT. */
310     /* The rotator is sqrt(2)*c(-6). */
311 
312     z2 = (INT32) wsptr[2];
313     z3 = (INT32) wsptr[6];
314 
315     z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
316     tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065);
317     tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865);
318 
319     tmp0 = ((INT32) wsptr[0] + (INT32) wsptr[4]) << CONST_BITS;
320     tmp1 = ((INT32) wsptr[0] - (INT32) wsptr[4]) << CONST_BITS;
321 
322     tmp10 = tmp0 + tmp3;
323     tmp13 = tmp0 - tmp3;
324     tmp11 = tmp1 + tmp2;
325     tmp12 = tmp1 - tmp2;
326 
327     /* Odd part per figure 8; the matrix is unitary and hence its
328      * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
329      */
330 
331     tmp0 = (INT32) wsptr[7];
332     tmp1 = (INT32) wsptr[5];
333     tmp2 = (INT32) wsptr[3];
334     tmp3 = (INT32) wsptr[1];
335 
336     z1 = tmp0 + tmp3;
337     z2 = tmp1 + tmp2;
338     z3 = tmp0 + tmp2;
339     z4 = tmp1 + tmp3;
340     z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
341 
342     tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
343     tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
344     tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
345     tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
346     z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
347     z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
348     z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
349     z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
350 
351     z3 += z5;
352     z4 += z5;
353 
354     tmp0 += z1 + z3;
355     tmp1 += z2 + z4;
356     tmp2 += z2 + z3;
357     tmp3 += z1 + z4;
358 
359     /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
360 
361     outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp3,
362 					  CONST_BITS+PASS1_BITS+3)
363 			    & RANGE_MASK];
364     outptr[7] = range_limit[(int) DESCALE(tmp10 - tmp3,
365 					  CONST_BITS+PASS1_BITS+3)
366 			    & RANGE_MASK];
367     outptr[1] = range_limit[(int) DESCALE(tmp11 + tmp2,
368 					  CONST_BITS+PASS1_BITS+3)
369 			    & RANGE_MASK];
370     outptr[6] = range_limit[(int) DESCALE(tmp11 - tmp2,
371 					  CONST_BITS+PASS1_BITS+3)
372 			    & RANGE_MASK];
373     outptr[2] = range_limit[(int) DESCALE(tmp12 + tmp1,
374 					  CONST_BITS+PASS1_BITS+3)
375 			    & RANGE_MASK];
376     outptr[5] = range_limit[(int) DESCALE(tmp12 - tmp1,
377 					  CONST_BITS+PASS1_BITS+3)
378 			    & RANGE_MASK];
379     outptr[3] = range_limit[(int) DESCALE(tmp13 + tmp0,
380 					  CONST_BITS+PASS1_BITS+3)
381 			    & RANGE_MASK];
382     outptr[4] = range_limit[(int) DESCALE(tmp13 - tmp0,
383 					  CONST_BITS+PASS1_BITS+3)
384 			    & RANGE_MASK];
385 
386     wsptr += DCTSIZE;		/* advance pointer to next row */
387   }
388 }
389 
390 #endif /* DCT_ISLOW_SUPPORTED */
391 
392 #endif //_FX_JPEG_TURBO_
393