1 /*
2 * jcdctmgr.c
3 *
4 * Copyright (C) 1994-1996, Thomas G. Lane.
5 * This file is part of the Independent JPEG Group's software.
6 * For conditions of distribution and use, see the accompanying README file.
7 *
8 * This file contains the forward-DCT management logic.
9 * This code selects a particular DCT implementation to be used,
10 * and it performs related housekeeping chores including coefficient
11 * quantization.
12 */
13
14 #define JPEG_INTERNALS
15 #include "jinclude.h"
16 #include "jpeglib.h"
17 #include "jdct.h" /* Private declarations for DCT subsystem */
18
19
20 /* Private subobject for this module */
21
22 typedef struct {
23 struct jpeg_forward_dct pub; /* public fields */
24
25 /* Pointer to the DCT routine actually in use */
26 forward_DCT_method_ptr do_dct;
27
28 /* The actual post-DCT divisors --- not identical to the quant table
29 * entries, because of scaling (especially for an unnormalized DCT).
30 * Each table is given in normal array order.
31 */
32 DCTELEM * divisors[NUM_QUANT_TBLS];
33
34 #ifdef DCT_FLOAT_SUPPORTED
35 /* Same as above for the floating-point case. */
36 float_DCT_method_ptr do_float_dct;
37 FAST_FLOAT * float_divisors[NUM_QUANT_TBLS];
38 #endif
39 } my_fdct_controller;
40
41 typedef my_fdct_controller * my_fdct_ptr;
42
43
44 /*
45 * Initialize for a processing pass.
46 * Verify that all referenced Q-tables are present, and set up
47 * the divisor table for each one.
48 * In the current implementation, DCT of all components is done during
49 * the first pass, even if only some components will be output in the
50 * first scan. Hence all components should be examined here.
51 */
52
53 METHODDEF(void)
start_pass_fdctmgr(j_compress_ptr cinfo)54 start_pass_fdctmgr (j_compress_ptr cinfo)
55 {
56 my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
57 int ci, qtblno, i;
58 jpeg_component_info *compptr;
59 JQUANT_TBL * qtbl;
60
61 for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
62 ci++, compptr++) {
63 qtblno = compptr->quant_tbl_no;
64 /* Make sure specified quantization table is present */
65 if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
66 cinfo->quant_tbl_ptrs[qtblno] == NULL)
67 ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
68 qtbl = cinfo->quant_tbl_ptrs[qtblno];
69 /* Compute divisors for this quant table */
70 /* We may do this more than once for same table, but it's not a big deal */
71 switch (cinfo->dct_method) {
72 #ifdef DCT_ISLOW_SUPPORTED
73 case JDCT_ISLOW:
74 /* For LL&M IDCT method, divisors are equal to raw quantization
75 * coefficients multiplied by 8 (to counteract scaling).
76 */
77 if (fdct->divisors[qtblno] == NULL) {
78 fdct->divisors[qtblno] = (DCTELEM *)
79 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
80 DCTSIZE2 * SIZEOF(DCTELEM));
81 }
82 dtbl = fdct->divisors[qtblno];
83 for (i = 0; i < DCTSIZE2; i++) {
84 dtbl[i] = ((DCTELEM) qtbl->quantval[i]) << 3;
85 }
86 break;
87 #endif
88 #ifdef DCT_IFAST_SUPPORTED
89 case JDCT_IFAST:
90 {
91 /* For AA&N IDCT method, divisors are equal to quantization
92 * coefficients scaled by scalefactor[row]*scalefactor[col], where
93 * scalefactor[0] = 1
94 * scalefactor[k] = cos(k*PI/16) * sqrt(2) for k=1..7
95 * We apply a further scale factor of 8.
96 */
97 #define CONST_BITS 14
98 static const INT16 aanscales[DCTSIZE2] = {
99 /* precomputed values scaled up by 14 bits */
100 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
101 22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
102 21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
103 19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
104 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
105 12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
106 8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
107 4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
108 };
109 SHIFT_TEMPS
110
111 if (fdct->divisors[qtblno] == NULL) {
112 fdct->divisors[qtblno] = (DCTELEM *)
113 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
114 DCTSIZE2 * SIZEOF(DCTELEM));
115 }
116 dtbl = fdct->divisors[qtblno];
117 for (i = 0; i < DCTSIZE2; i++) {
118 dtbl[i] = (DCTELEM)
119 DESCALE(MULTIPLY16V16((INT32) qtbl->quantval[i],
120 (INT32) aanscales[i]),
121 CONST_BITS-3);
122 }
123 }
124 break;
125 #endif
126 #ifdef DCT_FLOAT_SUPPORTED
127 case JDCT_FLOAT:
128 {
129 /* For float AA&N IDCT method, divisors are equal to quantization
130 * coefficients scaled by scalefactor[row]*scalefactor[col], where
131 * scalefactor[0] = 1
132 * scalefactor[k] = cos(k*PI/16) * sqrt(2) for k=1..7
133 * We apply a further scale factor of 8.
134 * What's actually stored is 1/divisor so that the inner loop can
135 * use a multiplication rather than a division.
136 */
137 FAST_FLOAT * fdtbl;
138 int row, col;
139 static const double aanscalefactor[DCTSIZE] = {
140 1.0, 1.387039845, 1.306562965, 1.175875602,
141 1.0, 0.785694958, 0.541196100, 0.275899379
142 };
143
144 if (fdct->float_divisors[qtblno] == NULL) {
145 fdct->float_divisors[qtblno] = (FAST_FLOAT *)
146 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
147 DCTSIZE2 * SIZEOF(FAST_FLOAT));
148 }
149 fdtbl = fdct->float_divisors[qtblno];
150 i = 0;
151 for (row = 0; row < DCTSIZE; row++) {
152 for (col = 0; col < DCTSIZE; col++) {
153 fdtbl[i] = (FAST_FLOAT)
154 (1.0 / (((double) qtbl->quantval[i] *
155 aanscalefactor[row] * aanscalefactor[col] * 8.0)));
156 i++;
157 }
158 }
159 }
160 break;
161 #endif
162 default:
163 ERREXIT(cinfo, JERR_NOT_COMPILED);
164 break;
165 }
166 }
167 }
168
169
170 /* code/jpeg-6b/jcdctmgr.c:184: warning: ‘forward_DCT’ defined but not used */
171 #if 0
172 /*
173 * Perform forward DCT on one or more blocks of a component.
174 *
175 * The input samples are taken from the sample_data[] array starting at
176 * position start_row/start_col, and moving to the right for any additional
177 * blocks. The quantized coefficients are returned in coef_blocks[].
178 */
179
180 METHODDEF(void)
181 forward_DCT (j_compress_ptr cinfo, jpeg_component_info * compptr,
182 JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
183 JDIMENSION start_row, JDIMENSION start_col,
184 JDIMENSION num_blocks)
185 /* This version is used for integer DCT implementations. */
186 {
187 /* This routine is heavily used, so it's worth coding it tightly. */
188 my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
189 forward_DCT_method_ptr do_dct = fdct->do_dct;
190 DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
191 DCTELEM workspace[DCTSIZE2]; /* work area for FDCT subroutine */
192 JDIMENSION bi;
193
194 sample_data += start_row; /* fold in the vertical offset once */
195
196 for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
197 /* Load data into workspace, applying unsigned->signed conversion */
198 { register DCTELEM *workspaceptr;
199 register JSAMPROW elemptr;
200 register int elemr;
201
202 workspaceptr = workspace;
203 for (elemr = 0; elemr < DCTSIZE; elemr++) {
204 elemptr = sample_data[elemr] + start_col;
205 #if DCTSIZE == 8 /* unroll the inner loop */
206 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
207 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
208 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
209 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
210 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
211 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
212 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
213 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
214 #else
215 { register int elemc;
216 for (elemc = DCTSIZE; elemc > 0; elemc--) {
217 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
218 }
219 }
220 #endif
221 }
222 }
223
224 /* Perform the DCT */
225 (*do_dct) (workspace);
226
227 /* Quantize/descale the coefficients, and store into coef_blocks[] */
228 { register DCTELEM temp, qval;
229 register int i;
230 register JCOEFPTR output_ptr = coef_blocks[bi];
231
232 for (i = 0; i < DCTSIZE2; i++) {
233 qval = divisors[i];
234 temp = workspace[i];
235 /* Divide the coefficient value by qval, ensuring proper rounding.
236 * Since C does not specify the direction of rounding for negative
237 * quotients, we have to force the dividend positive for portability.
238 *
239 * In most files, at least half of the output values will be zero
240 * (at default quantization settings, more like three-quarters...)
241 * so we should ensure that this case is fast. On many machines,
242 * a comparison is enough cheaper than a divide to make a special test
243 * a win. Since both inputs will be nonnegative, we need only test
244 * for a < b to discover whether a/b is 0.
245 * If your machine's division is fast enough, define FAST_DIVIDE.
246 */
247 #ifdef FAST_DIVIDE
248 #define DIVIDE_BY(a,b) a /= b
249 #else
250 #define DIVIDE_BY(a,b) if (a >= b) a /= b; else a = 0
251 #endif
252 if (temp < 0) {
253 temp = -temp;
254 temp += qval>>1; /* for rounding */
255 DIVIDE_BY(temp, qval);
256 temp = -temp;
257 } else {
258 temp += qval>>1; /* for rounding */
259 DIVIDE_BY(temp, qval);
260 }
261 output_ptr[i] = (JCOEF) temp;
262 }
263 }
264 }
265 }
266 #endif
267
268
269 #ifdef DCT_FLOAT_SUPPORTED
270
271 METHODDEF(void)
forward_DCT_float(j_compress_ptr cinfo,jpeg_component_info * compptr,JSAMPARRAY sample_data,JBLOCKROW coef_blocks,JDIMENSION start_row,JDIMENSION start_col,JDIMENSION num_blocks)272 forward_DCT_float (j_compress_ptr cinfo, jpeg_component_info * compptr,
273 JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
274 JDIMENSION start_row, JDIMENSION start_col,
275 JDIMENSION num_blocks)
276 /* This version is used for floating-point DCT implementations. */
277 {
278 /* This routine is heavily used, so it's worth coding it tightly. */
279 my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
280 float_DCT_method_ptr do_dct = fdct->do_float_dct;
281 FAST_FLOAT * divisors = fdct->float_divisors[compptr->quant_tbl_no];
282 FAST_FLOAT workspace[DCTSIZE2]; /* work area for FDCT subroutine */
283 JDIMENSION bi;
284
285 sample_data += start_row; /* fold in the vertical offset once */
286
287 for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
288 /* Load data into workspace, applying unsigned->signed conversion */
289 { register FAST_FLOAT *workspaceptr;
290 register JSAMPROW elemptr;
291 register int elemr;
292
293 workspaceptr = workspace;
294 for (elemr = 0; elemr < DCTSIZE; elemr++) {
295 elemptr = sample_data[elemr] + start_col;
296 #if DCTSIZE == 8 /* unroll the inner loop */
297 *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
298 *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
299 *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
300 *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
301 *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
302 *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
303 *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
304 *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
305 #else
306 { register int elemc;
307 for (elemc = DCTSIZE; elemc > 0; elemc--) {
308 *workspaceptr++ = (FAST_FLOAT)
309 (GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
310 }
311 }
312 #endif
313 }
314 }
315
316 /* Perform the DCT */
317 (*do_dct) (workspace);
318
319 /* Quantize/descale the coefficients, and store into coef_blocks[] */
320 { register FAST_FLOAT temp;
321 register int i;
322 register JCOEFPTR output_ptr = coef_blocks[bi];
323
324 for (i = 0; i < DCTSIZE2; i++) {
325 /* Apply the quantization and scaling factor */
326 temp = workspace[i] * divisors[i];
327 /* Round to nearest integer.
328 * Since C does not specify the direction of rounding for negative
329 * quotients, we have to force the dividend positive for portability.
330 * The maximum coefficient size is +-16K (for 12-bit data), so this
331 * code should work for either 16-bit or 32-bit ints.
332 */
333 output_ptr[i] = (JCOEF) ((int) (temp + (FAST_FLOAT) 16384.5) - 16384);
334 }
335 }
336 }
337 }
338
339 #endif /* DCT_FLOAT_SUPPORTED */
340
341
342 /*
343 * Initialize FDCT manager.
344 */
345
346 GLOBAL(void)
jinit_forward_dct(j_compress_ptr cinfo)347 jinit_forward_dct (j_compress_ptr cinfo)
348 {
349 my_fdct_ptr fdct;
350 int i;
351
352 fdct = (my_fdct_ptr)
353 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
354 SIZEOF(my_fdct_controller));
355 cinfo->fdct = (struct jpeg_forward_dct *) fdct;
356 fdct->pub.start_pass = start_pass_fdctmgr;
357
358 switch (cinfo->dct_method) {
359 #ifdef DCT_ISLOW_SUPPORTED
360 case JDCT_ISLOW:
361 fdct->pub.forward_DCT = forward_DCT;
362 fdct->do_dct = jpeg_fdct_islow;
363 break;
364 #endif
365 #ifdef DCT_IFAST_SUPPORTED
366 case JDCT_IFAST:
367 fdct->pub.forward_DCT = forward_DCT;
368 fdct->do_dct = jpeg_fdct_ifast;
369 break;
370 #endif
371 #ifdef DCT_FLOAT_SUPPORTED
372 case JDCT_FLOAT:
373 fdct->pub.forward_DCT = forward_DCT_float;
374 fdct->do_float_dct = jpeg_fdct_float;
375 break;
376 #endif
377 default:
378 ERREXIT(cinfo, JERR_NOT_COMPILED);
379 break;
380 }
381
382 /* Mark divisor tables unallocated */
383 for (i = 0; i < NUM_QUANT_TBLS; i++) {
384 fdct->divisors[i] = NULL;
385 #ifdef DCT_FLOAT_SUPPORTED
386 fdct->float_divisors[i] = NULL;
387 #endif
388 }
389 }
390