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.
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
14 #define JPEG_INTERNALS
17 #include "jdct.h" /* Private declarations for DCT subsystem */
20 /* Private subobject for this module */
23 struct jpeg_forward_dct pub; /* public fields */
25 /* Pointer to the DCT routine actually in use */
26 forward_DCT_method_ptr do_dct;
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.
32 DCTELEM * divisors[NUM_QUANT_TBLS];
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];
41 typedef my_fdct_controller * my_fdct_ptr;
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.
54 start_pass_fdctmgr (j_compress_ptr cinfo)
56 my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
58 jpeg_component_info *compptr;
62 for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
64 qtblno = compptr->quant_tbl_no;
65 /* Make sure specified quantization table is present */
66 if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
67 cinfo->quant_tbl_ptrs[qtblno] == NULL)
68 ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
69 qtbl = cinfo->quant_tbl_ptrs[qtblno];
70 /* Compute divisors for this quant table */
71 /* We may do this more than once for same table, but it's not a big deal */
72 switch (cinfo->dct_method) {
73 #ifdef DCT_ISLOW_SUPPORTED
75 /* For LL&M IDCT method, divisors are equal to raw quantization
76 * coefficients multiplied by 8 (to counteract scaling).
78 if (fdct->divisors[qtblno] == NULL) {
79 fdct->divisors[qtblno] = (DCTELEM *)
80 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
81 DCTSIZE2 * SIZEOF(DCTELEM));
83 dtbl = fdct->divisors[qtblno];
84 for (i = 0; i < DCTSIZE2; i++) {
85 dtbl[i] = ((DCTELEM) qtbl->quantval[i]) << 3;
89 #ifdef DCT_IFAST_SUPPORTED
92 /* For AA&N IDCT method, divisors are equal to quantization
93 * coefficients scaled by scalefactor[row]*scalefactor[col], where
95 * scalefactor[k] = cos(k*PI/16) * sqrt(2) for k=1..7
96 * We apply a further scale factor of 8.
99 static const INT16 aanscales[DCTSIZE2] = {
100 /* precomputed values scaled up by 14 bits */
101 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
102 22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
103 21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
104 19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
105 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
106 12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
107 8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
108 4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
112 if (fdct->divisors[qtblno] == NULL) {
113 fdct->divisors[qtblno] = (DCTELEM *)
114 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
115 DCTSIZE2 * SIZEOF(DCTELEM));
117 dtbl = fdct->divisors[qtblno];
118 for (i = 0; i < DCTSIZE2; i++) {
120 DESCALE(MULTIPLY16V16((INT32) qtbl->quantval[i],
121 (INT32) aanscales[i]),
127 #ifdef DCT_FLOAT_SUPPORTED
130 /* For float AA&N IDCT method, divisors are equal to quantization
131 * coefficients scaled by scalefactor[row]*scalefactor[col], where
133 * scalefactor[k] = cos(k*PI/16) * sqrt(2) for k=1..7
134 * We apply a further scale factor of 8.
135 * What's actually stored is 1/divisor so that the inner loop can
136 * use a multiplication rather than a division.
140 static const double aanscalefactor[DCTSIZE] = {
141 1.0, 1.387039845, 1.306562965, 1.175875602,
142 1.0, 0.785694958, 0.541196100, 0.275899379
145 if (fdct->float_divisors[qtblno] == NULL) {
146 fdct->float_divisors[qtblno] = (FAST_FLOAT *)
147 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
148 DCTSIZE2 * SIZEOF(FAST_FLOAT));
150 fdtbl = fdct->float_divisors[qtblno];
152 for (row = 0; row < DCTSIZE; row++) {
153 for (col = 0; col < DCTSIZE; col++) {
154 fdtbl[i] = (FAST_FLOAT)
155 (1.0 / (((double) qtbl->quantval[i] *
156 aanscalefactor[row] * aanscalefactor[col] * 8.0)));
164 ERREXIT(cinfo, JERR_NOT_COMPILED);
172 * Perform forward DCT on one or more blocks of a component.
174 * The input samples are taken from the sample_data[] array starting at
175 * position start_row/start_col, and moving to the right for any additional
176 * blocks. The quantized coefficients are returned in coef_blocks[].
180 forward_DCT (j_compress_ptr cinfo, jpeg_component_info * compptr,
181 JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
182 JDIMENSION start_row, JDIMENSION start_col,
183 JDIMENSION num_blocks)
184 /* This version is used for integer DCT implementations. */
186 /* This routine is heavily used, so it's worth coding it tightly. */
187 my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
188 forward_DCT_method_ptr do_dct = fdct->do_dct;
189 DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
190 DCTELEM workspace[DCTSIZE2]; /* work area for FDCT subroutine */
193 sample_data += start_row; /* fold in the vertical offset once */
195 for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
196 /* Load data into workspace, applying unsigned->signed conversion */
197 { register DCTELEM *workspaceptr;
198 register JSAMPROW elemptr;
201 workspaceptr = workspace;
202 for (elemr = 0; elemr < DCTSIZE; elemr++) {
203 elemptr = sample_data[elemr] + start_col;
204 #if DCTSIZE == 8 /* unroll the inner loop */
205 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
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;
214 { register int elemc;
215 for (elemc = DCTSIZE; elemc > 0; elemc--) {
216 *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
223 /* Perform the DCT */
224 (*do_dct) (workspace);
226 /* Quantize/descale the coefficients, and store into coef_blocks[] */
227 { register DCTELEM temp, qval;
229 register JCOEFPTR output_ptr = coef_blocks[bi];
231 for (i = 0; i < DCTSIZE2; i++) {
234 /* Divide the coefficient value by qval, ensuring proper rounding.
235 * Since C does not specify the direction of rounding for negative
236 * quotients, we have to force the dividend positive for portability.
238 * In most files, at least half of the output values will be zero
239 * (at default quantization settings, more like three-quarters...)
240 * so we should ensure that this case is fast. On many machines,
241 * a comparison is enough cheaper than a divide to make a special test
242 * a win. Since both inputs will be nonnegative, we need only test
243 * for a < b to discover whether a/b is 0.
244 * If your machine's division is fast enough, define FAST_DIVIDE.
247 #define DIVIDE_BY(a,b) a /= b
249 #define DIVIDE_BY(a,b) if (a >= b) a /= b; else a = 0
253 temp += qval>>1; /* for rounding */
254 DIVIDE_BY(temp, qval);
257 temp += qval>>1; /* for rounding */
258 DIVIDE_BY(temp, qval);
260 output_ptr[i] = (JCOEF) temp;
267 #ifdef DCT_FLOAT_SUPPORTED
270 forward_DCT_float (j_compress_ptr cinfo, jpeg_component_info * compptr,
271 JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
272 JDIMENSION start_row, JDIMENSION start_col,
273 JDIMENSION num_blocks)
274 /* This version is used for floating-point DCT implementations. */
276 /* This routine is heavily used, so it's worth coding it tightly. */
277 my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
278 float_DCT_method_ptr do_dct = fdct->do_float_dct;
279 FAST_FLOAT * divisors = fdct->float_divisors[compptr->quant_tbl_no];
280 FAST_FLOAT workspace[DCTSIZE2]; /* work area for FDCT subroutine */
283 sample_data += start_row; /* fold in the vertical offset once */
285 for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
286 /* Load data into workspace, applying unsigned->signed conversion */
287 { register FAST_FLOAT *workspaceptr;
288 register JSAMPROW elemptr;
291 workspaceptr = workspace;
292 for (elemr = 0; elemr < DCTSIZE; elemr++) {
293 elemptr = sample_data[elemr] + start_col;
294 #if DCTSIZE == 8 /* unroll the inner loop */
295 *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
296 *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
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);
304 { register int elemc;
305 for (elemc = DCTSIZE; elemc > 0; elemc--) {
306 *workspaceptr++ = (FAST_FLOAT)
307 (GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
314 /* Perform the DCT */
315 (*do_dct) (workspace);
317 /* Quantize/descale the coefficients, and store into coef_blocks[] */
318 { register FAST_FLOAT temp;
320 register JCOEFPTR output_ptr = coef_blocks[bi];
322 for (i = 0; i < DCTSIZE2; i++) {
323 /* Apply the quantization and scaling factor */
324 temp = workspace[i] * divisors[i];
325 /* Round to nearest integer.
326 * Since C does not specify the direction of rounding for negative
327 * quotients, we have to force the dividend positive for portability.
328 * The maximum coefficient size is +-16K (for 12-bit data), so this
329 * code should work for either 16-bit or 32-bit ints.
331 output_ptr[i] = (JCOEF) ((int) (temp + (FAST_FLOAT) 16384.5) - 16384);
337 #endif /* DCT_FLOAT_SUPPORTED */
341 * Initialize FDCT manager.
345 jinit_forward_dct (j_compress_ptr cinfo)
351 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
352 SIZEOF(my_fdct_controller));
353 cinfo->fdct = (struct jpeg_forward_dct *) fdct;
354 fdct->pub.start_pass = start_pass_fdctmgr;
356 switch (cinfo->dct_method) {
357 #ifdef DCT_ISLOW_SUPPORTED
359 fdct->pub.forward_DCT = forward_DCT;
360 fdct->do_dct = jpeg_fdct_islow;
363 #ifdef DCT_IFAST_SUPPORTED
365 fdct->pub.forward_DCT = forward_DCT;
366 fdct->do_dct = jpeg_fdct_ifast;
369 #ifdef DCT_FLOAT_SUPPORTED
371 fdct->pub.forward_DCT = forward_DCT_float;
372 fdct->do_float_dct = jpeg_fdct_float;
376 ERREXIT(cinfo, JERR_NOT_COMPILED);
380 /* Mark divisor tables unallocated */
381 for (i = 0; i < NUM_QUANT_TBLS; i++) {
382 fdct->divisors[i] = NULL;
383 #ifdef DCT_FLOAT_SUPPORTED
384 fdct->float_divisors[i] = NULL;