41
41
typedef my_fdct_controller * my_fdct_ptr;
44
/* The current scaled-DCT routines require ISLOW-style divisor tables,
45
* so be sure to compile that code if either ISLOW or SCALING is requested.
47
#ifdef DCT_ISLOW_SUPPORTED
48
#define PROVIDE_ISLOW_TABLES
50
#ifdef DCT_SCALING_SUPPORTED
51
#define PROVIDE_ISLOW_TABLES
57
* Perform forward DCT on one or more blocks of a component.
59
* The input samples are taken from the sample_data[] array starting at
60
* position start_row/start_col, and moving to the right for any additional
61
* blocks. The quantized coefficients are returned in coef_blocks[].
65
forward_DCT (j_compress_ptr cinfo, jpeg_component_info * compptr,
66
JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
67
JDIMENSION start_row, JDIMENSION start_col,
68
JDIMENSION num_blocks)
69
/* This version is used for integer DCT implementations. */
71
/* This routine is heavily used, so it's worth coding it tightly. */
72
my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
73
forward_DCT_method_ptr do_dct = fdct->do_dct[compptr->component_index];
74
DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
75
DCTELEM workspace[DCTSIZE2]; /* work area for FDCT subroutine */
78
sample_data += start_row; /* fold in the vertical offset once */
80
for (bi = 0; bi < num_blocks; bi++, start_col += compptr->DCT_h_scaled_size) {
82
(*do_dct) (workspace, sample_data, start_col);
84
/* Quantize/descale the coefficients, and store into coef_blocks[] */
85
{ register DCTELEM temp, qval;
87
register JCOEFPTR output_ptr = coef_blocks[bi];
89
for (i = 0; i < DCTSIZE2; i++) {
92
/* Divide the coefficient value by qval, ensuring proper rounding.
93
* Since C does not specify the direction of rounding for negative
94
* quotients, we have to force the dividend positive for portability.
96
* In most files, at least half of the output values will be zero
97
* (at default quantization settings, more like three-quarters...)
98
* so we should ensure that this case is fast. On many machines,
99
* a comparison is enough cheaper than a divide to make a special test
100
* a win. Since both inputs will be nonnegative, we need only test
101
* for a < b to discover whether a/b is 0.
102
* If your machine's division is fast enough, define FAST_DIVIDE.
105
#define DIVIDE_BY(a,b) a /= b
107
#define DIVIDE_BY(a,b) if (a >= b) a /= b; else a = 0
111
temp += qval>>1; /* for rounding */
112
DIVIDE_BY(temp, qval);
115
temp += qval>>1; /* for rounding */
116
DIVIDE_BY(temp, qval);
118
output_ptr[i] = (JCOEF) temp;
125
#ifdef DCT_FLOAT_SUPPORTED
128
forward_DCT_float (j_compress_ptr cinfo, jpeg_component_info * compptr,
129
JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
130
JDIMENSION start_row, JDIMENSION start_col,
131
JDIMENSION num_blocks)
132
/* This version is used for floating-point DCT implementations. */
134
/* This routine is heavily used, so it's worth coding it tightly. */
135
my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
136
float_DCT_method_ptr do_dct = fdct->do_float_dct[compptr->component_index];
137
FAST_FLOAT * divisors = fdct->float_divisors[compptr->quant_tbl_no];
138
FAST_FLOAT workspace[DCTSIZE2]; /* work area for FDCT subroutine */
141
sample_data += start_row; /* fold in the vertical offset once */
143
for (bi = 0; bi < num_blocks; bi++, start_col += compptr->DCT_h_scaled_size) {
144
/* Perform the DCT */
145
(*do_dct) (workspace, sample_data, start_col);
147
/* Quantize/descale the coefficients, and store into coef_blocks[] */
148
{ register FAST_FLOAT temp;
150
register JCOEFPTR output_ptr = coef_blocks[bi];
152
for (i = 0; i < DCTSIZE2; i++) {
153
/* Apply the quantization and scaling factor */
154
temp = workspace[i] * divisors[i];
155
/* Round to nearest integer.
156
* Since C does not specify the direction of rounding for negative
157
* quotients, we have to force the dividend positive for portability.
158
* The maximum coefficient size is +-16K (for 12-bit data), so this
159
* code should work for either 16-bit or 32-bit ints.
161
output_ptr[i] = (JCOEF) ((int) (temp + (FAST_FLOAT) 16384.5) - 16384);
167
#endif /* DCT_FLOAT_SUPPORTED */
45
171
* Initialize for a processing pass.
46
172
* Verify that all referenced Q-tables are present, and set up
56
182
my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
57
183
int ci, qtblno, i;
58
184
jpeg_component_info *compptr;
59
186
JQUANT_TBL * qtbl;
62
189
for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
63
190
ci++, compptr++) {
191
/* Select the proper DCT routine for this component's scaling */
192
switch ((compptr->DCT_h_scaled_size << 8) + compptr->DCT_v_scaled_size) {
193
#ifdef DCT_SCALING_SUPPORTED
195
fdct->do_dct[ci] = jpeg_fdct_1x1;
196
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
199
fdct->do_dct[ci] = jpeg_fdct_2x2;
200
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
203
fdct->do_dct[ci] = jpeg_fdct_3x3;
204
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
207
fdct->do_dct[ci] = jpeg_fdct_4x4;
208
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
211
fdct->do_dct[ci] = jpeg_fdct_5x5;
212
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
215
fdct->do_dct[ci] = jpeg_fdct_6x6;
216
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
219
fdct->do_dct[ci] = jpeg_fdct_7x7;
220
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
223
fdct->do_dct[ci] = jpeg_fdct_9x9;
224
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
226
case ((10 << 8) + 10):
227
fdct->do_dct[ci] = jpeg_fdct_10x10;
228
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
230
case ((11 << 8) + 11):
231
fdct->do_dct[ci] = jpeg_fdct_11x11;
232
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
234
case ((12 << 8) + 12):
235
fdct->do_dct[ci] = jpeg_fdct_12x12;
236
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
238
case ((13 << 8) + 13):
239
fdct->do_dct[ci] = jpeg_fdct_13x13;
240
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
242
case ((14 << 8) + 14):
243
fdct->do_dct[ci] = jpeg_fdct_14x14;
244
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
246
case ((15 << 8) + 15):
247
fdct->do_dct[ci] = jpeg_fdct_15x15;
248
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
250
case ((16 << 8) + 16):
251
fdct->do_dct[ci] = jpeg_fdct_16x16;
252
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
254
case ((16 << 8) + 8):
255
fdct->do_dct[ci] = jpeg_fdct_16x8;
256
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
258
case ((14 << 8) + 7):
259
fdct->do_dct[ci] = jpeg_fdct_14x7;
260
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
262
case ((12 << 8) + 6):
263
fdct->do_dct[ci] = jpeg_fdct_12x6;
264
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
266
case ((10 << 8) + 5):
267
fdct->do_dct[ci] = jpeg_fdct_10x5;
268
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
271
fdct->do_dct[ci] = jpeg_fdct_8x4;
272
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
275
fdct->do_dct[ci] = jpeg_fdct_6x3;
276
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
279
fdct->do_dct[ci] = jpeg_fdct_4x2;
280
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
283
fdct->do_dct[ci] = jpeg_fdct_2x1;
284
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
286
case ((8 << 8) + 16):
287
fdct->do_dct[ci] = jpeg_fdct_8x16;
288
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
290
case ((7 << 8) + 14):
291
fdct->do_dct[ci] = jpeg_fdct_7x14;
292
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
294
case ((6 << 8) + 12):
295
fdct->do_dct[ci] = jpeg_fdct_6x12;
296
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
298
case ((5 << 8) + 10):
299
fdct->do_dct[ci] = jpeg_fdct_5x10;
300
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
303
fdct->do_dct[ci] = jpeg_fdct_4x8;
304
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
307
fdct->do_dct[ci] = jpeg_fdct_3x6;
308
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
311
fdct->do_dct[ci] = jpeg_fdct_2x4;
312
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
315
fdct->do_dct[ci] = jpeg_fdct_1x2;
316
method = JDCT_ISLOW; /* jfdctint uses islow-style table */
319
case ((DCTSIZE << 8) + DCTSIZE):
320
switch (cinfo->dct_method) {
321
#ifdef DCT_ISLOW_SUPPORTED
323
fdct->do_dct[ci] = jpeg_fdct_islow;
327
#ifdef DCT_IFAST_SUPPORTED
329
fdct->do_dct[ci] = jpeg_fdct_ifast;
333
#ifdef DCT_FLOAT_SUPPORTED
335
fdct->do_float_dct[ci] = jpeg_fdct_float;
340
ERREXIT(cinfo, JERR_NOT_COMPILED);
345
ERREXIT2(cinfo, JERR_BAD_DCTSIZE,
346
compptr->DCT_h_scaled_size, compptr->DCT_v_scaled_size);
64
349
qtblno = compptr->quant_tbl_no;
65
350
/* Make sure specified quantization table is present */
66
351
if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
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
460
* Initialize FDCT manager.