119
118
* Perform dequantization and inverse DCT on one block of coefficients.
123
tinyjpeg_idct_float (struct component *compptr, uint8_t *output_buf, int stride)
121
void tinyjpeg_idct_float(struct component *compptr, uint8_t *output_buf, int stride)
125
FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
126
FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
127
FAST_FLOAT z5, z10, z11, z12, z13;
129
FAST_FLOAT *quantptr;
133
FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
135
/* Pass 1: process columns from input, store into work array. */
137
inptr = compptr->DCT;
138
quantptr = compptr->Q_table;
140
for (ctr = DCTSIZE; ctr > 0; ctr--) {
141
/* Due to quantization, we will usually find that many of the input
142
* coefficients are zero, especially the AC terms. We can exploit this
143
* by short-circuiting the IDCT calculation for any column in which all
144
* the AC terms are zero. In that case each output is equal to the
145
* DC coefficient (with scale factor as needed).
146
* With typical images and quantization tables, half or more of the
147
* column DCT calculations can be simplified this way.
150
if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
151
inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
152
inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
153
inptr[DCTSIZE*7] == 0) {
154
/* AC terms all zero */
155
FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
157
wsptr[DCTSIZE*0] = dcval;
158
wsptr[DCTSIZE*1] = dcval;
159
wsptr[DCTSIZE*2] = dcval;
160
wsptr[DCTSIZE*3] = dcval;
161
wsptr[DCTSIZE*4] = dcval;
162
wsptr[DCTSIZE*5] = dcval;
163
wsptr[DCTSIZE*6] = dcval;
164
wsptr[DCTSIZE*7] = dcval;
166
inptr++; /* advance pointers to next column */
174
tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
175
tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
176
tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
177
tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
179
tmp10 = tmp0 + tmp2; /* phase 3 */
182
tmp13 = tmp1 + tmp3; /* phases 5-3 */
183
tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
185
tmp0 = tmp10 + tmp13; /* phase 2 */
186
tmp3 = tmp10 - tmp13;
187
tmp1 = tmp11 + tmp12;
188
tmp2 = tmp11 - tmp12;
192
tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
193
tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
194
tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
195
tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
197
z13 = tmp6 + tmp5; /* phase 6 */
202
tmp7 = z11 + z13; /* phase 5 */
203
tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
205
z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
206
tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
207
tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
209
tmp6 = tmp12 - tmp7; /* phase 2 */
213
wsptr[DCTSIZE*0] = tmp0 + tmp7;
214
wsptr[DCTSIZE*7] = tmp0 - tmp7;
215
wsptr[DCTSIZE*1] = tmp1 + tmp6;
216
wsptr[DCTSIZE*6] = tmp1 - tmp6;
217
wsptr[DCTSIZE*2] = tmp2 + tmp5;
218
wsptr[DCTSIZE*5] = tmp2 - tmp5;
219
wsptr[DCTSIZE*4] = tmp3 + tmp4;
220
wsptr[DCTSIZE*3] = tmp3 - tmp4;
222
inptr++; /* advance pointers to next column */
227
/* Pass 2: process rows from work array, store into output array. */
228
/* Note that we must descale the results by a factor of 8 == 2**3. */
232
for (ctr = 0; ctr < DCTSIZE; ctr++) {
233
/* Rows of zeroes can be exploited in the same way as we did with columns.
234
* However, the column calculation has created many nonzero AC terms, so
235
* the simplification applies less often (typically 5% to 10% of the time).
236
* And testing floats for zero is relatively expensive, so we don't bother.
241
tmp10 = wsptr[0] + wsptr[4];
242
tmp11 = wsptr[0] - wsptr[4];
244
tmp13 = wsptr[2] + wsptr[6];
245
tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT) 1.414213562) - tmp13;
247
tmp0 = tmp10 + tmp13;
248
tmp3 = tmp10 - tmp13;
249
tmp1 = tmp11 + tmp12;
250
tmp2 = tmp11 - tmp12;
254
z13 = wsptr[5] + wsptr[3];
255
z10 = wsptr[5] - wsptr[3];
256
z11 = wsptr[1] + wsptr[7];
257
z12 = wsptr[1] - wsptr[7];
260
tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562);
262
z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
263
tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
264
tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
270
/* Final output stage: scale down by a factor of 8 and range-limit */
272
outptr[0] = descale_and_clamp((int)(tmp0 + tmp7), 3);
273
outptr[7] = descale_and_clamp((int)(tmp0 - tmp7), 3);
274
outptr[1] = descale_and_clamp((int)(tmp1 + tmp6), 3);
275
outptr[6] = descale_and_clamp((int)(tmp1 - tmp6), 3);
276
outptr[2] = descale_and_clamp((int)(tmp2 + tmp5), 3);
277
outptr[5] = descale_and_clamp((int)(tmp2 - tmp5), 3);
278
outptr[4] = descale_and_clamp((int)(tmp3 + tmp4), 3);
279
outptr[3] = descale_and_clamp((int)(tmp3 - tmp4), 3);
282
wsptr += DCTSIZE; /* advance pointer to next row */
123
FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
124
FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
125
FAST_FLOAT z5, z10, z11, z12, z13;
127
FAST_FLOAT *quantptr;
131
FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
133
/* Pass 1: process columns from input, store into work array. */
135
inptr = compptr->DCT;
136
quantptr = compptr->Q_table;
138
for (ctr = DCTSIZE; ctr > 0; ctr--) {
139
/* Due to quantization, we will usually find that many of the input
140
* coefficients are zero, especially the AC terms. We can exploit this
141
* by short-circuiting the IDCT calculation for any column in which all
142
* the AC terms are zero. In that case each output is equal to the
143
* DC coefficient (with scale factor as needed).
144
* With typical images and quantization tables, half or more of the
145
* column DCT calculations can be simplified this way.
148
if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
149
inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
150
inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
151
inptr[DCTSIZE*7] == 0) {
152
/* AC terms all zero */
153
FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
155
wsptr[DCTSIZE*0] = dcval;
156
wsptr[DCTSIZE*1] = dcval;
157
wsptr[DCTSIZE*2] = dcval;
158
wsptr[DCTSIZE*3] = dcval;
159
wsptr[DCTSIZE*4] = dcval;
160
wsptr[DCTSIZE*5] = dcval;
161
wsptr[DCTSIZE*6] = dcval;
162
wsptr[DCTSIZE*7] = dcval;
164
inptr++; /* advance pointers to next column */
172
tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
173
tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
174
tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
175
tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
177
tmp10 = tmp0 + tmp2; /* phase 3 */
180
tmp13 = tmp1 + tmp3; /* phases 5-3 */
181
tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
183
tmp0 = tmp10 + tmp13; /* phase 2 */
184
tmp3 = tmp10 - tmp13;
185
tmp1 = tmp11 + tmp12;
186
tmp2 = tmp11 - tmp12;
190
tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
191
tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
192
tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
193
tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
195
z13 = tmp6 + tmp5; /* phase 6 */
200
tmp7 = z11 + z13; /* phase 5 */
201
tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
203
z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
204
tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
205
tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
207
tmp6 = tmp12 - tmp7; /* phase 2 */
211
wsptr[DCTSIZE*0] = tmp0 + tmp7;
212
wsptr[DCTSIZE*7] = tmp0 - tmp7;
213
wsptr[DCTSIZE*1] = tmp1 + tmp6;
214
wsptr[DCTSIZE*6] = tmp1 - tmp6;
215
wsptr[DCTSIZE*2] = tmp2 + tmp5;
216
wsptr[DCTSIZE*5] = tmp2 - tmp5;
217
wsptr[DCTSIZE*4] = tmp3 + tmp4;
218
wsptr[DCTSIZE*3] = tmp3 - tmp4;
220
inptr++; /* advance pointers to next column */
225
/* Pass 2: process rows from work array, store into output array. */
226
/* Note that we must descale the results by a factor of 8 == 2**3. */
230
for (ctr = 0; ctr < DCTSIZE; ctr++) {
231
/* Rows of zeroes can be exploited in the same way as we did with columns.
232
* However, the column calculation has created many nonzero AC terms, so
233
* the simplification applies less often (typically 5% to 10% of the time).
234
* And testing floats for zero is relatively expensive, so we don't bother.
239
tmp10 = wsptr[0] + wsptr[4];
240
tmp11 = wsptr[0] - wsptr[4];
242
tmp13 = wsptr[2] + wsptr[6];
243
tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT) 1.414213562) - tmp13;
245
tmp0 = tmp10 + tmp13;
246
tmp3 = tmp10 - tmp13;
247
tmp1 = tmp11 + tmp12;
248
tmp2 = tmp11 - tmp12;
252
z13 = wsptr[5] + wsptr[3];
253
z10 = wsptr[5] - wsptr[3];
254
z11 = wsptr[1] + wsptr[7];
255
z12 = wsptr[1] - wsptr[7];
258
tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562);
260
z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
261
tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
262
tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
268
/* Final output stage: scale down by a factor of 8 and range-limit */
270
outptr[0] = descale_and_clamp((int)(tmp0 + tmp7), 3);
271
outptr[7] = descale_and_clamp((int)(tmp0 - tmp7), 3);
272
outptr[1] = descale_and_clamp((int)(tmp1 + tmp6), 3);
273
outptr[6] = descale_and_clamp((int)(tmp1 - tmp6), 3);
274
outptr[2] = descale_and_clamp((int)(tmp2 + tmp5), 3);
275
outptr[5] = descale_and_clamp((int)(tmp2 - tmp5), 3);
276
outptr[4] = descale_and_clamp((int)(tmp3 + tmp4), 3);
277
outptr[3] = descale_and_clamp((int)(tmp3 - tmp4), 3);
280
wsptr += DCTSIZE; /* advance pointer to next row */