1 /*
2 * idct_mmx.c
3 * Copyright (C) 2000-2003 Michel Lespinasse <walken@zoy.org>
4 * Copyright (C) 1999-2000 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
5 *
6 * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
7 * See http://libmpeg2.sourceforge.net/ for updates.
8 *
9 * mpeg2dec is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License as published by
11 * the Free Software Foundation; either version 2 of the License, or
12 * (at your option) any later version.
13 *
14 * mpeg2dec is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
18 *
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 */
23
24 #include "config.h"
25
26 #if defined(ARCH_X86) || defined(ARCH_X86_64)
27
28 #include <inttypes.h>
29
30 #include "mpeg2.h"
31 #include "attributes.h"
32 #include "mpeg2_internal.h"
33 #include "mmx.h"
34
35 #define ROW_SHIFT 15
36 #define COL_SHIFT 6
37
38 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
39 #define rounder(bias) {round (bias), round (bias)}
40 #define rounder_sse2(bias) {round (bias), round (bias), round (bias), round (bias)}
41
42
43 #if 0
44 /* C row IDCT - it is just here to document the MMXEXT and MMX versions */
45 static inline void idct_row (int16_t * row, int offset,
46 int16_t * table, int32_t * rounder)
47 {
48 int C1, C2, C3, C4, C5, C6, C7;
49 int a0, a1, a2, a3, b0, b1, b2, b3;
50
51 row += offset;
52
53 C1 = table[1];
54 C2 = table[2];
55 C3 = table[3];
56 C4 = table[4];
57 C5 = table[5];
58 C6 = table[6];
59 C7 = table[7];
60
61 a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
62 a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
63 a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
64 a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
65
66 b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
67 b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
68 b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
69 b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
70
71 row[0] = (a0 + b0) >> ROW_SHIFT;
72 row[1] = (a1 + b1) >> ROW_SHIFT;
73 row[2] = (a2 + b2) >> ROW_SHIFT;
74 row[3] = (a3 + b3) >> ROW_SHIFT;
75 row[4] = (a3 - b3) >> ROW_SHIFT;
76 row[5] = (a2 - b2) >> ROW_SHIFT;
77 row[6] = (a1 - b1) >> ROW_SHIFT;
78 row[7] = (a0 - b0) >> ROW_SHIFT;
79 }
80 #endif
81
82
83 /* SSE2 row IDCT */
84 #define sse2_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, c4, c6, \
85 c4, -c6, c4, -c2, \
86 c4, c6, -c4, -c2, \
87 -c4, c2, c4, -c6, \
88 c1, c3, c3, -c7, \
89 c5, -c1, c7, -c5, \
90 c5, c7, -c1, -c5, \
91 c7, c3, c3, -c1 }
92
93 #define SSE2_IDCT_2ROW(table, row1, row2, round1, round2) do { \
94 /* no scheduling: trust in out of order execution */ \
95 /* based on Intel AP-945 */ \
96 /* (http://cache-www.intel.com/cd/00/00/01/76/17680_w_idct.pdf) */ \
97 \
98 /* input */ /* 1: row1= x7 x5 x3 x1 x6 x4 x2 x0 */ \
99 pshufd_r2r (row1, xmm1, 0); /* 1: xmm1= x2 x0 x2 x0 x2 x0 x2 x0 */ \
100 pmaddwd_m2r (table[0], xmm1); /* 1: xmm1= x2*C + x0*C ... */ \
101 pshufd_r2r (row1, xmm3, 0xaa); /* 1: xmm3= x3 x1 x3 x1 x3 x1 x3 x1 */ \
102 pmaddwd_m2r (table[2*8], xmm3); /* 1: xmm3= x3*C + x1*C ... */ \
103 pshufd_r2r (row1, xmm2, 0x55); /* 1: xmm2= x6 x4 x6 x4 x6 x4 x6 x4 */ \
104 pshufd_r2r (row1, row1, 0xff); /* 1: row1= x7 x5 x7 x5 x7 x5 x7 x5 */ \
105 pmaddwd_m2r (table[1*8], xmm2); /* 1: xmm2= x6*C + x4*C ... */ \
106 paddd_m2r (round1, xmm1); /* 1: xmm1= x2*C + x0*C + round ... */ \
107 pmaddwd_m2r (table[3*8], row1); /* 1: row1= x7*C + x5*C ... */ \
108 pshufd_r2r (row2, xmm5, 0); /* 2: */ \
109 pshufd_r2r (row2, xmm6, 0x55); /* 2: */ \
110 pmaddwd_m2r (table[0], xmm5); /* 2: */ \
111 paddd_r2r (xmm2, xmm1); /* 1: xmm1= a[] */ \
112 movdqa_r2r (xmm1, xmm2); /* 1: xmm2= a[] */ \
113 pshufd_r2r (row2, xmm7, 0xaa); /* 2: */ \
114 pmaddwd_m2r (table[1*8], xmm6); /* 2: */ \
115 paddd_r2r (xmm3, row1); /* 1: row1= b[]= 7*C+5*C+3*C+1*C ... */ \
116 pshufd_r2r (row2, row2, 0xff); /* 2: */ \
117 psubd_r2r (row1, xmm2); /* 1: xmm2= a[] - b[] */ \
118 pmaddwd_m2r (table[2*8], xmm7); /* 2: */ \
119 paddd_r2r (xmm1, row1); /* 1: row1= a[] + b[] */ \
120 psrad_i2r (ROW_SHIFT, xmm2); /* 1: xmm2= result 4...7 */ \
121 paddd_m2r (round2, xmm5); /* 2: */ \
122 pmaddwd_m2r (table[3*8], row2); /* 2: */ \
123 paddd_r2r (xmm6, xmm5); /* 2: */ \
124 movdqa_r2r (xmm5, xmm6); /* 2: */ \
125 psrad_i2r (ROW_SHIFT, row1); /* 1: row1= result 0...4 */ \
126 pshufd_r2r (xmm2, xmm2, 0x1b); /* 1: [0 1 2 3] -> [3 2 1 0] */ \
127 packssdw_r2r (xmm2, row1); /* 1: row1= result[] */ \
128 paddd_r2r (xmm7, row2); /* 2: */ \
129 psubd_r2r (row2, xmm6); /* 2: */ \
130 paddd_r2r (xmm5, row2); /* 2: */ \
131 psrad_i2r (ROW_SHIFT, xmm6); /* 2: */ \
132 psrad_i2r (ROW_SHIFT, row2); /* 2: */ \
133 pshufd_r2r (xmm6, xmm6, 0x1b); /* 2: */ \
134 packssdw_r2r (xmm6, row2); /* 2: */ \
135 } while (0)
136
137
138 /* MMXEXT row IDCT */
139
140 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, -c4, -c2, \
141 c4, c6, c4, c6, \
142 c1, c3, -c1, -c5, \
143 c5, c7, c3, -c7, \
144 c4, -c6, c4, -c6, \
145 -c4, c2, c4, -c2, \
146 c5, -c1, c3, -c1, \
147 c7, c3, c7, -c5 }
148
mmxext_row_head(int16_t * const row,const int offset,const int16_t * const table)149 static inline void mmxext_row_head (int16_t * const row, const int offset,
150 const int16_t * const table)
151 {
152 movq_m2r (*(row+offset), mm2); /* mm2 = x6 x4 x2 x0 */
153
154 movq_m2r (*(row+offset+4), mm5); /* mm5 = x7 x5 x3 x1 */
155 movq_r2r (mm2, mm0); /* mm0 = x6 x4 x2 x0 */
156
157 movq_m2r (*table, mm3); /* mm3 = -C2 -C4 C2 C4 */
158 movq_r2r (mm5, mm6); /* mm6 = x7 x5 x3 x1 */
159
160 movq_m2r (*(table+4), mm4); /* mm4 = C6 C4 C6 C4 */
161 pmaddwd_r2r (mm0, mm3); /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
162
163 pshufw_r2r (mm2, mm2, 0x4e); /* mm2 = x2 x0 x6 x4 */
164 }
165
mmxext_row(const int16_t * const table,const int32_t * const rounder)166 static inline void mmxext_row (const int16_t * const table,
167 const int32_t * const rounder)
168 {
169 movq_m2r (*(table+8), mm1); /* mm1 = -C5 -C1 C3 C1 */
170 pmaddwd_r2r (mm2, mm4); /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
171
172 pmaddwd_m2r (*(table+16), mm0); /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
173 pshufw_r2r (mm6, mm6, 0x4e); /* mm6 = x3 x1 x7 x5 */
174
175 movq_m2r (*(table+12), mm7); /* mm7 = -C7 C3 C7 C5 */
176 pmaddwd_r2r (mm5, mm1); /* mm1 = -C1*x5-C5*x7 C1*x1+C3*x3 */
177
178 paddd_m2r (*rounder, mm3); /* mm3 += rounder */
179 pmaddwd_r2r (mm6, mm7); /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
180
181 pmaddwd_m2r (*(table+20), mm2); /* mm2 = C4*x0-C2*x2 -C4*x4+C2*x6 */
182 paddd_r2r (mm4, mm3); /* mm3 = a1 a0 + rounder */
183
184 pmaddwd_m2r (*(table+24), mm5); /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
185 movq_r2r (mm3, mm4); /* mm4 = a1 a0 + rounder */
186
187 pmaddwd_m2r (*(table+28), mm6); /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
188 paddd_r2r (mm7, mm1); /* mm1 = b1 b0 */
189
190 paddd_m2r (*rounder, mm0); /* mm0 += rounder */
191 psubd_r2r (mm1, mm3); /* mm3 = a1-b1 a0-b0 + rounder */
192
193 psrad_i2r (ROW_SHIFT, mm3); /* mm3 = y6 y7 */
194 paddd_r2r (mm4, mm1); /* mm1 = a1+b1 a0+b0 + rounder */
195
196 paddd_r2r (mm2, mm0); /* mm0 = a3 a2 + rounder */
197 psrad_i2r (ROW_SHIFT, mm1); /* mm1 = y1 y0 */
198
199 paddd_r2r (mm6, mm5); /* mm5 = b3 b2 */
200 movq_r2r (mm0, mm4); /* mm4 = a3 a2 + rounder */
201
202 paddd_r2r (mm5, mm0); /* mm0 = a3+b3 a2+b2 + rounder */
203 psubd_r2r (mm5, mm4); /* mm4 = a3-b3 a2-b2 + rounder */
204 }
205
mmxext_row_tail(int16_t * const row,const int store)206 static inline void mmxext_row_tail (int16_t * const row, const int store)
207 {
208 psrad_i2r (ROW_SHIFT, mm0); /* mm0 = y3 y2 */
209
210 psrad_i2r (ROW_SHIFT, mm4); /* mm4 = y4 y5 */
211
212 packssdw_r2r (mm0, mm1); /* mm1 = y3 y2 y1 y0 */
213
214 packssdw_r2r (mm3, mm4); /* mm4 = y6 y7 y4 y5 */
215
216 movq_r2m (mm1, *(row+store)); /* save y3 y2 y1 y0 */
217 pshufw_r2r (mm4, mm4, 0xb1); /* mm4 = y7 y6 y5 y4 */
218
219 /* slot */
220
221 movq_r2m (mm4, *(row+store+4)); /* save y7 y6 y5 y4 */
222 }
223
mmxext_row_mid(int16_t * const row,const int store,const int offset,const int16_t * const table)224 static inline void mmxext_row_mid (int16_t * const row, const int store,
225 const int offset,
226 const int16_t * const table)
227 {
228 movq_m2r (*(row+offset), mm2); /* mm2 = x6 x4 x2 x0 */
229 psrad_i2r (ROW_SHIFT, mm0); /* mm0 = y3 y2 */
230
231 movq_m2r (*(row+offset+4), mm5); /* mm5 = x7 x5 x3 x1 */
232 psrad_i2r (ROW_SHIFT, mm4); /* mm4 = y4 y5 */
233
234 packssdw_r2r (mm0, mm1); /* mm1 = y3 y2 y1 y0 */
235 movq_r2r (mm5, mm6); /* mm6 = x7 x5 x3 x1 */
236
237 packssdw_r2r (mm3, mm4); /* mm4 = y6 y7 y4 y5 */
238 movq_r2r (mm2, mm0); /* mm0 = x6 x4 x2 x0 */
239
240 movq_r2m (mm1, *(row+store)); /* save y3 y2 y1 y0 */
241 pshufw_r2r (mm4, mm4, 0xb1); /* mm4 = y7 y6 y5 y4 */
242
243 movq_m2r (*table, mm3); /* mm3 = -C2 -C4 C2 C4 */
244 movq_r2m (mm4, *(row+store+4)); /* save y7 y6 y5 y4 */
245
246 pmaddwd_r2r (mm0, mm3); /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
247
248 movq_m2r (*(table+4), mm4); /* mm4 = C6 C4 C6 C4 */
249 pshufw_r2r (mm2, mm2, 0x4e); /* mm2 = x2 x0 x6 x4 */
250 }
251
252
253 /* MMX row IDCT */
254
255 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, c4, c6, \
256 c4, c6, -c4, -c2, \
257 c1, c3, c3, -c7, \
258 c5, c7, -c1, -c5, \
259 c4, -c6, c4, -c2, \
260 -c4, c2, c4, -c6, \
261 c5, -c1, c7, -c5, \
262 c7, c3, c3, -c1 }
263
mmx_row_head(int16_t * const row,const int offset,const int16_t * const table)264 static inline void mmx_row_head (int16_t * const row, const int offset,
265 const int16_t * const table)
266 {
267 movq_m2r (*(row+offset), mm2); /* mm2 = x6 x4 x2 x0 */
268
269 movq_m2r (*(row+offset+4), mm5); /* mm5 = x7 x5 x3 x1 */
270 movq_r2r (mm2, mm0); /* mm0 = x6 x4 x2 x0 */
271
272 movq_m2r (*table, mm3); /* mm3 = C6 C4 C2 C4 */
273 movq_r2r (mm5, mm6); /* mm6 = x7 x5 x3 x1 */
274
275 punpckldq_r2r (mm0, mm0); /* mm0 = x2 x0 x2 x0 */
276
277 movq_m2r (*(table+4), mm4); /* mm4 = -C2 -C4 C6 C4 */
278 pmaddwd_r2r (mm0, mm3); /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
279
280 movq_m2r (*(table+8), mm1); /* mm1 = -C7 C3 C3 C1 */
281 punpckhdq_r2r (mm2, mm2); /* mm2 = x6 x4 x6 x4 */
282 }
283
mmx_row(const int16_t * const table,const int32_t * const rounder)284 static inline void mmx_row (const int16_t * const table,
285 const int32_t * const rounder)
286 {
287 pmaddwd_r2r (mm2, mm4); /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
288 punpckldq_r2r (mm5, mm5); /* mm5 = x3 x1 x3 x1 */
289
290 pmaddwd_m2r (*(table+16), mm0); /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
291 punpckhdq_r2r (mm6, mm6); /* mm6 = x7 x5 x7 x5 */
292
293 movq_m2r (*(table+12), mm7); /* mm7 = -C5 -C1 C7 C5 */
294 pmaddwd_r2r (mm5, mm1); /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
295
296 paddd_m2r (*rounder, mm3); /* mm3 += rounder */
297 pmaddwd_r2r (mm6, mm7); /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
298
299 pmaddwd_m2r (*(table+20), mm2); /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
300 paddd_r2r (mm4, mm3); /* mm3 = a1 a0 + rounder */
301
302 pmaddwd_m2r (*(table+24), mm5); /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
303 movq_r2r (mm3, mm4); /* mm4 = a1 a0 + rounder */
304
305 pmaddwd_m2r (*(table+28), mm6); /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
306 paddd_r2r (mm7, mm1); /* mm1 = b1 b0 */
307
308 paddd_m2r (*rounder, mm0); /* mm0 += rounder */
309 psubd_r2r (mm1, mm3); /* mm3 = a1-b1 a0-b0 + rounder */
310
311 psrad_i2r (ROW_SHIFT, mm3); /* mm3 = y6 y7 */
312 paddd_r2r (mm4, mm1); /* mm1 = a1+b1 a0+b0 + rounder */
313
314 paddd_r2r (mm2, mm0); /* mm0 = a3 a2 + rounder */
315 psrad_i2r (ROW_SHIFT, mm1); /* mm1 = y1 y0 */
316
317 paddd_r2r (mm6, mm5); /* mm5 = b3 b2 */
318 movq_r2r (mm0, mm7); /* mm7 = a3 a2 + rounder */
319
320 paddd_r2r (mm5, mm0); /* mm0 = a3+b3 a2+b2 + rounder */
321 psubd_r2r (mm5, mm7); /* mm7 = a3-b3 a2-b2 + rounder */
322 }
323
mmx_row_tail(int16_t * const row,const int store)324 static inline void mmx_row_tail (int16_t * const row, const int store)
325 {
326 psrad_i2r (ROW_SHIFT, mm0); /* mm0 = y3 y2 */
327
328 psrad_i2r (ROW_SHIFT, mm7); /* mm7 = y4 y5 */
329
330 packssdw_r2r (mm0, mm1); /* mm1 = y3 y2 y1 y0 */
331
332 packssdw_r2r (mm3, mm7); /* mm7 = y6 y7 y4 y5 */
333
334 movq_r2m (mm1, *(row+store)); /* save y3 y2 y1 y0 */
335 movq_r2r (mm7, mm4); /* mm4 = y6 y7 y4 y5 */
336
337 pslld_i2r (16, mm7); /* mm7 = y7 0 y5 0 */
338
339 psrld_i2r (16, mm4); /* mm4 = 0 y6 0 y4 */
340
341 por_r2r (mm4, mm7); /* mm7 = y7 y6 y5 y4 */
342
343 /* slot */
344
345 movq_r2m (mm7, *(row+store+4)); /* save y7 y6 y5 y4 */
346 }
347
mmx_row_mid(int16_t * const row,const int store,const int offset,const int16_t * const table)348 static inline void mmx_row_mid (int16_t * const row, const int store,
349 const int offset, const int16_t * const table)
350 {
351 movq_m2r (*(row+offset), mm2); /* mm2 = x6 x4 x2 x0 */
352 psrad_i2r (ROW_SHIFT, mm0); /* mm0 = y3 y2 */
353
354 movq_m2r (*(row+offset+4), mm5); /* mm5 = x7 x5 x3 x1 */
355 psrad_i2r (ROW_SHIFT, mm7); /* mm7 = y4 y5 */
356
357 packssdw_r2r (mm0, mm1); /* mm1 = y3 y2 y1 y0 */
358 movq_r2r (mm5, mm6); /* mm6 = x7 x5 x3 x1 */
359
360 packssdw_r2r (mm3, mm7); /* mm7 = y6 y7 y4 y5 */
361 movq_r2r (mm2, mm0); /* mm0 = x6 x4 x2 x0 */
362
363 movq_r2m (mm1, *(row+store)); /* save y3 y2 y1 y0 */
364 movq_r2r (mm7, mm1); /* mm1 = y6 y7 y4 y5 */
365
366 punpckldq_r2r (mm0, mm0); /* mm0 = x2 x0 x2 x0 */
367 psrld_i2r (16, mm7); /* mm7 = 0 y6 0 y4 */
368
369 movq_m2r (*table, mm3); /* mm3 = C6 C4 C2 C4 */
370 pslld_i2r (16, mm1); /* mm1 = y7 0 y5 0 */
371
372 movq_m2r (*(table+4), mm4); /* mm4 = -C2 -C4 C6 C4 */
373 por_r2r (mm1, mm7); /* mm7 = y7 y6 y5 y4 */
374
375 movq_m2r (*(table+8), mm1); /* mm1 = -C7 C3 C3 C1 */
376 punpckhdq_r2r (mm2, mm2); /* mm2 = x6 x4 x6 x4 */
377
378 movq_r2m (mm7, *(row+store+4)); /* save y7 y6 y5 y4 */
379 pmaddwd_r2r (mm0, mm3); /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
380 }
381
382
383 #if 0
384 /* C column IDCT - it is just here to document the MMXEXT and MMX versions */
385 static inline void idct_col (int16_t * col, int offset)
386 {
387 /* multiplication - as implemented on mmx */
388 #define F(c,x) (((c) * (x)) >> 16)
389
390 /* saturation - it helps us handle torture test cases */
391 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
392
393 int16_t x0, x1, x2, x3, x4, x5, x6, x7;
394 int16_t y0, y1, y2, y3, y4, y5, y6, y7;
395 int16_t a0, a1, a2, a3, b0, b1, b2, b3;
396 int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
397
398 col += offset;
399
400 x0 = col[0*8];
401 x1 = col[1*8];
402 x2 = col[2*8];
403 x3 = col[3*8];
404 x4 = col[4*8];
405 x5 = col[5*8];
406 x6 = col[6*8];
407 x7 = col[7*8];
408
409 u04 = S (x0 + x4);
410 v04 = S (x0 - x4);
411 u26 = S (F (T2, x6) + x2);
412 v26 = S (F (T2, x2) - x6);
413
414 a0 = S (u04 + u26);
415 a1 = S (v04 + v26);
416 a2 = S (v04 - v26);
417 a3 = S (u04 - u26);
418
419 u17 = S (F (T1, x7) + x1);
420 v17 = S (F (T1, x1) - x7);
421 u35 = S (F (T3, x5) + x3);
422 v35 = S (F (T3, x3) - x5);
423
424 b0 = S (u17 + u35);
425 b3 = S (v17 - v35);
426 u12 = S (u17 - u35);
427 v12 = S (v17 + v35);
428 u12 = S (2 * F (C4, u12));
429 v12 = S (2 * F (C4, v12));
430 b1 = S (u12 + v12);
431 b2 = S (u12 - v12);
432
433 y0 = S (a0 + b0) >> COL_SHIFT;
434 y1 = S (a1 + b1) >> COL_SHIFT;
435 y2 = S (a2 + b2) >> COL_SHIFT;
436 y3 = S (a3 + b3) >> COL_SHIFT;
437
438 y4 = S (a3 - b3) >> COL_SHIFT;
439 y5 = S (a2 - b2) >> COL_SHIFT;
440 y6 = S (a1 - b1) >> COL_SHIFT;
441 y7 = S (a0 - b0) >> COL_SHIFT;
442
443 col[0*8] = y0;
444 col[1*8] = y1;
445 col[2*8] = y2;
446 col[3*8] = y3;
447 col[4*8] = y4;
448 col[5*8] = y5;
449 col[6*8] = y6;
450 col[7*8] = y7;
451 }
452 #endif
453
454
455 #define T1 13036
456 #define T2 27146
457 #define T3 43790
458 #define C4 23170
459
460
461 /* SSE2 column IDCT */
sse2_idct_col(int16_t * const col)462 static inline void sse2_idct_col (int16_t * const col)
463 {
464 /* Almost identical to mmxext version: */
465 /* just do both 4x8 columns in paraller */
466
467 static const short t1_vector[] ATTR_ALIGN(16) = {T1,T1,T1,T1,T1,T1,T1,T1};
468 static const short t2_vector[] ATTR_ALIGN(16) = {T2,T2,T2,T2,T2,T2,T2,T2};
469 static const short t3_vector[] ATTR_ALIGN(16) = {T3,T3,T3,T3,T3,T3,T3,T3};
470 static const short c4_vector[] ATTR_ALIGN(16) = {C4,C4,C4,C4,C4,C4,C4,C4};
471
472 #if defined(__x86_64__)
473
474 /* INPUT: block in xmm8 ... xmm15 */
475
476 movdqa_m2r (*t1_vector, xmm0); /* xmm0 = T1 */
477 movdqa_r2r (xmm9, xmm1); /* xmm1 = x1 */
478
479 movdqa_r2r (xmm0, xmm2); /* xmm2 = T1 */
480 pmulhw_r2r (xmm1, xmm0); /* xmm0 = T1*x1 */
481
482 movdqa_m2r (*t3_vector, xmm5); /* xmm5 = T3 */
483 pmulhw_r2r (xmm15, xmm2); /* xmm2 = T1*x7 */
484
485 movdqa_r2r (xmm5, xmm7); /* xmm7 = T3-1 */
486 psubsw_r2r (xmm15, xmm0); /* xmm0 = v17 */
487
488 movdqa_m2r (*t2_vector, xmm9); /* xmm9 = T2 */
489 pmulhw_r2r (xmm11, xmm5); /* xmm5 = (T3-1)*x3 */
490
491 paddsw_r2r (xmm2, xmm1); /* xmm1 = u17 */
492 pmulhw_r2r (xmm13, xmm7); /* xmm7 = (T3-1)*x5 */
493
494 movdqa_r2r (xmm9, xmm2); /* xmm2 = T2 */
495 paddsw_r2r (xmm11, xmm5); /* xmm5 = T3*x3 */
496
497 pmulhw_r2r (xmm10, xmm9); /* xmm9 = T2*x2 */
498 paddsw_r2r (xmm13, xmm7); /* xmm7 = T3*x5 */
499
500 psubsw_r2r (xmm13, xmm5); /* xmm5 = v35 */
501 paddsw_r2r (xmm11, xmm7); /* xmm7 = u35 */
502
503 movdqa_r2r (xmm0, xmm6); /* xmm6 = v17 */
504 pmulhw_r2r (xmm14, xmm2); /* xmm2 = T2*x6 */
505
506 psubsw_r2r (xmm5, xmm0); /* xmm0 = b3 */
507 psubsw_r2r (xmm14, xmm9); /* xmm9 = v26 */
508
509 paddsw_r2r (xmm6, xmm5); /* xmm5 = v12 */
510 movdqa_r2r (xmm0, xmm11); /* xmm11 = b3 */
511
512 movdqa_r2r (xmm1, xmm6); /* xmm6 = u17 */
513 paddsw_r2r (xmm10, xmm2); /* xmm2 = u26 */
514
515 paddsw_r2r (xmm7, xmm6); /* xmm6 = b0 */
516 psubsw_r2r (xmm7, xmm1); /* xmm1 = u12 */
517
518 movdqa_r2r (xmm1, xmm7); /* xmm7 = u12 */
519 paddsw_r2r (xmm5, xmm1); /* xmm1 = u12+v12 */
520
521 movdqa_m2r (*c4_vector, xmm0); /* xmm0 = C4/2 */
522 psubsw_r2r (xmm5, xmm7); /* xmm7 = u12-v12 */
523
524 movdqa_r2r (xmm6, xmm4); /* xmm4 = b0 */
525 pmulhw_r2r (xmm0, xmm1); /* xmm1 = b1/2 */
526
527 movdqa_r2r (xmm9, xmm6); /* xmm6 = v26 */
528 pmulhw_r2r (xmm0, xmm7); /* xmm7 = b2/2 */
529
530 movdqa_r2r (xmm8, xmm10); /* xmm10 = x0 */
531 movdqa_r2r (xmm8, xmm0); /* xmm0 = x0 */
532
533 psubsw_r2r (xmm12, xmm10); /* xmm10 = v04 */
534 paddsw_r2r (xmm12, xmm0); /* xmm0 = u04 */
535
536 paddsw_r2r (xmm10, xmm9); /* xmm9 = a1 */
537 movdqa_r2r (xmm0, xmm8); /* xmm8 = u04 */
538
539 psubsw_r2r (xmm6, xmm10); /* xmm10 = a2 */
540 paddsw_r2r (xmm2, xmm8); /* xmm5 = a0 */
541
542 paddsw_r2r (xmm1, xmm1); /* xmm1 = b1 */
543 psubsw_r2r (xmm2, xmm0); /* xmm0 = a3 */
544
545 paddsw_r2r (xmm7, xmm7); /* xmm7 = b2 */
546 movdqa_r2r (xmm10, xmm13); /* xmm13 = a2 */
547
548 movdqa_r2r (xmm9, xmm14); /* xmm14 = a1 */
549 paddsw_r2r (xmm7, xmm10); /* xmm10 = a2+b2 */
550
551 psraw_i2r (COL_SHIFT,xmm10); /* xmm10 = y2 */
552 paddsw_r2r (xmm1, xmm9); /* xmm9 = a1+b1 */
553
554 psraw_i2r (COL_SHIFT, xmm9); /* xmm9 = y1 */
555 psubsw_r2r (xmm1, xmm14); /* xmm14 = a1-b1 */
556
557 psubsw_r2r (xmm7, xmm13); /* xmm13 = a2-b2 */
558 psraw_i2r (COL_SHIFT,xmm14); /* xmm14 = y6 */
559
560 movdqa_r2r (xmm8, xmm15); /* xmm15 = a0 */
561 psraw_i2r (COL_SHIFT,xmm13); /* xmm13 = y5 */
562
563 paddsw_r2r (xmm4, xmm8); /* xmm8 = a0+b0 */
564 psubsw_r2r (xmm4, xmm15); /* xmm15 = a0-b0 */
565
566 psraw_i2r (COL_SHIFT, xmm8); /* xmm8 = y0 */
567 movdqa_r2r (xmm0, xmm12); /* xmm12 = a3 */
568
569 psubsw_r2r (xmm11, xmm12); /* xmm12 = a3-b3 */
570 psraw_i2r (COL_SHIFT,xmm15); /* xmm15 = y7 */
571
572 paddsw_r2r (xmm0, xmm11); /* xmm11 = a3+b3 */
573 psraw_i2r (COL_SHIFT,xmm12); /* xmm12 = y4 */
574
575 psraw_i2r (COL_SHIFT,xmm11); /* xmm11 = y3 */
576
577 /* OUTPUT: block in xmm8 ... xmm15 */
578
579 #else
580 movdqa_m2r (*t1_vector, xmm0); /* xmm0 = T1 */
581
582 movdqa_m2r (*(col+1*8), xmm1); /* xmm1 = x1 */
583 movdqa_r2r (xmm0, xmm2); /* xmm2 = T1 */
584
585 movdqa_m2r (*(col+7*8), xmm4); /* xmm4 = x7 */
586 pmulhw_r2r (xmm1, xmm0); /* xmm0 = T1*x1 */
587
588 movdqa_m2r (*t3_vector, xmm5); /* xmm5 = T3 */
589 pmulhw_r2r (xmm4, xmm2); /* xmm2 = T1*x7 */
590
591 movdqa_m2r (*(col+5*8), xmm6); /* xmm6 = x5 */
592 movdqa_r2r (xmm5, xmm7); /* xmm7 = T3-1 */
593
594 movdqa_m2r (*(col+3*8), xmm3); /* xmm3 = x3 */
595 psubsw_r2r (xmm4, xmm0); /* xmm0 = v17 */
596
597 movdqa_m2r (*t2_vector, xmm4); /* xmm4 = T2 */
598 pmulhw_r2r (xmm3, xmm5); /* xmm5 = (T3-1)*x3 */
599
600 paddsw_r2r (xmm2, xmm1); /* xmm1 = u17 */
601 pmulhw_r2r (xmm6, xmm7); /* xmm7 = (T3-1)*x5 */
602
603 /* slot */
604
605 movdqa_r2r (xmm4, xmm2); /* xmm2 = T2 */
606 paddsw_r2r (xmm3, xmm5); /* xmm5 = T3*x3 */
607
608 pmulhw_m2r (*(col+2*8), xmm4); /* xmm4 = T2*x2 */
609 paddsw_r2r (xmm6, xmm7); /* xmm7 = T3*x5 */
610
611 psubsw_r2r (xmm6, xmm5); /* xmm5 = v35 */
612 paddsw_r2r (xmm3, xmm7); /* xmm7 = u35 */
613
614 movdqa_m2r (*(col+6*8), xmm3); /* xmm3 = x6 */
615 movdqa_r2r (xmm0, xmm6); /* xmm6 = v17 */
616
617 pmulhw_r2r (xmm3, xmm2); /* xmm2 = T2*x6 */
618 psubsw_r2r (xmm5, xmm0); /* xmm0 = b3 */
619
620 psubsw_r2r (xmm3, xmm4); /* xmm4 = v26 */
621 paddsw_r2r (xmm6, xmm5); /* xmm5 = v12 */
622
623 movdqa_r2m (xmm0, *(col+3*8)); /* save b3 in scratch0 */
624 movdqa_r2r (xmm1, xmm6); /* xmm6 = u17 */
625
626 paddsw_m2r (*(col+2*8), xmm2); /* xmm2 = u26 */
627 paddsw_r2r (xmm7, xmm6); /* xmm6 = b0 */
628
629 psubsw_r2r (xmm7, xmm1); /* xmm1 = u12 */
630 movdqa_r2r (xmm1, xmm7); /* xmm7 = u12 */
631
632 movdqa_m2r (*(col+0*8), xmm3); /* xmm3 = x0 */
633 paddsw_r2r (xmm5, xmm1); /* xmm1 = u12+v12 */
634
635 movdqa_m2r (*c4_vector, xmm0); /* xmm0 = C4/2 */
636 psubsw_r2r (xmm5, xmm7); /* xmm7 = u12-v12 */
637
638 movdqa_r2m (xmm6, *(col+5*8)); /* save b0 in scratch1 */
639 pmulhw_r2r (xmm0, xmm1); /* xmm1 = b1/2 */
640
641 movdqa_r2r (xmm4, xmm6); /* xmm6 = v26 */
642 pmulhw_r2r (xmm0, xmm7); /* xmm7 = b2/2 */
643
644 movdqa_m2r (*(col+4*8), xmm5); /* xmm5 = x4 */
645 movdqa_r2r (xmm3, xmm0); /* xmm0 = x0 */
646
647 psubsw_r2r (xmm5, xmm3); /* xmm3 = v04 */
648 paddsw_r2r (xmm5, xmm0); /* xmm0 = u04 */
649
650 paddsw_r2r (xmm3, xmm4); /* xmm4 = a1 */
651 movdqa_r2r (xmm0, xmm5); /* xmm5 = u04 */
652
653 psubsw_r2r (xmm6, xmm3); /* xmm3 = a2 */
654 paddsw_r2r (xmm2, xmm5); /* xmm5 = a0 */
655
656 paddsw_r2r (xmm1, xmm1); /* xmm1 = b1 */
657 psubsw_r2r (xmm2, xmm0); /* xmm0 = a3 */
658
659 paddsw_r2r (xmm7, xmm7); /* xmm7 = b2 */
660 movdqa_r2r (xmm3, xmm2); /* xmm2 = a2 */
661
662 movdqa_r2r (xmm4, xmm6); /* xmm6 = a1 */
663 paddsw_r2r (xmm7, xmm3); /* xmm3 = a2+b2 */
664
665 psraw_i2r (COL_SHIFT, xmm3); /* xmm3 = y2 */
666 paddsw_r2r (xmm1, xmm4); /* xmm4 = a1+b1 */
667
668 psraw_i2r (COL_SHIFT, xmm4); /* xmm4 = y1 */
669 psubsw_r2r (xmm1, xmm6); /* xmm6 = a1-b1 */
670
671 movdqa_m2r (*(col+5*8), xmm1); /* xmm1 = b0 */
672 psubsw_r2r (xmm7, xmm2); /* xmm2 = a2-b2 */
673
674 psraw_i2r (COL_SHIFT, xmm6); /* xmm6 = y6 */
675 movdqa_r2r (xmm5, xmm7); /* xmm7 = a0 */
676
677 movdqa_r2m (xmm4, *(col+1*8)); /* save y1 */
678 psraw_i2r (COL_SHIFT, xmm2); /* xmm2 = y5 */
679
680 movdqa_r2m (xmm3, *(col+2*8)); /* save y2 */
681 paddsw_r2r (xmm1, xmm5); /* xmm5 = a0+b0 */
682
683 movdqa_m2r (*(col+3*8), xmm4); /* xmm4 = b3 */
684 psubsw_r2r (xmm1, xmm7); /* xmm7 = a0-b0 */
685
686 psraw_i2r (COL_SHIFT, xmm5); /* xmm5 = y0 */
687 movdqa_r2r (xmm0, xmm3); /* xmm3 = a3 */
688
689 movdqa_r2m (xmm2, *(col+5*8)); /* save y5 */
690 psubsw_r2r (xmm4, xmm3); /* xmm3 = a3-b3 */
691
692 psraw_i2r (COL_SHIFT, xmm7); /* xmm7 = y7 */
693 paddsw_r2r (xmm0, xmm4); /* xmm4 = a3+b3 */
694
695 movdqa_r2m (xmm5, *(col+0*8)); /* save y0 */
696 psraw_i2r (COL_SHIFT, xmm3); /* xmm3 = y4 */
697
698 movdqa_r2m (xmm6, *(col+6*8)); /* save y6 */
699 psraw_i2r (COL_SHIFT, xmm4); /* xmm4 = y3 */
700
701 movdqa_r2m (xmm7, *(col+7*8)); /* save y7 */
702
703 movdqa_r2m (xmm3, *(col+4*8)); /* save y4 */
704
705 movdqa_r2m (xmm4, *(col+3*8)); /* save y3 */
706 #endif
707 }
708
709
710 /* MMX column IDCT */
idct_col(int16_t * const col,const int offset)711 static inline void idct_col (int16_t * const col, const int offset)
712 {
713 static const short t1_vector[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
714 static const short t2_vector[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
715 static const short t3_vector[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
716 static const short c4_vector[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
717
718 /* column code adapted from peter gubanov */
719 /* http://www.elecard.com/peter/idct.shtml */
720
721 movq_m2r (*t1_vector, mm0); /* mm0 = T1 */
722
723 movq_m2r (*(col+offset+1*8), mm1); /* mm1 = x1 */
724 movq_r2r (mm0, mm2); /* mm2 = T1 */
725
726 movq_m2r (*(col+offset+7*8), mm4); /* mm4 = x7 */
727 pmulhw_r2r (mm1, mm0); /* mm0 = T1*x1 */
728
729 movq_m2r (*t3_vector, mm5); /* mm5 = T3 */
730 pmulhw_r2r (mm4, mm2); /* mm2 = T1*x7 */
731
732 movq_m2r (*(col+offset+5*8), mm6); /* mm6 = x5 */
733 movq_r2r (mm5, mm7); /* mm7 = T3-1 */
734
735 movq_m2r (*(col+offset+3*8), mm3); /* mm3 = x3 */
736 psubsw_r2r (mm4, mm0); /* mm0 = v17 */
737
738 movq_m2r (*t2_vector, mm4); /* mm4 = T2 */
739 pmulhw_r2r (mm3, mm5); /* mm5 = (T3-1)*x3 */
740
741 paddsw_r2r (mm2, mm1); /* mm1 = u17 */
742 pmulhw_r2r (mm6, mm7); /* mm7 = (T3-1)*x5 */
743
744 /* slot */
745
746 movq_r2r (mm4, mm2); /* mm2 = T2 */
747 paddsw_r2r (mm3, mm5); /* mm5 = T3*x3 */
748
749 pmulhw_m2r (*(col+offset+2*8), mm4);/* mm4 = T2*x2 */
750 paddsw_r2r (mm6, mm7); /* mm7 = T3*x5 */
751
752 psubsw_r2r (mm6, mm5); /* mm5 = v35 */
753 paddsw_r2r (mm3, mm7); /* mm7 = u35 */
754
755 movq_m2r (*(col+offset+6*8), mm3); /* mm3 = x6 */
756 movq_r2r (mm0, mm6); /* mm6 = v17 */
757
758 pmulhw_r2r (mm3, mm2); /* mm2 = T2*x6 */
759 psubsw_r2r (mm5, mm0); /* mm0 = b3 */
760
761 psubsw_r2r (mm3, mm4); /* mm4 = v26 */
762 paddsw_r2r (mm6, mm5); /* mm5 = v12 */
763
764 movq_r2m (mm0, *(col+offset+3*8)); /* save b3 in scratch0 */
765 movq_r2r (mm1, mm6); /* mm6 = u17 */
766
767 paddsw_m2r (*(col+offset+2*8), mm2);/* mm2 = u26 */
768 paddsw_r2r (mm7, mm6); /* mm6 = b0 */
769
770 psubsw_r2r (mm7, mm1); /* mm1 = u12 */
771 movq_r2r (mm1, mm7); /* mm7 = u12 */
772
773 movq_m2r (*(col+offset+0*8), mm3); /* mm3 = x0 */
774 paddsw_r2r (mm5, mm1); /* mm1 = u12+v12 */
775
776 movq_m2r (*c4_vector, mm0); /* mm0 = C4/2 */
777 psubsw_r2r (mm5, mm7); /* mm7 = u12-v12 */
778
779 movq_r2m (mm6, *(col+offset+5*8)); /* save b0 in scratch1 */
780 pmulhw_r2r (mm0, mm1); /* mm1 = b1/2 */
781
782 movq_r2r (mm4, mm6); /* mm6 = v26 */
783 pmulhw_r2r (mm0, mm7); /* mm7 = b2/2 */
784
785 movq_m2r (*(col+offset+4*8), mm5); /* mm5 = x4 */
786 movq_r2r (mm3, mm0); /* mm0 = x0 */
787
788 psubsw_r2r (mm5, mm3); /* mm3 = v04 */
789 paddsw_r2r (mm5, mm0); /* mm0 = u04 */
790
791 paddsw_r2r (mm3, mm4); /* mm4 = a1 */
792 movq_r2r (mm0, mm5); /* mm5 = u04 */
793
794 psubsw_r2r (mm6, mm3); /* mm3 = a2 */
795 paddsw_r2r (mm2, mm5); /* mm5 = a0 */
796
797 paddsw_r2r (mm1, mm1); /* mm1 = b1 */
798 psubsw_r2r (mm2, mm0); /* mm0 = a3 */
799
800 paddsw_r2r (mm7, mm7); /* mm7 = b2 */
801 movq_r2r (mm3, mm2); /* mm2 = a2 */
802
803 movq_r2r (mm4, mm6); /* mm6 = a1 */
804 paddsw_r2r (mm7, mm3); /* mm3 = a2+b2 */
805
806 psraw_i2r (COL_SHIFT, mm3); /* mm3 = y2 */
807 paddsw_r2r (mm1, mm4); /* mm4 = a1+b1 */
808
809 psraw_i2r (COL_SHIFT, mm4); /* mm4 = y1 */
810 psubsw_r2r (mm1, mm6); /* mm6 = a1-b1 */
811
812 movq_m2r (*(col+offset+5*8), mm1); /* mm1 = b0 */
813 psubsw_r2r (mm7, mm2); /* mm2 = a2-b2 */
814
815 psraw_i2r (COL_SHIFT, mm6); /* mm6 = y6 */
816 movq_r2r (mm5, mm7); /* mm7 = a0 */
817
818 movq_r2m (mm4, *(col+offset+1*8)); /* save y1 */
819 psraw_i2r (COL_SHIFT, mm2); /* mm2 = y5 */
820
821 movq_r2m (mm3, *(col+offset+2*8)); /* save y2 */
822 paddsw_r2r (mm1, mm5); /* mm5 = a0+b0 */
823
824 movq_m2r (*(col+offset+3*8), mm4); /* mm4 = b3 */
825 psubsw_r2r (mm1, mm7); /* mm7 = a0-b0 */
826
827 psraw_i2r (COL_SHIFT, mm5); /* mm5 = y0 */
828 movq_r2r (mm0, mm3); /* mm3 = a3 */
829
830 movq_r2m (mm2, *(col+offset+5*8)); /* save y5 */
831 psubsw_r2r (mm4, mm3); /* mm3 = a3-b3 */
832
833 psraw_i2r (COL_SHIFT, mm7); /* mm7 = y7 */
834 paddsw_r2r (mm0, mm4); /* mm4 = a3+b3 */
835
836 movq_r2m (mm5, *(col+offset+0*8)); /* save y0 */
837 psraw_i2r (COL_SHIFT, mm3); /* mm3 = y4 */
838
839 movq_r2m (mm6, *(col+offset+6*8)); /* save y6 */
840 psraw_i2r (COL_SHIFT, mm4); /* mm4 = y3 */
841
842 movq_r2m (mm7, *(col+offset+7*8)); /* save y7 */
843
844 movq_r2m (mm3, *(col+offset+4*8)); /* save y4 */
845
846 movq_r2m (mm4, *(col+offset+3*8)); /* save y3 */
847 }
848
849
850 static const int32_t rounder0[] ATTR_ALIGN(8) =
851 rounder ((1 << (COL_SHIFT - 1)) - 0.5);
852 static const int32_t rounder4[] ATTR_ALIGN(8) = rounder (0);
853 static const int32_t rounder1[] ATTR_ALIGN(8) =
854 rounder (1.25683487303); /* C1*(C1/C4+C1+C7)/2 */
855 static const int32_t rounder7[] ATTR_ALIGN(8) =
856 rounder (-0.25); /* C1*(C7/C4+C7-C1)/2 */
857 static const int32_t rounder2[] ATTR_ALIGN(8) =
858 rounder (0.60355339059); /* C2 * (C6+C2)/2 */
859 static const int32_t rounder6[] ATTR_ALIGN(8) =
860 rounder (-0.25); /* C2 * (C6-C2)/2 */
861 static const int32_t rounder3[] ATTR_ALIGN(8) =
862 rounder (0.087788325588); /* C3*(-C3/C4+C3+C5)/2 */
863 static const int32_t rounder5[] ATTR_ALIGN(8) =
864 rounder (-0.441341716183); /* C3*(-C5/C4+C5-C3)/2 */
865
866
867 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
868 static inline void idct (int16_t * const block) \
869 { \
870 static const int16_t table04[] ATTR_ALIGN(16) = \
871 table (22725, 21407, 19266, 16384, 12873, 8867, 4520); \
872 static const int16_t table17[] ATTR_ALIGN(16) = \
873 table (31521, 29692, 26722, 22725, 17855, 12299, 6270); \
874 static const int16_t table26[] ATTR_ALIGN(16) = \
875 table (29692, 27969, 25172, 21407, 16819, 11585, 5906); \
876 static const int16_t table35[] ATTR_ALIGN(16) = \
877 table (26722, 25172, 22654, 19266, 15137, 10426, 5315); \
878 \
879 idct_row_head (block, 0*8, table04); \
880 idct_row (table04, rounder0); \
881 idct_row_mid (block, 0*8, 4*8, table04); \
882 idct_row (table04, rounder4); \
883 idct_row_mid (block, 4*8, 1*8, table17); \
884 idct_row (table17, rounder1); \
885 idct_row_mid (block, 1*8, 7*8, table17); \
886 idct_row (table17, rounder7); \
887 idct_row_mid (block, 7*8, 2*8, table26); \
888 idct_row (table26, rounder2); \
889 idct_row_mid (block, 2*8, 6*8, table26); \
890 idct_row (table26, rounder6); \
891 idct_row_mid (block, 6*8, 3*8, table35); \
892 idct_row (table35, rounder3); \
893 idct_row_mid (block, 3*8, 5*8, table35); \
894 idct_row (table35, rounder5); \
895 idct_row_tail (block, 5*8); \
896 \
897 idct_col (block, 0); \
898 idct_col (block, 4); \
899 }
900
sse2_idct(int16_t * const block)901 static inline void sse2_idct (int16_t * const block)
902 {
903 static const int16_t table04[] ATTR_ALIGN(16) =
904 sse2_table (22725, 21407, 19266, 16384, 12873, 8867, 4520);
905 static const int16_t table17[] ATTR_ALIGN(16) =
906 sse2_table (31521, 29692, 26722, 22725, 17855, 12299, 6270);
907 static const int16_t table26[] ATTR_ALIGN(16) =
908 sse2_table (29692, 27969, 25172, 21407, 16819, 11585, 5906);
909 static const int16_t table35[] ATTR_ALIGN(16) =
910 sse2_table (26722, 25172, 22654, 19266, 15137, 10426, 5315);
911
912 static const int32_t rounder0_128[] ATTR_ALIGN(16) =
913 rounder_sse2 ((1 << (COL_SHIFT - 1)) - 0.5);
914 static const int32_t rounder4_128[] ATTR_ALIGN(16) = rounder_sse2 (0);
915 static const int32_t rounder1_128[] ATTR_ALIGN(16) =
916 rounder_sse2 (1.25683487303); /* C1*(C1/C4+C1+C7)/2 */
917 static const int32_t rounder7_128[] ATTR_ALIGN(16) =
918 rounder_sse2 (-0.25); /* C1*(C7/C4+C7-C1)/2 */
919 static const int32_t rounder2_128[] ATTR_ALIGN(16) =
920 rounder_sse2 (0.60355339059); /* C2 * (C6+C2)/2 */
921 static const int32_t rounder6_128[] ATTR_ALIGN(16) =
922 rounder_sse2 (-0.25); /* C2 * (C6-C2)/2 */
923 static const int32_t rounder3_128[] ATTR_ALIGN(16) =
924 rounder_sse2 (0.087788325588); /* C3*(-C3/C4+C3+C5)/2 */
925 static const int32_t rounder5_128[] ATTR_ALIGN(16) =
926 rounder_sse2 (-0.441341716183); /* C3*(-C5/C4+C5-C3)/2 */
927
928 #if defined(__x86_64__)
929 movdqa_m2r (block[0*8], xmm8);
930 movdqa_m2r (block[4*8], xmm12);
931 SSE2_IDCT_2ROW (table04, xmm8, xmm12, *rounder0_128, *rounder4_128);
932
933 movdqa_m2r (block[1*8], xmm9);
934 movdqa_m2r (block[7*8], xmm15);
935 SSE2_IDCT_2ROW (table17, xmm9, xmm15, *rounder1_128, *rounder7_128);
936
937 movdqa_m2r (block[2*8], xmm10);
938 movdqa_m2r (block[6*8], xmm14);
939 SSE2_IDCT_2ROW (table26, xmm10, xmm14, *rounder2_128, *rounder6_128);
940
941 movdqa_m2r (block[3*8], xmm11);
942 movdqa_m2r (block[5*8], xmm13);
943 SSE2_IDCT_2ROW (table35, xmm11, xmm13, *rounder3_128, *rounder5_128);
944
945 /* OUTPUT: block in xmm8 ... xmm15 */
946
947 #else
948 movdqa_m2r (block[0*8], xmm0);
949 movdqa_m2r (block[4*8], xmm4);
950 SSE2_IDCT_2ROW (table04, xmm0, xmm4, *rounder0_128, *rounder4_128);
951 movdqa_r2m (xmm0, block[0*8]);
952 movdqa_r2m (xmm4, block[4*8]);
953
954 movdqa_m2r (block[1*8], xmm0);
955 movdqa_m2r (block[7*8], xmm4);
956 SSE2_IDCT_2ROW (table17, xmm0, xmm4, *rounder1_128, *rounder7_128);
957 movdqa_r2m (xmm0, block[1*8]);
958 movdqa_r2m (xmm4, block[7*8]);
959
960 movdqa_m2r (block[2*8], xmm0);
961 movdqa_m2r (block[6*8], xmm4);
962 SSE2_IDCT_2ROW (table26, xmm0, xmm4, *rounder2_128, *rounder6_128);
963 movdqa_r2m (xmm0, block[2*8]);
964 movdqa_r2m (xmm4, block[6*8]);
965
966 movdqa_m2r (block[3*8], xmm0);
967 movdqa_m2r (block[5*8], xmm4);
968 SSE2_IDCT_2ROW (table35, xmm0, xmm4, *rounder3_128, *rounder5_128);
969 movdqa_r2m (xmm0, block[3*8]);
970 movdqa_r2m (xmm4, block[5*8]);
971 #endif
972
973 sse2_idct_col (block);
974 }
975
sse2_block_copy(int16_t * const block,uint8_t * dest,const int stride)976 static void sse2_block_copy (int16_t * const block, uint8_t * dest,
977 const int stride)
978 {
979 #if defined(__x86_64__)
980 /* INPUT: block in xmm8 ... xmm15 */
981 packuswb_r2r (xmm8, xmm8);
982 packuswb_r2r (xmm9, xmm9);
983 movq_r2m (xmm8, *(dest+0*stride));
984 packuswb_r2r (xmm10, xmm10);
985 movq_r2m (xmm9, *(dest+1*stride));
986 packuswb_r2r (xmm11, xmm11);
987 movq_r2m (xmm10, *(dest+2*stride));
988 packuswb_r2r (xmm12, xmm12);
989 movq_r2m (xmm11, *(dest+3*stride));
990 packuswb_r2r (xmm13, xmm13);
991 movq_r2m (xmm12, *(dest+4*stride));
992 packuswb_r2r (xmm14, xmm14);
993 movq_r2m (xmm13, *(dest+5*stride));
994 packuswb_r2r (xmm15, xmm15);
995 movq_r2m (xmm14, *(dest+6*stride));
996 movq_r2m (xmm15, *(dest+7*stride));
997 #else
998 movdqa_m2r (*(block+0*8), xmm0);
999 movdqa_m2r (*(block+1*8), xmm1);
1000 movdqa_m2r (*(block+2*8), xmm2);
1001 packuswb_r2r (xmm0, xmm0);
1002 movdqa_m2r (*(block+3*8), xmm3);
1003 packuswb_r2r (xmm1, xmm1);
1004 movdqa_m2r (*(block+4*8), xmm4);
1005 packuswb_r2r (xmm2, xmm2);
1006 movdqa_m2r (*(block+5*8), xmm5);
1007 packuswb_r2r (xmm3, xmm3);
1008 movdqa_m2r (*(block+6*8), xmm6);
1009 packuswb_r2r (xmm4, xmm4);
1010 movdqa_m2r (*(block+7*8), xmm7);
1011 movq_r2m (xmm0, *(dest+0*stride));
1012 packuswb_r2r (xmm5, xmm5);
1013 movq_r2m (xmm1, *(dest+1*stride));
1014 packuswb_r2r (xmm6, xmm6);
1015 movq_r2m (xmm2, *(dest+2*stride));
1016 packuswb_r2r (xmm7, xmm7);
1017 movq_r2m (xmm3, *(dest+3*stride));
1018 movq_r2m (xmm4, *(dest+4*stride));
1019 movq_r2m (xmm5, *(dest+5*stride));
1020 movq_r2m (xmm6, *(dest+6*stride));
1021 movq_r2m (xmm7, *(dest+7*stride));
1022 #endif
1023 }
1024
1025 #define COPY_MMX(offset,r0,r1,r2) \
1026 do { \
1027 movq_m2r (*(block+offset), r0); \
1028 dest += stride; \
1029 movq_m2r (*(block+offset+4), r1); \
1030 movq_r2m (r2, *dest); \
1031 packuswb_r2r (r1, r0); \
1032 } while (0)
1033
block_copy(int16_t * const block,uint8_t * dest,const int stride)1034 static inline void block_copy (int16_t * const block, uint8_t * dest,
1035 const int stride)
1036 {
1037 movq_m2r (*(block+0*8), mm0);
1038 movq_m2r (*(block+0*8+4), mm1);
1039 movq_m2r (*(block+1*8), mm2);
1040 packuswb_r2r (mm1, mm0);
1041 movq_m2r (*(block+1*8+4), mm3);
1042 movq_r2m (mm0, *dest);
1043 packuswb_r2r (mm3, mm2);
1044 COPY_MMX (2*8, mm0, mm1, mm2);
1045 COPY_MMX (3*8, mm2, mm3, mm0);
1046 COPY_MMX (4*8, mm0, mm1, mm2);
1047 COPY_MMX (5*8, mm2, mm3, mm0);
1048 COPY_MMX (6*8, mm0, mm1, mm2);
1049 COPY_MMX (7*8, mm2, mm3, mm0);
1050 movq_r2m (mm2, *(dest+stride));
1051 }
1052
1053 #define ADD_SSE2_2ROW(op, block0, block1)\
1054 do { \
1055 movq_m2r (*(dest), xmm1); \
1056 movq_m2r (*(dest+stride), xmm2); \
1057 punpcklbw_r2r (xmm0, xmm1); \
1058 punpcklbw_r2r (xmm0, xmm2); \
1059 paddsw_##op (block0, xmm1); \
1060 paddsw_##op (block1, xmm2); \
1061 packuswb_r2r (xmm1, xmm1); \
1062 packuswb_r2r (xmm2, xmm2); \
1063 movq_r2m (xmm1, *(dest)); \
1064 movq_r2m (xmm2, *(dest+stride)); \
1065 dest += 2*stride; \
1066 } while (0)
1067
sse2_block_add(int16_t * const block,uint8_t * dest,const int stride)1068 static void sse2_block_add (int16_t * const block, uint8_t * dest,
1069 const int stride)
1070 {
1071 pxor_r2r(xmm0, xmm0);
1072 #if defined(__x86_64__)
1073 /* INPUT: block in xmm8 ... xmm15 */
1074 ADD_SSE2_2ROW(r2r, xmm8, xmm9);
1075 ADD_SSE2_2ROW(r2r, xmm10, xmm11);
1076 ADD_SSE2_2ROW(r2r, xmm12, xmm13);
1077 ADD_SSE2_2ROW(r2r, xmm14, xmm15);
1078 #else
1079 ADD_SSE2_2ROW(m2r, *(block+0*8), *(block+1*8));
1080 ADD_SSE2_2ROW(m2r, *(block+2*8), *(block+3*8));
1081 ADD_SSE2_2ROW(m2r, *(block+4*8), *(block+5*8));
1082 ADD_SSE2_2ROW(m2r, *(block+6*8), *(block+7*8));
1083 #endif
1084 }
1085
1086 #define ADD_MMX(offset,r1,r2,r3,r4) \
1087 do { \
1088 movq_m2r (*(dest+2*stride), r1); \
1089 packuswb_r2r (r4, r3); \
1090 movq_r2r (r1, r2); \
1091 dest += stride; \
1092 movq_r2m (r3, *dest); \
1093 punpcklbw_r2r (mm0, r1); \
1094 paddsw_m2r (*(block+offset), r1); \
1095 punpckhbw_r2r (mm0, r2); \
1096 paddsw_m2r (*(block+offset+4), r2); \
1097 } while (0)
1098
block_add(int16_t * const block,uint8_t * dest,const int stride)1099 static inline void block_add (int16_t * const block, uint8_t * dest,
1100 const int stride)
1101 {
1102 movq_m2r (*dest, mm1);
1103 pxor_r2r (mm0, mm0);
1104 movq_m2r (*(dest+stride), mm3);
1105 movq_r2r (mm1, mm2);
1106 punpcklbw_r2r (mm0, mm1);
1107 movq_r2r (mm3, mm4);
1108 paddsw_m2r (*(block+0*8), mm1);
1109 punpckhbw_r2r (mm0, mm2);
1110 paddsw_m2r (*(block+0*8+4), mm2);
1111 punpcklbw_r2r (mm0, mm3);
1112 paddsw_m2r (*(block+1*8), mm3);
1113 packuswb_r2r (mm2, mm1);
1114 punpckhbw_r2r (mm0, mm4);
1115 movq_r2m (mm1, *dest);
1116 paddsw_m2r (*(block+1*8+4), mm4);
1117 ADD_MMX (2*8, mm1, mm2, mm3, mm4);
1118 ADD_MMX (3*8, mm3, mm4, mm1, mm2);
1119 ADD_MMX (4*8, mm1, mm2, mm3, mm4);
1120 ADD_MMX (5*8, mm3, mm4, mm1, mm2);
1121 ADD_MMX (6*8, mm1, mm2, mm3, mm4);
1122 ADD_MMX (7*8, mm3, mm4, mm1, mm2);
1123 packuswb_r2r (mm4, mm3);
1124 movq_r2m (mm3, *(dest+stride));
1125 }
1126
1127
sse2_block_zero(int16_t * const block)1128 static inline void sse2_block_zero (int16_t * const block)
1129 {
1130 pxor_r2r (xmm0, xmm0);
1131 movdqa_r2m (xmm0, *(block+0*8));
1132 movdqa_r2m (xmm0, *(block+1*8));
1133 movdqa_r2m (xmm0, *(block+2*8));
1134 movdqa_r2m (xmm0, *(block+3*8));
1135 movdqa_r2m (xmm0, *(block+4*8));
1136 movdqa_r2m (xmm0, *(block+5*8));
1137 movdqa_r2m (xmm0, *(block+6*8));
1138 movdqa_r2m (xmm0, *(block+7*8));
1139 }
1140
block_zero(int16_t * const block)1141 static inline void block_zero (int16_t * const block)
1142 {
1143 pxor_r2r (mm0, mm0);
1144 movq_r2m (mm0, *(block+0*4));
1145 movq_r2m (mm0, *(block+1*4));
1146 movq_r2m (mm0, *(block+2*4));
1147 movq_r2m (mm0, *(block+3*4));
1148 movq_r2m (mm0, *(block+4*4));
1149 movq_r2m (mm0, *(block+5*4));
1150 movq_r2m (mm0, *(block+6*4));
1151 movq_r2m (mm0, *(block+7*4));
1152 movq_r2m (mm0, *(block+8*4));
1153 movq_r2m (mm0, *(block+9*4));
1154 movq_r2m (mm0, *(block+10*4));
1155 movq_r2m (mm0, *(block+11*4));
1156 movq_r2m (mm0, *(block+12*4));
1157 movq_r2m (mm0, *(block+13*4));
1158 movq_r2m (mm0, *(block+14*4));
1159 movq_r2m (mm0, *(block+15*4));
1160 }
1161
1162
1163 #define CPU_MMXEXT 0
1164 #define CPU_MMX 1
1165
1166 #define dup4(reg) \
1167 do { \
1168 if (cpu != CPU_MMXEXT) { \
1169 punpcklwd_r2r (reg, reg); \
1170 punpckldq_r2r (reg, reg); \
1171 } else \
1172 pshufw_r2r (reg, reg, 0x00); \
1173 } while (0)
1174
block_add_DC(int16_t * const block,uint8_t * dest,const int stride,const int cpu)1175 static inline void block_add_DC (int16_t * const block, uint8_t * dest,
1176 const int stride, const int cpu)
1177 {
1178 movd_v2r ((block[0] + 64) >> 7, mm0);
1179 pxor_r2r (mm1, mm1);
1180 movq_m2r (*dest, mm2);
1181 dup4 (mm0);
1182 psubsw_r2r (mm0, mm1);
1183 packuswb_r2r (mm0, mm0);
1184 paddusb_r2r (mm0, mm2);
1185 packuswb_r2r (mm1, mm1);
1186 movq_m2r (*(dest + stride), mm3);
1187 psubusb_r2r (mm1, mm2);
1188 block[0] = 0;
1189 paddusb_r2r (mm0, mm3);
1190 movq_r2m (mm2, *dest);
1191 psubusb_r2r (mm1, mm3);
1192 movq_m2r (*(dest + 2*stride), mm2);
1193 dest += stride;
1194 movq_r2m (mm3, *dest);
1195 paddusb_r2r (mm0, mm2);
1196 movq_m2r (*(dest + 2*stride), mm3);
1197 psubusb_r2r (mm1, mm2);
1198 dest += stride;
1199 paddusb_r2r (mm0, mm3);
1200 movq_r2m (mm2, *dest);
1201 psubusb_r2r (mm1, mm3);
1202 movq_m2r (*(dest + 2*stride), mm2);
1203 dest += stride;
1204 movq_r2m (mm3, *dest);
1205 paddusb_r2r (mm0, mm2);
1206 movq_m2r (*(dest + 2*stride), mm3);
1207 psubusb_r2r (mm1, mm2);
1208 dest += stride;
1209 paddusb_r2r (mm0, mm3);
1210 movq_r2m (mm2, *dest);
1211 psubusb_r2r (mm1, mm3);
1212 movq_m2r (*(dest + 2*stride), mm2);
1213 dest += stride;
1214 movq_r2m (mm3, *dest);
1215 paddusb_r2r (mm0, mm2);
1216 movq_m2r (*(dest + 2*stride), mm3);
1217 psubusb_r2r (mm1, mm2);
1218 block[63] = 0;
1219 paddusb_r2r (mm0, mm3);
1220 movq_r2m (mm2, *(dest + stride));
1221 psubusb_r2r (mm1, mm3);
1222 movq_r2m (mm3, *(dest + 2*stride));
1223 }
1224
mpeg2_idct_copy_sse2(int16_t * const block,uint8_t * const dest,const int stride)1225 void mpeg2_idct_copy_sse2 (int16_t * const block, uint8_t * const dest,
1226 const int stride)
1227 {
1228 sse2_idct (block);
1229 sse2_block_copy (block, dest, stride);
1230 sse2_block_zero (block);
1231 }
1232
mpeg2_idct_add_sse2(const int last,int16_t * const block,uint8_t * const dest,const int stride)1233 void mpeg2_idct_add_sse2 (const int last, int16_t * const block,
1234 uint8_t * const dest, const int stride)
1235 {
1236 if (last != 129 || (block[0] & (7 << 4)) == (4 << 4)) {
1237 sse2_idct (block);
1238 sse2_block_add (block, dest, stride);
1239 sse2_block_zero (block);
1240 } else
1241 block_add_DC (block, dest, stride, CPU_MMXEXT);
1242 }
1243
1244
declare_idct(mmxext_idct,mmxext_table,mmxext_row_head,mmxext_row,mmxext_row_tail,mmxext_row_mid)1245 declare_idct (mmxext_idct, mmxext_table,
1246 mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
1247
1248 void mpeg2_idct_copy_mmxext (int16_t * const block, uint8_t * const dest,
1249 const int stride)
1250 {
1251 mmxext_idct (block);
1252 block_copy (block, dest, stride);
1253 block_zero (block);
1254 }
1255
mpeg2_idct_add_mmxext(const int last,int16_t * const block,uint8_t * const dest,const int stride)1256 void mpeg2_idct_add_mmxext (const int last, int16_t * const block,
1257 uint8_t * const dest, const int stride)
1258 {
1259 if (last != 129 || (block[0] & (7 << 4)) == (4 << 4)) {
1260 mmxext_idct (block);
1261 block_add (block, dest, stride);
1262 block_zero (block);
1263 } else
1264 block_add_DC (block, dest, stride, CPU_MMXEXT);
1265 }
1266
1267
declare_idct(mmx_idct,mmx_table,mmx_row_head,mmx_row,mmx_row_tail,mmx_row_mid)1268 declare_idct (mmx_idct, mmx_table,
1269 mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
1270
1271 void mpeg2_idct_copy_mmx (int16_t * const block, uint8_t * const dest,
1272 const int stride)
1273 {
1274 mmx_idct (block);
1275 block_copy (block, dest, stride);
1276 block_zero (block);
1277 }
1278
mpeg2_idct_add_mmx(const int last,int16_t * const block,uint8_t * const dest,const int stride)1279 void mpeg2_idct_add_mmx (const int last, int16_t * const block,
1280 uint8_t * const dest, const int stride)
1281 {
1282 if (last != 129 || (block[0] & (7 << 4)) == (4 << 4)) {
1283 mmx_idct (block);
1284 block_add (block, dest, stride);
1285 block_zero (block);
1286 } else
1287 block_add_DC (block, dest, stride, CPU_MMX);
1288 }
1289
1290
mpeg2_idct_mmx_init(void)1291 void mpeg2_idct_mmx_init (void)
1292 {
1293 int i, j;
1294
1295 /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
1296
1297 for (i = 0; i < 64; i++) {
1298 j = mpeg2_scan_norm[i];
1299 mpeg2_scan_norm[i] = (j & 0x38) | ((j & 6) >> 1) | ((j & 1) << 2);
1300 j = mpeg2_scan_alt[i];
1301 mpeg2_scan_alt[i] = (j & 0x38) | ((j & 6) >> 1) | ((j & 1) << 2);
1302 }
1303 }
1304
1305 #endif
1306