Merge remote-tracking branch 'qatar/master'
[ffmpeg.git] / libavcodec / x86 / mpegvideo_mmx.c
1 /*
2  * The simplest mpeg encoder (well, it was the simplest!)
3  * Copyright (c) 2000,2001 Fabrice Bellard
4  *
5  * Optimized for ia32 CPUs by Nick Kurshev <nickols_k@mail.ru>
6  * h263, mpeg1, mpeg2 dequantizer & draw_edges by Michael Niedermayer <michaelni@gmx.at>
7  *
8  * This file is part of FFmpeg.
9  *
10  * FFmpeg is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU Lesser General Public
12  * License as published by the Free Software Foundation; either
13  * version 2.1 of the License, or (at your option) any later version.
14  *
15  * FFmpeg is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * Lesser General Public License for more details.
19  *
20  * You should have received a copy of the GNU Lesser General Public
21  * License along with FFmpeg; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23  */
24
25 #include "libavutil/cpu.h"
26 #include "libavutil/x86_cpu.h"
27 #include "libavcodec/avcodec.h"
28 #include "libavcodec/dsputil.h"
29 #include "libavcodec/mpegvideo.h"
30 #include "dsputil_mmx.h"
31
32 extern uint16_t ff_inv_zigzag_direct16[64];
33
34
35 static void dct_unquantize_h263_intra_mmx(MpegEncContext *s,
36                                   DCTELEM *block, int n, int qscale)
37 {
38     x86_reg level, qmul, qadd, nCoeffs;
39
40     qmul = qscale << 1;
41
42     assert(s->block_last_index[n]>=0 || s->h263_aic);
43
44     if (!s->h263_aic) {
45         if (n < 4)
46             level = block[0] * s->y_dc_scale;
47         else
48             level = block[0] * s->c_dc_scale;
49         qadd = (qscale - 1) | 1;
50     }else{
51         qadd = 0;
52         level= block[0];
53     }
54     if(s->ac_pred)
55         nCoeffs=63;
56     else
57         nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
58 //printf("%d %d  ", qmul, qadd);
59 __asm__ volatile(
60                 "movd %1, %%mm6                 \n\t" //qmul
61                 "packssdw %%mm6, %%mm6          \n\t"
62                 "packssdw %%mm6, %%mm6          \n\t"
63                 "movd %2, %%mm5                 \n\t" //qadd
64                 "pxor %%mm7, %%mm7              \n\t"
65                 "packssdw %%mm5, %%mm5          \n\t"
66                 "packssdw %%mm5, %%mm5          \n\t"
67                 "psubw %%mm5, %%mm7             \n\t"
68                 "pxor %%mm4, %%mm4              \n\t"
69                 ".p2align 4                     \n\t"
70                 "1:                             \n\t"
71                 "movq (%0, %3), %%mm0           \n\t"
72                 "movq 8(%0, %3), %%mm1          \n\t"
73
74                 "pmullw %%mm6, %%mm0            \n\t"
75                 "pmullw %%mm6, %%mm1            \n\t"
76
77                 "movq (%0, %3), %%mm2           \n\t"
78                 "movq 8(%0, %3), %%mm3          \n\t"
79
80                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
81                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
82
83                 "pxor %%mm2, %%mm0              \n\t"
84                 "pxor %%mm3, %%mm1              \n\t"
85
86                 "paddw %%mm7, %%mm0             \n\t"
87                 "paddw %%mm7, %%mm1             \n\t"
88
89                 "pxor %%mm0, %%mm2              \n\t"
90                 "pxor %%mm1, %%mm3              \n\t"
91
92                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
93                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
94
95                 "pandn %%mm2, %%mm0             \n\t"
96                 "pandn %%mm3, %%mm1             \n\t"
97
98                 "movq %%mm0, (%0, %3)           \n\t"
99                 "movq %%mm1, 8(%0, %3)          \n\t"
100
101                 "add $16, %3                    \n\t"
102                 "jng 1b                         \n\t"
103                 ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs))
104                 : "memory"
105         );
106         block[0]= level;
107 }
108
109
110 static void dct_unquantize_h263_inter_mmx(MpegEncContext *s,
111                                   DCTELEM *block, int n, int qscale)
112 {
113     x86_reg qmul, qadd, nCoeffs;
114
115     qmul = qscale << 1;
116     qadd = (qscale - 1) | 1;
117
118     assert(s->block_last_index[n]>=0 || s->h263_aic);
119
120     nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
121 //printf("%d %d  ", qmul, qadd);
122 __asm__ volatile(
123                 "movd %1, %%mm6                 \n\t" //qmul
124                 "packssdw %%mm6, %%mm6          \n\t"
125                 "packssdw %%mm6, %%mm6          \n\t"
126                 "movd %2, %%mm5                 \n\t" //qadd
127                 "pxor %%mm7, %%mm7              \n\t"
128                 "packssdw %%mm5, %%mm5          \n\t"
129                 "packssdw %%mm5, %%mm5          \n\t"
130                 "psubw %%mm5, %%mm7             \n\t"
131                 "pxor %%mm4, %%mm4              \n\t"
132                 ".p2align 4                     \n\t"
133                 "1:                             \n\t"
134                 "movq (%0, %3), %%mm0           \n\t"
135                 "movq 8(%0, %3), %%mm1          \n\t"
136
137                 "pmullw %%mm6, %%mm0            \n\t"
138                 "pmullw %%mm6, %%mm1            \n\t"
139
140                 "movq (%0, %3), %%mm2           \n\t"
141                 "movq 8(%0, %3), %%mm3          \n\t"
142
143                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
144                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
145
146                 "pxor %%mm2, %%mm0              \n\t"
147                 "pxor %%mm3, %%mm1              \n\t"
148
149                 "paddw %%mm7, %%mm0             \n\t"
150                 "paddw %%mm7, %%mm1             \n\t"
151
152                 "pxor %%mm0, %%mm2              \n\t"
153                 "pxor %%mm1, %%mm3              \n\t"
154
155                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
156                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
157
158                 "pandn %%mm2, %%mm0             \n\t"
159                 "pandn %%mm3, %%mm1             \n\t"
160
161                 "movq %%mm0, (%0, %3)           \n\t"
162                 "movq %%mm1, 8(%0, %3)          \n\t"
163
164                 "add $16, %3                    \n\t"
165                 "jng 1b                         \n\t"
166                 ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs))
167                 : "memory"
168         );
169 }
170
171
172 /*
173   NK:
174   Note: looking at PARANOID:
175   "enable all paranoid tests for rounding, overflows, etc..."
176
177 #ifdef PARANOID
178                 if (level < -2048 || level > 2047)
179                     fprintf(stderr, "unquant error %d %d\n", i, level);
180 #endif
181   We can suppose that result of two multiplications can't be greater than 0xFFFF
182   i.e. is 16-bit, so we use here only PMULLW instruction and can avoid
183   a complex multiplication.
184 =====================================================
185  Full formula for multiplication of 2 integer numbers
186  which are represent as high:low words:
187  input: value1 = high1:low1
188         value2 = high2:low2
189  output: value3 = value1*value2
190  value3=high3:low3 (on overflow: modulus 2^32 wrap-around)
191  this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4
192  but this algorithm will compute only 0x66cb0ce4
193  this limited by 16-bit size of operands
194  ---------------------------------
195  tlow1 = high1*low2
196  tlow2 = high2*low1
197  tlow1 = tlow1 + tlow2
198  high3:low3 = low1*low2
199  high3 += tlow1
200 */
201 static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s,
202                                      DCTELEM *block, int n, int qscale)
203 {
204     x86_reg nCoeffs;
205     const uint16_t *quant_matrix;
206     int block0;
207
208     assert(s->block_last_index[n]>=0);
209
210     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
211
212     if (n < 4)
213         block0 = block[0] * s->y_dc_scale;
214     else
215         block0 = block[0] * s->c_dc_scale;
216     /* XXX: only mpeg1 */
217     quant_matrix = s->intra_matrix;
218 __asm__ volatile(
219                 "pcmpeqw %%mm7, %%mm7           \n\t"
220                 "psrlw $15, %%mm7               \n\t"
221                 "movd %2, %%mm6                 \n\t"
222                 "packssdw %%mm6, %%mm6          \n\t"
223                 "packssdw %%mm6, %%mm6          \n\t"
224                 "mov %3, %%"REG_a"              \n\t"
225                 ".p2align 4                     \n\t"
226                 "1:                             \n\t"
227                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
228                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
229                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
230                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
231                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
232                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
233                 "pxor %%mm2, %%mm2              \n\t"
234                 "pxor %%mm3, %%mm3              \n\t"
235                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
236                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
237                 "pxor %%mm2, %%mm0              \n\t"
238                 "pxor %%mm3, %%mm1              \n\t"
239                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
240                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
241                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
242                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
243                 "pxor %%mm4, %%mm4              \n\t"
244                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
245                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
246                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
247                 "psraw $3, %%mm0                \n\t"
248                 "psraw $3, %%mm1                \n\t"
249                 "psubw %%mm7, %%mm0             \n\t"
250                 "psubw %%mm7, %%mm1             \n\t"
251                 "por %%mm7, %%mm0               \n\t"
252                 "por %%mm7, %%mm1               \n\t"
253                 "pxor %%mm2, %%mm0              \n\t"
254                 "pxor %%mm3, %%mm1              \n\t"
255                 "psubw %%mm2, %%mm0             \n\t"
256                 "psubw %%mm3, %%mm1             \n\t"
257                 "pandn %%mm0, %%mm4             \n\t"
258                 "pandn %%mm1, %%mm5             \n\t"
259                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
260                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
261
262                 "add $16, %%"REG_a"             \n\t"
263                 "js 1b                          \n\t"
264                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
265                 : "%"REG_a, "memory"
266         );
267     block[0]= block0;
268 }
269
270 static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s,
271                                      DCTELEM *block, int n, int qscale)
272 {
273     x86_reg nCoeffs;
274     const uint16_t *quant_matrix;
275
276     assert(s->block_last_index[n]>=0);
277
278     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
279
280         quant_matrix = s->inter_matrix;
281 __asm__ volatile(
282                 "pcmpeqw %%mm7, %%mm7           \n\t"
283                 "psrlw $15, %%mm7               \n\t"
284                 "movd %2, %%mm6                 \n\t"
285                 "packssdw %%mm6, %%mm6          \n\t"
286                 "packssdw %%mm6, %%mm6          \n\t"
287                 "mov %3, %%"REG_a"              \n\t"
288                 ".p2align 4                     \n\t"
289                 "1:                             \n\t"
290                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
291                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
292                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
293                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
294                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
295                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
296                 "pxor %%mm2, %%mm2              \n\t"
297                 "pxor %%mm3, %%mm3              \n\t"
298                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
299                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
300                 "pxor %%mm2, %%mm0              \n\t"
301                 "pxor %%mm3, %%mm1              \n\t"
302                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
303                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
304                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
305                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
306                 "paddw %%mm7, %%mm0             \n\t" // abs(block[i])*2 + 1
307                 "paddw %%mm7, %%mm1             \n\t" // abs(block[i])*2 + 1
308                 "pmullw %%mm4, %%mm0            \n\t" // (abs(block[i])*2 + 1)*q
309                 "pmullw %%mm5, %%mm1            \n\t" // (abs(block[i])*2 + 1)*q
310                 "pxor %%mm4, %%mm4              \n\t"
311                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
312                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
313                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
314                 "psraw $4, %%mm0                \n\t"
315                 "psraw $4, %%mm1                \n\t"
316                 "psubw %%mm7, %%mm0             \n\t"
317                 "psubw %%mm7, %%mm1             \n\t"
318                 "por %%mm7, %%mm0               \n\t"
319                 "por %%mm7, %%mm1               \n\t"
320                 "pxor %%mm2, %%mm0              \n\t"
321                 "pxor %%mm3, %%mm1              \n\t"
322                 "psubw %%mm2, %%mm0             \n\t"
323                 "psubw %%mm3, %%mm1             \n\t"
324                 "pandn %%mm0, %%mm4             \n\t"
325                 "pandn %%mm1, %%mm5             \n\t"
326                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
327                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
328
329                 "add $16, %%"REG_a"             \n\t"
330                 "js 1b                          \n\t"
331                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
332                 : "%"REG_a, "memory"
333         );
334 }
335
336 static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s,
337                                      DCTELEM *block, int n, int qscale)
338 {
339     x86_reg nCoeffs;
340     const uint16_t *quant_matrix;
341     int block0;
342
343     assert(s->block_last_index[n]>=0);
344
345     if(s->alternate_scan) nCoeffs= 63; //FIXME
346     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
347
348     if (n < 4)
349         block0 = block[0] * s->y_dc_scale;
350     else
351         block0 = block[0] * s->c_dc_scale;
352     quant_matrix = s->intra_matrix;
353 __asm__ volatile(
354                 "pcmpeqw %%mm7, %%mm7           \n\t"
355                 "psrlw $15, %%mm7               \n\t"
356                 "movd %2, %%mm6                 \n\t"
357                 "packssdw %%mm6, %%mm6          \n\t"
358                 "packssdw %%mm6, %%mm6          \n\t"
359                 "mov %3, %%"REG_a"              \n\t"
360                 ".p2align 4                     \n\t"
361                 "1:                             \n\t"
362                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
363                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
364                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
365                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
366                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
367                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
368                 "pxor %%mm2, %%mm2              \n\t"
369                 "pxor %%mm3, %%mm3              \n\t"
370                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
371                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
372                 "pxor %%mm2, %%mm0              \n\t"
373                 "pxor %%mm3, %%mm1              \n\t"
374                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
375                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
376                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
377                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
378                 "pxor %%mm4, %%mm4              \n\t"
379                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
380                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
381                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
382                 "psraw $3, %%mm0                \n\t"
383                 "psraw $3, %%mm1                \n\t"
384                 "pxor %%mm2, %%mm0              \n\t"
385                 "pxor %%mm3, %%mm1              \n\t"
386                 "psubw %%mm2, %%mm0             \n\t"
387                 "psubw %%mm3, %%mm1             \n\t"
388                 "pandn %%mm0, %%mm4             \n\t"
389                 "pandn %%mm1, %%mm5             \n\t"
390                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
391                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
392
393                 "add $16, %%"REG_a"             \n\t"
394                 "jng 1b                         \n\t"
395                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
396                 : "%"REG_a, "memory"
397         );
398     block[0]= block0;
399         //Note, we do not do mismatch control for intra as errors cannot accumulate
400 }
401
402 static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s,
403                                      DCTELEM *block, int n, int qscale)
404 {
405     x86_reg nCoeffs;
406     const uint16_t *quant_matrix;
407
408     assert(s->block_last_index[n]>=0);
409
410     if(s->alternate_scan) nCoeffs= 63; //FIXME
411     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
412
413         quant_matrix = s->inter_matrix;
414 __asm__ volatile(
415                 "pcmpeqw %%mm7, %%mm7           \n\t"
416                 "psrlq $48, %%mm7               \n\t"
417                 "movd %2, %%mm6                 \n\t"
418                 "packssdw %%mm6, %%mm6          \n\t"
419                 "packssdw %%mm6, %%mm6          \n\t"
420                 "mov %3, %%"REG_a"              \n\t"
421                 ".p2align 4                     \n\t"
422                 "1:                             \n\t"
423                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
424                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
425                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
426                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
427                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
428                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
429                 "pxor %%mm2, %%mm2              \n\t"
430                 "pxor %%mm3, %%mm3              \n\t"
431                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
432                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
433                 "pxor %%mm2, %%mm0              \n\t"
434                 "pxor %%mm3, %%mm1              \n\t"
435                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
436                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
437                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
438                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
439                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*2*q
440                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*2*q
441                 "paddw %%mm4, %%mm0             \n\t" // (abs(block[i])*2 + 1)*q
442                 "paddw %%mm5, %%mm1             \n\t" // (abs(block[i])*2 + 1)*q
443                 "pxor %%mm4, %%mm4              \n\t"
444                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
445                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
446                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
447                 "psrlw $4, %%mm0                \n\t"
448                 "psrlw $4, %%mm1                \n\t"
449                 "pxor %%mm2, %%mm0              \n\t"
450                 "pxor %%mm3, %%mm1              \n\t"
451                 "psubw %%mm2, %%mm0             \n\t"
452                 "psubw %%mm3, %%mm1             \n\t"
453                 "pandn %%mm0, %%mm4             \n\t"
454                 "pandn %%mm1, %%mm5             \n\t"
455                 "pxor %%mm4, %%mm7              \n\t"
456                 "pxor %%mm5, %%mm7              \n\t"
457                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
458                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
459
460                 "add $16, %%"REG_a"             \n\t"
461                 "jng 1b                         \n\t"
462                 "movd 124(%0, %3), %%mm0        \n\t"
463                 "movq %%mm7, %%mm6              \n\t"
464                 "psrlq $32, %%mm7               \n\t"
465                 "pxor %%mm6, %%mm7              \n\t"
466                 "movq %%mm7, %%mm6              \n\t"
467                 "psrlq $16, %%mm7               \n\t"
468                 "pxor %%mm6, %%mm7              \n\t"
469                 "pslld $31, %%mm7               \n\t"
470                 "psrlq $15, %%mm7               \n\t"
471                 "pxor %%mm7, %%mm0              \n\t"
472                 "movd %%mm0, 124(%0, %3)        \n\t"
473
474                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "r" (-2*nCoeffs)
475                 : "%"REG_a, "memory"
476         );
477 }
478
479 static void  denoise_dct_mmx(MpegEncContext *s, DCTELEM *block){
480     const int intra= s->mb_intra;
481     int *sum= s->dct_error_sum[intra];
482     uint16_t *offset= s->dct_offset[intra];
483
484     s->dct_count[intra]++;
485
486     __asm__ volatile(
487         "pxor %%mm7, %%mm7                      \n\t"
488         "1:                                     \n\t"
489         "pxor %%mm0, %%mm0                      \n\t"
490         "pxor %%mm1, %%mm1                      \n\t"
491         "movq (%0), %%mm2                       \n\t"
492         "movq 8(%0), %%mm3                      \n\t"
493         "pcmpgtw %%mm2, %%mm0                   \n\t"
494         "pcmpgtw %%mm3, %%mm1                   \n\t"
495         "pxor %%mm0, %%mm2                      \n\t"
496         "pxor %%mm1, %%mm3                      \n\t"
497         "psubw %%mm0, %%mm2                     \n\t"
498         "psubw %%mm1, %%mm3                     \n\t"
499         "movq %%mm2, %%mm4                      \n\t"
500         "movq %%mm3, %%mm5                      \n\t"
501         "psubusw (%2), %%mm2                    \n\t"
502         "psubusw 8(%2), %%mm3                   \n\t"
503         "pxor %%mm0, %%mm2                      \n\t"
504         "pxor %%mm1, %%mm3                      \n\t"
505         "psubw %%mm0, %%mm2                     \n\t"
506         "psubw %%mm1, %%mm3                     \n\t"
507         "movq %%mm2, (%0)                       \n\t"
508         "movq %%mm3, 8(%0)                      \n\t"
509         "movq %%mm4, %%mm2                      \n\t"
510         "movq %%mm5, %%mm3                      \n\t"
511         "punpcklwd %%mm7, %%mm4                 \n\t"
512         "punpckhwd %%mm7, %%mm2                 \n\t"
513         "punpcklwd %%mm7, %%mm5                 \n\t"
514         "punpckhwd %%mm7, %%mm3                 \n\t"
515         "paddd (%1), %%mm4                      \n\t"
516         "paddd 8(%1), %%mm2                     \n\t"
517         "paddd 16(%1), %%mm5                    \n\t"
518         "paddd 24(%1), %%mm3                    \n\t"
519         "movq %%mm4, (%1)                       \n\t"
520         "movq %%mm2, 8(%1)                      \n\t"
521         "movq %%mm5, 16(%1)                     \n\t"
522         "movq %%mm3, 24(%1)                     \n\t"
523         "add $16, %0                            \n\t"
524         "add $32, %1                            \n\t"
525         "add $16, %2                            \n\t"
526         "cmp %3, %0                             \n\t"
527             " jb 1b                             \n\t"
528         : "+r" (block), "+r" (sum), "+r" (offset)
529         : "r"(block+64)
530     );
531 }
532
533 static void  denoise_dct_sse2(MpegEncContext *s, DCTELEM *block){
534     const int intra= s->mb_intra;
535     int *sum= s->dct_error_sum[intra];
536     uint16_t *offset= s->dct_offset[intra];
537
538     s->dct_count[intra]++;
539
540     __asm__ volatile(
541         "pxor %%xmm7, %%xmm7                    \n\t"
542         "1:                                     \n\t"
543         "pxor %%xmm0, %%xmm0                    \n\t"
544         "pxor %%xmm1, %%xmm1                    \n\t"
545         "movdqa (%0), %%xmm2                    \n\t"
546         "movdqa 16(%0), %%xmm3                  \n\t"
547         "pcmpgtw %%xmm2, %%xmm0                 \n\t"
548         "pcmpgtw %%xmm3, %%xmm1                 \n\t"
549         "pxor %%xmm0, %%xmm2                    \n\t"
550         "pxor %%xmm1, %%xmm3                    \n\t"
551         "psubw %%xmm0, %%xmm2                   \n\t"
552         "psubw %%xmm1, %%xmm3                   \n\t"
553         "movdqa %%xmm2, %%xmm4                  \n\t"
554         "movdqa %%xmm3, %%xmm5                  \n\t"
555         "psubusw (%2), %%xmm2                   \n\t"
556         "psubusw 16(%2), %%xmm3                 \n\t"
557         "pxor %%xmm0, %%xmm2                    \n\t"
558         "pxor %%xmm1, %%xmm3                    \n\t"
559         "psubw %%xmm0, %%xmm2                   \n\t"
560         "psubw %%xmm1, %%xmm3                   \n\t"
561         "movdqa %%xmm2, (%0)                    \n\t"
562         "movdqa %%xmm3, 16(%0)                  \n\t"
563         "movdqa %%xmm4, %%xmm6                  \n\t"
564         "movdqa %%xmm5, %%xmm0                  \n\t"
565         "punpcklwd %%xmm7, %%xmm4               \n\t"
566         "punpckhwd %%xmm7, %%xmm6               \n\t"
567         "punpcklwd %%xmm7, %%xmm5               \n\t"
568         "punpckhwd %%xmm7, %%xmm0               \n\t"
569         "paddd (%1), %%xmm4                     \n\t"
570         "paddd 16(%1), %%xmm6                   \n\t"
571         "paddd 32(%1), %%xmm5                   \n\t"
572         "paddd 48(%1), %%xmm0                   \n\t"
573         "movdqa %%xmm4, (%1)                    \n\t"
574         "movdqa %%xmm6, 16(%1)                  \n\t"
575         "movdqa %%xmm5, 32(%1)                  \n\t"
576         "movdqa %%xmm0, 48(%1)                  \n\t"
577         "add $32, %0                            \n\t"
578         "add $64, %1                            \n\t"
579         "add $32, %2                            \n\t"
580         "cmp %3, %0                             \n\t"
581             " jb 1b                             \n\t"
582         : "+r" (block), "+r" (sum), "+r" (offset)
583         : "r"(block+64)
584           XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
585                             "%xmm4", "%xmm5", "%xmm6", "%xmm7")
586     );
587 }
588
589 #if HAVE_SSSE3
590 #define HAVE_SSSE3_BAK
591 #endif
592 #undef HAVE_SSSE3
593 #define HAVE_SSSE3 0
594
595 #undef HAVE_SSE2
596 #undef HAVE_MMX2
597 #define HAVE_SSE2 0
598 #define HAVE_MMX2 0
599 #define RENAME(a) a ## _MMX
600 #define RENAMEl(a) a ## _mmx
601 #include "mpegvideo_mmx_template.c"
602
603 #undef HAVE_MMX2
604 #define HAVE_MMX2 1
605 #undef RENAME
606 #undef RENAMEl
607 #define RENAME(a) a ## _MMX2
608 #define RENAMEl(a) a ## _mmx2
609 #include "mpegvideo_mmx_template.c"
610
611 #undef HAVE_SSE2
612 #define HAVE_SSE2 1
613 #undef RENAME
614 #undef RENAMEl
615 #define RENAME(a) a ## _SSE2
616 #define RENAMEl(a) a ## _sse2
617 #include "mpegvideo_mmx_template.c"
618
619 #ifdef HAVE_SSSE3_BAK
620 #undef HAVE_SSSE3
621 #define HAVE_SSSE3 1
622 #undef RENAME
623 #undef RENAMEl
624 #define RENAME(a) a ## _SSSE3
625 #define RENAMEl(a) a ## _sse2
626 #include "mpegvideo_mmx_template.c"
627 #endif
628
629 void ff_MPV_common_init_mmx(MpegEncContext *s)
630 {
631     int mm_flags = av_get_cpu_flags();
632
633     if (mm_flags & AV_CPU_FLAG_MMX) {
634         const int dct_algo = s->avctx->dct_algo;
635
636         s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_mmx;
637         s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_mmx;
638         s->dct_unquantize_mpeg1_intra = dct_unquantize_mpeg1_intra_mmx;
639         s->dct_unquantize_mpeg1_inter = dct_unquantize_mpeg1_inter_mmx;
640         if(!(s->flags & CODEC_FLAG_BITEXACT))
641             s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_mmx;
642         s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_mmx;
643
644         if (mm_flags & AV_CPU_FLAG_SSE2) {
645             s->denoise_dct= denoise_dct_sse2;
646         } else {
647                 s->denoise_dct= denoise_dct_mmx;
648         }
649
650         if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){
651 #if HAVE_SSSE3
652             if(mm_flags & AV_CPU_FLAG_SSSE3){
653                 s->dct_quantize= dct_quantize_SSSE3;
654             } else
655 #endif
656             if(mm_flags & AV_CPU_FLAG_SSE2){
657                 s->dct_quantize= dct_quantize_SSE2;
658             } else if(mm_flags & AV_CPU_FLAG_MMX2){
659                 s->dct_quantize= dct_quantize_MMX2;
660             } else {
661                 s->dct_quantize= dct_quantize_MMX;
662             }
663         }
664     }
665 }