cosmetics: Remove pointless period after copyright statement non-sentences.
[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/x86_cpu.h"
26 #include "libavcodec/avcodec.h"
27 #include "libavcodec/dsputil.h"
28 #include "libavcodec/mpegvideo.h"
29 #include "dsputil_mmx.h"
30
31 extern uint16_t inv_zigzag_direct16[64];
32
33
34 static void dct_unquantize_h263_intra_mmx(MpegEncContext *s,
35                                   DCTELEM *block, int n, int qscale)
36 {
37     x86_reg level, qmul, qadd, nCoeffs;
38
39     qmul = qscale << 1;
40
41     assert(s->block_last_index[n]>=0 || s->h263_aic);
42
43     if (!s->h263_aic) {
44         if (n < 4)
45             level = block[0] * s->y_dc_scale;
46         else
47             level = block[0] * s->c_dc_scale;
48         qadd = (qscale - 1) | 1;
49     }else{
50         qadd = 0;
51         level= block[0];
52     }
53     if(s->ac_pred)
54         nCoeffs=63;
55     else
56         nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
57 //printf("%d %d  ", qmul, qadd);
58 __asm__ volatile(
59                 "movd %1, %%mm6                 \n\t" //qmul
60                 "packssdw %%mm6, %%mm6          \n\t"
61                 "packssdw %%mm6, %%mm6          \n\t"
62                 "movd %2, %%mm5                 \n\t" //qadd
63                 "pxor %%mm7, %%mm7              \n\t"
64                 "packssdw %%mm5, %%mm5          \n\t"
65                 "packssdw %%mm5, %%mm5          \n\t"
66                 "psubw %%mm5, %%mm7             \n\t"
67                 "pxor %%mm4, %%mm4              \n\t"
68                 ASMALIGN(4)
69                 "1:                             \n\t"
70                 "movq (%0, %3), %%mm0           \n\t"
71                 "movq 8(%0, %3), %%mm1          \n\t"
72
73                 "pmullw %%mm6, %%mm0            \n\t"
74                 "pmullw %%mm6, %%mm1            \n\t"
75
76                 "movq (%0, %3), %%mm2           \n\t"
77                 "movq 8(%0, %3), %%mm3          \n\t"
78
79                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
80                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
81
82                 "pxor %%mm2, %%mm0              \n\t"
83                 "pxor %%mm3, %%mm1              \n\t"
84
85                 "paddw %%mm7, %%mm0             \n\t"
86                 "paddw %%mm7, %%mm1             \n\t"
87
88                 "pxor %%mm0, %%mm2              \n\t"
89                 "pxor %%mm1, %%mm3              \n\t"
90
91                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
92                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
93
94                 "pandn %%mm2, %%mm0             \n\t"
95                 "pandn %%mm3, %%mm1             \n\t"
96
97                 "movq %%mm0, (%0, %3)           \n\t"
98                 "movq %%mm1, 8(%0, %3)          \n\t"
99
100                 "add $16, %3                    \n\t"
101                 "jng 1b                         \n\t"
102                 ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs))
103                 : "memory"
104         );
105         block[0]= level;
106 }
107
108
109 static void dct_unquantize_h263_inter_mmx(MpegEncContext *s,
110                                   DCTELEM *block, int n, int qscale)
111 {
112     x86_reg qmul, qadd, nCoeffs;
113
114     qmul = qscale << 1;
115     qadd = (qscale - 1) | 1;
116
117     assert(s->block_last_index[n]>=0 || s->h263_aic);
118
119     nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
120 //printf("%d %d  ", qmul, qadd);
121 __asm__ volatile(
122                 "movd %1, %%mm6                 \n\t" //qmul
123                 "packssdw %%mm6, %%mm6          \n\t"
124                 "packssdw %%mm6, %%mm6          \n\t"
125                 "movd %2, %%mm5                 \n\t" //qadd
126                 "pxor %%mm7, %%mm7              \n\t"
127                 "packssdw %%mm5, %%mm5          \n\t"
128                 "packssdw %%mm5, %%mm5          \n\t"
129                 "psubw %%mm5, %%mm7             \n\t"
130                 "pxor %%mm4, %%mm4              \n\t"
131                 ASMALIGN(4)
132                 "1:                             \n\t"
133                 "movq (%0, %3), %%mm0           \n\t"
134                 "movq 8(%0, %3), %%mm1          \n\t"
135
136                 "pmullw %%mm6, %%mm0            \n\t"
137                 "pmullw %%mm6, %%mm1            \n\t"
138
139                 "movq (%0, %3), %%mm2           \n\t"
140                 "movq 8(%0, %3), %%mm3          \n\t"
141
142                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
143                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
144
145                 "pxor %%mm2, %%mm0              \n\t"
146                 "pxor %%mm3, %%mm1              \n\t"
147
148                 "paddw %%mm7, %%mm0             \n\t"
149                 "paddw %%mm7, %%mm1             \n\t"
150
151                 "pxor %%mm0, %%mm2              \n\t"
152                 "pxor %%mm1, %%mm3              \n\t"
153
154                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
155                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
156
157                 "pandn %%mm2, %%mm0             \n\t"
158                 "pandn %%mm3, %%mm1             \n\t"
159
160                 "movq %%mm0, (%0, %3)           \n\t"
161                 "movq %%mm1, 8(%0, %3)          \n\t"
162
163                 "add $16, %3                    \n\t"
164                 "jng 1b                         \n\t"
165                 ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs))
166                 : "memory"
167         );
168 }
169
170
171 /*
172   NK:
173   Note: looking at PARANOID:
174   "enable all paranoid tests for rounding, overflows, etc..."
175
176 #ifdef PARANOID
177                 if (level < -2048 || level > 2047)
178                     fprintf(stderr, "unquant error %d %d\n", i, level);
179 #endif
180   We can suppose that result of two multiplications can't be greater than 0xFFFF
181   i.e. is 16-bit, so we use here only PMULLW instruction and can avoid
182   a complex multiplication.
183 =====================================================
184  Full formula for multiplication of 2 integer numbers
185  which are represent as high:low words:
186  input: value1 = high1:low1
187         value2 = high2:low2
188  output: value3 = value1*value2
189  value3=high3:low3 (on overflow: modulus 2^32 wrap-around)
190  this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4
191  but this algorithm will compute only 0x66cb0ce4
192  this limited by 16-bit size of operands
193  ---------------------------------
194  tlow1 = high1*low2
195  tlow2 = high2*low1
196  tlow1 = tlow1 + tlow2
197  high3:low3 = low1*low2
198  high3 += tlow1
199 */
200 static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s,
201                                      DCTELEM *block, int n, int qscale)
202 {
203     x86_reg nCoeffs;
204     const uint16_t *quant_matrix;
205     int block0;
206
207     assert(s->block_last_index[n]>=0);
208
209     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
210
211     if (n < 4)
212         block0 = block[0] * s->y_dc_scale;
213     else
214         block0 = block[0] * s->c_dc_scale;
215     /* XXX: only mpeg1 */
216     quant_matrix = s->intra_matrix;
217 __asm__ volatile(
218                 "pcmpeqw %%mm7, %%mm7           \n\t"
219                 "psrlw $15, %%mm7               \n\t"
220                 "movd %2, %%mm6                 \n\t"
221                 "packssdw %%mm6, %%mm6          \n\t"
222                 "packssdw %%mm6, %%mm6          \n\t"
223                 "mov %3, %%"REG_a"              \n\t"
224                 ASMALIGN(4)
225                 "1:                             \n\t"
226                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
227                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
228                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
229                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
230                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
231                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
232                 "pxor %%mm2, %%mm2              \n\t"
233                 "pxor %%mm3, %%mm3              \n\t"
234                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
235                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
236                 "pxor %%mm2, %%mm0              \n\t"
237                 "pxor %%mm3, %%mm1              \n\t"
238                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
239                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
240                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
241                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
242                 "pxor %%mm4, %%mm4              \n\t"
243                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
244                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
245                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
246                 "psraw $3, %%mm0                \n\t"
247                 "psraw $3, %%mm1                \n\t"
248                 "psubw %%mm7, %%mm0             \n\t"
249                 "psubw %%mm7, %%mm1             \n\t"
250                 "por %%mm7, %%mm0               \n\t"
251                 "por %%mm7, %%mm1               \n\t"
252                 "pxor %%mm2, %%mm0              \n\t"
253                 "pxor %%mm3, %%mm1              \n\t"
254                 "psubw %%mm2, %%mm0             \n\t"
255                 "psubw %%mm3, %%mm1             \n\t"
256                 "pandn %%mm0, %%mm4             \n\t"
257                 "pandn %%mm1, %%mm5             \n\t"
258                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
259                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
260
261                 "add $16, %%"REG_a"             \n\t"
262                 "js 1b                          \n\t"
263                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
264                 : "%"REG_a, "memory"
265         );
266     block[0]= block0;
267 }
268
269 static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s,
270                                      DCTELEM *block, int n, int qscale)
271 {
272     x86_reg nCoeffs;
273     const uint16_t *quant_matrix;
274
275     assert(s->block_last_index[n]>=0);
276
277     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
278
279         quant_matrix = s->inter_matrix;
280 __asm__ volatile(
281                 "pcmpeqw %%mm7, %%mm7           \n\t"
282                 "psrlw $15, %%mm7               \n\t"
283                 "movd %2, %%mm6                 \n\t"
284                 "packssdw %%mm6, %%mm6          \n\t"
285                 "packssdw %%mm6, %%mm6          \n\t"
286                 "mov %3, %%"REG_a"              \n\t"
287                 ASMALIGN(4)
288                 "1:                             \n\t"
289                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
290                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
291                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
292                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
293                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
294                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
295                 "pxor %%mm2, %%mm2              \n\t"
296                 "pxor %%mm3, %%mm3              \n\t"
297                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
298                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
299                 "pxor %%mm2, %%mm0              \n\t"
300                 "pxor %%mm3, %%mm1              \n\t"
301                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
302                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
303                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
304                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
305                 "paddw %%mm7, %%mm0             \n\t" // abs(block[i])*2 + 1
306                 "paddw %%mm7, %%mm1             \n\t" // abs(block[i])*2 + 1
307                 "pmullw %%mm4, %%mm0            \n\t" // (abs(block[i])*2 + 1)*q
308                 "pmullw %%mm5, %%mm1            \n\t" // (abs(block[i])*2 + 1)*q
309                 "pxor %%mm4, %%mm4              \n\t"
310                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
311                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
312                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
313                 "psraw $4, %%mm0                \n\t"
314                 "psraw $4, %%mm1                \n\t"
315                 "psubw %%mm7, %%mm0             \n\t"
316                 "psubw %%mm7, %%mm1             \n\t"
317                 "por %%mm7, %%mm0               \n\t"
318                 "por %%mm7, %%mm1               \n\t"
319                 "pxor %%mm2, %%mm0              \n\t"
320                 "pxor %%mm3, %%mm1              \n\t"
321                 "psubw %%mm2, %%mm0             \n\t"
322                 "psubw %%mm3, %%mm1             \n\t"
323                 "pandn %%mm0, %%mm4             \n\t"
324                 "pandn %%mm1, %%mm5             \n\t"
325                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
326                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
327
328                 "add $16, %%"REG_a"             \n\t"
329                 "js 1b                          \n\t"
330                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
331                 : "%"REG_a, "memory"
332         );
333 }
334
335 static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s,
336                                      DCTELEM *block, int n, int qscale)
337 {
338     x86_reg nCoeffs;
339     const uint16_t *quant_matrix;
340     int block0;
341
342     assert(s->block_last_index[n]>=0);
343
344     if(s->alternate_scan) nCoeffs= 63; //FIXME
345     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
346
347     if (n < 4)
348         block0 = block[0] * s->y_dc_scale;
349     else
350         block0 = block[0] * s->c_dc_scale;
351     quant_matrix = s->intra_matrix;
352 __asm__ volatile(
353                 "pcmpeqw %%mm7, %%mm7           \n\t"
354                 "psrlw $15, %%mm7               \n\t"
355                 "movd %2, %%mm6                 \n\t"
356                 "packssdw %%mm6, %%mm6          \n\t"
357                 "packssdw %%mm6, %%mm6          \n\t"
358                 "mov %3, %%"REG_a"              \n\t"
359                 ASMALIGN(4)
360                 "1:                             \n\t"
361                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
362                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
363                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
364                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
365                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
366                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
367                 "pxor %%mm2, %%mm2              \n\t"
368                 "pxor %%mm3, %%mm3              \n\t"
369                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
370                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
371                 "pxor %%mm2, %%mm0              \n\t"
372                 "pxor %%mm3, %%mm1              \n\t"
373                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
374                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
375                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
376                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
377                 "pxor %%mm4, %%mm4              \n\t"
378                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
379                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
380                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
381                 "psraw $3, %%mm0                \n\t"
382                 "psraw $3, %%mm1                \n\t"
383                 "pxor %%mm2, %%mm0              \n\t"
384                 "pxor %%mm3, %%mm1              \n\t"
385                 "psubw %%mm2, %%mm0             \n\t"
386                 "psubw %%mm3, %%mm1             \n\t"
387                 "pandn %%mm0, %%mm4             \n\t"
388                 "pandn %%mm1, %%mm5             \n\t"
389                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
390                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
391
392                 "add $16, %%"REG_a"             \n\t"
393                 "jng 1b                         \n\t"
394                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
395                 : "%"REG_a, "memory"
396         );
397     block[0]= block0;
398         //Note, we do not do mismatch control for intra as errors cannot accumulate
399 }
400
401 static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s,
402                                      DCTELEM *block, int n, int qscale)
403 {
404     x86_reg nCoeffs;
405     const uint16_t *quant_matrix;
406
407     assert(s->block_last_index[n]>=0);
408
409     if(s->alternate_scan) nCoeffs= 63; //FIXME
410     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
411
412         quant_matrix = s->inter_matrix;
413 __asm__ volatile(
414                 "pcmpeqw %%mm7, %%mm7           \n\t"
415                 "psrlq $48, %%mm7               \n\t"
416                 "movd %2, %%mm6                 \n\t"
417                 "packssdw %%mm6, %%mm6          \n\t"
418                 "packssdw %%mm6, %%mm6          \n\t"
419                 "mov %3, %%"REG_a"              \n\t"
420                 ASMALIGN(4)
421                 "1:                             \n\t"
422                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
423                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
424                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
425                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
426                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
427                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
428                 "pxor %%mm2, %%mm2              \n\t"
429                 "pxor %%mm3, %%mm3              \n\t"
430                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
431                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
432                 "pxor %%mm2, %%mm0              \n\t"
433                 "pxor %%mm3, %%mm1              \n\t"
434                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
435                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
436                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
437                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
438                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*2*q
439                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*2*q
440                 "paddw %%mm4, %%mm0             \n\t" // (abs(block[i])*2 + 1)*q
441                 "paddw %%mm5, %%mm1             \n\t" // (abs(block[i])*2 + 1)*q
442                 "pxor %%mm4, %%mm4              \n\t"
443                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
444                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
445                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
446                 "psrlw $4, %%mm0                \n\t"
447                 "psrlw $4, %%mm1                \n\t"
448                 "pxor %%mm2, %%mm0              \n\t"
449                 "pxor %%mm3, %%mm1              \n\t"
450                 "psubw %%mm2, %%mm0             \n\t"
451                 "psubw %%mm3, %%mm1             \n\t"
452                 "pandn %%mm0, %%mm4             \n\t"
453                 "pandn %%mm1, %%mm5             \n\t"
454                 "pxor %%mm4, %%mm7              \n\t"
455                 "pxor %%mm5, %%mm7              \n\t"
456                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
457                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
458
459                 "add $16, %%"REG_a"             \n\t"
460                 "jng 1b                         \n\t"
461                 "movd 124(%0, %3), %%mm0        \n\t"
462                 "movq %%mm7, %%mm6              \n\t"
463                 "psrlq $32, %%mm7               \n\t"
464                 "pxor %%mm6, %%mm7              \n\t"
465                 "movq %%mm7, %%mm6              \n\t"
466                 "psrlq $16, %%mm7               \n\t"
467                 "pxor %%mm6, %%mm7              \n\t"
468                 "pslld $31, %%mm7               \n\t"
469                 "psrlq $15, %%mm7               \n\t"
470                 "pxor %%mm7, %%mm0              \n\t"
471                 "movd %%mm0, 124(%0, %3)        \n\t"
472
473                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "r" (-2*nCoeffs)
474                 : "%"REG_a, "memory"
475         );
476 }
477
478 static void  denoise_dct_mmx(MpegEncContext *s, DCTELEM *block){
479     const int intra= s->mb_intra;
480     int *sum= s->dct_error_sum[intra];
481     uint16_t *offset= s->dct_offset[intra];
482
483     s->dct_count[intra]++;
484
485     __asm__ volatile(
486         "pxor %%mm7, %%mm7                      \n\t"
487         "1:                                     \n\t"
488         "pxor %%mm0, %%mm0                      \n\t"
489         "pxor %%mm1, %%mm1                      \n\t"
490         "movq (%0), %%mm2                       \n\t"
491         "movq 8(%0), %%mm3                      \n\t"
492         "pcmpgtw %%mm2, %%mm0                   \n\t"
493         "pcmpgtw %%mm3, %%mm1                   \n\t"
494         "pxor %%mm0, %%mm2                      \n\t"
495         "pxor %%mm1, %%mm3                      \n\t"
496         "psubw %%mm0, %%mm2                     \n\t"
497         "psubw %%mm1, %%mm3                     \n\t"
498         "movq %%mm2, %%mm4                      \n\t"
499         "movq %%mm3, %%mm5                      \n\t"
500         "psubusw (%2), %%mm2                    \n\t"
501         "psubusw 8(%2), %%mm3                   \n\t"
502         "pxor %%mm0, %%mm2                      \n\t"
503         "pxor %%mm1, %%mm3                      \n\t"
504         "psubw %%mm0, %%mm2                     \n\t"
505         "psubw %%mm1, %%mm3                     \n\t"
506         "movq %%mm2, (%0)                       \n\t"
507         "movq %%mm3, 8(%0)                      \n\t"
508         "movq %%mm4, %%mm2                      \n\t"
509         "movq %%mm5, %%mm3                      \n\t"
510         "punpcklwd %%mm7, %%mm4                 \n\t"
511         "punpckhwd %%mm7, %%mm2                 \n\t"
512         "punpcklwd %%mm7, %%mm5                 \n\t"
513         "punpckhwd %%mm7, %%mm3                 \n\t"
514         "paddd (%1), %%mm4                      \n\t"
515         "paddd 8(%1), %%mm2                     \n\t"
516         "paddd 16(%1), %%mm5                    \n\t"
517         "paddd 24(%1), %%mm3                    \n\t"
518         "movq %%mm4, (%1)                       \n\t"
519         "movq %%mm2, 8(%1)                      \n\t"
520         "movq %%mm5, 16(%1)                     \n\t"
521         "movq %%mm3, 24(%1)                     \n\t"
522         "add $16, %0                            \n\t"
523         "add $32, %1                            \n\t"
524         "add $16, %2                            \n\t"
525         "cmp %3, %0                             \n\t"
526             " jb 1b                             \n\t"
527         : "+r" (block), "+r" (sum), "+r" (offset)
528         : "r"(block+64)
529     );
530 }
531
532 static void  denoise_dct_sse2(MpegEncContext *s, DCTELEM *block){
533     const int intra= s->mb_intra;
534     int *sum= s->dct_error_sum[intra];
535     uint16_t *offset= s->dct_offset[intra];
536
537     s->dct_count[intra]++;
538
539     __asm__ volatile(
540         "pxor %%xmm7, %%xmm7                    \n\t"
541         "1:                                     \n\t"
542         "pxor %%xmm0, %%xmm0                    \n\t"
543         "pxor %%xmm1, %%xmm1                    \n\t"
544         "movdqa (%0), %%xmm2                    \n\t"
545         "movdqa 16(%0), %%xmm3                  \n\t"
546         "pcmpgtw %%xmm2, %%xmm0                 \n\t"
547         "pcmpgtw %%xmm3, %%xmm1                 \n\t"
548         "pxor %%xmm0, %%xmm2                    \n\t"
549         "pxor %%xmm1, %%xmm3                    \n\t"
550         "psubw %%xmm0, %%xmm2                   \n\t"
551         "psubw %%xmm1, %%xmm3                   \n\t"
552         "movdqa %%xmm2, %%xmm4                  \n\t"
553         "movdqa %%xmm3, %%xmm5                  \n\t"
554         "psubusw (%2), %%xmm2                   \n\t"
555         "psubusw 16(%2), %%xmm3                 \n\t"
556         "pxor %%xmm0, %%xmm2                    \n\t"
557         "pxor %%xmm1, %%xmm3                    \n\t"
558         "psubw %%xmm0, %%xmm2                   \n\t"
559         "psubw %%xmm1, %%xmm3                   \n\t"
560         "movdqa %%xmm2, (%0)                    \n\t"
561         "movdqa %%xmm3, 16(%0)                  \n\t"
562         "movdqa %%xmm4, %%xmm6                  \n\t"
563         "movdqa %%xmm5, %%xmm0                  \n\t"
564         "punpcklwd %%xmm7, %%xmm4               \n\t"
565         "punpckhwd %%xmm7, %%xmm6               \n\t"
566         "punpcklwd %%xmm7, %%xmm5               \n\t"
567         "punpckhwd %%xmm7, %%xmm0               \n\t"
568         "paddd (%1), %%xmm4                     \n\t"
569         "paddd 16(%1), %%xmm6                   \n\t"
570         "paddd 32(%1), %%xmm5                   \n\t"
571         "paddd 48(%1), %%xmm0                   \n\t"
572         "movdqa %%xmm4, (%1)                    \n\t"
573         "movdqa %%xmm6, 16(%1)                  \n\t"
574         "movdqa %%xmm5, 32(%1)                  \n\t"
575         "movdqa %%xmm0, 48(%1)                  \n\t"
576         "add $32, %0                            \n\t"
577         "add $64, %1                            \n\t"
578         "add $32, %2                            \n\t"
579         "cmp %3, %0                             \n\t"
580             " jb 1b                             \n\t"
581         : "+r" (block), "+r" (sum), "+r" (offset)
582         : "r"(block+64)
583     );
584 }
585
586 #if HAVE_SSSE3
587 #define HAVE_SSSE3_BAK
588 #endif
589 #undef HAVE_SSSE3
590 #define HAVE_SSSE3 0
591
592 #undef HAVE_SSE2
593 #undef HAVE_MMX2
594 #define HAVE_SSE2 0
595 #define HAVE_MMX2 0
596 #define RENAME(a) a ## _MMX
597 #define RENAMEl(a) a ## _mmx
598 #include "mpegvideo_mmx_template.c"
599
600 #undef HAVE_MMX2
601 #define HAVE_MMX2 1
602 #undef RENAME
603 #undef RENAMEl
604 #define RENAME(a) a ## _MMX2
605 #define RENAMEl(a) a ## _mmx2
606 #include "mpegvideo_mmx_template.c"
607
608 #undef HAVE_SSE2
609 #define HAVE_SSE2 1
610 #undef RENAME
611 #undef RENAMEl
612 #define RENAME(a) a ## _SSE2
613 #define RENAMEl(a) a ## _sse2
614 #include "mpegvideo_mmx_template.c"
615
616 #ifdef HAVE_SSSE3_BAK
617 #undef HAVE_SSSE3
618 #define HAVE_SSSE3 1
619 #undef RENAME
620 #undef RENAMEl
621 #define RENAME(a) a ## _SSSE3
622 #define RENAMEl(a) a ## _sse2
623 #include "mpegvideo_mmx_template.c"
624 #endif
625
626 void MPV_common_init_mmx(MpegEncContext *s)
627 {
628     if (mm_flags & FF_MM_MMX) {
629         const int dct_algo = s->avctx->dct_algo;
630
631         s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_mmx;
632         s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_mmx;
633         s->dct_unquantize_mpeg1_intra = dct_unquantize_mpeg1_intra_mmx;
634         s->dct_unquantize_mpeg1_inter = dct_unquantize_mpeg1_inter_mmx;
635         if(!(s->flags & CODEC_FLAG_BITEXACT))
636             s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_mmx;
637         s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_mmx;
638
639         if (mm_flags & FF_MM_SSE2) {
640             s->denoise_dct= denoise_dct_sse2;
641         } else {
642                 s->denoise_dct= denoise_dct_mmx;
643         }
644
645         if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){
646 #if HAVE_SSSE3
647             if(mm_flags & FF_MM_SSSE3){
648                 s->dct_quantize= dct_quantize_SSSE3;
649             } else
650 #endif
651             if(mm_flags & FF_MM_SSE2){
652                 s->dct_quantize= dct_quantize_SSE2;
653             } else if(mm_flags & FF_MM_MMXEXT){
654                 s->dct_quantize= dct_quantize_MMX2;
655             } else {
656                 s->dct_quantize= dct_quantize_MMX;
657             }
658         }
659     }
660 }