1b7b1c19f44955549e311755e0b2ee2d53e30322
[ffmpeg.git] / libavcodec / i386 / mpegvideo_mmx.c
1 /*
2  * The simplest mpeg encoder (well, it was the simplest!)
3  * Copyright (c) 2000,2001 Fabrice Bellard.
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  *
21  * Optimized for ia32 cpus by Nick Kurshev <nickols_k@mail.ru>
22  * h263, mpeg1, mpeg2 dequantizer & draw_edges by Michael Niedermayer <michaelni@gmx.at>
23  */
24
25 #include "../dsputil.h"
26 #include "../mpegvideo.h"
27 #include "../avcodec.h"
28 #include "x86_cpu.h"
29
30 extern uint16_t inv_zigzag_direct16[64];
31
32 static const unsigned long long int mm_wabs __attribute__ ((aligned(8))) = 0xffffffffffffffffULL;
33 static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001ULL;
34
35
36 static void dct_unquantize_h263_intra_mmx(MpegEncContext *s,
37                                   DCTELEM *block, int n, int qscale)
38 {
39     long level, qmul, qadd, nCoeffs;
40
41     qmul = qscale << 1;
42
43     assert(s->block_last_index[n]>=0 || s->h263_aic);
44
45     if (!s->h263_aic) {
46         if (n < 4)
47             level = block[0] * s->y_dc_scale;
48         else
49             level = block[0] * s->c_dc_scale;
50         qadd = (qscale - 1) | 1;
51     }else{
52         qadd = 0;
53         level= block[0];
54     }
55     if(s->ac_pred)
56         nCoeffs=63;
57     else
58         nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
59 //printf("%d %d  ", qmul, qadd);
60 asm volatile(
61                 "movd %1, %%mm6                 \n\t" //qmul
62                 "packssdw %%mm6, %%mm6          \n\t"
63                 "packssdw %%mm6, %%mm6          \n\t"
64                 "movd %2, %%mm5                 \n\t" //qadd
65                 "pxor %%mm7, %%mm7              \n\t"
66                 "packssdw %%mm5, %%mm5          \n\t"
67                 "packssdw %%mm5, %%mm5          \n\t"
68                 "psubw %%mm5, %%mm7             \n\t"
69                 "pxor %%mm4, %%mm4              \n\t"
70                 ASMALIGN(4)
71                 "1:                             \n\t"
72                 "movq (%0, %3), %%mm0           \n\t"
73                 "movq 8(%0, %3), %%mm1          \n\t"
74
75                 "pmullw %%mm6, %%mm0            \n\t"
76                 "pmullw %%mm6, %%mm1            \n\t"
77
78                 "movq (%0, %3), %%mm2           \n\t"
79                 "movq 8(%0, %3), %%mm3          \n\t"
80
81                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
82                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
83
84                 "pxor %%mm2, %%mm0              \n\t"
85                 "pxor %%mm3, %%mm1              \n\t"
86
87                 "paddw %%mm7, %%mm0             \n\t"
88                 "paddw %%mm7, %%mm1             \n\t"
89
90                 "pxor %%mm0, %%mm2              \n\t"
91                 "pxor %%mm1, %%mm3              \n\t"
92
93                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
94                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
95
96                 "pandn %%mm2, %%mm0             \n\t"
97                 "pandn %%mm3, %%mm1             \n\t"
98
99                 "movq %%mm0, (%0, %3)           \n\t"
100                 "movq %%mm1, 8(%0, %3)          \n\t"
101
102                 "add $16, %3                    \n\t"
103                 "jng 1b                         \n\t"
104                 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(-nCoeffs))
105                 : "memory"
106         );
107         block[0]= level;
108 }
109
110
111 static void dct_unquantize_h263_inter_mmx(MpegEncContext *s,
112                                   DCTELEM *block, int n, int qscale)
113 {
114     long qmul, qadd, nCoeffs;
115
116     qmul = qscale << 1;
117     qadd = (qscale - 1) | 1;
118
119     assert(s->block_last_index[n]>=0 || s->h263_aic);
120
121     nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
122 //printf("%d %d  ", qmul, qadd);
123 asm volatile(
124                 "movd %1, %%mm6                 \n\t" //qmul
125                 "packssdw %%mm6, %%mm6          \n\t"
126                 "packssdw %%mm6, %%mm6          \n\t"
127                 "movd %2, %%mm5                 \n\t" //qadd
128                 "pxor %%mm7, %%mm7              \n\t"
129                 "packssdw %%mm5, %%mm5          \n\t"
130                 "packssdw %%mm5, %%mm5          \n\t"
131                 "psubw %%mm5, %%mm7             \n\t"
132                 "pxor %%mm4, %%mm4              \n\t"
133                 ASMALIGN(4)
134                 "1:                             \n\t"
135                 "movq (%0, %3), %%mm0           \n\t"
136                 "movq 8(%0, %3), %%mm1          \n\t"
137
138                 "pmullw %%mm6, %%mm0            \n\t"
139                 "pmullw %%mm6, %%mm1            \n\t"
140
141                 "movq (%0, %3), %%mm2           \n\t"
142                 "movq 8(%0, %3), %%mm3          \n\t"
143
144                 "pcmpgtw %%mm4, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
145                 "pcmpgtw %%mm4, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
146
147                 "pxor %%mm2, %%mm0              \n\t"
148                 "pxor %%mm3, %%mm1              \n\t"
149
150                 "paddw %%mm7, %%mm0             \n\t"
151                 "paddw %%mm7, %%mm1             \n\t"
152
153                 "pxor %%mm0, %%mm2              \n\t"
154                 "pxor %%mm1, %%mm3              \n\t"
155
156                 "pcmpeqw %%mm7, %%mm0           \n\t" // block[i] == 0 ? -1 : 0
157                 "pcmpeqw %%mm7, %%mm1           \n\t" // block[i] == 0 ? -1 : 0
158
159                 "pandn %%mm2, %%mm0             \n\t"
160                 "pandn %%mm3, %%mm1             \n\t"
161
162                 "movq %%mm0, (%0, %3)           \n\t"
163                 "movq %%mm1, 8(%0, %3)          \n\t"
164
165                 "add $16, %3                    \n\t"
166                 "jng 1b                         \n\t"
167                 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(-nCoeffs))
168                 : "memory"
169         );
170 }
171
172
173 /*
174   NK:
175   Note: looking at PARANOID:
176   "enable all paranoid tests for rounding, overflows, etc..."
177
178 #ifdef PARANOID
179                 if (level < -2048 || level > 2047)
180                     fprintf(stderr, "unquant error %d %d\n", i, level);
181 #endif
182   We can suppose that result of two multiplications can't be greate of 0xFFFF
183   i.e. is 16-bit, so we use here only PMULLW instruction and can avoid
184   a complex multiplication.
185 =====================================================
186  Full formula for multiplication of 2 integer numbers
187  which are represent as high:low words:
188  input: value1 = high1:low1
189         value2 = high2:low2
190  output: value3 = value1*value2
191  value3=high3:low3 (on overflow: modulus 2^32 wrap-around)
192  this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4
193  but this algorithm will compute only 0x66cb0ce4
194  this limited by 16-bit size of operands
195  ---------------------------------
196  tlow1 = high1*low2
197  tlow2 = high2*low1
198  tlow1 = tlow1 + tlow2
199  high3:low3 = low1*low2
200  high3 += tlow1
201 */
202 static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s,
203                                      DCTELEM *block, int n, int qscale)
204 {
205     long nCoeffs;
206     const uint16_t *quant_matrix;
207     int block0;
208
209     assert(s->block_last_index[n]>=0);
210
211     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
212
213     if (n < 4)
214         block0 = block[0] * s->y_dc_scale;
215     else
216         block0 = block[0] * s->c_dc_scale;
217     /* XXX: only mpeg1 */
218     quant_matrix = s->intra_matrix;
219 asm volatile(
220                 "pcmpeqw %%mm7, %%mm7           \n\t"
221                 "psrlw $15, %%mm7               \n\t"
222                 "movd %2, %%mm6                 \n\t"
223                 "packssdw %%mm6, %%mm6          \n\t"
224                 "packssdw %%mm6, %%mm6          \n\t"
225                 "mov %3, %%"REG_a"              \n\t"
226                 ASMALIGN(4)
227                 "1:                             \n\t"
228                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
229                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
230                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
231                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
232                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
233                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
234                 "pxor %%mm2, %%mm2              \n\t"
235                 "pxor %%mm3, %%mm3              \n\t"
236                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
237                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
238                 "pxor %%mm2, %%mm0              \n\t"
239                 "pxor %%mm3, %%mm1              \n\t"
240                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
241                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
242                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
243                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
244                 "pxor %%mm4, %%mm4              \n\t"
245                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
246                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
247                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
248                 "psraw $3, %%mm0                \n\t"
249                 "psraw $3, %%mm1                \n\t"
250                 "psubw %%mm7, %%mm0             \n\t"
251                 "psubw %%mm7, %%mm1             \n\t"
252                 "por %%mm7, %%mm0               \n\t"
253                 "por %%mm7, %%mm1               \n\t"
254                 "pxor %%mm2, %%mm0              \n\t"
255                 "pxor %%mm3, %%mm1              \n\t"
256                 "psubw %%mm2, %%mm0             \n\t"
257                 "psubw %%mm3, %%mm1             \n\t"
258                 "pandn %%mm0, %%mm4             \n\t"
259                 "pandn %%mm1, %%mm5             \n\t"
260                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
261                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
262
263                 "add $16, %%"REG_a"             \n\t"
264                 "js 1b                          \n\t"
265                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
266                 : "%"REG_a, "memory"
267         );
268     block[0]= block0;
269 }
270
271 static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s,
272                                      DCTELEM *block, int n, int qscale)
273 {
274     long nCoeffs;
275     const uint16_t *quant_matrix;
276
277     assert(s->block_last_index[n]>=0);
278
279     nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
280
281         quant_matrix = s->inter_matrix;
282 asm volatile(
283                 "pcmpeqw %%mm7, %%mm7           \n\t"
284                 "psrlw $15, %%mm7               \n\t"
285                 "movd %2, %%mm6                 \n\t"
286                 "packssdw %%mm6, %%mm6          \n\t"
287                 "packssdw %%mm6, %%mm6          \n\t"
288                 "mov %3, %%"REG_a"              \n\t"
289                 ASMALIGN(4)
290                 "1:                             \n\t"
291                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
292                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
293                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
294                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
295                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
296                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
297                 "pxor %%mm2, %%mm2              \n\t"
298                 "pxor %%mm3, %%mm3              \n\t"
299                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
300                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
301                 "pxor %%mm2, %%mm0              \n\t"
302                 "pxor %%mm3, %%mm1              \n\t"
303                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
304                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
305                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
306                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
307                 "paddw %%mm7, %%mm0             \n\t" // abs(block[i])*2 + 1
308                 "paddw %%mm7, %%mm1             \n\t" // abs(block[i])*2 + 1
309                 "pmullw %%mm4, %%mm0            \n\t" // (abs(block[i])*2 + 1)*q
310                 "pmullw %%mm5, %%mm1            \n\t" // (abs(block[i])*2 + 1)*q
311                 "pxor %%mm4, %%mm4              \n\t"
312                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
313                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
314                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
315                 "psraw $4, %%mm0                \n\t"
316                 "psraw $4, %%mm1                \n\t"
317                 "psubw %%mm7, %%mm0             \n\t"
318                 "psubw %%mm7, %%mm1             \n\t"
319                 "por %%mm7, %%mm0               \n\t"
320                 "por %%mm7, %%mm1               \n\t"
321                 "pxor %%mm2, %%mm0              \n\t"
322                 "pxor %%mm3, %%mm1              \n\t"
323                 "psubw %%mm2, %%mm0             \n\t"
324                 "psubw %%mm3, %%mm1             \n\t"
325                 "pandn %%mm0, %%mm4             \n\t"
326                 "pandn %%mm1, %%mm5             \n\t"
327                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
328                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
329
330                 "add $16, %%"REG_a"             \n\t"
331                 "js 1b                          \n\t"
332                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
333                 : "%"REG_a, "memory"
334         );
335 }
336
337 static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s,
338                                      DCTELEM *block, int n, int qscale)
339 {
340     long nCoeffs;
341     const uint16_t *quant_matrix;
342     int block0;
343
344     assert(s->block_last_index[n]>=0);
345
346     if(s->alternate_scan) nCoeffs= 63; //FIXME
347     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
348
349     if (n < 4)
350         block0 = block[0] * s->y_dc_scale;
351     else
352         block0 = block[0] * s->c_dc_scale;
353     quant_matrix = s->intra_matrix;
354 asm volatile(
355                 "pcmpeqw %%mm7, %%mm7           \n\t"
356                 "psrlw $15, %%mm7               \n\t"
357                 "movd %2, %%mm6                 \n\t"
358                 "packssdw %%mm6, %%mm6          \n\t"
359                 "packssdw %%mm6, %%mm6          \n\t"
360                 "mov %3, %%"REG_a"              \n\t"
361                 ASMALIGN(4)
362                 "1:                             \n\t"
363                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
364                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
365                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
366                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
367                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
368                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
369                 "pxor %%mm2, %%mm2              \n\t"
370                 "pxor %%mm3, %%mm3              \n\t"
371                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
372                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
373                 "pxor %%mm2, %%mm0              \n\t"
374                 "pxor %%mm3, %%mm1              \n\t"
375                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
376                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
377                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*q
378                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*q
379                 "pxor %%mm4, %%mm4              \n\t"
380                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
381                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
382                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
383                 "psraw $3, %%mm0                \n\t"
384                 "psraw $3, %%mm1                \n\t"
385                 "pxor %%mm2, %%mm0              \n\t"
386                 "pxor %%mm3, %%mm1              \n\t"
387                 "psubw %%mm2, %%mm0             \n\t"
388                 "psubw %%mm3, %%mm1             \n\t"
389                 "pandn %%mm0, %%mm4             \n\t"
390                 "pandn %%mm1, %%mm5             \n\t"
391                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
392                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
393
394                 "add $16, %%"REG_a"             \n\t"
395                 "jng 1b                         \n\t"
396                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
397                 : "%"REG_a, "memory"
398         );
399     block[0]= block0;
400         //Note, we dont do mismatch control for intra as errors cannot accumulate
401 }
402
403 static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s,
404                                      DCTELEM *block, int n, int qscale)
405 {
406     long nCoeffs;
407     const uint16_t *quant_matrix;
408
409     assert(s->block_last_index[n]>=0);
410
411     if(s->alternate_scan) nCoeffs= 63; //FIXME
412     else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
413
414         quant_matrix = s->inter_matrix;
415 asm volatile(
416                 "pcmpeqw %%mm7, %%mm7           \n\t"
417                 "psrlq $48, %%mm7               \n\t"
418                 "movd %2, %%mm6                 \n\t"
419                 "packssdw %%mm6, %%mm6          \n\t"
420                 "packssdw %%mm6, %%mm6          \n\t"
421                 "mov %3, %%"REG_a"              \n\t"
422                 ASMALIGN(4)
423                 "1:                             \n\t"
424                 "movq (%0, %%"REG_a"), %%mm0    \n\t"
425                 "movq 8(%0, %%"REG_a"), %%mm1   \n\t"
426                 "movq (%1, %%"REG_a"), %%mm4    \n\t"
427                 "movq 8(%1, %%"REG_a"), %%mm5   \n\t"
428                 "pmullw %%mm6, %%mm4            \n\t" // q=qscale*quant_matrix[i]
429                 "pmullw %%mm6, %%mm5            \n\t" // q=qscale*quant_matrix[i]
430                 "pxor %%mm2, %%mm2              \n\t"
431                 "pxor %%mm3, %%mm3              \n\t"
432                 "pcmpgtw %%mm0, %%mm2           \n\t" // block[i] < 0 ? -1 : 0
433                 "pcmpgtw %%mm1, %%mm3           \n\t" // block[i] < 0 ? -1 : 0
434                 "pxor %%mm2, %%mm0              \n\t"
435                 "pxor %%mm3, %%mm1              \n\t"
436                 "psubw %%mm2, %%mm0             \n\t" // abs(block[i])
437                 "psubw %%mm3, %%mm1             \n\t" // abs(block[i])
438                 "paddw %%mm0, %%mm0             \n\t" // abs(block[i])*2
439                 "paddw %%mm1, %%mm1             \n\t" // abs(block[i])*2
440                 "pmullw %%mm4, %%mm0            \n\t" // abs(block[i])*2*q
441                 "pmullw %%mm5, %%mm1            \n\t" // abs(block[i])*2*q
442                 "paddw %%mm4, %%mm0             \n\t" // (abs(block[i])*2 + 1)*q
443                 "paddw %%mm5, %%mm1             \n\t" // (abs(block[i])*2 + 1)*q
444                 "pxor %%mm4, %%mm4              \n\t"
445                 "pxor %%mm5, %%mm5              \n\t" // FIXME slow
446                 "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
447                 "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0
448                 "psrlw $4, %%mm0                \n\t"
449                 "psrlw $4, %%mm1                \n\t"
450                 "pxor %%mm2, %%mm0              \n\t"
451                 "pxor %%mm3, %%mm1              \n\t"
452                 "psubw %%mm2, %%mm0             \n\t"
453                 "psubw %%mm3, %%mm1             \n\t"
454                 "pandn %%mm0, %%mm4             \n\t"
455                 "pandn %%mm1, %%mm5             \n\t"
456                 "pxor %%mm4, %%mm7              \n\t"
457                 "pxor %%mm5, %%mm7              \n\t"
458                 "movq %%mm4, (%0, %%"REG_a")    \n\t"
459                 "movq %%mm5, 8(%0, %%"REG_a")   \n\t"
460
461                 "add $16, %%"REG_a"             \n\t"
462                 "jng 1b                         \n\t"
463                 "movd 124(%0, %3), %%mm0        \n\t"
464                 "movq %%mm7, %%mm6              \n\t"
465                 "psrlq $32, %%mm7               \n\t"
466                 "pxor %%mm6, %%mm7              \n\t"
467                 "movq %%mm7, %%mm6              \n\t"
468                 "psrlq $16, %%mm7               \n\t"
469                 "pxor %%mm6, %%mm7              \n\t"
470                 "pslld $31, %%mm7               \n\t"
471                 "psrlq $15, %%mm7               \n\t"
472                 "pxor %%mm7, %%mm0              \n\t"
473                 "movd %%mm0, 124(%0, %3)        \n\t"
474
475                 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "r" (-2*nCoeffs)
476                 : "%"REG_a, "memory"
477         );
478 }
479
480 /* draw the edges of width 'w' of an image of size width, height
481    this mmx version can only handle w==8 || w==16 */
482 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w)
483 {
484     uint8_t *ptr, *last_line;
485     int i;
486
487     last_line = buf + (height - 1) * wrap;
488     /* left and right */
489     ptr = buf;
490     if(w==8)
491     {
492         asm volatile(
493                 "1:                             \n\t"
494                 "movd (%0), %%mm0               \n\t"
495                 "punpcklbw %%mm0, %%mm0         \n\t"
496                 "punpcklwd %%mm0, %%mm0         \n\t"
497                 "punpckldq %%mm0, %%mm0         \n\t"
498                 "movq %%mm0, -8(%0)             \n\t"
499                 "movq -8(%0, %2), %%mm1         \n\t"
500                 "punpckhbw %%mm1, %%mm1         \n\t"
501                 "punpckhwd %%mm1, %%mm1         \n\t"
502                 "punpckhdq %%mm1, %%mm1         \n\t"
503                 "movq %%mm1, (%0, %2)           \n\t"
504                 "add %1, %0                     \n\t"
505                 "cmp %3, %0                     \n\t"
506                 " jb 1b                         \n\t"
507                 : "+r" (ptr)
508                 : "r" ((long)wrap), "r" ((long)width), "r" (ptr + wrap*height)
509         );
510     }
511     else
512     {
513         asm volatile(
514                 "1:                             \n\t"
515                 "movd (%0), %%mm0               \n\t"
516                 "punpcklbw %%mm0, %%mm0         \n\t"
517                 "punpcklwd %%mm0, %%mm0         \n\t"
518                 "punpckldq %%mm0, %%mm0         \n\t"
519                 "movq %%mm0, -8(%0)             \n\t"
520                 "movq %%mm0, -16(%0)            \n\t"
521                 "movq -8(%0, %2), %%mm1         \n\t"
522                 "punpckhbw %%mm1, %%mm1         \n\t"
523                 "punpckhwd %%mm1, %%mm1         \n\t"
524                 "punpckhdq %%mm1, %%mm1         \n\t"
525                 "movq %%mm1, (%0, %2)           \n\t"
526                 "movq %%mm1, 8(%0, %2)          \n\t"
527                 "add %1, %0                     \n\t"
528                 "cmp %3, %0                     \n\t"
529                 " jb 1b                         \n\t"
530                 : "+r" (ptr)
531                 : "r" ((long)wrap), "r" ((long)width), "r" (ptr + wrap*height)
532         );
533     }
534
535     for(i=0;i<w;i+=4) {
536         /* top and bottom (and hopefully also the corners) */
537         ptr= buf - (i + 1) * wrap - w;
538         asm volatile(
539                 "1:                             \n\t"
540                 "movq (%1, %0), %%mm0           \n\t"
541                 "movq %%mm0, (%0)               \n\t"
542                 "movq %%mm0, (%0, %2)           \n\t"
543                 "movq %%mm0, (%0, %2, 2)        \n\t"
544                 "movq %%mm0, (%0, %3)           \n\t"
545                 "add $8, %0                     \n\t"
546                 "cmp %4, %0                     \n\t"
547                 " jb 1b                         \n\t"
548                 : "+r" (ptr)
549                 : "r" ((long)buf - (long)ptr - w), "r" ((long)-wrap), "r" ((long)-wrap*3), "r" (ptr+width+2*w)
550         );
551         ptr= last_line + (i + 1) * wrap - w;
552         asm volatile(
553                 "1:                             \n\t"
554                 "movq (%1, %0), %%mm0           \n\t"
555                 "movq %%mm0, (%0)               \n\t"
556                 "movq %%mm0, (%0, %2)           \n\t"
557                 "movq %%mm0, (%0, %2, 2)        \n\t"
558                 "movq %%mm0, (%0, %3)           \n\t"
559                 "add $8, %0                     \n\t"
560                 "cmp %4, %0                     \n\t"
561                 " jb 1b                         \n\t"
562                 : "+r" (ptr)
563                 : "r" ((long)last_line - (long)ptr - w), "r" ((long)wrap), "r" ((long)wrap*3), "r" (ptr+width+2*w)
564         );
565     }
566 }
567
568 static void  denoise_dct_mmx(MpegEncContext *s, DCTELEM *block){
569     const int intra= s->mb_intra;
570     int *sum= s->dct_error_sum[intra];
571     uint16_t *offset= s->dct_offset[intra];
572
573     s->dct_count[intra]++;
574
575     asm volatile(
576         "pxor %%mm7, %%mm7                      \n\t"
577         "1:                                     \n\t"
578         "pxor %%mm0, %%mm0                      \n\t"
579         "pxor %%mm1, %%mm1                      \n\t"
580         "movq (%0), %%mm2                       \n\t"
581         "movq 8(%0), %%mm3                      \n\t"
582         "pcmpgtw %%mm2, %%mm0                   \n\t"
583         "pcmpgtw %%mm3, %%mm1                   \n\t"
584         "pxor %%mm0, %%mm2                      \n\t"
585         "pxor %%mm1, %%mm3                      \n\t"
586         "psubw %%mm0, %%mm2                     \n\t"
587         "psubw %%mm1, %%mm3                     \n\t"
588         "movq %%mm2, %%mm4                      \n\t"
589         "movq %%mm3, %%mm5                      \n\t"
590         "psubusw (%2), %%mm2                    \n\t"
591         "psubusw 8(%2), %%mm3                   \n\t"
592         "pxor %%mm0, %%mm2                      \n\t"
593         "pxor %%mm1, %%mm3                      \n\t"
594         "psubw %%mm0, %%mm2                     \n\t"
595         "psubw %%mm1, %%mm3                     \n\t"
596         "movq %%mm2, (%0)                       \n\t"
597         "movq %%mm3, 8(%0)                      \n\t"
598         "movq %%mm4, %%mm2                      \n\t"
599         "movq %%mm5, %%mm3                      \n\t"
600         "punpcklwd %%mm7, %%mm4                 \n\t"
601         "punpckhwd %%mm7, %%mm2                 \n\t"
602         "punpcklwd %%mm7, %%mm5                 \n\t"
603         "punpckhwd %%mm7, %%mm3                 \n\t"
604         "paddd (%1), %%mm4                      \n\t"
605         "paddd 8(%1), %%mm2                     \n\t"
606         "paddd 16(%1), %%mm5                    \n\t"
607         "paddd 24(%1), %%mm3                    \n\t"
608         "movq %%mm4, (%1)                       \n\t"
609         "movq %%mm2, 8(%1)                      \n\t"
610         "movq %%mm5, 16(%1)                     \n\t"
611         "movq %%mm3, 24(%1)                     \n\t"
612         "add $16, %0                            \n\t"
613         "add $32, %1                            \n\t"
614         "add $16, %2                            \n\t"
615         "cmp %3, %0                             \n\t"
616             " jb 1b                             \n\t"
617         : "+r" (block), "+r" (sum), "+r" (offset)
618         : "r"(block+64)
619     );
620 }
621
622 static void  denoise_dct_sse2(MpegEncContext *s, DCTELEM *block){
623     const int intra= s->mb_intra;
624     int *sum= s->dct_error_sum[intra];
625     uint16_t *offset= s->dct_offset[intra];
626
627     s->dct_count[intra]++;
628
629     asm volatile(
630         "pxor %%xmm7, %%xmm7                    \n\t"
631         "1:                                     \n\t"
632         "pxor %%xmm0, %%xmm0                    \n\t"
633         "pxor %%xmm1, %%xmm1                    \n\t"
634         "movdqa (%0), %%xmm2                    \n\t"
635         "movdqa 16(%0), %%xmm3                  \n\t"
636         "pcmpgtw %%xmm2, %%xmm0                 \n\t"
637         "pcmpgtw %%xmm3, %%xmm1                 \n\t"
638         "pxor %%xmm0, %%xmm2                    \n\t"
639         "pxor %%xmm1, %%xmm3                    \n\t"
640         "psubw %%xmm0, %%xmm2                   \n\t"
641         "psubw %%xmm1, %%xmm3                   \n\t"
642         "movdqa %%xmm2, %%xmm4                  \n\t"
643         "movdqa %%xmm3, %%xmm5                  \n\t"
644         "psubusw (%2), %%xmm2                   \n\t"
645         "psubusw 16(%2), %%xmm3                 \n\t"
646         "pxor %%xmm0, %%xmm2                    \n\t"
647         "pxor %%xmm1, %%xmm3                    \n\t"
648         "psubw %%xmm0, %%xmm2                   \n\t"
649         "psubw %%xmm1, %%xmm3                   \n\t"
650         "movdqa %%xmm2, (%0)                    \n\t"
651         "movdqa %%xmm3, 16(%0)                  \n\t"
652         "movdqa %%xmm4, %%xmm6                  \n\t"
653         "movdqa %%xmm5, %%xmm0                  \n\t"
654         "punpcklwd %%xmm7, %%xmm4               \n\t"
655         "punpckhwd %%xmm7, %%xmm6               \n\t"
656         "punpcklwd %%xmm7, %%xmm5               \n\t"
657         "punpckhwd %%xmm7, %%xmm0               \n\t"
658         "paddd (%1), %%xmm4                     \n\t"
659         "paddd 16(%1), %%xmm6                   \n\t"
660         "paddd 32(%1), %%xmm5                   \n\t"
661         "paddd 48(%1), %%xmm0                   \n\t"
662         "movdqa %%xmm4, (%1)                    \n\t"
663         "movdqa %%xmm6, 16(%1)                  \n\t"
664         "movdqa %%xmm5, 32(%1)                  \n\t"
665         "movdqa %%xmm0, 48(%1)                  \n\t"
666         "add $32, %0                            \n\t"
667         "add $64, %1                            \n\t"
668         "add $32, %2                            \n\t"
669         "cmp %3, %0                             \n\t"
670             " jb 1b                             \n\t"
671         : "+r" (block), "+r" (sum), "+r" (offset)
672         : "r"(block+64)
673     );
674 }
675
676 #undef HAVE_MMX2
677 #define RENAME(a) a ## _MMX
678 #define RENAMEl(a) a ## _mmx
679 #include "mpegvideo_mmx_template.c"
680
681 #define HAVE_MMX2
682 #undef RENAME
683 #undef RENAMEl
684 #define RENAME(a) a ## _MMX2
685 #define RENAMEl(a) a ## _mmx2
686 #include "mpegvideo_mmx_template.c"
687
688 #undef RENAME
689 #undef RENAMEl
690 #define RENAME(a) a ## _SSE2
691 #define RENAMEl(a) a ## _sse2
692 #include "mpegvideo_mmx_template.c"
693
694 void MPV_common_init_mmx(MpegEncContext *s)
695 {
696     if (mm_flags & MM_MMX) {
697         const int dct_algo = s->avctx->dct_algo;
698
699         s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_mmx;
700         s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_mmx;
701         s->dct_unquantize_mpeg1_intra = dct_unquantize_mpeg1_intra_mmx;
702         s->dct_unquantize_mpeg1_inter = dct_unquantize_mpeg1_inter_mmx;
703         if(!(s->flags & CODEC_FLAG_BITEXACT))
704             s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_mmx;
705         s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_mmx;
706
707         draw_edges = draw_edges_mmx;
708
709         if (mm_flags & MM_SSE2) {
710             s->denoise_dct= denoise_dct_sse2;
711         } else {
712                 s->denoise_dct= denoise_dct_mmx;
713         }
714
715         if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){
716             if(mm_flags & MM_SSE2){
717                 s->dct_quantize= dct_quantize_SSE2;
718             } else if(mm_flags & MM_MMXEXT){
719                 s->dct_quantize= dct_quantize_MMX2;
720             } else {
721                 s->dct_quantize= dct_quantize_MMX;
722             }
723         }
724     }
725 }