2 * The simplest mpeg encoder (well, it was the simplest!)
3 * Copyright (c) 2000,2001 Fabrice Bellard.
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>
8 * This file is part of FFmpeg.
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.
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.
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
26 #include "dsputil_mmx.h"
27 #include "mpegvideo.h"
31 extern uint16_t inv_zigzag_direct16[64];
34 static void dct_unquantize_h263_intra_mmx(MpegEncContext *s,
35 DCTELEM *block, int n, int qscale)
37 long level, qmul, qadd, nCoeffs;
41 assert(s->block_last_index[n]>=0 || s->h263_aic);
45 level = block[0] * s->y_dc_scale;
47 level = block[0] * s->c_dc_scale;
48 qadd = (qscale - 1) | 1;
56 nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
57 //printf("%d %d ", qmul, qadd);
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"
70 "movq (%0, %3), %%mm0 \n\t"
71 "movq 8(%0, %3), %%mm1 \n\t"
73 "pmullw %%mm6, %%mm0 \n\t"
74 "pmullw %%mm6, %%mm1 \n\t"
76 "movq (%0, %3), %%mm2 \n\t"
77 "movq 8(%0, %3), %%mm3 \n\t"
79 "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
80 "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
82 "pxor %%mm2, %%mm0 \n\t"
83 "pxor %%mm3, %%mm1 \n\t"
85 "paddw %%mm7, %%mm0 \n\t"
86 "paddw %%mm7, %%mm1 \n\t"
88 "pxor %%mm0, %%mm2 \n\t"
89 "pxor %%mm1, %%mm3 \n\t"
91 "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0
92 "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0
94 "pandn %%mm2, %%mm0 \n\t"
95 "pandn %%mm3, %%mm1 \n\t"
97 "movq %%mm0, (%0, %3) \n\t"
98 "movq %%mm1, 8(%0, %3) \n\t"
102 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(-nCoeffs))
109 static void dct_unquantize_h263_inter_mmx(MpegEncContext *s,
110 DCTELEM *block, int n, int qscale)
112 long qmul, qadd, nCoeffs;
115 qadd = (qscale - 1) | 1;
117 assert(s->block_last_index[n]>=0 || s->h263_aic);
119 nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
120 //printf("%d %d ", qmul, qadd);
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"
133 "movq (%0, %3), %%mm0 \n\t"
134 "movq 8(%0, %3), %%mm1 \n\t"
136 "pmullw %%mm6, %%mm0 \n\t"
137 "pmullw %%mm6, %%mm1 \n\t"
139 "movq (%0, %3), %%mm2 \n\t"
140 "movq 8(%0, %3), %%mm3 \n\t"
142 "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
143 "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
145 "pxor %%mm2, %%mm0 \n\t"
146 "pxor %%mm3, %%mm1 \n\t"
148 "paddw %%mm7, %%mm0 \n\t"
149 "paddw %%mm7, %%mm1 \n\t"
151 "pxor %%mm0, %%mm2 \n\t"
152 "pxor %%mm1, %%mm3 \n\t"
154 "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0
155 "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0
157 "pandn %%mm2, %%mm0 \n\t"
158 "pandn %%mm3, %%mm1 \n\t"
160 "movq %%mm0, (%0, %3) \n\t"
161 "movq %%mm1, 8(%0, %3) \n\t"
165 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(-nCoeffs))
173 Note: looking at PARANOID:
174 "enable all paranoid tests for rounding, overflows, etc..."
177 if (level < -2048 || level > 2047)
178 fprintf(stderr, "unquant error %d %d\n", i, level);
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
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 ---------------------------------
196 tlow1 = tlow1 + tlow2
197 high3:low3 = low1*low2
200 static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s,
201 DCTELEM *block, int n, int qscale)
204 const uint16_t *quant_matrix;
207 assert(s->block_last_index[n]>=0);
209 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
212 block0 = block[0] * s->y_dc_scale;
214 block0 = block[0] * s->c_dc_scale;
215 /* XXX: only mpeg1 */
216 quant_matrix = s->intra_matrix;
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"
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"
261 "add $16, %%"REG_a" \n\t"
263 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
269 static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s,
270 DCTELEM *block, int n, int qscale)
273 const uint16_t *quant_matrix;
275 assert(s->block_last_index[n]>=0);
277 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
279 quant_matrix = s->inter_matrix;
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"
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"
328 "add $16, %%"REG_a" \n\t"
330 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
335 static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s,
336 DCTELEM *block, int n, int qscale)
339 const uint16_t *quant_matrix;
342 assert(s->block_last_index[n]>=0);
344 if(s->alternate_scan) nCoeffs= 63; //FIXME
345 else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
348 block0 = block[0] * s->y_dc_scale;
350 block0 = block[0] * s->c_dc_scale;
351 quant_matrix = s->intra_matrix;
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"
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"
392 "add $16, %%"REG_a" \n\t"
394 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
398 //Note, we do not do mismatch control for intra as errors cannot accumulate
401 static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s,
402 DCTELEM *block, int n, int qscale)
405 const uint16_t *quant_matrix;
407 assert(s->block_last_index[n]>=0);
409 if(s->alternate_scan) nCoeffs= 63; //FIXME
410 else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
412 quant_matrix = s->inter_matrix;
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"
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"
459 "add $16, %%"REG_a" \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"
473 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "r" (-2*nCoeffs)
478 /* draw the edges of width 'w' of an image of size width, height
479 this mmx version can only handle w==8 || w==16 */
480 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w)
482 uint8_t *ptr, *last_line;
485 last_line = buf + (height - 1) * wrap;
492 "movd (%0), %%mm0 \n\t"
493 "punpcklbw %%mm0, %%mm0 \n\t"
494 "punpcklwd %%mm0, %%mm0 \n\t"
495 "punpckldq %%mm0, %%mm0 \n\t"
496 "movq %%mm0, -8(%0) \n\t"
497 "movq -8(%0, %2), %%mm1 \n\t"
498 "punpckhbw %%mm1, %%mm1 \n\t"
499 "punpckhwd %%mm1, %%mm1 \n\t"
500 "punpckhdq %%mm1, %%mm1 \n\t"
501 "movq %%mm1, (%0, %2) \n\t"
506 : "r" ((long)wrap), "r" ((long)width), "r" (ptr + wrap*height)
513 "movd (%0), %%mm0 \n\t"
514 "punpcklbw %%mm0, %%mm0 \n\t"
515 "punpcklwd %%mm0, %%mm0 \n\t"
516 "punpckldq %%mm0, %%mm0 \n\t"
517 "movq %%mm0, -8(%0) \n\t"
518 "movq %%mm0, -16(%0) \n\t"
519 "movq -8(%0, %2), %%mm1 \n\t"
520 "punpckhbw %%mm1, %%mm1 \n\t"
521 "punpckhwd %%mm1, %%mm1 \n\t"
522 "punpckhdq %%mm1, %%mm1 \n\t"
523 "movq %%mm1, (%0, %2) \n\t"
524 "movq %%mm1, 8(%0, %2) \n\t"
529 : "r" ((long)wrap), "r" ((long)width), "r" (ptr + wrap*height)
534 /* top and bottom (and hopefully also the corners) */
535 ptr= buf - (i + 1) * wrap - w;
538 "movq (%1, %0), %%mm0 \n\t"
539 "movq %%mm0, (%0) \n\t"
540 "movq %%mm0, (%0, %2) \n\t"
541 "movq %%mm0, (%0, %2, 2) \n\t"
542 "movq %%mm0, (%0, %3) \n\t"
547 : "r" ((long)buf - (long)ptr - w), "r" ((long)-wrap), "r" ((long)-wrap*3), "r" (ptr+width+2*w)
549 ptr= last_line + (i + 1) * wrap - w;
552 "movq (%1, %0), %%mm0 \n\t"
553 "movq %%mm0, (%0) \n\t"
554 "movq %%mm0, (%0, %2) \n\t"
555 "movq %%mm0, (%0, %2, 2) \n\t"
556 "movq %%mm0, (%0, %3) \n\t"
561 : "r" ((long)last_line - (long)ptr - w), "r" ((long)wrap), "r" ((long)wrap*3), "r" (ptr+width+2*w)
566 static void denoise_dct_mmx(MpegEncContext *s, DCTELEM *block){
567 const int intra= s->mb_intra;
568 int *sum= s->dct_error_sum[intra];
569 uint16_t *offset= s->dct_offset[intra];
571 s->dct_count[intra]++;
574 "pxor %%mm7, %%mm7 \n\t"
576 "pxor %%mm0, %%mm0 \n\t"
577 "pxor %%mm1, %%mm1 \n\t"
578 "movq (%0), %%mm2 \n\t"
579 "movq 8(%0), %%mm3 \n\t"
580 "pcmpgtw %%mm2, %%mm0 \n\t"
581 "pcmpgtw %%mm3, %%mm1 \n\t"
582 "pxor %%mm0, %%mm2 \n\t"
583 "pxor %%mm1, %%mm3 \n\t"
584 "psubw %%mm0, %%mm2 \n\t"
585 "psubw %%mm1, %%mm3 \n\t"
586 "movq %%mm2, %%mm4 \n\t"
587 "movq %%mm3, %%mm5 \n\t"
588 "psubusw (%2), %%mm2 \n\t"
589 "psubusw 8(%2), %%mm3 \n\t"
590 "pxor %%mm0, %%mm2 \n\t"
591 "pxor %%mm1, %%mm3 \n\t"
592 "psubw %%mm0, %%mm2 \n\t"
593 "psubw %%mm1, %%mm3 \n\t"
594 "movq %%mm2, (%0) \n\t"
595 "movq %%mm3, 8(%0) \n\t"
596 "movq %%mm4, %%mm2 \n\t"
597 "movq %%mm5, %%mm3 \n\t"
598 "punpcklwd %%mm7, %%mm4 \n\t"
599 "punpckhwd %%mm7, %%mm2 \n\t"
600 "punpcklwd %%mm7, %%mm5 \n\t"
601 "punpckhwd %%mm7, %%mm3 \n\t"
602 "paddd (%1), %%mm4 \n\t"
603 "paddd 8(%1), %%mm2 \n\t"
604 "paddd 16(%1), %%mm5 \n\t"
605 "paddd 24(%1), %%mm3 \n\t"
606 "movq %%mm4, (%1) \n\t"
607 "movq %%mm2, 8(%1) \n\t"
608 "movq %%mm5, 16(%1) \n\t"
609 "movq %%mm3, 24(%1) \n\t"
615 : "+r" (block), "+r" (sum), "+r" (offset)
620 static void denoise_dct_sse2(MpegEncContext *s, DCTELEM *block){
621 const int intra= s->mb_intra;
622 int *sum= s->dct_error_sum[intra];
623 uint16_t *offset= s->dct_offset[intra];
625 s->dct_count[intra]++;
628 "pxor %%xmm7, %%xmm7 \n\t"
630 "pxor %%xmm0, %%xmm0 \n\t"
631 "pxor %%xmm1, %%xmm1 \n\t"
632 "movdqa (%0), %%xmm2 \n\t"
633 "movdqa 16(%0), %%xmm3 \n\t"
634 "pcmpgtw %%xmm2, %%xmm0 \n\t"
635 "pcmpgtw %%xmm3, %%xmm1 \n\t"
636 "pxor %%xmm0, %%xmm2 \n\t"
637 "pxor %%xmm1, %%xmm3 \n\t"
638 "psubw %%xmm0, %%xmm2 \n\t"
639 "psubw %%xmm1, %%xmm3 \n\t"
640 "movdqa %%xmm2, %%xmm4 \n\t"
641 "movdqa %%xmm3, %%xmm5 \n\t"
642 "psubusw (%2), %%xmm2 \n\t"
643 "psubusw 16(%2), %%xmm3 \n\t"
644 "pxor %%xmm0, %%xmm2 \n\t"
645 "pxor %%xmm1, %%xmm3 \n\t"
646 "psubw %%xmm0, %%xmm2 \n\t"
647 "psubw %%xmm1, %%xmm3 \n\t"
648 "movdqa %%xmm2, (%0) \n\t"
649 "movdqa %%xmm3, 16(%0) \n\t"
650 "movdqa %%xmm4, %%xmm6 \n\t"
651 "movdqa %%xmm5, %%xmm0 \n\t"
652 "punpcklwd %%xmm7, %%xmm4 \n\t"
653 "punpckhwd %%xmm7, %%xmm6 \n\t"
654 "punpcklwd %%xmm7, %%xmm5 \n\t"
655 "punpckhwd %%xmm7, %%xmm0 \n\t"
656 "paddd (%1), %%xmm4 \n\t"
657 "paddd 16(%1), %%xmm6 \n\t"
658 "paddd 32(%1), %%xmm5 \n\t"
659 "paddd 48(%1), %%xmm0 \n\t"
660 "movdqa %%xmm4, (%1) \n\t"
661 "movdqa %%xmm6, 16(%1) \n\t"
662 "movdqa %%xmm5, 32(%1) \n\t"
663 "movdqa %%xmm0, 48(%1) \n\t"
669 : "+r" (block), "+r" (sum), "+r" (offset)
675 #define HAVE_SSSE3_BAK
681 #define RENAME(a) a ## _MMX
682 #define RENAMEl(a) a ## _mmx
683 #include "mpegvideo_mmx_template.c"
688 #define RENAME(a) a ## _MMX2
689 #define RENAMEl(a) a ## _mmx2
690 #include "mpegvideo_mmx_template.c"
695 #define RENAME(a) a ## _SSE2
696 #define RENAMEl(a) a ## _sse2
697 #include "mpegvideo_mmx_template.c"
699 #ifdef HAVE_SSSE3_BAK
703 #define RENAME(a) a ## _SSSE3
704 #define RENAMEl(a) a ## _sse2
705 #include "mpegvideo_mmx_template.c"
708 void MPV_common_init_mmx(MpegEncContext *s)
710 if (mm_flags & MM_MMX) {
711 const int dct_algo = s->avctx->dct_algo;
713 s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_mmx;
714 s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_mmx;
715 s->dct_unquantize_mpeg1_intra = dct_unquantize_mpeg1_intra_mmx;
716 s->dct_unquantize_mpeg1_inter = dct_unquantize_mpeg1_inter_mmx;
717 if(!(s->flags & CODEC_FLAG_BITEXACT))
718 s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_mmx;
719 s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_mmx;
721 draw_edges = draw_edges_mmx;
723 if (mm_flags & MM_SSE2) {
724 s->denoise_dct= denoise_dct_sse2;
726 s->denoise_dct= denoise_dct_mmx;
729 if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){
731 if(mm_flags & MM_SSSE3){
732 s->dct_quantize= dct_quantize_SSSE3;
735 if(mm_flags & MM_SSE2){
736 s->dct_quantize= dct_quantize_SSE2;
737 } else if(mm_flags & MM_MMXEXT){
738 s->dct_quantize= dct_quantize_MMX2;
740 s->dct_quantize= dct_quantize_MMX;