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
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"
32 extern uint16_t inv_zigzag_direct16[64];
35 static void dct_unquantize_h263_intra_mmx(MpegEncContext *s,
36 DCTELEM *block, int n, int qscale)
38 x86_reg level, qmul, qadd, nCoeffs;
42 assert(s->block_last_index[n]>=0 || s->h263_aic);
46 level = block[0] * s->y_dc_scale;
48 level = block[0] * s->c_dc_scale;
49 qadd = (qscale - 1) | 1;
57 nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
58 //printf("%d %d ", qmul, qadd);
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"
71 "movq (%0, %3), %%mm0 \n\t"
72 "movq 8(%0, %3), %%mm1 \n\t"
74 "pmullw %%mm6, %%mm0 \n\t"
75 "pmullw %%mm6, %%mm1 \n\t"
77 "movq (%0, %3), %%mm2 \n\t"
78 "movq 8(%0, %3), %%mm3 \n\t"
80 "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
81 "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
83 "pxor %%mm2, %%mm0 \n\t"
84 "pxor %%mm3, %%mm1 \n\t"
86 "paddw %%mm7, %%mm0 \n\t"
87 "paddw %%mm7, %%mm1 \n\t"
89 "pxor %%mm0, %%mm2 \n\t"
90 "pxor %%mm1, %%mm3 \n\t"
92 "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0
93 "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0
95 "pandn %%mm2, %%mm0 \n\t"
96 "pandn %%mm3, %%mm1 \n\t"
98 "movq %%mm0, (%0, %3) \n\t"
99 "movq %%mm1, 8(%0, %3) \n\t"
103 ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs))
110 static void dct_unquantize_h263_inter_mmx(MpegEncContext *s,
111 DCTELEM *block, int n, int qscale)
113 x86_reg qmul, qadd, nCoeffs;
116 qadd = (qscale - 1) | 1;
118 assert(s->block_last_index[n]>=0 || s->h263_aic);
120 nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
121 //printf("%d %d ", qmul, qadd);
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"
134 "movq (%0, %3), %%mm0 \n\t"
135 "movq 8(%0, %3), %%mm1 \n\t"
137 "pmullw %%mm6, %%mm0 \n\t"
138 "pmullw %%mm6, %%mm1 \n\t"
140 "movq (%0, %3), %%mm2 \n\t"
141 "movq 8(%0, %3), %%mm3 \n\t"
143 "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
144 "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
146 "pxor %%mm2, %%mm0 \n\t"
147 "pxor %%mm3, %%mm1 \n\t"
149 "paddw %%mm7, %%mm0 \n\t"
150 "paddw %%mm7, %%mm1 \n\t"
152 "pxor %%mm0, %%mm2 \n\t"
153 "pxor %%mm1, %%mm3 \n\t"
155 "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0
156 "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0
158 "pandn %%mm2, %%mm0 \n\t"
159 "pandn %%mm3, %%mm1 \n\t"
161 "movq %%mm0, (%0, %3) \n\t"
162 "movq %%mm1, 8(%0, %3) \n\t"
166 ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs))
174 Note: looking at PARANOID:
175 "enable all paranoid tests for rounding, overflows, etc..."
178 if (level < -2048 || level > 2047)
179 fprintf(stderr, "unquant error %d %d\n", i, level);
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
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 ---------------------------------
197 tlow1 = tlow1 + tlow2
198 high3:low3 = low1*low2
201 static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s,
202 DCTELEM *block, int n, int qscale)
205 const uint16_t *quant_matrix;
208 assert(s->block_last_index[n]>=0);
210 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
213 block0 = block[0] * s->y_dc_scale;
215 block0 = block[0] * s->c_dc_scale;
216 /* XXX: only mpeg1 */
217 quant_matrix = s->intra_matrix;
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"
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"
262 "add $16, %%"REG_a" \n\t"
264 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
270 static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s,
271 DCTELEM *block, int n, int qscale)
274 const uint16_t *quant_matrix;
276 assert(s->block_last_index[n]>=0);
278 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
280 quant_matrix = s->inter_matrix;
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"
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"
329 "add $16, %%"REG_a" \n\t"
331 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
336 static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s,
337 DCTELEM *block, int n, int qscale)
340 const uint16_t *quant_matrix;
343 assert(s->block_last_index[n]>=0);
345 if(s->alternate_scan) nCoeffs= 63; //FIXME
346 else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
349 block0 = block[0] * s->y_dc_scale;
351 block0 = block[0] * s->c_dc_scale;
352 quant_matrix = s->intra_matrix;
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"
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"
393 "add $16, %%"REG_a" \n\t"
395 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs)
399 //Note, we do not do mismatch control for intra as errors cannot accumulate
402 static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s,
403 DCTELEM *block, int n, int qscale)
406 const uint16_t *quant_matrix;
408 assert(s->block_last_index[n]>=0);
410 if(s->alternate_scan) nCoeffs= 63; //FIXME
411 else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
413 quant_matrix = s->inter_matrix;
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"
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"
460 "add $16, %%"REG_a" \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"
474 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "r" (-2*nCoeffs)
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];
484 s->dct_count[intra]++;
487 "pxor %%mm7, %%mm7 \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"
528 : "+r" (block), "+r" (sum), "+r" (offset)
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];
538 s->dct_count[intra]++;
541 "pxor %%xmm7, %%xmm7 \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"
582 : "+r" (block), "+r" (sum), "+r" (offset)
584 XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
585 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
590 #define HAVE_SSSE3_BAK
599 #define RENAME(a) a ## _MMX
600 #define RENAMEl(a) a ## _mmx
601 #include "mpegvideo_mmx_template.c"
607 #define RENAME(a) a ## _MMX2
608 #define RENAMEl(a) a ## _mmx2
609 #include "mpegvideo_mmx_template.c"
615 #define RENAME(a) a ## _SSE2
616 #define RENAMEl(a) a ## _sse2
617 #include "mpegvideo_mmx_template.c"
619 #ifdef HAVE_SSSE3_BAK
624 #define RENAME(a) a ## _SSSE3
625 #define RENAMEl(a) a ## _sse2
626 #include "mpegvideo_mmx_template.c"
629 void MPV_common_init_mmx(MpegEncContext *s)
631 int mm_flags = av_get_cpu_flags();
633 if (mm_flags & AV_CPU_FLAG_MMX) {
634 const int dct_algo = s->avctx->dct_algo;
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;
644 if (mm_flags & AV_CPU_FLAG_SSE2) {
645 s->denoise_dct= denoise_dct_sse2;
647 s->denoise_dct= denoise_dct_mmx;
650 if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){
652 if(mm_flags & AV_CPU_FLAG_SSSE3){
653 s->dct_quantize= dct_quantize_SSSE3;
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;
661 s->dct_quantize= dct_quantize_MMX;