2 * The simplest mpeg encoder (well, it was the simplest!)
3 * Copyright (c) 2000,2001 Fabrice Bellard.
5 * This library is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU Lesser General Public
7 * License as published by the Free Software Foundation; either
8 * version 2 of the License, or (at your option) any later version.
10 * This library is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * Lesser General Public License for more details.
15 * You should have received a copy of the GNU Lesser General Public
16 * License along with this library; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19 * Optimized for ia32 cpus by Nick Kurshev <nickols_k@mail.ru>
20 * h263, mpeg1, mpeg2 dequantizer & draw_edges by Michael Niedermayer <michaelni@gmx.at>
23 #include "../dsputil.h"
24 #include "../mpegvideo.h"
25 #include "../avcodec.h"
27 extern uint8_t zigzag_direct_noperm[64];
28 extern uint16_t inv_zigzag_direct16[64];
30 static const unsigned long long int mm_wabs __attribute__ ((aligned(8))) = 0xffffffffffffffffULL;
31 static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001ULL;
34 static void dct_unquantize_h263_mmx(MpegEncContext *s,
35 DCTELEM *block, int n, int qscale)
37 int level, qmul, qadd, nCoeffs;
40 qadd = (qscale - 1) | 1;
42 assert(s->block_last_index[n]>=0);
47 level = block[0] * s->y_dc_scale;
49 level = block[0] * s->c_dc_scale;
56 nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
57 level = 0;/* keep gcc quiet */
59 //printf("%d %d ", qmul, qadd);
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"
72 "movq (%0, %3), %%mm0 \n\t"
73 "movq 8(%0, %3), %%mm1 \n\t"
75 "pmullw %%mm6, %%mm0 \n\t"
76 "pmullw %%mm6, %%mm1 \n\t"
78 "movq (%0, %3), %%mm2 \n\t"
79 "movq 8(%0, %3), %%mm3 \n\t"
81 "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
82 "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
84 "pxor %%mm2, %%mm0 \n\t"
85 "pxor %%mm3, %%mm1 \n\t"
87 "paddw %%mm7, %%mm0 \n\t"
88 "paddw %%mm7, %%mm1 \n\t"
90 "pxor %%mm0, %%mm2 \n\t"
91 "pxor %%mm1, %%mm3 \n\t"
93 "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0
94 "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0
96 "pandn %%mm2, %%mm0 \n\t"
97 "pandn %%mm3, %%mm1 \n\t"
99 "movq %%mm0, (%0, %3) \n\t"
100 "movq %%mm1, 8(%0, %3) \n\t"
104 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(-nCoeffs))
114 Note: looking at PARANOID:
115 "enable all paranoid tests for rounding, overflows, etc..."
118 if (level < -2048 || level > 2047)
119 fprintf(stderr, "unquant error %d %d\n", i, level);
121 We can suppose that result of two multiplications can't be greate of 0xFFFF
122 i.e. is 16-bit, so we use here only PMULLW instruction and can avoid
123 a complex multiplication.
124 =====================================================
125 Full formula for multiplication of 2 integer numbers
126 which are represent as high:low words:
127 input: value1 = high1:low1
129 output: value3 = value1*value2
130 value3=high3:low3 (on overflow: modulus 2^32 wrap-around)
131 this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4
132 but this algorithm will compute only 0x66cb0ce4
133 this limited by 16-bit size of operands
134 ---------------------------------
137 tlow1 = tlow1 + tlow2
138 high3:low3 = low1*low2
141 static void dct_unquantize_mpeg1_mmx(MpegEncContext *s,
142 DCTELEM *block, int n, int qscale)
145 const uint16_t *quant_matrix;
147 assert(s->block_last_index[n]>=0);
149 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1;
154 block0 = block[0] * s->y_dc_scale;
156 block0 = block[0] * s->c_dc_scale;
157 /* XXX: only mpeg1 */
158 quant_matrix = s->intra_matrix;
160 "pcmpeqw %%mm7, %%mm7 \n\t"
161 "psrlw $15, %%mm7 \n\t"
162 "movd %2, %%mm6 \n\t"
163 "packssdw %%mm6, %%mm6 \n\t"
164 "packssdw %%mm6, %%mm6 \n\t"
165 "movl %3, %%eax \n\t"
168 "movq (%0, %%eax), %%mm0 \n\t"
169 "movq 8(%0, %%eax), %%mm1 \n\t"
170 "movq (%1, %%eax), %%mm4 \n\t"
171 "movq 8(%1, %%eax), %%mm5 \n\t"
172 "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i]
173 "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i]
174 "pxor %%mm2, %%mm2 \n\t"
175 "pxor %%mm3, %%mm3 \n\t"
176 "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
177 "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
178 "pxor %%mm2, %%mm0 \n\t"
179 "pxor %%mm3, %%mm1 \n\t"
180 "psubw %%mm2, %%mm0 \n\t" // abs(block[i])
181 "psubw %%mm3, %%mm1 \n\t" // abs(block[i])
182 "pmullw %%mm4, %%mm0 \n\t" // abs(block[i])*q
183 "pmullw %%mm5, %%mm1 \n\t" // abs(block[i])*q
184 "pxor %%mm4, %%mm4 \n\t"
185 "pxor %%mm5, %%mm5 \n\t" // FIXME slow
186 "pcmpeqw (%0, %%eax), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
187 "pcmpeqw 8(%0, %%eax), %%mm5 \n\t" // block[i] == 0 ? -1 : 0
188 "psraw $3, %%mm0 \n\t"
189 "psraw $3, %%mm1 \n\t"
190 "psubw %%mm7, %%mm0 \n\t"
191 "psubw %%mm7, %%mm1 \n\t"
192 "por %%mm7, %%mm0 \n\t"
193 "por %%mm7, %%mm1 \n\t"
194 "pxor %%mm2, %%mm0 \n\t"
195 "pxor %%mm3, %%mm1 \n\t"
196 "psubw %%mm2, %%mm0 \n\t"
197 "psubw %%mm3, %%mm1 \n\t"
198 "pandn %%mm0, %%mm4 \n\t"
199 "pandn %%mm1, %%mm5 \n\t"
200 "movq %%mm4, (%0, %%eax) \n\t"
201 "movq %%mm5, 8(%0, %%eax) \n\t"
203 "addl $16, %%eax \n\t"
205 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
211 quant_matrix = s->inter_matrix;
213 "pcmpeqw %%mm7, %%mm7 \n\t"
214 "psrlw $15, %%mm7 \n\t"
215 "movd %2, %%mm6 \n\t"
216 "packssdw %%mm6, %%mm6 \n\t"
217 "packssdw %%mm6, %%mm6 \n\t"
218 "movl %3, %%eax \n\t"
221 "movq (%0, %%eax), %%mm0 \n\t"
222 "movq 8(%0, %%eax), %%mm1 \n\t"
223 "movq (%1, %%eax), %%mm4 \n\t"
224 "movq 8(%1, %%eax), %%mm5 \n\t"
225 "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i]
226 "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i]
227 "pxor %%mm2, %%mm2 \n\t"
228 "pxor %%mm3, %%mm3 \n\t"
229 "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
230 "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
231 "pxor %%mm2, %%mm0 \n\t"
232 "pxor %%mm3, %%mm1 \n\t"
233 "psubw %%mm2, %%mm0 \n\t" // abs(block[i])
234 "psubw %%mm3, %%mm1 \n\t" // abs(block[i])
235 "paddw %%mm0, %%mm0 \n\t" // abs(block[i])*2
236 "paddw %%mm1, %%mm1 \n\t" // abs(block[i])*2
237 "paddw %%mm7, %%mm0 \n\t" // abs(block[i])*2 + 1
238 "paddw %%mm7, %%mm1 \n\t" // abs(block[i])*2 + 1
239 "pmullw %%mm4, %%mm0 \n\t" // (abs(block[i])*2 + 1)*q
240 "pmullw %%mm5, %%mm1 \n\t" // (abs(block[i])*2 + 1)*q
241 "pxor %%mm4, %%mm4 \n\t"
242 "pxor %%mm5, %%mm5 \n\t" // FIXME slow
243 "pcmpeqw (%0, %%eax), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
244 "pcmpeqw 8(%0, %%eax), %%mm5 \n\t" // block[i] == 0 ? -1 : 0
245 "psraw $4, %%mm0 \n\t"
246 "psraw $4, %%mm1 \n\t"
247 "psubw %%mm7, %%mm0 \n\t"
248 "psubw %%mm7, %%mm1 \n\t"
249 "por %%mm7, %%mm0 \n\t"
250 "por %%mm7, %%mm1 \n\t"
251 "pxor %%mm2, %%mm0 \n\t"
252 "pxor %%mm3, %%mm1 \n\t"
253 "psubw %%mm2, %%mm0 \n\t"
254 "psubw %%mm3, %%mm1 \n\t"
255 "pandn %%mm0, %%mm4 \n\t"
256 "pandn %%mm1, %%mm5 \n\t"
257 "movq %%mm4, (%0, %%eax) \n\t"
258 "movq %%mm5, 8(%0, %%eax) \n\t"
260 "addl $16, %%eax \n\t"
262 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
269 static void dct_unquantize_mpeg2_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 if(s->alternate_scan) nCoeffs= 63; //FIXME
278 else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
283 block0 = block[0] * s->y_dc_scale;
285 block0 = block[0] * s->c_dc_scale;
286 quant_matrix = s->intra_matrix;
288 "pcmpeqw %%mm7, %%mm7 \n\t"
289 "psrlw $15, %%mm7 \n\t"
290 "movd %2, %%mm6 \n\t"
291 "packssdw %%mm6, %%mm6 \n\t"
292 "packssdw %%mm6, %%mm6 \n\t"
293 "movl %3, %%eax \n\t"
296 "movq (%0, %%eax), %%mm0 \n\t"
297 "movq 8(%0, %%eax), %%mm1 \n\t"
298 "movq (%1, %%eax), %%mm4 \n\t"
299 "movq 8(%1, %%eax), %%mm5 \n\t"
300 "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i]
301 "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i]
302 "pxor %%mm2, %%mm2 \n\t"
303 "pxor %%mm3, %%mm3 \n\t"
304 "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
305 "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
306 "pxor %%mm2, %%mm0 \n\t"
307 "pxor %%mm3, %%mm1 \n\t"
308 "psubw %%mm2, %%mm0 \n\t" // abs(block[i])
309 "psubw %%mm3, %%mm1 \n\t" // abs(block[i])
310 "pmullw %%mm4, %%mm0 \n\t" // abs(block[i])*q
311 "pmullw %%mm5, %%mm1 \n\t" // abs(block[i])*q
312 "pxor %%mm4, %%mm4 \n\t"
313 "pxor %%mm5, %%mm5 \n\t" // FIXME slow
314 "pcmpeqw (%0, %%eax), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
315 "pcmpeqw 8(%0, %%eax), %%mm5 \n\t" // block[i] == 0 ? -1 : 0
316 "psraw $3, %%mm0 \n\t"
317 "psraw $3, %%mm1 \n\t"
318 "pxor %%mm2, %%mm0 \n\t"
319 "pxor %%mm3, %%mm1 \n\t"
320 "psubw %%mm2, %%mm0 \n\t"
321 "psubw %%mm3, %%mm1 \n\t"
322 "pandn %%mm0, %%mm4 \n\t"
323 "pandn %%mm1, %%mm5 \n\t"
324 "movq %%mm4, (%0, %%eax) \n\t"
325 "movq %%mm5, 8(%0, %%eax) \n\t"
327 "addl $16, %%eax \n\t"
329 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "g" (-2*nCoeffs)
333 //Note, we dont do mismatch control for intra as errors cannot accumulate
336 quant_matrix = s->inter_matrix;
338 "pcmpeqw %%mm7, %%mm7 \n\t"
339 "psrlq $48, %%mm7 \n\t"
340 "movd %2, %%mm6 \n\t"
341 "packssdw %%mm6, %%mm6 \n\t"
342 "packssdw %%mm6, %%mm6 \n\t"
343 "movl %3, %%eax \n\t"
346 "movq (%0, %%eax), %%mm0 \n\t"
347 "movq 8(%0, %%eax), %%mm1 \n\t"
348 "movq (%1, %%eax), %%mm4 \n\t"
349 "movq 8(%1, %%eax), %%mm5 \n\t"
350 "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i]
351 "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i]
352 "pxor %%mm2, %%mm2 \n\t"
353 "pxor %%mm3, %%mm3 \n\t"
354 "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
355 "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
356 "pxor %%mm2, %%mm0 \n\t"
357 "pxor %%mm3, %%mm1 \n\t"
358 "psubw %%mm2, %%mm0 \n\t" // abs(block[i])
359 "psubw %%mm3, %%mm1 \n\t" // abs(block[i])
360 "paddw %%mm0, %%mm0 \n\t" // abs(block[i])*2
361 "paddw %%mm1, %%mm1 \n\t" // abs(block[i])*2
362 "pmullw %%mm4, %%mm0 \n\t" // abs(block[i])*2*q
363 "pmullw %%mm5, %%mm1 \n\t" // abs(block[i])*2*q
364 "paddw %%mm4, %%mm0 \n\t" // (abs(block[i])*2 + 1)*q
365 "paddw %%mm5, %%mm1 \n\t" // (abs(block[i])*2 + 1)*q
366 "pxor %%mm4, %%mm4 \n\t"
367 "pxor %%mm5, %%mm5 \n\t" // FIXME slow
368 "pcmpeqw (%0, %%eax), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
369 "pcmpeqw 8(%0, %%eax), %%mm5 \n\t" // block[i] == 0 ? -1 : 0
370 "psrlw $4, %%mm0 \n\t"
371 "psrlw $4, %%mm1 \n\t"
372 "pxor %%mm2, %%mm0 \n\t"
373 "pxor %%mm3, %%mm1 \n\t"
374 "psubw %%mm2, %%mm0 \n\t"
375 "psubw %%mm3, %%mm1 \n\t"
376 "pandn %%mm0, %%mm4 \n\t"
377 "pandn %%mm1, %%mm5 \n\t"
378 "pxor %%mm4, %%mm7 \n\t"
379 "pxor %%mm5, %%mm7 \n\t"
380 "movq %%mm4, (%0, %%eax) \n\t"
381 "movq %%mm5, 8(%0, %%eax) \n\t"
383 "addl $16, %%eax \n\t"
385 "movd 124(%0, %3), %%mm0 \n\t"
386 "movq %%mm7, %%mm6 \n\t"
387 "psrlq $32, %%mm7 \n\t"
388 "pxor %%mm6, %%mm7 \n\t"
389 "movq %%mm7, %%mm6 \n\t"
390 "psrlq $16, %%mm7 \n\t"
391 "pxor %%mm6, %%mm7 \n\t"
392 "pslld $31, %%mm7 \n\t"
393 "psrlq $15, %%mm7 \n\t"
394 "pxor %%mm7, %%mm0 \n\t"
395 "movd %%mm0, 124(%0, %3) \n\t"
397 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "r" (-2*nCoeffs)
403 /* draw the edges of width 'w' of an image of size width, height
404 this mmx version can only handle w==8 || w==16 */
405 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w)
407 uint8_t *ptr, *last_line;
410 last_line = buf + (height - 1) * wrap;
417 "movd (%0), %%mm0 \n\t"
418 "punpcklbw %%mm0, %%mm0 \n\t"
419 "punpcklwd %%mm0, %%mm0 \n\t"
420 "punpckldq %%mm0, %%mm0 \n\t"
421 "movq %%mm0, -8(%0) \n\t"
422 "movq -8(%0, %2), %%mm1 \n\t"
423 "punpckhbw %%mm1, %%mm1 \n\t"
424 "punpckhwd %%mm1, %%mm1 \n\t"
425 "punpckhdq %%mm1, %%mm1 \n\t"
426 "movq %%mm1, (%0, %2) \n\t"
431 : "r" (wrap), "r" (width), "r" (ptr + wrap*height)
438 "movd (%0), %%mm0 \n\t"
439 "punpcklbw %%mm0, %%mm0 \n\t"
440 "punpcklwd %%mm0, %%mm0 \n\t"
441 "punpckldq %%mm0, %%mm0 \n\t"
442 "movq %%mm0, -8(%0) \n\t"
443 "movq %%mm0, -16(%0) \n\t"
444 "movq -8(%0, %2), %%mm1 \n\t"
445 "punpckhbw %%mm1, %%mm1 \n\t"
446 "punpckhwd %%mm1, %%mm1 \n\t"
447 "punpckhdq %%mm1, %%mm1 \n\t"
448 "movq %%mm1, (%0, %2) \n\t"
449 "movq %%mm1, 8(%0, %2) \n\t"
454 : "r" (wrap), "r" (width), "r" (ptr + wrap*height)
459 /* top and bottom (and hopefully also the corners) */
460 ptr= buf - (i + 1) * wrap - w;
463 "movq (%1, %0), %%mm0 \n\t"
464 "movq %%mm0, (%0) \n\t"
465 "movq %%mm0, (%0, %2) \n\t"
466 "movq %%mm0, (%0, %2, 2) \n\t"
467 "movq %%mm0, (%0, %3) \n\t"
472 : "r" ((int)buf - (int)ptr - w), "r" (-wrap), "r" (-wrap*3), "r" (ptr+width+2*w)
474 ptr= last_line + (i + 1) * wrap - w;
477 "movq (%1, %0), %%mm0 \n\t"
478 "movq %%mm0, (%0) \n\t"
479 "movq %%mm0, (%0, %2) \n\t"
480 "movq %%mm0, (%0, %2, 2) \n\t"
481 "movq %%mm0, (%0, %3) \n\t"
486 : "r" ((int)last_line - (int)ptr - w), "r" (wrap), "r" (wrap*3), "r" (ptr+width+2*w)
492 #define RENAME(a) a ## _MMX
493 #include "mpegvideo_mmx_template.c"
497 #define RENAME(a) a ## _MMX2
498 #include "mpegvideo_mmx_template.c"
500 void MPV_common_init_mmx(MpegEncContext *s)
502 if (mm_flags & MM_MMX) {
503 const int dct_algo = s->avctx->dct_algo;
505 s->dct_unquantize_h263 = dct_unquantize_h263_mmx;
506 s->dct_unquantize_mpeg1 = dct_unquantize_mpeg1_mmx;
507 s->dct_unquantize_mpeg2 = dct_unquantize_mpeg2_mmx;
509 draw_edges = draw_edges_mmx;
511 if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){
512 if(mm_flags & MM_MMXEXT){
513 s->dct_quantize= dct_quantize_MMX2;
515 s->dct_quantize= dct_quantize_MMX;