Mercurial > libavcodec.hg
comparison i386/vp3dsp_mmx.c @ 2696:9699d325049d libavcodec
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
author | michael |
---|---|
date | Tue, 17 May 2005 18:28:40 +0000 |
parents | 89422281f6f6 |
children | ba8ecddf5598 |
comparison
equal
deleted
inserted
replaced
2695:5d4a6edbc99c | 2696:9699d325049d |
---|---|
255 punpckldq_r2r(r5, r2); /* r2 = d2 c2 b2 a2 = R2 */ \ | 255 punpckldq_r2r(r5, r2); /* r2 = d2 c2 b2 a2 = R2 */ \ |
256 movq_r2m(r4, *I(3)); \ | 256 movq_r2m(r4, *I(3)); \ |
257 movq_r2m(r2, *I(2)); \ | 257 movq_r2m(r2, *I(2)); \ |
258 } | 258 } |
259 | 259 |
260 void vp3_dsp_init_mmx(void) | 260 void ff_vp3_dsp_init_mmx(void) |
261 { | 261 { |
262 int j = 16; | 262 int j = 16; |
263 uint16_t *p; | 263 uint16_t *p; |
264 | |
265 do { | |
266 idct_constants[--j] = 0; | |
267 } while (j); | |
268 | |
269 idct_constants[0] = idct_constants[5] = | |
270 idct_constants[10] = idct_constants[15] = 65535; | |
271 | 264 |
272 j = 1; | 265 j = 1; |
273 do { | 266 do { |
274 p = idct_constants + ((j + 3) << 2); | 267 p = idct_constants + ((j + 3) << 2); |
275 p[0] = p[1] = p[2] = p[3] = idct_cosine_table[j - 1]; | 268 p[0] = p[1] = p[2] = p[3] = idct_cosine_table[j - 1]; |
277 | 270 |
278 idct_constants[44] = idct_constants[45] = | 271 idct_constants[44] = idct_constants[45] = |
279 idct_constants[46] = idct_constants[47] = IdctAdjustBeforeShift; | 272 idct_constants[46] = idct_constants[47] = IdctAdjustBeforeShift; |
280 } | 273 } |
281 | 274 |
282 void vp3_idct_mmx(int16_t *input_data, int16_t *dequant_matrix, | 275 void ff_vp3_idct_mmx(int16_t *output_data) |
283 int coeff_count, int16_t *output_data) | |
284 { | 276 { |
285 /* eax = quantized input | 277 /* eax = quantized input |
286 * ebx = dequantizer matrix | 278 * ebx = dequantizer matrix |
287 * ecx = IDCT constants | 279 * ecx = IDCT constants |
288 * M(I) = ecx + MaskOffset(0) + I * 8 | 280 * M(I) = ecx + MaskOffset(0) + I * 8 |
289 * C(I) = ecx + CosineOffset(32) + (I-1) * 8 | 281 * C(I) = ecx + CosineOffset(32) + (I-1) * 8 |
290 * edx = output | 282 * edx = output |
291 * r0..r7 = mm0..mm7 | 283 * r0..r7 = mm0..mm7 |
292 */ | 284 */ |
293 | 285 |
294 #define M(x) (idct_constants + x * 4) | |
295 #define C(x) (idct_constants + 16 + (x - 1) * 4) | 286 #define C(x) (idct_constants + 16 + (x - 1) * 4) |
296 #define Eight (idct_constants + 44) | 287 #define Eight (idct_constants + 44) |
297 | 288 |
298 unsigned char *input_bytes = (unsigned char *)input_data; | |
299 unsigned char *dequant_matrix_bytes = (unsigned char *)dequant_matrix; | |
300 unsigned char *output_data_bytes = (unsigned char *)output_data; | |
301 | |
302 movq_m2r(*(input_bytes), r0); | |
303 pmullw_m2r(*(dequant_matrix_bytes), r0); /* r0 = 03 02 01 00 */ | |
304 movq_m2r(*(input_bytes+16), r1); | |
305 pmullw_m2r(*(dequant_matrix_bytes+16), r1); /* r1 = 13 12 11 10 */ | |
306 movq_m2r(*M(0), r2); /* r2 = __ __ __ FF */ | |
307 movq_r2r(r0, r3); /* r3 = 03 02 01 00 */ | |
308 movq_m2r(*(input_bytes+8), r4); | |
309 psrlq_i2r(16, r0); /* r0 = __ 03 02 01 */ | |
310 pmullw_m2r(*(dequant_matrix_bytes+8), r4); /* r4 = 07 06 05 04 */ | |
311 pand_r2r(r2, r3); /* r3 = __ __ __ 00 */ | |
312 movq_r2r(r0, r5); /* r5 = __ 03 02 01 */ | |
313 movq_r2r(r1, r6); /* r6 = 13 12 11 10 */ | |
314 pand_r2r(r2, r5); /* r5 = __ __ __ 01 */ | |
315 psllq_i2r(32, r6); /* r6 = 11 10 __ __ */ | |
316 movq_m2r(*M(3), r7); /* r7 = FF __ __ __ */ | |
317 pxor_r2r(r5, r0); /* r0 = __ 03 02 __ */ | |
318 pand_r2r(r6, r7); /* r7 = 11 __ __ __ */ | |
319 por_r2r(r3, r0); /* r0 = __ 03 02 00 */ | |
320 pxor_r2r(r7, r6); /* r6 = __ 10 __ __ */ | |
321 por_r2r(r7, r0); /* r0 = 11 03 02 00 = R0 */ | |
322 movq_m2r(*M(3), r7); /* r7 = FF __ __ __ */ | |
323 movq_r2r(r4, r3); /* r3 = 07 06 05 04 */ | |
324 movq_r2m(r0, *(output_data_bytes)); /* write R0 = r0 */ | |
325 pand_r2r(r2, r3); /* r3 = __ __ __ 04 */ | |
326 movq_m2r(*(input_bytes+32), r0); | |
327 psllq_i2r(16, r3); /* r3 = __ __ 04 __ */ | |
328 pmullw_m2r(*(dequant_matrix_bytes+32), r0); /* r0 = 23 22 21 20 */ | |
329 pand_r2r(r1, r7); /* r7 = 13 __ __ __ */ | |
330 por_r2r(r3, r5); /* r5 = __ __ 04 01 */ | |
331 por_r2r(r6, r7); /* r7 = 13 10 __ __ */ | |
332 movq_m2r(*(input_bytes+24), r3); | |
333 por_r2r(r5, r7); /* r7 = 13 10 04 01 = R1 */ | |
334 pmullw_m2r(*(dequant_matrix_bytes+24), r3); /* r3 = 17 16 15 14 */ | |
335 psrlq_i2r(16, r4); /* r4 = __ 07 06 05 */ | |
336 movq_r2m(r7, *(output_data_bytes+16)); /* write R1 = r7 */ | |
337 movq_r2r(r4, r5); /* r5 = __ 07 06 05 */ | |
338 movq_r2r(r0, r7); /* r7 = 23 22 21 20 */ | |
339 psrlq_i2r(16, r4); /* r4 = __ __ 07 06 */ | |
340 psrlq_i2r(48, r7); /* r7 = __ __ __ 23 */ | |
341 movq_r2r(r2, r6); /* r6 = __ __ __ FF */ | |
342 pand_r2r(r2, r5); /* r5 = __ __ __ 05 */ | |
343 pand_r2r(r4, r6); /* r6 = __ __ __ 06 */ | |
344 movq_r2m(r7, *(output_data_bytes+80)); /* partial R9 = __ __ __ 23 */ | |
345 pxor_r2r(r6, r4); /* r4 = __ __ 07 __ */ | |
346 psrlq_i2r(32, r1); /* r1 = __ __ 13 12 */ | |
347 por_r2r(r5, r4); /* r4 = __ __ 07 05 */ | |
348 movq_m2r(*M(3), r7); /* r7 = FF __ __ __ */ | |
349 pand_r2r(r2, r1); /* r1 = __ __ __ 12 */ | |
350 movq_m2r(*(input_bytes+48), r5); | |
351 psllq_i2r(16, r0); /* r0 = 22 21 20 __ */ | |
352 pmullw_m2r(*(dequant_matrix_bytes+48), r5); /* r5 = 33 32 31 30 */ | |
353 pand_r2r(r0, r7); /* r7 = 22 __ __ __ */ | |
354 movq_r2m(r1, *(output_data_bytes+64)); /* partial R8 = __ __ __ 12 */ | |
355 por_r2r(r4, r7); /* r7 = 22 __ 07 05 */ | |
356 movq_r2r(r3, r4); /* r4 = 17 16 15 14 */ | |
357 pand_r2r(r2, r3); /* r3 = __ __ __ 14 */ | |
358 movq_m2r(*M(2), r1); /* r1 = __ FF __ __ */ | |
359 psllq_i2r(32, r3); /* r3 = __ 14 __ __ */ | |
360 por_r2r(r3, r7); /* r7 = 22 14 07 05 = R2 */ | |
361 movq_r2r(r5, r3); /* r3 = 33 32 31 30 */ | |
362 psllq_i2r(48, r3); /* r3 = 30 __ __ __ */ | |
363 pand_r2r(r0, r1); /* r1 = __ 21 __ __ */ | |
364 movq_r2m(r7, *(output_data_bytes+32)); /* write R2 = r7 */ | |
365 por_r2r(r3, r6); /* r6 = 30 __ __ 06 */ | |
366 movq_m2r(*M(1), r7); /* r7 = __ __ FF __ */ | |
367 por_r2r(r1, r6); /* r6 = 30 21 __ 06 */ | |
368 movq_m2r(*(input_bytes+56), r1); | |
369 pand_r2r(r4, r7); /* r7 = __ __ 15 __ */ | |
370 pmullw_m2r(*(dequant_matrix_bytes+56), r1); /* r1 = 37 36 35 34 */ | |
371 por_r2r(r6, r7); /* r7 = 30 21 15 06 = R3 */ | |
372 pand_m2r(*M(1), r0); /* r0 = __ __ 20 __ */ | |
373 psrlq_i2r(32, r4); /* r4 = __ __ 17 16 */ | |
374 movq_r2m(r7, *(output_data_bytes+48)); /* write R3 = r7 */ | |
375 movq_r2r(r4, r6); /* r6 = __ __ 17 16 */ | |
376 movq_m2r(*M(3), r7); /* r7 = FF __ __ __ */ | |
377 pand_r2r(r2, r4); /* r4 = __ __ __ 16 */ | |
378 movq_m2r(*M(1), r3); /* r3 = __ __ FF __ */ | |
379 pand_r2r(r1, r7); /* r7 = 37 __ __ __ */ | |
380 pand_r2r(r5, r3); /* r3 = __ __ 31 __ */ | |
381 por_r2r(r4, r0); /* r0 = __ __ 20 16 */ | |
382 psllq_i2r(16, r3); /* r3 = __ 31 __ __ */ | |
383 por_r2r(r0, r7); /* r7 = 37 __ 20 16 */ | |
384 movq_m2r(*M(2), r4); /* r4 = __ FF __ __ */ | |
385 por_r2r(r3, r7); /* r7 = 37 31 20 16 = R4 */ | |
386 movq_m2r(*(input_bytes+80), r0); | |
387 movq_r2r(r4, r3); /* r3 = __ __ FF __ */ | |
388 pmullw_m2r(*(dequant_matrix_bytes+80), r0); /* r0 = 53 52 51 50 */ | |
389 pand_r2r(r5, r4); /* r4 = __ 32 __ __ */ | |
390 movq_r2m(r7, *(output_data_bytes+8)); /* write R4 = r7 */ | |
391 por_r2r(r4, r6); /* r6 = __ 32 17 16 */ | |
392 movq_r2r(r3, r4); /* r4 = __ FF __ __ */ | |
393 psrlq_i2r(16, r6); /* r6 = __ __ 32 17 */ | |
394 movq_r2r(r0, r7); /* r7 = 53 52 51 50 */ | |
395 pand_r2r(r1, r4); /* r4 = __ 36 __ __ */ | |
396 psllq_i2r(48, r7); /* r7 = 50 __ __ __ */ | |
397 por_r2r(r4, r6); /* r6 = __ 36 32 17 */ | |
398 movq_m2r(*(input_bytes+88), r4); | |
399 por_r2r(r6, r7); /* r7 = 50 36 32 17 = R5 */ | |
400 pmullw_m2r(*(dequant_matrix_bytes+88), r4); /* r4 = 57 56 55 54 */ | |
401 psrlq_i2r(16, r3); /* r3 = __ __ FF __ */ | |
402 movq_r2m(r7, *(output_data_bytes+24)); /* write R5 = r7 */ | |
403 pand_r2r(r1, r3); /* r3 = __ __ 35 __ */ | |
404 psrlq_i2r(48, r5); /* r5 = __ __ __ 33 */ | |
405 pand_r2r(r2, r1); /* r1 = __ __ __ 34 */ | |
406 movq_m2r(*(input_bytes+104), r6); | |
407 por_r2r(r3, r5); /* r5 = __ __ 35 33 */ | |
408 pmullw_m2r(*(dequant_matrix_bytes+104), r6); /* r6 = 67 66 65 64 */ | |
409 psrlq_i2r(16, r0); /* r0 = __ 53 52 51 */ | |
410 movq_r2r(r4, r7); /* r7 = 57 56 55 54 */ | |
411 movq_r2r(r2, r3); /* r3 = __ __ __ FF */ | |
412 psllq_i2r(48, r7); /* r7 = 54 __ __ __ */ | |
413 pand_r2r(r0, r3); /* r3 = __ __ __ 51 */ | |
414 pxor_r2r(r3, r0); /* r0 = __ 53 52 __ */ | |
415 psllq_i2r(32, r3); /* r3 = __ 51 __ __ */ | |
416 por_r2r(r5, r7); /* r7 = 54 __ 35 33 */ | |
417 movq_r2r(r6, r5); /* r5 = 67 66 65 64 */ | |
418 pand_m2r(*M(1), r6); /* r6 = __ __ 65 __ */ | |
419 por_r2r(r3, r7); /* r7 = 54 51 35 33 = R6 */ | |
420 psllq_i2r(32, r6); /* r6 = 65 __ __ __ */ | |
421 por_r2r(r1, r0); /* r0 = __ 53 52 34 */ | |
422 movq_r2m(r7, *(output_data_bytes+40)); /* write R6 = r7 */ | |
423 por_r2r(r6, r0); /* r0 = 65 53 52 34 = R7 */ | |
424 movq_m2r(*(input_bytes+120), r7); | |
425 movq_r2r(r5, r6); /* r6 = 67 66 65 64 */ | |
426 pmullw_m2r(*(dequant_matrix_bytes+120), r7); /* r7 = 77 76 75 74 */ | |
427 psrlq_i2r(32, r5); /* r5 = __ __ 67 66 */ | |
428 pand_r2r(r2, r6); /* r6 = __ __ __ 64 */ | |
429 movq_r2r(r5, r1); /* r1 = __ __ 67 66 */ | |
430 movq_r2m(r0, *(output_data_bytes+56)); /* write R7 = r0 */ | |
431 pand_r2r(r2, r1); /* r1 = __ __ __ 66 */ | |
432 movq_m2r(*(input_bytes+112), r0); | |
433 movq_r2r(r7, r3); /* r3 = 77 76 75 74 */ | |
434 pmullw_m2r(*(dequant_matrix_bytes+112), r0); /* r0 = 73 72 71 70 */ | |
435 psllq_i2r(16, r3); /* r3 = 76 75 74 __ */ | |
436 pand_m2r(*M(3), r7); /* r7 = 77 __ __ __ */ | |
437 pxor_r2r(r1, r5); /* r5 = __ __ 67 __ */ | |
438 por_r2r(r5, r6); /* r6 = __ __ 67 64 */ | |
439 movq_r2r(r3, r5); /* r5 = 76 75 74 __ */ | |
440 pand_m2r(*M(3), r5); /* r5 = 76 __ __ __ */ | |
441 por_r2r(r1, r7); /* r7 = 77 __ __ 66 */ | |
442 movq_m2r(*(input_bytes+96), r1); | |
443 pxor_r2r(r5, r3); /* r3 = __ 75 74 __ */ | |
444 pmullw_m2r(*(dequant_matrix_bytes+96), r1); /* r1 = 63 62 61 60 */ | |
445 por_r2r(r3, r7); /* r7 = 77 75 74 66 = R15 */ | |
446 por_r2r(r5, r6); /* r6 = 76 __ 67 64 */ | |
447 movq_r2r(r0, r5); /* r5 = 73 72 71 70 */ | |
448 movq_r2m(r7, *(output_data_bytes+120)); /* store R15 = r7 */ | |
449 psrlq_i2r(16, r5); /* r5 = __ 73 72 71 */ | |
450 pand_m2r(*M(2), r5); /* r5 = __ 73 __ __ */ | |
451 movq_r2r(r0, r7); /* r7 = 73 72 71 70 */ | |
452 por_r2r(r5, r6); /* r6 = 76 73 67 64 = R14 */ | |
453 pand_r2r(r2, r0); /* r0 = __ __ __ 70 */ | |
454 pxor_r2r(r0, r7); /* r7 = 73 72 71 __ */ | |
455 psllq_i2r(32, r0); /* r0 = __ 70 __ __ */ | |
456 movq_r2m(r6, *(output_data_bytes+104)); /* write R14 = r6 */ | |
457 psrlq_i2r(16, r4); /* r4 = __ 57 56 55 */ | |
458 movq_m2r(*(input_bytes+72), r5); | |
459 psllq_i2r(16, r7); /* r7 = 72 71 __ __ */ | |
460 pmullw_m2r(*(dequant_matrix_bytes+72), r5); /* r5 = 47 46 45 44 */ | |
461 movq_r2r(r7, r6); /* r6 = 72 71 __ __ */ | |
462 movq_m2r(*M(2), r3); /* r3 = __ FF __ __ */ | |
463 psllq_i2r(16, r6); /* r6 = 71 __ __ __ */ | |
464 pand_m2r(*M(3), r7); /* r7 = 72 __ __ __ */ | |
465 pand_r2r(r1, r3); /* r3 = __ 62 __ __ */ | |
466 por_r2r(r0, r7); /* r7 = 72 70 __ __ */ | |
467 movq_r2r(r1, r0); /* r0 = 63 62 61 60 */ | |
468 pand_m2r(*M(3), r1); /* r1 = 63 __ __ __ */ | |
469 por_r2r(r3, r6); /* r6 = 71 62 __ __ */ | |
470 movq_r2r(r4, r3); /* r3 = __ 57 56 55 */ | |
471 psrlq_i2r(32, r1); /* r1 = __ __ 63 __ */ | |
472 pand_r2r(r2, r3); /* r3 = __ __ __ 55 */ | |
473 por_r2r(r1, r7); /* r7 = 72 70 63 __ */ | |
474 por_r2r(r3, r7); /* r7 = 72 70 63 55 = R13 */ | |
475 movq_r2r(r4, r3); /* r3 = __ 57 56 55 */ | |
476 pand_m2r(*M(1), r3); /* r3 = __ __ 56 __ */ | |
477 movq_r2r(r5, r1); /* r1 = 47 46 45 44 */ | |
478 movq_r2m(r7, *(output_data_bytes+88)); /* write R13 = r7 */ | |
479 psrlq_i2r(48, r5); /* r5 = __ __ __ 47 */ | |
480 movq_m2r(*(input_bytes+64), r7); | |
481 por_r2r(r3, r6); /* r6 = 71 62 56 __ */ | |
482 pmullw_m2r(*(dequant_matrix_bytes+64), r7); /* r7 = 43 42 41 40 */ | |
483 por_r2r(r5, r6); /* r6 = 71 62 56 47 = R12 */ | |
484 pand_m2r(*M(2), r4); /* r4 = __ 57 __ __ */ | |
485 psllq_i2r(32, r0); /* r0 = 61 60 __ __ */ | |
486 movq_r2m(r6, *(output_data_bytes+72)); /* write R12 = r6 */ | |
487 movq_r2r(r0, r6); /* r6 = 61 60 __ __ */ | |
488 pand_m2r(*M(3), r0); /* r0 = 61 __ __ __ */ | |
489 psllq_i2r(16, r6); /* r6 = 60 __ __ __ */ | |
490 movq_m2r(*(input_bytes+40), r5); | |
491 movq_r2r(r1, r3); /* r3 = 47 46 45 44 */ | |
492 pmullw_m2r(*(dequant_matrix_bytes+40), r5); /* r5 = 27 26 25 24 */ | |
493 psrlq_i2r(16, r1); /* r1 = __ 47 46 45 */ | |
494 pand_m2r(*M(1), r1); /* r1 = __ __ 46 __ */ | |
495 por_r2r(r4, r0); /* r0 = 61 57 __ __ */ | |
496 pand_r2r(r7, r2); /* r2 = __ __ __ 40 */ | |
497 por_r2r(r1, r0); /* r0 = 61 57 46 __ */ | |
498 por_r2r(r2, r0); /* r0 = 61 57 46 40 = R11 */ | |
499 psllq_i2r(16, r3); /* r3 = 46 45 44 __ */ | |
500 movq_r2r(r3, r4); /* r4 = 46 45 44 __ */ | |
501 movq_r2r(r5, r2); /* r2 = 27 26 25 24 */ | |
502 movq_r2m(r0, *(output_data_bytes+112)); /* write R11 = r0 */ | |
503 psrlq_i2r(48, r2); /* r2 = __ __ __ 27 */ | |
504 pand_m2r(*M(2), r4); /* r4 = __ 45 __ __ */ | |
505 por_r2r(r2, r6); /* r6 = 60 __ __ 27 */ | |
506 movq_m2r(*M(1), r2); /* r2 = __ __ FF __ */ | |
507 por_r2r(r4, r6); /* r6 = 60 45 __ 27 */ | |
508 pand_r2r(r7, r2); /* r2 = __ __ 41 __ */ | |
509 psllq_i2r(32, r3); /* r3 = 44 __ __ __ */ | |
510 por_m2r(*(output_data_bytes+80), r3); /* r3 = 44 __ __ 23 */ | |
511 por_r2r(r2, r6); /* r6 = 60 45 41 27 = R10 */ | |
512 movq_m2r(*M(3), r2); /* r2 = FF __ __ __ */ | |
513 psllq_i2r(16, r5); /* r5 = 26 25 24 __ */ | |
514 movq_r2m(r6, *(output_data_bytes+96)); /* store R10 = r6 */ | |
515 pand_r2r(r5, r2); /* r2 = 26 __ __ __ */ | |
516 movq_m2r(*M(2), r6); /* r6 = __ FF __ __ */ | |
517 pxor_r2r(r2, r5); /* r5 = __ 25 24 __ */ | |
518 pand_r2r(r7, r6); /* r6 = __ 42 __ __ */ | |
519 psrlq_i2r(32, r2); /* r2 = __ __ 26 __ */ | |
520 pand_m2r(*M(3), r7); /* r7 = 43 __ __ __ */ | |
521 por_r2r(r2, r3); /* r3 = 44 __ 26 23 */ | |
522 por_m2r(*(output_data_bytes+64), r7); /* r7 = 43 __ __ 12 */ | |
523 por_r2r(r3, r6); /* r6 = 44 42 26 23 = R9 */ | |
524 por_r2r(r5, r7); /* r7 = 43 25 24 12 = R8 */ | |
525 movq_r2m(r6, *(output_data_bytes+80)); /* store R9 = r6 */ | |
526 movq_r2m(r7, *(output_data_bytes+64)); /* store R8 = r7 */ | |
527 | |
528 | |
529 #undef M | |
530 | |
531 /* at this point, function has completed dequantization + dezigzag + | 289 /* at this point, function has completed dequantization + dezigzag + |
532 * partial transposition; now do the idct itself */ | 290 * partial transposition; now do the idct itself */ |
533 | |
534 #define I(K) (output_data + K * 8) | 291 #define I(K) (output_data + K * 8) |
535 #define J(K) (output_data + ((K - 4) * 8) + 4) | 292 #define J(K) (output_data + ((K - 4) * 8) + 4) |
536 | 293 |
537 RowIDCT(); | 294 RowIDCT(); |
538 Transpose(); | 295 Transpose(); |