1 ;*****************************************************************************
2 ;* dct.asm: h264 encoder library
3 ;*****************************************************************************
4 ;* Copyright (C) 2003 x264 project
5 ;* $Id: dct.asm,v 1.1 2004/06/03 19:27:07 fenrir Exp $
7 ;* Authors: Min Chen <chenm001.163.com> (converted to nasm)
8 ;* Laurent Aimar <fenrir@via.ecp.fr> (initial version)
10 ;* This program is free software; you can redistribute it and/or modify
11 ;* it under the terms of the GNU General Public License as published by
12 ;* the Free Software Foundation; either version 2 of the License, or
13 ;* (at your option) any later version.
15 ;* This program 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
18 ;* GNU General Public License for more details.
20 ;* You should have received a copy of the GNU General Public License
21 ;* along with this program; if not, write to the Free Software
22 ;* Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111, USA.
23 ;*****************************************************************************
25 ;*****************************************************************************
27 ;* Revision history: *
29 ;* 2004.04.28 portab all 4x4 function to nasm (CM) *
31 ;*****************************************************************************
35 ;=============================================================================
36 ; Macros and other preprocessor constants
37 ;=============================================================================
40 %define GLOBAL wrt rip
58 %macro MMX_LOAD_DIFF_4P 5
66 %macro MMX_SUMSUB_BA 2
72 %macro MMX_SUMSUB_BADC 4
81 %macro MMX_SUMSUB2_AB 3
89 %macro MMX_SUMSUBD2_AB 4
104 %macro SBUTTERFLYdq 3
110 ;-----------------------------------------------------------------------------
111 ; input ABCD output ADTC
112 ;-----------------------------------------------------------------------------
113 %macro MMX_TRANSPOSE 5
114 SBUTTERFLYwd %1, %2, %5
115 SBUTTERFLYwd %3, %4, %2
116 SBUTTERFLYdq %1, %3, %4
117 SBUTTERFLYdq %5, %2, %3
120 %macro MMX_STORE_DIFF_4P 5
133 ;=============================================================================
134 ; Local Data (Read Only)
135 ;=============================================================================
143 ;-----------------------------------------------------------------------------
144 ; Various memory constants (trigonometric values or rounding values)
145 ;-----------------------------------------------------------------------------
148 x264_mmx_1: dw 1, 1, 1, 1
149 x264_mmx_32: dw 32, 32, 32, 32
150 x264_mmx_PPNN: dw 1, 1, -1, -1
151 x264_mmx_PNPN: dw 1, -1, 1, -1
152 x264_mmx_PNNP: dw 1, -1, -1, 1
153 x264_mmx_PPPN: dw 1, 1, 1, -1
154 x264_mmx_PPNP: dw 1, 1, -1, 1
155 x264_mmx_2121: dw 2, 1, 2, 1
156 x264_mmx_p2n2p1p1: dw 2, -2, 1, 1
158 ;=============================================================================
160 ;=============================================================================
164 cglobal x264_dct4x4dc_mmxext
167 ;-----------------------------------------------------------------------------
168 ; void __cdecl dct4x4dc( int16_t d[4][4] )
169 ;-----------------------------------------------------------------------------
170 x264_dct4x4dc_mmxext:
176 MMX_SUMSUB_BADC mm1, mm0, mm3, mm2 ; mm1=s01 mm0=d01 mm3=s23 mm2=d23
177 MMX_SUMSUB_BADC mm3, mm1, mm2, mm0 ; mm3=s01+s23 mm1=s01-s23 mm2=d01+d23 mm0=d01-d23
179 MMX_TRANSPOSE mm3, mm1, mm0, mm2, mm4 ; in: mm3, mm1, mm0, mm2 out: mm3, mm2, mm4, mm0
181 MMX_SUMSUB_BADC mm2, mm3, mm0, mm4 ; mm2=s01 mm3=d01 mm0=s23 mm4=d23
182 MMX_SUMSUB_BADC mm0, mm2, mm4, mm3 ; mm0=s01+s23 mm2=s01-s23 mm4=d01+d23 mm3=d01-d23
184 MMX_TRANSPOSE mm0, mm2, mm3, mm4, mm1 ; in: mm0, mm2, mm3, mm4 out: mm0, mm4, mm1, mm3
186 movq mm6, [x264_mmx_1 GLOBAL]
201 cglobal x264_idct4x4dc_mmxext
204 ;-----------------------------------------------------------------------------
205 ; void __cdecl x264_idct4x4dc_mmxext( int16_t d[4][4] )
206 ;-----------------------------------------------------------------------------
207 x264_idct4x4dc_mmxext:
213 MMX_SUMSUB_BADC mm1, mm0, mm3, mm2 ; mm1=s01 mm0=d01 mm3=s23 mm2=d23
214 MMX_SUMSUB_BADC mm3, mm1, mm2, mm0 ; mm3=s01+s23 mm1=s01-s23 mm2=d01+d23 mm0=d01-d23
216 MMX_TRANSPOSE mm3, mm1, mm0, mm2, mm4 ; in: mm3, mm1, mm0, mm2 out: mm3, mm2, mm4, mm0
218 MMX_SUMSUB_BADC mm2, mm3, mm0, mm4 ; mm2=s01 mm3=d01 mm0=s23 mm4=d23
219 MMX_SUMSUB_BADC mm0, mm2, mm4, mm3 ; mm0=s01+s23 mm2=s01-s23 mm4=d01+d23 mm3=d01-d23
221 MMX_TRANSPOSE mm0, mm2, mm3, mm4, mm1 ; in: mm0, mm2, mm3, mm4 out: mm0, mm4, mm1, mm3
229 cglobal x264_sub4x4_dct_mmxext
232 ;-----------------------------------------------------------------------------
233 ; void __cdecl x264_sub4x4_dct_mmxext( int16_t dct[4][4], uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 )
234 ;-----------------------------------------------------------------------------
235 x264_sub4x4_dct_mmxext:
238 movsxd rbx, edx ; i_pix1
239 ; mov rcx, rcx ; pix2
240 movsxd rdx, r8d ; i_pix2
245 MMX_LOAD_DIFF_4P mm0, mm6, mm7, [rax ], [rcx]
246 MMX_LOAD_DIFF_4P mm1, mm6, mm7, [rax+rbx ], [rcx+rdx]
247 MMX_LOAD_DIFF_4P mm2, mm6, mm7, [rax+rbx*2], [rcx+rdx*2]
250 MMX_LOAD_DIFF_4P mm3, mm6, mm7, [rax+rbx*2], [rcx+rdx*2]
252 MMX_SUMSUB_BADC mm3, mm0, mm2, mm1 ; mm3=s03 mm0=d03 mm2=s12 mm1=d12
254 MMX_SUMSUB_BA mm2, mm3 ; mm2=s03+s12 mm3=s03-s12
255 MMX_SUMSUB2_AB mm0, mm1, mm4 ; mm0=2.d03+d12 mm4=d03-2.d12
257 ; transpose in: mm2, mm0, mm3, mm4, out: mm2, mm4, mm1, mm3
258 MMX_TRANSPOSE mm2, mm0, mm3, mm4, mm1
260 MMX_SUMSUB_BADC mm3, mm2, mm1, mm4 ; mm3=s03 mm2=d03 mm1=s12 mm4=d12
262 MMX_SUMSUB_BA mm1, mm3 ; mm1=s03+s12 mm3=s03-s12
263 MMX_SUMSUB2_AB mm2, mm4, mm0 ; mm2=2.d03+d12 mm0=d03-2.d12
265 ; transpose in: mm1, mm2, mm3, mm0, out: mm1, mm0, mm4, mm3
266 MMX_TRANSPOSE mm1, mm2, mm3, mm0, mm4
268 movq [rdi+ 0], mm1 ; dct
276 cglobal x264_add4x4_idct_mmxext
279 ;-----------------------------------------------------------------------------
280 ; void __cdecl x264_add4x4_idct_mmxext( uint8_t *p_dst, int i_dst, int16_t dct[4][4] )
281 ;-----------------------------------------------------------------------------
282 x264_add4x4_idct_mmxext:
284 movq mm0, [rdx+ 0] ; dct
290 movsxd rcx, esi ; i_dst
293 ; out:mm0, mm1, mm2, mm3
294 MMX_TRANSPOSE mm0, mm4, mm3, mm1, mm2
296 MMX_SUMSUB_BA mm2, mm0 ; mm2=s02 mm0=d02
297 MMX_SUMSUBD2_AB mm1, mm3, mm5, mm4 ; mm1=s13 mm4=d13 ( well 1 + 3>>1 and 1>>1 + 3)
299 MMX_SUMSUB_BADC mm1, mm2, mm4, mm0 ; mm1=s02+s13 mm2=s02-s13 mm4=d02+d13 mm0=d02-d13
301 ; in: mm1, mm4, mm0, mm2 out: mm1, mm2, mm3, mm0
302 MMX_TRANSPOSE mm1, mm4, mm0, mm2, mm3
304 MMX_SUMSUB_BA mm3, mm1 ; mm3=s02 mm1=d02
305 MMX_SUMSUBD2_AB mm2, mm0, mm5, mm4 ; mm2=s13 mm4=d13 ( well 1 + 3>>1 and 1>>1 + 3)
307 MMX_SUMSUB_BADC mm2, mm3, mm4, mm1 ; mm2=s02+s13 mm3=s02-s13 mm4=d02+d13 mm1=d02-d13
310 movq mm6, [x264_mmx_32 GLOBAL]
312 MMX_STORE_DIFF_4P mm2, mm0, mm6, mm7, [rax]
313 MMX_STORE_DIFF_4P mm4, mm0, mm6, mm7, [rax+rcx]
314 MMX_STORE_DIFF_4P mm1, mm0, mm6, mm7, [rax+rcx*2]
315 MMX_STORE_DIFF_4P mm3, mm0, mm6, mm7, [rax+rdx]
321 ; =============================================================================
323 ; =============================================================================
325 ; -----------------------------------------------------------------------------
326 ; input 2x8 unsigned bytes (%5,%6), zero (%7) output: difference (%1,%2)
327 ; -----------------------------------------------------------------------------
328 %macro MMX_LOAD_DIFF_8P 7
341 %macro MMX_LOADSUMSUB 4 ; returns %1=%3+%4, %2=%3-%4
347 %macro MMX_STORE_DIFF_8P 6
358 cglobal x264_pixel_sub_8x8_mmx
359 cglobal x264_xdct8_mmxext
360 cglobal x264_ydct8_mmx
362 cglobal x264_xidct8_mmxext
363 cglobal x264_yidct8_mmx
364 cglobal x264_pixel_add_8x8_mmx
367 ;-----------------------------------------------------------------------------
368 ; void __cdecl x264_pixel_sub_8x8_mmx( int16_t *diff, uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 );
369 ;-----------------------------------------------------------------------------
370 x264_pixel_sub_8x8_mmx:
371 ; mov rdi, rdi ; diff
372 ; mov rsi, rsi ; pix1
373 movsxd rdx, edx ; i_pix1
374 ; mov rcx, rcx ; pix2
375 movsxd r8, r8d ; i_pix2
381 MMX_LOAD_DIFF_8P mm0, mm1, mm2, mm3, [rsi], [rcx], mm7
383 movq [rdi+disp+8], mm1
392 ;-----------------------------------------------------------------------------
393 ; void __cdecl x264_xdct8_mmxext( int16_t dest[8][8] );
394 ;-----------------------------------------------------------------------------
397 movq mm5, [x264_mmx_PPNN GLOBAL]
398 movq mm6, [x264_mmx_PNNP GLOBAL]
399 movq mm4, [x264_mmx_PPPN GLOBAL]
400 movq mm7, [x264_mmx_PPNP GLOBAL]
402 ;-------------------------------------------------------------------------
403 ; horizontal dct ( compute 1 row at a time -> 8 loops )
404 ;-------------------------------------------------------------------------
410 movq mm1, [rdi+disp+8]
412 pshufw mm2, mm1, 00011011b
414 paddw mm0, mm2 ; (low)s07/s16/d25/s34(high)
415 psubw mm1, mm2 ; (low)d07/d16/d25/d34(high)
417 pshufw mm2, mm0, 00011011b ; (low)s34/s25/s16/s07(high)
418 pmullw mm0, mm5 ; (low)s07/s16/-s25/-s34(high)
419 paddw mm0, mm2 ; (low)a0/a1/a3/a2(high)
422 psraw mm1, 1 ; (low)d07/d16/d25/d34(high) (x>>1)
423 pshufw mm2, mm3, 10110001b ; (low)d16/d07/d34/d25(high)
424 paddw mm1, mm3 ; (low)d07/d16/d25/d34(high) (x+(x>>1))
425 pshufw mm3, mm2, 00011011b ; (low)d25/d34/d07/d16(high)
426 pmullw mm2, mm5 ; (low)d16/d07/-d34/-d25(high)
427 pmullw mm1, mm6 ; (low)d07/-d16/-d25/d34(high) (x+(x>>1))
429 paddw mm1, mm3 ; (low)a4/a6/a5/a7(high)
432 pshufw mm2, mm0, 11001001b ; (low)a1/a3/a0/a2(high)
433 pshufw mm0, mm0, 10011100b ; (low)a0/a2/a1/a3(high)
434 pmullw mm2, [x264_mmx_2121 GLOBAL]
435 pmullw mm0, mm5 ; (low)a0/a2/-a1/-a3(high)
436 psraw mm2, 1 ; (low)a1/a3>>1/a0/a2>>1(high)
437 paddw mm0, mm2 ; (low)dst0/dst2/dst4/dst6(high)
439 pshufw mm1, mm1, 00100111b ; (low)a7/a6/a5/a4(high)
440 pshufw mm2, mm1, 00011011b ; (low)a4/a5/a6/a7(high)
441 psraw mm1, 2 ; (low)a7>>2/a6>>2/a5>>2/a4>>2(high)
442 pmullw mm2, mm4 ; (low)a4/a5/a6/-a7(high)
443 pmullw mm1, mm7 ; (low)a7>>2/a6>>2/-a5>>2/a4>>2(high)
444 paddw mm1, mm2 ; (low)dst1/dst3/dst5/dst7(high)
447 punpcklwd mm0, mm1 ; (low)dst0/dst1/dst2/dst3(high)
448 punpckhwd mm2, mm1 ; (low)dst4/dst5/dst6/dst7(high)
451 movq [rdi+disp+8], mm2
459 ;-----------------------------------------------------------------------------
460 ; void __cdecl x264_ydct8_mmx( int16_t dest[8][8] );
461 ;-----------------------------------------------------------------------------
464 ;-------------------------------------------------------------------------
465 ; vertical dct ( compute 4 columns at a time -> 2 loops )
466 ;-------------------------------------------------------------------------
471 MMX_LOADSUMSUB mm2, mm3, [rdi+disp+0*16], [rdi+disp+7*16] ; mm2 = s07, mm3 = d07
472 MMX_LOADSUMSUB mm1, mm5, [rdi+disp+1*16], [rdi+disp+6*16] ; mm1 = s16, mm5 = d16
473 MMX_LOADSUMSUB mm0, mm6, [rdi+disp+2*16], [rdi+disp+5*16] ; mm0 = s25, mm6 = d25
474 MMX_LOADSUMSUB mm4, mm7, [rdi+disp+3*16], [rdi+disp+4*16] ; mm4 = s34, mm7 = d34
476 MMX_SUMSUB_BA mm4, mm2 ; mm4 = a0, mm2 = a2
477 MMX_SUMSUB_BA mm0, mm1 ; mm0 = a1, mm1 = a3
478 MMX_SUMSUB_BA mm0, mm4 ; mm0 = dst0, mm1 = dst4
480 movq [rdi+disp+0*16], mm0
481 movq [rdi+disp+4*16], mm4
485 paddw mm0, mm2 ; a2 + (a3>>1)
487 psubw mm2, mm1 ; (a2>>1) - a3
489 movq [rdi+disp+2*16], mm0
490 movq [rdi+disp+6*16], mm2
494 paddw mm0, mm6 ; d25+(d25>>1)
496 psubw mm1, mm7 ; a5 = d07-d34-(d25+(d25>>1))
501 paddw mm0, mm5 ; d16+(d16>>1)
503 paddw mm2, mm7 ; a6 = d07+d34-(d16+(d16>>1))
508 paddw mm0, mm3 ; d07+(d07>>1)
510 paddw mm0, mm6 ; a4 = d16+d25+(d07+(d07>>1))
514 paddw mm3, mm7 ; d34+(d34>>1)
516 psubw mm3, mm6 ; a7 = d16-d25+(d34+(d34>>1))
520 paddw mm7, mm0 ; a4 + (a7>>2)
524 paddw mm6, mm1 ; a5 + (a6>>2)
528 psubw mm0, mm3 ; (a4>>2) - a7
529 psubw mm2, mm1 ; a6 - (a5>>2)
531 movq [rdi+disp+1*16], mm7
532 movq [rdi+disp+3*16], mm6
533 movq [rdi+disp+5*16], mm2
534 movq [rdi+disp+7*16], mm0
542 ;-----------------------------------------------------------------------------
543 ; void __cdecl x264_xidct8_mmxext( int16_t dest[8][8] );
544 ;-----------------------------------------------------------------------------
547 movq mm4, [x264_mmx_PPNN GLOBAL]
548 movq mm5, [x264_mmx_PNPN GLOBAL]
549 movq mm6, [x264_mmx_PPNP GLOBAL]
550 movq mm7, [x264_mmx_PPPN GLOBAL]
552 ;-------------------------------------------------------------------------
553 ; horizontal idct ( compute 1 row at a time -> 8 loops )
554 ;-------------------------------------------------------------------------
559 pshufw mm0, [rdi+disp], 11011000b ; (low)d0,d2,d1,d3(high)
560 pshufw mm2, [rdi+disp+8], 11011000b ; (low)d4,d6,d5,d7(high)
562 punpcklwd mm0, mm2 ; (low)d0,d4,d2,d6(high)
563 punpckhwd mm1, mm2 ; (low)d1,d5,d3,d7(high)
565 pshufw mm2, mm0, 10110001b ; (low)d4,d0,d6,d2(high)
566 pmullw mm0, [x264_mmx_p2n2p1p1 GLOBAL]
567 ; (low)2*d0,-2*d4,d2,d6(high)
568 pmullw mm2, mm6 ; (low)d4,d0,-d6,d2(high)
569 psraw mm0, 1 ; (low)d0,-d4,d2>>1,d6>>1(high)
570 paddw mm0, mm2 ; (low)e0,e2,e4,e6(high)
572 movq mm3, mm1 ; (low)d1,d5,d3,d7(high)
573 psraw mm1, 1 ; (low)d1>>1,d5>>1,d3>>1,d7>>1(high)
574 pshufw mm2, mm3, 10110001b ; (low)d5,d1,d7,d3(high)
575 paddw mm1, mm3 ; (low)d1+(d1>>1),d5+(d5>>1),d3+(d3>>1),d7+(d7>>1)(high)
576 pshufw mm3, mm2, 00011011b ; (low)d3,d7,d1,d5(high)
577 pmullw mm1, mm4 ; (low)d1+(d1>>1),d5+(d5>>1),-d3-(d3>>1),-d7-(d7>>1)(high)
578 pmullw mm2, mm5 ; (low)d5,-d1,d7,-d3(high)
580 paddw mm1, mm2 ; (low)e7,e5,e3,e1(high)
582 pshufw mm2, mm0, 00011011b ; (low)e6,e4,e2,e0(high)
583 pmullw mm0, mm4 ; (low)e0,e2,-e4,-e6(high)
584 pshufw mm3, mm1, 00011011b ; (low)e1,e3,e5,e7(high)
585 psraw mm1, 2 ; (low)e7>>2,e5>>2,e3>>2,e1>>2(high)
586 pmullw mm3, mm6 ; (low)e1,e3,-e5,e7(high)
587 pmullw mm1, mm7 ; (low)e7>>2,e5>>2,e3>>2,-e1>>2(high)
588 paddw mm0, mm2 ; (low)f0,f2,f4,f6(high)
589 paddw mm1, mm3 ; (low)f1,f3,f5,f7(high)
591 pshufw mm3, mm0, 00011011b ; (low)f6,f4,f2,f0(high)
592 pshufw mm2, mm1, 00011011b ; (low)f7,f5,f3,f1(high)
597 movq [rdi+disp+8], mm3
605 ;-----------------------------------------------------------------------------
606 ; void __cdecl x264_yidct8_mmx( int16_t dest[8][8] );
607 ;-----------------------------------------------------------------------------
610 ;-------------------------------------------------------------------------
611 ; vertical idct ( compute 4 columns at a time -> 2 loops )
612 ;-------------------------------------------------------------------------
617 movq mm1, [rdi+disp+1*16] ; mm1 = d1
618 movq mm3, [rdi+disp+3*16] ; mm3 = d3
619 movq mm5, [rdi+disp+5*16] ; mm5 = d5
620 movq mm7, [rdi+disp+7*16] ; mm7 = d7
627 psubw mm0, mm3 ; mm0 = e1
634 paddw mm2, mm1 ; mm2 = e3
640 psubw mm4, mm1 ; mm4 = e5
646 paddw mm6, mm3 ; mm6 = e7
656 paddw mm1, mm6 ; mm1 = f1
657 paddw mm3, mm2 ; mm3 = f3
658 psubw mm5, mm4 ; mm5 = f5
659 psubw mm7, mm0 ; mm7 = f7
661 movq mm2, [rdi+disp+2*16] ; mm2 = d2
662 movq mm6, [rdi+disp+6*16] ; mm6 = d6
667 psubw mm4, mm0 ; mm4 = a4
668 paddw mm6, mm2 ; mm6 = a6
670 movq mm2, [rdi+disp+0*16] ; mm2 = d0
671 movq mm0, [rdi+disp+4*16] ; mm0 = d4
672 MMX_SUMSUB_BA mm0, mm2 ; mm0 = a0, mm2 = a2
674 MMX_SUMSUB_BA mm6, mm0 ; mm6 = f0, mm0 = f6
675 MMX_SUMSUB_BA mm4, mm2 ; mm4 = f2, mm2 = f4
677 MMX_SUMSUB_BA mm7, mm6 ; mm7 = g0, mm6 = g7
678 MMX_SUMSUB_BA mm5, mm4 ; mm5 = g1, mm4 = g6
679 MMX_SUMSUB_BA mm3, mm2 ; mm3 = g2, mm2 = g5
680 MMX_SUMSUB_BA mm1, mm0 ; mm1 = g3, mm0 = g4
691 movq [rdi+disp+0*16], mm7
692 movq [rdi+disp+1*16], mm5
693 movq [rdi+disp+2*16], mm3
694 movq [rdi+disp+3*16], mm1
695 movq [rdi+disp+4*16], mm0
696 movq [rdi+disp+5*16], mm2
697 movq [rdi+disp+6*16], mm4
698 movq [rdi+disp+7*16], mm6
706 ;-----------------------------------------------------------------------------
707 ; void __cdecl x264_pixel_add_8x8_mmx( unit8_t *dst, int i_dst, int16_t src[8][8] );
708 ;-----------------------------------------------------------------------------
709 x264_pixel_add_8x8_mmx:
711 movsxd rsi, esi ; i_dst
718 MMX_STORE_DIFF_8P mm0, mm1, [rdi], [rdx+disp], [rdx+disp+8], mm7