1 ;******************************************************************************
2 ;* 36 point SSE-optimized IMDCT transform
3 ;* Copyright (c) 2011 Vitor Sessak
5 ;* This file is part of Libav.
7 ;* Libav is free software; you can redistribute it and/or
8 ;* modify it under the terms of the GNU Lesser General Public
9 ;* License as published by the Free Software Foundation; either
10 ;* version 2.1 of the License, or (at your option) any later version.
12 ;* Libav is distributed in the hope that it will be useful,
13 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
14 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 ;* Lesser General Public License for more details.
17 ;* You should have received a copy of the GNU Lesser General Public
18 ;* License along with Libav; if not, write to the Free Software
19 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 ;******************************************************************************
22 %include "libavutil/x86/x86inc.asm"
23 %include "libavutil/x86/x86util.asm"
28 ps_mask: dd 0, ~0, ~0, ~0
29 ps_mask2: dd 0, ~0, 0, ~0
30 ps_mask3: dd 0, 0, 0, ~0
31 ps_mask4: dd 0, ~0, 0, 0
33 ps_val1: dd -0.5, -0.5, -0.8660254038, -0.8660254038
34 ps_val2: dd 1.0, 1.0, 0.8660254038, 0.8660254038
35 ps_val3: dd 0.1736481777, 0.1736481777, 0.3420201433, 0.3420201433
36 ps_val4: dd -0.7660444431, -0.7660444431, 0.8660254038, 0.8660254038
37 ps_val5: dd -0.9396926208, -0.9396926208, -0.9848077530, -0.9848077530
38 ps_val6: dd 0.5, 0.5, -0.6427876097, -0.6427876097
39 ps_val7: dd 1.0, 1.0, -0.6427876097, -0.6427876097
41 ps_p1p1m1m1: dd 0, 0, 0x80000000, 0x80000000
42 ps_p1m1p1m1: dd 0, 0x80000000, 0, 0x80000000
44 ps_cosh: dd 1.0, 0.50190991877167369479, 1.0, 5.73685662283492756461
45 dd 1.0, 0.51763809020504152469, 1.0, 1.93185165257813657349
46 dd 1.0, 0.55168895948124587824, -1.0, -1.18310079157624925896
47 dd 1.0, 0.61038729438072803416, -1.0, -0.87172339781054900991
48 dd 1.0, 0.70710678118654752439, 0.0, 0.0
50 ps_cosh_sse3: dd 1.0, -0.50190991877167369479, 1.0, -5.73685662283492756461
51 dd 1.0, -0.51763809020504152469, 1.0, -1.93185165257813657349
52 dd 1.0, -0.55168895948124587824, -1.0, 1.18310079157624925896
53 dd 1.0, -0.61038729438072803416, -1.0, 0.87172339781054900991
54 dd 1.0, 0.70710678118654752439, 0.0, 0.0
56 costabs: times 4 dd 0.98480773
59 times 4 dd -0.76604444
60 times 4 dd -0.64278764
62 times 4 dd -0.50000000
63 times 4 dd -0.34202015
64 times 4 dd -0.17364818
79 %if cpuflag(sse2) && notcpuflag(avx)
86 ; input %2={x1,x2,x3,x4}, %3={y1,y2,y3,y4}
87 ; output %1={x3,x4,y1,y2}
88 %macro BUILDINVHIGHLOW 3
90 shufps %1, %2, %3, 0x4e
97 ; input %2={x1,x2,x3,x4}, %3={y1,y2,y3,y4}
98 ; output %1={x4,y1,y2,y3}
101 palignr %1, %3, %2, 12
103 BUILDINVHIGHLOW %1, %2, %3
104 shufps %1, %1, %3, 0x99
119 xorps %1, [ps_p1p1m1m1]
122 mulps %1, %1, [ps_cosh_sse3 + %3]
126 mulps %1, [ps_cosh + %3]
128 xorps %1, [ps_p1m1p1m1]
136 movss [%3 + 2*%4], %2
140 movss [%3 + 3*%4], %2
146 movlps %2, [%3 + 2*%4]
147 movhps %2, [%3 + 3*%4]
160 %macro DEFINE_IMDCT 0
161 cglobal imdct36_float, 4,4,9, out, buf, in, win
163 ; for(i=17;i>=1;i--) in[i] += in[i-1];
170 andps m6, m6, [ps_mask]
184 BUILDINVHIGHLOW m6, m3, m4
185 shufps m6, m6, m4, 0xa9
191 ; for(i=17;i>=3;i-=2) in[i] += in[i-2];
193 andps m5, m5, [ps_mask3]
195 BUILDINVHIGHLOW m7, m0, m1
196 andps m7, m7, [ps_mask2]
200 BUILDINVHIGHLOW m6, m1, m2
201 andps m6, m6, [ps_mask2]
205 BUILDINVHIGHLOW m7, m2, m3
206 andps m7, m7, [ps_mask2]
211 andps m6, m6, [ps_mask4]
217 movlhps m6, m1, m5 ; zero out high values
226 mulps m7, m2, [ps_val1]
229 mulps m5, m8, [ps_val2]
231 mulps m5, m5, [ps_val2]
235 mulps m5, m6, [ps_val1]
247 shufps m6, m4, m3, 0xe4
249 mulps m6, m6, [ps_val3]
252 mulps m4, m4, [ps_val4]
254 shufps m1, m1, m0, 0xe4
256 mulps m1, m1, [ps_val5]
258 mulps m3, m3, [ps_val6]
259 mulps m0, m0, [ps_val7]
262 xorps m2, m1, [ps_p1p1m1m1]
268 xorps m3, m3, [ps_p1p1m1m1]
270 shufps m0, m0, m4, 0xe4
274 BUILDINVHIGHLOW m4, m2, m3
275 shufps m3, m3, m2, 0x4e
277 ; we have tmp = {SwAPLH(m0), SwAPLH(m7), m3, m4, m5}
284 mulps m5, m5, [ps_cosh + 64]
286 xorps m5, m5, [ps_p1m1p1m1]
290 ; m0 0 1 2 3 => 2 6 10 14 m1
291 ; m7 4 5 6 7 => 3 7 11 15 m2
292 ; m3 8 9 10 11 => 17 13 9 5 m3
293 ; m4 12 13 14 15 => 16 12 8 4 m5
294 ; m5 16 17 xx xx => 0 1 xx xx m0
309 movss m4, [bufq + 4*68]
310 movss m7, [bufq + 4*64]
312 mulps m6, m6, [winq + 16*4]
314 movss [outq + 64*SBLIMIT], m6
315 shufps m6, m6, m6, 0xb1
316 movss [outq + 68*SBLIMIT], m6
318 mulps m6, m3, [winq + 4*4]
319 LOAD m4, m7, bufq + 4*16, 16
321 STORE m6, m7, outq + 16*SBLIMIT, 4*SBLIMIT
323 shufps m4, m0, m3, 0xb5
324 mulps m4, m4, [winq + 8*4]
325 LOAD m7, m6, bufq + 4*32, 16
327 STORE m4, m6, outq + 32*SBLIMIT, 4*SBLIMIT
329 shufps m3, m3, m2, 0xb1
330 mulps m3, m3, [winq + 12*4]
331 LOAD m7, m6, bufq + 4*48, 16
333 STORE m3, m7, outq + 48*SBLIMIT, 4*SBLIMIT
336 LOAD m6, m7, bufq, 16
338 STORE m2, m7, outq, 4*SBLIMIT
340 mulps m4, m1, [winq + 20*4]
341 STORE m4, m7, bufq, 16
343 mulps m3, m5, [winq + 24*4]
344 STORE m3, m7, bufq + 4*16, 16
346 shufps m0, m0, m5, 0xb0
347 mulps m0, m0, [winq + 28*4]
348 STORE m0, m7, bufq + 4*32, 16
350 shufps m5, m5, m1, 0xb1
351 mulps m5, m5, [winq + 32*4]
352 STORE m5, m7, bufq + 4*48, 16
354 shufps m1, m1, m1, 0xb1
355 mulps m1, m1, [winq + 36*4]
356 movss [bufq + 4*64], m1
358 movss [bufq + 4*68], m1
382 %define SPILLED(x) m %+ x
384 %define SPILLED(x) [tmpq+(x-8)*16 + 32*4]
385 %macro SPILL 2 ; xmm#, mempos
386 movaps SPILLED(%2), m%1
389 movaps m%1, SPILLED(%2)
393 %macro DEFINE_FOUR_IMDCT 0
394 cglobal four_imdct36_float, 5,5,8, out, buf, in, win, tmp
396 movhps m0, [inq+64 + 72]
397 movlps m3, [inq+64 + 2*72]
398 movhps m3, [inq+64 + 3*72]
400 shufps m5, m0, m3, 0xdd
401 shufps m0, m0, m3, 0x88
404 movu m6, [inq+48 + 72]
405 mova m7, [inq+48 + 2*72]
406 movu m3, [inq+48 + 3*72]
408 TRANSPOSE4x4PS 1, 6, 7, 3, 4
423 movu m5, [inq+32 + 72]
424 mova m2, [inq+32 + 2*72]
425 movu m7, [inq+32 + 3*72]
427 TRANSPOSE4x4PS 4, 5, 2, 7, 3
443 movu m7, [inq+16 + 72]
444 mova m1, [inq+16 + 2*72]
445 movu m6, [inq+16 + 3*72]
447 TRANSPOSE4x4PS 2, 7, 1, 6, 3
456 mulps m6, [costabs + 16*2]
463 mova m3, [inq + 2*72]
464 movu m5, [inq + 3*72]
466 TRANSPOSE4x4PS 1, 6, 3, 5, 0
476 addps m6, m4, SPILLED(12)
481 mulps m7, [costabs + 16*5]
483 mulps m0, m6, [costabs + 16*6]
489 mulps m6, [costabs + 16*1]
490 subps m4, SPILLED(12)
491 mulps m4, [costabs + 16*8]
492 addps m2, SPILLED(12)
493 mulps m2, [costabs + 16*3]
505 mulps m5, [costabs + 16*7]
507 mulps m1, [costabs + 16*2]
509 mulps m4, [costabs + 16*4]
514 mulps m3, [costabs + 16*2]
525 addps m1, m0, SPILLED(15)
527 mova m4, [costabs + 16*5]
532 mulps m5, m1, [costabs + 16*6]
540 mulps m5, [costabs + 16*1]
541 mulps m7, [costabs + 16*8]
543 mulps m0, [costabs + 16*3]
555 subps m0, SPILLED(11)
556 mulps m0, [costabs + 16*2]
557 addps m4, m7, SPILLED(11)
560 mulps m7, [costabs + 16*7]
561 addps m2, SPILLED(11)
562 mulps m2, [costabs + 16*4]
563 addps m1, m7, [tmpq+4*8]
570 addps m4, m6, SPILLED(10)
571 subps m6, SPILLED(10)
573 mulps m2, [costabs + 16*9]
575 mulps m5, [costabs + 16*17]
578 mulps m2, m1, [winq+4*36]
579 addps m2, [bufq+4*36]
581 mulps m1, [winq+4*32]
582 addps m1, [bufq+4*32]
584 mulps m1, m4, [winq+4*116]
586 mulps m4, [winq+4*112]
590 mulps m1, m6, [winq+4*68]
591 addps m1, [bufq+4*68]
596 mulps m1, m2, [winq+4*148]
598 mulps m2, [winq+4*80]
600 addps m5, m3, [tmpq+4*24]
605 mulps m1, [costabs + 16*10]
607 mulps m0, [costabs + 16*16]
610 mulps m3, m5, [winq+4*40]
611 addps m3, [bufq+4*40]
613 mulps m5, [winq+4*28]
614 addps m5, [bufq+4*28]
616 mulps m1, m6, [winq+4*120]
618 mulps m6, [winq+4*108]
622 mulps m5, m2, [winq+4*64]
623 addps m5, [bufq+4*64]
628 mulps m0, m1, [winq+4*144]
630 mulps m1, [winq+4*84]
634 addps m1, SPILLED(13)
635 subps m5, SPILLED(13)
638 mulps m2, [costabs + 16*11]
640 mulps m3, [costabs + 16*15]
644 mulps m6, m1, [winq+4*44]
645 addps m6, [bufq+4*44]
647 mulps m1, [winq+4*24]
648 addps m1, [bufq+4*24]
650 mulps m0, m2, [winq+4*124]
652 mulps m2, [winq+4*104]
656 mulps m1, m5, [winq+4*60]
657 addps m1, [bufq+4*60]
662 mulps m1, m0, [winq+4*140]
664 mulps m0, [winq+4*88]
667 addps m1, SPILLED(12)
669 subps m2, SPILLED(12)
671 subps m0, m7, SPILLED(11)
672 addps m7, SPILLED(11)
673 mulps m4, m7, [costabs + 16*12]
674 mulps m0, [costabs + 16*14]
677 mulps m7, m1, [winq+4*48]
678 addps m7, [bufq+4*48]
680 mulps m1, [winq+4*20]
681 addps m1, [bufq+4*20]
683 mulps m1, m5, [winq+4*128]
685 mulps m5, [winq+4*100]
689 mulps m1, m2, [winq+4*56]
690 addps m1, [bufq+4*56]
692 mulps m2, [winq+4*12]
693 addps m2, [bufq+4*12]
695 mulps m0, m6, [winq+4*136]
697 mulps m6, [winq+4*92]
700 mulps m0, [costabs + 16*13]
704 mulps m0, m3, [winq+4*52]
705 addps m0, [bufq+4*52]
707 mulps m3, [winq+4*16]
708 addps m3, [bufq+4*16]
710 mulps m0, m2, [winq+4*132]
712 mulps m2, [winq+4*96]