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/x86util.asm"
27 ps_mask: dd 0, ~0, ~0, ~0
28 ps_mask2: dd 0, ~0, 0, ~0
29 ps_mask3: dd 0, 0, 0, ~0
30 ps_mask4: dd 0, ~0, 0, 0
32 ps_val1: dd -0.5, -0.5, -0.8660254038, -0.8660254038
33 ps_val2: dd 1.0, 1.0, 0.8660254038, 0.8660254038
34 ps_val3: dd 0.1736481777, 0.1736481777, 0.3420201433, 0.3420201433
35 ps_val4: dd -0.7660444431, -0.7660444431, 0.8660254038, 0.8660254038
36 ps_val5: dd -0.9396926208, -0.9396926208, -0.9848077530, -0.9848077530
37 ps_val6: dd 0.5, 0.5, -0.6427876097, -0.6427876097
38 ps_val7: dd 1.0, 1.0, -0.6427876097, -0.6427876097
40 ps_p1p1m1m1: dd 0, 0, 0x80000000, 0x80000000
41 ps_p1m1p1m1: dd 0, 0x80000000, 0, 0x80000000
43 ps_cosh: dd 1.0, 0.50190991877167369479, 1.0, 5.73685662283492756461
44 dd 1.0, 0.51763809020504152469, 1.0, 1.93185165257813657349
45 dd 1.0, 0.55168895948124587824, -1.0, -1.18310079157624925896
46 dd 1.0, 0.61038729438072803416, -1.0, -0.87172339781054900991
47 dd 1.0, 0.70710678118654752439, 0.0, 0.0
49 ps_cosh_sse3: dd 1.0, -0.50190991877167369479, 1.0, -5.73685662283492756461
50 dd 1.0, -0.51763809020504152469, 1.0, -1.93185165257813657349
51 dd 1.0, -0.55168895948124587824, -1.0, 1.18310079157624925896
52 dd 1.0, -0.61038729438072803416, -1.0, 0.87172339781054900991
53 dd 1.0, 0.70710678118654752439, 0.0, 0.0
55 costabs: times 4 dd 0.98480773
58 times 4 dd -0.76604444
59 times 4 dd -0.64278764
61 times 4 dd -0.50000000
62 times 4 dd -0.34202015
63 times 4 dd -0.17364818
78 %if cpuflag(sse2) && notcpuflag(avx)
85 ; input %2={x1,x2,x3,x4}, %3={y1,y2,y3,y4}
86 ; output %1={x3,x4,y1,y2}
87 %macro BUILDINVHIGHLOW 3
89 shufps %1, %2, %3, 0x4e
96 ; input %2={x1,x2,x3,x4}, %3={y1,y2,y3,y4}
97 ; output %1={x4,y1,y2,y3}
100 palignr %1, %3, %2, 12
102 BUILDINVHIGHLOW %1, %2, %3
103 shufps %1, %1, %3, 0x99
118 xorps %1, [ps_p1p1m1m1]
121 mulps %1, %1, [ps_cosh_sse3 + %3]
125 mulps %1, [ps_cosh + %3]
127 xorps %1, [ps_p1m1p1m1]
135 movss [%3 + 2*%4], %2
139 movss [%3 + 3*%4], %2
145 movlps %2, [%3 + 2*%4]
146 movhps %2, [%3 + 3*%4]
159 %macro DEFINE_IMDCT 0
160 cglobal imdct36_float, 4,4,9, out, buf, in, win
162 ; for(i=17;i>=1;i--) in[i] += in[i-1];
169 andps m6, m6, [ps_mask]
183 BUILDINVHIGHLOW m6, m3, m4
184 shufps m6, m6, m4, 0xa9
190 ; for(i=17;i>=3;i-=2) in[i] += in[i-2];
192 andps m5, m5, [ps_mask3]
194 BUILDINVHIGHLOW m7, m0, m1
195 andps m7, m7, [ps_mask2]
199 BUILDINVHIGHLOW m6, m1, m2
200 andps m6, m6, [ps_mask2]
204 BUILDINVHIGHLOW m7, m2, m3
205 andps m7, m7, [ps_mask2]
210 andps m6, m6, [ps_mask4]
216 movlhps m6, m1, m5 ; zero out high values
225 mulps m7, m2, [ps_val1]
228 mulps m5, m8, [ps_val2]
230 mulps m5, m5, [ps_val2]
234 mulps m5, m6, [ps_val1]
246 shufps m6, m4, m3, 0xe4
248 mulps m6, m6, [ps_val3]
251 mulps m4, m4, [ps_val4]
253 shufps m1, m1, m0, 0xe4
255 mulps m1, m1, [ps_val5]
257 mulps m3, m3, [ps_val6]
258 mulps m0, m0, [ps_val7]
261 xorps m2, m1, [ps_p1p1m1m1]
267 xorps m3, m3, [ps_p1p1m1m1]
269 shufps m0, m0, m4, 0xe4
273 BUILDINVHIGHLOW m4, m2, m3
274 shufps m3, m3, m2, 0x4e
276 ; we have tmp = {SwAPLH(m0), SwAPLH(m7), m3, m4, m5}
283 mulps m5, m5, [ps_cosh + 64]
285 xorps m5, m5, [ps_p1m1p1m1]
289 ; m0 0 1 2 3 => 2 6 10 14 m1
290 ; m7 4 5 6 7 => 3 7 11 15 m2
291 ; m3 8 9 10 11 => 17 13 9 5 m3
292 ; m4 12 13 14 15 => 16 12 8 4 m5
293 ; m5 16 17 xx xx => 0 1 xx xx m0
308 movss m4, [bufq + 4*68]
309 movss m7, [bufq + 4*64]
311 mulps m6, m6, [winq + 16*4]
313 movss [outq + 64*SBLIMIT], m6
314 shufps m6, m6, m6, 0xb1
315 movss [outq + 68*SBLIMIT], m6
317 mulps m6, m3, [winq + 4*4]
318 LOAD m4, m7, bufq + 4*16, 16
320 STORE m6, m7, outq + 16*SBLIMIT, 4*SBLIMIT
322 shufps m4, m0, m3, 0xb5
323 mulps m4, m4, [winq + 8*4]
324 LOAD m7, m6, bufq + 4*32, 16
326 STORE m4, m6, outq + 32*SBLIMIT, 4*SBLIMIT
328 shufps m3, m3, m2, 0xb1
329 mulps m3, m3, [winq + 12*4]
330 LOAD m7, m6, bufq + 4*48, 16
332 STORE m3, m7, outq + 48*SBLIMIT, 4*SBLIMIT
335 LOAD m6, m7, bufq, 16
337 STORE m2, m7, outq, 4*SBLIMIT
339 mulps m4, m1, [winq + 20*4]
340 STORE m4, m7, bufq, 16
342 mulps m3, m5, [winq + 24*4]
343 STORE m3, m7, bufq + 4*16, 16
345 shufps m0, m0, m5, 0xb0
346 mulps m0, m0, [winq + 28*4]
347 STORE m0, m7, bufq + 4*32, 16
349 shufps m5, m5, m1, 0xb1
350 mulps m5, m5, [winq + 32*4]
351 STORE m5, m7, bufq + 4*48, 16
353 shufps m1, m1, m1, 0xb1
354 mulps m1, m1, [winq + 36*4]
355 movss [bufq + 4*64], m1
357 movss [bufq + 4*68], m1
381 %define SPILLED(x) m %+ x
383 %define SPILLED(x) [tmpq+(x-8)*16 + 32*4]
384 %macro SPILL 2 ; xmm#, mempos
385 movaps SPILLED(%2), m%1
388 movaps m%1, SPILLED(%2)
392 %macro DEFINE_FOUR_IMDCT 0
393 cglobal four_imdct36_float, 5,5,16, out, buf, in, win, tmp
395 movhps m0, [inq+64 + 72]
396 movlps m3, [inq+64 + 2*72]
397 movhps m3, [inq+64 + 3*72]
399 shufps m5, m0, m3, 0xdd
400 shufps m0, m0, m3, 0x88
403 movu m6, [inq+48 + 72]
404 mova m7, [inq+48 + 2*72]
405 movu m3, [inq+48 + 3*72]
407 TRANSPOSE4x4PS 1, 6, 7, 3, 4
422 movu m5, [inq+32 + 72]
423 mova m2, [inq+32 + 2*72]
424 movu m7, [inq+32 + 3*72]
426 TRANSPOSE4x4PS 4, 5, 2, 7, 3
442 movu m7, [inq+16 + 72]
443 mova m1, [inq+16 + 2*72]
444 movu m6, [inq+16 + 3*72]
446 TRANSPOSE4x4PS 2, 7, 1, 6, 3
455 mulps m6, [costabs + 16*2]
462 mova m3, [inq + 2*72]
463 movu m5, [inq + 3*72]
465 TRANSPOSE4x4PS 1, 6, 3, 5, 0
475 addps m6, m4, SPILLED(12)
480 mulps m7, [costabs + 16*5]
482 mulps m0, m6, [costabs + 16*6]
488 mulps m6, [costabs + 16*1]
489 subps m4, SPILLED(12)
490 mulps m4, [costabs + 16*8]
491 addps m2, SPILLED(12)
492 mulps m2, [costabs + 16*3]
504 mulps m5, [costabs + 16*7]
506 mulps m1, [costabs + 16*2]
508 mulps m4, [costabs + 16*4]
513 mulps m3, [costabs + 16*2]
524 addps m1, m0, SPILLED(15)
526 mova m4, [costabs + 16*5]
531 mulps m5, m1, [costabs + 16*6]
539 mulps m5, [costabs + 16*1]
540 mulps m7, [costabs + 16*8]
542 mulps m0, [costabs + 16*3]
554 subps m0, SPILLED(11)
555 mulps m0, [costabs + 16*2]
556 addps m4, m7, SPILLED(11)
559 mulps m7, [costabs + 16*7]
560 addps m2, SPILLED(11)
561 mulps m2, [costabs + 16*4]
562 addps m1, m7, [tmpq+4*8]
569 addps m4, m6, SPILLED(10)
570 subps m6, SPILLED(10)
572 mulps m2, [costabs + 16*9]
574 mulps m5, [costabs + 16*17]
577 mulps m2, m1, [winq+4*36]
578 addps m2, [bufq+4*36]
580 mulps m1, [winq+4*32]
581 addps m1, [bufq+4*32]
583 mulps m1, m4, [winq+4*116]
585 mulps m4, [winq+4*112]
589 mulps m1, m6, [winq+4*68]
590 addps m1, [bufq+4*68]
595 mulps m1, m2, [winq+4*148]
597 mulps m2, [winq+4*80]
599 addps m5, m3, [tmpq+4*24]
604 mulps m1, [costabs + 16*10]
606 mulps m0, [costabs + 16*16]
609 mulps m3, m5, [winq+4*40]
610 addps m3, [bufq+4*40]
612 mulps m5, [winq+4*28]
613 addps m5, [bufq+4*28]
615 mulps m1, m6, [winq+4*120]
617 mulps m6, [winq+4*108]
621 mulps m5, m2, [winq+4*64]
622 addps m5, [bufq+4*64]
627 mulps m0, m1, [winq+4*144]
629 mulps m1, [winq+4*84]
633 addps m1, SPILLED(13)
634 subps m5, SPILLED(13)
637 mulps m2, [costabs + 16*11]
639 mulps m3, [costabs + 16*15]
643 mulps m6, m1, [winq+4*44]
644 addps m6, [bufq+4*44]
646 mulps m1, [winq+4*24]
647 addps m1, [bufq+4*24]
649 mulps m0, m2, [winq+4*124]
651 mulps m2, [winq+4*104]
655 mulps m1, m5, [winq+4*60]
656 addps m1, [bufq+4*60]
661 mulps m1, m0, [winq+4*140]
663 mulps m0, [winq+4*88]
666 addps m1, SPILLED(12)
668 subps m2, SPILLED(12)
670 subps m0, m7, SPILLED(11)
671 addps m7, SPILLED(11)
672 mulps m4, m7, [costabs + 16*12]
673 mulps m0, [costabs + 16*14]
676 mulps m7, m1, [winq+4*48]
677 addps m7, [bufq+4*48]
679 mulps m1, [winq+4*20]
680 addps m1, [bufq+4*20]
682 mulps m1, m5, [winq+4*128]
684 mulps m5, [winq+4*100]
688 mulps m1, m2, [winq+4*56]
689 addps m1, [bufq+4*56]
691 mulps m2, [winq+4*12]
692 addps m2, [bufq+4*12]
694 mulps m0, m6, [winq+4*136]
696 mulps m6, [winq+4*92]
699 mulps m0, [costabs + 16*13]
703 mulps m0, m3, [winq+4*52]
704 addps m0, [bufq+4*52]
706 mulps m3, [winq+4*16]
707 addps m3, [bufq+4*16]
709 mulps m0, m2, [winq+4*132]
711 mulps m2, [winq+4*96]