1 ;******************************************************************************
2 ;* Copyright (c) 2012 Michael Niedermayer
4 ;* This file is part of FFmpeg.
6 ;* FFmpeg is free software; you can redistribute it and/or
7 ;* modify it under the terms of the GNU Lesser General Public
8 ;* License as published by the Free Software Foundation; either
9 ;* version 2.1 of the License, or (at your option) any later version.
11 ;* FFmpeg is distributed in the hope that it will be useful,
12 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
13 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 ;* Lesser General Public License for more details.
16 ;* You should have received a copy of the GNU Lesser General Public
17 ;* License along with FFmpeg; if not, write to the Free Software
18 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 ;******************************************************************************
21 %include "libavutil/x86/x86util.asm"
24 flt2pm31: times 8 dd 4.6566129e-10
25 flt2p31 : times 8 dd 2147483648.0
26 flt2p15 : times 8 dd 32768.0
28 word_unpack_shuf : db 0, 1, 4, 5, 8, 9,12,13, 2, 3, 6, 7,10,11,14,15
33 ;to, from, a/u, log2_outsize, log_intsize, const
35 cglobal pack_2ch_%2_to_%1_%3, 3, 4, 6, dst, src, len, src2
36 mov src2q , [srcq+gprsize]
41 jne pack_2ch_%2_to_%1_u_int %+ SUFFIX
43 jne pack_2ch_%2_to_%1_u_int %+ SUFFIX
45 jne pack_2ch_%2_to_%1_u_int %+ SUFFIX
47 pack_2ch_%2_to_%1_u_int %+ SUFFIX
49 lea srcq , [srcq + (1<<%5)*lenq]
50 lea src2q, [src2q + (1<<%5)*lenq]
51 lea dstq , [dstq + (2<<%4)*lenq]
56 mov%3 m0, [ srcq +(1<<%5)*lenq]
58 mov%3 m2, [ src2q+(1<<%5)*lenq]
68 mov%3 m0, [ srcq +(1<<%5)*lenq]
69 mov%3 m1, [mmsize + srcq +(1<<%5)*lenq]
70 mov%3 m2, [ src2q+(1<<%5)*lenq]
71 mov%3 m3, [mmsize + src2q+(1<<%5)*lenq]
78 mov%3 [ dstq+(2<<%4)*lenq], m0
79 mov%3 [ mmsize + dstq+(2<<%4)*lenq], m1
81 mov%3 [2*mmsize + dstq+(2<<%4)*lenq], m2
82 mov%3 [3*mmsize + dstq+(2<<%4)*lenq], m3
83 add lenq, 4*mmsize/(2<<%4)
85 add lenq, 2*mmsize/(2<<%4)
92 cglobal unpack_2ch_%2_to_%1_%3, 3, 4, 7, dst, src, len, dst2
93 mov dst2q , [dstq+gprsize]
98 jne unpack_2ch_%2_to_%1_u_int %+ SUFFIX
100 jne unpack_2ch_%2_to_%1_u_int %+ SUFFIX
102 jne unpack_2ch_%2_to_%1_u_int %+ SUFFIX
104 unpack_2ch_%2_to_%1_u_int %+ SUFFIX
106 lea srcq , [srcq + (2<<%5)*lenq]
107 lea dstq , [dstq + (1<<%4)*lenq]
108 lea dst2q, [dst2q + (1<<%4)*lenq]
111 mova m6, [word_unpack_shuf]
113 mov%3 m0, [ srcq +(2<<%5)*lenq]
114 mov%3 m2, [ mmsize + srcq +(2<<%5)*lenq]
116 %ifidn SUFFIX, _ssse3
137 shufps m0, m2, 10001000b
138 shufps m1, m2, 11011101b
141 mov%3 m2, [2*mmsize + srcq +(2<<%5)*lenq]
143 mov%3 m4, [3*mmsize + srcq +(2<<%5)*lenq]
144 shufps m2, m4, 10001000b
145 shufps m3, m4, 11011101b
149 mov%3 [ dstq+(1<<%4)*lenq], m0
151 mov%3 [ dst2q+(1<<%4)*lenq], m2
152 mov%3 [ mmsize + dstq+(1<<%4)*lenq], m1
153 mov%3 [ mmsize + dst2q+(1<<%4)*lenq], m3
154 add lenq, 2*mmsize/(1<<%4)
156 mov%3 [ dst2q+(1<<%4)*lenq], m1
157 add lenq, mmsize/(1<<%4)
164 cglobal %2_to_%1_%3, 3, 3, 6, dst, src, len
169 jne %2_to_%1_u_int %+ SUFFIX
171 jne %2_to_%1_u_int %+ SUFFIX
173 %2_to_%1_u_int %+ SUFFIX
175 lea srcq , [srcq + (1<<%5)*lenq]
176 lea dstq , [dstq + (1<<%4)*lenq]
180 mov%3 m0, [ srcq +(1<<%5)*lenq]
181 mov%3 m1, [ mmsize + srcq +(1<<%5)*lenq]
183 mov%3 m2, [2*mmsize + srcq +(1<<%5)*lenq]
184 mov%3 m3, [3*mmsize + srcq +(1<<%5)*lenq]
187 mov%3 [ dstq+(1<<%4)*lenq], m0
188 mov%3 [ mmsize + dstq+(1<<%4)*lenq], m1
190 mov%3 [2*mmsize + dstq+(1<<%4)*lenq], m2
191 mov%3 [3*mmsize + dstq+(1<<%4)*lenq], m3
192 add lenq, 4*mmsize/(1<<%4)
194 add lenq, 2*mmsize/(1<<%4)
206 cglobal pack_6ch_%2_to_%1_%3, 2,8,7, dst, src, src1, src2, src3, src4, src5, len
210 %define lend dword r2m
212 mov src1q, [srcq+1*gprsize]
213 mov src2q, [srcq+2*gprsize]
214 mov src3q, [srcq+3*gprsize]
215 mov src4q, [srcq+4*gprsize]
216 mov src5q, [srcq+5*gprsize]
221 jne pack_6ch_%2_to_%1_u_int %+ SUFFIX
223 jne pack_6ch_%2_to_%1_u_int %+ SUFFIX
225 jne pack_6ch_%2_to_%1_u_int %+ SUFFIX
227 jne pack_6ch_%2_to_%1_u_int %+ SUFFIX
229 jne pack_6ch_%2_to_%1_u_int %+ SUFFIX
231 jne pack_6ch_%2_to_%1_u_int %+ SUFFIX
233 jne pack_6ch_%2_to_%1_u_int %+ SUFFIX
235 pack_6ch_%2_to_%1_u_int %+ SUFFIX
244 mov%3 m1, [srcq+src1q]
245 mov%3 m2, [srcq+src2q]
246 mov%3 m3, [srcq+src3q]
247 mov%3 m4, [srcq+src4q]
248 mov%3 m5, [srcq+src5q]
256 blendps m6, m4, m0, 1100b
265 blendps m2, m5, m1, 1100b
278 mov %+ %3 %+ ps [dstq ], m0
279 mov %+ %3 %+ ps [dstq+16], m6
280 mov %+ %3 %+ ps [dstq+32], m4
281 mov %+ %3 %+ ps [dstq+48], m1
282 mov %+ %3 %+ ps [dstq+64], m2
283 mov %+ %3 %+ ps [dstq+80], m5
285 SBUTTERFLY dq, 0, 1, 6
286 SBUTTERFLY dq, 2, 3, 6
287 SBUTTERFLY dq, 4, 5, 6
309 cglobal pack_8ch_%2_to_%1_%3, 2,10,10, ARCH_X86_32*32, dst, src, len, src1, src2, src3, src4, src5, src6, src7
312 DEFINE_ARGS dst, src, src2, src3, src4, src5, src6
313 %define lend dword r2m
320 mov src7q, [srcq+7*gprsize]
321 mov src6q, [srcq+6*gprsize]
322 mov src5q, [srcq+5*gprsize]
323 mov src4q, [srcq+4*gprsize]
327 mov src3q, [srcq+3*gprsize]
328 mov src2q, [srcq+2*gprsize]
329 mov src1q, [srcq+1*gprsize]
337 jne pack_8ch_%2_to_%1_u_int %+ SUFFIX
339 jne pack_8ch_%2_to_%1_u_int %+ SUFFIX
341 jne pack_8ch_%2_to_%1_u_int %+ SUFFIX
343 jne pack_8ch_%2_to_%1_u_int %+ SUFFIX
345 jne pack_8ch_%2_to_%1_u_int %+ SUFFIX
347 jne pack_8ch_%2_to_%1_u_int %+ SUFFIX
349 jne pack_8ch_%2_to_%1_u_int %+ SUFFIX
351 jne pack_8ch_%2_to_%1_u_int %+ SUFFIX
357 jne pack_8ch_%2_to_%1_u_int %+ SUFFIX
359 pack_8ch_%2_to_%1_u_int %+ SUFFIX
379 %define m9 [flt2pm31]
384 mov%3 m1, [srcq+src1q]
385 mov%3 m2, [srcq+src2q]
386 mov%3 m3, [srcq+src3q]
390 mov%3 m4, [srcq+src4q]
391 mov%3 m5, [srcq+src5q]
392 mov%3 m6, [srcq+src6q]
393 mov%3 m7, [srcq+src7q]
396 TRANSPOSE8x4D 0, 1, 2, 3, 4, 5, 6, 7, 8
407 TRANSPOSE8x4D 0, 1, 2, 3, 4, 5, 6, 7, [rsp], [rsp+16], 1
437 %macro INT16_TO_INT32_N 6
449 %macro INT32_TO_INT16_N 6
459 %macro INT32_TO_FLOAT_INIT 6
462 %macro INT32_TO_FLOAT_N 6
469 %macro FLOAT_TO_INT32_INIT 6
472 %macro FLOAT_TO_INT32_N 6
483 %macro INT16_TO_FLOAT_INIT 6
486 %macro INT16_TO_FLOAT_N 6
487 INT16_TO_INT32_N %1,%2,%3,%4,%5,%6
498 %macro FLOAT_TO_INT16_INIT 6
501 %macro FLOAT_TO_INT16_N 6
518 CONV int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N
519 CONV int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N
520 CONV int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N
521 CONV int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N
523 PACK_6CH float, float, u, 2, 2, NOP_N, NOP_N
524 PACK_6CH float, float, a, 2, 2, NOP_N, NOP_N
527 PACK_6CH float, float, u, 2, 2, NOP_N, NOP_N
528 PACK_6CH float, float, a, 2, 2, NOP_N, NOP_N
531 CONV int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N
532 CONV int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N
533 CONV int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N
534 CONV int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N
536 PACK_2CH int16, int16, u, 1, 1, NOP_N, NOP_N
537 PACK_2CH int16, int16, a, 1, 1, NOP_N, NOP_N
538 PACK_2CH int32, int32, u, 2, 2, NOP_N, NOP_N
539 PACK_2CH int32, int32, a, 2, 2, NOP_N, NOP_N
540 PACK_2CH int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N
541 PACK_2CH int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N
542 PACK_2CH int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N
543 PACK_2CH int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N
545 UNPACK_2CH int16, int16, u, 1, 1, NOP_N, NOP_N
546 UNPACK_2CH int16, int16, a, 1, 1, NOP_N, NOP_N
547 UNPACK_2CH int32, int32, u, 2, 2, NOP_N, NOP_N
548 UNPACK_2CH int32, int32, a, 2, 2, NOP_N, NOP_N
549 UNPACK_2CH int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N
550 UNPACK_2CH int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N
551 UNPACK_2CH int16, int32, u, 1, 2, INT32_TO_INT16_N, NOP_N
552 UNPACK_2CH int16, int32, a, 1, 2, INT32_TO_INT16_N, NOP_N
554 CONV float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
555 CONV float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
556 CONV int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
557 CONV int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
558 CONV float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
559 CONV float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
560 CONV int16, float, u, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT
561 CONV int16, float, a, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT
563 PACK_2CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
564 PACK_2CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
565 PACK_2CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
566 PACK_2CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
567 PACK_2CH float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
568 PACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
569 PACK_2CH int16, float, u, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT
570 PACK_2CH int16, float, a, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT
572 UNPACK_2CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
573 UNPACK_2CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
574 UNPACK_2CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
575 UNPACK_2CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
576 UNPACK_2CH float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
577 UNPACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
578 UNPACK_2CH int16, float, u, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT
579 UNPACK_2CH int16, float, a, 1, 2, FLOAT_TO_INT16_N, FLOAT_TO_INT16_INIT
581 PACK_6CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
582 PACK_6CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
583 PACK_6CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
584 PACK_6CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
586 %if HAVE_ALIGNED_STACK
587 PACK_8CH float, float, u, 2, 2, NOP_N, NOP_N
588 PACK_8CH float, float, a, 2, 2, NOP_N, NOP_N
590 PACK_8CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
591 PACK_8CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
592 PACK_8CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
593 PACK_8CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
597 UNPACK_2CH int16, int16, u, 1, 1, NOP_N, NOP_N
598 UNPACK_2CH int16, int16, a, 1, 1, NOP_N, NOP_N
599 UNPACK_2CH int32, int16, u, 2, 1, INT16_TO_INT32_N, NOP_N
600 UNPACK_2CH int32, int16, a, 2, 1, INT16_TO_INT32_N, NOP_N
601 UNPACK_2CH float, int16, u, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
602 UNPACK_2CH float, int16, a, 2, 1, INT16_TO_FLOAT_N, INT16_TO_FLOAT_INIT
604 %if HAVE_AVX_EXTERNAL
606 PACK_6CH float, float, u, 2, 2, NOP_N, NOP_N
607 PACK_6CH float, float, a, 2, 2, NOP_N, NOP_N
609 PACK_6CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
610 PACK_6CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
611 PACK_6CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
612 PACK_6CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
614 %if HAVE_ALIGNED_STACK
615 PACK_8CH float, float, u, 2, 2, NOP_N, NOP_N
616 PACK_8CH float, float, a, 2, 2, NOP_N, NOP_N
618 PACK_8CH float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
619 PACK_8CH float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
620 PACK_8CH int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
621 PACK_8CH int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
625 CONV float, int32, u, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
626 CONV float, int32, a, 2, 2, INT32_TO_FLOAT_N, INT32_TO_FLOAT_INIT
629 %if HAVE_AVX2_EXTERNAL
631 CONV int32, float, u, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT
632 CONV int32, float, a, 2, 2, FLOAT_TO_INT32_N, FLOAT_TO_INT32_INIT