1 ;******************************************************************************
2 ;* x86-optimized horizontal/vertical line scaling functions
3 ;* Copyright (c) 2011 Ronald S. Bultje <rsbultje@gmail.com>
4 ;* Kieran Kunhya <kieran@kunhya.com>
6 ;* This file is part of Libav.
8 ;* Libav is free software; you can redistribute it and/or
9 ;* modify it under the terms of the GNU Lesser General Public
10 ;* License as published by the Free Software Foundation; either
11 ;* version 2.1 of the License, or (at your option) any later version.
13 ;* Libav is distributed in the hope that it will be useful,
14 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
15 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 ;* Lesser General Public License for more details.
18 ;* You should have received a copy of the GNU Lesser General Public
19 ;* License along with Libav; if not, write to the Free Software
20 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 ;******************************************************************************
24 %include "x86util.asm"
28 max_19bit_int: times 4 dd 0x7ffff
29 max_19bit_flt: times 4 dd 524287.0
30 minshort: times 8 dw 0x8000
31 unicoeff: times 4 dd 0x20000000
32 yuv2yuvX_16_start: times 4 dd 0x4000 - 0x40000000
33 yuv2yuvX_10_start: times 4 dd 0x10000
34 yuv2yuvX_9_start: times 4 dd 0x20000
35 yuv2yuvX_10_upper: times 8 dw 0x3ff
36 yuv2yuvX_9_upper: times 8 dw 0x1ff
40 ;-----------------------------------------------------------------------------
41 ; horizontal line scaling
43 ; void hscale<source_width>to<intermediate_nbits>_<filterSize>_<opt>
44 ; (SwsContext *c, int{16,32}_t *dst,
45 ; int dstW, const uint{8,16}_t *src,
46 ; const int16_t *filter,
47 ; const int16_t *filterPos, int filterSize);
49 ; Scale one horizontal line. Input is either 8-bits width or 16-bits width
50 ; ($source_width can be either 8, 9, 10 or 16, difference is whether we have to
51 ; downscale before multiplying). Filter is 14-bits. Output is either 15bits
52 ; (in int16_t) or 19bits (in int32_t), as given in $intermediate_nbits. Each
53 ; output pixel is generated from $filterSize input pixels, the position of
54 ; the first pixel is given in filterPos[nOutputPixel].
55 ;-----------------------------------------------------------------------------
57 ; SCALE_FUNC source_width, intermediate_nbits, filtersize, filtersuffix, opt, n_args, n_xmm
59 cglobal hscale%1to%2_%4_%5, %6, 7, %7
65 mova m2, [max_19bit_int]
67 mova m2, [max_19bit_int]
69 mova m2, [max_19bit_flt]
70 %endif ; mmx/sse2/ssse3/sse4
93 shl r2, 1 ; this allows *16 (i.e. now *8) in lea instructions for the 8-tap filter
100 lea r1, [r1+r2*(2>>r2shr)]
102 lea r1, [r1+r2*(4>>r2shr)]
104 lea r5, [r5+r2*(2>>r2shr)]
108 %if %3 == 4 ; filterSize == 4 scaling
109 ; load 2x4 or 4x4 source pixels into m0/m1
110 movsx r0, word [r5+r2*2+0] ; filterPos[0]
111 movsx r6, word [r5+r2*2+2] ; filterPos[1]
112 movlh m0, [r3+r0*srcmul] ; src[filterPos[0] + {0,1,2,3}]
114 movlh m1, [r3+r6*srcmul] ; src[filterPos[1] + {0,1,2,3}]
117 movhps m0, [r3+r6*srcmul] ; src[filterPos[1] + {0,1,2,3}]
119 movd m4, [r3+r6*srcmul] ; src[filterPos[1] + {0,1,2,3}]
121 movsx r0, word [r5+r2*2+4] ; filterPos[2]
122 movsx r6, word [r5+r2*2+6] ; filterPos[3]
123 movlh m1, [r3+r0*srcmul] ; src[filterPos[2] + {0,1,2,3}]
125 movhps m1, [r3+r6*srcmul] ; src[filterPos[3] + {0,1,2,3}]
127 movd m5, [r3+r6*srcmul] ; src[filterPos[3] + {0,1,2,3}]
130 %endif ; %1 == 8 && %5 <= ssse
131 %endif ; mmsize == 8/16
133 punpcklbw m0, m3 ; byte -> word
134 punpcklbw m1, m3 ; byte -> word
137 ; multiply with filter coefficients
138 %if %1 == 16 ; pmaddwd needs signed adds, so this moves unsigned -> signed, we'll
139 ; add back 0x8000 * sum(coeffs) after the horizontal add
143 pmaddwd m0, [r4+r2*8+mmsize*0] ; *= filter[{0,1,..,6,7}]
144 pmaddwd m1, [r4+r2*8+mmsize*1] ; *= filter[{8,9,..,14,15}]
146 ; add up horizontally (4 srcpix * 4 coefficients -> 1 dstpix)
147 %if mmsize == 8 ; mmx
154 shufps m0, m1, 10001000b
155 shufps m4, m1, 11011101b
158 phaddd m0, m1 ; filter[{ 0, 1, 2, 3}]*src[filterPos[0]+{0,1,2,3}],
159 ; filter[{ 4, 5, 6, 7}]*src[filterPos[1]+{0,1,2,3}],
160 ; filter[{ 8, 9,10,11}]*src[filterPos[2]+{0,1,2,3}],
161 ; filter[{12,13,14,15}]*src[filterPos[3]+{0,1,2,3}]
162 %endif ; mmx/sse2/ssse3/sse4
163 %else ; %3 == 8, i.e. filterSize == 8 scaling
164 ; load 2x8 or 4x8 source pixels into m0, m1, m4 and m5
165 movsx r0, word [r5+r2*1+0] ; filterPos[0]
166 movsx r6, word [r5+r2*1+2] ; filterPos[1]
167 movbh m0, [r3+ r0 *srcmul] ; src[filterPos[0] + {0,1,2,3,4,5,6,7}]
169 movbh m1, [r3+(r0+4)*srcmul] ; src[filterPos[0] + {4,5,6,7}]
170 movbh m4, [r3+ r6 *srcmul] ; src[filterPos[1] + {0,1,2,3}]
171 movbh m5, [r3+(r6+4)*srcmul] ; src[filterPos[1] + {4,5,6,7}]
173 movbh m1, [r3+ r6 *srcmul] ; src[filterPos[1] + {0,1,2,3,4,5,6,7}]
174 movsx r0, word [r5+r2*1+4] ; filterPos[2]
175 movsx r6, word [r5+r2*1+6] ; filterPos[3]
176 movbh m4, [r3+ r0 *srcmul] ; src[filterPos[2] + {0,1,2,3,4,5,6,7}]
177 movbh m5, [r3+ r6 *srcmul] ; src[filterPos[3] + {0,1,2,3,4,5,6,7}]
178 %endif ; mmsize == 8/16
180 punpcklbw m0, m3 ; byte -> word
181 punpcklbw m1, m3 ; byte -> word
182 punpcklbw m4, m3 ; byte -> word
183 punpcklbw m5, m3 ; byte -> word
187 %if %1 == 16 ; pmaddwd needs signed adds, so this moves unsigned -> signed, we'll
188 ; add back 0x8000 * sum(coeffs) after the horizontal add
194 pmaddwd m0, [r4+r2*8+mmsize*0] ; *= filter[{0,1,..,6,7}]
195 pmaddwd m1, [r4+r2*8+mmsize*1] ; *= filter[{8,9,..,14,15}]
196 pmaddwd m4, [r4+r2*8+mmsize*2] ; *= filter[{16,17,..,22,23}]
197 pmaddwd m5, [r4+r2*8+mmsize*3] ; *= filter[{24,25,..,30,31}]
199 ; add up horizontally (8 srcpix * 8 coefficients -> 1 dstpix)
213 ; emulate horizontal add as transpose + vertical add
227 ; FIXME if we rearrange the filter in pairs of 4, we can
228 ; load pixels likewise and use 2 x paddd + phaddd instead
229 ; of 3 x phaddd here, faster on older cpus
232 phaddd m0, m4 ; filter[{ 0, 1,..., 6, 7}]*src[filterPos[0]+{0,1,...,6,7}],
233 ; filter[{ 8, 9,...,14,15}]*src[filterPos[1]+{0,1,...,6,7}],
234 ; filter[{16,17,...,22,23}]*src[filterPos[2]+{0,1,...,6,7}],
235 ; filter[{24,25,...,30,31}]*src[filterPos[3]+{0,1,...,6,7}]
236 %endif ; mmx/sse2/ssse3/sse4
239 %else ; %3 == X, i.e. any filterSize scaling
243 %else ; %4 == X || %4 == X8
248 movsxd r6, r6d ; filterSize
249 lea r12, [r3+(r6-r6sub)*srcmul] ; &src[filterSize&~4]
254 lea r0, [r3+(r6-r6sub)*srcmul] ; &src[filterSize&~4]
270 movsx r0, word [r5+r2*2+0] ; filterPos[0]
271 movsx r1x, word [r5+r2*2+2] ; filterPos[1]
272 ; FIXME maybe do 4px/iteration on x86-64 (x86-32 wouldn't have enough regs)?
278 ; load 2x4 (mmx) or 2x8 (sse) source pixels into m0/m1 -> m4/m5
279 movbh m0, [src_reg+r0 *srcmul] ; src[filterPos[0] + {0,1,2,3(,4,5,6,7)}]
280 movbh m1, [src_reg+(r1x+r6sub)*srcmul] ; src[filterPos[1] + {0,1,2,3(,4,5,6,7)}]
287 %if %1 == 16 ; pmaddwd needs signed adds, so this moves unsigned -> signed, we'll
288 ; add back 0x8000 * sum(coeffs) after the horizontal add
292 pmaddwd m0, [r4 ] ; filter[{0,1,2,3(,4,5,6,7)}]
293 pmaddwd m1, [r4+(r6+r6sub)*2] ; filter[filtersize+{0,1,2,3(,4,5,6,7)}]
297 add src_reg, srcmul*mmsize/2
298 cmp src_reg, filter2 ; while (src += 4) < &src[filterSize]
302 movsx r1x, word [r5+r2*2+2] ; filterPos[1]
303 movlh m0, [src_reg+r0 *srcmul] ; split last 4 srcpx of dstpx[0]
304 sub r1x, r6 ; and first 4 srcpx of dstpx[1]
306 movhps m0, [src_reg+(r1x+r6sub)*srcmul]
308 movd m1, [src_reg+(r1x+r6sub)*srcmul]
310 %endif ; %1 == 8 && %5 <= ssse
314 %if %1 == 16 ; pmaddwd needs signed adds, so this moves unsigned -> signed, we'll
315 ; add back 0x8000 * sum(coeffs) after the horizontal add
321 lea r4, [r4+(r6+r6sub)*2]
323 %if mmsize == 8 ; mmx
336 %endif ; sse2/ssse3/sse4
341 pshufd m4, m4, 11011000b
347 %endif ; sse2/ssse3/sse4
348 %endif ; mmsize == 8/16
351 %if %1 == 16 ; add 0x8000 * sum(coeffs), i.e. back from signed -> unsigned
356 psrad m0, 14 + %1 - %2
363 movh [r1+r2*(2>>r2shr)], m0
369 PMINSD_MMX m0, m2, m4
376 %endif ; mmx/sse2/ssse3/sse4
378 mova [r1+r2*(4>>r2shr)], m0
384 add r2, (mmsize<<r2shr)/4 ; both 8tap and 4tap really only do 4 pixels (or for mmx: 2 pixels)
385 ; per iteration. see "shl r2,1" above as for why we do this
402 ; SCALE_FUNCS source_width, intermediate_nbits, opt, n_xmm
404 SCALE_FUNC %1, %2, 4, 4, %3, 6, %4
405 SCALE_FUNC %1, %2, 8, 8, %3, 6, %4
407 SCALE_FUNC %1, %2, X, X, %3, 7, %4
409 SCALE_FUNC %1, %2, X, X4, %3, 7, %4
410 SCALE_FUNC %1, %2, X, X8, %3, 7, %4
414 ; SCALE_FUNCS2 opt, 8_xmm_args, 9to10_xmm_args, 16_xmm_args
415 %macro SCALE_FUNCS2 4
417 SCALE_FUNCS 8, 15, %1, %2
418 SCALE_FUNCS 9, 15, %1, %3
419 SCALE_FUNCS 10, 15, %1, %3
420 SCALE_FUNCS 14, 15, %1, %3
421 SCALE_FUNCS 16, 15, %1, %4
423 SCALE_FUNCS 8, 19, %1, %2
424 SCALE_FUNCS 9, 19, %1, %3
425 SCALE_FUNCS 10, 19, %1, %3
426 SCALE_FUNCS 14, 19, %1, %3
427 SCALE_FUNCS 16, 19, %1, %4
432 SCALE_FUNCS2 mmx, 0, 0, 0
435 SCALE_FUNCS2 sse2, 6, 7, 8
436 SCALE_FUNCS2 ssse3, 6, 6, 8
437 SCALE_FUNCS2 sse4, 6, 6, 8
439 ;-----------------------------------------------------------------------------
440 ; vertical line scaling
442 ; void yuv2plane1_<output_size>_<opt>(const int16_t *src, uint8_t *dst, int dstW,
443 ; const uint8_t *dither, int offset)
445 ; void yuv2planeX_<output_size>_<opt>(const int16_t *filter, int filterSize,
446 ; const int16_t **src, uint8_t *dst, int dstW,
447 ; const uint8_t *dither, int offset)
449 ; Scale one or $filterSize lines of source data to generate one line of output
450 ; data. The input is 15-bit in int16_t if $output_size is [8,10] and 19-bit in
451 ; int32_t if $output_size is 16. $filter is 12-bits. $filterSize is a multiple
452 ; of 2. $offset is either 0 or 3. $dither holds 8 values.
453 ;-----------------------------------------------------------------------------
455 %macro yuv2planeX_fn 4
465 cglobal yuv2planeX_%2_%1, %4, 7, %3
466 %if %2 == 8 || %2 == 9 || %2 == 10
468 %endif ; %2 == 8/9/10
472 %assign pad 0x2c - (stack_offset & 15)
479 ; create registers holding dither
480 movq m_dith, [r5] ; dither
484 punpcklqdq m_dith, m_dith
485 %endif ; mmsize == 16
486 PALIGNR m_dith, m_dith, 3, m0
491 punpcklwd m8, m_dith, m6
494 punpcklwd m5, m_dith, m6
501 mova [rsp+16], m_dith
504 punpcklbw m5, m_dith, m6
508 punpcklwd m3, m_dith, m6
517 mova [rsp+24], m_dith
518 %endif ; mmsize == 8/16
525 ; the rep here is for the 8bit output mmx case, where dither covers
526 ; 8 pixels but we can only handle 2 pixels per register, and thus 4
527 ; pixels per iteration. In order to not have to keep track of where
528 ; we are w.r.t. dithering, we unroll the mmx/8bit loop x2.
535 mova m2, [rsp+mmsize*(0+%%i)]
536 mova m1, [rsp+mmsize*(1+%%i)]
541 %else ; %2 == 9/10/16
542 mova m1, [yuv2yuvX_%2_start]
544 %endif ; %2 == 8/9/10/16
548 mov r6, [r2+gprsize*cntr_reg-2*gprsize]
551 mova m5, [r6+r5*4+mmsize]
554 %endif ; %2 == 8/9/10/16
555 mov r6, [r2+gprsize*cntr_reg-gprsize]
558 mova m6, [r6+r5*4+mmsize]
561 %endif ; %2 == 8/9/10/16
564 movd m0, [r0+2*cntr_reg-4]; coeff[0], coeff[1]
566 pshuflw m7, m0, 0 ; coeff[0]
567 pshuflw m0, m0, 0x55 ; coeff[1]
568 pmovsxwd m7, m7 ; word -> dword
569 pmovsxwd m0, m0 ; word -> dword
590 %endif ; %2 == 8/9/10/16
593 jg .filterloop_ %+ %%i
601 %endif ; %2 == 8/9/10/16
607 %else ; %2 == 9/10/16
619 %endif ; mmx2/sse2/sse4/avx
620 pminsw m2, [yuv2yuvX_%2_upper]
621 %endif ; %2 == 9/10/16
623 %endif ; %2 == 8/9/10/16
640 %else ; %2 == 9/10/16
642 %endif ; %2 == 8/9/10/16
645 %define PALIGNR PALIGNR_MMX
648 yuv2planeX_fn mmx, 8, 0, 7
649 yuv2planeX_fn mmx2, 9, 0, 5
650 yuv2planeX_fn mmx2, 10, 0, 5
654 yuv2planeX_fn sse2, 8, 10, 7
655 yuv2planeX_fn sse2, 9, 7, 5
656 yuv2planeX_fn sse2, 10, 7, 5
658 %define PALIGNR PALIGNR_SSSE3
659 yuv2planeX_fn sse4, 8, 10, 7
660 yuv2planeX_fn sse4, 9, 7, 5
661 yuv2planeX_fn sse4, 10, 7, 5
662 yuv2planeX_fn sse4, 16, 8, 5
665 yuv2planeX_fn avx, 8, 10, 7
666 yuv2planeX_fn avx, 9, 7, 5
667 yuv2planeX_fn avx, 10, 7, 5