1 ;*****************************************************************************
2 ;* SSE2-optimized HEVC deblocking code
3 ;*****************************************************************************
4 ;* Copyright (C) 2013 VTT
6 ;* Authors: Seppo Tomperi <seppo.tomperi@vtt.fi>
8 ;* This file is part of FFmpeg.
10 ;* FFmpeg is free software; you can redistribute it and/or
11 ;* modify it under the terms of the GNU Lesser General Public
12 ;* License as published by the Free Software Foundation; either
13 ;* version 2.1 of the License, or (at your option) any later version.
15 ;* FFmpeg 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 GNU
18 ;* Lesser General Public License for more details.
20 ;* You should have received a copy of the GNU Lesser General Public
21 ;* License along with FFmpeg; if not, write to the Free Software
22 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 ;******************************************************************************
25 %include "libavutil/x86/x86util.asm"
29 pw_pixel_max_12: times 8 dw ((1 << 12)-1)
30 pw_pixel_max_10: times 8 dw ((1 << 10)-1)
41 ; expands to [base],...,[base+7*stride]
42 %define PASS8ROWS(base, base3, stride, stride3) \
43 [base], [base+stride], [base+stride*2], [base3], \
44 [base3+stride], [base3+stride*2], [base3+stride3], [base3+stride*4]
46 ; in: 8 rows of 4 bytes in %4..%11
47 ; out: 4 rows of 8 words in m0..m3
48 %macro TRANSPOSE4x8B_LOAD 8
77 ; in: 4 rows of 8 words in m0..m3
78 ; out: 8 rows of 4 bytes in %1..%8
79 %macro TRANSPOSE8x4B_STORE 8
108 ; in: 8 rows of 4 words in %4..%11
109 ; out: 4 rows of 8 words in m0..m3
110 %macro TRANSPOSE4x8W_LOAD 8
131 punpckhqdq m1, m0, m4
133 punpckhqdq m3, m2, m6
138 ; in: 4 rows of 8 words in m0..m3
139 ; out: 8 rows of 4 words in %1..%8
140 %macro TRANSPOSE8x4W_STORE 9
141 pxor m5, m5; zeros reg
168 ; in: 8 rows of 8 bytes in %1..%8
169 ; out: 8 rows of 8 words in m0..m7
170 %macro TRANSPOSE8x8B_LOAD 8
191 punpckldq m1, m3, m9; 0, 1
192 punpckhdq m3, m9; 2, 3
194 punpckldq m5, m7, m4; 4, 5
195 punpckhdq m7, m4; 6, 7
199 punpcklbw m0, m1, m13; 0 in 16 bit
200 punpckhbw m1, m13; 1 in 16 bit
202 punpcklbw m2, m3, m13; 2
205 punpcklbw m4, m5, m13; 4
208 punpcklbw m6, m7, m13; 6
213 ; in: 8 rows of 8 words in m0..m8
214 ; out: 8 rows of 8 bytes in %1..%8
215 %macro TRANSPOSE8x8B_STORE 8
237 punpckhdq m10, m0, m4; 2, 3
238 punpckldq m0, m4; 0, 1
240 punpckldq m11, m8, m9; 4, 5
241 punpckhdq m8, m9; 6, 7
252 ; in: 8 rows of 8 words in %1..%8
253 ; out: 8 rows of 8 words in m0..m7
254 %macro TRANSPOSE8x8W_LOAD 8
263 TRANSPOSE8x8W 0, 1, 2, 3, 4, 5, 6, 7, 8
266 ; in: 8 rows of 8 words in m0..m8
267 ; out: 8 rows of 8 words in %1..%8
268 %macro TRANSPOSE8x8W_STORE 9
269 TRANSPOSE8x8W 0, 1, 2, 3, 4, 5, 6, 7, 8
297 pand %2, m11 ; and mask
298 pandn m10, m11, %1; and -mask
305 ; mask in %3, will be clobbered
306 %macro MASKED_COPY2 3
307 pand %2, %3 ; and mask
308 pandn %3, %1; and -mask
314 ; input in m0 ... m3 and tcs in r2. Output in m1 and m2
315 %macro CHROMA_DEBLOCK_BODY 1
316 psubw m4, m2, m1; q0 - p0
317 psubw m5, m0, m3; p1 - q1
324 movd m4, [tcq+4]; tc1
326 shufps m6, m4, 0; tc0, tc1
327 pmullw m4, m6, [pw_m1]; -tc0, -tc1
334 psllw m4, %1-8; << (BIT_DEPTH - 8)
335 psllw m6, %1-8; << (BIT_DEPTH - 8)
339 paddw m1, m5; p0 + delta0
340 psubw m2, m5; q0 - delta0
343 ; input in m0 ... m7, beta in r2 tcs in r3. Output in m1...m6
344 %macro LUMA_DEBLOCK_BODY 2
348 ABS1 m10, m11 ; 0dp0, 0dp3 , 1dp0, 1dp3
353 ABS1 m11, m13 ; 0dq0, 0dq3 , 1dq0, 1dq3
361 ;end beta calculations
363 paddw m9, m10, m11; 0d0, 0d3 , 1d0, 1d3
365 pshufhw m14, m9, 0x0f ;0b00001111; 0d3 0d3 0d0 0d0 in high
366 pshuflw m14, m14, 0x0f ;0b00001111; 1d3 1d3 1d0 1d0 in low
368 pshufhw m9, m9, 0xf0 ;0b11110000; 0d0 0d0 0d3 0d3
369 pshuflw m9, m9, 0xf0 ;0b11110000; 1d0 1d0 1d3 1d3
371 paddw m14, m9; 0d0+0d3, 1d0+1d3
374 pcmpgtw m15, m13, m14
375 movmskps r13, m15 ;filtering mask 0d0 + 0d3 < beta0 (bit 2 or 3) , 1d0 + 1d3 < beta1 (bit 0 or 1)
379 ;weak / strong decision compare to beta_2
380 psraw m15, m13, 2; beta >> 2
382 pcmpgtw m15, m8; (d0 << 1) < beta_2, (d3 << 1) < beta_2
384 ;end weak / strong decision
386 ; weak filter nd_p/q calculation
391 and r7, 0xffff; 1dp0 + 1dp3
394 and r8, 0xffff; 0dp0 + 0dp3
400 and r9, 0xffff; 1dq0 + 1dq3
403 and r10, 0xffff; 0dq0 + 0dq3
404 ; end calc for weak filter
416 pcmpeqd m11, [pd_1]; filtering mask
418 ;decide between strong and weak filtering
431 add r11d, r3d; tc0 + tc1
435 shufps m8, m9, 0; tc0, tc1
438 pavgw m8, m9; tc25 = ((tc * 5 + 1) >> 1)
439 ;end tc25 calculations
441 ;----beta_3 comparison-----
442 psubw m12, m0, m3; p3 - p0
443 ABS1 m12, m14; abs(p3 - p0)
445 psubw m15, m7, m4; q3 - q0
446 ABS1 m15, m14; abs(q3 - q0)
448 paddw m12, m15; abs(p3 - p0) + abs(q3 - q0)
450 pshufhw m12, m12, 0xf0 ;0b11110000;
451 pshuflw m12, m12, 0xf0 ;0b11110000;
453 psraw m13, 3; beta >> 3
456 and r14, r11; strong mask , beta_2 and beta_3 comparisons
457 ;----beta_3 comparison end-----
458 ;----tc25 comparison---
459 psubw m12, m3, m4; p0 - q0
460 ABS1 m12, m14; abs(p0 - q0)
462 pshufhw m12, m12, 0xf0 ;0b11110000;
463 pshuflw m12, m12, 0xf0 ;0b11110000;
465 pcmpgtw m8, m12; tc25 comparisons
467 and r14, r11; strong mask, beta_2, beta_3 and tc25 comparisons
468 ;----tc25 comparison end---
471 and r14, r11; strong mask, bits 2 and 0
473 pmullw m14, m9, [pw_m2]; -tc * 2
477 mov r11, r14; strong mask
479 movd m12, r14d; store to xmm for mask generation
482 movd m10, r11d; store to xmm for mask generation
483 or r14, r11; final strong mask, bits 1 and 0
487 pcmpeqd m10, [pd_1]; strong mask
489 mova m13, [pw_4]; 4 in every cell
490 pand m11, m10; combine filtering mask and strong mask
491 paddw m12, m2, m3; p1 + p0
492 paddw m12, m4; p1 + p0 + q0
494 paddw m12, m12; 2*p1 + 2*p0 + 2*q0
495 paddw m12, m1; p2 + 2*p1 + 2*p0 + 2*q0
496 paddw m12, m5; p2 + 2*p1 + 2*p0 + 2*q0 + q1
497 paddw m12, m13; p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4
498 psraw m12, 3; ((p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4) >> 3)
499 psubw m12, m3; ((p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4) >> 3) - p0
501 pminsw m12, m9; av_clip( , -2 * tc, 2 * tc)
504 paddw m15, m1, m10; p2 + p1 + p0 + q0
505 psrlw m13, 1; 2 in every cell
506 paddw m15, m13; p2 + p1 + p0 + q0 + 2
507 psraw m15, 2; (p2 + p1 + p0 + q0 + 2) >> 2
508 psubw m15, m2;((p2 + p1 + p0 + q0 + 2) >> 2) - p1
510 pminsw m15, m9; av_clip( , -2 * tc, 2 * tc)
513 paddw m8, m1, m0; p3 + p2
514 paddw m8, m8; 2*p3 + 2*p2
515 paddw m8, m1; 2*p3 + 3*p2
516 paddw m8, m10; 2*p3 + 3*p2 + p1 + p0 + q0
518 paddw m8, m13; 2*p3 + 3*p2 + p1 + p0 + q0 + 4
519 psraw m8, 3; (2*p3 + 3*p2 + p1 + p0 + q0 + 4) >> 3
520 psubw m8, m1; ((2*p3 + 3*p2 + p1 + p0 + q0 + 4) >> 3) - p2
522 pminsw m8, m9; av_clip( , -2 * tc, 2 * tc)
526 paddw m8, m3, m4; p0 + q0
527 paddw m8, m5; p0 + q0 + q1
528 paddw m8, m8; 2*p0 + 2*q0 + 2*q1
529 paddw m8, m2; p1 + 2*p0 + 2*q0 + 2*q1
530 paddw m8, m6; p1 + 2*p0 + 2*q0 + 2*q1 + q2
531 paddw m8, m13; p1 + 2*p0 + 2*q0 + 2*q1 + q2 + 4
532 psraw m8, 3; (p1 + 2*p0 + 2*q0 + 2*q1 + q2 + 4) >>3
535 pminsw m8, m9; av_clip( , -2 * tc, 2 * tc)
539 paddw m15, m3, m4; p0 + q0
540 paddw m15, m5; p0 + q0 + q1
542 paddw m15, m6; p0 + q0 + q1 + q2
543 psrlw m13, 1; 2 in every cell
544 paddw m15, m13; p0 + q0 + q1 + q2 + 2
545 psraw m15, 2; (p0 + q0 + q1 + q2 + 2) >> 2
546 psubw m15, m5; ((p0 + q0 + q1 + q2 + 2) >> 2) - q1
548 pminsw m15, m9; av_clip( , -2 * tc, 2 * tc)
551 paddw m13, m7; q3 + 2
552 paddw m13, m6; q3 + q2 + 2
553 paddw m13, m13; 2*q3 + 2*q2 + 4
554 paddw m13, m6; 2*q3 + 3*q2 + 4
555 paddw m13, m10; 2*q3 + 3*q2 + q1 + q0 + p0 + 4
556 psraw m13, 3; (2*q3 + 3*q2 + q1 + q0 + p0 + 4) >> 3
557 psubw m13, m6; ((2*q3 + 3*q2 + q1 + q0 + p0 + 4) >> 3) - q2
559 pminsw m13, m9; av_clip( , -2 * tc, 2 * tc)
568 not r14; strong mask -> weak mask
569 and r14, r13; final weak filtering mask, bits 0 and 1
572 ; weak filtering mask
579 pcmpeqd m11, [pd_1]; filtering mask
584 shr betaq, 3; ((beta + (beta >> 1)) >> 3))
587 psubw m12, m4, m3 ; q0 - p0
588 psllw m10, m12, 3; 8 * (q0 - p0)
589 paddw m12, m10 ; 9 * (q0 - p0)
591 psubw m10, m5, m2 ; q1 - p1
592 psllw m8, m10, 1; 2 * ( q1 - p1 )
593 paddw m10, m8; 3 * ( q1 - p1 )
594 psubw m12, m10; 9 * (q0 - p0) - 3 * ( q1 - p1 )
596 psraw m12, 4; >> 4 , delta0
597 PABSW m13, m12; abs(delta0)
600 psllw m10, m9, 2; 8 * tc
601 paddw m10, m9; 10 * tc
605 psraw m9, 1; tc * 2 -> tc
606 psraw m14, 1; -tc * 2 -> -tc
609 pminsw m12, m9; av_clip(delta0, -tc, tc)
611 psraw m9, 1; tc -> tc / 2
612 pmullw m14, m9, [pw_m1]; -tc / 2
614 pavgw m15, m1, m3; (p2 + p0 + 1) >> 1
615 psubw m15, m2; ((p2 + p0 + 1) >> 1) - p1
616 paddw m15, m12; ((p2 + p0 + 1) >> 1) - p1 + delta0
617 psraw m15, 1; (((p2 + p0 + 1) >> 1) - p1 + delta0) >> 1
619 pminsw m15, m9; av_clip(deltap1, -tc/2, tc/2)
626 movd m13, r7d; 1dp0 + 1dp3
627 movd m8, r8d; 0dp0 + 0dp3
633 ;end beta calculations
634 MASKED_COPY2 m2, m15, m8; write p1'
636 pavgw m8, m6, m4; (q2 + q0 + 1) >> 1
637 psubw m8, m5; ((q2 + q0 + 1) >> 1) - q1
638 psubw m8, m12; ((q2 + q0 + 1) >> 1) - q1 - delta0)
639 psraw m8, 1; ((q2 + q0 + 1) >> 1) - q1 - delta0) >> 1
641 pminsw m8, m9; av_clip(deltaq1, -tc/2, tc/2)
648 shufps m13, m15, 0; dq0 + dq3
650 pcmpgtw m10, m13; compare to ((beta+(beta>>1))>>3)
652 MASKED_COPY2 m5, m8, m10; write q1'
654 paddw m15, m3, m12 ; p0 + delta0
657 psubw m8, m4, m12 ; q0 - delta0
662 ;-----------------------------------------------------------------------------
663 ; void ff_hevc_v_loop_filter_chroma(uint8_t *_pix, ptrdiff_t _stride, int *_tc,
664 ; uint8_t *_no_p, uint8_t *_no_q);
665 ;-----------------------------------------------------------------------------
666 cglobal hevc_v_loop_filter_chroma_8, 3, 5, 7, pix, stride, tc, pix0, r3stride
668 lea r3strideq, [3*strideq]
671 TRANSPOSE4x8B_LOAD PASS8ROWS(pix0q, pixq, strideq, r3strideq)
672 CHROMA_DEBLOCK_BODY 8
673 TRANSPOSE8x4B_STORE PASS8ROWS(pix0q, pixq, strideq, r3strideq)
676 cglobal hevc_v_loop_filter_chroma_10, 3, 5, 7, pix, stride, tc, pix0, r3stride
678 lea r3strideq, [3*strideq]
681 TRANSPOSE4x8W_LOAD PASS8ROWS(pix0q, pixq, strideq, r3strideq)
682 CHROMA_DEBLOCK_BODY 10
683 TRANSPOSE8x4W_STORE PASS8ROWS(pix0q, pixq, strideq, r3strideq), [pw_pixel_max_10]
686 cglobal hevc_v_loop_filter_chroma_12, 3, 5, 7, pix, stride, tc, pix0, r3stride
688 lea r3strideq, [3*strideq]
691 TRANSPOSE4x8W_LOAD PASS8ROWS(pix0q, pixq, strideq, r3strideq)
692 CHROMA_DEBLOCK_BODY 12
693 TRANSPOSE8x4W_STORE PASS8ROWS(pix0q, pixq, strideq, r3strideq), [pw_pixel_max_12]
696 ;-----------------------------------------------------------------------------
697 ; void ff_hevc_h_loop_filter_chroma(uint8_t *_pix, ptrdiff_t _stride, int *_tc,
698 ; uint8_t *_no_p, uint8_t *_no_q);
699 ;-----------------------------------------------------------------------------
700 cglobal hevc_h_loop_filter_chroma_8, 3, 4, 7, pix, stride, tc, pix0
705 movq m1, [pix0q+strideq]; p0
707 movq m3, [pixq+strideq]; q1
708 pxor m5, m5; zeros reg
713 CHROMA_DEBLOCK_BODY 8
715 movh[pix0q+strideq], m1
719 cglobal hevc_h_loop_filter_chroma_10, 3, 4, 7, pix, stride, tc, pix0
724 movu m1, [pix0q+strideq]; p0
726 movu m3, [pixq+strideq]; q1
727 CHROMA_DEBLOCK_BODY 10
728 pxor m5, m5; zeros reg
729 CLIPW m1, m5, [pw_pixel_max_10]
730 CLIPW m2, m5, [pw_pixel_max_10]
731 movu [pix0q+strideq], m1
735 cglobal hevc_h_loop_filter_chroma_12, 3, 4, 7, pix, stride, tc, pix0
740 movu m1, [pix0q+strideq]; p0
742 movu m3, [pixq+strideq]; q1
743 CHROMA_DEBLOCK_BODY 12
744 pxor m5, m5; zeros reg
745 CLIPW m1, m5, [pw_pixel_max_12]
746 CLIPW m2, m5, [pw_pixel_max_12]
747 movu [pix0q+strideq], m1
752 %macro LOOP_FILTER_LUMA 0
753 ;-----------------------------------------------------------------------------
754 ; void ff_hevc_v_loop_filter_luma(uint8_t *_pix, ptrdiff_t _stride, int beta,
755 ; int *_tc, uint8_t *_no_p, uint8_t *_no_q);
756 ;-----------------------------------------------------------------------------
757 cglobal hevc_v_loop_filter_luma_8, 4, 15, 16, pix, stride, beta, tc
762 TRANSPOSE8x8B_LOAD PASS8ROWS(r6, r0, r1, r5)
763 LUMA_DEBLOCK_BODY 8, v
765 TRANSPOSE8x8B_STORE PASS8ROWS(r6, r0, r1, r5)
769 cglobal hevc_v_loop_filter_luma_10, 4, 15, 16, pix, stride, beta, tc
771 lea r5, [3 * strideq]
774 TRANSPOSE8x8W_LOAD PASS8ROWS(r6, pixq, strideq, r5)
775 LUMA_DEBLOCK_BODY 10, v
777 TRANSPOSE8x8W_STORE PASS8ROWS(r6, r0, r1, r5), [pw_pixel_max_10]
781 cglobal hevc_v_loop_filter_luma_12, 4, 15, 16, pix, stride, beta, tc
783 lea r5, [3 * strideq]
786 TRANSPOSE8x8W_LOAD PASS8ROWS(r6, pixq, strideq, r5)
787 LUMA_DEBLOCK_BODY 12, v
789 TRANSPOSE8x8W_STORE PASS8ROWS(r6, r0, r1, r5), [pw_pixel_max_12]
793 ;-----------------------------------------------------------------------------
794 ; void ff_hevc_h_loop_filter_luma(uint8_t *_pix, ptrdiff_t _stride, int beta,
795 ; int *_tc, uint8_t *_no_p, uint8_t *_no_q);
796 ;-----------------------------------------------------------------------------
797 cglobal hevc_h_loop_filter_luma_8, 4, 15, 16, pix, stride, beta, tc, count, pix0, src3stride
798 lea src3strideq, [3 * strideq]
800 sub pix0q, src3strideq
803 movq m1, [pix0q + strideq]; p2
804 movq m2, [pix0q + 2 * strideq]; p1
805 movq m3, [pix0q + src3strideq]; p0
807 movq m5, [pixq + strideq]; q1
808 movq m6, [pixq + 2 * strideq]; q2
809 movq m7, [pixq + src3strideq]; q3
819 LUMA_DEBLOCK_BODY 8, h
825 movhps [r5 + 2 * r1], m1
829 movhps [r0 + 2 * r1], m5
833 cglobal hevc_h_loop_filter_luma_10, 4, 15, 16, pix, stride, beta, tc, count, pix0, src3stride
834 lea src3strideq, [3 * strideq]
836 sub pix0q, src3strideq
838 movdqu m0, [pix0q]; p3
839 movdqu m1, [pix0q + strideq]; p2
840 movdqu m2, [pix0q + 2 * strideq]; p1
841 movdqu m3, [pix0q + src3strideq]; p0
842 movdqu m4, [pixq]; q0
843 movdqu m5, [pixq + strideq]; q1
844 movdqu m6, [pixq + 2 * strideq]; q2
845 movdqu m7, [pixq + src3strideq]; q3
846 LUMA_DEBLOCK_BODY 10, h
848 pxor m8, m8; zeros reg
849 CLIPW m1, m8, [pw_pixel_max_10]
850 CLIPW m2, m8, [pw_pixel_max_10]
851 CLIPW m3, m8, [pw_pixel_max_10]
852 CLIPW m4, m8, [pw_pixel_max_10]
853 CLIPW m5, m8, [pw_pixel_max_10]
854 CLIPW m6, m8, [pw_pixel_max_10]
855 movdqu [pix0q + strideq], m1; p2
856 movdqu [pix0q + 2 * strideq], m2; p1
857 movdqu [pix0q + src3strideq], m3; p0
858 movdqu [pixq ], m4; q0
859 movdqu [pixq + strideq], m5; q1
860 movdqu [pixq + 2 * strideq], m6; q2
864 cglobal hevc_h_loop_filter_luma_12, 4, 15, 16, pix, stride, beta, tc, count, pix0, src3stride
865 lea src3strideq, [3 * strideq]
867 sub pix0q, src3strideq
869 movdqu m0, [pix0q]; p3
870 movdqu m1, [pix0q + strideq]; p2
871 movdqu m2, [pix0q + 2 * strideq]; p1
872 movdqu m3, [pix0q + src3strideq]; p0
873 movdqu m4, [pixq]; q0
874 movdqu m5, [pixq + strideq]; q1
875 movdqu m6, [pixq + 2 * strideq]; q2
876 movdqu m7, [pixq + src3strideq]; q3
877 LUMA_DEBLOCK_BODY 12, h
879 pxor m8, m8; zeros reg
880 CLIPW m1, m8, [pw_pixel_max_12]
881 CLIPW m2, m8, [pw_pixel_max_12]
882 CLIPW m3, m8, [pw_pixel_max_12]
883 CLIPW m4, m8, [pw_pixel_max_12]
884 CLIPW m5, m8, [pw_pixel_max_12]
885 CLIPW m6, m8, [pw_pixel_max_12]
886 movdqu [pix0q + strideq], m1; p2
887 movdqu [pix0q + 2 * strideq], m2; p1
888 movdqu [pix0q + src3strideq], m3; p0
889 movdqu [pixq ], m4; q0
890 movdqu [pixq + strideq], m5; q1
891 movdqu [pixq + 2 * strideq], m6; q2