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 TRANSPOSE4x4W 0, 1, 2, 3, 4
143 pxor m5, m5; zeros reg
159 ; in: 8 rows of 8 bytes in %1..%8
160 ; out: 8 rows of 8 words in m0..m7
161 %macro TRANSPOSE8x8B_LOAD 8
182 punpckldq m1, m3, m9; 0, 1
183 punpckhdq m3, m9; 2, 3
185 punpckldq m5, m7, m4; 4, 5
186 punpckhdq m7, m4; 6, 7
190 punpcklbw m0, m1, m13; 0 in 16 bit
191 punpckhbw m1, m13; 1 in 16 bit
193 punpcklbw m2, m3, m13; 2
196 punpcklbw m4, m5, m13; 4
199 punpcklbw m6, m7, m13; 6
204 ; in: 8 rows of 8 words in m0..m8
205 ; out: 8 rows of 8 bytes in %1..%8
206 %macro TRANSPOSE8x8B_STORE 8
228 punpckhdq m10, m0, m4; 2, 3
229 punpckldq m0, m4; 0, 1
231 punpckldq m11, m8, m9; 4, 5
232 punpckhdq m8, m9; 6, 7
243 ; in: 8 rows of 8 words in %1..%8
244 ; out: 8 rows of 8 words in m0..m7
245 %macro TRANSPOSE8x8W_LOAD 8
254 TRANSPOSE8x8W 0, 1, 2, 3, 4, 5, 6, 7, 8
257 ; in: 8 rows of 8 words in m0..m8
258 ; out: 8 rows of 8 words in %1..%8
259 %macro TRANSPOSE8x8W_STORE 9
260 TRANSPOSE8x8W 0, 1, 2, 3, 4, 5, 6, 7, 8
288 pand %2, m11 ; and mask
289 pandn m10, m11, %1; and -mask
296 ; mask in %3, will be clobbered
297 %macro MASKED_COPY2 3
298 pand %2, %3 ; and mask
299 pandn %3, %1; and -mask
305 ; input in m0 ... m3 and tcs in r2. Output in m1 and m2
306 %macro CHROMA_DEBLOCK_BODY 1
307 psubw m4, m2, m1; q0 - p0
308 psubw m5, m0, m3; p1 - q1
315 pshufd m6, m6, 0xA0; tc0, tc1
317 psignw m4, m6, [pw_m1]; -tc0, -tc1
319 pmullw m4, m6, [pw_m1]; -tc0, -tc1
327 psllw m4, %1-8; << (BIT_DEPTH - 8)
328 psllw m6, %1-8; << (BIT_DEPTH - 8)
332 paddw m1, m5; p0 + delta0
333 psubw m2, m5; q0 - delta0
336 ; input in m0 ... m7, beta in r2 tcs in r3. Output in m1...m6
337 %macro LUMA_DEBLOCK_BODY 2
341 ABS1 m10, m11 ; 0dp0, 0dp3 , 1dp0, 1dp3
346 ABS1 m11, m13 ; 0dq0, 0dq3 , 1dq0, 1dq3
354 ;end beta calculations
356 paddw m9, m10, m11; 0d0, 0d3 , 1d0, 1d3
358 pshufhw m14, m9, 0x0f ;0b00001111; 0d3 0d3 0d0 0d0 in high
359 pshuflw m14, m14, 0x0f ;0b00001111; 1d3 1d3 1d0 1d0 in low
361 pshufhw m9, m9, 0xf0 ;0b11110000; 0d0 0d0 0d3 0d3
362 pshuflw m9, m9, 0xf0 ;0b11110000; 1d0 1d0 1d3 1d3
364 paddw m14, m9; 0d0+0d3, 1d0+1d3
367 pcmpgtw m15, m13, m14
368 movmskps r13, m15 ;filtering mask 0d0 + 0d3 < beta0 (bit 2 or 3) , 1d0 + 1d3 < beta1 (bit 0 or 1)
372 ;weak / strong decision compare to beta_2
373 psraw m15, m13, 2; beta >> 2
375 pcmpgtw m15, m8; (d0 << 1) < beta_2, (d3 << 1) < beta_2
377 ;end weak / strong decision
379 ; weak filter nd_p/q calculation
384 and r7, 0xffff; 1dp0 + 1dp3
387 and r8, 0xffff; 0dp0 + 0dp3
393 and r9, 0xffff; 1dq0 + 1dq3
396 and r10, 0xffff; 0dq0 + 0dq3
397 ; end calc for weak filter
409 pcmpeqd m11, [pd_1]; filtering mask
411 ;decide between strong and weak filtering
422 add r11d, r3d; tc0 + tc1
427 shufps m8, m9, 0; tc0, tc1
430 pavgw m8, m9; tc25 = ((tc * 5 + 1) >> 1)
431 ;end tc25 calculations
433 ;----beta_3 comparison-----
434 psubw m12, m0, m3; p3 - p0
435 ABS1 m12, m14; abs(p3 - p0)
437 psubw m15, m7, m4; q3 - q0
438 ABS1 m15, m14; abs(q3 - q0)
440 paddw m12, m15; abs(p3 - p0) + abs(q3 - q0)
442 pshufhw m12, m12, 0xf0 ;0b11110000;
443 pshuflw m12, m12, 0xf0 ;0b11110000;
445 psraw m13, 3; beta >> 3
448 and r6, r11; strong mask , beta_2 and beta_3 comparisons
449 ;----beta_3 comparison end-----
450 ;----tc25 comparison---
451 psubw m12, m3, m4; p0 - q0
452 ABS1 m12, m14; abs(p0 - q0)
454 pshufhw m12, m12, 0xf0 ;0b11110000;
455 pshuflw m12, m12, 0xf0 ;0b11110000;
457 pcmpgtw m8, m12; tc25 comparisons
459 and r6, r11; strong mask, beta_2, beta_3 and tc25 comparisons
460 ;----tc25 comparison end---
463 and r6, r11; strong mask, bits 2 and 0
465 pmullw m14, m9, [pw_m2]; -tc * 2
469 mov r11, r6; strong mask
471 movd m12, r6d; store to xmm for mask generation
474 movd m10, r11d; store to xmm for mask generation
475 or r6, r11; final strong mask, bits 1 and 0
479 pcmpeqd m10, [pd_1]; strong mask
481 mova m13, [pw_4]; 4 in every cell
482 pand m11, m10; combine filtering mask and strong mask
483 paddw m12, m2, m3; p1 + p0
484 paddw m12, m4; p1 + p0 + q0
486 paddw m12, m12; 2*p1 + 2*p0 + 2*q0
487 paddw m12, m1; p2 + 2*p1 + 2*p0 + 2*q0
488 paddw m12, m5; p2 + 2*p1 + 2*p0 + 2*q0 + q1
489 paddw m12, m13; p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4
490 psraw m12, 3; ((p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4) >> 3)
491 psubw m12, m3; ((p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4) >> 3) - p0
493 pminsw m12, m9; av_clip( , -2 * tc, 2 * tc)
496 paddw m15, m1, m10; p2 + p1 + p0 + q0
497 psrlw m13, 1; 2 in every cell
498 paddw m15, m13; p2 + p1 + p0 + q0 + 2
499 psraw m15, 2; (p2 + p1 + p0 + q0 + 2) >> 2
500 psubw m15, m2;((p2 + p1 + p0 + q0 + 2) >> 2) - p1
502 pminsw m15, m9; av_clip( , -2 * tc, 2 * tc)
505 paddw m8, m1, m0; p3 + p2
506 paddw m8, m8; 2*p3 + 2*p2
507 paddw m8, m1; 2*p3 + 3*p2
508 paddw m8, m10; 2*p3 + 3*p2 + p1 + p0 + q0
510 paddw m8, m13; 2*p3 + 3*p2 + p1 + p0 + q0 + 4
511 psraw m8, 3; (2*p3 + 3*p2 + p1 + p0 + q0 + 4) >> 3
512 psubw m8, m1; ((2*p3 + 3*p2 + p1 + p0 + q0 + 4) >> 3) - p2
514 pminsw m8, m9; av_clip( , -2 * tc, 2 * tc)
518 paddw m8, m3, m4; p0 + q0
519 paddw m8, m5; p0 + q0 + q1
520 paddw m8, m8; 2*p0 + 2*q0 + 2*q1
521 paddw m8, m2; p1 + 2*p0 + 2*q0 + 2*q1
522 paddw m8, m6; p1 + 2*p0 + 2*q0 + 2*q1 + q2
523 paddw m8, m13; p1 + 2*p0 + 2*q0 + 2*q1 + q2 + 4
524 psraw m8, 3; (p1 + 2*p0 + 2*q0 + 2*q1 + q2 + 4) >>3
527 pminsw m8, m9; av_clip( , -2 * tc, 2 * tc)
531 paddw m15, m3, m4; p0 + q0
532 paddw m15, m5; p0 + q0 + q1
534 paddw m15, m6; p0 + q0 + q1 + q2
535 psrlw m13, 1; 2 in every cell
536 paddw m15, m13; p0 + q0 + q1 + q2 + 2
537 psraw m15, 2; (p0 + q0 + q1 + q2 + 2) >> 2
538 psubw m15, m5; ((p0 + q0 + q1 + q2 + 2) >> 2) - q1
540 pminsw m15, m9; av_clip( , -2 * tc, 2 * tc)
543 paddw m13, m7; q3 + 2
544 paddw m13, m6; q3 + q2 + 2
545 paddw m13, m13; 2*q3 + 2*q2 + 4
546 paddw m13, m6; 2*q3 + 3*q2 + 4
547 paddw m13, m10; 2*q3 + 3*q2 + q1 + q0 + p0 + 4
548 psraw m13, 3; (2*q3 + 3*q2 + q1 + q0 + p0 + 4) >> 3
549 psubw m13, m6; ((2*q3 + 3*q2 + q1 + q0 + p0 + 4) >> 3) - q2
551 pminsw m13, m9; av_clip( , -2 * tc, 2 * tc)
560 not r6; strong mask -> weak mask
561 and r6, r13; final weak filtering mask, bits 0 and 1
564 ; weak filtering mask
571 pcmpeqd m11, [pd_1]; filtering mask
576 shr betaq, 3; ((beta + (beta >> 1)) >> 3))
579 psubw m12, m4, m3 ; q0 - p0
580 psllw m10, m12, 3; 8 * (q0 - p0)
581 paddw m12, m10 ; 9 * (q0 - p0)
583 psubw m10, m5, m2 ; q1 - p1
584 psllw m8, m10, 1; 2 * ( q1 - p1 )
585 paddw m10, m8; 3 * ( q1 - p1 )
586 psubw m12, m10; 9 * (q0 - p0) - 3 * ( q1 - p1 )
588 psraw m12, 4; >> 4 , delta0
589 PABSW m13, m12; abs(delta0)
592 psllw m10, m9, 2; 8 * tc
593 paddw m10, m9; 10 * tc
597 psraw m9, 1; tc * 2 -> tc
598 psraw m14, 1; -tc * 2 -> -tc
601 pminsw m12, m9; av_clip(delta0, -tc, tc)
603 psraw m9, 1; tc -> tc / 2
605 psignw m14, m9, [pw_m1]; -tc / 2
607 pmullw m14, m9, [pw_m1]; -tc / 2
610 pavgw m15, m1, m3; (p2 + p0 + 1) >> 1
611 psubw m15, m2; ((p2 + p0 + 1) >> 1) - p1
612 paddw m15, m12; ((p2 + p0 + 1) >> 1) - p1 + delta0
613 psraw m15, 1; (((p2 + p0 + 1) >> 1) - p1 + delta0) >> 1
615 pminsw m15, m9; av_clip(deltap1, -tc/2, tc/2)
622 movd m13, r7d; 1dp0 + 1dp3
623 movd m8, r8d; 0dp0 + 0dp3
629 ;end beta calculations
630 MASKED_COPY2 m2, m15, m8; write p1'
632 pavgw m8, m6, m4; (q2 + q0 + 1) >> 1
633 psubw m8, m5; ((q2 + q0 + 1) >> 1) - q1
634 psubw m8, m12; ((q2 + q0 + 1) >> 1) - q1 - delta0)
635 psraw m8, 1; ((q2 + q0 + 1) >> 1) - q1 - delta0) >> 1
637 pminsw m8, m9; av_clip(deltaq1, -tc/2, tc/2)
644 shufps m13, m15, 0; dq0 + dq3
646 pcmpgtw m10, m13; compare to ((beta+(beta>>1))>>3)
648 MASKED_COPY2 m5, m8, m10; write q1'
650 paddw m15, m3, m12 ; p0 + delta0
653 psubw m8, m4, m12 ; q0 - delta0
657 ;-----------------------------------------------------------------------------
658 ; void ff_hevc_v_loop_filter_chroma(uint8_t *_pix, ptrdiff_t _stride, int *_tc,
659 ; uint8_t *_no_p, uint8_t *_no_q);
660 ;-----------------------------------------------------------------------------
661 %macro LOOP_FILTER_CHROMA 0
662 cglobal hevc_v_loop_filter_chroma_8, 3, 5, 7, pix, stride, tc, pix0, r3stride
664 lea r3strideq, [3*strideq]
667 TRANSPOSE4x8B_LOAD PASS8ROWS(pix0q, pixq, strideq, r3strideq)
668 CHROMA_DEBLOCK_BODY 8
669 TRANSPOSE8x4B_STORE PASS8ROWS(pix0q, pixq, strideq, r3strideq)
672 cglobal hevc_v_loop_filter_chroma_10, 3, 5, 7, pix, stride, tc, pix0, r3stride
674 lea r3strideq, [3*strideq]
677 TRANSPOSE4x8W_LOAD PASS8ROWS(pix0q, pixq, strideq, r3strideq)
678 CHROMA_DEBLOCK_BODY 10
679 TRANSPOSE8x4W_STORE PASS8ROWS(pix0q, pixq, strideq, r3strideq), [pw_pixel_max_10]
682 cglobal hevc_v_loop_filter_chroma_12, 3, 5, 7, pix, stride, tc, pix0, r3stride
684 lea r3strideq, [3*strideq]
687 TRANSPOSE4x8W_LOAD PASS8ROWS(pix0q, pixq, strideq, r3strideq)
688 CHROMA_DEBLOCK_BODY 12
689 TRANSPOSE8x4W_STORE PASS8ROWS(pix0q, pixq, strideq, r3strideq), [pw_pixel_max_12]
692 ;-----------------------------------------------------------------------------
693 ; void ff_hevc_h_loop_filter_chroma(uint8_t *_pix, ptrdiff_t _stride, int *_tc,
694 ; uint8_t *_no_p, uint8_t *_no_q);
695 ;-----------------------------------------------------------------------------
696 cglobal hevc_h_loop_filter_chroma_8, 3, 4, 7, pix, stride, tc, pix0
701 movq m1, [pix0q+strideq]; p0
703 movq m3, [pixq+strideq]; q1
704 pxor m5, m5; zeros reg
709 CHROMA_DEBLOCK_BODY 8
711 movh[pix0q+strideq], m1
715 cglobal hevc_h_loop_filter_chroma_10, 3, 4, 7, pix, stride, tc, pix0
720 movu m1, [pix0q+strideq]; p0
722 movu m3, [pixq+strideq]; q1
723 CHROMA_DEBLOCK_BODY 10
724 pxor m5, m5; zeros reg
725 CLIPW m1, m5, [pw_pixel_max_10]
726 CLIPW m2, m5, [pw_pixel_max_10]
727 movu [pix0q+strideq], m1
731 cglobal hevc_h_loop_filter_chroma_12, 3, 4, 7, pix, stride, tc, pix0
736 movu m1, [pix0q+strideq]; p0
738 movu m3, [pixq+strideq]; q1
739 CHROMA_DEBLOCK_BODY 12
740 pxor m5, m5; zeros reg
741 CLIPW m1, m5, [pw_pixel_max_12]
742 CLIPW m2, m5, [pw_pixel_max_12]
743 movu [pix0q+strideq], m1
754 %macro LOOP_FILTER_LUMA 0
755 ;-----------------------------------------------------------------------------
756 ; void ff_hevc_v_loop_filter_luma(uint8_t *_pix, ptrdiff_t _stride, int beta,
757 ; int *_tc, uint8_t *_no_p, uint8_t *_no_q);
758 ;-----------------------------------------------------------------------------
759 cglobal hevc_v_loop_filter_luma_8, 4, 14, 16, pix, stride, beta, tc, pix0, src3stride
762 mov src3strideq, pixq
764 TRANSPOSE8x8B_LOAD PASS8ROWS(src3strideq, pixq, r1, pix0q)
765 LUMA_DEBLOCK_BODY 8, v
767 TRANSPOSE8x8B_STORE PASS8ROWS(src3strideq, pixq, r1, pix0q)
771 cglobal hevc_v_loop_filter_luma_10, 4, 14, 16, pix, stride, beta, tc, pix0, src3stride
773 lea pix0q, [3 * strideq]
774 mov src3strideq, pixq
776 TRANSPOSE8x8W_LOAD PASS8ROWS(src3strideq, pixq, strideq, pix0q)
777 LUMA_DEBLOCK_BODY 10, v
779 TRANSPOSE8x8W_STORE PASS8ROWS(src3strideq, pixq, r1, pix0q), [pw_pixel_max_10]
783 cglobal hevc_v_loop_filter_luma_12, 4, 14, 16, pix, stride, beta, tc, pix0, src3stride
785 lea pix0q, [3 * strideq]
786 mov src3strideq, pixq
788 TRANSPOSE8x8W_LOAD PASS8ROWS(src3strideq, pixq, strideq, pix0q)
789 LUMA_DEBLOCK_BODY 12, v
791 TRANSPOSE8x8W_STORE PASS8ROWS(src3strideq, pixq, r1, pix0q), [pw_pixel_max_12]
795 ;-----------------------------------------------------------------------------
796 ; void ff_hevc_h_loop_filter_luma(uint8_t *_pix, ptrdiff_t _stride, int beta,
797 ; int *_tc, uint8_t *_no_p, uint8_t *_no_q);
798 ;-----------------------------------------------------------------------------
799 cglobal hevc_h_loop_filter_luma_8, 4, 14, 16, pix, stride, beta, tc, pix0, src3stride
800 lea src3strideq, [3 * strideq]
802 sub pix0q, src3strideq
805 movq m1, [pix0q + strideq]; p2
806 movq m2, [pix0q + 2 * strideq]; p1
807 movq m3, [pix0q + src3strideq]; p0
809 movq m5, [pixq + strideq]; q1
810 movq m6, [pixq + 2 * strideq]; q2
811 movq m7, [pixq + src3strideq]; q3
821 LUMA_DEBLOCK_BODY 8, h
826 movh [pix0q + strideq], m1
827 movhps [pix0q + 2 * strideq], m1
828 movh [pix0q + src3strideq], m3
830 movh [pixq + strideq], m5
831 movhps [pixq + 2 * strideq], m5
835 cglobal hevc_h_loop_filter_luma_10, 4, 14, 16, pix, stride, beta, tc, pix0, src3stride
836 lea src3strideq, [3 * strideq]
838 sub pix0q, src3strideq
840 movdqu m0, [pix0q]; p3
841 movdqu m1, [pix0q + strideq]; p2
842 movdqu m2, [pix0q + 2 * strideq]; p1
843 movdqu m3, [pix0q + src3strideq]; p0
844 movdqu m4, [pixq]; q0
845 movdqu m5, [pixq + strideq]; q1
846 movdqu m6, [pixq + 2 * strideq]; q2
847 movdqu m7, [pixq + src3strideq]; q3
848 LUMA_DEBLOCK_BODY 10, h
850 pxor m8, m8; zeros reg
851 CLIPW m1, m8, [pw_pixel_max_10]
852 CLIPW m2, m8, [pw_pixel_max_10]
853 CLIPW m3, m8, [pw_pixel_max_10]
854 CLIPW m4, m8, [pw_pixel_max_10]
855 CLIPW m5, m8, [pw_pixel_max_10]
856 CLIPW m6, m8, [pw_pixel_max_10]
857 movdqu [pix0q + strideq], m1; p2
858 movdqu [pix0q + 2 * strideq], m2; p1
859 movdqu [pix0q + src3strideq], m3; p0
860 movdqu [pixq ], m4; q0
861 movdqu [pixq + strideq], m5; q1
862 movdqu [pixq + 2 * strideq], m6; q2
866 cglobal hevc_h_loop_filter_luma_12, 4, 14, 16, pix, stride, beta, tc, pix0, src3stride
867 lea src3strideq, [3 * strideq]
869 sub pix0q, src3strideq
871 movdqu m0, [pix0q]; p3
872 movdqu m1, [pix0q + strideq]; p2
873 movdqu m2, [pix0q + 2 * strideq]; p1
874 movdqu m3, [pix0q + src3strideq]; p0
875 movdqu m4, [pixq]; q0
876 movdqu m5, [pixq + strideq]; q1
877 movdqu m6, [pixq + 2 * strideq]; q2
878 movdqu m7, [pixq + src3strideq]; q3
879 LUMA_DEBLOCK_BODY 12, h
881 pxor m8, m8; zeros reg
882 CLIPW m1, m8, [pw_pixel_max_12]
883 CLIPW m2, m8, [pw_pixel_max_12]
884 CLIPW m3, m8, [pw_pixel_max_12]
885 CLIPW m4, m8, [pw_pixel_max_12]
886 CLIPW m5, m8, [pw_pixel_max_12]
887 CLIPW m6, m8, [pw_pixel_max_12]
888 movdqu [pix0q + strideq], m1; p2
889 movdqu [pix0q + 2 * strideq], m2; p1
890 movdqu [pix0q + src3strideq], m3; p0
891 movdqu [pixq ], m4; q0
892 movdqu [pixq + strideq], m5; q1
893 movdqu [pixq + 2 * strideq], m6; q2