Mercurial > libavcodec.hg

comparison x86/vp8dsp.asm @ 12082:8527154f6e81 libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
SSSE3 versions of vp8 width4 bilinear MC functions
author darkshikari
date Sat, 03 Jul 2010 00:48:12 +0000
parents b8f80fe02861
children d780ae746855
comparison
equal deleted inserted replaced
12081:812e23197d64 12082:8527154f6e81
768 INIT_MMX 768 INIT_MMX
769 FILTER_BILINEAR mmxext, 4, 0 769 FILTER_BILINEAR mmxext, 4, 0
770 INIT_XMM 770 INIT_XMM
771 FILTER_BILINEAR sse2, 8, 7 771 FILTER_BILINEAR sse2, 8, 7
772 772
773 cglobal put_vp8_bilinear8_v_ssse3, 7,7,5 773 %macro FILTER_BILINEAR_SSSE3 1
774 cglobal put_vp8_bilinear%1_v_ssse3, 7,7
774 shl r6d, 4 775 shl r6d, 4
775 %ifdef PIC 776 %ifdef PIC
776 lea r11, [bilinear_filter_vb_m] 777 lea r11, [bilinear_filter_vb_m]
777 %endif 778 %endif
778 pxor m4, m4 779 pxor m4, m4
787 pmaddubsw m1, m3 788 pmaddubsw m1, m3
788 psraw m0, 2 789 psraw m0, 2
789 psraw m1, 2 790 psraw m1, 2
790 pavgw m0, m4 791 pavgw m0, m4
791 pavgw m1, m4 792 pavgw m1, m4
793 %if mmsize==8
794 packuswb m0, m0
795 packuswb m1, m1
796 movh [r0+r1*0], m0
797 movh [r0+r1*1], m1
798 %else
792 packuswb m0, m1 799 packuswb m0, m1
793 movh [r0+r1*0], m0 800 movh [r0+r1*0], m0
794 movhps [r0+r1*1], m0 801 movhps [r0+r1*1], m0
802 %endif
795 803
796 lea r0, [r0+r1*2] 804 lea r0, [r0+r1*2]
797 lea r2, [r2+r3*2] 805 lea r2, [r2+r3*2]
798 sub r4, 2 806 sub r4, 2
799 jg .nextrow 807 jg .nextrow
800 REP_RET 808 REP_RET
801 809
802 cglobal put_vp8_bilinear8_h_ssse3, 7,7,5 810 cglobal put_vp8_bilinear%1_h_ssse3, 7,7
803 shl r5d, 4 811 shl r5d, 4
804 %ifdef PIC 812 %ifdef PIC
805 lea r11, [bilinear_filter_vb_m] 813 lea r11, [bilinear_filter_vb_m]
806 %endif 814 %endif
807 pxor m4, m4 815 pxor m4, m4
816 pmaddubsw m1, m3 824 pmaddubsw m1, m3
817 psraw m0, 2 825 psraw m0, 2
818 psraw m1, 2 826 psraw m1, 2
819 pavgw m0, m4 827 pavgw m0, m4
820 pavgw m1, m4 828 pavgw m1, m4
829 %if mmsize==8
830 packuswb m0, m0
831 packuswb m1, m1
832 movh [r0+r1*0], m0
833 movh [r0+r1*1], m1
834 %else
821 packuswb m0, m1 835 packuswb m0, m1
822 movh [r0+r1*0], m0 836 movh [r0+r1*0], m0
823 movhps [r0+r1*1], m0 837 movhps [r0+r1*1], m0
838 %endif
824 839
825 lea r0, [r0+r1*2] 840 lea r0, [r0+r1*2]
826 lea r2, [r2+r3*2] 841 lea r2, [r2+r3*2]
827 sub r4, 2 842 sub r4, 2
828 jg .nextrow 843 jg .nextrow
829 REP_RET 844 REP_RET
845 %endmacro
846
847 INIT_MMX
848 FILTER_BILINEAR_SSSE3 4
849 INIT_XMM
850 FILTER_BILINEAR_SSSE3 8
830 851
831 cglobal put_vp8_pixels8_mmx, 5,5 852 cglobal put_vp8_pixels8_mmx, 5,5
832 .nextrow: 853 .nextrow:
833 movq mm0, [r2+r3*0] 854 movq mm0, [r2+r3*0]
834 movq mm1, [r2+r3*1] 855 movq mm1, [r2+r3*1]