comparison x86/vp8dsp.asm @ 11992:da388061b227 libavcodec

Add x86 asm functions for VP8 put_pixels
author darkshikari
date Mon, 28 Jun 2010 19:14:40 +0000
parents a6d24fc1deb7
children a717c1a93036
comparison
equal deleted inserted replaced
11991:a6d24fc1deb7 11992:da388061b227
808 lea r2, [r2+r3*2] 808 lea r2, [r2+r3*2]
809 sub r4, 2 809 sub r4, 2
810 jg .nextrow 810 jg .nextrow
811 REP_RET 811 REP_RET
812 812
813 cglobal put_vp8_pixels8_mmx, 5,5
814 .nextrow:
815 movq mm0, [r2+r3*0]
816 movq mm1, [r2+r3*1]
817 lea r2, [r2+r3*2]
818 movq [r0+r1*0], mm0
819 movq [r0+r1*1], mm1
820 lea r0, [r0+r1*2]
821 sub r4d, 2
822 jg .nextrow
823 REP_RET
824
825 cglobal put_vp8_pixels16_mmx, 5,5
826 .nextrow:
827 movq mm0, [r2+r3*0+0]
828 movq mm1, [r2+r3*0+8]
829 movq mm2, [r2+r3*1+0]
830 movq mm3, [r2+r3*1+8]
831 lea r2, [r2+r3*2]
832 movq [r0+r1*0+0], mm0
833 movq [r0+r1*0+8], mm1
834 movq [r0+r1*1+0], mm2
835 movq [r0+r1*1+8], mm3
836 lea r0, [r0+r1*2]
837 sub r4d, 2
838 jg .nextrow
839 REP_RET
840
841 cglobal put_vp8_pixels16_sse, 5,5,2
842 .nextrow:
843 movups xmm0, [r2+r3*0]
844 movups xmm1, [r2+r3*1]
845 lea r2, [r2+r3*2]
846 movaps [r0+r1*0], xmm0
847 movaps [r0+r1*1], xmm1
848 lea r0, [r0+r1*2]
849 sub r4d, 2
850 jg .nextrow
851 REP_RET
852
813 ;----------------------------------------------------------------------------- 853 ;-----------------------------------------------------------------------------
814 ; IDCT functions: 854 ; IDCT functions:
815 ; 855 ;
816 ; void vp8_idct_dc_add_<opt>(uint8_t *dst, DCTELEM block[16], int stride); 856 ; void vp8_idct_dc_add_<opt>(uint8_t *dst, DCTELEM block[16], int stride);
817 ;----------------------------------------------------------------------------- 857 ;-----------------------------------------------------------------------------