Mercurial > libavcodec.hg
comparison i386/dsputil_mmx.c @ 2899:d3a726717baf libavcodec
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
faster 8x8 mmx ssd (77=>70 cycles)
author | lorenm |
---|---|
date | Fri, 30 Sep 2005 02:31:47 +0000 |
parents | 41315d0120b3 |
children | 3c79bc9f3aa9 |
comparison
equal
deleted
inserted
replaced
2898:95f469274a1d | 2899:d3a726717baf |
---|---|
742 | 742 |
743 static int sse8_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { | 743 static int sse8_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
744 int tmp; | 744 int tmp; |
745 asm volatile ( | 745 asm volatile ( |
746 "movl %4,%%ecx\n" | 746 "movl %4,%%ecx\n" |
747 "shr $1,%%ecx\n" | |
747 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */ | 748 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */ |
748 "pxor %%mm7,%%mm7\n" /* mm7 holds the sum */ | 749 "pxor %%mm7,%%mm7\n" /* mm7 holds the sum */ |
749 "1:\n" | 750 "1:\n" |
750 "movq (%0),%%mm1\n" /* mm1 = pix1[0-7] */ | 751 "movq (%0),%%mm1\n" /* mm1 = pix1[0][0-7] */ |
751 "movq (%1),%%mm2\n" /* mm2 = pix2[0-7] */ | 752 "movq (%1),%%mm2\n" /* mm2 = pix2[0][0-7] */ |
752 | 753 "movq (%0,%3),%%mm3\n" /* mm3 = pix1[1][0-7] */ |
754 "movq (%1,%3),%%mm4\n" /* mm4 = pix2[1][0-7] */ | |
755 | |
756 /* todo: mm1-mm2, mm3-mm4 */ | |
757 /* algo: substract mm1 from mm2 with saturation and vice versa */ | |
758 /* OR the results to get absolute difference */ | |
753 "movq %%mm1,%%mm5\n" | 759 "movq %%mm1,%%mm5\n" |
760 "movq %%mm3,%%mm6\n" | |
754 "psubusb %%mm2,%%mm1\n" | 761 "psubusb %%mm2,%%mm1\n" |
762 "psubusb %%mm4,%%mm3\n" | |
755 "psubusb %%mm5,%%mm2\n" | 763 "psubusb %%mm5,%%mm2\n" |
764 "psubusb %%mm6,%%mm4\n" | |
756 | 765 |
757 "por %%mm1,%%mm2\n" | 766 "por %%mm1,%%mm2\n" |
758 | 767 "por %%mm3,%%mm4\n" |
768 | |
769 /* now convert to 16-bit vectors so we can square them */ | |
759 "movq %%mm2,%%mm1\n" | 770 "movq %%mm2,%%mm1\n" |
771 "movq %%mm4,%%mm3\n" | |
760 | 772 |
761 "punpckhbw %%mm0,%%mm2\n" | 773 "punpckhbw %%mm0,%%mm2\n" |
774 "punpckhbw %%mm0,%%mm4\n" | |
762 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */ | 775 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */ |
776 "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (mm3,mm4) */ | |
763 | 777 |
764 "pmaddwd %%mm2,%%mm2\n" | 778 "pmaddwd %%mm2,%%mm2\n" |
779 "pmaddwd %%mm4,%%mm4\n" | |
765 "pmaddwd %%mm1,%%mm1\n" | 780 "pmaddwd %%mm1,%%mm1\n" |
766 | 781 "pmaddwd %%mm3,%%mm3\n" |
767 "add %3,%0\n" | 782 |
768 "add %3,%1\n" | 783 "lea (%0,%3,2), %0\n" /* pix1 += 2*line_size */ |
784 "lea (%1,%3,2), %1\n" /* pix2 += 2*line_size */ | |
769 | 785 |
770 "paddd %%mm2,%%mm1\n" | 786 "paddd %%mm2,%%mm1\n" |
787 "paddd %%mm4,%%mm3\n" | |
771 "paddd %%mm1,%%mm7\n" | 788 "paddd %%mm1,%%mm7\n" |
789 "paddd %%mm3,%%mm7\n" | |
772 | 790 |
773 "decl %%ecx\n" | 791 "decl %%ecx\n" |
774 "jnz 1b\n" | 792 "jnz 1b\n" |
775 | 793 |
776 "movq %%mm7,%%mm1\n" | 794 "movq %%mm7,%%mm1\n" |
838 "paddd %%mm7,%%mm1\n" | 856 "paddd %%mm7,%%mm1\n" |
839 "movd %%mm1,%2\n" | 857 "movd %%mm1,%2\n" |
840 : "+r" (pix1), "+r" (pix2), "=r"(tmp) | 858 : "+r" (pix1), "+r" (pix2), "=r"(tmp) |
841 : "r" ((long)line_size) , "m" (h) | 859 : "r" ((long)line_size) , "m" (h) |
842 : "%ecx"); | 860 : "%ecx"); |
861 return tmp; | |
862 } | |
863 | |
864 static int sse16_sse2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { | |
865 int tmp; | |
866 asm volatile ( | |
867 "shr $1,%2\n" | |
868 "pxor %%xmm0,%%xmm0\n" /* mm0 = 0 */ | |
869 "pxor %%xmm7,%%xmm7\n" /* mm7 holds the sum */ | |
870 "1:\n" | |
871 "movdqu (%0),%%xmm1\n" /* mm1 = pix1[0][0-15] */ | |
872 "movdqu (%1),%%xmm2\n" /* mm2 = pix2[0][0-15] */ | |
873 "movdqu (%0,%4),%%xmm3\n" /* mm3 = pix1[1][0-15] */ | |
874 "movdqu (%1,%4),%%xmm4\n" /* mm4 = pix2[1][0-15] */ | |
875 | |
876 /* todo: mm1-mm2, mm3-mm4 */ | |
877 /* algo: substract mm1 from mm2 with saturation and vice versa */ | |
878 /* OR the results to get absolute difference */ | |
879 "movdqa %%xmm1,%%xmm5\n" | |
880 "movdqa %%xmm3,%%xmm6\n" | |
881 "psubusb %%xmm2,%%xmm1\n" | |
882 "psubusb %%xmm4,%%xmm3\n" | |
883 "psubusb %%xmm5,%%xmm2\n" | |
884 "psubusb %%xmm6,%%xmm4\n" | |
885 | |
886 "por %%xmm1,%%xmm2\n" | |
887 "por %%xmm3,%%xmm4\n" | |
888 | |
889 /* now convert to 16-bit vectors so we can square them */ | |
890 "movdqa %%xmm2,%%xmm1\n" | |
891 "movdqa %%xmm4,%%xmm3\n" | |
892 | |
893 "punpckhbw %%xmm0,%%xmm2\n" | |
894 "punpckhbw %%xmm0,%%xmm4\n" | |
895 "punpcklbw %%xmm0,%%xmm1\n" /* mm1 now spread over (mm1,mm2) */ | |
896 "punpcklbw %%xmm0,%%xmm3\n" /* mm4 now spread over (mm3,mm4) */ | |
897 | |
898 "pmaddwd %%xmm2,%%xmm2\n" | |
899 "pmaddwd %%xmm4,%%xmm4\n" | |
900 "pmaddwd %%xmm1,%%xmm1\n" | |
901 "pmaddwd %%xmm3,%%xmm3\n" | |
902 | |
903 "lea (%0,%4,2), %0\n" /* pix1 += 2*line_size */ | |
904 "lea (%1,%4,2), %1\n" /* pix2 += 2*line_size */ | |
905 | |
906 "paddd %%xmm2,%%xmm1\n" | |
907 "paddd %%xmm4,%%xmm3\n" | |
908 "paddd %%xmm1,%%xmm7\n" | |
909 "paddd %%xmm3,%%xmm7\n" | |
910 | |
911 "decl %2\n" | |
912 "jnz 1b\n" | |
913 | |
914 "movdqa %%xmm7,%%xmm1\n" | |
915 "psrldq $8, %%xmm7\n" /* shift hi qword to lo */ | |
916 "paddd %%xmm1,%%xmm7\n" | |
917 "movdqa %%xmm7,%%xmm1\n" | |
918 "psrldq $4, %%xmm7\n" /* shift hi dword to lo */ | |
919 "paddd %%xmm1,%%xmm7\n" | |
920 "movd %%xmm7,%3\n" | |
921 : "+r" (pix1), "+r" (pix2), "+r"(h), "=r"(tmp) | |
922 : "r" ((long)line_size)); | |
843 return tmp; | 923 return tmp; |
844 } | 924 } |
845 | 925 |
846 static int hf_noise8_mmx(uint8_t * pix1, int line_size, int h) { | 926 static int hf_noise8_mmx(uint8_t * pix1, int line_size, int h) { |
847 int tmp; | 927 int tmp; |
2624 | 2704 |
2625 c->hadamard8_diff[0]= hadamard8_diff16_mmx; | 2705 c->hadamard8_diff[0]= hadamard8_diff16_mmx; |
2626 c->hadamard8_diff[1]= hadamard8_diff_mmx; | 2706 c->hadamard8_diff[1]= hadamard8_diff_mmx; |
2627 | 2707 |
2628 c->pix_norm1 = pix_norm1_mmx; | 2708 c->pix_norm1 = pix_norm1_mmx; |
2629 c->sse[0] = sse16_mmx; | 2709 c->sse[0] = (mm_flags & MM_SSE2) ? sse16_sse2 : sse16_mmx; |
2630 c->sse[1] = sse8_mmx; | 2710 c->sse[1] = sse8_mmx; |
2631 c->vsad[4]= vsad_intra16_mmx; | 2711 c->vsad[4]= vsad_intra16_mmx; |
2632 | 2712 |
2633 c->nsse[0] = nsse16_mmx; | 2713 c->nsse[0] = nsse16_mmx; |
2634 c->nsse[1] = nsse8_mmx; | 2714 c->nsse[1] = nsse8_mmx; |