[MPlayer-cvslog] CVS: main/libmpcodecs vf_pp7.c, NONE, 1.1 vf.c, 1.115, 1.116 Makefile, 1.144, 1.145

[Rich Felker]

On Sat, Jul 09, 2005 at 09:49:53AM +0200, Michael Niedermayer CVS wrote:
> static void dctB_mmx(DCTELEM *dst, DCTELEM *src){
>     asm volatile (
>         "movq  (%0), %%mm0      \n\t"
>         "movq  1*4*2(%0), %%mm1 \n\t"
>         "paddw 6*4*2(%0), %%mm0 \n\t"
>         "paddw 5*4*2(%0), %%mm1 \n\t"
>         "movq  2*4*2(%0), %%mm2 \n\t"
>         "movq  3*4*2(%0), %%mm3 \n\t"
>         "paddw 4*4*2(%0), %%mm2 \n\t"
>         "paddw %%mm3, %%mm3     \n\t" //s
>         "movq %%mm3, %%mm4      \n\t" //s
>         "psubw %%mm0, %%mm3     \n\t" //s-s0
>         "paddw %%mm0, %%mm4     \n\t" //s+s0
>         "movq %%mm2, %%mm0      \n\t" //s2
>         "psubw %%mm1, %%mm2     \n\t" //s2-s1
>         "paddw %%mm1, %%mm0     \n\t" //s2+s1
>         "movq %%mm4, %%mm1      \n\t" //s0'
>         "psubw %%mm0, %%mm4     \n\t" //s0'-s'
>         "paddw %%mm0, %%mm1     \n\t" //s0'+s'
>         "movq %%mm3, %%mm0      \n\t" //s3'
>         "psubw %%mm2, %%mm3     \n\t"
>         "psubw %%mm2, %%mm3     \n\t"
>         "paddw %%mm0, %%mm2     \n\t"
>         "paddw %%mm0, %%mm2     \n\t"
>         "movq %%mm1, (%1)       \n\t"
>         "movq %%mm4, 2*4*2(%1)  \n\t"
>         "movq %%mm2, 1*4*2(%1)  \n\t"
>         "movq %%mm3, 3*4*2(%1)  \n\t"
>         :: "r" (src), "r"(dst)
>     );
> }

needs #ifdef around it... (breaks compile on non-x86!!)

> static void get_image(struct vf_instance_s* vf, mp_image_t *mpi){
>     if(mpi->flags&MP_IMGFLAG_PRESERVE) return; // don't change
>     // ok, we can do pp in-place (or pp disabled):
>     vf->dmpi=vf_get_image(vf->next,mpi->imgfmt,
>         mpi->type, mpi->flags, mpi->w, mpi->h);
>     mpi->planes[0]=vf->dmpi->planes[0];
>     mpi->stride[0]=vf->dmpi->stride[0];
>     mpi->width=vf->dmpi->width;
>     if(mpi->flags&MP_IMGFLAG_PLANAR){
>         mpi->planes[1]=vf->dmpi->planes[1];
>         mpi->planes[2]=vf->dmpi->planes[2];
>         mpi->stride[1]=vf->dmpi->stride[1];
>         mpi->stride[2]=vf->dmpi->stride[2];
>     }
>     mpi->flags|=MP_IMGFLAG_DIRECT;
> }

this is incorrect. dr must be disabled if the dmpi buffer obtained is
not in readable memory, otherwise it will be slow as hell.. i would
recommend disabling dr entirely if you cannot ensure this since it
barely helps anyway.

rich