On Sun, Jul 05, 2009 at 12:02:27PM +0000, Jai Menon wrote: > On Sun, Jun 28, 2009 at 12:06 PM, Michael Niedermayer<[email protected]> wrote: > > On Sat, Jun 27, 2009 at 08:25:43PM +0000, Jai Menon wrote: > >> On Thu, Jun 25, 2009 at 9:51 PM, Michael Niedermayer<[email protected]> > >> wrote: > >> > On Wed, Jun 24, 2009 at 05:59:19PM +0000, Jai Menon wrote: > >> >> On Wed, Jun 24, 2009 at 3:58 PM, Michael Niedermayer<[email protected]> > >> >> wrote: > >> >> > On Wed, Jun 24, 2009 at 01:42:08PM +0000, Jai Menon wrote: > >> >> >> On Wed, Jun 24, 2009 at 1:27 PM, Michael > >> >> >> Niedermayer<[email protected]> wrote: > >> >> >> > On Sun, Jun 21, 2009 at 04:35:20PM +0000, Jai Menon wrote: > >> >> > [...] > >> >> >> > [...] > >> >> >> >> @@ -806,6 +815,26 @@ > >> >> >> >> > >> >> >> >> line += s->picture.linesize[0]; > >> >> >> >> } > >> >> >> >> + } else { > >> >> >> >> + for (; y < tile->comp[0].coord[1][1] - > >> >> >> >> s->image_offset_y; y++) { > >> >> >> >> + uint16_t *dst; > >> >> >> >> + x = tile->comp[0].coord[0][0] - s->image_offset_x; > >> >> >> >> + dst = line + x * s->ncomponents * 2; > >> >> >> >> + for (; x < tile->comp[0].coord[0][1] - > >> >> >> >> s->image_offset_x; x++) { > >> >> >> >> + for (compno = 0; compno < s->ncomponents; > >> >> >> >> compno++) { > >> >> >> > > >> >> >> >> + *src[compno] = av_rescale(*src[compno], (1 > >> >> >> >> << 16) - 1, > >> >> >> >> + (1 << > >> >> >> >> s->cbps[compno]) - 1); > >> >> >> > > >> >> >> > av_rescale is too slow > >> >> >> > >> >> >> So just (*src[compno]/((1 << s->cbps[compno]) - 1)) * ((1 << 16) - > >> >> >> 1) ? > >> >> > > >> >> > * is slow > >> >> > / s slower > >> >> > > >> >> > "src" << C > >> >> > it should be > >> >> > >> >> <possibly dumb question ahead> > >> >> > >> >> I understand that * and / are slower but how can I achieve the same > >> >> effect with a single <<? > >> > > >> > well, not the same but close enough IMHO > >> > src<<C > >> > or > >> > (src<<C) + (src>>(16-C)) > >> > should be close enough, my point was mainly that av_rescale() is too slow > >> > to be done per pixel and anything else is better > >> > >> Okay, modified patch attached. > > > > [...] > >> @@ -806,6 +815,22 @@ > >> > >> line += s->picture.linesize[0]; > >> } > >> + } else { > >> + for (; y < tile->comp[0].coord[1][1] - s->image_offset_y; y++) { > >> + uint16_t *dst; > >> + x = tile->comp[0].coord[0][0] - s->image_offset_x; > >> + dst = line + x * s->ncomponents * 2; > >> + for (; x < tile->comp[0].coord[0][1] - s->image_offset_x; > >> x++) { > >> + for (compno = 0; compno < s->ncomponents; compno++) { > >> + *src[compno] = *src[compno] << (16 - s->cbps[compno]); > >> + *src[compno] += 1 << 15; > >> + *src[compno] = av_clip(*src[compno], 0, (1 << 16) - > >> 1); > >> + *dst++ = *src[compno]++; > > > > i dont think using *src[compno] as a temporary is a good choice > > You mean *src[compno] should be copied to dst and all operations > should be done on dst? > Current approach seemed correct because this a part of level shifting. > Or did i misunderstand?
int val= src << ... val += ... val = av_clip(...) *dst++= val; its easy for the compiler to put val in a register, doing t with src is not because it would have to proof that src is not read after it [...] -- Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB Awnsering whenever a program halts or runs forever is On a turing machine, in general impossible (turings halting problem). On any real computer, always possible as a real computer has a finite number of states N, and will either halt in less than N cycles or never halt.
signature.asc
Description: Digital signature
_______________________________________________ FFmpeg-soc mailing list [email protected] https://lists.mplayerhq.hu/mailman/listinfo/ffmpeg-soc
