On Mon, Apr 15, 2013 at 09:25:22AM +0200, Kostya Shishkov wrote:
> On Mon, Apr 15, 2013 at 08:31:46AM +0200, Nicolas BERTRAND wrote:
> > Le 15/04/2013 06:15, Kostya Shishkov a écrit :
> > >a) make FATE test account for fuzziness of the result
> > ??
> 
> We have tests for audio that allow result to be slightly different from the
> reference - exactly because of floats.
> 
> > >b) add lots of round() and forget about performance in principle
> > Performance is already crap
> 
> But then it won't be possible to improve it much even with SIMD.
> 
> > >c) add an integer version of IDWT (even if only for bitexact tests)
> > Thats may be not possible. in JPEG2000 2 cases ( lossy, and
> > lossless) . lossy float IDWT lossless integer IDWt. And its 2
> > separate algos ( duabechies 9/7 and daubechies 5/3). Dcinema case is
> > always lossy
> 
> One can always approximate. With JPEG-1992 DCT is also defined with
> floating-point coefficients yet many implementations use integers.

To give a more specific example:

p[2 * i + 1] -= LFTG_GAMMA * (p[2 * i] + p[2 * i + 2]);

can be replaced with

p[2 * i + 1] -= (57862 * (p[2 * i] + p[2 * i + 2]) + (1 << 15)) >> 16;
for integer p

(57862 = LFTG_GAMMA * (1 << 16) if it's not clear)
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel

Reply via email to