WEIRD: convert -distort Affine producing black images

[NicolasRobidoux]

Cristy's patch looks like a good one: In exact arithmetic, we are never taking a square root of something negative. But we're not in exact arithmetic.
From svn diff:

Code: Select all

-  const double sqrt_discriminant = sqrt(discriminant);
+  const double sqrt_discriminant =
+    sqrt(discriminant < -0.0 ? 0.0 : discriminant);

If more problems crop up, I should carefully put some MagickEpsilons in there.

[NicolasRobidoux]

This bug is an immediate illustration of "direct" SVD decompositions being ill-conditioned. Fast, but better be extra careful. And I was not.

[NicolasRobidoux]

Forgot to use changelist again with IM6 (Makefile.PL, magick-baseconfig.h and version.h changed when I did not intend to).
-----
In any case, I made the test this one to minimize the impact of "various computation branches on floating point computation":

Code: Select all

 /*
   * In exact arithmetic, discriminant can't be negative. In floating
   * point, it can, because of the bad conditioning of SVD
   * decompositions done through the associated normal matrix.
   */
  const double sqrt_discriminant =
    sqrt(discriminant > 0.0 ? discriminant : 0.0);

[anthony]

The patch seems to have solved the problem.

[NicolasRobidoux]

I'm really impressed at how fast Cristy, Anthony and Fred got the issue figured out and solved.

[fmw42]

I'm really impressed at how fast Cristy, Anthony and Fred got the issue figured out and solved.

Thanks goes to Anthony and Cristy and you. I really had nothing to contribute on this one.

[tsanders]

Is there any progress on the new beta release? Thanks.

[tsanders]

Found it! Checked it out and it works 100% OK! Thanks everyone, GREAT JOB!

[NicolasRobidoux]

GREAT JOB!

... to you too. Thank you for your very informative bug report.

[anthony]

I have mapped it into IMv7 too.

[NicolasRobidoux]

I have mapped it into IMv7 too.

Apologies if I forgot. Juggling -> dropping balls.