Correct lossless issue on linux x86

[mayeut]

Update #571

[malaterre]

this fixes the symptoms, not the bug

[mayeut]

@malaterre,

opj_tcd_makelayer takes a threshold that's computed in opj_tcd_rateallocate

the computation for the threshold in case of lossless transform is identical to the computation in opj_tcd_makelayer. Thus, it expects "computation" == stored_threshold to be true at some point.
The stored threshold will always be rounded to 64bits. "computation" precision will depend on compilation options, architectures, ...

IMHO, this PR removes the dependency on 64bits precision.

If you fill it does not fix the issue properly, I'll revert the "invalid fix" & only merge the tests in.

[mayeut]

@malaterre,

Given the definition of DBL_EPSILON:

#define DBL_EPSILON     2.2204460492503131e-016 /* smallest such that 1.0+DBL_EPSILON != 1.0 */

It might not be a good fix.

We might want to make the check with epsilon directly in the test in opj_tcd_makelayer

if ((thresh - (dd / dr)) <= DBL_EPSILON) /* do not rely on float equality, check with DBL_EPSILON margin */
    n = passno + 1;

I think the later would be OK, do you agree ?

[malaterre]

@mayeut that's a lot better ! I still do not like the equality (does not work for floating point anyway). So I would write:

if ((thresh - (dd / dr)) < DBL_EPSILON)

Do we need an fabs here also ?

[mayeut]

@malaterre,

I removed = from the test. It's now:

if ((thresh - (dd / dr)) < DBL_EPSILON)

fabs would have been needed if original test was ==. This is not the case here.

[malaterre]

ACK. Everything is fine for me. Thx