[Pkg-opencl-devel] Bug#767961: unblock pre-approval: beignet/0.9.3~really.0.8+dfsg-1

[Rebecca N. Palmer]

> I didn't look at the details of the patch for #768090, but the bug log
> suggests that there are remaining failures. Is that still the case with this
> patch?

Assuming you mean
> The remaining test failures are:
> -cospi/sinpi/tanpi, powr/pown/pow, tgamma are less accurate than the
> OpenCL spec requires (at least the first group explicitly use the
> fast-but-inaccurate path, 2e-5 instead of 1e-8 typical relative error
> but 10-20x faster; one reason 0.9 is "faster" is that it has sin/cos/tan
> do the same by default).
> -sub_bufffer_check sometimes crashes.
yes, it's the same patch so has the same failures: I don't claim this is 
perfect, just the best we can reasonably do within the freeze rules.

In current upstream (1.0), the crash is gone and a non-default "strict 
conformance" mode is added where cospi/sinpi/tanpi meet the accuracy 
standard, but powr/pown/pow and tgamma still fail.

> you can use a version like 1:0.8+dfsg-1 in unstable if you don't like
> 0.9.3~really.0.8+dfsg-1
Please don't do that: I'd rather have a temporary ~really than a 
(permanent) epoch.