[Pkg-opencl-devel] beignet 1.2.0

[Rebecca N. Palmer]

What I just pushed is probably close to what will be 1.2.0-2, but it 
hasn't been fully tested yet, so please don't upload it.

As jessie-backports now has a mesa that uses LLVM/Clang 3.8, we also 
need to decide whether jessie-backports beignet should match jessie mesa 
(LLVM/Clang 3.5) or jessie-backports mesa, and if the latter whether it 
should have a Breaks: on non-backports mesa.  (I've never seen the 
mismatched-versions crash myself, but I don't usually link beignet and 
mesa into the same program.)

> I'm also considering whether to ship the testing tool in beignet-dev, as
> they now claim this can be used to test ICDs other than beignet
> (https://www.freedesktop.org/wiki/Software/Beignet/howto/stand-alone-utest-howto/).

Done as /usr/lib/<arch>/utest_run (i.e. an "internal binary" by 
FHS/Policy); it may make sense to move it to /usr/bin with a less 
generic name (e.g. beignet-opencl-test), but I'd like to discuss that 
with upstream first.  (It would also imply either giving up multiarch on 
beignet-dev, or splitting it into two packages.)

> They say these are "Khronos official extensions",

They are (https://www.khronos.org/registry/cl/#otherextspecs 23 and 24).

> My patch moves these two to the (existing, in beignet-dev)
> Intel-specific header cl_intel.h; they work in the test suite (linked
> directly to beignet/libcl.so), but this requires #including that file
> (as noted in update-docs.patch), and I don't know whether they work via
> ICD.

They do work via ICD (at least to the extent tested by 
builtin_kernel_block_motion_estimate_intel).