0.9.3 usually passes all its tests with 2 warnings:
double_precision_check()
- WARN: GPU doesn't have correct double precision. Got 9.995699E-05,
expected 0.0001000001
[SUCCESS]
test_printf()Warning: Have a int parameter for %f like specifier, take
care of it
it once failed one test (not sure if it was runtime_barrier_list() or
runtime_marker_list(): race condition?), but usually passes that one.
This is better than 0.8, which always fails builtin_{cospi,erf,erfc,
nextafter,pow,pown,powr,rootn,sinpi,tanpi,tgamma}_float{,2,4,8,16}() and
often crashes with a bus error on sub_buffer_check().
My beignet_test script (see #767148) finds that multiplication is ~1.5x
faster in 0.9.3 than in 0.8, and cos+sin+sqrt ~25x (!) faster but less
accurate. The test suite takes 41-45s on 0.9.3 and 46-48s on 0.8, but
probably isn't the best speed measure.
It would be easy to create an 0.8+dfsg1-1 that just fixed the two RC
bugs and would be allowed under freeze, but given the above, and that
new hardware needs new beignet versions (0.8 is Ivy Bridge _only_,
nothing newer), it may make more sense to accept removal from jessie and
provide a regularly-updated version in jessie-backports.
At least the code that decides that beignet is not supported on the
platform should be compiled without these switches.
That's at least src/cl_api.c, src/cl_device_id.c,
src/intel/intel_driver.c (from when I added that error return: #745363).
--
To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org
with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
