Closed GoogleCodeExporter closed 9 years ago
The fast-but-inaccurate math functions break OpenCL numerical compliance. This
is non-standard unless -cl-fast-relaxed-math is specified. I've edited math.h
to check if __FAST_RELAXED_MATH__ is set before using the fast-but-inaccurate
functions (otherwise using the previous slow-but-accurate functions).
Original comment by sctinc...@gmail.com
on 11 Aug 2014 at 4:28
Attachments:
Tweak clSetKernelArg to fix crash and return proper errors for when arg_size ==
0 in __local
Original comment by sctinc...@gmail.com
on 13 Aug 2014 at 12:38
Attachments:
in clGetProgramInfo: edge case where SET_VAR is not used, so required to check
if param_value is null
Original comment by sctinc...@gmail.com
on 13 Aug 2014 at 1:47
Attachments:
clEnqueue{Read,Write}Buffer edge case argument validation
Original comment by sctinc...@gmail.com
on 13 Aug 2014 at 9:29
Attachments:
That's impressive work! Thanks!
I've been reviewing and merging half of your patches. If you are interested I
can give you developer access to the SVN repository.
Original comment by zuzu...@gmail.com
on 13 Aug 2014 at 9:34
I've merged all your patches but the 0013. I've been reading the OpenCL 1.2
specifications (revision 19) and I could not find any constraint on the
accuracy of the native_* functions. As far as I understand this document all
these functions could return 0 no matter their inputs. Am I reading it
correctly ?
Original comment by zuzu...@gmail.com
on 13 Aug 2014 at 10:09
Thanks! And sure, I'm interested :]
You are correct. I was going by one of piglit's tests (GEGL gamma kernel),
which complained about not meeting tolerance, but indeed the spec does state
native_* functions have implementation-defined accuracy. This seems to be a
case where piglit either should not be testing with native_pow, or set a much
higher tolerance. My apologies.
Original comment by sctinc...@gmail.com
on 14 Aug 2014 at 12:00
You should have developer access to SVN repository now ;)
Original comment by zuzu...@gmail.com
on 14 Aug 2014 at 11:47
There are still a few tests that are failing/crashing. But these seems to be
symptoms of other problems.
In the very least, freeocl "conforms" as far as I'm concerned, so I am closing
this :]
Original comment by sctinc...@gmail.com
on 2 Mar 2015 at 5:16
Original issue reported on code.google.com by
sctinc...@gmail.com
on 8 Aug 2014 at 8:45Attachments: