https://bugs.documentfoundation.org/show_bug.cgi?id=167365

--- Comment #17 from Eike Rathke <[email protected]> ---
The runtime tests are there for a reason, and if they fail that specific OpenCL
driver remains disabled on purpose. While ChatGPT in general is right on "the
test and Calc use double precision (double, IEEE-754 64-bit). With only 23 bits
of mantissa, the rounding error of a float quickly exceeds 1e-12", I don't know
if "kernel returns values in single precision (float)" is actually correct
(though it seems it is here); why would the OpenCL driver use float single
precision even though RUSTICL_FEATURES=fp64 was requested that allegedly should
use 64-bit floating-point double precision? Of course lowering precision
requirements eventually will let the test pass, but the purpose of the test is
to ensure that OpenCL results are sufficiently close to non-OpenCL results.
Other known-good OpenCL drivers pass the test, so lowering the test threshold
to make this driver pass sounds much ill-advised to me.

-- 
You are receiving this mail because:
You are the assignee for the bug.

Reply via email to