On Fri, Oct 25, 2013 at 11:08 PM, Tom Stellard <[email protected]> wrote: > On Fri, Oct 25, 2013 at 05:40:47PM -0500, Aaron Watry wrote: >> Pretty basic, but we didn't have any tests for this before... >> >> Tested on r600g w/ CEDAR. >> --- >> generated_tests/generate-cl-math-builtins.py | 8 ++++++++ >> 1 file changed, 8 insertions(+) >> >> diff --git a/generated_tests/generate-cl-math-builtins.py >> b/generated_tests/generate-cl-math-builtins.py >> index b6b4e43..150c412 100644 >> --- a/generated_tests/generate-cl-math-builtins.py >> +++ b/generated_tests/generate-cl-math-builtins.py >> @@ -57,6 +57,14 @@ tests = { >> [1.0, -1.0, 0.0, -0.0, 0.0], >> [0.5, -0.5, 0.0, -0.0, float("nan")] >> ] >> + }, >> + 'sqrt' : { >> + 'arg_types': [F, F], >> + 'function_type': 'ttt', >> + 'values': [ >> + [1.0, 2.0, 6.0, 2.5 , float("nan"), 4.0], # Result >> + [1.0, 4.0, 36.0, 6.25, float("nan"), 16.0], # Arg1 > > I think you should also add one or more cases here that require a lot of > precision. I'm pretty sure we get this wrong on r600g, since we are > using the instruction that calculates the approximate sqrt. >
Alright. > Also, I think it would be good to add a TODO somewhere in the python > script for checking that the results of the builtins satisfy the ulp > requirements for the spec. Yeah, I wasn't sure how to specify tolerances in a cl test case, so I had specifically chosen values that I knew would come out exact given my limited understanding of floating point encoding. We definitely need to test the precision of float instructions, and I'll add a TODO for that. I realized that I also forgot to send the hunk that sets CLC_VERSION_MIN for sqrt to 10. I'll make sure to include that bit when I push. --Aaron > > With these changes, the patch is: > Reviewed-by: Tom Stellard <[email protected]> > >> + ] >> } >> } >> >> -- >> 1.8.3.2 >> >> _______________________________________________ >> Piglit mailing list >> [email protected] >> http://lists.freedesktop.org/mailman/listinfo/piglit _______________________________________________ Piglit mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/piglit
