This series upgrades the ARB_shader_precision test generator with
support for complex function tolerances.
The ARB_shader_precision spec does not give specific guidance in regard
to acceptable error ranges for complex functions, supplying only the
murky "Built-in functions defined in the specification with an equation
built from the above operations inherit the above errors." Currently,
the shader_precision tests allow zero tolerance for complex functions,
but this is causing many false negatives and needs to be improved.
This series is my best attempt at due diligence - any discussion/input
is certainly welcomed
The theory behind this is probably worth a blog post and I won't belabor
the point here, but this series basically implements a filter for the
test_suite dict that will drop any test vectors that would push a
function's output into regions of greatly multiplied error - the
"badlands". If the output does not trespass on the badlands, then the
error estimate in ulps is calculated and attached to the current test
vector for the shader_test to use as the allowable error range when
analyzing the output generated on the GPU during test run-time.
The bigfloat package is a necessary dependency for this filter, as it
provides both arbitrary-size floating point operations and it allows all
floating point values and operations in a code block to be forced into a
particular precision, leaving no question as to what precision was
employed for an intermediate calculation. This comes in handy when
running the reference versions of the complex built-ins. Performance
impact is small (~6s for the entire gen_shader_precision_tests.py script
on IvyBridge i7 @ 2.9GHz) and is localized to build-time.
The technique described above actually takes very little python code to
construct - most of the complexity is in dealing with the polymorphic
nature of the GLSL built-ins - componentwise vs. non-componentwise
parameters, etc.
Micah Fedke (3):
arb_shader_precision: support scalar values in shader_runner_format
arb_shader_precision: add framework for calculating tolerances for
complex functions
arb_shader_precision: enable calculation of complex function tolerances
CMakeLists.txt | 1 +
cmake/Modules/FindPythonBigfloat.cmake | 43
++++++++++++++
generated_tests/gen_shader_precision_tests.py | 402
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++----------
generated_tests/templates/gen_shader_precision_tests/fs.mako | 51
++++++++++------
generated_tests/templates/gen_shader_precision_tests/gs.mako | 51
++++++++++------
generated_tests/templates/gen_shader_precision_tests/vs.mako | 51
++++++++++------
6 files changed, 519 insertions(+), 80 deletions(-)
Changes since original:
Review comments addressed
- updated _is_sequence to use isinstance()
- simplified shader_runner_format(), added asserts
- added justification for bigfloat dep to commit msg
- added cmake check for bigfloat
- forced all args to be lists, to avoid scalar vs list madness
- removed ambiguous use of any()
- clarified the simple_fns vs complex_fns by moving mults into their own
dict with associated logic updates and added a comment block
- eliminated dict of errors/badlands/component_tolerances, now just
return the three as a tuple
- added comments to _gen_tolerance()
- added spacing around all math operators
- combined the framework and reference function commits, in order to
avoid the cross-references between them
Public repo:
http://cgit.collabora.com/git/user/fedke.m/piglit.git/log/?h=complex_tolerances_v2
--
Micah Fedke
Collabora Ltd.
+44 1223 362967
https://www.collabora.com/
https://twitter.com/collaboraltd
_______________________________________________
Piglit mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/piglit
