TSC Meeting Notes 2020-11-19

Owned by Cary Phillips
Nov 04, 2022

Attending:

  • Cary Phillips
  • Christina Tempelaar-Lietz
  • Joseph Goldstone
  • Kimball Thurston
  • Larry Gritz
  • Nick Porcino
  • Owen Thompson
  • Peter Hillman

Discussion:

  • Imath PR #66: Use likely/unlikely to improve certain vector ops.

    • This applies the builtin_expect to if statements to assist with optimization. This is the result of engineers at Intel going over things with a fine-tooth comb.

    • The general rule is that the directive is appropriate when you’re detecting an error that is truly exceptional.

    • Does MSVC have the same feature? Maybe, someone should check. But at least now all we have to do is extend the macro.

    • #66 also includes a change to a line to invoke std::abs(), technically shouldn't go in the same PR, but fine as is.

  • Imath #65 Deprecate Math in favor of std::

    • In the name of portability, std:: is better.

    • Kimball: There are some cases where I didn’t want to call the std version because it does extra error checking. But you can tell the compiler to use the unsafe function

    • Should we get rid of the Math class entirely? This PR removes all uses of the class but leaves it in the code.

    • Nick: Every line that goes is less debt.

    • Kimball: IlmThread in OpenEXR has the deprecation stiff, for mutex and lock.

    • Larry will port that over to this class.

  • #60 Make Matrix implementation more SIMD friendly

    • Unrolled assignments are faster than memset/memcpy.

    • These came about by an attempt to autofectorize the code.

    • Also included: a fix to Matrix33::setScale(), which attempted to mimic Matrix44::setScale() by not affecting the final row, which for M44 is the translation. But that doesn't ake sense for a 3x3 matrix. Nick wrote the original code at ILM years ago, an abandoned effort. All agree to regard the current behavior as a bug and fix it to conform to the most logical behavior and affect all rows.

  • Imath #59 Hide Vec and Matrix exceptions from Cuda

    • This PR #ifdef's out the throw calls under CUDA.

    • Would it be better to #ifdef out the entire function?

    • Or alternatively, remove the HOST_DEVICE decorator, which will prevent them from bei ng compiled under CUDA altogether.

    • Nick: A little churn in CUDA code doesn’t set off alarm bells, it’s expected to be a bit volitile, so this shouldn't cause a problem.

    • One gotcha: M44::inverse(bool singExc) if invoked explicitly via inverse(false). But the argument-less version is available.

  • Imath PR #58 Make separate test calls for each test
  • Not sure why the CLA test failed. Cary will fix it and merge.

  • openexr PR #865: Add a conditional to prevent analysis on forks

    • Larry: GitHub says they’re disabling actions on all forks, which should address this.

    • Kimball: this was the suggested fix, although it wasn't the most thorough of investigations.

  • openexr PR #862 E2K: added initial support of MCST Elbrus 2000 CPU architecture

    • Elbrus 2000 is a Russian supercomputer. It has all the same registers, but not the same assembly syntax.

    • Nick: I was hoping that the directives could be specified at cmake time. The code in question is a part of highly nested bits at compile-time. Can those pathways be selected at cmake time? I couldn’t tell.

    • It’s only used in the test of the DWA compressor.

    • It's minimal extra debt, so fine to accept, subject to the CLA.

  • Larry: We haven’t yet marked things as noexcept.

    • Ideally, we do it across the board everywhere that's applicable. Comprehensive.

    • Certainly, in the cases of dual functions, it helps with documentation.

    • The compiler ought to be able to tell you if a function throws, but it doesn't.

    • In C++21(?) noexcept becomes a part of the name of the function, so it would be an ABI break.

    • Larry will pick something and do it, then that can be used as an example.

  • Larry: OIIO builds and supports OpenEXR back to version 2.x.

    • Every downstream project is going to need to change. At source and build.

    • Nick: USD ws doing detection on OpenEXR version, to fix dependency on OIIO.

    • Larry will write up a short guide for how downstream projects will need to adjust to 3.0