Audit of the OpenCL/CUDA compute path: 13 defects with fixes, and tests for the sensitivity path

[LawrenceMulholland]

Hello,

I'm Lawrence Mulholland at NAG (Numerical Algorithms Group). We have been using ORE as a case study for auditing
accelerated risk calculations, and the work turned up defects we would like to give back.

On current master (3d75a69):

• 13 defects in the OpenCL and CUDA backends**, six of which fail silently (exit code 0 with wrong or empty results,
  for example a CVA of 0 and an empty sensitivity file). Each has a cause, a fix and a reproduction. With the fixes,
  the CPU, OpenCL double and CUDA double give identical CVA and sensitivities on the cvasensi example.
• Smoothed indicatorGt / indicatorGeq (eps > 0) compare x with zero rather than with y, which affects the AD
  Min and Max. We have not established whether any production path is affected.
• Four latent defects without patches, and a documentation point (the tutorial runs the test suites from
  build/, but they must run from the source test directories).
• Seven new test cases that fail when we inject 15 sensitivity-path faults the existing tests did not catch.

It is written up as issues with patches, and as two pull requests (fixes; tests). Following the contributions page,
we would like to ask before opening them: would you prefer the issues first and the pull requests once you have
looked, or something else? Please also tell us if there is a process or agreement you would like us to follow.

Thank you for ORE.

[pcaspers]

Hi @LawrenceMulholland, you can open the pull requests directly and add the description of the issue there. This is much appreciated, many thanks in advance. Best, Peter