SPIR-V 1.6 Testing

[bashbaug]

Here is one more issue to track SPIR-V 1.6 testing.

• New Behavior:

  • [x] Adds a new Nontemporal image operand. https://github.com/KhronosGroup/OpenCL-CTS/pull/2249
  • The FP Fast Math Mode can apply to many more instructions:
    • [ ] OpFNegate
    • [ ] Comparison instructions such as OpOrdered, OpUnordered, OpFOrdEqual, etc.
    • [ ] Extended instructions - TBD, which ones?
  • [x] Adds the UniformDecoration capability, and Uniform and UniformId decorations. https://github.com/KhronosGroup/OpenCL-CTS/pull/2254
  • [ ] DotProduct and associated instructions may now be available without SPV_KHR_integer_dot_product. https://github.com/KhronosGroup/OpenCL-CTS/pull/2258
  • "NonSemantic" extended instruction sets are available. Not defining a task for this, because it should probably be tested as part of the non-semantic instruction set itself.

• Restrictions (informational):

  • OpTypeSampledImage and OpSampledImage do not work on Buffer images.
  • OpBranchConditional cannot have the same id for the True Lable and False Label.

[bashbaug]

For the FP Fast Math Mode testing:

1. Do we want to test all of the different flags, or just Fast? Do we want to test an explicit None?
2. Do we need to do anything more than ignoring results where the inputs or outputs are NaNs or infinities, or ignoring the sign of zero? I don't think we'll want to relax the behavior of any OpFNegate or any of the comparisons beyond this, for example.
3. We still need to determine which extended instructions can have FP Fast Math Modes. For extended instructions with Fast especially, do we want to define any accuracy requirements?