[spirv] Add Vk::RawBufferLoad intrinsic

[yuriy-odonnell-epic]

This PR adds a basic implementation of memory loads through buffer_reference mechanism in SPIR-V. Partially addresses #3772.

Current implementation only handles basic uint loads, with similar semantics as HLSL ByteAddressBuffer (4 byte alignment).

Example shader:

uint64_t Address;
float4 main() : SV_Target0 {
  uint Value = vk::RawBufferLoad(Address);
  return asfloat(Value);
}

Proposed change implements the minimal functionality required for Unreal Engine and I would like to keep the scope of this particular PR minimal. Custom alignment, large/vector loads, templated load syntax, stores, etc. would also be great to implement (and I could do the work, if needed), but I would prefer to do them separately.

Update: The PR is now rebased and squashed, ready for review.

[yuriy-odonnell-epic]

Potentially, custom alignment tagging on load ops could be implemented using SPIR-V intrinsics that are being added in #3919.

[AppVeyorBot]

❌ Build DirectXShaderCompiler 1.0.808 failed (commit f09c2de60f by @yuriy-odonnell-epic)

[AppVeyorBot]

✅ Build DirectXShaderCompiler 1.0.809 completed (commit 4643e792c8 by @yuriy-odonnell-epic)

[AppVeyorBot]

❌ Build DirectXShaderCompiler 1.0.811 failed (commit 2db2a43323 by @yuriy-odonnell-epic)

[yuriy-odonnell-epic]

The functionality, docs and test are now implemented. I am not planning to make any significant further changes, so now would be a good time to review.

[AppVeyorBot]

✅ Build DirectXShaderCompiler 1.0.812 completed (commit 8047778760 by @yuriy-odonnell-epic)

[AppVeyorBot]

✅ Build DirectXShaderCompiler 1.0.813 completed (commit 125010354e by @yuriy-odonnell-epic)

[AppVeyorBot]

✅ Build DirectXShaderCompiler 1.0.819 completed (commit a4ac7bce28 by @yuriy-odonnell-epic)

[AppVeyorBot]

✅ Build DirectXShaderCompiler 1.0.836 completed (commit ba160cb6b0 by @yuriy-odonnell-epic)

[jaebaek]

Looks good to me. Just a nit.

By the way, this PR clearly and precisely addresses a non-trivial issue, which is pretty impressive. I appreciate your contribution!

[jaebaek]

Just out of a curiosity, is there any reason that you do not use

  const SpirvPointerType *bufferType =
      spvBuilder.getPhysicalStorageBufferType(spvContext.getUIntType(32));

?

It is a minor issue, but we prefer SpirvPointerType because const SpirvPointerType *SpirvContext::getPointerType(..) internally keeps the pointers we already created and reuses it while const HybridPointerType *SpirvContext::getPointerType(..) does not.

If there is no special reason, could you please replace it with SpirvPointerType?

[yuriy-odonnell-epic]

Done.

[jaebaek]

By the way, I observed that running clang-format slightly changes the code format.
Could you please apply this patch?
diff.txt

I got it by running

git clang-format master
git diff > diff.txt

[yuriy-odonnell-epic]

Applied the getUIntType and clang format changes.

[AppVeyorBot]

✅ Build DirectXShaderCompiler 1.0.846 completed (commit 7ebbbaf7b7 by @yuriy-odonnell-epic)

[jaebaek]

Thanks again for your contribution!
This PR clearly addresses the buffer address support!

[yuriy-odonnell-epic]

Thanks for reviewing and merging! I will work on the vectorized, templated and aligned loads next.