David Neto

When converting from a wider floating type to a narrower floating type, check inputs that have magnitude that are just a little bigger than the biggest-magnitude normal numbers in the...

We never updated the WGSL spec to make compat errors into shader-creation errors. So the resolution is to not change the WGSL spec. Closing.

Summary: Yes, it's ok for pipeline creation to set the value of an override that is not statically used by the shader. From this part of the API spec: >...

I've forked the diagnostic feature request into its own issue #4643 This issue should be closed once we've landed the clarification to the original question.

> I'm not aware of apis in Vulkan/D3D12 which expose it. Vulkan 1.1 added [VkPhysicalDeviceSubgroupProperties](https://www.khronos.org/registry/vulkan/specs/1.1/html/vkspec.html#VkPhysicalDeviceSubgroupProperties) which has a subgroupSize member. That will match the value of the SubgroupSize-builtin-decorated input variable...

Wait a second, I think there's some confusion here. There are two concepts at play: - OpenCL/Vulkan workgroup =~ Metal threadgroup - OpenCL/Vulkan subgroup =~ Metal SIMD group. The size...

Our OpenCL C to Vulkan compute shader (https://github.com/google/clspv) compiler supports 3 ways for passing plain-old-data (POD) arguments: - buffer per argument - collect them into a single UBO - collect...

One of the sharp edges is that space available for push-constants can be very limited, e.g. 128bytes is the minimum required to be supported by a Vulkan implementation.

Vulkan permits use of ConstOffset with ImageFetch. I suspect there is a missing rule to ban ConstOffset with ImageRead, ImageWrite.

Yes, the implementation has to ensure that these instructions behave as if clamped. The shader coming into the implementation doesn't have to do anything special to get this behaviour.