[SPIR-V] Texture2D used with sampleCmp not tagged as depth texture about directxshadercompiler HOT 2 CLOSED

I don't think we will be able to match the code generated by slang. Our primary concern is to generate SPIR-V for Vulkan, and supporting other APIs through spirv-cross or other tools is a secondary. In this case, slang generates invalid spir-v.

%13 = OpTypeImage %float 2D 0 0 0 1 Unknown  ; Not a depth image
%19 = OpTypeImage %float 2D 1 0 0 1 Unknown ;  a depth image
%20 = OpTypeSampledImage %19

%29 = OpLoad %13 %tex_0 ; tex_0 is not a depth image.
%31 = OpSampledImage %20 %29 %30  ; The type of %29 implicitly changes to a depth image. The type of %29 is %13, but the image type in %20 is %19.

The spir-v spec for OpSampledImage says:

Result Type must be the OpTypeSampledImage type whose Image Type operand is the type of Image.

I'm double checking with people who know the Vulkan spec better than I do that this is not overridden by the Vulkan specification.

In general, DXC cannot tell if an image is a depth image or not, so it is best to use 2 for the depth in the type. If we started changing the type based on how the variable is used, we could easily run into problem keeping track of the type. For example, Slang is generating what I believe to be invalid code because of this problem.

I view this more as a spirv-cross problem. They are the ones that have to worry about metal, and doing the correct analysis to get the right types for metal. That is not DXC's job.

from directxshadercompiler.

Tint has a SPIR-V to WGSL conversion path, and faces similar translation challenges as spirv-cross.
Tint tracks usages by operations rather than types, to infer the resulting type for the target language (WGSL).

from directxshadercompiler.