[Vulkan, Glslang]: Storage texture format not patched in SPIR-V compiled from HLSL via glslang

[MikhailGorobets]

On the Vulkan backend, the storage texture format isn’t being patched in SPIR-V compiled from HLSL via glslang.
For a texture defined as VK_IMAGE_FORMAT("r11g11b10f") RWTexture2D<float3 /*format=r11f_g11f_b10f*/> g_TextureRadiance;
I get the following error from the validation layers:

Diligent Engine: ERROR: Vulkan debug message (validation): VUID-RuntimeSpirv-OpImageWrite-07112
                 Validation Error: [ VUID-RuntimeSpirv-OpImageWrite-07112 ] Object 0: handle = 0x9fba48000000204a, name = ComputeGroundTruthRadiance-CS [C448EE8E6962D9F26E992A5B37246626], type = VK_OBJECT_TYPE_SHADER_MODULE; | MessageID = 0x7fcb0409 | vkCreateComputePipelines(): pCreateInfos[0].stage SPIR-V OpImageWrite Texel operand only contains 3 components, but the OpImage format mapping to VK_FORMAT_R32G32B32A32_SFLOAT has 4 components.
OpImageWrite %2333 %2358 %5779
%2206 = OpTypeImage %6 1 0 0 0 2 1
. The Vulkan spec states: OpImageWrite to any Image whose Image Format is not Unknown must have the Texel operand contain at least as many components as the corresponding VkFormat as given in the SPIR-V Image Format compatibility table (https://vulkan.lunarg.com/doc/view/1.3.280.0/windows/1.3-extensions/vkspec.html#VUID-RuntimeSpirv-OpImageWrite-07112)
                 Object[0] (shader module): Handle 0x9fba48000000204a, Name: 'ComputeGroundTruthRadiance-CS [C448EE8E6962D9F26E992A5B37246626]'

[TheMostDiligent]

Yes, this is unfortunately the missing glslang feature that will unlikely be implemented. The workaround is to use the SHADER_COMPILE_FLAG_HLSL_TO_SPIRV_VIA_GLSL shader compile flag.

https://github.com/DiligentGraphics/DiligentSamples/blob/c1a37afb8e0e1cb917063db52c72d544befea492/Tutorials/Tutorial23_CommandQueues/src/Terrain.cpp#L244-L248

[MikhailGorobets]

Is it possible to use ExtractGLSLImageFormatsFromHLSL by default and then patch the SPIR-V directly? That would remove the need to translate to GLSL (the main issue is not the translation speed, but that the translator is not fully featured, so strict attention has to be paid to types, initialization patterns, and so on).

[TheMostDiligent]

Is it possible to use ExtractGLSLImageFormatsFromHLSL by default and then patch the SPIR-V directly

It should be possible, but sounds complicated: first we need to parse HLSL to find the format attribute, and then patch the SPIRV. I think that it may actually be easier to contribute this feature to glslang.

the main issue is not the translation speed, but that the translator is not fully featured

Still, the translator can handle quite complicated shaders, and it is sufficiently reliable. For rare corner cases, there usually exist relatively simple workarounds. Parsing HLSL to extract format information may also be tricky and prone to potential errors.

I think existing approach with intermediate GLSL conversion is acceptable, though obviously not ideal.