HLSL: binding start offset feature

[TheArheus]

Hi, I have small user experience issue, not really bug or anything. I am interested if spirv-cross have a feature so that when it cross compile from spv to hlsl, it would start, for example, CBVs not from 0, but from another user specified value. For example, right now, I have this cbuffer, generated from spirv:

cbuffer pushConstant
{
    row_major float4x4 _75_ShadowMatrix : packoffset(c0);
};

And I want to know if it spirv-cross could generate the bindings for specific resource type to be started from another value(for example, starting from 1) and not from zero

[HansKristian-Work]

What do you mean here, do you want to offset packoffset() (c0), or register(b#)? Of the latter, you can use the resource binding APIs to achieve that. Random resource shifts is jarring and not supported. You can override the bindings in many different ways.

[TheArheus]

Yeah, I was about packoffset(c#) starts from some arbitrary value and register(b#, space#) to start from some value and was thinking about this feature or how to do this if already implemed features

[HansKristian-Work]

You want to look at CompilerHLSL::add_hlsl_resource_binding. It can override register/space as needed. packoffset is really something you don't want to mess with. If you really want to do you overwrite the decorations with set_member_decoration(type.self, DecorationOffset, offset) on the type.

[TheArheus]

@HansKristian-Work there is another thing with generating actual hlsl code. It is similar to my initial issue, but different, and I think this should be a bug. For example, I have this piece of glsl shader bindings defines:

layout(set = 0, binding = 8)  uniform sampler2D GBuffer[GBUFFER_COUNT]; // << Based on error, this one is the issue, GBUFFER_COUNT is equal to 5
layout(set = 0, binding = 9)  uniform writeonly image2D ColorTarget;
layout(set = 0, binding = 10) uniform writeonly image2D BrightTarget;
layout(set = 0, binding = 11) uniform sampler2D AmbientOcclusionBuffer;
layout(set = 0, binding = 12) uniform sampler2D ShadowMap[DEPTH_CASCADES_COUNT];

And the issue is that it will translate to something like this:

Texture2D<float4> GBuffer[GBUFFER_COUNT] : register(t8, space0); // <-- The issue will be after this
SamplerState _GBuffer_sampler[GBUFFER_COUNT] : register(s8, space0); // <-- The issue will be after this
RWTexture2D<float4> ColorTarget : register(u9, space0);
RWTexture2D<float4> BrightTarget : register(u10, space0);
Texture2D<float4> AmbientOcclusionBuffer : register(t11, space0);
SamplerState _AmbientOcclusionBuffer_sampler : register(s11, space0);
Texture2D<float4> ShadowMap[DEPTH_CASCADES_COUNT] : register(t12, space0);
SamplerState _ShadowMap_sampler[DEPTH_CASCADES_COUNT] : register(s12, space0);

The issue is that after translating glsl to hlsl, there will be overlapping registers and I wont be able to create a pipeline state. If I have a register, which is actually an array of samplers (like GBuffer in my code), then, the bindings after that should come like this:

Texture2D<float4> GBuffer[GBUFFER_COUNT] : register(t8, space0); // <-- Now there shouldn't be an issue after that
SamplerState _GBuffer_sampler[GBUFFER_COUNT] : register(s8, space0); // <-- Now there shouldn't be an issue after that
RWTexture2D<float4> ColorTarget : register(u13, space0);
RWTexture2D<float4> BrightTarget : register(u14, space0);
Texture2D<float4> AmbientOcclusionBuffer : register(t15, space0);
SamplerState _AmbientOcclusionBuffer_sampler : register(s15, space0);
Texture2D<float4> ShadowMap[DEPTH_CASCADES_COUNT] : register(t16, space0);
SamplerState _ShadowMap_sampler[DEPTH_CASCADES_COUNT] : register(s16, space0);

Or it shpuld just generate Texture2DArray, but then without binding offset