[Migrated] MCP: Bindless support

Empieza comparando el ejemplo de HLSL/DXC SPIR-V con el intento actual de Uniform<&[u32]> en el issue. Traza cómo se representan los arrays en tiempo de ejecución, el acceso mutable y la indexación de recursos; después, define la compatibilidad con bindless arrays y tipos wrapper globales para los tipos de recursos mencionados. Se considera terminado cuando el ejemplo puede compilarse con la estructura esperada de runtime-array e indexación de descriptores.

Issue automatically imported from old repo: https://github.com/EmbarkStudios/rust-gpu/issues/389
Old labels: mcp: accepted,a: ark
Originally creatd by Jasper-Bekkers on 2021-01-21T10:12:41Z

Let's start off with the simplest hlsl shader I can more or less come up with to do bindless support, which I would like to port to rust-gpu.

Notice that instead of ByteArrayBuffer here, for the full feature one should be able to have (RW)Texture2D and other types as well.

ByteAddressBuffer g_byteAddressBuffer[] : register(t0, space3);
RWByteAddressBuffer g_rwByteAddressBuffer[] : register(u0, space4);

[numthreads(64, 1, 1)]
void main(int threadId: SV_DispatchThreadID)
{
    g_rwByteAddressBuffer[0].Store(threadId, g_byteAddressBuffer[0].Load(threadId));
}

Emitted SPIR-V from DXC:

; SPIR-V
; Version: 1.0
; Generator: Google spiregg; 0
; Bound: 30
; Schema: 0
               OpCapability Shader
               OpCapability RuntimeDescriptorArray
               OpExtension "SPV_EXT_descriptor_indexing"
               OpMemoryModel Logical GLSL450
               OpEntryPoint GLCompute %main "main" %gl_GlobalInvocationID
               OpExecutionMode %main LocalSize 64 1 1
               OpSource HLSL 660
               OpName %type_ByteAddressBuffer "type.ByteAddressBuffer"
               OpName %g_byteAddressBuffer "g_byteAddressBuffer"
               OpName %type_RWByteAddressBuffer "type.RWByteAddressBuffer"
               OpName %g_rwByteAddressBuffer "g_rwByteAddressBuffer"
               OpName %main "main"
               OpDecorate %gl_GlobalInvocationID BuiltIn GlobalInvocationId
               OpDecorate %g_byteAddressBuffer DescriptorSet 3
               OpDecorate %g_byteAddressBuffer Binding 0
               OpDecorate %g_rwByteAddressBuffer DescriptorSet 4
               OpDecorate %g_rwByteAddressBuffer Binding 0
               OpDecorate %_runtimearr_uint ArrayStride 4
               OpMemberDecorate %type_ByteAddressBuffer 0 Offset 0
               OpMemberDecorate %type_ByteAddressBuffer 0 NonWritable
               OpDecorate %type_ByteAddressBuffer BufferBlock
               OpMemberDecorate %type_RWByteAddressBuffer 0 Offset 0
               OpDecorate %type_RWByteAddressBuffer BufferBlock
        %int = OpTypeInt 32 1
      %int_0 = OpConstant %int 0
       %uint = OpTypeInt 32 0
     %uint_2 = OpConstant %uint 2
     %uint_0 = OpConstant %uint 0
%_runtimearr_uint = OpTypeRuntimeArray %uint
%type_ByteAddressBuffer = OpTypeStruct %_runtimearr_uint
%_runtimearr_type_ByteAddressBuffer = OpTypeRuntimeArray %type_ByteAddressBuffer
%_ptr_Uniform__runtimearr_type_ByteAddressBuffer = OpTypePointer Uniform %_runtimearr_type_ByteAddressBuffer
%type_RWByteAddressBuffer = OpTypeStruct %_runtimearr_uint
%_runtimearr_type_RWByteAddressBuffer = OpTypeRuntimeArray %type_RWByteAddressBuffer
%_ptr_Uniform__runtimearr_type_RWByteAddressBuffer = OpTypePointer Uniform %_runtimearr_type_RWByteAddressBuffer
      %v3int = OpTypeVector %int 3
%_ptr_Input_v3int = OpTypePointer Input %v3int
       %void = OpTypeVoid
         %20 = OpTypeFunction %void
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%g_byteAddressBuffer = OpVariable %_ptr_Uniform__runtimearr_type_ByteAddressBuffer Uniform
%g_rwByteAddressBuffer = OpVariable %_ptr_Uniform__runtimearr_type_RWByteAddressBuffer Uniform
%gl_GlobalInvocationID = OpVariable %_ptr_Input_v3int Input
       %main = OpFunction %void None %20
         %22 = OpLabel
         %23 = OpLoad %v3int %gl_GlobalInvocationID
         %24 = OpCompositeExtract %int %23 0
         %25 = OpBitcast %uint %24
         %26 = OpShiftRightLogical %uint %25 %uint_2
         %27 = OpAccessChain %_ptr_Uniform_uint %g_byteAddressBuffer %int_0 %uint_0 %26
         %28 = OpLoad %uint %27
         %29 = OpAccessChain %_ptr_Uniform_uint %g_rwByteAddressBuffer %int_0 %uint_0 %26
               OpStore %29 %28
               OpReturn
               OpFunctionEnd

First attempt:

#[allow(unused_attributes)]
#[spirv(gl_compute)]
pub fn main_cs(
    #[spirv(descriptor_set = 0, binding = 0)] img: Uniform<&[u32]>,
) {
    let img = img.load();
    let stuff = &img[0];
}

This seems to emit a OpRuntimeArray but then proceed to do some things wrong. I had even more trouble making the slice mutable (for storing data) since that leads to a bunch of compiler errors down the line around .load not being available etc.

Ideally we would also like to declare the bindless arrays as globals, so it's nicer to create our own wrapper types around Image2d and other (so they can do the indirection through the bindless array).