Compute Shader: lack of support for most basic data types and operations in OpenGL / Vulcan

Bug Report Case: 1275152

Windows 10; Unity 2019.4.8f1 also 2020.1.4f1

Compute Shader support for OpenGL, Vulkan (Metal untested) seems non-existant.

When the type of a buffer is anything more complicated than <int> for example <float4> or <SomeStruct> it appears to be translated to an array of uints[] and then every time a value is being read or stored it's translated using uintBitsToFloat(). Unfortunately this causes all kinds of problems. For example InterlockedMin() is translated to uint variant of atomicMin(uint, uint). There's nothing I can do to force the int variant (atomicMin(int, int)) even if I cast everything explicitly to int. That obviously fails to properly compare negative numbers and returns incorrect results.

Simple example demonstrating the problem.

Code (CSharp):

1.  
2. struct ScopeMinMax
3. {
4.     int min;
5. };
6.  
7. RWStructuredBuffer<float4> _Buff;
8. RWStructuredBuffer<ScopeMinMax> _Min;
9.  
10.  
11. #pragma kernel BugRepro
12.  
13. [numthreads(8, 8, 1)]
14. void BugRepro(uint3 id : SV_DispatchThreadID)
15. {
16.     InterlockedMin((int)_Min[0].min, (int)_Buff[0].x);
17. }
18.  

Below is the buggy glsl code Unity currently generates (notice the forced cast to uint in AtomicMin() call):

Code (CSharp):

1.  
2. **** Platform OpenGL Core:
3. Compiled code for kernel BugRepro
4. keywords: <none>
5. #version 430
6. #extension GL_ARB_shading_language_420pack : require
7.  
8. #define HLSLCC_ENABLE_UNIFORM_BUFFERS 1
9. #if HLSLCC_ENABLE_UNIFORM_BUFFERS
10. #define UNITY_UNIFORM
11. #else
12. #define UNITY_UNIFORM uniform
13. #endif
14. #define UNITY_SUPPORTS_UNIFORM_LOCATION 1
15. #if UNITY_SUPPORTS_UNIFORM_LOCATION
16. #define UNITY_LOCATION(x) layout(location = x)
17. #define UNITY_BINDING(x) layout(binding = x, std140)
18. #else
19. #define UNITY_LOCATION(x)
20. #define UNITY_BINDING(x) layout(std140)
21. #endif
22.  struct _Buff_type {
23.     uint[4] value;
24. };
25.  
26. layout(std430, binding = 0) buffer _Buff {
27.     _Buff_type _Buff_buf[];
28. };
29.  struct _Min_type {
30.     uint[1] value;
31. };
32.  
33. layout(std430, binding = 1) buffer _Min {
34.     _Min_type _Min_buf[];
35. };
36. float u_xlat0;
37. int u_xlati0;
38. layout(local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
39. void main()
40. {
41.     u_xlat0 = uintBitsToFloat(_Buff_buf[0].value[(0 >> 2) + 0]);
42.     u_xlati0 = int(u_xlat0);
43.     atomicMin(_Min_buf[int(0)].value[int(0) >> 2], uint(u_xlati0));
44.     return;
45. }
46.  

Do you plan to properly support types like vec3 instead of this strange float3 <-> uint[] conversion in your HLSL to GLSL converter?

Without this you're effectively forcing everyone to write GLSL compute shaders from scratch manually.

PS. I don't have the auto generated vulcan code here but returned results seem to indicate similar if not identical bugs.