Question How to avoid race conditions in group shared memory

Hi, there.

I want to use group shared memory in a kernel function of a compute shader, so I tried to test with the following shader. My expectation was that it should return all (100, 0, 0), but the output was (1, 0, 0).

I think this is due to race conditions, but is there a good solution?

Code (CSharp):

1.  
2. using System.Runtime.InteropServices;
3. using UnityEngine;
4.  
5. public class TestScript : MonoBehaviour
6. {
7.     uint _objNum = 10;
8.  
9.     ComputeShader _shader;
10.     GraphicsBuffer _buffer;
11.  
12.     void Start()
13.     {
14.         _shader = (ComputeShader)Resources.Load("Test");
15.         _buffer = new GraphicsBuffer(
16.             GraphicsBuffer.Target.Structured,
17.             (int)_objNum,
18.             Marshal.SizeOf(typeof(Vector3)));
19.  
20.         DispatchShader(isDebugLog: true);
21.     }
22.  
23.     private void OnDestroy()
24.     {
25.         _buffer.Release();
26.     }
27.  
28.     void DispatchShader(bool isDebugLog)
29.     {
30.         int kernelID = _shader.FindKernel("CSMain");
31.         _shader.SetBuffer(kernelID, "test", _buffer);
32.         _shader.GetKernelThreadGroupSizes(kernelID, out var x, out var y, out var z);
33.         _shader.Dispatch(kernelID, 1, 1, 1);
34.  
35.         if (!isDebugLog) return;
36.  
37.         BufferUtils.DebugBuffer<Vector3>(_buffer, (int)_objNum);    // <- same to "Debug.Log()"
38.     }
39. }
40.  

Code (ComputeShader):

1.  
2. #pragma kernel CSMain
3. #define OBJECT_NUM 10
4. RWStructuredBuffer<float3> test;
5. groupshared float3 sharedBuffer[OBJECT_NUM];
6. [numthreads(10,10,10)]
7. void CSMain(uint3 groupID : SV_GroupID, uint3 threadID : SV_GroupThreadID, uint3 dispatchID : SV_DispatchThreadID, uint gi : SV_GroupIndex)
8. {
9.     const uint ID = dispatchID.x;
10.     sharedBuffer[threadID.x] = float3(0, 0, 0);
11.     GroupMemoryBarrierWithGroupSync();
12.    
13.     sharedBuffer[threadID.x] += float3(1, 0, 0);
14.     GroupMemoryBarrierWithGroupSync();
15.    
16.     test[threadID.x] = sharedBuffer[threadID.x];
17. }
18.