Metal Compute Shader error: Requesting blit command encoder while having another active encoder

I'm trying to get some compute shader stuff working but seeing this error emitted for every thread.

Assertion failed: Requesting blit command encoder while having another active encoder

Since the same code works fine under DX11 I assume there're some additional constraints for Metal. The problem seems to be down to some conflict caused by using an InterlockedAdd to increment a counter within a structured buffer whilst also having other structured buffers being accessed in the same thread. ...Which is clearly ridiculous - it must be possible to use simple atomics under Metal, right?

Code (CSharp):

1. #pragma kernel CSMain
2.  
3. StructuredBuffer<float4> _ChunkDensities;
4. RWStructuredBuffer<uint> _EdgeResults;
5. RWBuffer<uint> _Counter;
6.  
7. uint3 _ChunkMultipliers;        // Per axis sizes to create a flattened index.
8.  
9. groupshared uint groupEdgeCount[9 * 9 * 3];
10.  
11. void ReadCornerDensities(uint fi, inout float2 v[2][2][2])
12. {
13.     // Read 8 values from _ChunkDensities, assign to input array.
14.     // NOTE: If these reads are commented out the error goes away.
15. }
16.  
17. uint EvaluateSurface(float2 v[2][2][2], uint groupIndex)
18. {
19.     uint res = ...            // Build a bit-field of edge states.
20.     uint c = countbits(res);            // How many edge bits were just set?
21.     groupEdgeCount[groupIndex] = c;
22.     return res | (c << 12);
23. }
24.  
25. [numthreads(9, 9, 3)]
26. void CSMain(uint3 id: SV_DispatchThreadID, uint groupIndex : SV_GroupIndex)
27. {
28.     uint3 scId = id * _ChunkMultipliers;
29.     uint fi = scId.x + scId.y + scId.z;
30.     float2 v[2][2][2];
31.     ReadCornerDensities(fi, v);
32.     _EdgeResults[fi] = EvaluateSurface(v, groupIndex);
33.  
34.     GroupMemoryBarrierWithGroupSync();
35.  
36.     if (groupIndex == 0) {
37.         uint sum = 0;
38.         for (int i = 0; i < (9 * 9 * 3); i++) {
39.             sum += groupEdgeCount[i];
40.         }
41.         InterlockedAdd(_Counter[0], sum);    // Increment edge pt count
42.     }
43. }
44.  

I've tried rewriting this a few different ways, this version using groupshared memory so that the atomic add is only used once per thread-group and with the Densities buffer as a StructuredBuffer rather than the Texture3D that it initially was. ...But no change What's especially driving me crazy is that I have a couple of other sample projects which are doing exactly the same kinds of things I'm trying to do and working just fine!

 

Seems I've finally found the culprit. The innocuous looking line:

 uint c = countbits(res); 

...If I don't use 'countbits' all is well.

So whatever that intrinsic expands to on Metal appears to make it basically unusable in combination with any other meaningful compute work. Having never gotten this code running I don't know if it was actually returning a valid result or not...