[Metal] Invalid DispatchThreadID value reported due to wrong pipeline group size setup

[Keenuts]

Found while working on #648

The shader outputs every threads values for some compute semantics like DispatchThreadID, GroupID, GroupThreadID and GroupIndex.
The values are written to a large buffer with a unique key as index computed from the GroupID and the GroupThreadId. Reading https://learn.microsoft.com/en-us/windows/win32/direct3dhlsl/sv-groupid this should allow me to get a stable index no matter the order of scheduling: the index depends on the XYZ position, not the order.

The result is stable across DX and VK, but for MSL, the results are slightly off:
https://github.com/llvm/offload-test-suite/actions/runs/21940854135/job/63365919110?pr=648

In this run, the value 0x210008 is reported by Metal, while the expected value is 0x210002.
0x21 is the unique key computed in the shader.

Key is computed as follows:
FID = thread index in the flat layout
KEY = FID * 12 + N, with N being a value in [0; 12[
(each thread outputs 12 values).

So we can compute back the thread index from this key:

• raw key: 0x21
• 0x21 == 2 * 12 + 9
• this is thread 2 in the flat representation.

dispatch: 2x2x2
numthreads: 2x2x2.
This means we have 4 threads per line, 16 threads per plane.
This means the global thread coordinate here is: x=2, y=0, z=0
Which should be: GroupID: x=1, y=0, z=0, GroupThreadID: x=0, y=0, z=0, GroupIndex: 2, DispatchThreadID: x=2, y=0, z=0

The expected value in the test file is 0x210002, which means we expect DispatchThreadID.x to be 2. But on Metal, this value is 8. Which is I think impossible given dispatch.x is 2 and numthreads.x is 2.

[Keenuts]

Found out in Metal, the numthreads attribute is not used, but the GroupSize is set during the dispatch call.

In the offload-test-suite, we have the following code:

    const NS::UInteger TGS =
        IS.ComputePipeline->maxTotalThreadsPerThreadgroup();
    const MTL::Size GroupSize(TGS, 1, 1);
    CmdEncoder->dispatchThreads(GridSize, GroupSize);

This means we have the grid size set to 2x2x2, but the group size is probably set to something higher by default. Looking at the short log extract, it seams we also have a TID.x reported as 10, meaning we have a groupSize > 10

[Keenuts]

Looking at Metal a bit, it seems MSL has no direct equivalent to numthreads in the shader itself so this cannot be done this way.

Solution would be to keep the current behavior as default, but add a new key in the YAML: MetalNumThreads to set it to a specific value when important for the test.

[Keenuts]

Looks like there is a PR in progress to add a json file with some reflection information:
microsoft/DirectXShaderCompiler#8159

The information we need is:

{
    "state" : {
      "tg_size" : [
         2,
         2,
         2
      ],
      [...]
   }
   [...]
}