Implementation of the CoopVec Inference and Training builtin intrinisics

[anupamachandra]

Implements
HLSL:
__builtin_MatVecMul
__builtin_MatVecMulAdd
__builtin_OuterProductAccumulate
__builtin_VectorAccumulate

Lowered to
DXIL:
@dx.op.matVecMul
@dx.op.matVecMulAdd
@dx.op.outerProductAccumulate
@dx.op.vectorAccumulate

[github-actions]

✅ With the latest revision this PR passed the Python code formatter.

[damyanp]

NOTE: this is a general issue with long vectors tracked by #7297. I'll keep this comment here since it has an interesting case we might want to test in it.

This applies to all the builtins I've tried so far, but the VectorAccumulate example is quite minimal. Given this code:

export void TruncatedVector(vector<half, 254> Input254, vector<half, 255> Input255) {
  __builtin_VectorAccumulate(Input254, RWBuf, 0);
  __builtin_VectorAccumulate(Input255, RWBuf, 0);
}```

This generates:

```llvm
; Function Attrs: nounwind
define void @"\01?TruncatedVector@@YAXV?$vector@$halff@$0PO@@@V?$vector@$halff@$0PP@@@@Z"(<254 x float> %Input254, <255 x float> %Input255) #0 {
  %1 = load %dx.types.Handle, %dx.types.Handle* @"\01?RWBuf@@3URWByteAddressBuffer@@A", align 4
  %2 = call %dx.types.Handle @dx.op.createHandleForLib.dx.types.Handle(i32 160, %dx.types.Handle %1)  ; CreateHandleForLib(Resource)
  %3 = call %dx.types.Handle @dx.op.annotateHandle(i32 216, %dx.types.Handle %2, %dx.types.ResourceProperties { i32 4107, i32 0 })  ; AnnotateHandle(res,props)  resource: RWByteAddressBuffer
  call void @dx.op.vectorAccumulate.v254f32(i32 308, <254 x float> %Input254, %dx.types.Handle %3, i32 0)  ; VectorAccumulate(inputVector,arrayBuffer,arrayOffset)
  %4 = shufflevector <255 x float> %Input255, <255 x float> undef, <1 x i32> zeroinitializer
  %5 = call %dx.types.Handle @dx.op.createHandleForLib.dx.types.Handle(i32 160, %dx.types.Handle %1)  ; CreateHandleForLib(Resource)
  %6 = call %dx.types.Handle @dx.op.annotateHandle(i32 216, %dx.types.Handle %5, %dx.types.ResourceProperties { i32 4107, i32 0 })  ; AnnotateHandle(res,props)  resource: RWByteAddressBuffer
  call void @dx.op.vectorAccumulate.v1f32(i32 308, <1 x float> %4, %dx.types.Handle %6, i32 0)  ; VectorAccumulate(inputVector,arrayBuffer,arrayOffset)
  ret void
}

Note how Input255 is explicitly truncated to 1xfloat before the vectorAccumulate is called. Input254 is not truncated.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[tex3d]

Besides the generated content, I think this looks good. Just one small nit regarding DXIL Op descriptions.

I believe the generated content is out of date/incorrect, since I noticed some deleted operations and a missing .json file update. In any case, generated files will need to be updated before the final PR is ready for merging.

[tex3d]

Just an FYI:
Our gcc pipelines started failing today because of a docker image update that updates Ubuntu (will be ultimately to Ubunutu v24.4 by 5/9), which will require us to bump our gcc version a few. A fix is in the works, but in the meantime, I think we should override this failure and merge if there are no other failures, once the other pipelines are complete.