[GPUToLLVM] `gpu.launch_func` lowering in gpu-to-llvm only supports 1 async dependency

[mshahneo]

Currently, gpu-to-llvm lowering only supports one async dependency on gpu.launc_func. More than 1 dependency results in failure. (https://github.com/llvm/llvm-project/blob/main/mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp#L954).

Since, the gpu.launch_func itself does not employ the restriction, updating the gpu-to-llvm to handle more than 1 async dependency would provide a more robust solution.

[llvmbot]

@llvm/issue-subscribers-mlir

Author: Md Abdullah Shahneous Bari (mshahneo)

Currently, `gpu-to-llvm` lowering only supports one async dependency on `gpu.launc_func`. More than 1 dependency results in failure. (https://github.com/llvm/llvm-project/blob/main/mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp#L954).

However, if a gpu.launch_func already has a dependency, and we run gpu-async-region pass on it. It may end up with more than one async dependency.

So, we either have to,

1. Update the gpu-to-llvm to handle more than 1 async dependency on gpu.launch_func, or
2. Update the gpu-async-region pass to make sure, no gpu.launch_func op ends up with more than 1 async dependency.

1 is more suitable, since, the gpu.launch_func itself does not employ the restriction, solving 1 would provide a more robust solution.