FusionLayerNormSharedMemoryBuffer_CUDA validation

[naoyam]

https://github.com/NVIDIA/Fuser/blob/main/tests/cpp/test_gpu3.cpp#L7371

FusionLayerNormSharedMemoryBuffer_CUDA seems to be trying to check if the inner persistent scheduler is used, but just calling computeHeuristics doesn't seem to guarantee FusionExecutorCache indeed uses the scheduler but it seems it just computes the heuristics parameters assuming the fusion passes the canSchedule functions. Indeed, running the tests seems to indicate the fusion is segmented in some sizes.

@liqiangxl What is this part meant to check?

[liqiangxl]

That part is used to check whether register or shared memory is used to store persistent buffers.
You are right, some cases may segment when the hidden size is very large and exceeds the avialable shared memory.

[naoyam]

Please update the test so that its intended behavior is indeed verified. The result through FusionExecutorCache is what actually matters, and that's not validated.