Enable tests previously disabled due to an aliasing bug

[naoyam]

The bug was fixed by #1792

Both tests just require about 33 KB of SMEM, not 70 KB anymore, so they should be fine on most devices. Still making sure not exceeding the available capacity.

[zasdfgbnm]

Not necessary in this PR, but it would be good to use the inferred shared memory size from FusionExecutor::computeSharedMemory, instead of manually calculating it here.

[naoyam]

I think that's part of the check actually. We want to make sure that the aliasing is accurately analyzed and only this size is required. If aliasing analysis fails somehow, FusionExecutor::computeSharedMemory could just return a larger size, and the test would be just skipped, but in that case we would want the test to fail. So in this case checking with this required_smem_size is more appropriate.

[zasdfgbnm]

I see, makes sense

[zasdfgbnm]

Question, what is the difference between sharedMemPerBlock and sharedMemPerBlockOptin? I didn't find the official documentation helpful:

size_t cudaDeviceProp::sharedMemPerBlock [inherited]
  Shared memory available per block in bytes

size_t cudaDeviceProp::sharedMemPerBlockOptin [inherited]
  Per device maximum shared memory per block usable by special opt in

[naoyam]

I didn't check thoroughly, but I assume the former is the default size, whereas the latter is the maximum size that can be configured. Since we don't do any smem size configuration, I think the former is the right limit.

[zasdfgbnm]

Is it related to this part of the code?

pytorch/torch/csrc/jit/codegen/cuda/executor.cpp

Lines 647 to 676 in a43cb20

	// Check that requested smem size can be dynamically allocated.
	// This check is only done once a kernel has been compiled, since
	// maybe_available_dynamic_smem_ needs to be evaluated on
	// a compiled kernel.
	if (maybe_available_dynamic_smem_.has_value()) {
	// Dynamic shared memory space that we can allocate without
	// carving more space from L1.
	const uint64_t available_dynamic_smem_without_reconfiguration =
	maybe_available_dynamic_smem_.value();
	// Maximum additional shared memory size we could request
	// if we do re-configuration.
	const uint64_t additional_dynamic_smem_available_through_reconfiguration =
	device_smem_limit_ - configured_device_smem_;
	TORCH_INTERNAL_ASSERT(
	(dynamic_smem_size) <
	(available_dynamic_smem_without_reconfiguration +
	additional_dynamic_smem_available_through_reconfiguration),
	"The total shared memory allocation is larger than available memory.",
	" Dynamic size: ",
	dynamic_smem_size,
	". Available size: ",
	maybe_available_dynamic_smem_.value(),
	". Configured smem size: ",
	configured_device_smem_,
	". Device limit size: ",
	device_smem_limit_);
	}
	launch_params.setSmem(dynamic_smem_size);

[naoyam]

Oh, I didn't know we actually do configure the size when necessary.

pytorch/torch/csrc/jit/codegen/cuda/executor.cpp

Line 1207 in a43cb20

// Increase limit of dynamic shared memory if needed.

I'll change the check to use Optin.

[naoyam]

FYI, those values on Titan RTX are:

sharedMemPerBlock: 49152 
sharedMemPerBlockOptin: 65536