[Bug] binary arithmetic with CUDA scalar #95
Description
In
It looks like in order to calculate a binary operator with a scalar on GPU, we need to first copy it to CPU (Is it expected? Will it affect the performance because of synchronization?). And the cpu() function raises an error saying we should first detach the variable in
hidet/python/hidet/graph/tensor.py
Line 486 in 80a35d6
If we add detach here, here comes the second problem. It looks like the .numpy(or to_dlpack) method does not support a tensor with only one element and shape []. The error message is like:
File "/home/su/accdiff/thirdparty/hidet/python/hidet/graph/ops/definitions/arithmetic.py", line 529, in divide
return binary_arithmetic(
File "/home/su/accdiff/thirdparty/hidet/python/hidet/graph/ops/definitions/arithmetic.py", line 479, in binary_arithmetic
x = x.dtype(x.item())
File "/home/su/accdiff/thirdparty/hidet/python/hidet/graph/tensor.py", line 528, in item
ret = self.squeeze(dims=list(range(len(self.shape)))).tolist()
File "/home/su/accdiff/thirdparty/hidet/python/hidet/graph/tensor.py", line 488, in tolist
ret = ret.numpy()
File "/home/su/accdiff/thirdparty/hidet/python/hidet/graph/tensor.py", line 930, in numpy
return np.from_dlpack(self)
File "/home/su/accdiff/thirdparty/hidet/python/hidet/graph/tensor.py", line 467, in __dlpack__
return to_dlpack(self)
File "/home/su/accdiff/thirdparty/hidet/python/hidet/graph/impl/dlpack.py", line 251, in to_dlpack
return DLManagedTensorContext(tensor).capsuled_dltensor()
File "/home/su/accdiff/thirdparty/hidet/python/hidet/graph/impl/dlpack.py", line 229, in __init__
data=tensor.storage.addr,
AttributeError: 'NoneType' object has no attribute 'addr'