bitwise and kernel #156
bitwise and kernel #156
Conversation
|
I think you mentioned that bitwise and works for booleans as well, right? We probably need a test for booleans as well. |
| Tensor and_kernel_hb(const Tensor& self, const Tensor& other) { | ||
| TORCH_CHECK(self.numel() == other.numel(), "The size of two tensors should match."); | ||
| // TORCH_CHECK(self.scalar_type() == ScalarType::Int || self.scalar_type() == ScalarType::Bool, "HammerBlade and is implemented for Int and Bool only"); | ||
| // TORCH_CHECK(other.scalar_type() == ScalarType::Int || other.scalar_type() == ScalarType::Bool, "HammerBlade and is implemented for Int and Bool only"); |
There was a problem hiding this comment.
If you want to check the data type of tensors, the canonical way is like this
./aten/src/ATen/native/BinaryOps.cpp:94: TORCH_CHECK(self.scalar_type() != kBool || other.scalar_type() != kBool,
There was a problem hiding this comment.
Actually, with boolean data the emulator segfaults when you want to quit ... I kind of think the reason is your kernel uses 32bit int internally, and I'm not too sure about the size of booleans
There was a problem hiding this comment.
>>> x = torch.ones(10, dtype=torch.bool)
>>> x.element_size()
1
yeah so a booleans is a single byte in pytorch
There was a problem hiding this comment.
Actually, with boolean data the emulator segfaults when you want to quit ... I kind of think the reason is your kernel uses 32bit int internally, and I'm not too sure about the size of booleans
how can I get around with the segfault?
There was a problem hiding this comment.
I made two functions in kernel_and (one for int and one for bool). If type is bool, do bool fun; if type is int do int fun. But I still have segfault for bool type.
2 participants
bitwise and kernel with two tensor inputs, only accept integer and boolean
type checks needed