Skip to content

Fix Tensor.type(dtype) not preserving device #7474

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
colesbury merged 1 commit into from
May 10, 2018
Merged

Fix Tensor.type(dtype) not preserving device #7474

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions test/test_cuda.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -817,6 +817,8 @@ def test_type_conversions(self):
def test_type_conversions_same_gpu(self):
x = torch.randn(5, 5).cuda(1)
self.assertEqual(x.int().get_device(), 1)
self.assertEqual(x.type(torch.int).get_device(), 1)
self.assertEqual(x.to(torch.int).get_device(), 1)

def test_neg(self):
TestTorch._test_neg(self, lambda t: t.cuda())
Expand Down
2 changes: 1 addition & 1 deletion tools/autograd/templates/python_variable_methods.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -628,7 +628,7 @@ static PyObject * THPVariable_type(PyObject* self, PyObject* args, PyObject* kwa
auto self_device_type = torch::getDeviceType(self_.type());
auto& type = is_dtype ? torch::getType(r.scalartype(0), *torch::getLayout(self_.type().backend()), self_device_type) :
torch::utils::type_from_string(type_name);
return THPVariable_Wrap(torch::utils::dispatch_type_conversion(self_, type, -1, r.toBool(1)));
return THPVariable_Wrap(torch::utils::dispatch_type_conversion(self_, type, at::nullopt, r.toBool(1)));
END_HANDLE_TH_ERRORS
}

Expand Down
16 changes: 8 additions & 8 deletions torch/csrc/utils/tensor_conversion_dispatch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,19 +8,19 @@

namespace torch { namespace utils {

at::Tensor dispatch_type_conversion(const at::Tensor & self, const at::Type & type) {
int64_t device = self.is_cuda() ? self.get_device() : -1;
return dispatch_type_conversion(self, type, device, false);
}

at::Tensor dispatch_type_conversion(const at::Tensor & self, const at::Type & type,
int device, bool non_blocking) {
at::Tensor dispatch_type_conversion(
const at::Tensor & self,
const at::Type & type,
at::optional<int> device,
bool non_blocking) {
if (type.is_cuda()) {
torch::utils::cuda_lazy_init();
}
AutoNoGIL no_gil;
AutoGPU auto_gpu(device);

int64_t tensor_device = self.is_cuda() ? self.get_device() : -1;
AutoGPU auto_gpu(device.value_or(tensor_device));

if (self.is_cuda() && type.is_cuda() && tensor_device != at::current_device()) {
// copy if the devices are different even if the types are the same
return type.copy(self, non_blocking);
Expand Down
22 changes: 19 additions & 3 deletions torch/csrc/utils/tensor_conversion_dispatch.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,8 +6,24 @@

namespace torch { namespace utils {

at::Tensor dispatch_type_conversion(const at::Tensor & self, const at::Type & type);
at::Tensor dispatch_type_conversion(const at::Tensor & self, const at::Type & type,
int device, bool non_blocking);
// Returns a tensor with the same data as `self` and the specified type and
// device. Returns `self` unmodified if neither the type nor device change;
// otherwise a copy is made.
//
// The `device` argument is only relevant if `type` is a CUDA type. There are
// a few special cases for device:
//
// - if device is -1 then the returned tensor will be on the current device
// - if device is nullopt then the returned tensor will be on the same device
// as `self` if possible; otherwise it will be on the current device.
//
// If `non_blocking` is true, then the copy may be performed asynchronously
// w.r.t the host if `self` is a CPU tensor in pinned memory and `type` is a
// CUDA type. Note that copies between CUDA devices are always asynchronous
// w.r.t the host.
at::Tensor dispatch_type_conversion(const at::Tensor & self,
const at::Type & type,
at::optional<int> device=at::nullopt,
bool non_blocking=false);

}} // namespace torch::utils