Port `resize_as_` and `clone` from TH to Aten (#23027)
Summary:
API operators now routed to `at::native::resize_as_*_` and `at::native::clone` accordingly.
Internal `THTensor_(resizeAs)`, `THCTensor_(resizeAs)`, `THTensor_(newClone)` and `THCTensor_(newClone)` remains to support older TH code.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23027
Differential Revision: D16362304
Pulled By: VitalyFedyunin
fbshipit-source-id: 4c1e8516da685f3fdea632ff791d143f27aeebeb
diff --git a/aten/src/ATen/Declarations.cwrap b/aten/src/ATen/Declarations.cwrap
index 57f3df7..f558390 100644
--- a/aten/src/ATen/Declarations.cwrap
+++ b/aten/src/ATen/Declarations.cwrap
@@ -182,19 +182,6 @@
- THTensor* self
]]
[[
- name: _th_clone
- cname: newClone
- return: THTensor*
- variants:
- - function
- cpu_half: True
- cpu_bool: True
- cuda_bool: True
- cpu_bfloat16: True
- arguments:
- - THTensor* self
-]]
-[[
name: _th_view
cname: newView
cpu_half: True
@@ -211,21 +198,6 @@
long_args: True
]]
[[
- name: _th_resize_as_
- cname: resizeAs
- cpu_half: True
- cpu_bool: True
- cuda_bool: True
- cpu_bfloat16: True
- variants:
- - function
- return: self
- scalar_check: the_template_->dim() == 0
- arguments:
- - THTensor* self
- - THTensor* the_template
-]]
-[[
name: _th_index_select
cpu_bool: True
cuda_bool: True
diff --git a/aten/src/ATen/native/Resize.cpp b/aten/src/ATen/native/Resize.cpp
index d04d695..e342aae 100644
--- a/aten/src/ATen/native/Resize.cpp
+++ b/aten/src/ATen/native/Resize.cpp
@@ -10,4 +10,8 @@
return self;
}
+Tensor& resize_as_cpu_(Tensor& self, const Tensor& the_template) {
+ return resize_cpu_(self, the_template.sizes());
+}
+
}}
diff --git a/aten/src/ATen/native/TensorFactories.cpp b/aten/src/ATen/native/TensorFactories.cpp
index d9f5403..c08caca 100644
--- a/aten/src/ATen/native/TensorFactories.cpp
+++ b/aten/src/ATen/native/TensorFactories.cpp
@@ -826,5 +826,13 @@
return tensor;
}
+// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ clone ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Tensor clone(const Tensor& src) {
+ auto self = at::empty_like(src);
+ self.copy_(src);
+ return self;
+}
+
} // namespace native
} // namespace at
diff --git a/aten/src/ATen/native/cuda/Resize.cu b/aten/src/ATen/native/cuda/Resize.cu
index 6e00b4f..1fda687 100644
--- a/aten/src/ATen/native/cuda/Resize.cu
+++ b/aten/src/ATen/native/cuda/Resize.cu
@@ -12,4 +12,8 @@
return self;
}
+Tensor& resize_as_cuda_(Tensor& self, const Tensor& the_template) {
+ return resize_cuda_(self, the_template.sizes());
+}
+
}}
diff --git a/aten/src/ATen/native/native_functions.yaml b/aten/src/ATen/native/native_functions.yaml
index 6d70850..d58ed6e 100644
--- a/aten/src/ATen/native/native_functions.yaml
+++ b/aten/src/ATen/native/native_functions.yaml
@@ -2381,8 +2381,8 @@
- func: clone(Tensor self) -> Tensor
variants: function, method
dispatch:
- CPU: legacy::cpu::_th_clone
- CUDA: legacy::cuda::_th_clone
+ CPU: clone
+ CUDA: clone
SparseCPU: clone_sparse
SparseCUDA: clone_sparse
MkldnnCPU: mkldnn_clone
@@ -2391,8 +2391,8 @@
- func: resize_as_(Tensor(a!) self, Tensor the_template) -> Tensor(a!)
variants: function, method
dispatch:
- CPU: legacy::cpu::_th_resize_as_
- CUDA: legacy::cuda::_th_resize_as_
+ CPU: resize_as_cpu_
+ CUDA: resize_as_cuda_
SparseCPU: resize_as_sparse_
SparseCUDA: resize_as_sparse_
diff --git a/aten/src/TH/generic/THTensor.cpp b/aten/src/TH/generic/THTensor.cpp
index 9ca0c9e..4742f5c 100644
--- a/aten/src/TH/generic/THTensor.cpp
+++ b/aten/src/TH/generic/THTensor.cpp
@@ -153,10 +153,11 @@
THTensor *THTensor_(newClone)(THTensor *self)
{
+ // already available in Aten as at::clone()
THTensor *tensor = THTensor_(new)();
- THTensor_(resizeAs)(tensor, self);
at::Tensor tensor_wrap = THTensor_wrap(tensor);
at::Tensor self_wrap = THTensor_wrap(self);
+ tensor_wrap.resize_as_(self_wrap);
at::native::copy_(tensor_wrap, self_wrap, false);
return tensor;
}
@@ -217,6 +218,7 @@
void THTensor_(resizeAs)(THTensor *self, THTensor *src)
{
+ // already available in Aten as at::resize_as_()
if(!THTensor_(isSameSizeAs)(self, src))
THTensor_(resizeNd)(self, src->dim(), THTensor_getSizePtr(src), NULL);
}
diff --git a/aten/src/THC/generic/THCTensor.cpp b/aten/src/THC/generic/THCTensor.cpp
index 65aaf08..14aaeaf 100644
--- a/aten/src/THC/generic/THCTensor.cpp
+++ b/aten/src/THC/generic/THCTensor.cpp
@@ -160,8 +160,11 @@
THCTensor *THCTensor_(newClone)(THCState *state, THCTensor *self)
{
+ // already available in Aten as at::clone()
THCTensor *tensor = THCTensor_(new)(state);
- THCTensor_(resizeAs)(state, tensor, self);
+ at::Tensor tensor_wrap = THTensor_wrap(tensor);
+ at::Tensor self_wrap = THTensor_wrap(self);
+ tensor_wrap.resize_as_(self_wrap);
THCTensor_(copy)(state, tensor, self);
return tensor;
}
@@ -249,6 +252,7 @@
void THCTensor_(resizeAs)(THCState *state, THCTensor *self, THCTensor *src)
{
+ // already available in Aten as at::resize_as_()
THCTensor_resizeAs(state, self, src);
}
diff --git a/test/test_torch.py b/test/test_torch.py
index 09f9378..04e23b4 100644
--- a/test/test_torch.py
+++ b/test/test_torch.py
@@ -3479,15 +3479,12 @@
for device in torch.testing.get_all_device_types():
for dt in torch.testing.get_all_dtypes():
x = torch.tensor([1, 2, 3, 4], dtype=dt, device=device)
-
- if (device == 'cuda' and dt == torch.bfloat16):
- self.assertRaises(RuntimeError, lambda: x.clone())
- continue
-
x_clone = x.clone()
+ if (device == 'cuda' and dt == torch.bfloat16):
+ self.assertRaises(RuntimeError, lambda: copy(x))
+ continue
y = copy(x)
y.fill_(1)
-
# copy is a shallow copy, only copies the tensor view,
# not the data
self.assertEqual(x, y)
@@ -3505,9 +3502,6 @@
for dt in torch.testing.get_all_dtypes():
x = torch.tensor([[1, 2], [3, 4], [5, 6]], dtype=dt, device=device)
y = torch.tensor([[1, 2, 3], [4, 5, 6]], dtype=dt, device=device)
- if (device == 'cuda' and dt == torch.bfloat16):
- self.assertRaises(RuntimeError, lambda: x.resize_as_(y))
- continue
x.resize_as_(y)
self.assertEqual(y.shape, x.shape)
@@ -3536,11 +3530,11 @@
for device in torch.testing.get_all_device_types():
for dt in torch.testing.get_all_dtypes():
x = torch.tensor((1, 1), dtype=dt, device=device)
- if (device == 'cuda' and dt == torch.bfloat16):
- self.assertRaises(RuntimeError, lambda: x.clone())
- continue
-
y = x.clone()
+ if (device == 'cuda' and dt == torch.bfloat16):
+ # `x - y` is used inside of the assertEqual
+ self.assertRaises(RuntimeError, lambda: x - y)
+ continue
self.assertEqual(x, y)
def test_cat_all_dtypes_and_devices(self):
