pin_memory support for NT (#110404)

jbschlosser · pytorchmergebot · commit 3597325bc7f0 · 2023-10-05T16:33:22.000Z
Pull Request resolved: #110404 Approved by: https://github.com/cpuhrsch, https://github.com/albanD ghstack dependencies: #110292
diff --git a/aten/src/ATen/native/native_functions.yaml b/aten/src/ATen/native/native_functions.yaml
@@ -4405,7 +4405,7 @@
 - func: is_pinned(Tensor self, Device? device=None) -> bool
   variants: method
   dispatch:
-    CUDA: is_pinned_cuda
+    NestedTensorCUDA, CUDA: is_pinned_cuda
     MPS: is_pinned_mps
     CompositeExplicitAutograd: is_pinned_default
 
@@ -4419,6 +4419,7 @@
   dispatch:
     CUDA: _pin_memory_cuda
     MPS: _pin_memory_mps
+    NestedTensorCUDA, NestedTensorCPU: _pin_memory_nested
   autogen: _pin_memory.out
 
 - func: pinverse(Tensor self, float rcond=1e-15) -> Tensor
diff --git a/aten/src/ATen/native/nested/NestedTensorUnaryOps.cpp b/aten/src/ATen/native/nested/NestedTensorUnaryOps.cpp
@@ -132,5 +132,15 @@ Tensor cos_nested(const Tensor& self) {
   return map_nt(self, at::cos);
 }
 
+Tensor _pin_memory_nested(const Tensor& self, c10::optional<Device> device) {
+  auto* nt_input = get_nested_tensor_impl(self);
+  const auto& input_buffer = nt_input->get_unsafe_storage_as_tensor();
+  return wrap_buffer(
+      at::_pin_memory(input_buffer, device),
+      nt_input->get_nested_sizes(),
+      nt_input->get_nested_strides(),
+      nt_input->get_storage_offsets());
+}
+
 } // namespace native
 } // namespace at
diff --git a/aten/src/ATen/templates/RegisterBackendSelect.cpp b/aten/src/ATen/templates/RegisterBackendSelect.cpp
@@ -36,6 +36,11 @@ bool is_pinned(const Tensor& self, c10::optional<at::Device> device) {
 at::Tensor _pin_memory(const Tensor& self, c10::optional<at::Device> device) {
   TORCH_CHECK(self.device().is_cpu(), "cannot pin '", self.toString(), "' only dense CPU tensors can be pinned");
   DispatchKeySet _dk = c10::DispatchKeySet(c10::computeDispatchKey(c10::nullopt, self.layout(), device.value_or(at::kCUDA)));
+  if (self.is_nested()) {
+    constexpr auto nested_key_set = c10::DispatchKeySet(
+        {c10::DispatchKey::NestedTensor, c10::DispatchKey::AutogradNestedTensor});
+    _dk = _dk.add(self.key_set() & nested_key_set);
+  }
   return at::_ops::_pin_memory::redispatch(_dk, self, device);
 }
 
diff --git a/test/test_dataloader.py b/test/test_dataloader.py
@@ -2282,16 +2282,22 @@ def test_nested_tensor_multiprocessing(self, device, context):
 
         dataset = [torch.nested.nested_tensor([torch.randn(5)], device=device) for _ in range(10)]
 
-        loader = torch.utils.data.DataLoader(
-            dataset,
-            batch_size=1,
-            num_workers=4,
-            collate_fn=_clone_collate,
-            multiprocessing_context=context,
-        )
+        pin_memory_settings = [False]
+        if device == 'cpu' and torch.cuda.is_available():
+            pin_memory_settings.append(True)
+
+        for pin_memory in pin_memory_settings:
+            loader = torch.utils.data.DataLoader(
+                dataset,
+                batch_size=1,
+                num_workers=4,
+                collate_fn=_clone_collate,
+                pin_memory=pin_memory,
+                multiprocessing_context=context,
+            )
 
-        for i, batch in enumerate(loader):
-            self.assertEqual(batch[0], dataset[i])
+            for i, batch in enumerate(loader):
+                self.assertEqual(batch[0], dataset[i])
 
         # Error case: default collate_fn doesn't currently support batches of nested tensors.
         # Following the current semantics, we'd need to stack them, which isn't possible atm.
diff --git a/test/test_nestedtensor.py b/test/test_nestedtensor.py
@@ -15,6 +15,7 @@
     onlyCPU,
     onlyCUDA,
     skipMeta,
+    PYTORCH_CUDA_MEMCHECK,
 )
 from torch.testing._internal.common_dtype import floating_types_and_half
 from torch.testing._internal.common_utils import (
@@ -2941,6 +2942,20 @@ def compare_metadata(nt1, nt2):
             self.assertEqual(b, nt_contiguous)
             self.assertEqual(b, nt_noncontiguous)
 
+    @unittest.skipIf(PYTORCH_CUDA_MEMCHECK, "is_pinned uses failure to detect pointer property")
+    @onlyCUDA
+    def test_pin_memory(self, device):
+        nt_contiguous, nt_noncontiguous = random_nt_noncontiguous_pair((2, 3, 6, 7))
+        for nt in [nt_contiguous, nt_noncontiguous]:
+            self.assertFalse(nt.is_pinned())
+            pinned = nt.pin_memory(device)
+            self.assertTrue(pinned.is_pinned())
+            self.assertEqual(nt, pinned)
+            self.assertNotEqual(nt.data_ptr(), pinned.data_ptr())
+            # test that pin_memory on already pinned tensor has no effect
+            self.assertIs(pinned, pinned.pin_memory())
+            self.assertEqual(pinned.data_ptr(), pinned.pin_memory().data_ptr())
+
 
 instantiate_parametrized_tests(TestNestedTensor)
 instantiate_device_type_tests(TestNestedTensorDeviceType, globals())
