Decompose/add reference for view_as_complex (#108005)

Aten source: https://github.com/pytorch/pytorch/blob/d4a99631dd589afbb972b3401e56f029192c9b0f/aten/src/ATen/native/ComplexHelper.h#L78

Documentation reference:
https://pytorch.org/docs/stable/generated/torch.view_as_complex.html

Note: this adds a new primitive `view_of_dtype`, which is trivially implemented, as its meta function is already implemented elsewhere.

Finally, this is not registered as a decomposition (yet), because TorchInductor does not yet support complex types. It should be added once we do.

Closes #108020 as well.

Pull Request resolved: #108005
Approved by: https://github.com/peterbell10, https://github.com/ezyang

aten/src/ATen/native/TensorConversions.cpp

@@ -695,15 +695,15 @@ Tensor sparse_compressed_to_dense(
 
 // Computes the strides for view_dtype output when the view dtype is
 // smaller than the original dtype
-inline DimVector compute_strides_for_view_dtype_downsize(IntArrayRef old_strides, int64_t size_ratio, ScalarType old_dtype, ScalarType new_dtype) {
+inline SymDimVector compute_strides_for_view_dtype_downsize(SymIntArrayRef old_strides, int64_t size_ratio, ScalarType old_dtype, ScalarType new_dtype) {
   const int64_t ndim = old_strides.size();
 
   TORCH_CHECK(
     old_strides[ndim - 1] == 1,
     "self.stride(-1) must be 1 to view ", old_dtype, " as ", new_dtype,
     " (different element sizes), but got ", old_strides[ndim - 1]);
 
-  DimVector new_strides(ndim);
+  SymDimVector new_strides(ndim);
   for (int64_t dim_idx = 0; dim_idx < ndim - 1; dim_idx++) {
     new_strides[dim_idx] = old_strides[dim_idx] * size_ratio;
   }
@@ -713,14 +713,14 @@ inline DimVector compute_strides_for_view_dtype_downsize(IntArrayRef old_strides
 
 // Computes the strides for view_dtype output when the view dtype is
 // larger than the original dtype
-inline DimVector compute_strides_for_view_dtype_upsize(IntArrayRef old_strides, int64_t size_ratio, ScalarType old_dtype, ScalarType new_dtype) {
+inline SymDimVector compute_strides_for_view_dtype_upsize(SymIntArrayRef old_strides, int64_t size_ratio, ScalarType old_dtype, ScalarType new_dtype) {
   const int64_t ndim = old_strides.size();
   TORCH_CHECK(
     old_strides[ndim - 1] == 1,
     "self.stride(-1) must be 1 to view ", old_dtype, " as ", new_dtype,
     " (different element sizes), but got ", old_strides[ndim - 1]);
 
-  DimVector new_strides(ndim);
+  SymDimVector new_strides(ndim);
   for (int64_t dim_idx = 0; dim_idx < ndim - 1; dim_idx++) {
     TORCH_CHECK(
       (old_strides[dim_idx] % size_ratio) == 0,
@@ -753,8 +753,7 @@ Tensor view_dtype(const Tensor& self, ScalarType dtype) {
   auto* impl = new_tensor.unsafeGetTensorImpl();
 
   if (self_element_size == new_element_size) {
-    impl->set_storage_offset(self.storage_offset());
-    impl->set_sizes_and_strides(self.sizes(), self.strides());
+    impl->set_sizes_and_strides(self.sym_sizes(), self.sym_strides(), self.sym_storage_offset());
 
   } else if (self.dim() == 0) {
     TORCH_CHECK(false,
@@ -766,47 +765,45 @@ Tensor view_dtype(const Tensor& self, ScalarType dtype) {
 
     int64_t size_ratio = self_element_size / new_element_size;
     auto new_strides = compute_strides_for_view_dtype_downsize(
-      self.strides(), size_ratio, self.scalar_type(), dtype);
+      self.sym_strides(), size_ratio, self.scalar_type(), dtype);
 
-    auto old_sizes = self.sizes();
-    DimVector new_sizes(self.dim());
+    auto old_sizes = self.sym_sizes();
+    SymDimVector new_sizes(self.dim());
     std::copy(old_sizes.begin(), old_sizes.end(), new_sizes.begin());
     new_sizes[self.dim() - 1] *= size_ratio;
 
-    auto new_storage_offset = size_ratio * self.storage_offset();
+    auto new_storage_offset = size_ratio * self.sym_storage_offset();
 
-    impl->set_storage_offset(new_storage_offset);
-    impl->set_sizes_and_strides(new_sizes, new_strides);
+    impl->set_sizes_and_strides(new_sizes, new_strides, new_storage_offset);
 
   } else {
     // Upsizing element size
 
     int64_t size_ratio = new_element_size / self_element_size;
 
     TORCH_CHECK(
-      (self.size(-1) % size_ratio) == 0,
+      (self.sym_size(-1) % size_ratio) == 0,
       "self.size(-1) must be divisible by ", size_ratio, " to view ",
       self.scalar_type(), " as ", dtype, " (different element sizes), ",
-      "but got ", self.size(-1));
+      "but got ", self.sym_size(-1));
 
     TORCH_CHECK(
-      (self.storage_offset() % size_ratio) == 0,
+      (self.sym_storage_offset() % size_ratio) == 0,
       "self.storage_offset() must be divisible by ", size_ratio, " to view ",
       self.scalar_type(), " as ", dtype, " (different element sizes), but got ",
-      self.storage_offset());
+      self.sym_storage_offset());
 
     auto new_strides = compute_strides_for_view_dtype_upsize(
-      self.strides(), size_ratio, self.scalar_type(), dtype);
+      self.sym_strides(), size_ratio, self.scalar_type(), dtype);
 
-    auto old_sizes = self.sizes();
-    DimVector new_sizes(self.dim());
+    auto old_sizes = self.sym_sizes();
+    SymDimVector new_sizes(self.dim());
     std::copy(old_sizes.begin(), old_sizes.end(), new_sizes.begin());
     new_sizes[self.dim() - 1] /= size_ratio;
 
-    auto new_storage_offset = self.storage_offset() / size_ratio;
+    auto new_storage_offset = self.sym_storage_offset() / size_ratio;
 
-    impl->set_storage_offset(new_storage_offset);
-    impl->set_sizes_and_strides(new_sizes, new_strides);
+    impl->set_sizes_and_strides(new_sizes, new_strides, new_storage_offset);
   }
 
   return new_tensor;

test/test_ops.py

@@ -1876,6 +1876,7 @@ class TestRefsOpsInfo(TestCase):
         '_refs.imag',
         '_refs.reshape_as',
         '_refs.view_as',
+        '_refs.view_as_complex'  # TorchInductor does not support complex at the moment.
     }
 
     @parametrize("op", ref_ops_names)

torch/_decomp/__init__.py

@@ -373,6 +373,7 @@ def core_aten_decompositions() -> Dict[torch._ops.OperatorBase, Callable]:
             aten._unsafe_index,
             aten.upsample_bilinear2d,
             aten.upsample_nearest2d_backward,
+            aten.view_as_complex,
             aten.xlogy,
             aten.xlogy_,
             aten.zero,

torch/_inductor/lowering.py

@@ -1886,6 +1886,7 @@ def apply_constraint(arg, fx_arg):
 make_fallback(aten.upsample_bicubic2d_backward, require_contiguous)
 make_fallback(aten._scaled_mm.default)
 
+# TODO: This is done, just need to enable support in TorchInductor for complex types.
 make_fallback(aten.view_as_complex, require_contiguous)
 
 # The following were added as a result of https://github.com/pytorch/pytorch/pull/94039 to pass tests

torch/_meta_registrations.py

@@ -2642,11 +2642,6 @@ def meta_complex(real, imag):
     return real.new_empty(out_shape, dtype=corresponding_complex_dtype(real.dtype))
 
 
-@register_meta(aten.view.dtype)
-def view_dtype(self, dtype):
-    return utils.clone_preserve_strides(self).to(dtype)
-
-
 @register_meta([aten.nonzero_static.default, aten.nonzero_static.out])
 def nonzero_static(self, *, size: int, fill_value: int = -1):
     return self.new_empty((size, self.dim()), dtype=torch.long)

torch/_prims/__init__.py

@@ -148,6 +148,7 @@
     "squeeze",
     "transpose",
     "view_of",
+    "view_element_type",
     #
     # Functionalized view mutations
     #
@@ -172,7 +173,6 @@
     "item",
     "maximum_value",
     "minimum_value",
-    "to_dtype",
     "copy_strided",
     #
     # Inplace prims
@@ -1780,6 +1780,27 @@ def _view_of_aten(a: Tensor) -> Tensor:
     doc=_view_of_doc,
 )
 
+
+def _view_element_type_meta(a: TensorLikeType, dtype: torch.dtype) -> TensorLikeType:
+    return a.view(dtype)
+
+
+def _view_element_type_aten(a: Tensor, dtype: torch.dtype) -> Tensor:
+    return a.view(dtype)
+
+
+_view_element_type_doc = """
+    Creates a view of the tensor with a different dtype.
+    """
+
+view_element_type = _make_prim(
+    schema="view_of_dtype(Tensor(a) a, ScalarType dtype) -> Tensor",
+    meta=_view_element_type_meta,
+    impl_aten=_view_element_type_aten,
+    return_type=RETURN_TYPE.VIEW,
+    doc=_view_element_type_doc,
+)
+
 #
 # Functionalized view mutations
 #

torch/_refs/__init__.py

@@ -289,6 +289,7 @@
     "view_as",
     "vsplit",
     "vstack",
+    "view_as_complex",
     "unflatten",
     "unbind",
     "triu",
@@ -949,6 +950,43 @@ def trunc(a):
     return prims.trunc(a)
 
 
+# TODO: register this as a real ref/decomposition once TorchInductor supports complex!
+def view_as_complex(self: TensorLikeType) -> TensorLikeType:
+    input_dtype = self.dtype
+    torch._check(
+        utils.is_float_dtype(input_dtype),
+        lambda: f"view_as_complex is only supported for floating point"
+        f"tensors, but got a tensor of scalar type: {input_dtype}",
+    )
+    sizes = self.size()
+    torch._check(
+        len(sizes) != 0,
+        lambda: "Input tensor must have one or more dimensions",
+    )
+    torch._check(
+        sizes[-1] == 2,
+        lambda: "Tensor must have a last dimension of size 2",
+    )
+
+    old_strides = self.stride()
+    torch._check(
+        old_strides[-1] == 1,
+        lambda: "Tensor must have a last dimension with stride 1",
+    )
+    dims = old_strides[:-1]
+    torch._check(
+        py_all(stride % 2 == 0 for stride in dims),
+        lambda: "Tensor must have a stride divisible by 2 for all but last dimension",
+    )
+    torch._check(
+        self.storage_offset() % 2 == 0,
+        lambda: "Tensor must have a storage_offset divisible by 2",
+    )
+    return prims.view_element_type(
+        self, utils.corresponding_complex_dtype(input_dtype)
+    ).squeeze(-1)
+
+
 def _make_elementwise_binary_reference(
     type_promotion_kind,
     aten_op=infer_aten_op,

torch/testing/_internal/common_methods_invocations.py

@@ -21576,6 +21576,10 @@ def reference_flatten(input, start_dim=0, end_dim=-1):
             ),
         ],
     ),
+    PythonRefInfo(
+        "_refs.view_as_complex",
+        torch_opinfo_name="view_as_complex",
+    ),
 ]
 python_ref_db += opinfo.definitions.python_ref_db