[WIP] Partially Codegen Atan2 (#4184)

* partially codegen atan2

* clang-format-7

* clang-format-7 again

Author: ManfeiBai

torch_xla/csrc/aten_xla_type.cpp

@@ -688,11 +688,15 @@ at::Tensor XLANativeFunctions::atan2(const at::Tensor& self,
     return at::native::call_fallback_fn<&xla_cpu_fallback,
                                         ATEN_OP(atan2)>::call(self, other);
   }
-  return DoBinaryOp(self, other,
-                    [&](const XLATensorPtr& xself, const XLATensorPtr& xother,
-                        at::ScalarType dtype) {
-                      return XLATensor::atan2(xself, xother, dtype);
-                    });
+
+  auto common_device = torch_xla::bridge::GetXlaDevice(self, other);
+  XLA_CHECK(common_device);
+  torch::lazy::NodePtr node =
+      torch::lazy::MakeNode<Atan2>(bridge::GetXlaTensor(self)->GetIrValue(),
+                                   bridge::GetXlaTensor(other)->GetIrValue());
+
+  return torch_xla::bridge::AtenFromXlaTensor(
+      torch_xla::XLATensor::Create(std::move(node), *common_device));
 }
 
 at::Tensor XLANativeFunctions::avg_pool2d(

torch_xla/csrc/ops/ops.cpp

@@ -77,7 +77,6 @@ PTXLA_UNARY_OP(Sqrt, at::aten::sqrt, xla::Sqrt);
 PTXLA_BINARY_OP(Min, at::aten::min, xla::Min);
 PTXLA_BINARY_OP(Pow, at::aten::pow, xla::Pow);
 PTXLA_BINARY_OP(Fmod, at::aten::fmod, xla::Rem);
-PTXLA_BINARY_OP(Atan2, at::aten::atan2, xla::Atan2);
 
 torch::lazy::NodePtr LogBase(const torch::lazy::Value& input,
                              torch::lazy::OpKind op, double base) {

torch_xla/csrc/ops/ops.h

@@ -75,9 +75,6 @@ torch::lazy::NodePtr Sin(const torch::lazy::Value& input);
 
 torch::lazy::NodePtr Sinh(const torch::lazy::Value& input);
 
-torch::lazy::NodePtr Atan2(const torch::lazy::Value& input,
-                           const torch::lazy::Value& other);
-
 torch::lazy::NodePtr Tan(const torch::lazy::Value& input);
 
 torch::lazy::NodePtr Neg(const torch::lazy::Value& input);

torch_xla/csrc/ops/ops_lower_fn.cpp

@@ -120,6 +120,13 @@ torch_xla::XlaOpVector Atan::Lower(LoweringContext* loctx) const {
   return ReturnOp(xla::Atan(xla_input), loctx);
 }
 
+torch_xla::XlaOpVector Atan2::Lower(LoweringContext* loctx) const {
+  xla::XlaOp xla_input = loctx->GetOutputOp(operand(0));
+  xla::XlaOp xla_other = loctx->GetOutputOp(operand(1));
+  auto promoted = XlaHelpers::Promote(xla_input, xla_other);
+  return ReturnOp(xla::Atan2(promoted.first, promoted.second), loctx);
+}
+
 torch_xla::XlaOpVector Atanh::Lower(LoweringContext* loctx) const {
   xla::XlaOp xla_input = loctx->GetOutputOp(operand(0));
   return ReturnOp(xla::Atanh(xla_input), loctx);

torch_xla/csrc/ops/ops_xla_shape_fn.cpp

@@ -193,6 +193,17 @@ xla::Shape AtanOutputShape(const torch::lazy::Value& input) {
   return GetXlaShape(input);
 }
 
+xla::Shape Atan2OutputShape(const torch::lazy::Value& input,
+                            const torch::lazy::Value& other) {
+  auto lower_for_shape_fn =
+      [&](absl::Span<const xla::XlaOp> operands) -> xla::XlaOp {
+    auto promoted = XlaHelpers::Promote(operands[0], operands[1]);
+    return xla::Atan2(promoted.first, promoted.second);
+  };
+  return InferOutputShape({GetXlaShape(input), GetXlaShape(other)},
+                          lower_for_shape_fn);
+}
+
 xla::Shape AtanhOutputShape(const torch::lazy::Value& input) {
   return GetXlaShape(input);
 }

torch_xla/csrc/ops/ops_xla_shape_fn.h

@@ -53,6 +53,9 @@ xla::Shape AsinhOutputShape(const torch::lazy::Value& input);
 
 xla::Shape AtanOutputShape(const torch::lazy::Value& input);
 
+xla::Shape Atan2OutputShape(const torch::lazy::Value& input,
+                            const torch::lazy::Value& other);
+
 xla::Shape AtanhOutputShape(const torch::lazy::Value& input);
 
 xla::Shape BinaryCrossEntropyOutputShape(

torch_xla/csrc/tensor.h

@@ -428,10 +428,6 @@ class XLATensor : public c10::intrusive_ptr_target {
                           std::vector<int64_t> stride,
                           c10::optional<int64_t> storage_offset);
 
-  static XLATensorPtr atan2(
-      const XLATensorPtr& input, const XLATensorPtr& other,
-      c10::optional<at::ScalarType> logical_element_type = c10::nullopt);
-
   static XLATensorPtr avg_pool_nd(const XLATensorPtr& input,
                                   int64_t spatial_dim_count,
                                   std::vector<int64_t> kernel_size,

torch_xla/csrc/tensor_methods.cpp

@@ -711,13 +711,6 @@ void XLATensor::as_strided_(XLATensorPtr& input, std::vector<int64_t> size,
   }
 }
 
-XLATensorPtr XLATensor::atan2(
-    const XLATensorPtr& input, const XLATensorPtr& other,
-    c10::optional<at::ScalarType> logical_element_type) {
-  return input->CreateFrom(Atan2(input->GetIrValue(), other->GetIrValue()),
-                           logical_element_type);
-}
-
 XLATensorPtr XLATensor::avg_pool_nd(const XLATensorPtr& input,
                                     int64_t spatial_dim_count,
                                     std::vector<int64_t> kernel_size,

xla_native_functions.yaml

@@ -87,6 +87,7 @@ ir_gen:
   - _adaptive_avg_pool2d_backward
   - _adaptive_avg_pool3d
   - _adaptive_avg_pool3d_backward
+  - atan2
   - bitwise_and.Tensor
   - bitwise_or.Tensor
   - bitwise_xor.Tensor