Promote int to float for tanh operation (consistent with Pytorch)

pytorch · Dec 14, 2023 · 5a6e57f · 5a6e57f
1 parent c4f8772
commit 5a6e57f
Show file tree

Hide file tree

Showing 4 changed files with 24 additions and 2 deletions.
diff --git a/test/cpp/test_aten_xla_tensor_2.cpp b/test/cpp/test_aten_xla_tensor_2.cpp
@@ -2411,6 +2411,20 @@ TEST_F(AtenXlaTensorTest, TestTanh) {
   });
 }
 
+// In torch, tanh works with integer inputs. The same should be true for
+// torch_xla
+TEST_F(AtenXlaTensorTest, TestTanhWithInt) {
+  torch::Tensor a = torch::rand({2, 2});
+  torch::Tensor b = torch::tanh(a);
+  ForEachDevice([&](const torch::Device& device) {
+    torch::Tensor xla_a = CopyToDevice(a, device);
+    torch::Tensor xla_b = torch::tanh(xla_a);
+    AllClose(b, xla_b, /*rtol=*/1e-3, /*atol=*/1e-5);
+  });
+  ExpectCounterNotChanged("aten::.*", cpp_test::GetIgnoredCounters());
+  ExpectCounterChanged("xla::tanh", cpp_test::GetIgnoredCounters());
+}
+
 TEST_F(AtenXlaTensorTest, TestClampMinMax) {
   torch::Tensor a = torch::rand({2, 2}, torch::TensorOptions(torch::kFloat));
   torch::Scalar min_val(0.311);

diff --git a/test/test_core_aten_ops.py b/test/test_core_aten_ops.py
@@ -4419,7 +4419,6 @@ def test_aten_tanh_1(self):
     kwargs = dict()
     run_export_and_compare(self, torch.ops.aten.tanh, args, kwargs)
 
-  @unittest.skip
   def test_aten_tanh_2(self):
     args = (torch.randint(0, 10, (10, 10)).to(torch.int32),)
     kwargs = dict()

diff --git a/torch_xla/csrc/ops/ops_lower_fn.cpp b/torch_xla/csrc/ops/ops_lower_fn.cpp
@@ -751,6 +751,11 @@ torch_xla::XlaOpVector Tan::Lower(LoweringContext* loctx) const {
 
 torch_xla::XlaOpVector Tanh::Lower(LoweringContext* loctx) const {
   xla::XlaOp xla_input = loctx->GetOutputOp(operand(0));
+  if (xla::primitive_util::IsIntegralType(XlaHelpers::TypeOfXlaOp(xla_input))) {
+    xla::PrimitiveType input_type = XlaHelpers::TypeOfXlaOp(xla_input);
+    xla_input = ConvertTo(xla_input, input_type, xla::PrimitiveType::F32,
+                          /*device=*/nullptr);
+  }
   return ReturnOp(xla::Tanh(xla_input), loctx);
 }
 

diff --git a/torch_xla/csrc/ops/ops_xla_shape_fn.cpp b/torch_xla/csrc/ops/ops_xla_shape_fn.cpp
@@ -818,7 +818,11 @@ xla::Shape TakeOutputShape(const torch::lazy::Value& input,
 }
 
 xla::Shape TanhOutputShape(const torch::lazy::Value& input) {
-  return GetXlaShape(input);
+  xla::Shape result_shape = GetXlaShape(input);
+  if (xla::primitive_util::IsIntegralType(result_shape.element_type())) {
+    result_shape.set_element_type(xla::PrimitiveType::F32);
+  }
+  return result_shape;
 }
 
 xla::Shape TrilOutputShape(const torch::lazy::Value& input) {