Fix type promotion for trace on CPU.

aten/src/ATen/native/ReduceOps.cpp

@@ -1,6 +1,7 @@
 #include <ATen/native/ReduceOps.h>
 
 #include <ATen/ATen.h>
+#include <ATen/AccumulateType.h>
 #include <ATen/ExpandUtils.h>
 #include <ATen/NativeFunctions.h>
 #include <ATen/WrapDimUtils.h>
@@ -473,6 +474,40 @@ static Tensor& prod_out_impl(Tensor& result, const Tensor& self, IntArrayRef dim
   return result;
 }
 
+// NOTE: this could be implemented via diag and sum, but this has perf problems,
+// see https://github.com/pytorch/pytorch/pull/47305,
+Tensor trace_cpu(const Tensor& self) {
+  Tensor result;
+  ScalarType dtype = get_dtype(result, self, c10::nullopt, true);
+  result = at::empty({}, self.options().dtype(dtype));
+  AT_DISPATCH_ALL_TYPES(self.scalar_type(), "trace", [&] {
+    using accscalar_t = at::acc_type<scalar_t, false>;
+    accscalar_t sum = 0;
+    const auto* t_data = self.data_ptr<scalar_t>();
+
+    int64_t t_stride_0, t_stride_1, t_diag_size;
+
+    TORCH_CHECK(self.dim() == 2, "trace: expected a matrix, but got tensor with dim ", self.dim());
+
+    t_stride_0 = self.stride(0);
+    t_stride_1 = self.stride(1);
+
+    t_diag_size = std::min(self.size(0), self.size(1));
+    for (int64_t i = 0; i < t_diag_size; i++) {
+      sum += t_data[i * (t_stride_0 + t_stride_1)];
+    }
+
+    // all integer types get promoted to kLong
+    if (result.scalar_type() == at::kLong) {
+      *result.data_ptr<int64_t>() = sum;
+    } else {
+      *result.data_ptr<scalar_t>() = sum;
+    }
+  });
+
+  return result;
+}
+
 Tensor prod(const Tensor& self, int64_t dim, bool keepdim, c10::optional<ScalarType> dtype) {
   Tensor result;
   native::prod_out_impl(result, self, dim, keepdim, dtype);

aten/src/ATen/native/TensorShape.cpp

@@ -1985,29 +1985,4 @@ Tensor movedim(const Tensor& self, int64_t src, int64_t dst) {
   return at::movedim(self, IntArrayRef{src}, IntArrayRef{dst});
 }
 
-Tensor trace_cpu(const Tensor& self) {
-  Tensor result = at::empty({}, self.options());
-  AT_DISPATCH_ALL_TYPES(self.scalar_type(), "trace", [&] {
-    using accscalar_t = at::acc_type<scalar_t, false>;
-    accscalar_t sum = 0;
-    const auto* t_data = self.data_ptr<scalar_t>();
-
-    int64_t t_stride_0, t_stride_1, t_diag_size;
-
-    TORCH_CHECK(self.dim() == 2, "trace: expected a matrix, but got tensor with dim ", self.dim());
-
-    t_stride_0 = self.stride(0);
-    t_stride_1 = self.stride(1);
-
-    t_diag_size = std::min(self.size(0), self.size(1));
-    for (int64_t i = 0; i < t_diag_size; i++) {
-      sum += t_data[i * (t_stride_0 + t_stride_1)];
-    }
-
-    *result.data_ptr<scalar_t>() = sum;
-  });
-
-  return result;
-}
-
 }} // at::native

test/test_torch.py

@@ -6048,15 +6048,14 @@ def test_diagonal_multidim(self, device, dtype):
         self.assertEqual(expected.shape, result.shape)
         self.assertEqual(expected, result)
 
-    def _test_trace(self, device, dtype, legacy):
+    @onlyOnCPUAndCUDA
+    @dtypesIfCPU(*torch.testing.get_all_dtypes(include_complex=False, include_bool=False, include_half=False,
+                                               include_bfloat16=False))
+    @dtypesIfCUDA(*torch.testing.get_all_dtypes(include_complex=False, include_bool=False, include_bfloat16=False))
+    def test_trace(self, device, dtype):
         def test(shape):
             tensor = make_tensor(shape, device, dtype, low=-9, high=9)
-            diag = tensor.diag()
-            if legacy:
-                # NB: trace on cpu doesn't do type promotion... #47127
-                expected_dtype = dtype
-            else:
-                expected_dtype = tensor.sum().dtype
+            expected_dtype = tensor.sum().dtype
             expected_dtype = torch_to_numpy_dtype_dict[expected_dtype]
 
             result = np.trace(tensor.cpu().numpy(), dtype=expected_dtype)
@@ -6073,16 +6072,6 @@ def test(shape):
         for shape in shapes:
             test(shape)
 
-    @onlyCPU
-    @dtypes(*torch.testing.get_all_dtypes(include_complex=False, include_bool=False, include_half=False, include_bfloat16=False))
-    def test_trace_legacy(self, device, dtype):
-        self._test_trace(device, dtype, legacy=True)
-
-    @onlyCUDA
-    @dtypes(*torch.testing.get_all_dtypes(include_complex=False, include_bool=False, include_bfloat16=False))
-    def test_trace(self, device, dtype):
-        self._test_trace(device, dtype, legacy=False)
-
     @onlyCPU
     @dtypes(torch.float)
     def test_broadcast_tensors(self, device, dtype):