Add sort util for 1D tensors (#2786)

Summary:
Pull Request resolved: #2786

This diff adds a simple sort utility that sorts a tensor's values and returns the sorted values and the sorted indices in the out tensors that are provided.
There are currently two limitations to this sort:
- It only supports 1D tensors currently, has to be extended to support 2D and greater tensors.
- Input types are assumed to be float and it currently asserts on that. This has to be templatized to support all dtypes.

Reviewed By: iseeyuan

Differential Revision: D55577025

Author: tarun292

kernels/portable/cpu/util/sort_util.cpp

@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include "executorch/kernels/portable/cpu/util/sort_util.h"
+#include <executorch/runtime/kernel/kernel_includes.h>
+#include <algorithm>
+
+namespace torch {
+namespace executor {
+
+using Tensor = exec_aten::Tensor;
+
+Error sort_tensor(
+    const Tensor& tensor,
+    Tensor& sorted_tensor,
+    Tensor& sorted_indices,
+    bool descending) {
+  // Check if the input tensor is a valid input
+  ET_CHECK_MSG(tensor.dim() == 1, "Input tensor must be 1D");
+
+  // Check if the output tensors are valid
+  ET_CHECK_MSG(sorted_tensor.dim() == 1, "Output tensor must be 1D");
+  ET_CHECK_MSG(sorted_indices.dim() == 1, "Output tensor must be 1D");
+
+  // Check if the output tensors have the same dtype
+  ET_CHECK_MSG(
+      tensor.scalar_type() == sorted_tensor.scalar_type(),
+      "Input and output tensors must have the same dtype");
+  ET_CHECK_MSG(
+      tensor.scalar_type() == ScalarType::Float,
+      "Only float inputs are supported currently");
+  ET_CHECK_MSG(
+      sorted_indices.scalar_type() == exec_aten::ScalarType::Long,
+      "Output tensor must be of type int64");
+
+  // Get the number of elements in the tensor
+  int size = tensor.numel();
+
+  // Create a tensor to store the indices
+  for (int i = 0; i < size; i++) {
+    sorted_indices.mutable_data_ptr<int64_t>()[i] = i;
+  }
+
+  // Sort the indices based on the corresponding tensor values
+  std::sort(
+      sorted_indices.mutable_data_ptr<int64_t>(),
+      sorted_indices.mutable_data_ptr<int64_t>() + size,
+      [&tensor, descending](int64_t i, int64_t j) {
+        if (descending) {
+          return tensor.const_data_ptr<float>()[i] >
+              tensor.const_data_ptr<float>()[j];
+        } else {
+          return tensor.const_data_ptr<float>()[i] <
+              tensor.const_data_ptr<float>()[j];
+        }
+      });
+
+  // Rearrange the tensor values based on the sorted indices
+  for (int i = 0; i < size; i++) {
+    sorted_tensor.mutable_data_ptr<float>()[i] = tensor.const_data_ptr<
+        float>()[sorted_indices.const_data_ptr<int64_t>()[i]];
+  }
+
+  return Error::Ok;
+}
+
+} // namespace executor
+} // namespace torch

kernels/portable/cpu/util/sort_util.h

@@ -0,0 +1,25 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
+
+namespace torch {
+namespace executor {
+
+using Tensor = exec_aten::Tensor;
+
+Error sort_tensor(
+    const Tensor& tensor,
+    Tensor& sorted_tensor,
+    Tensor& sorted_indice,
+    bool descending = false);
+
+} // namespace executor
+} // namespace torch

kernels/portable/cpu/util/targets.bzl

@@ -247,6 +247,17 @@ def define_common_targets():
         visibility = ["//executorch/kernels/portable/cpu/..."],
     )
 
+    runtime.cxx_library(
+        name = "sort_util",
+        srcs = ["sort_util.cpp"],
+        exported_headers = ["sort_util.h"],
+        deps = [
+            "//executorch/runtime/kernel:kernel_includes",
+            "//executorch/runtime/core/exec_aten/util:tensor_util",
+        ],
+        visibility = ["//executorch/kernels/portable/cpu/...", "//executorch/kernels/torchvision/..."],
+    )
+
     # Utility functions that can be used by operators that perform reduction
     for aten_mode in [True, False]:
         suffix = "_aten" if aten_mode else ""

kernels/portable/cpu/util/test/sort_util_test.cpp

@@ -0,0 +1,45 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/kernels/portable/cpu/util/sort_util.h>
+#include <executorch/runtime/core/exec_aten/testing_util/tensor_factory.h>
+#include <executorch/runtime/core/exec_aten/testing_util/tensor_util.h>
+#include <executorch/runtime/core/exec_aten/util/tensor_util.h>
+#include <executorch/test/utils/DeathTest.h>
+
+#include <gtest/gtest.h>
+
+using namespace ::testing;
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
+using torch::executor::ArrayRef;
+using torch::executor::testing::TensorFactory;
+
+TEST(SortUtilTest, SortTensorTest) {
+  TensorFactory<ScalarType::Float> tf;
+  TensorFactory<ScalarType::Long> lf;
+
+  Tensor a = tf.make({4}, {3, 2, 1, 4});
+  Tensor b = tf.zeros({4});
+  Tensor c = lf.zeros({4});
+
+  // Ascending order sort test
+  sort_tensor(a, b, c);
+
+  Tensor expected = tf.make({4}, {1, 2, 3, 4});
+  Tensor expected_indices = lf.make({4}, {2, 1, 0, 3});
+  EXPECT_TENSOR_EQ(b, expected);
+  EXPECT_TENSOR_EQ(c, expected_indices);
+
+  // Descending order sort test
+  sort_tensor(a, b, c, true);
+  expected = tf.make({4}, {4, 3, 2, 1});
+  expected_indices = lf.make({4}, {3, 0, 1, 2});
+  EXPECT_TENSOR_EQ(b, expected);
+  EXPECT_TENSOR_EQ(c, expected_indices);
+}

kernels/portable/cpu/util/test/targets.bzl

@@ -29,5 +29,13 @@ def define_common_targets():
             "//executorch/runtime/core/exec_aten:lib",
             "//executorch/kernels/portable/cpu/util:allocate_tensor_util",
             "//executorch/runtime/kernel:kernel_includes",
+        
+    runtime.cxx_test(
+        name = "sort_util_test",
+        srcs = ["sort_util_test.cpp"],
+        deps = [
+            "//executorch/runtime/core/exec_aten:lib",
+            "//executorch/runtime/core/exec_aten/testing_util:tensor_util",
+            "//executorch/kernels/portable/cpu/util:sort_util",
         ],
     )