kernels/test/op_permute_copy_test.cpp

/*
 * 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/test/FunctionHeaderWrapper.h> // Declares the operator
#include <executorch/kernels/test/TestUtil.h>
#include <executorch/kernels/test/supported_features.h>
#include <executorch/runtime/core/exec_aten/exec_aten.h>
#include <executorch/runtime/core/exec_aten/testing_util/tensor_factory.h>
#include <executorch/runtime/core/exec_aten/testing_util/tensor_util.h>

#include <gtest/gtest.h>

using namespace ::testing;
using exec_aten::ArrayRef;
using exec_aten::IntArrayRef;
using exec_aten::ScalarType;
using exec_aten::Tensor;
using torch::executor::testing::TensorFactory;

class OpPermuteCopyTest : public OperatorTest {
 protected:
  Tensor&
  op_permute_copy_out(const Tensor& self, IntArrayRef dims, Tensor& out) {
    return torch::executor::aten::permute_copy_outf(context_, self, dims, out);
  }
};

TEST_F(OpPermuteCopyTest, OneDPermute) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {0};

  const std::vector<int32_t> sizes = {2};
  Tensor t_int = tf.make(sizes, {1, 2});

  Tensor out = tf.zeros(sizes);

  op_permute_copy_out(
      t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out);
  EXPECT_TENSOR_EQ(out, tf.make(sizes, {1, 2}));
}

TEST_F(OpPermuteCopyTest, PermuteWithNoDataReorder) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {1, 0, 2};

  // clang-format off
  Tensor t_int = tf.make({1,4,5}, {
    0,  1,  2,  3,  4,
    5,  6,  7,  8,  9,
    10, 11, 12, 13, 14,
    15, 16, 17,18, 19});
  // clang-format on

  const std::vector<int32_t> new_sizes = {4, 1, 5};
  Tensor out = tf.zeros(new_sizes);

  op_permute_copy_out(
      t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out);
  // clang-format off
  EXPECT_TENSOR_EQ(out, tf.make(new_sizes, {
    0,  1,  2,  3,  4,
    5,  6,  7,  8,  9,
    10, 11, 12, 13, 14,
    15, 16, 17, 18, 19}));
  // clang-format on
}

TEST_F(OpPermuteCopyTest, TwoDPermute) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {1, 0};

  // clang-format off
  Tensor t_int = tf.make({2, 3}, {
    // 2x3 data block
    0, 1, 2,
    3, 4, 5
  });
  // clang-format on

  const std::vector<int32_t> new_sizes = {3, 2};
  Tensor out = tf.zeros(new_sizes);

  op_permute_copy_out(
      t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out);
  // clang-format off
  EXPECT_TENSOR_EQ(out, tf.make(new_sizes, {
    // 3x2 data block
    0, 3,
    1, 4,
    2, 5
  }));
  // clang-format on
}

TEST_F(OpPermuteCopyTest, ThreeDPermute) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {2, 0, 1};

  // clang-format off
  Tensor t_int = tf.make({2, 1, 3}, {
    // 2 1x3 data blocks
    0, 1, 2,

    3, 4, 5
  });
  // clang-format on

  const std::vector<int32_t> new_sizes = {3, 2, 1};
  Tensor out = tf.zeros(new_sizes);

  op_permute_copy_out(
      t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out);
  // clang-format off
  EXPECT_TENSOR_EQ(out, tf.make(new_sizes, {
  // 3 2x1 data blocks
    0,
    3,

    1,
    4,

    2,
    5
  }));
  // clang-format on
}

TEST_F(OpPermuteCopyTest, FourDPermute) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {0, 3, 2, 1};

  // clang-format off
  Tensor t_int = tf.make(
      {2, 3, 3, 4},
      // 2 groupings of 3 3x4 data blocks
      {0,  1,  2,  3,
       4,  5,  6,  7,
       8,  9,  10, 11,

       12, 13, 14, 15,
       16, 17, 18, 19,
       20, 21, 22, 23,

       24, 25, 26, 27,
       28, 29, 30, 31,
       32, 33, 34, 35,


       36, 37, 38, 39,
       40, 41, 42, 43,
       44, 45, 46, 47,

       48, 49, 50, 51,
       52, 53, 54, 55,
       56, 57, 58, 59,

       60, 61, 62, 63,
       64, 65, 66, 67,
       68, 69, 70, 71});
  // clang-format on

  const std::vector<int32_t> new_sizes = {2, 4, 3, 3};
  Tensor out = tf.zeros(new_sizes);

  // Long results like this are gotten by running torch.permute in a notebook
  // and copy pasting the result. Ex:
  //   import torch
  //   x = torch.arange(0, 72, 1).view(2, 3, 3, 4).contiguous()
  //   print(x.flatten().contiguous())
  //   z = torch.permute(x, (0, 3, 2, 1))
  //   print(z.flatten().contiguous())
  op_permute_copy_out(
      t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out);
  EXPECT_TENSOR_EQ(
      out,
      // clang-format off
      tf.make(new_sizes, {
        // 2 groupings of 4 3x3 data blocks
        0, 12, 24,
        4, 16, 28,
        8, 20, 32,

        1, 13, 25,
        5, 17, 29,
        9, 21, 33,

        2, 14, 26,
        6, 18, 30,
        10, 22, 34,

        3, 15, 27,
        7, 19, 31,
        11, 23, 35,


        36, 48, 60,
        40, 52, 64,
        44, 56, 68,

        37, 49, 61,
        41, 53, 65,
        45, 57, 69,

        38, 50, 62,
        42, 54, 66,
        46, 58, 70,

        39, 51, 63,
        43, 55, 67,
        47, 59, 71}));
  // clang-format on
}

TEST_F(OpPermuteCopyTest, FiveDPermute) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {4, 3, 2, 1, 0};

  const std::vector<int32_t> sizes = {2, 2, 2, 2, 2};
  // clang-format off
  Tensor t_int = tf.make(
      sizes, {
        0,  1,
        2,  3,

        4,  5,
        6,  7,


        8,  9,
        10, 11,

        12, 13,
        14, 15,


        16, 17,
        18, 19,

        20, 21,
        22, 23,


        24, 25,
        26, 27,

        28, 29,
        30, 31});
  // clang-format on

  Tensor out = tf.zeros(sizes);

  op_permute_copy_out(
      t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out);
  // Long results like this are gotten by running torch.permute in a notebook
  // and copy pasting the result. Ex:
  //   import torch
  //   x = torch.arange(0, 32, 1).view(2, 2, 2, 2, 2).contiguous()
  //   print(x.flatten().contiguous())
  //   z = torch.permute(x, (4, 3, 2, 1, 0))
  //   print(z.flatten().contiguous())
  // clang-format off
  EXPECT_TENSOR_EQ(
      out, tf.make(sizes, {
        0,  16,
        8,  24,

        4,  20,
        12, 28,


        2,  18,
        10, 26,

        6,  22,
        14, 30,


        1,  17,
        9,  25,

        5,  21,
        13, 29,


        3,  19,
        11, 27,

        7,  23,
        15, 31}));
  // clang-format on
}

TEST_F(OpPermuteCopyTest, AllDimensionsSizeOne) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {4, 3, 2, 1, 0};

  const std::vector<int32_t> sizes = {1, 1, 1, 1, 1};
  Tensor t_int = tf.make(sizes, {1});

  Tensor out = tf.zeros(sizes);

  op_permute_copy_out(
      t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out);
  EXPECT_TENSOR_EQ(out, tf.make(sizes, {1}));
}

TEST_F(OpPermuteCopyTest, DupeDimensionPos) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {0, 1, 1};

  const std::vector<int32_t> sizes = {1, 1, 1};
  Tensor t_int = tf.make(sizes, {1});

  Tensor out = tf.zeros(sizes);

  ET_EXPECT_KERNEL_FAILURE(
      context_,
      op_permute_copy_out(
          t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out));
}

TEST_F(OpPermuteCopyTest, DupeDimensionPos2) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {1, 1, 1};

  const std::vector<int32_t> sizes = {1, 1, 1};
  Tensor t_int = tf.make(sizes, {1});

  Tensor out = tf.zeros(sizes);

  ET_EXPECT_KERNEL_FAILURE(
      context_,
      op_permute_copy_out(
          t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out));
}

TEST_F(OpPermuteCopyTest, DupeDimensionNeg) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {0, 1, -2};

  const std::vector<int32_t> sizes = {1, 1, 1};
  Tensor t_int = tf.make(sizes, {1});

  Tensor out = tf.zeros(sizes);

  ET_EXPECT_KERNEL_FAILURE(
      context_,
      op_permute_copy_out(
          t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out));
}

TEST_F(OpPermuteCopyTest, DupeDimensionNeg2) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {0, 1, -5};

  const std::vector<int32_t> sizes = {1, 1, 1};
  Tensor t_int = tf.make(sizes, {1});

  Tensor out = tf.zeros(sizes);

  ET_EXPECT_KERNEL_FAILURE(
      context_,
      op_permute_copy_out(
          t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out));
}

TEST_F(OpPermuteCopyTest, MismatchDim) {
  TensorFactory<ScalarType::Int> tf;

  const std::vector<int64_t> new_dim = {0, 1, 2};

  const std::vector<int32_t> sizes = {1, 1};
  Tensor t_int = tf.make(sizes, {1});

  Tensor out = tf.zeros(sizes);

  ET_EXPECT_KERNEL_FAILURE(
      context_,
      op_permute_copy_out(
          t_int, ArrayRef<int64_t>(new_dim.data(), new_dim.size()), out));
}

/* %python
import torch
torch.manual_seed(0)
x = torch.randint(10, (2, 3, 4))
res = torch.permute(x, (2, 0, 1))
op = "op_permute_copy_out"
opt_setup_params = f"""
  {declare_array_ref([2, 0, 1], "int64_t", "perm_aref")}
"""
opt_extra_params = "perm_aref,"
dtype = "ScalarType::Int"
check = "EXPECT_TENSOR_EQ" */

TEST_F(OpPermuteCopyTest, DynamicShapeUpperBoundSameAsExpected) {
  /* %python
  out_args = "{4, 2, 3}, torch::executor::TensorShapeDynamism::DYNAMIC_BOUND"
  %rewrite(unary_op) */

  TensorFactory<ScalarType::Int> tf;

  Tensor x = tf.make({2, 3, 4}, {4, 9, 3, 0, 3, 9, 7, 3, 7, 3, 1, 6,
                                 6, 9, 8, 6, 6, 8, 4, 3, 6, 9, 1, 4});
  Tensor expected = tf.make({4, 2, 3}, {4, 3, 7, 6, 6, 6, 9, 9, 3, 9, 8, 9,
                                        3, 7, 1, 8, 4, 1, 0, 3, 6, 6, 3, 4});

  std::vector<int64_t> perm_arefv = {2, 0, 1};
  ArrayRef<int64_t> perm_aref(perm_arefv.data(), perm_arefv.size());

  Tensor out =
      tf.zeros({4, 2, 3}, torch::executor::TensorShapeDynamism::DYNAMIC_BOUND);
  op_permute_copy_out(x, perm_aref, out);
  EXPECT_TENSOR_EQ(out, expected);
}

TEST_F(OpPermuteCopyTest, DynamicShapeUpperBoundLargerThanExpected) {
  /* %python
  out_args = "{5, 5, 5}, torch::executor::TensorShapeDynamism::DYNAMIC_BOUND"
  %rewrite(unary_op) */

  TensorFactory<ScalarType::Int> tf;

  Tensor x = tf.make({2, 3, 4}, {4, 9, 3, 0, 3, 9, 7, 3, 7, 3, 1, 6,
                                 6, 9, 8, 6, 6, 8, 4, 3, 6, 9, 1, 4});
  Tensor expected = tf.make({4, 2, 3}, {4, 3, 7, 6, 6, 6, 9, 9, 3, 9, 8, 9,
                                        3, 7, 1, 8, 4, 1, 0, 3, 6, 6, 3, 4});

  std::vector<int64_t> perm_arefv = {2, 0, 1};
  ArrayRef<int64_t> perm_aref(perm_arefv.data(), perm_arefv.size());

  Tensor out =
      tf.zeros({5, 5, 5}, torch::executor::TensorShapeDynamism::DYNAMIC_BOUND);
  op_permute_copy_out(x, perm_aref, out);
  EXPECT_TENSOR_EQ(out, expected);
}

TEST_F(OpPermuteCopyTest, DynamicShapeUnbound) {
  if (!torch::executor::testing::SupportedFeatures::get()->output_resize) {
    GTEST_SKIP() << "Dynamic shape unbound not supported";
  }
  /* %python
  out_args = "{1, 1, 1}, torch::executor::TensorShapeDynamism::DYNAMIC_UNBOUND"
  %rewrite(unary_op) */

  TensorFactory<ScalarType::Int> tf;

  Tensor x = tf.make({2, 3, 4}, {4, 9, 3, 0, 3, 9, 7, 3, 7, 3, 1, 6,
                                 6, 9, 8, 6, 6, 8, 4, 3, 6, 9, 1, 4});
  Tensor expected = tf.make({4, 2, 3}, {4, 3, 7, 6, 6, 6, 9, 9, 3, 9, 8, 9,
                                        3, 7, 1, 8, 4, 1, 0, 3, 6, 6, 3, 4});

  std::vector<int64_t> perm_arefv = {2, 0, 1};
  ArrayRef<int64_t> perm_aref(perm_arefv.data(), perm_arefv.size());

  Tensor out = tf.zeros(
      {1, 1, 1}, torch::executor::TensorShapeDynamism::DYNAMIC_UNBOUND);
  op_permute_copy_out(x, perm_aref, out);
  EXPECT_TENSOR_EQ(out, expected);
}