test/concatenate4.cc

// Copyright 2022 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <algorithm>
#include <array>
#include <cstddef>
#include <cstdint>
#include <limits>
#include <memory>
#include <numeric>
#include <random>

#include <xnnpack.h>
#include <xnnpack/node-type.h>
#include <xnnpack/operator.h>
#include <xnnpack/subgraph.h>

#include <gtest/gtest.h>

template <typename T> class Concatenate4Test : public ::testing::Test {
protected:
  Concatenate4Test()
  {
    random_device = std::unique_ptr<std::random_device>(new std::random_device());
    rng = std::mt19937((*random_device)());
    shape_dist = std::uniform_int_distribution<size_t>(1, XNN_MAX_TENSOR_DIMS);
    dim_dist = std::uniform_int_distribution<size_t>(1, 9);
    f32dist = std::uniform_real_distribution<float>();
    i8dist =
      std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max());
    u8dist =
      std::uniform_int_distribution<int32_t>(std::numeric_limits<uint8_t>::min(), std::numeric_limits<uint8_t>::max());
    scale_dist = std::uniform_real_distribution<float>(0.1f, 5.0f);

    input1_dims = RandomShape();
    axis = RandomAxis(input1_dims);
    input2_dims = RandomShape(input1_dims, axis);
    input3_dims = RandomShape(input1_dims, axis);
    input4_dims = RandomShape(input1_dims, axis);
    output_dims = input1_dims;
    output_dims[axis] = input1_dims[axis] + input2_dims[axis] + input3_dims[axis] + input4_dims[axis];

    input1 = std::vector<T>(NumElements(input1_dims));
    input2 = std::vector<T>(NumElements(input2_dims));
    input3 = std::vector<T>(NumElements(input3_dims));
    input4 = std::vector<T>(NumElements(input4_dims));
    operator_output = std::vector<T>(NumElements(output_dims));
    subgraph_output = std::vector<T>(NumElements(output_dims));

    signed_zero_point = i8dist(rng);
    unsigned_zero_point = u8dist(rng);
    scale = scale_dist(rng);

    batch_size = 1;
    channels_1 = 1;
    channels_2 = 1;
    channels_3 = 1;
    channels_4 = 1;
    for (size_t i = 0; i < axis; i++) {
      batch_size *= output_dims[i];
    }

    for (size_t i = axis; i < input1_dims.size(); i++) {
      channels_1 *= input1_dims[i];
      channels_2 *= input2_dims[i];
      channels_3 *= input3_dims[i];
      channels_4 *= input4_dims[i];
    }
    output_stride = channels_1 + channels_2 + channels_3 + channels_4;
  }

  std::vector<size_t> RandomShape()
  {
    std::vector<size_t> dims(shape_dist(rng));
    std::generate(dims.begin(), dims.end(), [&] { return dim_dist(rng); });
    return dims;
  }

  std::vector<size_t> RandomShape(const std::vector<size_t> base_dims, size_t axis)
  {
    auto dims = base_dims;
    dims[axis] = dim_dist(rng);
    return dims;
  }

  size_t RandomAxis(const std::vector<size_t>& dims)
  {
    return std::uniform_int_distribution<size_t>(0, dims.size() - 1)(rng);
  }

  size_t NumElements(const std::vector<size_t>& dims)
  {
    return std::accumulate(dims.begin(), dims.end(), size_t(1), std::multiplies<size_t>());
  }

  std::unique_ptr<std::random_device> random_device;
  std::mt19937 rng;
  std::uniform_int_distribution<size_t> shape_dist;
  std::uniform_int_distribution<size_t> dim_dist;
  std::uniform_real_distribution<float> f32dist;
  std::uniform_int_distribution<int32_t> i8dist;
  std::uniform_int_distribution<int32_t> u8dist;
  std::uniform_real_distribution<float> scale_dist;

  uint32_t input1_id;
  uint32_t input2_id;
  uint32_t input3_id;
  uint32_t input4_id;
  uint32_t output_id;

  std::vector<size_t> input1_dims;
  std::vector<size_t> input2_dims;
  std::vector<size_t> input3_dims;
  std::vector<size_t> input4_dims;
  std::vector<size_t> output_dims;

  size_t axis;
  size_t batch_size;
  size_t channels_1;
  size_t channels_2;
  size_t channels_3;
  size_t channels_4;
  size_t output_stride;

  int32_t signed_zero_point;
  int32_t unsigned_zero_point;
  float scale;

  std::vector<T> input1;
  std::vector<T> input2;
  std::vector<T> input3;
  std::vector<T> input4;
  std::vector<T> operator_output;
  std::vector<T> subgraph_output;
};

using Concatenate4TestQS8 = Concatenate4Test<int8_t>;
using Concatenate4TestQU8 = Concatenate4Test<uint8_t>;
using Concatenate4TestF32 = Concatenate4Test<float>;

TEST_F(Concatenate4TestQS8, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/5, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  input1_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_qint8, signed_zero_point, scale, input1_dims.size(), input1_dims.data(), nullptr, 0,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input1_id));
  ASSERT_NE(input1_id, XNN_INVALID_NODE_ID);

  input2_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_qint8, signed_zero_point, scale, input2_dims.size(), input2_dims.data(), nullptr, 1,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input2_id));
  ASSERT_NE(input2_id, XNN_INVALID_NODE_ID);

  input3_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_qint8, signed_zero_point, scale, input3_dims.size(), input3_dims.data(), nullptr, 2,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input3_id));
  ASSERT_NE(input3_id, XNN_INVALID_NODE_ID);

  input4_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_qint8, signed_zero_point, scale, input4_dims.size(), input4_dims.data(), nullptr, 3,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input4_id));
  ASSERT_NE(input4_id, XNN_INVALID_NODE_ID);

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_qint8, signed_zero_point, scale, output_dims.size(), output_dims.data(), nullptr, 4,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_concatenate4(subgraph, axis, input1_id, input2_id, input3_id, input4_id, output_id, /*flags=*/0));

  ASSERT_EQ(subgraph->num_nodes, 1);
  const struct xnn_node* node = &subgraph->nodes[0];
  ASSERT_EQ(node->type, xnn_node_type_concatenate4);
  ASSERT_EQ(node->compute_type, xnn_compute_type_qs8);
  ASSERT_EQ(node->params.concatenate.axis, axis);
  ASSERT_EQ(node->num_inputs, 4);
  ASSERT_EQ(node->inputs[0], input1_id);
  ASSERT_EQ(node->inputs[1], input2_id);
  ASSERT_EQ(node->inputs[2], input3_id);
  ASSERT_EQ(node->inputs[3], input4_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(Concatenate4TestQU8, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/5, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  input1_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input1_dims.size(), input1_dims.data(), nullptr, 0,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input1_id));
  ASSERT_NE(input1_id, XNN_INVALID_NODE_ID);

  input2_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input2_dims.size(), input2_dims.data(), nullptr, 1,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input2_id));
  ASSERT_NE(input2_id, XNN_INVALID_NODE_ID);

  input3_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input3_dims.size(), input3_dims.data(), nullptr, 2,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input3_id));
  ASSERT_NE(input3_id, XNN_INVALID_NODE_ID);

  input4_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input4_dims.size(), input4_dims.data(), nullptr, 3,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input4_id));
  ASSERT_NE(input4_id, XNN_INVALID_NODE_ID);

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, output_dims.size(), output_dims.data(), nullptr, 4,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_concatenate4(subgraph, axis, input1_id, input2_id, input3_id, input4_id, output_id, /*flags=*/0));

  ASSERT_EQ(subgraph->num_nodes, 1);
  const struct xnn_node* node = &subgraph->nodes[0];
  ASSERT_EQ(node->type, xnn_node_type_concatenate4);
  ASSERT_EQ(node->compute_type, xnn_compute_type_qu8);
  ASSERT_EQ(node->params.concatenate.axis, axis);
  ASSERT_EQ(node->num_inputs, 4);
  ASSERT_EQ(node->inputs[0], input1_id);
  ASSERT_EQ(node->inputs[1], input2_id);
  ASSERT_EQ(node->inputs[2], input3_id);
  ASSERT_EQ(node->inputs[3], input4_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(Concatenate4TestF32, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/5, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  input1_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input1_dims.size(), input1_dims.data(), nullptr, 0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input1_id));
  ASSERT_NE(input1_id, XNN_INVALID_NODE_ID);

  input2_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input2_dims.size(), input2_dims.data(), nullptr, 1,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input2_id));
  ASSERT_NE(input2_id, XNN_INVALID_NODE_ID);

  input3_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input3_dims.size(), input3_dims.data(), nullptr, 2,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input3_id));
  ASSERT_NE(input3_id, XNN_INVALID_NODE_ID);

  input4_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input4_dims.size(), input4_dims.data(), nullptr, 3,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input4_id));
  ASSERT_NE(input4_id, XNN_INVALID_NODE_ID);

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, output_dims.size(), output_dims.data(), nullptr, 4,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_concatenate4(subgraph, axis, input1_id, input2_id, input3_id, input4_id, output_id, /*flags=*/0));

  ASSERT_EQ(subgraph->num_nodes, 1);
  const struct xnn_node* node = &subgraph->nodes[0];
  ASSERT_EQ(node->type, xnn_node_type_concatenate4);
  ASSERT_EQ(node->compute_type, xnn_compute_type_fp32);
  ASSERT_EQ(node->params.concatenate.axis, axis);
  ASSERT_EQ(node->num_inputs, 4);
  ASSERT_EQ(node->inputs[0], input1_id);
  ASSERT_EQ(node->inputs[1], input2_id);
  ASSERT_EQ(node->inputs[2], input3_id);
  ASSERT_EQ(node->inputs[3], input4_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(Concatenate4TestQS8, matches_operator_api)
{
  std::generate(input1.begin(), input1.end(), [&]() { return i8dist(rng); });
  std::generate(input2.begin(), input2.end(), [&]() { return i8dist(rng); });
  std::generate(input3.begin(), input3.end(), [&]() { return i8dist(rng); });
  std::generate(input4.begin(), input4.end(), [&]() { return i8dist(rng); });
  std::fill(operator_output.begin(), operator_output.end(), INT8_C(0xA5));
  std::fill(subgraph_output.begin(), subgraph_output.end(), INT8_C(0xA5));

  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_operator_t op1 = nullptr;
  xnn_operator_t op2 = nullptr;
  xnn_operator_t op3 = nullptr;
  xnn_operator_t op4 = nullptr;

  // Call operator API.
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels_1, channels_1, output_stride, /*flags=*/0, &op1));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op1(op1, xnn_delete_operator);
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels_2, channels_2, output_stride, /*flags=*/0, &op2));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op2(op2, xnn_delete_operator);
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels_3, channels_3, output_stride, /*flags=*/0, &op3));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op3(op3, xnn_delete_operator);
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels_4, channels_4, output_stride, /*flags=*/0, &op4));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op4(op4, xnn_delete_operator);

  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x8(op1, batch_size, input1.data(), operator_output.data(), nullptr /* thread pool */));
  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x8(
      op2, batch_size, input2.data(), (uint8_t*) operator_output.data() + op1->channels, nullptr /* thread pool */));
  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x8(
      op3, batch_size, input3.data(), (uint8_t*) operator_output.data() + op1->channels + op2->channels,
      nullptr /* thread pool */));
  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x8(
      op4, batch_size, input4.data(), (uint8_t*) operator_output.data() + op1->channels + op2->channels + op3->channels,
      nullptr /* thread pool */));

  ASSERT_EQ(xnn_status_success, xnn_run_operator(op1, nullptr /* thread pool */));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(op2, nullptr /* thread pool */));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(op3, nullptr /* thread pool */));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(op4, nullptr /* thread pool */));

  // Call subgraph API.
  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/5, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  input1_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_qint8, signed_zero_point, scale, input1_dims.size(), input1_dims.data(), nullptr, 0,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input1_id));
  ASSERT_NE(input1_id, XNN_INVALID_NODE_ID);

  input2_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_qint8, signed_zero_point, scale, input2_dims.size(), input2_dims.data(), nullptr, 1,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input2_id));
  ASSERT_NE(input2_id, XNN_INVALID_NODE_ID);

  input3_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_qint8, signed_zero_point, scale, input3_dims.size(), input3_dims.data(), nullptr, 2,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input3_id));
  ASSERT_NE(input3_id, XNN_INVALID_NODE_ID);

  input4_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_qint8, signed_zero_point, scale, input4_dims.size(), input4_dims.data(), nullptr, 3,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input4_id));
  ASSERT_NE(input4_id, XNN_INVALID_NODE_ID);

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_qint8, signed_zero_point, scale, output_dims.size(), output_dims.data(), nullptr, 4,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_concatenate4(subgraph, axis, input1_id, input2_id, input3_id, input4_id, output_id, /*flags=*/0));

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
  ASSERT_NE(nullptr, runtime);
  std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
  std::array<xnn_external_value, 5> external = {
    xnn_external_value{input1_id, input1.data()}, xnn_external_value{input2_id, input2.data()},
    xnn_external_value{input3_id, input3.data()}, xnn_external_value{input4_id, input4.data()},
    xnn_external_value{output_id, subgraph_output.data()}};
  ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

  // Check outputs match.
  ASSERT_EQ(subgraph_output, operator_output);
}

TEST_F(Concatenate4TestQU8, matches_operator_api)
{
  std::generate(input1.begin(), input1.end(), [&]() { return u8dist(rng); });
  std::generate(input2.begin(), input2.end(), [&]() { return u8dist(rng); });
  std::generate(input3.begin(), input3.end(), [&]() { return u8dist(rng); });
  std::generate(input4.begin(), input4.end(), [&]() { return u8dist(rng); });
  std::fill(operator_output.begin(), operator_output.end(), UINT8_C(0xA5));
  std::fill(subgraph_output.begin(), subgraph_output.end(), UINT8_C(0xA5));

  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_operator_t op1 = nullptr;
  xnn_operator_t op2 = nullptr;
  xnn_operator_t op3 = nullptr;
  xnn_operator_t op4 = nullptr;

  // Call operator API.
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels_1, channels_1, output_stride, /*flags=*/0, &op1));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op1(op1, xnn_delete_operator);
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels_2, channels_2, output_stride, /*flags=*/0, &op2));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op2(op2, xnn_delete_operator);
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels_3, channels_3, output_stride, /*flags=*/0, &op3));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op3(op3, xnn_delete_operator);
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels_4, channels_4, output_stride, /*flags=*/0, &op4));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op4(op4, xnn_delete_operator);

  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x8(op1, batch_size, input1.data(), operator_output.data(), nullptr /* thread pool */));
  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x8(
      op2, batch_size, input2.data(), (uint8_t*) operator_output.data() + op1->channels, nullptr /* thread pool */));
  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x8(
      op3, batch_size, input3.data(), (uint8_t*) operator_output.data() + op1->channels + op2->channels,
      nullptr /* thread pool */));
  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x8(
      op4, batch_size, input4.data(), (uint8_t*) operator_output.data() + op1->channels + op2->channels + op3->channels,
      nullptr /* thread pool */));

  ASSERT_EQ(xnn_status_success, xnn_run_operator(op1, nullptr /* thread pool */));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(op2, nullptr /* thread pool */));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(op3, nullptr /* thread pool */));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(op4, nullptr /* thread pool */));

  // Call subgraph API.
  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/5, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  input1_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input1_dims.size(), input1_dims.data(), nullptr, 0,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input1_id));
  ASSERT_NE(input1_id, XNN_INVALID_NODE_ID);

  input2_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input2_dims.size(), input2_dims.data(), nullptr, 1,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input2_id));
  ASSERT_NE(input2_id, XNN_INVALID_NODE_ID);

  input3_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input3_dims.size(), input3_dims.data(), nullptr, 2,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input3_id));
  ASSERT_NE(input3_id, XNN_INVALID_NODE_ID);

  input4_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input4_dims.size(), input4_dims.data(), nullptr, 3,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input4_id));
  ASSERT_NE(input4_id, XNN_INVALID_NODE_ID);

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_quantized_tensor_value(
      subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, output_dims.size(), output_dims.data(), nullptr, 4,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_concatenate4(subgraph, axis, input1_id, input2_id, input3_id, input4_id, output_id, /*flags=*/0));

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
  ASSERT_NE(nullptr, runtime);
  std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
  std::array<xnn_external_value, 5> external = {
    xnn_external_value{input1_id, input1.data()}, xnn_external_value{input2_id, input2.data()},
    xnn_external_value{input3_id, input3.data()}, xnn_external_value{input4_id, input4.data()},
    xnn_external_value{output_id, subgraph_output.data()}};
  ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

  // Check outputs match.
  ASSERT_EQ(subgraph_output, operator_output);
}

TEST_F(Concatenate4TestF32, matches_operator_api)
{
  std::generate(input1.begin(), input1.end(), [&]() { return f32dist(rng); });
  std::generate(input2.begin(), input2.end(), [&]() { return f32dist(rng); });
  std::generate(input3.begin(), input3.end(), [&]() { return f32dist(rng); });
  std::generate(input4.begin(), input4.end(), [&]() { return f32dist(rng); });
  std::fill(operator_output.begin(), operator_output.end(), std::nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), std::nanf(""));

  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_operator_t op1 = nullptr;
  xnn_operator_t op2 = nullptr;
  xnn_operator_t op3 = nullptr;
  xnn_operator_t op4 = nullptr;

  // Call operator API.
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x32(channels_1, channels_1, output_stride, /*flags=*/0, &op1));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op1(op1, xnn_delete_operator);
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x32(channels_2, channels_2, output_stride, /*flags=*/0, &op2));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op2(op2, xnn_delete_operator);
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x32(channels_3, channels_3, output_stride, /*flags=*/0, &op3));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op3(op3, xnn_delete_operator);
  ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x32(channels_4, channels_4, output_stride, /*flags=*/0, &op4));
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op4(op4, xnn_delete_operator);

  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x32(op1, batch_size, input1.data(), operator_output.data(), nullptr /* thread pool */));
  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x32(
      op2, batch_size, input2.data(), (float*) operator_output.data() + op1->channels, nullptr /* thread pool */));
  ASSERT_EQ(
    xnn_status_success, xnn_setup_copy_nc_x32(
                          op3, batch_size, input3.data(),
                          (float*) operator_output.data() + op1->channels + op2->channels, nullptr /* thread pool */));
  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_copy_nc_x32(
      op4, batch_size, input4.data(), (float*) operator_output.data() + op1->channels + op2->channels + op3->channels,
      nullptr /* thread pool */));

  ASSERT_EQ(xnn_status_success, xnn_run_operator(op1, nullptr /* thread pool */));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(op2, nullptr /* thread pool */));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(op3, nullptr /* thread pool */));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(op4, nullptr /* thread pool */));

  // Call subgraph API.
  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/5, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  input1_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input1_dims.size(), input1_dims.data(), nullptr, 0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input1_id));
  ASSERT_NE(input1_id, XNN_INVALID_NODE_ID);

  input2_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input2_dims.size(), input2_dims.data(), nullptr, 1,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input2_id));
  ASSERT_NE(input2_id, XNN_INVALID_NODE_ID);

  input3_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input3_dims.size(), input3_dims.data(), nullptr, 2,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input3_id));
  ASSERT_NE(input3_id, XNN_INVALID_NODE_ID);

  input4_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input4_dims.size(), input4_dims.data(), nullptr, 3,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input4_id));
  ASSERT_NE(input4_id, XNN_INVALID_NODE_ID);

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, output_dims.size(), output_dims.data(), nullptr, 4,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_concatenate4(subgraph, axis, input1_id, input2_id, input3_id, input4_id, output_id, /*flags=*/0));

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
  ASSERT_NE(nullptr, runtime);
  std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
  std::array<xnn_external_value, 5> external = {
    xnn_external_value{input1_id, input1.data()}, xnn_external_value{input2_id, input2.data()},
    xnn_external_value{input3_id, input3.data()}, xnn_external_value{input4_id, input4.data()},
    xnn_external_value{output_id, subgraph_output.data()}};
  ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

  // Check outputs match.
  ASSERT_EQ(subgraph_output, operator_output);
}