// 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>  // For std::generate, std::min.
#include <array>      // For std::array.
#include <cmath>      // For std::lrintf.
#include <cstddef>    // For size_t.
#include <cstdint>    // For uint32_t.
#include <limits>     // For std::numeric_limits.
#include <memory>     // For std::unique_ptr.
#include <numeric>    // For std::accumulate.
#include <random>     // For std::random_device, std::mt19937, std::uniform_real_distribution.
#include <vector>     // For std::vector.

#include <xnnpack.h>
#include <xnnpack/operator.h>
#include <xnnpack/requantization.h>
#include <xnnpack/subgraph.h>

#include <gtest/gtest.h>

template <class T, class BiasType = T> class FullyConnectedTestBase : public ::testing::Test {
protected:
  FullyConnectedTestBase()
  {
    random_device = std::unique_ptr<std::random_device>(new std::random_device());
    rng = std::mt19937((*random_device)());
    input_size_dist = std::uniform_int_distribution<uint32_t>(10, 15);
    kernel_size_dist = std::uniform_int_distribution<uint32_t>(1, 5);
    stride_dist = std::uniform_int_distribution<uint32_t>(1, 2);
    f32dist = std::uniform_real_distribution<float>(0.1f, 1.0f);
    scale_dist = std::uniform_real_distribution<float>(1.0f, 5.0f);
    i32dist = std::uniform_int_distribution<int32_t>(-10000, 10000);
    auto shape_dist = std::uniform_int_distribution<size_t>(2, XNN_MAX_TENSOR_DIMS);
    dim_dist = std::uniform_int_distribution<size_t>(5, 15);
    i8dist =
      std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max());
    w8dist =
      std::uniform_int_distribution<int32_t>(-std::numeric_limits<uint8_t>::max(), std::numeric_limits<uint8_t>::max());

    output_min = -std::numeric_limits<float>::infinity();
    output_max = std::numeric_limits<float>::infinity();

    size_t num_input_dims = shape_dist(rng);
    input_dims = RandomShape(num_input_dims);
    assert(input_dims.size() >= 2);
    output_channels = dim_dist(rng);
    input_channels = input_dims.back();
    kernel_dims = {output_channels, input_channels};
    output_dims = input_dims;
    output_dims[output_dims.size() - 1] = output_channels;

    batch_size = NumElements(input_dims) / input_channels;

    input = std::vector<T>(XNN_EXTRA_BYTES / sizeof(T) + NumElements(input_dims));
    kernel = std::vector<T>(input_channels * output_channels);
    bias = std::vector<BiasType>(output_channels);
    operator_output = std::vector<T>(NumElements(output_dims));
    subgraph_output = std::vector<T>(operator_output.size());
    accumulators = std::vector<int32_t>(batch_size * output_channels);
  }

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

  size_t NumElements(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<uint32_t> input_size_dist;
  std::uniform_int_distribution<uint32_t> kernel_size_dist;
  std::uniform_int_distribution<uint32_t> stride_dist;
  std::uniform_int_distribution<int32_t> i32dist;
  std::uniform_real_distribution<float> f32dist;
  std::uniform_real_distribution<float> scale_dist;
  std::uniform_int_distribution<size_t> dim_dist;
  std::uniform_int_distribution<int32_t> i8dist;
  std::uniform_int_distribution<int32_t> u8dist;
  std::uniform_int_distribution<int32_t> w8dist;

  uint32_t batch_size;
  size_t input_channels;
  size_t output_channels;

  float output_min;
  float output_max;

  std::vector<size_t> input_dims;
  std::vector<size_t> kernel_dims;
  std::vector<size_t> bias_dims;
  std::vector<size_t> output_dims;

  std::vector<T> input;
  std::vector<T> kernel;
  std::vector<BiasType> bias;
  std::vector<T> operator_output;
  std::vector<T> subgraph_output;
  std::vector<int32_t> accumulators;
};

template <class T> class QuantizedFullyConnectedTestBase : public FullyConnectedTestBase<T, int32_t> {
protected:
  void initialize_accumulators_from_bias()
  {
    for (size_t i = 0; i < this->batch_size; i++) {
      for (size_t oc = 0; oc < this->output_channels; oc++) {
        this->accumulators[i * this->output_channels + oc] = this->bias[oc];
      }
    }
  }
};

using FullyConnectedTestQS8 = QuantizedFullyConnectedTestBase<int8_t>;
using FullyConnectedTestQU8 = QuantizedFullyConnectedTestBase<uint8_t>;
using FullyConnectedTestF32 = FullyConnectedTestBase<float>;

TEST_F(FullyConnectedTestQS8, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(4, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_qint8, 0, 1.0f, input_dims.size(), input_dims.data(), nullptr,
                          /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t kernel_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_qint8, 0, 1.0f, kernel_dims.size(), kernel_dims.data(), kernel.data(),
                          /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_qint32, 0, 1.0f, bias_dims.size(), bias_dims.data(), bias.data(),
                          /*external_id=*/2, /*flags=*/0, &bias_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_qint8, 0, 1.0f, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, /*flags=*/0, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_fully_connected(subgraph, output_min, output_max, input_id, kernel_id, bias_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_fully_connected);
  ASSERT_EQ(node->compute_type, xnn_compute_type_qs8);
  ASSERT_EQ(node->activation.output_min, output_min);
  ASSERT_EQ(node->activation.output_max, output_max);
  ASSERT_EQ(node->num_inputs, 3);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->inputs[1], kernel_id);
  ASSERT_EQ(node->inputs[2], bias_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(FullyConnectedTestQU8, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(4, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_quint8, 0, 1.0f, input_dims.size(), input_dims.data(), nullptr,
                          /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t kernel_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_quint8, 0, 1.0f, kernel_dims.size(), kernel_dims.data(), kernel.data(),
                          /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_qint32, 0, 1.0f, bias_dims.size(), bias_dims.data(), bias.data(),
                          /*external_id=*/2, /*flags=*/0, &bias_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_quint8, 0, 1.0f, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, /*flags=*/0, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(
    xnn_status_success, xnn_define_fully_connected(
                          subgraph, output_min, output_max, input_id, kernel_id, bias_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_fully_connected);
  ASSERT_EQ(node->compute_type, xnn_compute_type_qu8);
  ASSERT_EQ(node->activation.output_min, output_min);
  ASSERT_EQ(node->activation.output_max, output_max);
  ASSERT_EQ(node->num_inputs, 3);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->inputs[1], kernel_id);
  ASSERT_EQ(node->inputs[2], bias_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(FullyConnectedTestF32, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(4, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(), nullptr,
                          /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t kernel_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_tensor_value(
      subgraph, xnn_datatype_fp32, kernel_dims.size(), kernel_dims.data(), kernel.data(), /*external_id=*/1,
      /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, bias_dims.size(), bias_dims.data(), bias.data(),
                          /*external_id=*/2, /*flags=*/0, &bias_id));

  uint32_t 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,
                          /*external_id=*/3, /*flags=*/0, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_fully_connected(subgraph, output_min, output_max, input_id, kernel_id, bias_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_fully_connected);
  ASSERT_EQ(node->compute_type, xnn_compute_type_fp32);
  ASSERT_EQ(node->activation.output_min, output_min);
  ASSERT_EQ(node->activation.output_max, output_max);
  ASSERT_EQ(node->num_inputs, 3);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->inputs[1], kernel_id);
  ASSERT_EQ(node->inputs[2], bias_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(FullyConnectedTestQS8, matches_operator_api)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_operator_t op = nullptr;

  std::generate(input.begin(), input.end(), [&]() { return i8dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return i32dist(rng); });
  std::fill(operator_output.begin(), operator_output.end(), INT8_C(0xA5));
  std::fill(subgraph_output.begin(), subgraph_output.end(), INT8_C(0xA5));
  const int8_t input_zero_point = -1;
  const float input_scale = scale_dist(rng);
  const float kernel_scale = scale_dist(rng);

  // Compute reference results, without renormalization.
  initialize_accumulators_from_bias();
  for (size_t i = 0; i < batch_size; i++) {
    for (size_t oc = 0; oc < output_channels; oc++) {
      for (size_t ic = 0; ic < input_channels; ic++) {
        accumulators[i * output_channels + oc] +=
          (int32_t(input[i * input_channels + ic]) - int32_t(input_zero_point)) *
          int32_t(kernel[oc * input_channels + ic]);
      }
    }
  }

  // Compute renormalization parameters.
  const int32_t accumulated_min = *std::min_element(accumulators.cbegin(), accumulators.cend());
  const int32_t accumulated_max = *std::max_element(accumulators.cbegin(), accumulators.cend());

  float output_scale = double(uint32_t(accumulated_max - accumulated_min)) / 255.0;
  int8_t output_zero_point = int8_t(std::max(
    std::min(
      lrint(-0.5 - 0.5 * double(accumulated_min + accumulated_max) / output_scale),
      long(std::numeric_limits<int8_t>::max())),
    long(std::numeric_limits<int8_t>::min())));
  const int8_t quantized_output_min = xnn_qs8_quantize(output_min, output_scale, output_zero_point);
  const int8_t quantized_output_max = xnn_qs8_quantize(output_max, output_scale, output_zero_point);

  // Call operator API.
  const xnn_status status = xnn_create_fully_connected_nc_qs8(
    input_channels, output_channels, input_channels, output_channels, input_zero_point, input_scale, kernel_scale,
    kernel.data(), bias.data(), output_zero_point, output_scale, quantized_output_min, quantized_output_max,
    /*flags=*/0, nullptr, &op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);
  ASSERT_EQ(
    xnn_status_success, xnn_setup_fully_connected_nc_qs8(
                          op, batch_size, input.data(), operator_output.data(),
                          /*threadpool=*/nullptr));

  ASSERT_EQ(xnn_status_success, xnn_run_operator(op, /*threadpool=*/nullptr));

  // Call subgraph API.
  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(4, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_qint8, input_zero_point, input_scale, input_dims.size(),
                          input_dims.data(), nullptr, /*external_id=*/0, XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t kernel_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_qint8, 0, kernel_scale, kernel_dims.size(), kernel_dims.data(),
                          kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_qint32, 0, kernel_scale, bias_dims.size(), bias_dims.data(),
                          bias.data(), /*external_id=*/2, /*flags=*/0, &bias_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_qint8, output_zero_point, output_scale, output_dims.size(),
                          output_dims.data(), nullptr, /*external_id=*/3, XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_fully_connected(subgraph, output_min, output_max, input_id, kernel_id, bias_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, 2> external = {
    xnn_external_value{input_id, input.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.
  for (size_t i = 0; i < operator_output.size(); i++) {
    ASSERT_EQ(subgraph_output[i], operator_output[i]);
  }
}

TEST_F(FullyConnectedTestQU8, matches_operator_api)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_operator_t op = nullptr;

  std::generate(input.begin(), input.end(), [&]() { return u8dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return u8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return i32dist(rng); });
  std::fill(operator_output.begin(), operator_output.end(), UINT8_C(0xA5));
  std::fill(subgraph_output.begin(), subgraph_output.end(), UINT8_C(0xA5));
  const uint8_t input_zero_point = u8dist(rng);
  const uint8_t kernel_zero_point = 0;
  const float input_scale = scale_dist(rng);
  const float kernel_scale = scale_dist(rng);

  // Compute reference results, without renormalization.
  initialize_accumulators_from_bias();
  for (size_t i = 0; i < batch_size; i++) {
    for (size_t oc = 0; oc < output_channels; oc++) {
      for (size_t ic = 0; ic < input_channels; ic++) {
        accumulators[i * output_channels + oc] +=
          (int32_t(input[i * input_channels + ic]) - int32_t(input_zero_point)) *
          (int32_t(kernel[oc * input_channels + ic]) - int32_t(kernel_zero_point));
      }
    }
  }

  // Compute renormalization parameters.
  const int32_t accumulated_min = *std::min_element(accumulators.cbegin(), accumulators.cend());
  const int32_t accumulated_max = *std::max_element(accumulators.cbegin(), accumulators.cend());

  const double output_scale = double(uint32_t(accumulated_max - accumulated_min)) / 255.0;
  const uint8_t output_zero_point = uint8_t(std::max(
    std::min(
      lrint(127.5 - 0.5 * double(accumulated_min + accumulated_max) / output_scale),
      long(std::numeric_limits<uint8_t>::max())),
    long(std::numeric_limits<uint8_t>::min())));
  const uint8_t quantized_output_min = xnn_qu8_quantize(output_min, output_scale, output_zero_point);
  const uint8_t quantized_output_max = xnn_qu8_quantize(output_max, output_scale, output_zero_point);

  // Call operator API.
  const xnn_status status = xnn_create_fully_connected_nc_qu8(
    input_channels, output_channels, input_channels, output_channels, input_zero_point, input_scale, kernel_zero_point,
    kernel_scale, kernel.data(), bias.data(), output_zero_point, output_scale, quantized_output_min,
    quantized_output_max, /*flags=*/0, nullptr, &op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);
  ASSERT_EQ(
    xnn_status_success, xnn_setup_fully_connected_nc_qu8(
                          op, batch_size, input.data(), operator_output.data(),
                          /*threadpool=*/nullptr));

  ASSERT_EQ(xnn_status_success, xnn_run_operator(op, /*threadpool=*/nullptr));

  // Call subgraph API.
  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(4, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_quint8, input_zero_point, input_scale, input_dims.size(),
                          input_dims.data(), nullptr, /*external_id=*/0, XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t kernel_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_quint8, 0, kernel_scale, kernel_dims.size(), kernel_dims.data(),
                          kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_qint32, 0, kernel_scale, bias_dims.size(), bias_dims.data(),
                          bias.data(), /*external_id=*/2, /*flags=*/0, &bias_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_quantized_tensor_value(
                          subgraph, xnn_datatype_quint8, output_zero_point, output_scale, output_dims.size(),
                          output_dims.data(), nullptr, /*external_id=*/3, XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_fully_connected(subgraph, output_min, output_max, input_id, kernel_id, bias_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, 2> external = {
    xnn_external_value{input_id, input.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.
  for (size_t i = 0; i < operator_output.size(); i++) {
    ASSERT_EQ(subgraph_output[i], operator_output[i]);
  }
}

TEST_F(FullyConnectedTestF32, matches_operator_api)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_operator_t op = nullptr;

  std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return f32dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));

  // Call operator API.
  const xnn_status status = xnn_create_fully_connected_nc_f32(
    input_channels, output_channels, input_channels, output_channels, kernel.data(), bias.data(), output_min,
    output_max,
    /*flags=*/0, nullptr, &op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);
  ASSERT_EQ(
    xnn_status_success, xnn_setup_fully_connected_nc_f32(
                          op, batch_size, input.data(), operator_output.data(),
                          /*threadpool=*/nullptr));

  ASSERT_EQ(xnn_status_success, xnn_run_operator(op, /*threadpool=*/nullptr));

  // Call subgraph API.
  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(4, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  uint32_t input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(), nullptr,
                          /*external_id=*/0, XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t kernel_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, kernel_dims.size(), kernel_dims.data(), kernel.data(),
                          /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, bias_dims.size(), bias_dims.data(), bias.data(),
                          /*external_id=*/2, /*flags=*/0, &bias_id));

  uint32_t 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,
                          /*external_id=*/3, XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_fully_connected(subgraph, output_min, output_max, input_id, kernel_id, bias_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, 2> external = {
    xnn_external_value{input_id, input.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.
  for (size_t i = 0; i < operator_output.size(); i++) {
    ASSERT_EQ(subgraph_output[i], operator_output[i]);
  }
}