// 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 <random>

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

#include "subgraph-unary-tester.h"
#include <gtest/gtest.h>

using ConvertTestF32ToQS8 = UnaryTest<float, int8_t>;
using ConvertTestF32ToQU8 = UnaryTest<float, uint8_t>;
using ConvertTestQS8ToQS8 = UnaryTest<int8_t, int8_t>;
using ConvertTestQS8ToF32 = UnaryTest<int8_t, float>;
using ConvertTestQU8ToQU8 = UnaryTest<uint8_t, uint8_t>;
using ConvertTestQU8ToF32 = UnaryTest<uint8_t, float>;

TEST_F(ConvertTestF32ToQS8, 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=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, 0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_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, dims.size(), dims.data(), nullptr, 1,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_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_convert);
  ASSERT_EQ(node->compute_type, xnn_compute_type_fp32_to_qs8);
  ASSERT_EQ(node->num_inputs, 1);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(ConvertTestF32ToQU8, 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=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, 0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_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, dims.size(), dims.data(), nullptr, 1,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_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_convert);
  ASSERT_EQ(node->compute_type, xnn_compute_type_fp32_to_qu8);
  ASSERT_EQ(node->num_inputs, 1);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(ConvertTestQS8ToF32, 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=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

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

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, 1,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_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_convert);
  ASSERT_EQ(node->compute_type, xnn_compute_type_qs8_to_fp32);
  ASSERT_EQ(node->num_inputs, 1);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(ConvertTestQS8ToQS8, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  const int32_t input_zero_point = i8dist(rng);
  const int32_t output_zero_point = i8dist(rng);
  // Scale distributions chosen to guarantee 2**-8 <= input_scale / output_scale <= 2**7
  const float input_scale = std::uniform_real_distribution<float>(0.0883883f, 11.3137f)(rng);
  const float output_scale = std::uniform_real_distribution<float>(0.0883883f, 11.3137f)(rng);

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  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, dims.size(), dims.data(), nullptr, 0,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_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, output_zero_point, output_scale, dims.size(), dims.data(), nullptr, 1,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_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_convert);
  ASSERT_EQ(node->compute_type, xnn_compute_type_qs8);
  ASSERT_EQ(node->num_inputs, 1);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(ConvertTestQU8ToF32, 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=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

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

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, 1,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_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_convert);
  ASSERT_EQ(node->compute_type, xnn_compute_type_qu8_to_fp32);
  ASSERT_EQ(node->num_inputs, 1);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(ConvertTestQU8ToQU8, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  const int32_t input_zero_point = u8dist(rng);
  const int32_t output_zero_point = u8dist(rng);
  // Scale distributions chosen to guarantee 2**-8 <= input_scale / output_scale <= 2**7
  const float input_scale = std::uniform_real_distribution<float>(0.0883883f, 11.3137f)(rng);
  const float output_scale = std::uniform_real_distribution<float>(0.0883883f, 11.3137f)(rng);

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  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, dims.size(), dims.data(), nullptr, 0,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_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, output_zero_point, output_scale, dims.size(), dims.data(), nullptr, 1,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_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_convert);
  ASSERT_EQ(node->compute_type, xnn_compute_type_qu8);
  ASSERT_EQ(node->num_inputs, 1);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(ConvertTestF32ToQS8, matches_operator_api)
{
  std::uniform_real_distribution<float> f32dist(-1.0f, 1.0f);
  std::generate(input.begin(), input.end(), [&]() { return f32dist(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));

  // Call operator API.
  xnn_operator_t op = nullptr;
  const xnn_status status = xnn_create_convert_nc_f32_qs8(
    channels, channels, channels, scale, signed_zero_point, INT8_MIN, INT8_MAX, /*flags=*/0, &op);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_convert_nc_f32_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(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);
  input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, /*external_id=*/0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_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, dims.size(), dims.data(), nullptr, /*external_id=*/1,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_id, output_id, /*flags=*/0));
  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));

  ASSERT_EQ(subgraph_output, operator_output);
}

TEST_F(ConvertTestF32ToQU8, matches_operator_api)
{
  std::uniform_real_distribution<float> f32dist(-1.0f, 1.0f);
  std::generate(input.begin(), input.end(), [&]() { return f32dist(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));

  // Call operator API.
  xnn_operator_t op = nullptr;
  const xnn_status status = xnn_create_convert_nc_f32_qu8(
    channels, channels, channels, scale, unsigned_zero_point, 0, UINT8_MAX, /*flags=*/0, &op);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_convert_nc_f32_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(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);
  input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, /*external_id=*/0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_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, dims.size(), dims.data(), nullptr, /*external_id=*/1,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_id, output_id, /*flags=*/0));
  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));

  ASSERT_EQ(subgraph_output, operator_output);
}

TEST_F(ConvertTestQS8ToF32, matches_operator_api)
{
  std::generate(input.begin(), input.end(), [&]() { return i8dist(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));

  // Call operator API.
  xnn_operator_t op = nullptr;
  const xnn_status status =
    xnn_create_convert_nc_qs8_f32(channels, channels, channels, scale, signed_zero_point, /*flags=*/0, &op);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_convert_nc_qs8_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(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

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

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, 1,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_id, output_id, /*flags=*/0));
  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));

  ASSERT_EQ(subgraph_output, operator_output);
}

TEST_F(ConvertTestQS8ToQS8, matches_operator_api)
{
  const int8_t input_zero_point = i8dist(rng);
  const int8_t output_zero_point = i8dist(rng);
  const float input_scale = std::uniform_real_distribution<float>(0.25f, 4.0f)(rng);
  const float output_scale = std::uniform_real_distribution<float>(0.25f, 4.0f)(rng);

  std::generate(input.begin(), input.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(0x5A));

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

  // Call operator API.
  xnn_operator_t op = nullptr;
  const xnn_status status =
    xnn_create_convert_nc_qs8(channels, channels, channels, input_scale, input_zero_point, output_scale, output_zero_point, /*flags=*/0, &op);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_convert_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(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  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,
                          dims.size(), dims.data(), nullptr, 0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_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, output_zero_point, output_scale,
                          dims.size(), dims.data(), nullptr, 1,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_id, output_id, /*flags=*/0));
  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);
  const 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));

  ASSERT_EQ(subgraph_output, operator_output);
}

TEST_F(ConvertTestQU8ToF32, matches_operator_api)
{
  std::generate(input.begin(), input.end(), [&]() { return u8dist(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));

  // Call operator API.
  xnn_operator_t op = nullptr;
  const xnn_status status =
    xnn_create_convert_nc_qu8_f32(channels, channels, channels, scale, unsigned_zero_point, /*flags=*/0, &op);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_convert_nc_qu8_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(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

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

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, dims.size(), dims.data(), nullptr, 1,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_id, output_id, /*flags=*/0));
  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));

  ASSERT_EQ(subgraph_output, operator_output);
}

TEST_F(ConvertTestQU8ToQU8, matches_operator_api)
{
  const uint8_t input_zero_point = u8dist(rng);
  const uint8_t output_zero_point = u8dist(rng);
  const float input_scale = std::uniform_real_distribution<float>(0.25f, 4.0f)(rng);
  const float output_scale = std::uniform_real_distribution<float>(0.25f, 4.0f)(rng);

  std::generate(input.begin(), input.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(0x5A));

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

  // Call operator API.
  xnn_operator_t op = nullptr;
  const xnn_status status =
    xnn_create_convert_nc_qu8(channels, channels, channels, input_scale, input_zero_point, output_scale, output_zero_point, /*flags=*/0, &op);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  ASSERT_EQ(
    xnn_status_success,
    xnn_setup_convert_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(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  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,
                          dims.size(), dims.data(), nullptr, 0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_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, output_zero_point, output_scale,
                          dims.size(), dims.data(), nullptr, 1,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_define_convert(subgraph, input_id, output_id, /*flags=*/0));
  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);
  const 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));

  ASSERT_EQ(subgraph_output, operator_output);
}