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

#include <gmock/gmock.h>
#include <gtest/gtest.h>
#include <fp16/fp16.h>
#include "xnnpack.h"
#include "xnnpack/aligned-allocator.h"
#include "xnnpack/common.h"
#include "xnnpack/config.h"
#include "xnnpack/internal.h"
#include "xnnpack/math.h"
#include "xnnpack/node-type.h"
#include "xnnpack/operator.h"
#include "xnnpack/packq.h"
#include "xnnpack/requantization.h"
#include "xnnpack/subgraph.h"
#include "replicable_random_device.h"

using testing::ElementsAreArray;

template <class InputType, class KernelType = InputType,
          class BiasType = InputType, class OutputType = InputType, bool even_channels = false>
class FullyConnectedTestBase : public ::testing::TestWithParam<bool> {
 protected:
  FullyConnectedTestBase() {
    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();
    // Adjust number of kernel elements for QC4W. input_channels should be padded to byte boundary, hence even.
    if (even_channels) {
      input_channels = round_up_po2(input_channels, 2);
      input_dims.back() = input_channels;
    }
    kernel_dims = {output_channels, input_channels};
    kernel_dims_tranposed = {input_channels, output_channels};
    bias_dims = {output_channels};
    output_dims = input_dims;
    output_dims[output_dims.size() - 1] = output_channels;

    batch_size = NumElements(input_dims) / input_channels;

    input = std::vector<InputType>(XNN_EXTRA_BYTES / sizeof(InputType) + NumElements(input_dims));
    kernel = std::vector<KernelType>(input_channels * output_channels);
    kernel_fp16 = std::vector<xnn_float16>(input_channels * output_channels);
    bias = std::vector<BiasType>(output_channels);
    bias_fp16 = std::vector<xnn_float16>(output_channels);
    operator_output = std::vector<OutputType>(NumElements(output_dims));
    subgraph_output = std::vector<OutputType>(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>());
  }

  xnnpack::ReplicableRandomDevice rng;
  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> kernel_dims_tranposed;
  std::vector<size_t> bias_dims;
  std::vector<size_t> output_dims;

  std::vector<InputType> input;
  std::vector<KernelType> kernel;
  std::vector<BiasType> bias;
  std::vector<xnn_float16> kernel_fp16;
  std::vector<xnn_float16> bias_fp16;
  std::vector<OutputType> operator_output;
  std::vector<OutputType> subgraph_output;
  std::vector<int32_t> accumulators;
};

class FullyConnectedTestQP8F32QC4W
    : public FullyConnectedTestBase<int8_t, uint8_t, float, float, true> {};

TEST_F(FullyConnectedTestQP8F32QC4W, define) {
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  if (xnn_init_qp8_f32_qc4w_gemm_config() == nullptr) {
    GTEST_SKIP();
  }

  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_VALUE_ID;
  ASSERT_EQ(xnn_status_success,
            xnn_define_dynamically_quantized_tensor_value(
                subgraph, xnn_datatype_qpint8, input_dims.size(),
                /*num_nonbatch_dims=*/1, input_dims.data(),
                /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  // Adjust number of kernel elements for QC4W. input_channels should be padded
  // to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);

  const uint8_t kernel_zero_point = 8;
  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(),
                [&]() { return scale_dist(rng); });
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(xnn_status_success,
            xnn_define_channelwise_quantized_tensor_value_v2(
                subgraph, xnn_datatype_qcint4, kernel_zero_point,
                kernel_scale.data(), kernel_dims.size(),
                /*channel_dim=*/0, kernel_dims.data(), kernel.data(),
                /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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_qp8_to_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(FullyConnectedTestQP8F32QC4W, matches_qd8_f32_qc4w) {
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  if (xnn_init_qp8_f32_qc4w_gemm_config() == nullptr) {
    GTEST_SKIP();
  }

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/4,
                                                    /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(
      subgraph, xnn_delete_subgraph);
  std::vector<float> convert_input(batch_size * input_channels +
                                   XNN_EXTRA_BYTES / sizeof(float));
  std::vector<int8_t> operator_qp8_data(
      xnn_x8_packq_f32qp8_gemm_packed_size(batch_size, input_channels) +
      XNN_EXTRA_BYTES);
  std::vector<int8_t> operator_qd8_data(batch_size * input_channels +
                                        XNN_EXTRA_BYTES);
  std::vector<float> qp8_operator_output(batch_size * output_channels);
  std::vector<float> qd8_operator_output(batch_size * output_channels);
  std::fill(qp8_operator_output.begin(), qp8_operator_output.end(), nanf(""));
  std::fill(qd8_operator_output.begin(), qd8_operator_output.end(), nanf(""));

  // Adjust number of kernel elements for QC4W. input_channels should be padded
  // to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);

  std::vector<float> kernel_scale(output_channels);
  std::vector<xnn_dynamic_quantization_params> quantization_params(
      batch_size + XNN_EXTRA_QUANTIZATION_PARAMS);
  std::generate(kernel_scale.begin(), kernel_scale.end(),
                [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::generate(convert_input.begin(), convert_input.end(),
                [&]() { return f32dist(rng); });

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

  const uint8_t kernel_zero_point = 8;

  // Call operator API for `qp8`.
  xnn_operator_t qp8_convert_op = nullptr;
  xnn_operator_t qp8_fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f32_qp8(
      /*flags=*/0, &qp8_convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)>
      auto_qp8_convert_op(qp8_convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, qp8_convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f32_qp8(
                                    qp8_convert_op, batch_size, input_channels,
                                    input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_convert_nc_f32_qp8(qp8_convert_op, convert_input.data(),
                                         operator_qp8_data.data()));
  ASSERT_EQ(xnn_status_success,
            xnn_run_operator(qp8_convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qp8_f32_qc4w(
      input_channels, output_channels, input_channels, output_channels,
      kernel_zero_point, kernel_scale.data(), kernel.data(), bias.data(),
      output_min, output_max,
      /*flags=*/0, nullptr, nullptr, &qp8_fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_qp8_fc_op(
      qp8_fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, qp8_fc_op);
  ASSERT_EQ(xnn_status_success,
            xnn_reshape_fully_connected_nc_qp8_f32_qc4w(
                qp8_fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_qp8_f32_qc4w(
                                    qp8_fc_op, operator_qp8_data.data(),
                                    qp8_operator_output.data()));
  ASSERT_EQ(xnn_status_success,
            xnn_run_operator(qp8_fc_op, /*threadpool=*/nullptr));

  // Call operator API for `qd8`.
  xnn_operator_t qd8_convert_op = nullptr;
  xnn_operator_t qd8_fc_op = nullptr;
  status = xnn_create_convert_nc_f32_qd8(
      /*flags=*/0, &qd8_convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)>
      auto_qd8_convert_op(qd8_convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, qd8_convert_op);
  ASSERT_EQ(xnn_status_success,
            xnn_reshape_convert_nc_f32_qd8(
                qd8_convert_op, batch_size, input_channels, input_channels,
                input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_convert_nc_f32_qd8(qd8_convert_op, convert_input.data(),
                                         operator_qd8_data.data(),
                                         quantization_params.data()));
  ASSERT_EQ(xnn_status_success,
            xnn_run_operator(qd8_convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qd8_f32_qc4w(
      input_channels, output_channels, input_channels, output_channels,
      kernel_zero_point, kernel_scale.data(), kernel.data(), bias.data(),
      output_min, output_max,
      /*flags=*/0, nullptr, nullptr, &qd8_fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_qd8_fc_op(
      qd8_fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, qd8_fc_op);
  ASSERT_EQ(xnn_status_success,
            xnn_reshape_fully_connected_nc_qd8_f32_qc4w(
                qd8_fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qd8_f32_qc4w(
                qd8_fc_op, operator_qd8_data.data(), qd8_operator_output.data(),
                quantization_params.data()));
  ASSERT_EQ(xnn_status_success,
            xnn_run_operator(qd8_fc_op, /*threadpool=*/nullptr));

  // Compare the outputs. Note that the values will not be exactly the same
  // since the `qd8` quantization rounds to zero, whereas the `qp8` quantization
  // does not.
  float max_abs_val = 0.0f;
  for (size_t i = 0; i < batch_size * output_channels; i++) {
    max_abs_val = std::max(max_abs_val, std::abs(qd8_operator_output[i]));
  }
  for (size_t i = 0; i < batch_size * output_channels; i++) {
    ASSERT_NEAR(qp8_operator_output[i], qd8_operator_output[i],
                max_abs_val * 1e-2);
  }
}

TEST_F(FullyConnectedTestQP8F32QC4W, matches_operator_api) {
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  if (xnn_init_qp8_f32_qc4w_gemm_config() == nullptr) {
    GTEST_SKIP();
  }

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/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;
  std::vector<float> convert_input(batch_size * input_channels +
                                   XNN_EXTRA_BYTES / sizeof(float));
  std::vector<int8_t> operator_dq_data(
      xnn_x8_packq_f32qp8_gemm_packed_size(batch_size, input_channels));
  std::vector<float> subgraph_output(batch_size * output_channels);
  std::vector<float> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));

  // Adjust number of kernel elements for QC4W. input_channels should be padded
  // to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);

  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(),
                [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::generate(convert_input.begin(), convert_input.end(),
                [&]() { return f32dist(rng); });

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

  const uint8_t kernel_zero_point = 8;

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f32_qp8(
      /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(
      convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f32_qp8(
                                    convert_op, batch_size, input_channels,
                                    input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_convert_nc_f32_qp8(convert_op, convert_input.data(),
                                         operator_dq_data.data()));
  ASSERT_EQ(xnn_status_success,
            xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qp8_f32_qc4w(
      input_channels, output_channels, input_channels, output_channels,
      kernel_zero_point, kernel_scale.data(), kernel.data(), bias.data(),
      output_min, output_max,
      /*flags=*/0, nullptr, nullptr, &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(
      fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qp8_f32_qc4w(
                                    fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qp8_f32_qc4w(
                fc_op, operator_dq_data.data(), operator_output.data()));
  ASSERT_EQ(xnn_status_success,
            xnn_run_operator(fc_op, /*threadpool=*/nullptr));

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

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(xnn_status_success,
            xnn_define_dynamically_quantized_tensor_value(
                subgraph, xnn_datatype_qdint8, input_dims.size(),
                /*num_nonbatch_dims=*/1, input_dims.data(),
                XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(xnn_status_success,
            xnn_define_channelwise_quantized_tensor_value_v2(
                subgraph, xnn_datatype_qcint4, kernel_zero_point,
                kernel_scale.data(), kernel_dims.size(),
                /*channel_dim=*/0, kernel_dims.data(), kernel.data(),
                /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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=*/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, dq_quantized_id,
                               /*flags=*/XNN_FLAG_MAYBE_PACK_FOR_GEMM));
  ASSERT_EQ(xnn_status_success,
            xnn_define_fully_connected(subgraph, output_min, output_max,
                                       dq_quantized_id, kernel_id, bias_id,
                                       output_id, /*flags=*/0));

  // Make sure the quantized inputs were coerced to `qpint8`.
  ASSERT_EQ(subgraph->num_nodes, 2);
  const struct xnn_node* fc_node = &subgraph->nodes[1];
  ASSERT_EQ(fc_node->type, xnn_node_type_fully_connected);
  ASSERT_EQ(fc_node->compute_type, xnn_compute_type_qp8_to_fp32);

  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, convert_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));

  EXPECT_THAT(subgraph_output, operator_output);
}

TEST_F(FullyConnectedTestQP8F32QC4W, matches_operator_api_with_reshape) {
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  if (xnn_init_qp8_f32_qc4w_gemm_config() == nullptr) {
    GTEST_SKIP();
  }

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/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;
  std::vector<float> convert_input;
  convert_input.reserve(5 * batch_size * input_channels +
                        XNN_EXTRA_BYTES / sizeof(float));
  convert_input.resize(batch_size * input_channels +
                       XNN_EXTRA_BYTES / sizeof(float));
  std::vector<float> subgraph_output;
  subgraph_output.reserve(5 * batch_size * output_channels);
  subgraph_output.resize(batch_size * output_channels);
  std::vector<int8_t> operator_dq_data(
      xnn_x8_packq_f32qp8_gemm_packed_size(batch_size, input_channels) +
      XNN_EXTRA_BYTES);
  std::vector<float> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));

  // Adjust number of kernel elements for QC4W. input_channels should be padded
  // to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);

  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(),
                [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(convert_input.begin(), convert_input.end(),
                [&]() { return f32dist(rng); });

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

  const uint8_t kernel_zero_point = 8;

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f32_qp8(
      /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(
      convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f32_qp8(
                                    convert_op, batch_size, input_channels,
                                    input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_convert_nc_f32_qp8(convert_op, convert_input.data(),
                                         operator_dq_data.data()));
  ASSERT_EQ(xnn_status_success,
            xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qp8_f32_qc4w(
      input_channels, output_channels, input_channels, output_channels,
      kernel_zero_point, kernel_scale.data(), kernel.data(), nullptr,
      output_min, output_max,
      /*flags=*/0, nullptr, nullptr, &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(
      fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qp8_f32_qc4w(
                                    fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qp8_f32_qc4w(
                fc_op, operator_dq_data.data(), operator_output.data()));
  ASSERT_EQ(xnn_status_success,
            xnn_run_operator(fc_op, /*threadpool=*/nullptr));

  // Call subgraph API.
  //
  // dim[0] size increments:
  //
  //  0..........2......3.......4
  //  ^          ^      ^       ^
  //  `..Input1  |      |       |
  //     Input2..'      |       |
  //     Subgraph.......'       |
  //     Input3.................'

  std::vector<size_t> subgraph_input_dims(input_dims);
  std::vector<size_t> subgraph_output_dims(output_dims);
  subgraph_input_dims[0] += 3;
  subgraph_output_dims[0] += 3;

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

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(xnn_status_success,
            xnn_define_dynamically_quantized_tensor_value(
                subgraph, xnn_datatype_qdint8, subgraph_input_dims.size(),
                /*num_nonbatch_dims=*/1, subgraph_input_dims.data(),
                XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(xnn_status_success,
            xnn_define_channelwise_quantized_tensor_value_v2(
                subgraph, xnn_datatype_qcint4, kernel_zero_point,
                kernel_scale.data(), kernel_dims.size(),
                /*channel_dim=*/0, kernel_dims.data(), kernel.data(),
                /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(xnn_status_success,
            xnn_define_tensor_value(
                subgraph, xnn_datatype_fp32, subgraph_output_dims.size(),
                subgraph_output_dims.data(), nullptr,
                /*external_id=*/3, /*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, dq_quantized_id,
                               /*flags=*/XNN_FLAG_MAYBE_PACK_FOR_GEMM));
  ASSERT_EQ(xnn_status_success,
            xnn_define_fully_connected(
                subgraph, output_min, output_max, dq_quantized_id, kernel_id,
                XNN_INVALID_NODE_ID, output_id, /*flags=*/0));

  // Make sure the quantized inputs were coerced to `qpint8`.
  ASSERT_EQ(subgraph->num_nodes, 2);
  const struct xnn_node* fc_node = &subgraph->nodes[1];
  ASSERT_EQ(fc_node->type, xnn_node_type_fully_connected);
  ASSERT_EQ(fc_node->compute_type, xnn_compute_type_qp8_to_fp32);

  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);

  struct xnn_node* node = &subgraph->nodes[0];
  ASSERT_EQ(node->type, xnn_node_type_convert);

  // 1st inference: lets start smaller than we planned memory for
  std::array<xnn_external_value, 2> external = {
      xnn_external_value{input_id, convert_input.data()},
      xnn_external_value{output_id, subgraph_output.data()}};
  ASSERT_EQ(xnn_status_success,
            xnn_reshape_external_value(runtime, input_id, input_dims.size(),
                                       input_dims.data()));
  ASSERT_EQ(xnn_status_success, xnn_reshape_runtime(runtime));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_runtime_v2(runtime, external.size(), external.data()));
  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));
  EXPECT_THAT(subgraph_output, operator_output);

  // 2nd inference: The dq-params should be properly allocated to handle a
  // resize without memory retrigger
  input_dims[0] += 2;
  size_t batch_size2 = std::accumulate(input_dims.begin(), input_dims.end() - 1,
                                       1, std::multiplies<size_t>());
  convert_input.resize(batch_size2 * input_channels +
                       XNN_EXTRA_BYTES / sizeof(float));
  subgraph_output.resize(batch_size2 * output_channels);
  ASSERT_EQ(xnn_status_success,
            xnn_reshape_external_value(runtime, input_id, input_dims.size(),
                                       input_dims.data()));
  ASSERT_EQ(xnn_status_success, xnn_reshape_runtime(runtime));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_runtime_v2(runtime, external.size(), external.data()));
  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

  // 3rd inference: The dq-params should be properly allocated even with memory
  // retrigger
  input_dims[0] += 2;  // +4 total
  size_t batch_size3 = std::accumulate(input_dims.begin(), input_dims.end() - 1,
                                       1, std::multiplies<size_t>());
  convert_input.resize(batch_size3 * input_channels +
                       XNN_EXTRA_BYTES / sizeof(float));
  subgraph_output.resize(batch_size3 * output_channels);
  ASSERT_EQ(xnn_status_success,
            xnn_reshape_external_value(runtime, input_id, input_dims.size(),
                                       input_dims.data()));
  ASSERT_EQ(xnn_status_success, xnn_reshape_runtime(runtime));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_runtime_v2(runtime, external.size(), external.data()));
  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));
}

// TODO(b/355416339): Re-enable once we can handle strides again
TEST_F(FullyConnectedTestQP8F32QC4W, DISABLED_matches_operator_api_transposed_weights) {
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  if (xnn_init_qp8_f32_qc4w_gemm_config() == nullptr) {
    GTEST_SKIP();
  }

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/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;
  std::vector<float> convert_input(batch_size * input_channels +
                                   XNN_EXTRA_BYTES / sizeof(float));
  std::vector<int8_t> operator_dq_data(
      xnn_x8_packq_f32qp8_gemm_packed_size(batch_size, input_channels) +
      XNN_EXTRA_BYTES);
  std::vector<float> subgraph_output(batch_size * output_channels);
  std::vector<float> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));

  // Adjust number of kernel elements for QC4W. input_channels should be padded
  // to byte boundary, hence even.
  const size_t rounded_output_channels = round_up_po2(output_channels, 2);
  kernel = std::vector<uint8_t>(input_channels * rounded_output_channels);

  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(),
                [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::generate(convert_input.begin(), convert_input.end(),
                [&]() { return f32dist(rng); });

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

  const uint8_t kernel_zero_point = 8;

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f32_qp8(
      /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(
      convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f32_qp8(
                                    convert_op, batch_size, input_channels,
                                    input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_convert_nc_f32_qp8(convert_op, convert_input.data(),
                                         operator_dq_data.data()));
  ASSERT_EQ(xnn_status_success,
            xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qp8_f32_qc4w(
      input_channels, output_channels, input_channels, output_channels,
      kernel_zero_point, kernel_scale.data(), kernel.data(), bias.data(),
      output_min, output_max, XNN_FLAG_TRANSPOSE_WEIGHTS, nullptr, nullptr,
      &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(
      fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qp8_f32_qc4w(
                                    fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qp8_f32_qc4w(
                fc_op, operator_dq_data.data(), operator_output.data()));
  ASSERT_EQ(xnn_status_success,
            xnn_run_operator(fc_op, /*threadpool=*/nullptr));

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

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(xnn_status_success,
            xnn_define_dynamically_quantized_tensor_value(
                subgraph, xnn_datatype_qdint8, input_dims.size(),
                /*num_nonbatch_dims=*/1, input_dims.data(),
                XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(xnn_status_success,
            xnn_define_channelwise_quantized_tensor_value_v2(
                subgraph, xnn_datatype_qcint4, kernel_zero_point,
                kernel_scale.data(), kernel_dims_tranposed.size(),
                /*channel_dim=*/1, kernel_dims_tranposed.data(), kernel.data(),
                /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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=*/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, dq_quantized_id,
                               /*flags=*/XNN_FLAG_MAYBE_PACK_FOR_GEMM));
  ASSERT_EQ(xnn_status_success,
            xnn_define_fully_connected(subgraph, output_min, output_max,
                                       dq_quantized_id, kernel_id, bias_id,
                                       output_id, XNN_FLAG_TRANSPOSE_WEIGHTS));

  // Make sure the quantized inputs were coerced to `qpint8`.
  ASSERT_EQ(subgraph->num_nodes, 2);
  const struct xnn_node* fc_node = &subgraph->nodes[1];
  ASSERT_EQ(fc_node->type, xnn_node_type_fully_connected);
  ASSERT_EQ(fc_node->compute_type, xnn_compute_type_qp8_to_fp32);

  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, convert_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));

  EXPECT_THAT(subgraph_output, operator_output);
}

template <class T> class QuantizedFullyConnectedTestBase : public FullyConnectedTestBase<T, 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];
      }
    }
  }
};

class FullyConnectedTestF32QC4W : public FullyConnectedTestBase<float, uint8_t, float> {
};

class FullyConnectedTestF32QC8W : public FullyConnectedTestBase<float, int8_t, float> {
};

using FullyConnectedTestQC8 = QuantizedFullyConnectedTestBase<int8_t>;
using FullyConnectedTestQS8 = QuantizedFullyConnectedTestBase<int8_t>;
using FullyConnectedTestQU8 = QuantizedFullyConnectedTestBase<uint8_t>;
using FullyConnectedTestF16 = FullyConnectedTestBase<xnn_float16, float, float>;
using FullyConnectedTestF32 = FullyConnectedTestBase<float>;
using DynamicFullyConnectedTestF32 = FullyConnectedTestBase<float>;

TEST_F(FullyConnectedTestQC8, 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_VALUE_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_VALUE_ID);

  std::vector<float> scale(output_channels, 1.0f);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_channelwise_quantized_tensor_value(
        subgraph, xnn_datatype_qcint8, scale.data(), kernel_dims.size(), 0, kernel_dims.data(), kernel.data(),
        /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint32, scale.data(), bias_dims.size(), 0, bias_dims.data(), bias.data(),
                          /*external_id=*/2, /*flags=*/0, &bias_id));

  uint32_t output_id = XNN_INVALID_VALUE_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_VALUE_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_qc8);
  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, 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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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_VALUE_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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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_VALUE_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_P(FullyConnectedTestF16, define)
{
  bool use_bias = GetParam();
  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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, input_dims.size(), input_dims.data(), nullptr,
                          /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_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_VALUE_ID;
  if (use_bias) {
    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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, /*flags=*/0, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_VALUE_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_fp16);
  ASSERT_EQ(node->activation.output_min, output_min);
  ASSERT_EQ(node->activation.output_max, output_max);
  ASSERT_EQ(node->num_inputs, use_bias ? 3 : 2);
  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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_VALUE_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(FullyConnectedTestF32QC4W, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
  std::vector<float> requantization_scales(output_channels, 1.0f);

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value_v2(
                          subgraph, xnn_datatype_qcint4, /*zero_point=*/8, requantization_scales.data(),
                          kernel_dims.size(), /*channel_dim=*/0, kernel_dims.data(), kernel.data(), /*external_id=*/1,
                          /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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(FullyConnectedTestF32QC4W, define_without_bias)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
  std::vector<float> requantization_scales(output_channels, 1.0f);

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value_v2(
                          subgraph, xnn_datatype_qcint4, /*zero_point=*/8, requantization_scales.data(),
                          kernel_dims.size(), /*channel_dim=*/0, kernel_dims.data(), kernel.data(), /*external_id=*/1,
                          /*flags=*/0, &kernel_id));

  uint32_t output_id = XNN_INVALID_VALUE_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_VALUE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_fully_connected(
      subgraph, output_min, output_max, input_id, kernel_id, XNN_INVALID_VALUE_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, 2);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->inputs[1], kernel_id);
  ASSERT_EQ(node->inputs[2], XNN_INVALID_VALUE_ID);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(FullyConnectedTestF32QC8W, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
  std::vector<float> requantization_scales(output_channels, 1.0f);

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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(FullyConnectedTestF32QC8W, define_without_bias)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
  std::vector<float> requantization_scales(output_channels, 1.0f);

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t output_id = XNN_INVALID_VALUE_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_VALUE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_fully_connected(subgraph, output_min, output_max, input_id, kernel_id, XNN_INVALID_VALUE_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, 2);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->inputs[1], kernel_id);
  ASSERT_EQ(node->inputs[2], XNN_INVALID_VALUE_ID);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(DynamicFullyConnectedTestF32, 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_VALUE_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_VALUE_ID);

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

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

  uint32_t output_id = XNN_INVALID_VALUE_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_VALUE_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(FullyConnectedTestQC8, 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);
  std::vector<float> requantization_scales(output_channels, 1.0f);
  std::generate(requantization_scales.begin(), requantization_scales.end(), [&]() { return f32dist(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_qc8w(
    input_channels, output_channels, input_channels, output_channels, input_zero_point, input_scale,
    requantization_scales.data(), kernel.data(),
    bias.data(), output_zero_point, output_scale, quantized_output_min, quantized_output_max,
    /*flags=*/0, nullptr, 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_reshape_fully_connected_nc_qs8_qc8w(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_qs8_qc8w(op, input.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8,
                          requantization_scales.data(), kernel_dims.size(), 0, kernel_dims.data(),
                          kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint32, requantization_scales.data(), bias_dims.size(), 0, bias_dims.data(),
                          bias.data(), /*external_id=*/2, /*flags=*/0, &bias_id));

  uint32_t output_id = XNN_INVALID_VALUE_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_VALUE_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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

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, 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_reshape_fully_connected_nc_qs8(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_qs8(op, input.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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_VALUE_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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

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, 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_reshape_fully_connected_nc_qu8(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_qu8(op, input.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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_VALUE_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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

TEST_P(FullyConnectedTestF16, matches_operator_api)
{
  bool use_bias = GetParam();
  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); });
  if (use_bias) {
    std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  }
  std::fill(operator_output.begin(), operator_output.end(), std::nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), std::nanf(""));

  // Call operator API.
  const xnn_status status = xnn_create_fully_connected_nc_f16(
    input_channels, output_channels, input_channels, output_channels, kernel.data(), use_bias ? bias.data() : nullptr, output_min,
    output_max, XNN_FLAG_FP32_STATIC_WEIGHTS, nullptr, 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_reshape_fully_connected_nc_f16(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_f16(op, input.data(), operator_output.data()));

  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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, input_dims.size(), input_dims.data(), nullptr,
                          /*external_id=*/0, XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_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_VALUE_ID;
  if (use_bias) {
    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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_VALUE_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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

INSTANTIATE_TEST_SUITE_P(
    UseBias,
    FullyConnectedTestF16,
    ::testing::Bool());

TEST_P(FullyConnectedTestF16, matches_operator_api_f16_weights)
{
  bool use_bias = GetParam();
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_operator_t op = nullptr;

  std::generate(input.begin(), input.end(), [&]() { return xnn_float16_from_float(f32dist(rng)); });
  std::generate(kernel_fp16.begin(), kernel_fp16.end(), [&]() { return xnn_float16_from_float(f32dist(rng)); });
  if (use_bias) {
    std::generate(bias_fp16.begin(), bias_fp16.end(), [&]() { return xnn_float16_from_float(f32dist(rng)); });
  }
  std::fill(operator_output.begin(), operator_output.end(), UINT16_C(0x7E00) /* NaN */);
  std::fill(subgraph_output.begin(), subgraph_output.end(), UINT16_C(0x7E00) /* NaN */);

  // Call operator API.
  const xnn_status status = xnn_create_fully_connected_nc_f16(
    input_channels, output_channels, input_channels, output_channels, kernel_fp16.data(), use_bias ? bias_fp16.data() : nullptr, output_min,
    output_max, /*flags=*/0, nullptr, 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_reshape_fully_connected_nc_f16(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_f16(op, input.data(), operator_output.data()));

  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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, input_dims.size(), input_dims.data(), nullptr,
                          /*external_id=*/0, XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, kernel_dims.size(), kernel_dims.data(), kernel_fp16.data(),
                          /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_ID;
  if (use_bias) {
    ASSERT_EQ(
      xnn_status_success, xnn_define_tensor_value(
                            subgraph, xnn_datatype_fp16, bias_dims.size(), bias_dims.data(), bias_fp16.data(),
                            /*external_id=*/2, /*flags=*/0, &bias_id));
  }

  uint32_t output_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_VALUE_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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

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, 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_reshape_fully_connected_nc_f32(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_f32(op, input.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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_VALUE_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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

TEST_F(FullyConnectedTestF32QC4W, 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); });
  // Adjust number of kernel elements for QC4W. input_channels should be padded to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);
  std::generate(kernel.begin(), kernel.end(), [&]() { return i8dist(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(""));
  std::vector<float> requantization_scales(output_channels);
  std::generate(requantization_scales.begin(), requantization_scales.end(), [&]() { return scale_dist(rng); });
  uint8_t kernel_zero_point = 8;

  // Call operator API.
  const xnn_status status = xnn_create_fully_connected_nc_f32_qc4w(
    input_channels, output_channels, input_channels, output_channels,
    kernel_zero_point, requantization_scales.data(), kernel.data(), bias.data(),
    output_min, output_max,
    /*flags=*/0, nullptr, 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_reshape_fully_connected_nc_f32_qc4w(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_f32_qc4w(op, input.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value_v2(
                          subgraph, xnn_datatype_qcint4, kernel_zero_point, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(),
                          /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

TEST_F(FullyConnectedTestF32QC4W, matches_operator_api_without_bias)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_operator_t op = nullptr;

  std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
  // Adjust number of kernel elements for QC4W. input_channels should be padded to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);
  std::generate(kernel.begin(), kernel.end(), [&]() { return i8dist(rng); });
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));
  std::vector<float> requantization_scales(output_channels);
  std::generate(requantization_scales.begin(), requantization_scales.end(), [&]() { return scale_dist(rng); });
  uint8_t kernel_zero_point = 8;

  // Call operator API.
  const xnn_status status = xnn_create_fully_connected_nc_f32_qc4w(
    input_channels, output_channels, input_channels, output_channels,
    kernel_zero_point, requantization_scales.data(), kernel.data(), nullptr,
    output_min, output_max,
    /*flags=*/0, nullptr, 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_reshape_fully_connected_nc_f32_qc4w(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_f32_qc4w(op, input.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value_v2(
                          subgraph, xnn_datatype_qcint4, kernel_zero_point, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(),
                          /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t output_id = XNN_INVALID_VALUE_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_VALUE_ID);
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_fully_connected(subgraph, output_min, output_max, input_id, kernel_id, XNN_INVALID_VALUE_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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

TEST_F(FullyConnectedTestF32QC8W, 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 i8dist(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(""));
  std::vector<float> requantization_scales(output_channels);
  std::generate(requantization_scales.begin(), requantization_scales.end(), [&]() { return scale_dist(rng); });

  // Call operator API.
  const xnn_status status = xnn_create_fully_connected_nc_f32_qc8w(
    input_channels, output_channels, input_channels, output_channels, requantization_scales.data(), kernel.data(), bias.data(), output_min,
    output_max,
    /*flags=*/0, nullptr, 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_reshape_fully_connected_nc_f32_qc8w(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_f32_qc8w(op, input.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(),
                          /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

TEST_F(FullyConnectedTestF32QC8W, matches_operator_api_without_bias)
{
  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 i8dist(rng); });
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));
  std::vector<float> requantization_scales(output_channels);
  std::generate(requantization_scales.begin(), requantization_scales.end(), [&]() { return scale_dist(rng); });

  // Call operator API.
  const xnn_status status = xnn_create_fully_connected_nc_f32_qc8w(
    input_channels, output_channels, input_channels, output_channels, requantization_scales.data(), kernel.data(),
    /*bias=*/nullptr, output_min, output_max,
    /*flags=*/0, nullptr, 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_reshape_fully_connected_nc_f32_qc8w(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_f32_qc8w(op, input.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(),
                          /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t output_id = XNN_INVALID_VALUE_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_VALUE_ID);
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_fully_connected(
      subgraph, output_min, output_max, input_id, kernel_id, XNN_INVALID_VALUE_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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

INSTANTIATE_TEST_SUITE_P(
    UseBias,
    DynamicFullyConnectedTestF32,
    ::testing::Bool());

TEST_P(DynamicFullyConnectedTestF32, matches_operator_api_dynamic_kernel)
{
  bool use_bias = GetParam();
  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); });
  if (use_bias) {
    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_dynamic_fully_connected_nc_f32(output_min, output_max, /*flags=*/0, &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);

  size_t workspace_size = 0;
  size_t workspace_alignment = 0;
  ASSERT_EQ(
    xnn_status_success, xnn_reshape_dynamic_fully_connected_nc_f32(
                          op, batch_size, input_channels, output_channels, input_channels, output_channels,
                          &workspace_size, &workspace_alignment, /*threadpool=*/nullptr));
  ASSERT_NE(workspace_size, 0);
  ASSERT_LE(workspace_alignment, XNN_ALLOCATION_ALIGNMENT);

  std::vector<char, AlignedAllocator<char, XNN_ALLOCATION_ALIGNMENT>> workspace(workspace_size);
  ASSERT_EQ(
    xnn_status_success, xnn_setup_dynamic_fully_connected_nc_f32(
                          op, workspace.data(), input.data(), kernel.data(), use_bias ? bias.data() : nullptr, operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, kernel_dims.size(), kernel_dims.data(), nullptr,
                          /*external_id=*/1, XNN_VALUE_FLAG_EXTERNAL_INPUT, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_ID;
  if (use_bias) {
    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_VALUE_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_VALUE_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, 3> external = {
    xnn_external_value{input_id, input.data()},
    xnn_external_value{kernel_id, kernel.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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

TEST_F(DynamicFullyConnectedTestF32, matches_operator_api_dynamic_bias)
{
  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_dynamic_fully_connected_nc_f32(output_min, output_max, /*flags=*/0, &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);

  size_t workspace_size = 0;
  size_t workspace_alignment = 0;
  ASSERT_EQ(
    xnn_status_success, xnn_reshape_dynamic_fully_connected_nc_f32(
                          op, batch_size, input_channels, output_channels, input_channels, output_channels,
                          &workspace_size, &workspace_alignment, /*threadpool=*/nullptr));
  ASSERT_NE(workspace_size, 0);
  ASSERT_LE(workspace_alignment, XNN_ALLOCATION_ALIGNMENT);

  std::vector<char, AlignedAllocator<char, XNN_ALLOCATION_ALIGNMENT>> workspace(workspace_size);
  ASSERT_EQ(
    xnn_status_success, xnn_setup_dynamic_fully_connected_nc_f32(
                          op, workspace.data(), input.data(), kernel.data(), bias.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, bias_dims.size(), bias_dims.data(), nullptr,
                          /*external_id=*/2, XNN_VALUE_FLAG_EXTERNAL_INPUT, &bias_id));

  uint32_t output_id = XNN_INVALID_VALUE_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_VALUE_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, 3> external = {
    xnn_external_value{input_id, input.data()},
    xnn_external_value{bias_id, bias.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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

TEST_F(DynamicFullyConnectedTestF32, matches_operator_api_dynamic_kernel_and_bias)
{
  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_dynamic_fully_connected_nc_f32(output_min, output_max, /*flags=*/0, &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);

  size_t workspace_size = 0;
  size_t workspace_alignment = 0;
  ASSERT_EQ(
    xnn_status_success, xnn_reshape_dynamic_fully_connected_nc_f32(
                          op, batch_size, input_channels, output_channels, input_channels, output_channels,
                          &workspace_size, &workspace_alignment, /*threadpool=*/nullptr));
  ASSERT_NE(workspace_size, 0);
  ASSERT_LE(workspace_alignment, XNN_ALLOCATION_ALIGNMENT);

  std::vector<char, AlignedAllocator<char, XNN_ALLOCATION_ALIGNMENT>> workspace(workspace_size);
  ASSERT_EQ(
    xnn_status_success, xnn_setup_dynamic_fully_connected_nc_f32(
                          op, workspace.data(), input.data(), kernel.data(), bias.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, kernel_dims.size(), kernel_dims.data(), nullptr,
                          /*external_id=*/1, XNN_VALUE_FLAG_EXTERNAL_INPUT, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, bias_dims.size(), bias_dims.data(), nullptr,
                          /*external_id=*/2, XNN_VALUE_FLAG_EXTERNAL_INPUT, &bias_id));

  uint32_t output_id = XNN_INVALID_VALUE_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_VALUE_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, 4> external = {
    xnn_external_value{input_id, input.data()},
    xnn_external_value{kernel_id, kernel.data()},
    xnn_external_value{bias_id, bias.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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

TEST_F(FullyConnectedTestF32QC8W, matches_operator_api_transposed_weights)
{
  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 i8dist(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(""));
  std::vector<float> requantization_scales(output_channels);
  std::generate(requantization_scales.begin(), requantization_scales.end(), [&]() { return scale_dist(rng); });

  // Call operator API.
  const xnn_status status = xnn_create_fully_connected_nc_f32_qc8w(
    input_channels, output_channels, input_channels, output_channels, requantization_scales.data(), kernel.data(),
    bias.data(), output_min, output_max, XNN_FLAG_TRANSPOSE_WEIGHTS, nullptr, 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_reshape_fully_connected_nc_f32_qc8w(op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_fully_connected_nc_f32_qc8w(op, input.data(), operator_output.data()));

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims_tranposed.size(),
                          /*channel_dim=*/1, kernel_dims_tranposed.data(), kernel.data(),
                          /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_ID);
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_fully_connected(
      subgraph, output_min, output_max, input_id, kernel_id, bias_id, output_id, XNN_FLAG_TRANSPOSE_WEIGHTS));

  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.
  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

TEST_F(FullyConnectedTestF32QC8W, non_static_kernel_is_invalid_parameter)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
  std::vector<float> requantization_scales(output_channels, 1.0f);

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), nullptr, /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_ID);

  ASSERT_EQ(
    xnn_status_invalid_parameter,
    xnn_define_fully_connected(subgraph, output_min, output_max, input_id, kernel_id, bias_id, output_id, /*flags=*/0));
}

TEST_F(FullyConnectedTestF32QC8W, non_static_bias_is_invalid_parameter)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
  std::vector<float> requantization_scales(output_channels, 1.0f);

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

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

  uint32_t output_id = XNN_INVALID_VALUE_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_VALUE_ID);

  ASSERT_EQ(
    xnn_status_invalid_parameter,
    xnn_define_fully_connected(subgraph, output_min, output_max, input_id, kernel_id, bias_id, output_id, /*flags=*/0));
}

TEST_F(FullyConnectedTestF32QC8W, invalid_channel_dimension)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
  const size_t channel_dim = 1;
  std::vector<float> requantization_scales(kernel_dims[channel_dim], 1.0f);

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims.size(), channel_dim,
                          kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_ID);

  ASSERT_EQ(
    xnn_status_invalid_parameter,
    xnn_define_fully_connected(subgraph, output_min, output_max, input_id, kernel_id, bias_id, output_id, /*flags=*/0));
}

TEST_F(FullyConnectedTestF32QC8W, transposed_weights_invalid_channel_dimension)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
  const size_t channel_dim = 0;
  std::vector<float> requantization_scales(kernel_dims_tranposed[channel_dim], 1.0f);

  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_VALUE_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_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_channelwise_quantized_tensor_value(
      subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims_tranposed.size(), channel_dim,
      kernel_dims_tranposed.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_ID);

  ASSERT_EQ(
    xnn_status_invalid_parameter,
    xnn_define_fully_connected(
      subgraph, output_min, output_max, input_id, kernel_id, bias_id, output_id, XNN_FLAG_TRANSPOSE_WEIGHTS));
}

class FullyConnectedTestQD8F16QC4W : public FullyConnectedTestBase<int8_t, int8_t, float, xnn_float16, true> {
};

TEST_F(FullyConnectedTestQD8F16QC4W, define)
{
  std::vector<float> requantization_scales(output_channels, 1.0f);

  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);

  const uint8_t kernel_zero_point = 8;

  uint32_t input_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value_v2(
                          subgraph, xnn_datatype_qcint4, kernel_zero_point, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, /*flags=*/0, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_VALUE_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_qd8_to_fp16);
  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(FullyConnectedTestQD8F16QC4W, internally_allocated_dynamic_quantization_parameters)
{
  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=*/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;
  std::vector<xnn_float16> convert_input(batch_size * input_channels + XNN_EXTRA_BYTES / sizeof(xnn_float16));
  std::vector<int8_t> operator_dq_data(batch_size * input_channels + XNN_EXTRA_BYTES);
  std::vector<xnn_float16> subgraph_output(batch_size * output_channels);
  std::vector<xnn_float16> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), UINT16_C(0xDEAD));
  std::fill(subgraph_output.begin(), subgraph_output.end(), UINT16_C(0xDEAD));
  std::vector<xnn_dynamic_quantization_params> quantization_params(batch_size + XNN_EXTRA_QUANTIZATION_PARAMS);

  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::generate(quantization_params.begin(), quantization_params.end(), [&]() { return xnn_dynamic_quantization_params{w8dist(rng), f32dist(rng)}; });
  std::generate(convert_input.begin(), convert_input.end(), [&]() { return f32dist(rng); });

  const float output_min = -std::numeric_limits<float>::infinity();
  const float output_max = std::numeric_limits<float>::infinity();
  const uint8_t kernel_zero_point = 8;

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f16_qd8(
    /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f16_qd8(convert_op, batch_size, input_channels, input_channels, input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_convert_nc_f16_qd8(convert_op, convert_input.data(),
                                                             operator_dq_data.data(), quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qd8_f16_qc4w(
    input_channels, output_channels, input_channels, output_channels, kernel_zero_point, kernel_scale.data(),
    kernel.data(), bias.data(), output_min, output_max,
    /*flags=*/0, nullptr, nullptr, &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qd8_f16_qc4w(fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qd8_f16_qc4w(fc_op, operator_dq_data.data(), operator_output.data(),
                                                      quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(fc_op, /*threadpool=*/nullptr));

  // Call subgraph API.
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, input_dims.size(), input_dims.data(), nullptr, /*external_id=*/0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value_v2(
                          subgraph, xnn_datatype_qcint4, kernel_zero_point, kernel_scale.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_fp16, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, /*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, dq_quantized_id, /*flags=*/0));
  ASSERT_EQ(xnn_status_success, xnn_define_fully_connected(subgraph, output_min, output_max, dq_quantized_id,
                                                           kernel_id, bias_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, convert_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));

  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

class FullyConnectedTestQD8F16QB4W : public FullyConnectedTestBase<int8_t, uint8_t, float, xnn_float16, true> {
};

TEST_F(FullyConnectedTestQD8F16QB4W, define)
{
  size_t block_size = 32;
  input_channels = round_up_po2(input_channels, block_size);

  input_dims[input_dims.size() - 1] = input_channels;
  kernel_dims[kernel_dims.size() - 1] = input_channels;

  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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  // Adjust number of kernel elements for QB4W. input_channels should be padded to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);
  const uint8_t kernel_zero_point = 8;
  std::vector<xnn_bfloat16> kernel_scale(output_channels * block_size);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_blockwise_quantized_tensor_value(
        subgraph, xnn_datatype_qbint4, kernel_zero_point, reinterpret_cast<const uint16_t*>(kernel_scale.data()), kernel_dims.size(),
        /*channel_dim=*/0, block_size, kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, /*flags=*/0, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_VALUE_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_qd8_to_fp16);
  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(FullyConnectedTestQD8F16QB4W, internally_allocated_dynamic_quantization_parameters)
{
  size_t block_size = 32;
  input_channels = round_up_po2(input_channels, block_size);

  input_dims[input_dims.size() - 1] = input_channels;
  kernel_dims[kernel_dims.size() - 1] = input_channels;

  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=*/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;
  std::vector<float> convert_input(batch_size * input_channels + XNN_EXTRA_BYTES / sizeof(float));
  std::vector<int8_t> operator_dq_data(batch_size * input_channels + XNN_EXTRA_BYTES);
  std::vector<float> subgraph_output(batch_size * output_channels);
  std::vector<float> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));
  std::vector<xnn_dynamic_quantization_params> quantization_params(batch_size + XNN_EXTRA_QUANTIZATION_PARAMS);

  std::vector<xnn_bfloat16> kernel_scale(output_channels * block_size);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::generate(convert_input.begin(), convert_input.end(), [&]() { return f32dist(rng); });
  std::generate(quantization_params.begin(), quantization_params.end(), [&]() { return xnn_dynamic_quantization_params{w8dist(rng), f32dist(rng)}; });

  // Adjust number of kernel elements for QC4W. input_channels should be padded to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);

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

  const uint8_t kernel_zero_point = 8;

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f16_qd8(
    /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f16_qd8(convert_op, batch_size, input_channels, input_channels, input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_convert_nc_f16_qd8(convert_op, convert_input.data(),
                                                             operator_dq_data.data(), quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qd8_f16_qb4w(
    input_channels, output_channels, input_channels, output_channels, block_size, kernel_zero_point, reinterpret_cast<const uint16_t*>(kernel_scale.data()),
    kernel.data(), bias.data(), output_min, output_max,
    /*flags=*/0, nullptr, nullptr, &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qd8_f16_qb4w(fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qd8_f16_qb4w(fc_op, operator_dq_data.data(), operator_output.data(),
                                                      quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(fc_op, /*threadpool=*/nullptr));

  // Call subgraph API.
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, input_dims.size(), input_dims.data(), nullptr, /*external_id=*/0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_blockwise_quantized_tensor_value(
        subgraph, xnn_datatype_qbint4, kernel_zero_point, reinterpret_cast<const uint16_t*>(kernel_scale.data()), kernel_dims.size(),
        /*channel_dim=*/0, block_size, kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_fp16, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, /*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, dq_quantized_id, /*flags=*/0));
  ASSERT_EQ(xnn_status_success, xnn_define_fully_connected(subgraph, output_min, output_max, dq_quantized_id,
                                                           kernel_id, bias_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, convert_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));
}

class FullyConnectedTestQD8F16QC8W : public FullyConnectedTestBase<int8_t, int8_t, float, xnn_float16> {
};

TEST_F(FullyConnectedTestQD8F16QC8W, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  std::vector<float> requantization_scales(output_channels, 1.0f);

  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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, /*flags=*/0, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_VALUE_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_qd8_to_fp16);
  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(FullyConnectedTestQD8F16QC8W, internally_allocated_dynamic_quantization_parameters)
{
  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=*/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;
  std::vector<xnn_float16> convert_input(batch_size * input_channels + XNN_EXTRA_BYTES / sizeof(xnn_float16));
  std::vector<int8_t> operator_dq_data(batch_size * input_channels + XNN_EXTRA_BYTES);
  std::vector<xnn_float16> subgraph_output(batch_size * output_channels);
  std::vector<xnn_float16> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), UINT16_C(0xDEAD));
  std::fill(subgraph_output.begin(), subgraph_output.end(), UINT16_C(0xDEAD));
  std::vector<xnn_dynamic_quantization_params> quantization_params(batch_size + XNN_EXTRA_QUANTIZATION_PARAMS);

  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::generate(quantization_params.begin(), quantization_params.end(), [&]() { return xnn_dynamic_quantization_params{w8dist(rng), f32dist(rng)}; });
  std::generate(convert_input.begin(), convert_input.end(), [&]() { return f32dist(rng); });

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

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f16_qd8(
    /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f16_qd8(convert_op, batch_size, input_channels, input_channels, input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_convert_nc_f16_qd8(convert_op, convert_input.data(),
                                                             operator_dq_data.data(), quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qd8_f16_qc8w(
    input_channels, output_channels, input_channels, output_channels, kernel_scale.data(),
    kernel.data(), bias.data(), output_min, output_max,
    /*flags=*/0, nullptr, nullptr, &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qd8_f16_qc8w(fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qd8_f16_qc8w(fc_op, operator_dq_data.data(), operator_output.data(),
                                                      quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(fc_op, /*threadpool=*/nullptr));

  // Call subgraph API.
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp16, input_dims.size(), input_dims.data(), nullptr, /*external_id=*/0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, kernel_scale.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_fp16, output_dims.size(), output_dims.data(), nullptr,
                          /*external_id=*/3, /*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, dq_quantized_id, /*flags=*/0));
  ASSERT_EQ(xnn_status_success, xnn_define_fully_connected(subgraph, output_min, output_max, dq_quantized_id,
                                                           kernel_id, bias_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, convert_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));

  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

class FullyConnectedTestQD8F32QC8W : public FullyConnectedTestBase<int8_t, int8_t, float, float> {
};

TEST_F(FullyConnectedTestQD8F32QC8W, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  std::vector<float> requantization_scales(output_channels, 1.0f);

  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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, requantization_scales.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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_qd8_to_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(FullyConnectedTestQD8F32QC8W, internally_allocated_dynamic_quantization_parameters)
{
  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=*/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;
  std::vector<float> convert_input(batch_size * input_channels + XNN_EXTRA_BYTES / sizeof(float));
  std::vector<int8_t> operator_dq_data(batch_size * input_channels + XNN_EXTRA_BYTES);
  std::vector<float> subgraph_output(batch_size * output_channels);
  std::vector<float> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));
  std::vector<xnn_dynamic_quantization_params> quantization_params(batch_size + XNN_EXTRA_QUANTIZATION_PARAMS);

  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::generate(quantization_params.begin(), quantization_params.end(), [&]() { return xnn_dynamic_quantization_params{w8dist(rng), f32dist(rng)}; });
  std::generate(convert_input.begin(), convert_input.end(), [&]() { return f32dist(rng); });

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

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f32_qd8(
    /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f32_qd8(convert_op, batch_size, input_channels, input_channels, input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_convert_nc_f32_qd8(convert_op, convert_input.data(),
                                                             operator_dq_data.data(), quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qd8_f32_qc8w(
    input_channels, output_channels, input_channels, output_channels, kernel_scale.data(),
    kernel.data(), bias.data(), output_min, output_max,
    /*flags=*/0, nullptr, nullptr, &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qd8_f32_qc8w(fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qd8_f32_qc8w(fc_op, operator_dq_data.data(), operator_output.data(),
                                                      quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(fc_op, /*threadpool=*/nullptr));

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

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value(
                          subgraph, xnn_datatype_qcint8, kernel_scale.data(), kernel_dims.size(), /*channel_dim=*/0,
                          kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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=*/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, dq_quantized_id, /*flags=*/0));
  ASSERT_EQ(xnn_status_success, xnn_define_fully_connected(subgraph, output_min, output_max, dq_quantized_id,
                                                           kernel_id, bias_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, convert_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));

  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

class FullyConnectedTestQD8F32QC4W : public FullyConnectedTestBase<int8_t, uint8_t, float, float, true> {
};

TEST_F(FullyConnectedTestQD8F32QC4W, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  std::vector<float> requantization_scales(output_channels, 1.0f);

  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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  // Adjust number of kernel elements for QC4W. input_channels should be padded to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);
  const uint8_t kernel_zero_point = 8;
  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value_v2(
        subgraph, xnn_datatype_qcint4, kernel_zero_point, kernel_scale.data(), kernel_dims.size(),
        /*channel_dim=*/0, kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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_qd8_to_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(FullyConnectedTestQD8F32QC4W, internally_allocated_dynamic_quantization_parameters)
{
  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=*/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;
  std::vector<float> convert_input(batch_size * input_channels + XNN_EXTRA_BYTES / sizeof(float));
  std::vector<int8_t> operator_dq_data(batch_size * input_channels + XNN_EXTRA_BYTES);
  std::vector<float> subgraph_output(batch_size * output_channels);
  std::vector<float> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));
  std::vector<xnn_dynamic_quantization_params> quantization_params(batch_size + XNN_EXTRA_QUANTIZATION_PARAMS);

  // Adjust number of kernel elements for QC4W. input_channels should be padded to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);

  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::generate(convert_input.begin(), convert_input.end(), [&]() { return f32dist(rng); });
  std::generate(quantization_params.begin(), quantization_params.end(), [&]() { return xnn_dynamic_quantization_params{w8dist(rng), f32dist(rng)}; });

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

  const uint8_t kernel_zero_point = 8;

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f32_qd8(
    /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f32_qd8(convert_op, batch_size, input_channels, input_channels, input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_convert_nc_f32_qd8(convert_op, convert_input.data(),
                                                             operator_dq_data.data(), quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qd8_f32_qc4w(
    input_channels, output_channels, input_channels, output_channels, kernel_zero_point, kernel_scale.data(),
    kernel.data(), bias.data(), output_min, output_max,
    /*flags=*/0, nullptr, nullptr, &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qd8_f32_qc4w(fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qd8_f32_qc4w(fc_op, operator_dq_data.data(), operator_output.data(),
                                                      quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(fc_op, /*threadpool=*/nullptr));

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

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value_v2(
        subgraph, xnn_datatype_qcint4, kernel_zero_point, kernel_scale.data(), kernel_dims.size(),
        /*channel_dim=*/0, kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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=*/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, dq_quantized_id, /*flags=*/0));
  ASSERT_EQ(xnn_status_success, xnn_define_fully_connected(subgraph, output_min, output_max, dq_quantized_id,
                                                           kernel_id, bias_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, convert_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));

  EXPECT_THAT(subgraph_output, operator_output);
}

TEST_F(FullyConnectedTestQD8F32QC4W, internally_allocated_dynamic_quantization_parameters_with_reshape)
{
  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=*/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;
  std::vector<float> convert_input;
  convert_input.reserve(5 * batch_size * input_channels + XNN_EXTRA_BYTES / sizeof(float));
  convert_input.resize(batch_size * input_channels + XNN_EXTRA_BYTES / sizeof(float));
  std::vector<float> subgraph_output;
  subgraph_output.reserve(5 * batch_size * output_channels);
  subgraph_output.resize(batch_size * output_channels);
  std::vector<int8_t> operator_dq_data(batch_size * input_channels + XNN_EXTRA_BYTES);
  std::vector<float> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));
  std::vector<xnn_dynamic_quantization_params> quantization_params(batch_size + XNN_EXTRA_QUANTIZATION_PARAMS);

  // Adjust number of kernel elements for QC4W. input_channels should be padded to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);

  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(convert_input.begin(), convert_input.end(), [&]() { return f32dist(rng); });
  std::generate(quantization_params.begin(), quantization_params.end(), [&]() { return xnn_dynamic_quantization_params{w8dist(rng), f32dist(rng)}; });

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

  const uint8_t kernel_zero_point = 8;

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f32_qd8(
    /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f32_qd8(convert_op, batch_size, input_channels, input_channels, input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_convert_nc_f32_qd8(convert_op, convert_input.data(),
                                                             operator_dq_data.data(), quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qd8_f32_qc4w(
    input_channels, output_channels, input_channels, output_channels, kernel_zero_point, kernel_scale.data(),
    kernel.data(), nullptr, output_min, output_max,
    /*flags=*/0, nullptr, nullptr, &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qd8_f32_qc4w(fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qd8_f32_qc4w(fc_op, operator_dq_data.data(), operator_output.data(),
                                                      quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(fc_op, /*threadpool=*/nullptr));

  // Call subgraph API.
  //
  // dim[0] size increments:
  //
  //  0..........2......3.......4
  //  ^          ^      ^       ^
  //  `..Input1  |      |       |
  //     Input2..'      |       |
  //     Subgraph.......'       |
  //     Input3.................'

  std::vector<size_t> subgraph_input_dims(input_dims);
  std::vector<size_t> subgraph_output_dims(output_dims);
  subgraph_input_dims[0] += 3;
  subgraph_output_dims[0] += 3;

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

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, subgraph_input_dims.size(), /*num_nonbatch_dims=*/1, subgraph_input_dims.data(),
                          XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value_v2(
        subgraph, xnn_datatype_qcint4, kernel_zero_point, kernel_scale.data(), kernel_dims.size(),
        /*channel_dim=*/0, kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ( xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, subgraph_output_dims.size(), subgraph_output_dims.data(), nullptr,
                          /*external_id=*/3, /*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, dq_quantized_id, /*flags=*/0));
  ASSERT_EQ(xnn_status_success, xnn_define_fully_connected(subgraph, output_min, output_max, dq_quantized_id,
                                                           kernel_id, XNN_INVALID_NODE_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);

  struct xnn_node* node = &subgraph->nodes[0];
  ASSERT_EQ(node->type, xnn_node_type_convert);
  const size_t dynamic_param_size = runtime->values[node->outputs[0]].quantization.dynamic_params_size;
  ASSERT_GT(dynamic_param_size, 0);

  // 1st inference: lets start smaller than we planned memory for
  std::array<xnn_external_value, 2> external = {
    xnn_external_value{input_id, convert_input.data()}, xnn_external_value{output_id, subgraph_output.data()}};
  ASSERT_EQ(xnn_status_success, xnn_reshape_external_value(runtime, input_id, input_dims.size(), input_dims.data()));
  ASSERT_EQ(xnn_status_success, xnn_reshape_runtime(runtime));
  ASSERT_EQ(xnn_status_success, xnn_setup_runtime_v2(runtime, external.size(), external.data()));
  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));
  EXPECT_THAT(subgraph_output, operator_output);
  const size_t dynamic_param_size1 = runtime->values[node->outputs[0]].quantization.dynamic_params_size;
  // No change in dynamic param size after the first inference
  ASSERT_EQ(dynamic_param_size, dynamic_param_size1);

  // 2nd inference: The dq-params should be properly allocated to handle a resize without memory retrigger
  input_dims[0] += 2;
  size_t batch_size2 = std::accumulate(input_dims.begin(), input_dims.end() - 1, 1, std::multiplies<size_t>());
  convert_input.resize(batch_size2 * input_channels + XNN_EXTRA_BYTES / sizeof(float));
  subgraph_output.resize(batch_size2 * output_channels);
  ASSERT_EQ(xnn_status_success, xnn_reshape_external_value(runtime, input_id, input_dims.size(), input_dims.data()));
  ASSERT_EQ(xnn_status_success, xnn_reshape_runtime(runtime));
  ASSERT_EQ(xnn_status_success, xnn_setup_runtime_v2(runtime, external.size(), external.data()));
  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));
  const size_t dynamic_param_size2 = runtime->values[node->outputs[0]].quantization.dynamic_params_size;
  // No change after the second inference
  ASSERT_EQ(dynamic_param_size1, dynamic_param_size2);

  // 3rd inference: The dq-params should be properly allocated even with memory retrigger
  input_dims[0] += 2; // +4 total
  size_t batch_size3 = std::accumulate(input_dims.begin(), input_dims.end() - 1, 1, std::multiplies<size_t>());
  convert_input.resize(batch_size3 * input_channels + XNN_EXTRA_BYTES / sizeof(float));
  subgraph_output.resize(batch_size3 * output_channels);
  ASSERT_EQ(xnn_status_success, xnn_reshape_external_value(runtime, input_id, input_dims.size(), input_dims.data()));
  ASSERT_EQ(xnn_status_success, xnn_reshape_runtime(runtime));
  ASSERT_EQ(xnn_status_success, xnn_setup_runtime_v2(runtime, external.size(), external.data()));
  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));
  const size_t dynamic_param_size3 = runtime->values[node->outputs[0]].quantization.dynamic_params_size;
  // It should be larger after the third inference
  ASSERT_LT(dynamic_param_size2, dynamic_param_size3);
}

TEST_F(FullyConnectedTestQD8F32QC4W, internally_allocated_dynamic_quantization_parameters_transposed_weights)
{
  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=*/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;
  std::vector<float> convert_input(batch_size * input_channels + XNN_EXTRA_BYTES / sizeof(float));
  std::vector<int8_t> operator_dq_data(batch_size * input_channels + XNN_EXTRA_BYTES);
  std::vector<float> subgraph_output(batch_size * output_channels);
  std::vector<float> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));
  std::vector<xnn_dynamic_quantization_params> quantization_params(batch_size + XNN_EXTRA_QUANTIZATION_PARAMS);

  // Adjust number of kernel elements for QC4W. input_channels should be padded to byte boundary, hence even.
  const size_t rounded_output_channels = round_up_po2(output_channels, 2);
  kernel = std::vector<uint8_t>(input_channels * rounded_output_channels);

  std::vector<float> kernel_scale(output_channels);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::generate(convert_input.begin(), convert_input.end(), [&]() { return f32dist(rng); });
  std::generate(quantization_params.begin(), quantization_params.end(), [&]() { return xnn_dynamic_quantization_params{w8dist(rng), f32dist(rng)}; });

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

  const uint8_t kernel_zero_point = 8;

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f32_qd8(
    /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f32_qd8(convert_op, batch_size, input_channels, input_channels, input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_convert_nc_f32_qd8(convert_op, convert_input.data(),
                                                             operator_dq_data.data(), quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qd8_f32_qc4w(
    input_channels, output_channels, input_channels, output_channels, kernel_zero_point, kernel_scale.data(),
    kernel.data(), bias.data(), output_min, output_max,
    XNN_FLAG_TRANSPOSE_WEIGHTS, nullptr, nullptr, &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qd8_f32_qc4w(fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qd8_f32_qc4w(fc_op, operator_dq_data.data(), operator_output.data(),
                                                      quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(fc_op, /*threadpool=*/nullptr));

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

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_channelwise_quantized_tensor_value_v2(
        subgraph, xnn_datatype_qcint4, kernel_zero_point, kernel_scale.data(), kernel_dims_tranposed.size(),
        /*channel_dim=*/1, kernel_dims_tranposed.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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=*/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, dq_quantized_id, /*flags=*/0));
  ASSERT_EQ(xnn_status_success, xnn_define_fully_connected(subgraph, output_min, output_max, dq_quantized_id,
                                                           kernel_id, bias_id, output_id, XNN_FLAG_TRANSPOSE_WEIGHTS));
  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, convert_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));

  EXPECT_THAT(subgraph_output, ElementsAreArray(operator_output));
}

class FullyConnectedTestQD8F32QB4W : public FullyConnectedTestBase<int8_t, uint8_t, float, float> {
};

TEST_F(FullyConnectedTestQD8F32QB4W, define)
{
  size_t block_size = 32;
  input_channels = round_up_po2(input_channels, block_size);

  input_dims[input_dims.size() - 1] = input_channels;
  kernel_dims[kernel_dims.size() - 1] = input_channels;

  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_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          /*external_id=*/0, /*flags=*/0, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  // Adjust number of kernel elements for QB4W. input_channels should be padded to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);
  const uint8_t kernel_zero_point = 8;
  std::vector<xnn_bfloat16> kernel_scale(output_channels * block_size);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_blockwise_quantized_tensor_value(
        subgraph, xnn_datatype_qbint4, kernel_zero_point, reinterpret_cast<const uint16_t*>(kernel_scale.data()), kernel_dims.size(),
        /*channel_dim=*/0, block_size, kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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_qd8_to_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(FullyConnectedTestQD8F32QB4W, internally_allocated_dynamic_quantization_parameters)
{
  size_t block_size = 32;
  input_channels = round_up_po2(input_channels, block_size);

  input_dims[input_dims.size() - 1] = input_channels;
  kernel_dims[kernel_dims.size() - 1] = input_channels;

  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=*/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;
  std::vector<float> convert_input(batch_size * input_channels + XNN_EXTRA_BYTES / sizeof(float));
  std::vector<int8_t> operator_dq_data(batch_size * input_channels + XNN_EXTRA_BYTES);
  std::vector<float> subgraph_output(batch_size * output_channels);
  std::vector<float> operator_output(batch_size * output_channels);
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));
  std::vector<xnn_dynamic_quantization_params> quantization_params(batch_size + XNN_EXTRA_QUANTIZATION_PARAMS);

  std::vector<xnn_bfloat16> kernel_scale(output_channels * block_size);
  std::generate(kernel_scale.begin(), kernel_scale.end(), [&]() { return scale_dist(rng); });
  std::generate(kernel.begin(), kernel.end(), [&]() { return w8dist(rng); });
  std::generate(bias.begin(), bias.end(), [&]() { return f32dist(rng); });
  std::generate(convert_input.begin(), convert_input.end(), [&]() { return f32dist(rng); });
  std::generate(quantization_params.begin(), quantization_params.end(), [&]() { return xnn_dynamic_quantization_params{w8dist(rng), f32dist(rng)}; });

  // Adjust number of kernel elements for QC4W. input_channels should be padded to byte boundary, hence even.
  const size_t rounded_input_channels = round_up_po2(input_channels, 2);
  kernel = std::vector<uint8_t>(output_channels * rounded_input_channels);

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

  const uint8_t kernel_zero_point = 8;

  // Call operator API.
  xnn_operator_t convert_op = nullptr;
  xnn_operator_t fc_op = nullptr;
  xnn_status status = xnn_create_convert_nc_f32_qd8(
    /*flags=*/0, &convert_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convert_op(convert_op, xnn_delete_operator);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }
  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, convert_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_convert_nc_f32_qd8(convert_op, batch_size, input_channels, input_channels, input_channels, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_convert_nc_f32_qd8(convert_op, convert_input.data(),
                                                             operator_dq_data.data(), quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(convert_op, /*threadpool=*/nullptr));

  status = xnn_create_fully_connected_nc_qd8_f32_qb4w(
    input_channels, output_channels, input_channels, output_channels, block_size, kernel_zero_point, reinterpret_cast<const uint16_t*>(kernel_scale.data()),
    kernel.data(), bias.data(), output_min, output_max,
    /*flags=*/0, nullptr, nullptr, &fc_op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fc_op(fc_op, xnn_delete_operator);

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

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, fc_op);
  ASSERT_EQ(xnn_status_success, xnn_reshape_fully_connected_nc_qd8_f32_qb4w(fc_op, batch_size, /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success,
            xnn_setup_fully_connected_nc_qd8_f32_qb4w(fc_op, operator_dq_data.data(), operator_output.data(),
                                                      quantization_params.data()));
  ASSERT_EQ(xnn_status_success, xnn_run_operator(fc_op, /*threadpool=*/nullptr));

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

  uint32_t dq_quantized_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_dynamically_quantized_tensor_value(
                          subgraph, xnn_datatype_qdint8, input_dims.size(), /*num_nonbatch_dims=*/1, input_dims.data(),
                          XNN_INVALID_VALUE_ID, /*flags=*/0, &dq_quantized_id));
  ASSERT_NE(dq_quantized_id, XNN_INVALID_NODE_ID);
  uint32_t kernel_id = XNN_INVALID_VALUE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_blockwise_quantized_tensor_value(
        subgraph, xnn_datatype_qbint4, kernel_zero_point, reinterpret_cast<const uint16_t*>(kernel_scale.data()), kernel_dims.size(),
        /*channel_dim=*/0, block_size, kernel_dims.data(), kernel.data(), /*external_id=*/1, /*flags=*/0, &kernel_id));

  uint32_t bias_id = XNN_INVALID_VALUE_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=*/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, dq_quantized_id, /*flags=*/0));
  ASSERT_EQ(xnn_status_success, xnn_define_fully_connected(subgraph, output_min, output_max, dq_quantized_id,
                                                           kernel_id, bias_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, convert_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));
}

TEST_F(FullyConnectedTestF32, reshape)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/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(subgraph_output.begin(), subgraph_output.end(), nanf(""));

  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_VALUE_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=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_VALUE_ID);

  uint32_t kernel_id = XNN_INVALID_VALUE_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_VALUE_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_VALUE_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=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_VALUE_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);
  ASSERT_EQ(subgraph->num_values, 4);
  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);

  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));

  std::vector<size_t> new_input_dims(input_dims.begin(), input_dims.end());
  const xnn_shape* output_shape = &runtime->values[node->outputs[0]].shape;
  const xnn_shape* kernel_shape = &runtime->values[node->inputs[1]].shape;


  // case 1 : no change in input dims
  ASSERT_EQ(xnn_status_success, node->reshape(&runtime->opdata[0], subgraph->values, subgraph->num_values, nullptr /*threadpool*/));

  // case 2: Resize input to a smaller size. This should not require memory planning
  for (size_t i=0; i<new_input_dims.size() - 1; ++i) {
    new_input_dims[i] = input_dims[i] - 1;
  }

  ASSERT_EQ(xnn_status_success, xnn_reshape_external_value(runtime, input_id, new_input_dims.size(), new_input_dims.data()));

  ASSERT_EQ(xnn_status_success, node->reshape(&runtime->opdata[0], runtime->values, runtime->num_values, nullptr /*threadpool*/));

  // Check that the output shape is correct
  for (size_t i=0; i<output_shape->num_dims - 1; ++i) {
    ASSERT_EQ(output_shape->dim[i], new_input_dims[i]);
  }
  ASSERT_EQ(output_shape->dim[output_shape->num_dims-1], kernel_shape->dim[0]);

  // case 3: Resize input to a larger size. This should require memory planning
  for (size_t i=0; i<new_input_dims.size() - 1; ++i) {
    new_input_dims[i] = input_dims[i] + 1;
  }

  ASSERT_EQ(xnn_status_success, xnn_reshape_external_value(runtime, input_id, new_input_dims.size(), new_input_dims.data()));

  ASSERT_EQ(xnn_status_reallocation_required, node->reshape(&runtime->opdata[0], runtime->values, runtime->num_values, nullptr /*threadpool*/));

  // Check that the output shape is correct
  for (size_t i=0; i<output_shape->num_dims - 1; ++i) {
    ASSERT_EQ(output_shape->dim[i], new_input_dims[i]);
  }
  ASSERT_EQ(output_shape->dim[output_shape->num_dims-1], kernel_shape->dim[0]);

  size_t num_output_elements = std::accumulate(new_input_dims.begin(), new_input_dims.end() - 1, size_t{1}, std::multiplies<size_t>()) * kernel_shape->dim[0];
  ASSERT_EQ(runtime->values[node->outputs[0]].size, num_output_elements * sizeof(float));
}