XNNPACK/test/fully-connected-operator-tester.h

// Copyright (c) Facebook, Inc. and its affiliates.
// All rights reserved.
//
// Copyright 2019 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.

#pragma once

#include <gtest/gtest.h>

#include <cassert>
#include <cstddef>
#include <cstdlib>
#include <algorithm>
#include <cmath>
#include <limits>
#include <random>
#include <vector>

#include <fp16.h>

#include <xnnpack.h>
#include <xnnpack/cache.h>


class FullyConnectedOperatorTester {
 public:
  enum class WeightsType {
    Default,
    FP32,
  };

  inline FullyConnectedOperatorTester& input_channels(size_t input_channels) {
    assert(input_channels >= 1);
    this->input_channels_ = input_channels;
    return *this;
  }

  inline size_t input_channels() const {
    return this->input_channels_;
  }

  inline FullyConnectedOperatorTester& output_channels(size_t output_channels) {
    assert(output_channels >= 1);
    this->output_channels_ = output_channels;
    return *this;
  }

  inline size_t output_channels() const {
    return this->output_channels_;
  }

  inline FullyConnectedOperatorTester& batch_size(size_t batch_size) {
    assert(batch_size >= 1);
    this->batch_size_ = batch_size;
    return *this;
  }

  inline size_t batch_size() const {
    return this->batch_size_;
  }

  inline FullyConnectedOperatorTester& input_stride(size_t input_stride) {
    assert(input_stride >= 1);
    this->input_stride_ = input_stride;
    return *this;
  }

  inline size_t input_stride() const {
    if (this->input_stride_ == 0) {
      return input_channels();
    } else {
      assert(this->input_stride_ >= input_channels());
      return this->input_stride_;
    }
  }

  inline FullyConnectedOperatorTester& output_stride(size_t output_stride) {
    assert(output_stride >= 1);
    this->output_stride_ = output_stride;
    return *this;
  }

  inline size_t output_stride() const {
    if (this->output_stride_ == 0) {
      return output_channels();
    } else {
      assert(this->output_stride_ >= output_channels());
      return this->output_stride_;
    }
  }

  inline FullyConnectedOperatorTester& qmin(uint8_t qmin) {
    this->qmin_ = qmin;
    return *this;
  }

  inline uint8_t qmin() const {
    return this->qmin_;
  }

  inline FullyConnectedOperatorTester& qmax(uint8_t qmax) {
    this->qmax_ = qmax;
    return *this;
  }

  inline uint8_t qmax() const {
    return this->qmax_;
  }

  inline FullyConnectedOperatorTester& transpose_weights(bool transpose_weights) {
    this->transpose_weights_ = transpose_weights;
    return *this;
  }

  inline bool transpose_weights() const {
    return this->transpose_weights_;
  }

  inline FullyConnectedOperatorTester& has_bias(bool has_bias) {
    this->has_bias_ = has_bias;
    return *this;
  }

  inline bool has_bias() const {
    return this->has_bias_;
  }

  inline FullyConnectedOperatorTester& weights_type(WeightsType weights_type) {
    this->weights_type_ = weights_type;
    return *this;
  }

  inline WeightsType weights_type() const {
    return this->weights_type_;
  }

  inline FullyConnectedOperatorTester& use_weights_cache(bool use_weights_cache) {
    this->use_weights_cache_ = use_weights_cache;
    return *this;
  }

  inline bool use_weights_cache() const {
    return this->use_weights_cache_;
  }

  inline FullyConnectedOperatorTester& iterations(size_t iterations) {
    this->iterations_ = iterations;
    return *this;
  }

  inline size_t iterations() const {
    return this->iterations_;
  }

  void TestQS8() const {
    ASSERT_EQ(weights_type(), WeightsType::Default);

    std::random_device random_device;
    auto rng = std::mt19937(random_device());
    std::uniform_int_distribution<int32_t> i32dist(-10000, 10000);
    std::uniform_int_distribution<int32_t> i8dist(
      std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max());
    std::uniform_int_distribution<int32_t> w8dist(
      -std::numeric_limits<int8_t>::max(), std::numeric_limits<int8_t>::max());

    std::vector<int8_t> input(XNN_EXTRA_BYTES / sizeof(int8_t) +
      (batch_size() - 1) * input_stride() + input_channels());
    std::vector<int8_t> kernel(output_channels() * input_channels());
    std::vector<int32_t> bias(output_channels());
    std::vector<int8_t> output((batch_size() - 1) * output_stride() + output_channels());
    std::vector<int32_t> accumulators(batch_size() * output_channels());
    std::vector<double> output_ref(batch_size() * output_channels());

    const int8_t input_zero_point = 127;

    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      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(output.begin(), output.end(), INT8_C(0xA5));

      // Compute reference results, without renormalization.
      if (has_bias()) {
        for (size_t i = 0; i < batch_size(); i++) {
          for (size_t oc = 0; oc < output_channels(); oc++) {
            accumulators[i * output_channels() + oc] = bias[oc];
          }
        }
      } else {
        std::fill(accumulators.begin(), accumulators.end(), 0);
      }
      if (transpose_weights()) {
        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_stride() + ic]) - int32_t(input_zero_point)) *
                int32_t(kernel[ic * output_channels() + oc]);
            }
          }
        }
      } else {
        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_stride() + 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());

      const double output_scale = double(uint32_t(accumulated_max - accumulated_min)) / 255.0;
      const 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())));

      // Renormalize reference results.
      std::transform(accumulators.cbegin(), accumulators.cend(), output_ref.begin(),
        [this, output_scale, output_zero_point](int32_t x) -> double {
          return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax() - 0x80) - output_zero_point), double(qmin() - 0x80) - output_zero_point);
        });

      // Create, setup, run, and destroy Fully Connected operator.
      ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */));
      xnn_operator_t fully_connected_op = nullptr;

      xnn_caches caches = {
        .code_cache = NULL,
        .weights_cache = NULL,
      };
      xnn_weights_cache weights_cache;
      if (use_weights_cache()) {
        xnn_init_weights_cache(&weights_cache);
        caches.weights_cache = &weights_cache;
      }

      const xnn_status status = xnn_create_fully_connected_nc_qs8(
          input_channels(), output_channels(),
          input_stride(), output_stride(),
          input_zero_point, 1.0f /* input scale */,
          1.0f /* kernel scale */,
          kernel.data(), has_bias() ? bias.data() : nullptr,
          output_zero_point, output_scale, int8_t(qmin() - 0x80), int8_t(qmax() - 0x80),
          transpose_weights() ? XNN_FLAG_TRANSPOSE_WEIGHTS : 0,
          &caches,
          &fully_connected_op);
      if (status == xnn_status_unsupported_hardware) {
        GTEST_SKIP();
      }
      ASSERT_EQ(xnn_status_success, status);
      ASSERT_NE(nullptr, fully_connected_op);
      if (use_weights_cache()) {
        ASSERT_EQ(xnn_status_success,
                  xnn_finalize_weights_cache(&weights_cache, xnn_weights_cache_finalization_kind_soft));
      }

      // Smart pointer to automatically delete fully_connected_op.
      std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fully_connected_op(fully_connected_op, xnn_delete_operator);

      ASSERT_EQ(xnn_status_success,
        xnn_setup_fully_connected_nc_qs8(
          fully_connected_op,
          batch_size(),
          input.data(), output.data(),
          nullptr /* thread pool */));

      ASSERT_EQ(xnn_status_success,
        xnn_run_operator(fully_connected_op, nullptr /* thread pool */));

      // Verify results.
      VerifyQS8(output, output_ref, double(output_zero_point));

      if (use_weights_cache()) {
        // Create another operator with the same weights cache.
        xnn_operator_t fully_connected_op2 = nullptr;
        size_t old_weights_cache_size = weights_cache.cache.weights.size;

        ASSERT_EQ(xnn_status_success,
                  xnn_create_fully_connected_nc_qs8(
                      input_channels(), output_channels(), input_stride(),
                      output_stride(), input_zero_point, 1.0f /* input scale */,
                      1.0f /* kernel scale */, kernel.data(),
                      has_bias() ? bias.data() : nullptr, output_zero_point,
                      output_scale, int8_t(qmin() - 0x80),
                      int8_t(qmax() - 0x80),
                      transpose_weights() ? XNN_FLAG_TRANSPOSE_WEIGHTS : 0,
                      &caches, &fully_connected_op2));
        ASSERT_NE(nullptr, fully_connected_op2);

        // Smart pointer to automatically delete fully_connected_op.
        std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)>
            auto_fully_connected_op(fully_connected_op2, xnn_delete_operator);
        std::vector<int8_t> output2(output.size(), INT8_C(0xA5));

        ASSERT_EQ(xnn_status_success,
                  xnn_setup_fully_connected_nc_qs8(
                      fully_connected_op2,
                      batch_size(),
                      input.data(), output2.data(),
                      nullptr /* thread pool */));

        ASSERT_EQ(
            xnn_status_success,
            xnn_run_operator(fully_connected_op2, nullptr /* thread pool */));

        VerifyWeightsCache(weights_cache, old_weights_cache_size);
        xnn_release_weights_cache(&weights_cache);

        VerifyQS8(output, output_ref, double(output_zero_point));
      }
    }
  }

  void VerifyQS8(const std::vector<int8_t>& output,
                 const std::vector<double>& output_ref,
                 double output_zero_point) const {
    for (size_t i = 0; i < batch_size(); i++) {
      for (size_t c = 0; c < output_channels(); c++) {
        ASSERT_LE(int32_t(output[i * output_stride() + c]), int32_t(qmax() - 0x80))
            << "batch index = " << i << ", channel = " << c;
        ASSERT_GE(int32_t(output[i * output_stride() + c]), int32_t(qmin() - 0x80))
            << "batch index = " << i << ", channel = " << c;
        ASSERT_NEAR(output_ref[i * output_channels() + c],
                    double(output[i * output_stride() + c]) - output_zero_point,
                    0.9)
            << "batch index = " << i << ", channel = " << c;
      }
    }
  }

  void TestQU8() const {
    ASSERT_EQ(weights_type(), WeightsType::Default);

    std::random_device random_device;
    auto rng = std::mt19937(random_device());
    std::uniform_int_distribution<int32_t> i32dist(-10000, 10000);
    std::uniform_int_distribution<int32_t> u8dist(
      std::numeric_limits<uint8_t>::min(), std::numeric_limits<uint8_t>::max());

    std::vector<uint8_t> input(XNN_EXTRA_BYTES / sizeof(uint8_t) +
      (batch_size() - 1) * input_stride() + input_channels());
    std::vector<uint8_t> kernel(output_channels() * input_channels());
    std::vector<int32_t> bias(output_channels());
    std::vector<uint8_t> output((batch_size() - 1) * output_stride() + output_channels());
    std::vector<int32_t> accumulators(batch_size() * output_channels());
    std::vector<double> output_ref(batch_size() * output_channels());

    const uint8_t input_zero_point = 127;
    const uint8_t kernel_zero_point = 127;

    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      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(output.begin(), output.end(), UINT8_C(0xA5));

      // Compute reference results, without renormalization.
      if (has_bias()) {
        for (size_t i = 0; i < batch_size(); i++) {
          for (size_t oc = 0; oc < output_channels(); oc++) {
            accumulators[i * output_channels() + oc] = bias[oc];
          }
        }
      } else {
        std::fill(accumulators.begin(), accumulators.end(), 0);
      }
      if (transpose_weights()) {
        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_stride() + ic]) - int32_t(input_zero_point)) *
                (int32_t(kernel[ic * output_channels() + oc]) - int32_t(kernel_zero_point));
            }
          }
        }
      } else {
        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_stride() + 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())));

      // Renormalize reference results.
      std::transform(accumulators.cbegin(), accumulators.cend(), output_ref.begin(),
        [this, output_scale, output_zero_point](int32_t x) -> double {
          return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax()) - output_zero_point), double(qmin()) - output_zero_point);
        });

      // Create, setup, run, and destroy Fully Connected operator.
      ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */));
      xnn_operator_t fully_connected_op = nullptr;

      xnn_caches caches = {
        .code_cache = NULL,
        .weights_cache = NULL,
      };
      xnn_weights_cache weights_cache;
      if (use_weights_cache()) {
        xnn_init_weights_cache(&weights_cache);
        caches.weights_cache = &weights_cache;
      }

      const xnn_status status = xnn_create_fully_connected_nc_qu8(
          input_channels(), output_channels(),
          input_stride(), output_stride(),
          input_zero_point, 1.0f /* input scale */,
          kernel_zero_point, 1.0f /* kernel scale */,
          kernel.data(), has_bias() ? bias.data() : nullptr,
          output_zero_point, output_scale, qmin(), qmax(),
          transpose_weights() ? XNN_FLAG_TRANSPOSE_WEIGHTS : 0,
          &caches,
          &fully_connected_op);
      if (status == xnn_status_unsupported_hardware) {
        GTEST_SKIP();
      }
      ASSERT_EQ(xnn_status_success, status);
      ASSERT_NE(nullptr, fully_connected_op);
      if (use_weights_cache()) {
        ASSERT_EQ(xnn_status_success,
                  xnn_finalize_weights_cache(&weights_cache, xnn_weights_cache_finalization_kind_soft));
      }

      // Smart pointer to automatically delete fully_connected_op.
      std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fully_connected_op(fully_connected_op, xnn_delete_operator);

      ASSERT_EQ(xnn_status_success,
        xnn_setup_fully_connected_nc_qu8(
          fully_connected_op,
          batch_size(),
          input.data(), output.data(),
          nullptr /* thread pool */));

      ASSERT_EQ(xnn_status_success,
        xnn_run_operator(fully_connected_op, nullptr /* thread pool */));

      VerifyQU8(output, output_ref, double(output_zero_point));

      if (use_weights_cache()) {
        // Create another operator with the same weights cache.
        xnn_operator_t fully_connected_op2 = nullptr;
        size_t old_weights_cache_size = weights_cache.cache.weights.size;

        ASSERT_EQ(xnn_status_success,
                  xnn_create_fully_connected_nc_qu8(
                      input_channels(), output_channels(), input_stride(),
                      output_stride(), input_zero_point, 1.0f /* input scale */,
                      kernel_zero_point, 1.0f /* kernel scale */, kernel.data(),
                      has_bias() ? bias.data() : nullptr, output_zero_point,
                      output_scale, qmin(), qmax(),
                      transpose_weights() ? XNN_FLAG_TRANSPOSE_WEIGHTS : 0,
                      &caches, &fully_connected_op2));
        ASSERT_NE(nullptr, fully_connected_op2);

        // Smart pointer to automatically delete fully_connected_op.
        std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)>
            auto_fully_connected_op(fully_connected_op2, xnn_delete_operator);
        std::vector<uint8_t> output2(output.size(), UINT8_C(0xA5));

        ASSERT_EQ(xnn_status_success,
                  xnn_setup_fully_connected_nc_qu8(
                      fully_connected_op2, batch_size(), input.data(),
                      output2.data(), nullptr /* thread pool */));

        ASSERT_EQ(
            xnn_status_success,
            xnn_run_operator(fully_connected_op2, nullptr /* thread pool */));

        VerifyWeightsCache(weights_cache, old_weights_cache_size);
        xnn_release_weights_cache(&weights_cache);

        VerifyQU8(output2, output_ref, double(output_zero_point));
      }

    }
  }

  void VerifyQU8(const std::vector<uint8_t>& output,
                 const std::vector<double>& output_ref,
                 double output_zero_point) const {
    for (size_t i = 0; i < batch_size(); i++) {
      for (size_t c = 0; c < output_channels(); c++) {
        ASSERT_LE(int32_t(output[i * output_stride() + c]), int32_t(qmax()))
            << "batch index = " << i << ", channel = " << c;
        ASSERT_GE(int32_t(output[i * output_stride() + c]), int32_t(qmin()))
            << "batch index = " << i << ", channel = " << c;
        ASSERT_NEAR(output_ref[i * output_channels() + c],
                    double(output[i * output_stride() + c]) - output_zero_point,
                    0.9)
            << "batch index = " << i << ", channel = " << c;
      }
    }
  }

  void TestF32() const {
    ASSERT_EQ(weights_type(), WeightsType::Default);

    std::random_device random_device;
    auto rng = std::mt19937(random_device());
    std::uniform_real_distribution<float> f32dist(0.1f, 1.0f);

    std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) +
      (batch_size() - 1) * input_stride() + input_channels());
    std::vector<float> kernel(output_channels() * input_channels());
    std::vector<float> bias(output_channels());
    std::vector<float> output((batch_size() - 1) * output_stride() + output_channels());
    std::vector<float> output_ref(batch_size() * output_channels());

    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      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(output.begin(), output.end(), nanf(""));

      // Compute reference results, without renormalization.
      if (has_bias()) {
        for (size_t i = 0; i < batch_size(); i++) {
          for (size_t oc = 0; oc < output_channels(); oc++) {
            output_ref[i * output_channels() + oc] = bias[oc];
          }
        }
      } else {
        std::fill(output_ref.begin(), output_ref.end(), 0.0f);
      }
      if (transpose_weights()) {
        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++) {
              output_ref[i * output_channels() + oc] +=
                input[i * input_stride() + ic] * kernel[ic * output_channels() + oc];
            }
          }
        }
      } else {
        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++) {
              output_ref[i * output_channels() + oc] +=
                input[i * input_stride() + ic] * kernel[oc * input_channels() + ic];
            }
          }
        }
      }

      // Compute clamping parameters.
      const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend());
      const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend());

      const float output_min = qmin() == 0 ? -std::numeric_limits<float>::infinity() :
        accumulated_min + (accumulated_max - accumulated_min) / 255.0f * float(qmin());
      const float output_max = qmax() == 255 ? std::numeric_limits<float>::infinity() :
        accumulated_max - (accumulated_max - accumulated_min) / 255.0f * float(255 - qmax());

      // Clamp reference results.
      for (float& value : output_ref) {
        value = std::max(std::min(value, output_max), output_min);
      }

      // Create, setup, run, and destroy Fully Connected operator.
      ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */));
      xnn_operator_t fully_connected_op = nullptr;

      xnn_caches caches = {
        .code_cache = NULL,
        .weights_cache = NULL,
      };
      xnn_weights_cache weights_cache;
      if (use_weights_cache()) {
        xnn_init_weights_cache(&weights_cache);
        caches.weights_cache = &weights_cache;
      }

      const xnn_status status = xnn_create_fully_connected_nc_f32(
          input_channels(), output_channels(),
          input_stride(), output_stride(),
          kernel.data(), has_bias() ? bias.data() : nullptr,
          output_min, output_max,
          transpose_weights() ? XNN_FLAG_TRANSPOSE_WEIGHTS : 0,
          &caches,
          &fully_connected_op);
      if (status == xnn_status_unsupported_hardware) {
        GTEST_SKIP();
      }
      ASSERT_EQ(xnn_status_success, status);
      ASSERT_NE(nullptr, fully_connected_op);
      if (use_weights_cache()) {
        ASSERT_EQ(xnn_status_success,
                  xnn_finalize_weights_cache(&weights_cache, xnn_weights_cache_finalization_kind_soft));
      }

      // Smart pointer to automatically delete fully_connected_op.
      std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fully_connected_op(fully_connected_op, xnn_delete_operator);

      ASSERT_EQ(xnn_status_success,
        xnn_setup_fully_connected_nc_f32(
          fully_connected_op,
          batch_size(),
          input.data(), output.data(),
          nullptr /* thread pool */));

      ASSERT_EQ(xnn_status_success,
        xnn_run_operator(fully_connected_op, nullptr /* thread pool */));

      VerifyF32(output, output_ref, output_max, output_min);

      if (use_weights_cache()) {
        // Create another operator with the same weights cache.
        xnn_operator_t fully_connected_op2 = nullptr;
        size_t old_weights_cache_size = weights_cache.cache.weights.size;
        ASSERT_EQ(xnn_status_success,
                  xnn_create_fully_connected_nc_f32(
                      input_channels(), output_channels(), input_stride(),
                      output_stride(), kernel.data(),
                      has_bias() ? bias.data() : nullptr, output_min,
                      output_max,
                      transpose_weights() ? XNN_FLAG_TRANSPOSE_WEIGHTS : 0,
                      &caches, &fully_connected_op2));
        ASSERT_NE(nullptr, fully_connected_op2);

        std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fully_connected_op(fully_connected_op2, xnn_delete_operator);

        std::vector<float> output2(output.size(), nanf(""));
        ASSERT_EQ(xnn_status_success,
                  xnn_setup_fully_connected_nc_f32(
                      fully_connected_op2,
                      batch_size(),
                      input.data(), output2.data(),
                      nullptr /* thread pool */));

        ASSERT_EQ(xnn_status_success,
                  xnn_run_operator(fully_connected_op2, nullptr /* thread pool */));
        VerifyWeightsCache(weights_cache, old_weights_cache_size);
        xnn_release_weights_cache(&weights_cache);

        VerifyF32(output, output_ref, output_max, output_min);
      }
    }
  }

  void VerifyF32(const std::vector<float>& output,
                 const std::vector<float>& output_ref,
                 float output_max,
                 float output_min) const {
    // Verify results.
    for (size_t i = 0; i < batch_size(); i++) {
      for (size_t c = 0; c < output_channels(); c++) {
        ASSERT_LE(output[i * output_stride() + c], output_max)
            << "batch index = " << i << ", channel = " << c;
        ASSERT_GE(output[i * output_stride() + c], output_min)
            << "batch index = " << i << ", channel = " << c;
        ASSERT_NEAR(output_ref[i * output_channels() + c],
                    output[i * output_stride() + c],
                    1.0e-4 * std::abs(output_ref[i * output_channels() + c]))
            << "batch index = " << i << ", channel = " << c;
      }
    }
  }

  void TestF16() const {
    switch (weights_type()) {
      case WeightsType::Default:
        break;
      case WeightsType::FP32:
        break;
      default:
        GTEST_FAIL() << "unexpected weights type";
    }

    std::random_device random_device;
    auto rng = std::mt19937(random_device());
    std::uniform_real_distribution<float> f32dist(0.1f, 1.0f);

    std::vector<uint16_t> input(XNN_EXTRA_BYTES / sizeof(uint16_t) +
      (batch_size() - 1) * input_stride() + input_channels());
    std::vector<uint16_t> kernel(output_channels() * input_channels());
    std::vector<float> kernel_as_float(kernel.size());
    std::vector<uint16_t> bias(output_channels());
    std::vector<float> bias_as_float(bias.size());
    std::vector<uint16_t> output((batch_size() - 1) * output_stride() + output_channels());
    std::vector<float> output_ref(batch_size() * output_channels());

    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      std::generate(input.begin(), input.end(), [&]() { return fp16_ieee_from_fp32_value(f32dist(rng)); });
      std::generate(kernel.begin(), kernel.end(), [&]() { return fp16_ieee_from_fp32_value(f32dist(rng)); });
      std::transform(kernel.cbegin(), kernel.cend(), kernel_as_float.begin(), fp16_ieee_to_fp32_value);
      std::generate(bias.begin(), bias.end(), [&]() { return fp16_ieee_from_fp32_value(f32dist(rng)); });
      std::transform(bias.cbegin(), bias.cend(), bias_as_float.begin(), fp16_ieee_to_fp32_value);
      std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */);

      // Compute reference results, without renormalization.
      if (has_bias()) {
        for (size_t i = 0; i < batch_size(); i++) {
          for (size_t oc = 0; oc < output_channels(); oc++) {
            output_ref[i * output_channels() + oc] = fp16_ieee_to_fp32_value(bias[oc]);
          }
        }
      } else {
        std::fill(output_ref.begin(), output_ref.end(), 0.0f);
      }
      if (transpose_weights()) {
        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++) {
              output_ref[i * output_channels() + oc] +=
                fp16_ieee_to_fp32_value(input[i * input_stride() + ic]) * fp16_ieee_to_fp32_value(kernel[ic * output_channels() + oc]);
            }
          }
        }
      } else {
        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++) {
              output_ref[i * output_channels() + oc] +=
                fp16_ieee_to_fp32_value(input[i * input_stride() + ic]) * fp16_ieee_to_fp32_value(kernel[oc * input_channels() + ic]);
            }
          }
        }
      }

      // Compute clamping parameters.
      const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend());
      const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend());
      const float accumulated_range = accumulated_max - accumulated_min;
      const float scaled_min = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_min + accumulated_range / 255.0f * float(qmin())));
      const float scaled_max = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_max - accumulated_range / 255.0f * float(255 - qmax())));
      const float output_min = scaled_min == scaled_max ? -std::numeric_limits<float>::infinity() : scaled_min;
      const float output_max = scaled_min == scaled_max ? +std::numeric_limits<float>::infinity() : scaled_max;

      // Clamp reference results.
      for (float& value : output_ref) {
        value = std::max(std::min(value, output_max), output_min);
      }

      // Create, setup, run, and destroy Fully Connected operator.
      ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */));
      xnn_operator_t fully_connected_op = nullptr;

      xnn_caches caches = {
        .code_cache = NULL,
        .weights_cache = NULL,
      };
      xnn_weights_cache weights_cache;
      if (use_weights_cache()) {
        xnn_init_weights_cache(&weights_cache);
        caches.weights_cache = &weights_cache;
      }

      const void* kernel_data = kernel.data();
      const void* bias_data = bias.data();
      if (weights_type() == WeightsType::FP32) {
        kernel_data = kernel_as_float.data();
        bias_data = bias_as_float.data();
      }
      uint32_t flags = 0;
      if (transpose_weights()) {
        flags |= XNN_FLAG_TRANSPOSE_WEIGHTS;
      }
      if (weights_type() == WeightsType::FP32) {
        flags |= XNN_FLAG_FP32_STATIC_WEIGHTS;
      }
      const xnn_status status = xnn_create_fully_connected_nc_f16(
          input_channels(), output_channels(),
          input_stride(), output_stride(),
          kernel_data, has_bias() ? bias_data : nullptr,
          output_min, output_max,
          flags,
          &caches,
          &fully_connected_op);
      if (status == xnn_status_unsupported_hardware) {
        GTEST_SKIP();
      }
      ASSERT_EQ(xnn_status_success, status);
      ASSERT_NE(nullptr, fully_connected_op);
      if (use_weights_cache()) {
        ASSERT_EQ(xnn_status_success,
                  xnn_finalize_weights_cache(&weights_cache, xnn_weights_cache_finalization_kind_soft));
      }

      // Smart pointer to automatically delete fully_connected_op.
      std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fully_connected_op(fully_connected_op, xnn_delete_operator);

      ASSERT_EQ(xnn_status_success,
        xnn_setup_fully_connected_nc_f16(
          fully_connected_op,
          batch_size(),
          input.data(), output.data(),
          nullptr /* thread pool */));

      ASSERT_EQ(xnn_status_success,
        xnn_run_operator(fully_connected_op, nullptr /* thread pool */));

      // Verify results.
      VerifyF16(output, output_ref, output_max, output_min);

      if (use_weights_cache()) {
        xnn_operator_t fully_connected_op2 = nullptr;
        size_t old_weights_cache_size = weights_cache.cache.weights.size;
        ASSERT_EQ(xnn_status_success,
                  xnn_create_fully_connected_nc_f16(
                      input_channels(), output_channels(), input_stride(),
                      output_stride(), kernel_data,
                      has_bias() ? bias_data : nullptr, output_min, output_max,
                      flags, &caches, &fully_connected_op2));
        if (status == xnn_status_unsupported_hardware) {
          GTEST_SKIP();
        }
        ASSERT_EQ(xnn_status_success, status);
        ASSERT_NE(nullptr, fully_connected_op2);

        // Smart pointer to automatically delete fully_connected_op2.
        std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_fully_connected_op(fully_connected_op2, xnn_delete_operator);
        std::vector<uint16_t> output2(output.size(), UINT16_C(0x7E00) /* NaN */);

        ASSERT_EQ(xnn_status_success,
                  xnn_setup_fully_connected_nc_f16(
                      fully_connected_op2,
                      batch_size(),
                      input.data(), output2.data(),
                      nullptr /* thread pool */));

        ASSERT_EQ(xnn_status_success,
                  xnn_run_operator(fully_connected_op2, nullptr /* thread pool */));

        // Verify results.
        VerifyF16(output2, output_ref, output_max, output_min);
        VerifyWeightsCache(weights_cache, old_weights_cache_size);
        xnn_release_weights_cache(&weights_cache);
      }
    }
  }

  void VerifyF16(const std::vector<uint16_t>& output,
                 const std::vector<float>& output_ref,
                 const float output_max,
                 const float output_min) const {
    for (size_t i = 0; i < batch_size(); i++) {
      for (size_t c = 0; c < output_channels(); c++) {
        ASSERT_LE(fp16_ieee_to_fp32_value(output[i * output_stride() + c]), output_max)
          << "batch index = " << i << ", channel = " << c;
        ASSERT_GE(fp16_ieee_to_fp32_value(output[i * output_stride() + c]), output_min)
          << "batch index = " << i << ", channel = " << c;
        ASSERT_NEAR(
            output_ref[i * output_channels() + c],
            fp16_ieee_to_fp32_value(output[i * output_stride() + c]),
            1.0e-2f * std::abs(output_ref[i * output_channels() + c]))
          << "batch index = " << i << ", channel = " << c;
      }
    }
  }

  void VerifyWeightsCache(const xnn_weights_cache& weights_cache, size_t old_size) const {
    ASSERT_EQ(weights_cache.cache.hits, 1);
    // Ensure that we did not write more weights to the cache because it was a cache hit.
    ASSERT_EQ(old_size, weights_cache.cache.weights.size);
  };

 private:
  size_t input_channels_{1};
  size_t input_stride_{0};
  size_t output_channels_{1};
  size_t output_stride_{0};
  size_t batch_size_{1};
  uint8_t qmin_{0};
  uint8_t qmax_{255};
  bool transpose_weights_{false};
  bool has_bias_{true};
  WeightsType weights_type_{WeightsType::Default};
  bool use_weights_cache_{false};
  size_t iterations_{1};
};