ibilinear-microkernel-tester.h

// 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 <algorithm>
#include <cassert>
#include <cmath>
#include <cstddef>
#include <cstdint>
#include <limits>
#include <random>
#include <vector>

#include <gtest/gtest.h>
#include <fp16/fp16.h>
#include "xnnpack.h"
#include "xnnpack/aligned-allocator.h"
#include "xnnpack/math.h"
#include "xnnpack/microfnptr.h"
#include "replicable_random_device.h"

class IBilinearMicrokernelTester {
 public:
  IBilinearMicrokernelTester& pixels(uint32_t pixels) {
    assert(pixels >= 1);
    this->pixels_ = pixels;
    return *this;
  }

  uint32_t pixels() const {
    return this->pixels_;
  }

  IBilinearMicrokernelTester& channels(uint32_t channels) {
    assert(channels >= 1);
    this->channels_ = channels;
    return *this;
  }

  uint32_t channels() const {
    return this->channels_;
  }

  IBilinearMicrokernelTester& input_offset(uint32_t input_offset) {
    this->input_offset_ = input_offset;
    return *this;
  }

  uint32_t input_offset() const {
    return this->input_offset_;
  }

  IBilinearMicrokernelTester& output_stride(uint32_t output_stride) {
    assert(output_stride != 0);
    this->output_stride_ = output_stride;
    return *this;
  }

  uint32_t output_stride() const {
    if (this->output_stride_ == 0) {
      return channels();
    } else {
      assert(this->output_stride_ >= channels());
      return this->output_stride_;
    }
  }

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

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

  IBilinearMicrokernelTester& input_stride(uint32_t input_stride) {
    assert(input_stride != 0);
    this->input_stride_ = input_stride;
    return *this;
  }

  uint32_t input_stride() const {
    if (this->input_stride_ == 0) {
      return 4 * pixels();
    } else {
      assert(this->input_stride_ >= 4 * pixels());
      return this->input_stride_;
    }
  }

  void Test(xnn_f16_ibilinear_ukernel_fn ibilinear) const {
    xnnpack::ReplicableRandomDevice rng;
    std::uniform_real_distribution<float> f32dist(0.1f, 1.0f);

    std::vector<const xnn_float16*> indirection(pixels() * 4);
    std::vector<xnn_float16> input(XNN_EXTRA_BYTES / sizeof(xnn_float16) + indirection.size() * channels());
    std::vector<xnn_float16, AlignedAllocator<xnn_float16, 64>> packed_weights(pixels() * 2);
    std::vector<xnn_float16> output((pixels() - 1) * output_stride() + channels());
    std::vector<float> output_ref(pixels() * channels());

    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
      std::generate(packed_weights.begin(), packed_weights.end(), [&]() { return f32dist(rng); });
      std::fill(output.begin(), output.end(), std::nanf(""));

      for (size_t i = 0; i < indirection.size(); i++) {
        indirection[i] = input.data() + i * channels() - input_offset();
      }
      std::shuffle(indirection.begin(), indirection.end(), rng);

      // Compute reference results.
      for (size_t i = 0; i < pixels(); i++) {
        for (size_t c = 0; c < channels(); c++) {
          const float alpha_h = packed_weights[i * 2 + 0];
          const float alpha_v = packed_weights[i * 2 + 1];
          output_ref[i * channels() + c] =
            indirection[i * 4 + 0][c + input_offset()] * (1.0f - alpha_h) * (1.0f - alpha_v) +
            indirection[i * 4 + 1][c + input_offset()] * alpha_h * (1.0f - alpha_v) +
            indirection[i * 4 + 2][c + input_offset()] * (1.0f - alpha_h) * alpha_v +
            indirection[i * 4 + 3][c + input_offset()] * alpha_h * alpha_v;
        }
      }

      // Call optimized micro-kernel.
      ibilinear(
        pixels(), channels() * sizeof(xnn_float16),
        reinterpret_cast<const xnn_float16**>(indirection.data()), input_offset() * sizeof(xnn_float16),
        packed_weights.data(), output.data(),
        (output_stride() - channels()) * sizeof(xnn_float16));

      // Verify results.
      for (size_t i = 0; i < pixels(); i++) {
        for (size_t c = 0; c < channels(); c++) {
          ASSERT_NEAR(
              output[i * output_stride() + c],
              output_ref[i * channels() + c],
              std::abs(output_ref[i * channels() + c]) * 1.0e-2f)
            << "pixel " << i << " / " << pixels() << ", channel " << c << " / " << channels();
        }
      }
    }
  }

  void Test(xnn_f32_ibilinear_ukernel_fn ibilinear) const {
    xnnpack::ReplicableRandomDevice rng;
    std::uniform_real_distribution<float> f32dist;

    std::vector<const float*> indirection(pixels() * 4);
    std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + indirection.size() * channels());
    std::vector<float, AlignedAllocator<float, 64>> packed_weights(pixels() * 2);
    std::vector<float> output((pixels() - 1) * output_stride() + channels());
    std::vector<float> output_ref(pixels() * channels());

    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
      std::generate(packed_weights.begin(), packed_weights.end(), [&]() { return f32dist(rng); });
      std::fill(output.begin(), output.end(), nanf(""));

      for (size_t i = 0; i < indirection.size(); i++) {
        indirection[i] = input.data() + i * channels() - input_offset();
      }
      std::shuffle(indirection.begin(), indirection.end(), rng);

      // Compute reference results.
      for (size_t i = 0; i < pixels(); i++) {
        for (size_t c = 0; c < channels(); c++) {
          const float alpha_h = packed_weights[i * 2 + 0];
          const float alpha_v = packed_weights[i * 2 + 1];
          output_ref[i * channels() + c] =
            indirection[i * 4 + 0][c + input_offset()] * (1.0f - alpha_h) * (1.0f - alpha_v) +
            indirection[i * 4 + 1][c + input_offset()] * alpha_h * (1.0f - alpha_v) +
            indirection[i * 4 + 2][c + input_offset()] * (1.0f - alpha_h) * alpha_v +
            indirection[i * 4 + 3][c + input_offset()] * alpha_h * alpha_v;
        }
      }

      // Call optimized micro-kernel.
      ibilinear(
        pixels(), channels() * sizeof(float),
        indirection.data(), input_offset() * sizeof(float),
        packed_weights.data(), output.data(),
        (output_stride() - channels()) * sizeof(float));

      // Verify results.
      for (size_t i = 0; i < pixels(); i++) {
        for (size_t c = 0; c < channels(); c++) {
          EXPECT_NEAR(
              output_ref[i * channels() + c],
              output[i * output_stride() + c],
              std::abs(output_ref[i * channels() + c]) * 1.0e-4)
            << "pixel " << i << " / " << pixels() << ", channel " << c << " / " << channels();
        }
      }
    }
  }

  void Test(xnn_s8_ibilinear_ukernel_fn ibilinear) const {
    xnnpack::ReplicableRandomDevice rng;
    std::uniform_int_distribution<int32_t> i8dist(
      std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max());
    std::uniform_int_distribution<int16_t> w11dist(0, 2047);

    std::vector<const int8_t*> indirection(pixels() * 4);
    std::vector<int8_t> input(XNN_EXTRA_BYTES / sizeof(int8_t) + indirection.size() * channels());
    std::vector<int16_t, AlignedAllocator<int16_t, 64>> packed_weights(pixels() * 2);
    std::vector<int8_t> output((pixels() - 1) * output_stride() + channels());
    std::vector<int8_t> output_ref(pixels() * channels());

    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      std::generate(input.begin(), input.end(), [&]() { return i8dist(rng); });
      std::generate(packed_weights.begin(), packed_weights.end(), [&]() { return w11dist(rng); });
      std::fill(output.begin(), output.end(), INT8_C(0xFA));

      for (size_t i = 0; i < indirection.size(); i++) {
        indirection[i] = input.data() + i * channels() - input_offset();
      }
      std::shuffle(indirection.begin(), indirection.end(), rng);

      // Compute reference results.
      for (size_t i = 0; i < pixels(); i++) {
        for (size_t c = 0; c < channels(); c++) {
          const int32_t alpha_h = packed_weights[i * 2 + 0];
          const int32_t alpha_v = packed_weights[i * 2 + 1];
          const int32_t acc = math_asr_s32(
            int32_t(indirection[i * 4 + 0][c + input_offset()]) * (2048 - alpha_h) * (2048 - alpha_v) +
            int32_t(indirection[i * 4 + 1][c + input_offset()]) * alpha_h * (2048 - alpha_v) +
            int32_t(indirection[i * 4 + 2][c + input_offset()]) * (2048 - alpha_h) * alpha_v +
            int32_t(indirection[i * 4 + 3][c + input_offset()]) * alpha_h * alpha_v +
            2097152, 22);
          ASSERT_GE(acc, std::numeric_limits<int8_t>::min());
          ASSERT_LE(acc, std::numeric_limits<int8_t>::max());
          output_ref[i * channels() + c] = (int8_t) acc;
        }
      }

      // Call optimized micro-kernel.
      ibilinear(
        pixels(), channels() * sizeof(int8_t),
        indirection.data(), input_offset() * sizeof(int8_t),
        packed_weights.data(), output.data(),
        (output_stride() - channels()) * sizeof(int8_t));

      // Verify results.
      for (size_t i = 0; i < pixels(); i++) {
        for (size_t c = 0; c < channels(); c++) {
          EXPECT_EQ(int32_t(output_ref[i * channels() + c]), int32_t(output[i * output_stride() + c]))
            << "pixel " << i << " / " << pixels() << ", channel " << c << " / " << channels();
        }
      }
    }
  }

  void Test(xnn_u8_ibilinear_ukernel_fn ibilinear) const {
    xnnpack::ReplicableRandomDevice rng;
    std::uniform_int_distribution<int32_t> u8dist(
      std::numeric_limits<uint8_t>::min(), std::numeric_limits<uint8_t>::max());
    std::uniform_int_distribution<int16_t> w11dist(0, 2047);

    std::vector<const uint8_t*> indirection(pixels() * 4);
    std::vector<uint8_t> input(XNN_EXTRA_BYTES / sizeof(uint8_t) + indirection.size() * channels());
    std::vector<int16_t, AlignedAllocator<int16_t, 64>> packed_weights(pixels() * 2);
    std::vector<uint8_t> output((pixels() - 1) * output_stride() + channels());
    std::vector<uint8_t> output_ref(pixels() * channels());

    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      std::generate(input.begin(), input.end(), [&]() { return u8dist(rng); });
      std::generate(packed_weights.begin(), packed_weights.end(), [&]() { return w11dist(rng); });
      std::fill(output.begin(), output.end(), UINT8_C(0xFA));

      for (size_t i = 0; i < indirection.size(); i++) {
        indirection[i] = input.data() + i * channels() - input_offset();
      }
      std::shuffle(indirection.begin(), indirection.end(), rng);

      // Compute reference results.
      for (size_t i = 0; i < pixels(); i++) {
        for (size_t c = 0; c < channels(); c++) {
          const uint32_t alpha_h = uint32_t(int32_t(packed_weights[i * 2 + 0]));
          const uint32_t alpha_v = uint32_t(int32_t(packed_weights[i * 2 + 1]));
          const uint32_t acc = (2097152 +
            int32_t(indirection[i * 4 + 0][c + input_offset()]) * (2048 - alpha_h) * (2048 - alpha_v) +
            int32_t(indirection[i * 4 + 1][c + input_offset()]) * alpha_h * (2048 - alpha_v) +
            int32_t(indirection[i * 4 + 2][c + input_offset()]) * (2048 - alpha_h) * alpha_v +
            int32_t(indirection[i * 4 + 3][c + input_offset()]) * alpha_h * alpha_v) >> 22;
          ASSERT_LE(acc, std::numeric_limits<uint8_t>::max());
          output_ref[i * channels() + c] = (uint8_t) acc;
        }
      }

      // Call optimized micro-kernel.
      ibilinear(
        pixels(), channels() * sizeof(uint8_t),
        indirection.data(), input_offset() * sizeof(uint8_t),
        packed_weights.data(), output.data(),
        (output_stride() - channels()) * sizeof(uint8_t));

      // Verify results.
      for (size_t i = 0; i < pixels(); i++) {
        for (size_t c = 0; c < channels(); c++) {
          EXPECT_EQ(uint32_t(output_ref[i * channels() + c]), uint32_t(output[i * output_stride() + c]))
            << "pixel " << i << " / " << pixels() << ", channel " << c << " / " << channels();
        }
      }
    }
  }

  void TestCHW(xnn_f16_ibilinear_chw_ukernel_fn ibilinear) const {
    xnnpack::ReplicableRandomDevice rng;
    std::uniform_real_distribution<float> f32dist(0.1f, 1.0f);

    std::vector<const xnn_float16*> indirection(pixels() * 2);
    std::vector<xnn_float16> input(XNN_EXTRA_BYTES / sizeof(xnn_float16) + (channels() - 1) * input_stride() + 4 * pixels());
    std::vector<xnn_float16, AlignedAllocator<xnn_float16, 64>> packed_weights(pixels() * 2);
    std::vector<xnn_float16> output(pixels() * channels());
    std::vector<float> output_ref(pixels() * channels());

    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
      std::generate(packed_weights.begin(), packed_weights.end(), [&]() { return f32dist(rng); });
      std::fill(output.begin(), output.end(), std::nanf(""));

      // Indirection will point to the even ("left") pixels of the input.
      // The kernels will expect "right" pixels to be placed right next to them.
      for (size_t i = 0; i < indirection.size(); i++) {
        const xnn_float16* left_corner = input.data() + 2 * i - input_offset();
        indirection[i] = left_corner;
      }
      std::shuffle(indirection.begin(), indirection.end(), rng);

      // Compute reference results.
      for (size_t i = 0; i < pixels(); i++) {
        for (size_t c = 0; c < channels(); c++) {
          const float alpha_h = packed_weights[i * 2 + 0];
          const float alpha_v = packed_weights[i * 2 + 1];
          // `c * pixels() + i` because the output is NCHW.
          output_ref[c * pixels() + i] =
            // `c * indirection.size()` because the input is NCHW.
            (indirection[i * 2 + 0] + 0)[c * input_stride() + input_offset()] * (1.0f - alpha_h) * (1.0f - alpha_v) +
            (indirection[i * 2 + 0] + 1)[c * input_stride() + input_offset()] * alpha_h * (1.0f - alpha_v) +
            (indirection[i * 2 + 1] + 0)[c * input_stride() + input_offset()] * (1.0f - alpha_h) * alpha_v +
            (indirection[i * 2 + 1] + 1)[c * input_stride() + input_offset()] * alpha_h * alpha_v;
        }
      }

      // Call optimized micro-kernel.
      ibilinear(
        pixels(), channels(),
        reinterpret_cast<const xnn_float16**>(indirection.data()), input_offset() * sizeof(xnn_float16),
        packed_weights.data(), output.data(), input_stride() * sizeof(xnn_float16));

      // Verify results.
      for (size_t c = 0; c < channels(); c++) {
        for (size_t i = 0; i < pixels(); i++) {
          ASSERT_NEAR(
              output[c * pixels() + i],
              output_ref[c * pixels() + i],
              std::abs(output_ref[c * pixels() + i]) * 1.0e-2f)
            << "i = " << i << ", channel = " << c;
        }
      }
    }
  }

  void TestCHW(xnn_f32_ibilinear_chw_ukernel_fn ibilinear) const {
    xnnpack::ReplicableRandomDevice rng;
    std::uniform_real_distribution<float> f32dist;

    std::vector<const float*> indirection(pixels() * 2);
    std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + (channels() - 1) * input_stride() + 4 * pixels());
    std::vector<float, AlignedAllocator<float, 64>> packed_weights(pixels() * 2);
    std::vector<float> output(pixels() * channels());
    std::vector<float> output_ref(pixels() * channels());

    for (size_t iteration = 0; iteration < iterations(); iteration++) {
      std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
      std::generate(packed_weights.begin(), packed_weights.end(), [&]() { return f32dist(rng); });
      std::fill(output.begin(), output.end(), nanf(""));

      // Indirection will point to the even ("left") pixels of the input.
      // The kernels will expect "right" pixels to be placed right next to them.
      for (size_t i = 0; i < indirection.size(); i++) {
        const float* left_corner = input.data() + 2 * i - input_offset();
        indirection[i] = left_corner;
      }
      std::shuffle(indirection.begin(), indirection.end(), rng);

      // Compute reference results.
      for (size_t i = 0; i < pixels(); i++) {
        for (size_t c = 0; c < channels(); c++) {
          const float alpha_h = packed_weights[i * 2 + 0];
          const float alpha_v = packed_weights[i * 2 + 1];
          // `c * pixels() + i` because the output is NCHW.
          output_ref[c * pixels() + i] =
            // `c * indirection.size()` because the input is NCHW.
            (indirection[i * 2 + 0] + 0)[c * input_stride() + input_offset()] * (1.0f - alpha_h) * (1.0f - alpha_v) +
            (indirection[i * 2 + 0] + 1)[c * input_stride() + input_offset()] * alpha_h * (1.0f - alpha_v) +
            (indirection[i * 2 + 1] + 0)[c * input_stride() + input_offset()] * (1.0f - alpha_h) * alpha_v +
            (indirection[i * 2 + 1] + 1)[c * input_stride() + input_offset()] * alpha_h * alpha_v;
        }
      }

      // Call optimized micro-kernel.
      ibilinear(
        pixels(), channels(),
        indirection.data(), input_offset() * sizeof(float),
        packed_weights.data(), output.data(), input_stride() * sizeof(float));

      // Verify results.
      for (size_t c = 0; c < channels(); c++) {
        for (size_t i = 0; i < pixels(); i++) {
          EXPECT_NEAR(
              output_ref[c * pixels() + i],
              output[c * pixels() + i],
              std::abs(output_ref[c * pixels() + i]) * 1.0e-4)
            << "i = " << i << ", channel = " << c;
        }
      }
    }
  }

 private:
  uint32_t channels_{1};
  uint32_t pixels_{1};
  uint32_t output_stride_{0};
  uint32_t input_stride_{0};
  uint32_t input_offset_{0};
  size_t iterations_{3};
};