delegate/test/Convolution2dTest.cpp

//
// Copyright © 2020, 2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include "ConvolutionTestHelper.hpp"

#include <armnn_delegate.hpp>

#include <flatbuffers/flatbuffers.h>
#include <tensorflow/lite/interpreter.h>
#include <tensorflow/lite/kernels/register.h>
#include <tensorflow/lite/model.h>
#include <tensorflow/lite/version.h>

#include <doctest/doctest.h>

namespace armnnDelegate
{

void Conv2DWithBiasesFp32Test()
{
    // Set input data
    std::vector<int32_t> inputShape { 1, 5, 5, 1 };
    std::vector<int32_t> filterShape { 1, 3, 3, 1 };
    std::vector<int32_t> biasShape { 1 };
    std::vector<int32_t> outputShape { 1, 3, 3, 1 };

    static std::vector<float> inputValues =
        {
            1, 5, 2, 3, 5,
            8, 7, 3, 6, 3,
            3, 3, 9, 1, 9,
            4, 1, 8, 1, 3,
            6, 8, 1, 9, 2
        };

    std::vector<float> filterValues =
        {
            4, 5, 6,
            0, 0, 0,
            3, 2, 1
        };

    std::vector<float> biasValues = { 0 };

    std::vector<float> expectedOutputValues =
        {
            23, 33, 24,
            91, 99, 48,
            26, 50, 19
        };

    tflite::Padding padding = tflite::Padding_SAME;

    ConvolutionTest<float>(tflite::BuiltinOperator_CONV_2D,
                           ::tflite::TensorType_FLOAT32,
                           2, // strideX
                           2, // strideY
                           1, // dilationX
                           1, // dilationY
                           padding,
                           tflite::ActivationFunctionType_NONE,
                           inputShape,
                           filterShape,
                           outputShape,
                           inputValues,
                           filterValues,
                           expectedOutputValues,
                           biasShape,
                           biasValues);
}

void Conv2DWithBiasesInt8Test()
{
    // Set input data
    std::vector<int32_t> inputShape { 1, 2, 2, 1 };
    std::vector<int32_t> filterShape { 1, 2, 2, 1 };
    std::vector<int32_t> biasShape { 1 };
    std::vector<int32_t> outputShape { 1, 2, 2, 1 };

    static std::vector<int8_t> inputValues = { 1, 2, 3, 4 };

    std::vector<int8_t> filterValues = { 2, 1, 0, 6 };

    std::vector<int32_t> biasValues = { 10 };

    std::vector<int8_t> expectedOutputValues =
        {
            (1 * 2 + 2 * 1 + 3 * 0 + 4 * 6 + 10) / 2, // 19
            (2 * 2 + 0 * 1 + 4 * 0 + 0 * 6 + 10) / 2, // 7
            (3 * 2 + 4 * 1 + 0 * 0 + 0 * 6 + 10) / 2, // 10
            (4 * 2 + 0 * 1 + 0 * 0 + 0 * 6 + 10) / 2,  // 9
        };

    tflite::Padding padding = tflite::Padding_SAME;

    ConvolutionTest<int8_t, int32_t>(tflite::BuiltinOperator_CONV_2D,
                                     ::tflite::TensorType_INT8,
                                     1, // strideX
                                     1, // strideY
                                     1, // dilationX
                                     1, // dilationY
                                     padding,
                                     tflite::ActivationFunctionType_NONE,
                                     inputShape,
                                     filterShape,
                                     outputShape,
                                     inputValues,
                                     filterValues,
                                     expectedOutputValues,
                                     biasShape,
                                     biasValues);
}

void Conv2DWithBiasesReluUint8Test()
{
    // Set input data
    std::vector<int32_t> inputShape { 1, 2, 2, 1 };
    std::vector<int32_t> filterShape { 1, 2, 2, 1 };
    std::vector<int32_t> biasShape { 1 };
    std::vector<int32_t> outputShape { 1, 2, 2, 1 };

    static std::vector<uint8_t> inputValues = { 1, 2, 4, 8 };

    std::vector<uint8_t> filterValues = { 2, 1, 0, 6 };

    std::vector<int32_t> biasValues = { 16 };

    // factors to consider:
    // - the filter zero point is non zero, hence the (x-fz)
    // - the output scale is 2 hence the /2
    // - output zero point is non zero, hence the +outZero
    // - RELU cuts negative values and then we add the output zero point
    uint8_t bias = 16;
    uint8_t outZero = 20;
    uint8_t fz = 4; // filter zero point

    std::vector<uint8_t> expectedOutputValues =
        {
            std::max(outZero, static_cast<uint8_t>((1*(2-fz) + 2*(1-fz) + 4*(0-fz) + 8*(6-fz) + bias)/2 + outZero)),
            std::max(outZero, static_cast<uint8_t>((2*(2-fz) + 0*(1-fz) + 8*(0-fz) + 0*(6-fz) + bias)/2 + outZero)),
            std::max(outZero, static_cast<uint8_t>((4*(2-fz) + 8*(1-fz) + 0*(0-fz) + 0*(6-fz) + bias)/2 + outZero)),
            std::max(outZero, static_cast<uint8_t>((8*(2-fz) + 0*(1-fz) + 0*(0-fz) + 0*(6-fz) + bias)/2 + outZero))
        };

    tflite::Padding padding = tflite::Padding_SAME;

    ConvolutionTest<uint8_t, int32_t>(tflite::BuiltinOperator_CONV_2D,
                                      ::tflite::TensorType_UINT8,
                                     1, // strideX
                                     1, // strideY
                                     1, // dilationX
                                     1, // dilationY
                                     padding,
                                     tflite::ActivationFunctionType_RELU,
                                     inputShape,
                                     filterShape,
                                     outputShape,
                                     inputValues,
                                     filterValues,
                                     expectedOutputValues,
                                     biasShape,
                                     biasValues,
                                     {1.0f}, // biasScale
                                     {0},    // biasOffset
                                     {1.0f}, // filterScale
                                     {4},    // filterOffsets
                                     2, // output scale
                                     20); // output offset
}

void Conv2DWithBiasesRelu6Uint8Test()
{
    // Set input data
    std::vector<int32_t> inputShape { 1, 2, 2, 1 };
    std::vector<int32_t> filterShape { 1, 2, 2, 1 };
    std::vector<int32_t> biasShape { 1 };
    std::vector<int32_t> outputShape { 1, 2, 2, 1 };

    static std::vector<uint8_t> inputValues = { 1, 2, 4, 1 };

    std::vector<uint8_t> filterValues = { 2, 1, 0, 6 };

    std::vector<int32_t> biasValues = { 0 };

    // factors to consider:
    // - the output scale is 2 hence the /2
    // - RELU6 cuts output values at +6
    uint8_t relu6Min = 6 / 2; // divide by output scale

    std::vector<uint8_t> expectedOutputValues =
        {
            std::min(relu6Min, static_cast<uint8_t>((1 * 2 + 2 * 1 + 4 * 0 + 1 * 6) / 2)),
            std::min(relu6Min, static_cast<uint8_t>((2 * 2 + 0 * 1 + 1 * 0 + 0 * 6) / 2)),
            std::min(relu6Min, static_cast<uint8_t>((4 * 2 + 1 * 1 + 0 * 0 + 0 * 6) / 2)),
            std::min(relu6Min, static_cast<uint8_t>((1 * 2 + 0 * 1 + 0 * 0 + 0 * 6) / 2))
        };

    tflite::Padding padding = tflite::Padding_SAME;

    ConvolutionTest<uint8_t, int32_t>(tflite::BuiltinOperator_CONV_2D,
                                     ::tflite::TensorType_UINT8,
                                     1, // strideX
                                     1, // strideY
                                     1, // dilationX
                                     1, // dilationY
                                     padding,
                                     tflite::ActivationFunctionType_RELU6,
                                     inputShape,
                                     filterShape,
                                     outputShape,
                                     inputValues,
                                     filterValues,
                                     expectedOutputValues,
                                     biasShape,
                                     biasValues);
}


void Conv2DPerChannelInt8Test()
{
    // Set input data
    std::vector<int32_t> inputShape  { 1,4,4,2 };
    std::vector<int32_t> filterShape { 4,2,2,2 };
    std::vector<int32_t> biasShape   { 4 };
    std::vector<int32_t> outputShape { 1,4,4,4 };

    static std::vector<int8_t> inputValues =
        {
            -11, 40,-26, 11,-28,  8,  0, -8,
            -10, 34, 47,  0,-33,-14, 28, 35,
              6,-28,-26,  8, 13, 33,-31,-41,
             31,-20,-31,-16,  8,-18,-44,  0
        };

    std::vector<float>  filterScales = { 1.858268, 2.0, 1.992126, 1.905512 };
    int32_t filterQuantizationDim    = 0;
    std::vector<int8_t> filterValues =
        {
             13,-44,  5,-14, 21,-45, 36,-25,
            -42, -2, 24,-30,-31, 35, 43,-30,
            -20, -5, 25, 17, 18, 20,  4,-46,
            -49,  9, -3,-20, 46,  5,  7,-15
        };

    std::vector<int32_t> biasValues = { 0,0,0,0 };
    std::vector<float>   biasScales = { 0.721445, 0.7764700055, 0.773414, 0.739787 };

    std::vector<int8_t> expectedOutputValues =
        {
               -1,  9,  3, 5, 1, -1,  5,  9,
                2,  7, -1, 2, 2,  4,  5,  6,
                1,  1,  4, 4, 2,  0, -4, -3,
                0,  6, 12, 6, 3,  0, -1, -2,
                7, -4,  4, 4, 3,  6,  6,  2,
                0, -3, -1, 4, 4,  8,  3,  1,
                5,  0,  0, 1, 4,  7,  4,  6,
                4,  0,  1, 2, 2,  7,  5,  7
        };
    float outputQuantScale  = 401.960785f;
    int   outputQuantOffset = 3;
    float inputQuantScale   = 0.388235f;
    int   inputQuantOffset  = 1;

    tflite::Padding padding = tflite::Padding_SAME;

    ConvolutionTest<int8_t, int32_t>(tflite::BuiltinOperator_CONV_2D,
                                     ::tflite::TensorType_INT8,
                                     1, // strideX
                                     1, // strideY
                                     1, // dilationX
                                     1, // dilationY
                                     padding,
                                     tflite::ActivationFunctionType_NONE,
                                     inputShape,
                                     filterShape,
                                     outputShape,
                                     inputValues,
                                     filterValues,
                                     expectedOutputValues,
                                     biasShape,
                                     biasValues,
                                     biasScales,
                                     {0,0,0,0},
                                     filterScales,
                                     {0,0,0,0},
                                     outputQuantScale,
                                     outputQuantOffset,
                                     inputQuantScale,
                                     inputQuantOffset,
                                     1, // depth_multiplier is ignored for conv2d value doesn't matter
                                     filterQuantizationDim);
}

TEST_SUITE("Convolution2dTest_Tests")
{

TEST_CASE ("Conv2DWithBiases_Fp32_Test")
{
    Conv2DWithBiasesFp32Test();
}

TEST_CASE ("Conv2DWithBiases_Int8_Test")
{
    Conv2DWithBiasesInt8Test();
}

TEST_CASE ("Conv2DPerChannel_Int8_Test")
{
    Conv2DPerChannelInt8Test();
}

} //End of TEST_SUITE("Convolution2dTest_Tests")

} // namespace armnnDelegate