EthosnRefConvImpl.cpp

Go to the documentation of this file.

//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include "EthosnRefConvImpl.hpp"

#include <armnn/utility/Assert.hpp>

#include <cmath>
#include <limits>

namespace armnn
{

EthosnRefQuantizedMultiplierSmallerThanOne::EthosnRefQuantizedMultiplierSmallerThanOne(float multiplier)
{
    ARMNN_ASSERT(multiplier >= 0.0f && multiplier < 1.0f);
    if (multiplier == 0.0f)
    {
        m_Multiplier = 0;
        m_RightShift = 0;
    }
    else
    {
        const double q = std::frexp(multiplier, &m_RightShift);
        m_RightShift = -m_RightShift;
        int64_t qFixed = static_cast<int64_t>(std::round(q * (1ll << 16)));
        ARMNN_ASSERT(qFixed <= (1ll << 16));
        if (qFixed == (1ll << 16))
        {
            qFixed /= 2;
            --m_RightShift;
        }
        ARMNN_ASSERT(m_RightShift >= 0);
        ARMNN_ASSERT(qFixed <= std::numeric_limits<int32_t>::max());
        m_Multiplier = static_cast<int32_t>(qFixed);
    }
}

#define SATURATE_VAL_S_32S(WIDTH, VAL)                                                                                 \
    (static_cast<int32_t>(VAL > ((static_cast<int64_t>(1) << (WIDTH - 1)) - 1)                                         \
                              ? ((static_cast<int64_t>(1) << (WIDTH - 1)) - 1)                                         \
                              : VAL < (-1 * (static_cast<int64_t>(1) << (WIDTH - 1)))                                  \
                                    ? (-1 * (static_cast<int64_t>(1) << (WIDTH - 1)))                                  \
                                    : VAL))

int32_t EthosnRefQuantizedMultiplierSmallerThanOne::operator*(int32_t rhs) const
{
    // Align with the implementation in NPU model
    // See function WinogradFinalTransform::postscale_uint8 in winograd_final_transform.cpp
    int64_t a_64(rhs);
    int64_t b_64(m_Multiplier);
    int64_t ab_64 = a_64 * b_64;

    ab_64                 = ab_64 >> 15;
    int32_t roundBit      = (ab_64 >> m_RightShift) & 0x1;
    int64_t resultShifted = (ab_64 >> (m_RightShift + 1)) + roundBit;

    return SATURATE_VAL_S_32S(9, resultShifted);
}

int64_t EthosnRefQuantizedMultiplierSmallerThanOne::operator*(int64_t rhs) const
{
    // Align with the implementation in NPU model
    // See function WinogradFinalTransform::postscale_uint8 in winograd_final_transform.cpp
    int64_t a_64(rhs);
    int64_t b_64(m_Multiplier);
    int64_t ab_64 = a_64 * b_64;

    ab_64                 = ab_64 >> 15;
    int32_t roundBit      = (ab_64 >> m_RightShift) & 0x1;
    int64_t resultShifted = (ab_64 >> (m_RightShift + 1)) + roundBit;

    return SATURATE_VAL_S_32S(16, resultShifted);
}

} //namespace armnn