Method computeScaleADMM

tools/quantization/TensorStatistic.cpp:236–277

}

std::pair<float, int8_t> TensorStatistic::computeScaleADMM() {
    const int count         = mOriginTensor->elementSize();
    float max               = 0;
    const float bound       = mFeatureClampValue;
    const float* originData = mOriginTensor->host<float>();

    for (int i = 0; i < count; i++) {
        float absData = std::fabs(originData[i]);
        if (absData > max) {
            max = absData;
        }
    }
    float alpha = max / (bound * 2.5);

    // DLOG(INFO) << "alpha init: " << alpha;

    const int maxStep = 300;
    float sum1        = 0;
    float sum2        = 0;
    float invAlpha;

    for (int i = 0; i < maxStep; i++) {
        sum1     = 0;
        sum2     = 0;
        invAlpha = 1 / alpha;

        for (int i = 0; i < count; i++) {
            auto origin    = originData[i];
            auto dataQuant = std::roundf(origin * invAlpha);
            dataQuant      = std::fmin(bound, std::fmax(-bound, dataQuant));
            sum1 += (dataQuant * origin);
            sum2 += (dataQuant * dataQuant);
        }

        alpha = sum1 / sum2;
    }
    // DLOG(INFO) << "alpha final: " << alpha;
    mScale = alpha;
    mVisited = true;
    mZeroPoint = 0;
    return std::make_pair(mScale, mZeroPoint);
}

std::pair<std::vector<float>, float> TensorStatistic::fakeQuantFeature() {
    const int count         = mOriginTensor->elementSize();