Function choose_quantization_params

examples/neon_gemm_qasymm8.cpp:52–85

Return reasonable quantisation parameters to use for an array of floats based on min and max values

// Return reasonable quantisation parameters to use for an array of floats
// based on min and max values
QuantizationInfo choose_quantization_params(float min, float max)
{
    // Extend the [min,max] interval to contain 0 so we can represent it exactly
    min = std::min(min, 0.f);
    max = std::max(max, 0.f);

    // Set the quantized min and max in float values
    const float qmin = 0;
    const float qmax = 255;

    // Determine the scale
    const float scale = (max - min) / (qmax - qmin);

    // Determine the zero-point; using affine equation val = (qval-zerop) * scale
    const float zero_point_real = qmin - min / scale;

    // But we need to nudge the zero_point to an integer (exact quantized value)
    std::uint8_t zero_point_nudged = 0;
    if (zero_point_real < qmin)
    {
        zero_point_nudged = qmin;
    }
    else if (zero_point_real > qmax)
    {
        zero_point_nudged = qmax;
    }
    else
    {
        zero_point_nudged = static_cast<std::uint8_t>(support::cpp11::round(zero_point_real));
    }

    QuantizationInfo qinfo = QuantizationInfo(scale, zero_point_nudged);
    return qinfo;
}

void quantize_values(int size, qasymm8_t *output, float *input, const QuantizationInfo qinfo)
{