float.h

Go to the documentation of this file.

// SPDX-License-Identifier: Apache-2.0
// Copyright Contributors to the OpenQMC Project.
 
 
#pragma once
 
#include "gpu.h"
 
#include <cmath>
#include <cstdint>
 
 
namespace oqmc
{
 
constexpr auto floatOneOverUintMax = 2.3283064365386962890625e-10f; 
constexpr auto floatOneMinusEpsilon = 0.999999940395355224609375f;  
 
OQMC_HOST_DEVICE inline float uintToFloat(std::uint32_t value)
{
    // This method is inspired by the method used by Matt Pharr in PBRT v4. It
    // has the undesirable property of floating point values rounding up to the
    // nearest representable number to reduce error. This gives the potential
    // for an output equal to one, requiring a min operation. However, this was
    // considered the best tradeoff when compared to other methods.
 
    // Marc Reynolds gives an option where 8 bits of precision is first removed
    // (http://marc-b-reynolds.github.io/distribution/2017/01/17/DenseFloat.html)
    // prior to division. This removes the need for any min operations, but also
    // looses us a good amount of precision sub one half.
 
    // An alternative algorithm using a clz operation is detailed in Section 2.1
    // of 'Quasi-Monte Carlo Algorithms (not only) for Graphics Software'. This
    // optimally maps the unsinged integer bits to the floating point exponent
    // and mantissa. However, after benchmarking this alternative method the
    // cost when drawing samples was a 2x time increase for some samplers.
 
    return std::fmin(value * floatOneOverUintMax, floatOneMinusEpsilon);
}
 
} // namespace oqmc