#ifndef _TCUFLOAT_HPP #define _TCUFLOAT_HPP /*------------------------------------------------------------------------- * drawElements Quality Program Tester Core * ---------------------------------------- * * Copyright 2014 The Android Open Source Project * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. * *//*! * \file * \brief Reconfigurable floating-point value template. *//*--------------------------------------------------------------------*/ #include "tcuDefs.hpp" // For memcpy(). #include namespace tcu { enum FloatFlags { FLOAT_HAS_SIGN = (1<<0), FLOAT_SUPPORT_DENORM = (1<<1) }; /*--------------------------------------------------------------------*//*! * \brief Floating-point format template * * This template implements arbitrary floating-point handling. Template * can be used for conversion between different formats and checking * various properties of floating-point values. *//*--------------------------------------------------------------------*/ template class Float { public: typedef StorageType_ StorageType; enum { EXPONENT_BITS = ExponentBits, MANTISSA_BITS = MantissaBits, EXPONENT_BIAS = ExponentBias, FLAGS = Flags, }; Float (void); explicit Float (StorageType value); explicit Float (float v); explicit Float (double v); template static Float convert (const Float& src); static inline Float convert (const Float& src) { return src; } /*--------------------------------------------------------------------*//*! * \brief Construct floating point value * \param sign Sign. Must be +1/-1 * \param exponent Exponent in range [1-ExponentBias, ExponentBias+1] * \param mantissa Mantissa bits with implicit leading bit explicitly set * \return The specified float * * This function constructs a floating point value from its inputs. * The normally implicit leading bit of the mantissa must be explicitly set. * The exponent normally used for zero/subnormals is an invalid input. Such * values are specified with the leading mantissa bit of zero and the lowest * normal exponent (1-ExponentBias). Additionally having both exponent and * mantissa set to zero is a shorthand notation for the correctly signed * floating point zero. Inf and NaN must be specified directly with an * exponent of ExponentBias+1 and the appropriate mantissa (with leading * bit set) *//*--------------------------------------------------------------------*/ static inline Float construct (int sign, int exponent, StorageType mantissa); /*--------------------------------------------------------------------*//*! * \brief Construct floating point value. Explicit version * \param sign Sign. Must be +1/-1 * \param exponent Exponent in range [-ExponentBias, ExponentBias+1] * \param mantissa Mantissa bits * \return The specified float * * This function constructs a floating point value from its inputs with * minimal intervention. * The sign is turned into a sign bit and the exponent bias is added. * See IEEE-754 for additional information on the inputs and * the encoding of special values. *//*--------------------------------------------------------------------*/ static Float constructBits (int sign, int exponent, StorageType mantissaBits); StorageType bits (void) const { return m_value; } float asFloat (void) const; double asDouble (void) const; inline int signBit (void) const { return (m_value >> (ExponentBits+MantissaBits)) & 1; } inline StorageType exponentBits (void) const { return (m_value >> MantissaBits) & ((StorageType(1)< Float16; //!< IEEE 754-2008 16-bit floating-point value typedef Float Float32; //!< IEEE 754 32-bit floating-point value typedef Float Float64; //!< IEEE 754 64-bit floating-point value template inline Float::Float (void) : m_value(0) { } template inline Float::Float (StorageType value) : m_value(value) { } template inline Float::Float (float value) : m_value(0) { deUint32 u32; memcpy(&u32, &value, sizeof(deUint32)); *this = convert(Float32(u32)); } template inline Float::Float (double value) : m_value(0) { deUint64 u64; memcpy(&u64, &value, sizeof(deUint64)); *this = convert(Float64(u64)); } template inline float Float::asFloat (void) const { float v; deUint32 u32 = Float32::convert(*this).bits(); memcpy(&v, &u32, sizeof(deUint32)); return v; } template inline double Float::asDouble (void) const { double v; deUint64 u64 = Float64::convert(*this).bits(); memcpy(&v, &u64, sizeof(deUint64)); return v; } template inline Float Float::zero (int sign) { DE_ASSERT(sign == 1 || ((Flags & FLOAT_HAS_SIGN) && sign == -1)); return Float(StorageType((sign > 0 ? 0ull : 1ull) << (ExponentBits+MantissaBits))); } template inline Float Float::inf (int sign) { DE_ASSERT(sign == 1 || ((Flags & FLOAT_HAS_SIGN) && sign == -1)); return Float(StorageType(((sign > 0 ? 0ull : 1ull) << (ExponentBits+MantissaBits)) | (((1ull< inline Float Float::nan (void) { return Float(StorageType((1ull<<(ExponentBits+MantissaBits))-1)); } template Float Float::construct (int sign, int exponent, StorageType mantissa) { // Repurpose this otherwise invalid input as a shorthand notation for zero (no need for caller to care about internal representation) const bool isShorthandZero = exponent == 0 && mantissa == 0; // Handles the typical notation for zero (min exponent, mantissa 0). Note that the exponent usually used exponent (-ExponentBias) for zero/subnormals is not used. // Instead zero/subnormals have the (normally implicit) leading mantissa bit set to zero. const bool isDenormOrZero = (exponent == 1 - ExponentBias) && (mantissa >> MantissaBits == 0); const StorageType s = StorageType(sign < 0 ? 1 : 0) << (ExponentBits+MantissaBits); const StorageType exp = (isShorthandZero || isDenormOrZero) ? StorageType(0) : StorageType(exponent + ExponentBias); DE_ASSERT(sign == +1 || sign == -1); DE_ASSERT(isShorthandZero || isDenormOrZero || mantissa >> MantissaBits == 1); DE_ASSERT(exp >> ExponentBits == 0); return Float(StorageType(s | (exp << MantissaBits) | (mantissa & ((StorageType(1)< Float Float::constructBits (int sign, int exponent, StorageType mantissaBits) { const StorageType signBit = sign < 0 ? 1 : 0; const StorageType exponentBits = exponent + ExponentBias; DE_ASSERT(sign == +1 || sign == -1 ); DE_ASSERT(exponentBits >> ExponentBits == 0); DE_ASSERT(mantissaBits >> MantissaBits == 0); return Float(StorageType((signBit << (ExponentBits+MantissaBits)) | (exponentBits << MantissaBits) | (mantissaBits))); } template template Float Float::convert (const Float& other) { if (!(Flags & FLOAT_HAS_SIGN) && other.sign() < 0) { // Negative number, truncate to zero. return zero(+1); } else if (other.isInf()) { return inf(other.sign()); } else if (other.isNaN()) { return nan(); } else if (other.isZero()) { return zero(other.sign()); } else { const int eMin = 1 - ExponentBias; const int eMax = ((1<> bitDiff) & 1; return Float(StorageType(s | (m + half + bias) >> bitDiff)); } else return zero(other.sign()); } else { // Remove leading 1. m = m & ~(1ull<> bitDiff) & 1; m = (m + half + bias) >> bitDiff; if (m & (1ull< eMax) { // Overflow. return inf(other.sign()); } else { DE_ASSERT(de::inRange(e, eMin, eMax)); DE_ASSERT(((e + ExponentBias) & ~((1ull<