external/vulkancts/modules/vulkan/shaderexecutor/vktShaderBuiltinPrecisionTests.cpp

/*------------------------------------------------------------------------
 * Vulkan Conformance Tests
 * ------------------------
 *
 * Copyright (c) 2018 The Khronos Group Inc.
 * Copyright (c) 2015 Samsung Electronics Co., Ltd.
 * Copyright (c) 2016 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 Precision and range tests for builtins and types.
 *
 *//*--------------------------------------------------------------------*/

#include "vktShaderBuiltinPrecisionTests.hpp"
#include "vktShaderExecutor.hpp"
#include "amber/vktAmberTestCase.hpp"

#include "deMath.h"
#include "deMemory.h"
#include "deFloat16.h"
#include "deDefs.hpp"
#include "deRandom.hpp"
#include "deSTLUtil.hpp"
#include "deStringUtil.hpp"
#include "deUniquePtr.hpp"
#include "deSharedPtr.hpp"
#include "deArrayUtil.hpp"

#include "tcuCommandLine.hpp"
#include "tcuFloatFormat.hpp"
#include "tcuInterval.hpp"
#include "tcuTestLog.hpp"
#include "tcuVector.hpp"
#include "tcuMatrix.hpp"
#include "tcuResultCollector.hpp"
#include "tcuMaybe.hpp"

#include "gluContextInfo.hpp"
#include "gluVarType.hpp"
#include "gluRenderContext.hpp"
#include "glwDefs.hpp"

#include <cmath>
#include <string>
#include <sstream>
#include <iostream>
#include <map>
#include <utility>
#include <limits>

// Uncomment this to get evaluation trace dumps to std::cerr
// #define GLS_ENABLE_TRACE

// set this to true to dump even passing results
#define GLS_LOG_ALL_RESULTS false

#define FLOAT16_1_0		0x3C00 //1.0 float16bit
#define FLOAT16_180_0	0x59A0 //180.0 float16bit
#define FLOAT16_2_0		0x4000 //2.0 float16bit
#define FLOAT16_3_0		0x4200 //3.0 float16bit
#define FLOAT16_0_5		0x3800 //0.5 float16bit
#define FLOAT16_0_0		0x0000 //0.0 float16bit


using tcu::Vector;
typedef Vector<deFloat16, 1>	Vec1_16Bit;
typedef Vector<deFloat16, 2>	Vec2_16Bit;
typedef Vector<deFloat16, 3>	Vec3_16Bit;
typedef Vector<deFloat16, 4>	Vec4_16Bit;

typedef Vector<double, 1>		Vec1_64Bit;
typedef Vector<double, 2>		Vec2_64Bit;
typedef Vector<double, 3>		Vec3_64Bit;
typedef Vector<double, 4>		Vec4_64Bit;

enum
{
	// Computing reference intervals can take a non-trivial amount of time, especially on
	// platforms where toggling floating-point rounding mode is slow (emulated arm on x86).
	// As a workaround watchdog is kept happy by touching it periodically during reference
	// interval computation.
	TOUCH_WATCHDOG_VALUE_FREQUENCY	= 512
};

namespace vkt
{
namespace shaderexecutor
{

using std::string;
using std::map;
using std::ostream;
using std::ostringstream;
using std::pair;
using std::vector;
using std::set;

using de::MovePtr;
using de::Random;
using de::SharedPtr;
using de::UniquePtr;
using tcu::Interval;
using tcu::FloatFormat;
using tcu::MessageBuilder;
using tcu::TestLog;
using tcu::Vector;
using tcu::Matrix;
using glu::Precision;
using glu::VarType;
using glu::DataType;
using glu::ShaderType;

enum PrecisionTestFeatureBits
{
	PRECISION_TEST_FEATURES_NONE									= 0u,
	PRECISION_TEST_FEATURES_16BIT_BUFFER_ACCESS						= (1u << 1),
	PRECISION_TEST_FEATURES_16BIT_UNIFORM_AND_STORAGE_BUFFER_ACCESS	= (1u << 2),
	PRECISION_TEST_FEATURES_16BIT_PUSH_CONSTANT						= (1u << 3),
	PRECISION_TEST_FEATURES_16BIT_INPUT_OUTPUT						= (1u << 4),
	PRECISION_TEST_FEATURES_16BIT_SHADER_FLOAT						= (1u << 5),
	PRECISION_TEST_FEATURES_64BIT_SHADER_FLOAT						= (1u << 6),
};
typedef deUint32 PrecisionTestFeatures;


void areFeaturesSupported (const Context& context, deUint32 toCheck)
{
	if (toCheck == PRECISION_TEST_FEATURES_NONE) return;

	const vk::VkPhysicalDevice16BitStorageFeatures& extensionFeatures = context.get16BitStorageFeatures();

	if ((toCheck & PRECISION_TEST_FEATURES_16BIT_BUFFER_ACCESS) != 0 && extensionFeatures.storageBuffer16BitAccess == VK_FALSE)
		TCU_THROW(NotSupportedError, "Requested 16bit storage features not supported");

	if ((toCheck & PRECISION_TEST_FEATURES_16BIT_UNIFORM_AND_STORAGE_BUFFER_ACCESS) != 0 && extensionFeatures.uniformAndStorageBuffer16BitAccess == VK_FALSE)
		TCU_THROW(NotSupportedError, "Requested 16bit storage features not supported");

	if ((toCheck & PRECISION_TEST_FEATURES_16BIT_PUSH_CONSTANT) != 0 && extensionFeatures.storagePushConstant16 == VK_FALSE)
		TCU_THROW(NotSupportedError, "Requested 16bit storage features not supported");

	if ((toCheck & PRECISION_TEST_FEATURES_16BIT_INPUT_OUTPUT) != 0 && extensionFeatures.storageInputOutput16 == VK_FALSE)
		TCU_THROW(NotSupportedError, "Requested 16bit storage features not supported");

	if ((toCheck & PRECISION_TEST_FEATURES_16BIT_SHADER_FLOAT) != 0 && context.getShaderFloat16Int8Features().shaderFloat16 == VK_FALSE)
		TCU_THROW(NotSupportedError, "Requested 16-bit floats (halfs) are not supported in shader code");

	if ((toCheck & PRECISION_TEST_FEATURES_64BIT_SHADER_FLOAT) != 0 && context.getDeviceFeatures().shaderFloat64 == VK_FALSE)
		TCU_THROW(NotSupportedError, "Requested 64-bit floats are not supported in shader code");
}

/*--------------------------------------------------------------------*//*!
 * \brief Generic singleton creator.
 *
 * instance<T>() returns a reference to a unique default-constructed instance
 * of T. This is mainly used for our GLSL function implementations: each
 * function is implemented by an object, and each of the objects has a
 * distinct class. It would be extremely toilsome to maintain a separate
 * context object that contained individual instances of the function classes,
 * so we have to resort to global singleton instances.
 *
 *//*--------------------------------------------------------------------*/
template <typename T>
const T& instance (void)
{
	static const T s_instance = T();
	return s_instance;
}

/*--------------------------------------------------------------------*//*!
 * \brief Empty placeholder type for unused template parameters.
 *
 * In the precision tests we are dealing with functions of different arities.
 * To minimize code duplication, we only define templates with the maximum
 * number of arguments, currently four. If a function's arity is less than the
 * maximum, Void us used as the type for unused arguments.
 *
 * Although Voids are not used at run-time, they still must be compilable, so
 * they must support all operations that other types do.
 *
 *//*--------------------------------------------------------------------*/
struct Void
{
	typedef	Void		Element;
	enum
	{
		SIZE = 0,
	};

	template <typename T>
	explicit			Void			(const T&)		{}
						Void			(void)			{}
						operator double	(void)	const	{ return TCU_NAN; }

	// These are used to make Voids usable as containers in container-generic code.
	Void&				operator[]		(int)			{ return *this; }
	const Void&			operator[]		(int)	const	{ return *this; }
};

ostream& operator<< (ostream& os, Void) { return os << "()"; }

//! Returns true for all other types except Void
template <typename T>	bool isTypeValid		(void)	{ return true;	}
template <>				bool isTypeValid<Void>	(void)	{ return false;	}

template <typename T>	bool isInteger				(void)	{ return false;	}
template <>				bool isInteger<int>			(void)	{ return true;	}
template <>				bool isInteger<tcu::IVec2>	(void)	{ return true;	}
template <>				bool isInteger<tcu::IVec3>	(void)	{ return true; }
template <>				bool isInteger<tcu::IVec4>	(void)	{ return true; }

//! Utility function for getting the name of a data type.
//! This is used in vector and matrix constructors.
template <typename T>
const char* dataTypeNameOf (void)
{
	return glu::getDataTypeName(glu::dataTypeOf<T>());
}

template <>
const char* dataTypeNameOf<Void> (void)
{
	DE_FATAL("Impossible");
	return DE_NULL;
}

template <typename T>
VarType getVarTypeOf (Precision prec = glu::PRECISION_LAST)
{
	return glu::varTypeOf<T>(prec);
}

//! A hack to get Void support for VarType.
template <>
VarType getVarTypeOf<Void> (Precision)
{
	DE_FATAL("Impossible");
	return VarType();
}

/*--------------------------------------------------------------------*//*!
 * \brief Type traits for generalized interval types.
 *
 * We are trying to compute sets of acceptable values not only for
 * float-valued expressions but also for compound values: vectors and
 * matrices. We approximate a set of vectors as a vector of intervals and
 * likewise for matrices.
 *
 * We now need generalized operations for each type and its interval
 * approximation. These are given in the type Traits<T>.
 *
 * The type Traits<T>::IVal is the approximation of T: it is `Interval` for
 * scalar types, and a vector or matrix of intervals for container types.
 *
 * To allow template inference to take place, there are function wrappers for
 * the actual operations in Traits<T>. Hence we can just use:
 *
 * makeIVal(someFloat)
 *
 * instead of:
 *
 * Traits<float>::doMakeIVal(value)
 *
 *//*--------------------------------------------------------------------*/

template <typename T> struct Traits;

//! Create container from elementwise singleton values.
template <typename T>
typename Traits<T>::IVal makeIVal (const T& value)
{
	return Traits<T>::doMakeIVal(value);
}

//! Elementwise union of intervals.
template <typename T>
typename Traits<T>::IVal unionIVal (const typename Traits<T>::IVal& a,
									const typename Traits<T>::IVal& b)
{
	return Traits<T>::doUnion(a, b);
}

//! Returns true iff every element of `ival` contains the corresponding element of `value`.
template <typename T, typename U = Void>
bool contains (const typename Traits<T>::IVal& ival, const T& value, bool is16Bit = false, const tcu::Maybe<U>& modularDivisor = tcu::nothing<U>())
{
	return Traits<T>::doContains(ival, value, is16Bit, modularDivisor);
}

//! Print out an interval with the precision of `fmt`.
template <typename T>
void printIVal (const FloatFormat& fmt, const typename Traits<T>::IVal& ival, ostream& os)
{
	Traits<T>::doPrintIVal(fmt, ival, os);
}

template <typename T>
string intervalToString (const FloatFormat& fmt, const typename Traits<T>::IVal& ival)
{
	ostringstream oss;
	printIVal<T>(fmt, ival, oss);
	return oss.str();
}

//! Print out a value with the precision of `fmt`.
template <typename T>
void printValue16 (const FloatFormat& fmt, const T& value, ostream& os)
{
	Traits<T>::doPrintValue16(fmt, value, os);
}

template <typename T>
string value16ToString(const FloatFormat& fmt, const T& val)
{
	ostringstream oss;
	printValue16(fmt, val, oss);
	return oss.str();
}

const std::string getComparisonOperation(const int ndx)
{
	const int operationCount = 10;
	DE_ASSERT(de::inBounds(ndx, 0, operationCount));
	const std::string operations[operationCount] =
	{
		"OpFOrdEqual\t\t\t",
		"OpFOrdGreaterThan\t",
		"OpFOrdLessThan\t\t",
		"OpFOrdGreaterThanEqual",
		"OpFOrdLessThanEqual\t",
		"OpFUnordEqual\t\t",
		"OpFUnordGreaterThan\t",
		"OpFUnordLessThan\t",
		"OpFUnordGreaterThanEqual",
		"OpFUnordLessThanEqual"
	};
	return operations[ndx];
}

template <typename T>
string comparisonMessage(const T& val)
{
	DE_UNREF(val);
	return "";
}

template <>
string comparisonMessage(const int& val)
{
	ostringstream oss;

	int flags = val;
	for(int ndx = 0; ndx < 10; ++ndx)
	{
		oss << getComparisonOperation(ndx) << "\t:\t" << ((flags & 1) == 1 ? "TRUE" : "FALSE") << "\n";
		flags = flags >> 1;
	}
	return oss.str();
}

template <>
string comparisonMessage(const tcu::IVec2& val)
{
	ostringstream oss;
	tcu::IVec2 flags = val;
	for (int ndx = 0; ndx < 10; ++ndx)
	{
		oss << getComparisonOperation(ndx) << "\t:\t" << ((flags.x() & 1) == 1 ? "TRUE" : "FALSE") << "\t" << ((flags.y() & 1) == 1 ? "TRUE" : "FALSE") << "\n";
		flags.x() = flags.x() >> 1;
		flags.y() = flags.y() >> 1;
	}
	return oss.str();
}

template <>
string comparisonMessage(const tcu::IVec3& val)
{
	ostringstream oss;
	tcu::IVec3 flags = val;
	for (int ndx = 0; ndx < 10; ++ndx)
	{
		oss << getComparisonOperation(ndx) << "\t:\t" << ((flags.x() & 1) == 1 ? "TRUE" : "FALSE") << "\t"
								<< ((flags.y() & 1) == 1 ? "TRUE" : "FALSE") << "\t"
								<< ((flags.z() & 1) == 1 ? "TRUE" : "FALSE") << "\n";
		flags.x() = flags.x() >> 1;
		flags.y() = flags.y() >> 1;
		flags.z() = flags.z() >> 1;
	}
	return oss.str();
}

template <>
string comparisonMessage(const tcu::IVec4& val)
{
	ostringstream oss;
	tcu::IVec4 flags = val;
	for (int ndx = 0; ndx < 10; ++ndx)
	{
		oss << getComparisonOperation(ndx) << "\t:\t" << ((flags.x() & 1) == 1 ? "TRUE" : "FALSE") << "\t"
			<< ((flags.y() & 1) == 1 ? "TRUE" : "FALSE") << "\t"
			<< ((flags.z() & 1) == 1 ? "TRUE" : "FALSE") << "\t"
			<< ((flags.w() & 1) == 1 ? "TRUE" : "FALSE") << "\n";
		flags.x() = flags.x() >> 1;
		flags.y() = flags.y() >> 1;
		flags.z() = flags.z() >> 1;
		flags.w() = flags.z() >> 1;
	}
	return oss.str();
}
//! Print out a value with the precision of `fmt`.
template <typename T>
void printValue32 (const FloatFormat& fmt, const T& value, ostream& os)
{
	Traits<T>::doPrintValue32(fmt, value, os);
}

template <typename T>
string value32ToString (const FloatFormat& fmt, const T& val)
{
	ostringstream oss;
	printValue32(fmt, val, oss);
	return oss.str();
}

template <typename T>
void printValue64 (const FloatFormat& fmt, const T& value, ostream& os)
{
	Traits<T>::doPrintValue64(fmt, value, os);
}

template <typename T>
string value64ToString (const FloatFormat& fmt, const T& val)
{
	ostringstream oss;
	printValue64(fmt, val, oss);
	return oss.str();
}

//! Approximate `value` elementwise to the float precision defined in `fmt`.
//! The resulting interval might not be a singleton if rounding in both
//! directions is allowed.
template <typename T>
typename Traits<T>::IVal round (const FloatFormat& fmt, const T& value)
{
	return Traits<T>::doRound(fmt, value);
}

template <typename T>
typename Traits<T>::IVal convert (const FloatFormat&				fmt,
								  const typename Traits<T>::IVal&	value)
{
	return Traits<T>::doConvert(fmt, value);
}

// Matching input and output types. We may be in a modulo case and modularDivisor may have an actual value.
template <typename T>
bool intervalContains (const Interval& interval, T value, const tcu::Maybe<T>& modularDivisor)
{
	bool contained = interval.contains(value);

	if (!contained && modularDivisor)
	{
		const T divisor = modularDivisor.get();

		// In a modulo operation, if the calculated answer contains the divisor, allow exactly 0.0 as a replacement. Alternatively,
		// if the calculated answer contains 0.0, allow exactly the divisor as a replacement.
		if (interval.contains(static_cast<double>(divisor)))
			contained |= (value == 0.0);
		if (interval.contains(0.0))
			contained |= (value == divisor);
	}
	return contained;
}

// When the input and output types do not match, we are not in a real modulo operation. Do not take the divisor into account. This
// version is provided for syntactical compatibility only.
template <typename T, typename U>
bool intervalContains (const Interval& interval, T value, const tcu::Maybe<U>& modularDivisor)
{
	DE_UNREF(modularDivisor);		// For release builds.
	DE_ASSERT(!modularDivisor);
	return interval.contains(value);
}

//! Common traits for scalar types.
template <typename T>
struct ScalarTraits
{
	typedef				Interval		IVal;

	static Interval		doMakeIVal		(const T& value)
	{
		// Thankfully all scalar types have a well-defined conversion to `double`,
		// hence Interval can represent their ranges without problems.
		return Interval(double(value));
	}

	static Interval		doUnion			(const Interval& a, const Interval& b)
	{
		return a | b;
	}

	static bool			doContains		(const Interval& a, T value)
	{
		return a.contains(double(value));
	}

	static Interval		doConvert		(const FloatFormat& fmt, const IVal& ival)
	{
		return fmt.convert(ival);
	}

	static Interval		doConvert		(const FloatFormat& fmt, const IVal& ival, bool is16Bit)
	{
		DE_UNREF(is16Bit);
		return fmt.convert(ival);
	}

	static Interval		doRound			(const FloatFormat& fmt, T value)
	{
		return fmt.roundOut(double(value), false);
	}
};

template <>
struct ScalarTraits<deUint16>
{
	typedef				Interval		IVal;

	static Interval		doMakeIVal		(const deUint16& value)
	{
		// Thankfully all scalar types have a well-defined conversion to `double`,
		// hence Interval can represent their ranges without problems.
		return Interval(double(deFloat16To32(value)));
	}

	static Interval		doUnion			(const Interval& a, const Interval& b)
	{
		return a | b;
	}

	static Interval		doConvert		(const FloatFormat& fmt, const IVal& ival)
	{
		return fmt.convert(ival);
	}

	static Interval		doRound			(const FloatFormat& fmt, deUint16 value)
	{
		return fmt.roundOut(double(deFloat16To32(value)), false);
	}
};

template<>
struct Traits<float> : ScalarTraits<float>
{
	static void			doPrintIVal		(const FloatFormat&	fmt,
										 const Interval&	ival,
										 ostream&			os)
	{
		os << fmt.intervalToHex(ival);
	}

	static void			doPrintValue16	(const FloatFormat&	fmt,
										 const float&		value,
										 ostream&			os)
	{
		const deUint32 iRep = reinterpret_cast<const deUint32 & >(value);
		float res0 = deFloat16To32((deFloat16)(iRep & 0xFFFF));
		float res1 = deFloat16To32((deFloat16)(iRep >> 16));
		os << fmt.floatToHex(res0) << " " << fmt.floatToHex(res1);
	}

	static void			doPrintValue32	(const FloatFormat&	fmt,
										 const float&		value,
										 ostream&			os)
	{
		os << fmt.floatToHex(value);
	}

	static void			doPrintValue64	(const FloatFormat&	fmt,
										 const float&		value,
										 ostream&			os)
	{
		os << fmt.floatToHex(value);
	}

	template <typename U>
	static bool			doContains		(const Interval& a, const float& value, bool is16Bit, const tcu::Maybe<U>& modularDivisor)
	{
		if(is16Bit)
		{
			// Note: for deFloat16s packed in 32 bits, the original divisor is provided as a float to the shader in the input
			// buffer, so U is also float here and we call the right interlvalContains() version.
			const deUint32 iRep = reinterpret_cast<const deUint32&>(value);
			float res0 = deFloat16To32((deFloat16)(iRep & 0xFFFF));
			float res1 = deFloat16To32((deFloat16)(iRep >> 16));
			return intervalContains(a, res0, modularDivisor) && (res1 == -1.0);
		}
		return intervalContains(a, value, modularDivisor);
	}
};

template<>
struct Traits<double> : ScalarTraits<double>
{
	static void			doPrintIVal		(const FloatFormat&	fmt,
										 const Interval&	ival,
										 ostream&			os)
	{
		os << fmt.intervalToHex(ival);
	}

	static void			doPrintValue16	(const FloatFormat&	fmt,
										 const double&		value,
										 ostream&			os)
	{
		const deUint64 iRep = reinterpret_cast<const deUint64&>(value);
		double byte0 = deFloat16To64((deFloat16)((iRep      ) & 0xffff));
		double byte1 = deFloat16To64((deFloat16)((iRep >> 16) & 0xffff));
		double byte2 = deFloat16To64((deFloat16)((iRep >> 32) & 0xffff));
		double byte3 = deFloat16To64((deFloat16)((iRep >> 48) & 0xffff));
		os << fmt.floatToHex(byte0) << " " << fmt.floatToHex(byte1) << " " << fmt.floatToHex(byte2) << " " << fmt.floatToHex(byte3);
	}

	static void			doPrintValue32	(const FloatFormat&	fmt,
										 const double&		value,
										 ostream&			os)
	{
		const deUint64 iRep = reinterpret_cast<const deUint64&>(value);
		double res0 = static_cast<double>((float)((iRep      ) & 0xffffffff));
		double res1 = static_cast<double>((float)((iRep >> 32) & 0xffffffff));
		os << fmt.floatToHex(res0) << " " << fmt.floatToHex(res1);
	}

	static void			doPrintValue64	(const FloatFormat&	fmt,
										 const double&		value,
										 ostream&			os)
	{
		os << fmt.floatToHex(value);
	}

	template <class U>
	static bool			doContains		(const Interval& a, const double& value, bool is16Bit, const tcu::Maybe<U>& modularDivisor)
	{
		DE_UNREF(is16Bit);
		DE_ASSERT(!is16Bit);
		return intervalContains(a, value, modularDivisor);
	}
};

template<>
struct Traits<deFloat16> : ScalarTraits<deFloat16>
{
	static void			doPrintIVal		(const FloatFormat&	fmt,
										 const Interval&	ival,
										 ostream&			os)
	{
		os << fmt.intervalToHex(ival);
	}

	static void			doPrintValue16	(const FloatFormat&	fmt,
										 const deFloat16&	value,
										 ostream&			os)
	{
		const float res0 = deFloat16To32(value);
		os << fmt.floatToHex(static_cast<double>(res0));
	}
	static void			doPrintValue32	(const FloatFormat&	fmt,
										 const deFloat16&	value,
										 ostream&			os)
	{
		const float res0 = deFloat16To32(value);
		os << fmt.floatToHex(static_cast<double>(res0));
	}

	static void			doPrintValue64	(const FloatFormat&	fmt,
										 const deFloat16&	value,
										 ostream&			os)
	{
		const double res0 = deFloat16To64(value);
		os << fmt.floatToHex(res0);
	}

	// When the value and divisor are both deFloat16, convert both to float to call the right intervalContains version.
	static bool			doContains		(const Interval& a, const deFloat16& value, bool is16Bit, const tcu::Maybe<deFloat16>& modularDivisor)
	{
		DE_UNREF(is16Bit);
		float res0 = deFloat16To32(value);
		const tcu::Maybe<float> convertedDivisor = (modularDivisor ? tcu::just(deFloat16To32(modularDivisor.get())) : tcu::nothing<float>());
		return intervalContains(a, res0, convertedDivisor);
	}

	// If the types don't match we should not be in a modulo operation, so no conversion should take place.
	template <class U>
	static bool			doContains		(const Interval& a, const deFloat16& value, bool is16Bit, const tcu::Maybe<U>& modularDivisor)
	{
		DE_UNREF(is16Bit);
		float res0 = deFloat16To32(value);
		return intervalContains(a, res0, modularDivisor);
	}
};

template<>
struct Traits<bool> : ScalarTraits<bool>
{
	static void			doPrintValue16	(const FloatFormat&,
										 const float&		value,
										 ostream&			os)
	{
		os << (value != 0.0f ? "true" : "false");
	}

	static void		doPrintValue32	(const			FloatFormat&,
									 const float&	value,
									 ostream&		os)
	{
		os << (value != 0.0f ? "true" : "false");
	}

	static void		doPrintValue64	(const			FloatFormat&,
									 const float&	value,
									 ostream&		os)
	{
		os << (value != 0.0f ? "true" : "false");
	}

	static void			doPrintIVal		(const FloatFormat&,
										 const Interval&	ival,
										 ostream&			os)
	{
		os << "{";
		if (ival.contains(false))
			os << "false";
		if (ival.contains(false) && ival.contains(true))
			os << ", ";
		if (ival.contains(true))
			os << "true";
		os << "}";
	}
};

template<>
struct Traits<int> : ScalarTraits<int>
{
	static void			doPrintValue16	(const FloatFormat&,
										 const int&			value,
										 ostream&			os)
	{
		int res0 = value & 0xFFFF;
		int res1 = value >> 16;
		os << res0 << " " << res1;
	}

	static void		doPrintValue32		(const FloatFormat&,
										 const int&			value,
										 ostream&			os)
	{
		os << value;
	}

	static void		doPrintValue64		(const FloatFormat&,
										 const int&			value,
										 ostream&			os)
	{
		os << value;
	}

	static void			doPrintIVal		(const FloatFormat&,
										 const Interval&	ival,
										 ostream&			os)
	{
		os << "[" << int(ival.lo()) << ", " << int(ival.hi()) << "]";
	}

	template <typename U>
	static bool			doContains		(const Interval& a, const int& value, bool is16Bit, const tcu::Maybe<U>& modularDivisor)
	{
		DE_UNREF(is16Bit);
		return intervalContains(a, value, modularDivisor);
	}
};

//! Common traits for containers, i.e. vectors and matrices.
//! T is the container type itself, I is the same type with interval elements.
template <typename T, typename I>
struct ContainerTraits
{
	typedef typename	T::Element		Element;
	typedef				I				IVal;

	static IVal			doMakeIVal		(const T& value)
	{
		IVal ret;

		for (int ndx = 0; ndx < T::SIZE; ++ndx)
			ret[ndx] = makeIVal(value[ndx]);

		return ret;
	}

	static IVal			doUnion			(const IVal& a, const IVal& b)
	{
		IVal ret;

		for (int ndx = 0; ndx < T::SIZE; ++ndx)
			ret[ndx] = unionIVal<Element>(a[ndx], b[ndx]);

		return ret;
	}

	// When the input and output types match, we may be in a modulo operation. If the divisor is provided, use each of its
	// components to determine if the obtained result is fine.
	static bool			doContains		(const IVal& ival, const T& value, bool is16Bit, const tcu::Maybe<T>& modularDivisor)
	{
		using DivisorElement = typename T::Element;

		for (int ndx = 0; ndx < T::SIZE; ++ndx)
		{
			const tcu::Maybe<DivisorElement> divisorElement = (modularDivisor ? tcu::just((*modularDivisor)[ndx]) : tcu::nothing<DivisorElement>());
			if (!contains(ival[ndx], value[ndx], is16Bit, divisorElement))
				return false;
		}

		return true;
	}

	// When the input and output types do not match we should not be in a modulo operation. This version is provided for syntactical
	// compatibility.
	template <typename U>
	static bool			doContains		(const IVal& ival, const T& value, bool is16Bit, const tcu::Maybe<U>& modularDivisor)
	{
		for (int ndx = 0; ndx < T::SIZE; ++ndx)
		{
			if (!contains(ival[ndx], value[ndx], is16Bit, modularDivisor))
				return false;
		}

		return true;
	}

	static void			doPrintIVal		(const FloatFormat& fmt, const IVal ival, ostream& os)
	{
		os << "(";

		for (int ndx = 0; ndx < T::SIZE; ++ndx)
		{
			if (ndx > 0)
				os << ", ";

			printIVal<Element>(fmt, ival[ndx], os);
		}

		os << ")";
	}

	static void			doPrintValue16	(const FloatFormat& fmt, const T& value, ostream& os)
	{
		os << dataTypeNameOf<T>() << "(";

		for (int ndx = 0; ndx < T::SIZE; ++ndx)
		{
			if (ndx > 0)
				os << ", ";

			printValue16<Element>(fmt, value[ndx], os);
		}

		os << ")";
	}

	static void			doPrintValue32	(const FloatFormat& fmt, const T& value, ostream& os)
	{
		os << dataTypeNameOf<T>() << "(";

		for (int ndx = 0; ndx < T::SIZE; ++ndx)
		{
			if (ndx > 0)
				os << ", ";

			printValue32<Element>(fmt, value[ndx], os);
		}

		os << ")";
	}

	static void			doPrintValue64	(const FloatFormat& fmt, const T& value, ostream& os)
	{
		os << dataTypeNameOf<T>() << "(";

		for (int ndx = 0; ndx < T::SIZE; ++ndx)
		{
			if (ndx > 0)
				os << ", ";

			printValue64<Element>(fmt, value[ndx], os);
		}

		os << ")";
	}

	static IVal			doConvert		(const FloatFormat& fmt, const IVal& value)
	{
		IVal ret;

		for (int ndx = 0; ndx < T::SIZE; ++ndx)
			ret[ndx] = convert<Element>(fmt, value[ndx]);

		return ret;
	}

	static IVal			doRound			(const FloatFormat& fmt, T value)
	{
		IVal ret;

		for (int ndx = 0; ndx < T::SIZE; ++ndx)
			ret[ndx] = round(fmt, value[ndx]);

		return ret;
	}
};

template <typename T, int Size>
struct Traits<Vector<T, Size> > :
	ContainerTraits<Vector<T, Size>, Vector<typename Traits<T>::IVal, Size> >
{
};

template <typename T, int Rows, int Cols>
struct Traits<Matrix<T, Rows, Cols> > :
	ContainerTraits<Matrix<T, Rows, Cols>, Matrix<typename Traits<T>::IVal, Rows, Cols> >
{
};

//! Void traits. These are just dummies, but technically valid: a Void is a
//! unit type with a single possible value.
template<>
struct Traits<Void>
{
	typedef		Void			IVal;

	static Void	doMakeIVal		(const Void& value)										{ return value; }
	static Void	doUnion			(const Void&, const Void&)								{ return Void(); }
	static bool	doContains		(const Void&, Void)										{ return true; }
	template <typename U>
	static bool	doContains		(const Void&, const Void& value, bool is16Bit, const tcu::Maybe<U>& modularDivisor) { DE_UNREF(value); DE_UNREF(is16Bit); DE_UNREF(modularDivisor); return true; }
	static Void	doRound			(const FloatFormat&, const Void& value)					{ return value; }
	static Void	doConvert		(const FloatFormat&, const Void& value)					{ return value; }

	static void	doPrintValue16	(const FloatFormat&, const Void&, ostream& os)
	{
		os << "()";
	}

	static void	doPrintValue32	(const FloatFormat&, const Void&, ostream& os)
	{
		os << "()";
	}

	static void	doPrintValue64	(const FloatFormat&, const Void&, ostream& os)
	{
		os << "()";
	}

	static void	doPrintIVal		(const FloatFormat&, const Void&, ostream& os)
	{
		os << "()";
	}
};

//! This is needed for container-generic operations.
//! We want a scalar type T to be its own "one-element vector".
template <typename T, int Size> struct ContainerOf	{ typedef Vector<T, Size>	Container; };

template <typename T>			struct ContainerOf<T, 1>		{ typedef T		Container; };
template <int Size>				struct ContainerOf<Void, Size>	{ typedef Void	Container; };

// This is a kludge that is only needed to get the ExprP::operator[] syntactic sugar to work.
template <typename T>	struct ElementOf		{ typedef	typename T::Element	Element; };
template <>				struct ElementOf<float>	{ typedef	void				Element; };
template <>				struct ElementOf<double>{ typedef	void				Element; };
template <>				struct ElementOf<bool>	{ typedef	void				Element; };
template <>				struct ElementOf<int>	{ typedef	void				Element; };

template <typename T>
string comparisonMessageInterval(const typename Traits<T>::IVal& val)
{
	DE_UNREF(val);
	return "";
}

template <>
string comparisonMessageInterval<int>(const Traits<int>::IVal& val)
{
	return comparisonMessage(static_cast<int>(val.lo()));
}

template <>
string comparisonMessageInterval<float>(const Traits<float>::IVal& val)
{
	return comparisonMessage(static_cast<int>(val.lo()));
}

template <>
string comparisonMessageInterval<tcu::Vector<int, 2> >(const tcu::Vector<tcu::Interval, 2> & val)
{
	tcu::IVec2 result(static_cast<int>(val[0].lo()), static_cast<int>(val[1].lo()));
	return comparisonMessage(result);
}

template <>
string comparisonMessageInterval<tcu::Vector<int, 3> >(const tcu::Vector<tcu::Interval, 3> & val)
{
	tcu::IVec3 result(static_cast<int>(val[0].lo()), static_cast<int>(val[1].lo()), static_cast<int>(val[2].lo()));
	return comparisonMessage(result);
}

template <>
string comparisonMessageInterval<tcu::Vector<int, 4> >(const tcu::Vector<tcu::Interval, 4> & val)
{
	tcu::IVec4 result(static_cast<int>(val[0].lo()), static_cast<int>(val[1].lo()), static_cast<int>(val[2].lo()), static_cast<int>(val[3].lo()));
	return comparisonMessage(result);
}

/*--------------------------------------------------------------------*//*!
 *
 * \name Abstract syntax for expressions and statements.
 *
 * We represent GLSL programs as syntax objects: an Expr<T> represents an
 * expression whose GLSL type corresponds to the C++ type T, and a Statement
 * represents a statement.
 *
 * To ease memory management, we use shared pointers to refer to expressions
 * and statements. ExprP<T> is a shared pointer to an Expr<T>, and StatementP
 * is a shared pointer to a Statement.
 *
 * \{
 *
 *//*--------------------------------------------------------------------*/

class ExprBase;
class ExpandContext;
class Statement;
class StatementP;
class FuncBase;
template <typename T> class ExprP;
template <typename T> class Variable;
template <typename T> class VariableP;
template <typename T> class DefaultSampling;

typedef set<const FuncBase*> FuncSet;

template <typename T>
VariableP<T>	variable			(const string& name);
StatementP		compoundStatement	(const vector<StatementP>& statements);

/*--------------------------------------------------------------------*//*!
 * \brief A variable environment.
 *
 * An Environment object maintains the mapping between variables of the
 * abstract syntax tree and their values.
 *
 * \todo [2014-03-28 lauri] At least run-time type safety.
 *
 *//*--------------------------------------------------------------------*/
class Environment
{
public:
	template<typename T>
	void						bind	(const Variable<T>&					variable,
										 const typename Traits<T>::IVal&	value)
	{
		deUint8* const data = new deUint8[sizeof(value)];

		deMemcpy(data, &value, sizeof(value));
		de::insert(m_map, variable.getName(), SharedPtr<deUint8>(data, de::ArrayDeleter<deUint8>()));
	}

	template<typename T>
	typename Traits<T>::IVal&	lookup	(const Variable<T>& variable) const
	{
		deUint8* const data = de::lookup(m_map, variable.getName()).get();

		return *reinterpret_cast<typename Traits<T>::IVal*>(data);
	}

private:
	map<string, SharedPtr<deUint8> >	m_map;
};

/*--------------------------------------------------------------------*//*!
 * \brief Evaluation context.
 *
 * The evaluation context contains everything that separates one execution of
 * an expression from the next. Currently this means the desired floating
 * point precision and the current variable environment.
 *
 *//*--------------------------------------------------------------------*/
struct EvalContext
{
	EvalContext (const FloatFormat&	format_,
				 Precision			floatPrecision_,
				 Environment&		env_,
				 int				callDepth_)
		: format				(format_)
		, floatPrecision		(floatPrecision_)
		, env					(env_)
		, callDepth				(callDepth_) {}

	FloatFormat		format;
	Precision		floatPrecision;
	Environment&	env;
	int				callDepth;
};

/*--------------------------------------------------------------------*//*!
 * \brief Simple incremental counter.
 *
 * This is used to make sure that different ExpandContexts will not produce
 * overlapping temporary names.
 *
 *//*--------------------------------------------------------------------*/
class Counter
{
public:
			Counter		(int count = 0) : m_count(count) {}
	int		operator()	(void) { return m_count++; }

private:
	int		m_count;
};

class ExpandContext
{
public:
						ExpandContext	(Counter& symCounter) : m_symCounter(symCounter) {}
						ExpandContext	(const ExpandContext& parent)
							: m_symCounter(parent.m_symCounter) {}

	template<typename T>
	VariableP<T>		genSym			(const string& baseName)
	{
		return variable<T>(baseName + de::toString(m_symCounter()));
	}

	void				addStatement	(const StatementP& stmt)
	{
		m_statements.push_back(stmt);
	}

	vector<StatementP>	getStatements	(void) const
	{
		return m_statements;
	}
private:
	Counter&			m_symCounter;
	vector<StatementP>	m_statements;
};

/*--------------------------------------------------------------------*//*!
 * \brief A statement or declaration.
 *
 * Statements have no values. Instead, they are executed for their side
 * effects only: the execute() method should modify at least one variable in
 * the environment.
 *
 * As a bit of a kludge, a Statement object can also represent a declaration:
 * when it is evaluated, it can add a variable binding to the environment
 * instead of modifying a current one.
 *
 *//*--------------------------------------------------------------------*/
class Statement
{
public:
	virtual			~Statement		(void)							{								 }
	//! Execute the statement, modifying the environment of `ctx`
	void			execute			(EvalContext&	ctx)	const	{ this->doExecute(ctx);			 }
	void			print			(ostream&		os)		const	{ this->doPrint(os);			 }
	//! Add the functions used in this statement to `dst`.
	void			getUsedFuncs	(FuncSet& dst)			const	{ this->doGetUsedFuncs(dst);	 }
	void			failed			(EvalContext& ctx)		const	{ this->doFail(ctx);			 }

protected:
	virtual void	doPrint			(ostream& os)			const	= 0;
	virtual void	doExecute		(EvalContext& ctx)		const	= 0;
	virtual void	doGetUsedFuncs	(FuncSet& dst)			const	= 0;
	virtual void	doFail			(EvalContext& ctx)		const	{ DE_UNREF(ctx); }
};

ostream& operator<<(ostream& os, const Statement& stmt)
{
	stmt.print(os);
	return os;
}

/*--------------------------------------------------------------------*//*!
 * \brief Smart pointer for statements (and declarations)
 *
 *//*--------------------------------------------------------------------*/
class StatementP : public SharedPtr<const Statement>
{
public:
	typedef		SharedPtr<const Statement>	Super;

				StatementP			(void) {}
	explicit	StatementP			(const Statement* ptr)	: Super(ptr) {}
				StatementP			(const Super& ptr)		: Super(ptr) {}
};

/*--------------------------------------------------------------------*//*!
 * \brief
 *
 * A statement that modifies a variable or a declaration that binds a variable.
 *
 *//*--------------------------------------------------------------------*/
template <typename T>
class VariableStatement : public Statement
{
public:
					VariableStatement	(const VariableP<T>& variable, const ExprP<T>& value,
										 bool isDeclaration)
						: m_variable		(variable)
						, m_value			(value)
						, m_isDeclaration	(isDeclaration) {}

protected:
	void			doPrint				(ostream& os)							const
	{
		if (m_isDeclaration)
			os << glu::declare(getVarTypeOf<T>(), m_variable->getName());
		else
			os << m_variable->getName();

		os << " = ";
		os<< *m_value << ";\n";
	}

	void			doExecute			(EvalContext& ctx)						const
	{
		if (m_isDeclaration)
			ctx.env.bind(*m_variable, m_value->evaluate(ctx));
		else
			ctx.env.lookup(*m_variable) = m_value->evaluate(ctx);
	}

	void			doGetUsedFuncs		(FuncSet& dst)							const
	{
		m_value->getUsedFuncs(dst);
	}

	virtual void	doFail			(EvalContext& ctx)		const
	{
		if (m_isDeclaration)
			ctx.env.bind(*m_variable, m_value->fails(ctx));
		else
			ctx.env.lookup(*m_variable) = m_value->fails(ctx);
	}

	VariableP<T>	m_variable;
	ExprP<T>		m_value;
	bool			m_isDeclaration;
};

template <typename T>
StatementP variableStatement (const VariableP<T>&	variable,
							  const ExprP<T>&		value,
							  bool					isDeclaration)
{
	return StatementP(new VariableStatement<T>(variable, value, isDeclaration));
}

template <typename T>
StatementP variableDeclaration (const VariableP<T>& variable, const ExprP<T>& definiens)
{
	return variableStatement(variable, definiens, true);
}

template <typename T>
StatementP variableAssignment (const VariableP<T>& variable, const ExprP<T>& value)
{
	return variableStatement(variable, value, false);
}

/*--------------------------------------------------------------------*//*!
 * \brief A compound statement, i.e. a block.
 *
 * A compound statement is executed by executing its constituent statements in
 * sequence.
 *
 *//*--------------------------------------------------------------------*/
class CompoundStatement : public Statement
{
public:
						CompoundStatement	(const vector<StatementP>& statements)
							: m_statements	(statements) {}

protected:
	void				doPrint				(ostream&		os)						const
	{
		os << "{\n";

		for (size_t ndx = 0; ndx < m_statements.size(); ++ndx)
			os << *m_statements[ndx];

		os << "}\n";
	}

	void				doExecute			(EvalContext&	ctx)					const
	{
		for (size_t ndx = 0; ndx < m_statements.size(); ++ndx)
			m_statements[ndx]->execute(ctx);
	}

	void				doGetUsedFuncs		(FuncSet& dst)							const
	{
		for (size_t ndx = 0; ndx < m_statements.size(); ++ndx)
			m_statements[ndx]->getUsedFuncs(dst);
	}

	vector<StatementP>	m_statements;
};

StatementP compoundStatement(const vector<StatementP>& statements)
{
	return StatementP(new CompoundStatement(statements));
}

//! Common base class for all expressions regardless of their type.
class ExprBase
{
public:
	virtual				~ExprBase		(void)									{}
	void				printExpr		(ostream& os) const { this->doPrintExpr(os); }

	//! Output the functions that this expression refers to
	void				getUsedFuncs	(FuncSet& dst) const
	{
		this->doGetUsedFuncs(dst);
	}

protected:
	virtual void		doPrintExpr		(ostream&)	const	{}
	virtual void		doGetUsedFuncs	(FuncSet&)	const	{}
};

//! Type-specific operations for an expression representing type T.
template <typename T>
class Expr : public ExprBase
{
public:
	typedef				T				Val;
	typedef typename	Traits<T>::IVal	IVal;

	IVal				evaluate		(const EvalContext&	ctx) const;
	IVal				fails			(const EvalContext&	ctx) const	{ return this->doFails(ctx); }

protected:
	virtual IVal		doEvaluate		(const EvalContext&	ctx) const = 0;
	virtual IVal		doFails			(const EvalContext&	ctx) const {return doEvaluate(ctx);}
};

//! Evaluate an expression with the given context, optionally tracing the calls to stderr.
template <typename T>
typename Traits<T>::IVal Expr<T>::evaluate (const EvalContext& ctx) const
{
#ifdef GLS_ENABLE_TRACE
	static const FloatFormat	highpFmt	(-126, 127, 23, true,
											 tcu::MAYBE,
											 tcu::YES,
											 tcu::MAYBE);
	EvalContext					newCtx		(ctx.format, ctx.floatPrecision,
											 ctx.env, ctx.callDepth + 1);
	const IVal					ret			= this->doEvaluate(newCtx);

	if (isTypeValid<T>())
	{
		std::cerr << string(ctx.callDepth, ' ');
		this->printExpr(std::cerr);
		std::cerr << " -> " << intervalToString<T>(highpFmt, ret) << std::endl;
	}
	return ret;
#else
	return this->doEvaluate(ctx);
#endif
}

template <typename T>
class ExprPBase : public SharedPtr<const Expr<T> >
{
public:
};

ostream& operator<< (ostream& os, const ExprBase& expr)
{
	expr.printExpr(os);
	return os;
}

/*--------------------------------------------------------------------*//*!
 * \brief Shared pointer to an expression of a container type.
 *
 * Container types (i.e. vectors and matrices) support the subscription
 * operator. This class provides a bit of syntactic sugar to allow us to use
 * the C++ subscription operator to create a subscription expression.
 *//*--------------------------------------------------------------------*/
template <typename T>
class ContainerExprPBase : public ExprPBase<T>
{
public:
	ExprP<typename T::Element>	operator[]	(int i) const;
};

template <typename T>
class ExprP : public ExprPBase<T> {};

// We treat Voids as containers since the unused parameters in generalized
// vector functions are represented as Voids.
template <>
class ExprP<Void> : public ContainerExprPBase<Void> {};

template <typename T, int Size>
class ExprP<Vector<T, Size> > : public ContainerExprPBase<Vector<T, Size> > {};

template <typename T, int Rows, int Cols>
class ExprP<Matrix<T, Rows, Cols> > : public ContainerExprPBase<Matrix<T, Rows, Cols> > {};

template <typename T> ExprP<T> exprP (void)
{
	return ExprP<T>();
}

template <typename T>
ExprP<T> exprP (const SharedPtr<const Expr<T> >& ptr)
{
	ExprP<T> ret;
	static_cast<SharedPtr<const Expr<T> >&>(ret) = ptr;
	return ret;
}

template <typename T>
ExprP<T> exprP (const Expr<T>* ptr)
{
	return exprP(SharedPtr<const Expr<T> >(ptr));
}

/*--------------------------------------------------------------------*//*!
 * \brief A shared pointer to a variable expression.
 *
 * This is just a narrowing of ExprP for the operations that require a variable
 * instead of an arbitrary expression.
 *
 *//*--------------------------------------------------------------------*/
template <typename T>
class VariableP : public SharedPtr<const Variable<T> >
{
public:
	typedef		SharedPtr<const Variable<T> >	Super;
	explicit	VariableP	(const Variable<T>* ptr) : Super(ptr) {}
				VariableP	(void) {}
				VariableP	(const Super& ptr) : Super(ptr) {}

	operator	ExprP<T>	(void) const { return exprP(SharedPtr<const Expr<T> >(*this)); }
};

/*--------------------------------------------------------------------*//*!
 * \name Syntactic sugar operators for expressions.
 *
 * @{
 *
 * These operators allow the use of C++ syntax to construct GLSL expressions
 * containing operators: e.g. "a+b" creates an addition expression with
 * operands a and b, and so on.
 *
 *//*--------------------------------------------------------------------*/
ExprP<float>						operator+ (const ExprP<float>&						arg0,
											  const ExprP<float>&						arg1);
ExprP<deFloat16>					operator+ (const ExprP<deFloat16>&					arg0,
											  const ExprP<deFloat16>&					arg1);
ExprP<double>						operator+ (const ExprP<double>&						arg0,
											  const ExprP<double>&						arg1);
template <typename T>
ExprP<T>							operator- (const ExprP<T>& arg0);
template <typename T>
ExprP<T>							operator- (const ExprP<T>&							arg0,
											  const ExprP<T>&							arg1);
template<int Left, int Mid, int Right, typename T>
ExprP<Matrix<T, Left, Right> >		operator* (const ExprP<Matrix<T, Left, Mid> >&		left,
											   const ExprP<Matrix<T, Mid, Right> >&		right);
ExprP<float>						operator* (const ExprP<float>&						arg0,
											  const ExprP<float>&						arg1);
ExprP<deFloat16>					operator* (const ExprP<deFloat16>&					arg0,
											   const ExprP<deFloat16>&					arg1);
ExprP<double>						operator* (const ExprP<double>&						arg0,
											  const ExprP<double>&						arg1);
template <typename T>
ExprP<T>							operator/ (const ExprP<T>&							arg0,
											  const ExprP<T>&							arg1);
template<typename T, int Size>
ExprP<Vector<T, Size> >				operator- (const ExprP<Vector<T, Size> >&			arg0);
template<typename T, int Size>
ExprP<Vector<T, Size> >				operator- (const ExprP<Vector<T, Size> >&			arg0,
											   const ExprP<Vector<T, Size> >&			arg1);
template<int Size, typename T>
ExprP<Vector<T, Size> >				operator* (const ExprP<Vector<T, Size> >&			arg0,
											   const ExprP<T>&							arg1);
template<typename T, int Size>
ExprP<Vector<T, Size> >				operator* (const ExprP<Vector<T, Size> >&			arg0,
											   const ExprP<Vector<T, Size> >&			arg1);
template<int Rows, int Cols, typename T>
ExprP<Vector<T, Rows> >				operator* (const ExprP<Vector<T, Cols> >&			left,
											   const ExprP<Matrix<T, Rows, Cols> >&		right);
template<int Rows, int Cols, typename T>
ExprP<Vector<T, Cols> >				operator* (const ExprP<Matrix<T, Rows, Cols> >&		left,
											   const ExprP<Vector<T, Rows> >&			right);
template<int Rows, int Cols, typename T>
ExprP<Matrix<T, Rows, Cols> >		operator* (const ExprP<Matrix<T, Rows, Cols> >&		left,
											   const ExprP<T>&							right);
template<int Rows, int Cols>
ExprP<Matrix<float, Rows, Cols> >	operator+ (const ExprP<Matrix<float, Rows, Cols> >&	left,
											   const ExprP<Matrix<float, Rows, Cols> >&	right);
template<int Rows, int Cols>
ExprP<Matrix<deFloat16, Rows, Cols> >	operator+ (const ExprP<Matrix<deFloat16, Rows, Cols> >&	left,
												   const ExprP<Matrix<deFloat16, Rows, Cols> >&	right);
template<int Rows, int Cols>
ExprP<Matrix<double, Rows, Cols> >	operator+ (const ExprP<Matrix<double, Rows, Cols> >&	left,
											   const ExprP<Matrix<double, Rows, Cols> >&	right);
template<typename T, int Rows, int Cols>
ExprP<Matrix<T, Rows, Cols> >	operator- (const ExprP<Matrix<T, Rows, Cols> >&	mat);

//! @}

/*--------------------------------------------------------------------*//*!
 * \brief Variable expression.
 *
 * A variable is evaluated by looking up its range of possible values from an
 * environment.
 *//*--------------------------------------------------------------------*/
template <typename T>
class Variable : public Expr<T>
{
public:
	typedef typename Expr<T>::IVal IVal;

					Variable	(const string& name) : m_name (name) {}
	string			getName		(void)							const { return m_name; }

protected:
	void			doPrintExpr	(ostream& os)					const { os << m_name; }
	IVal			doEvaluate	(const EvalContext& ctx)		const
	{
		return ctx.env.lookup<T>(*this);
	}

private:
	string	m_name;
};

template <typename T>
VariableP<T> variable (const string& name)
{
	return VariableP<T>(new Variable<T>(name));
}

template <typename T>
VariableP<T> bindExpression (const string& name, ExpandContext& ctx, const ExprP<T>& expr)
{
	VariableP<T> var = ctx.genSym<T>(name);
	ctx.addStatement(variableDeclaration(var, expr));
	return var;
}

/*--------------------------------------------------------------------*//*!
 * \brief Constant expression.
 *
 * A constant is evaluated by rounding it to a set of possible values allowed
 * by the current floating point precision.
 *//*--------------------------------------------------------------------*/
template <typename T>
class Constant : public Expr<T>
{
public:
	typedef typename Expr<T>::IVal IVal;

			Constant		(const T& value) : m_value(value) {}

protected:
	void	doPrintExpr		(ostream& os) const			{ os << m_value; }
	IVal	doEvaluate		(const EvalContext&) const	{ return makeIVal(m_value); }

private:
	T		m_value;
};

template <typename T>
ExprP<T> constant (const T& value)
{
	return exprP(new Constant<T>(value));
}

//! Return a reference to a singleton void constant.
const ExprP<Void>& voidP (void)
{
	static const ExprP<Void> singleton = constant(Void());

	return singleton;
}

/*--------------------------------------------------------------------*//*!
 * \brief Four-element tuple.
 *
 * This is used for various things where we need one thing for each possible
 * function parameter. Currently the maximum supported number of parameters is
 * four.
 *//*--------------------------------------------------------------------*/
template <typename T0 = Void, typename T1 = Void, typename T2 = Void, typename T3 = Void>
struct Tuple4
{
	explicit Tuple4 (const T0 e0 = T0(),
					 const T1 e1 = T1(),
					 const T2 e2 = T2(),
					 const T3 e3 = T3())
		: a	(e0)
		, b	(e1)
		, c	(e2)
		, d	(e3)
	{
	}

	T0 a;
	T1 b;
	T2 c;
	T3 d;
};

/*--------------------------------------------------------------------*//*!
 * \brief Function signature.
 *
 * This is a purely compile-time structure used to bundle all types in a
 * function signature together. This makes passing the signature around in
 * templates easier, since we only need to take and pass a single Sig instead
 * of a bunch of parameter types and a return type.
 *
 *//*--------------------------------------------------------------------*/
template <typename R,
		  typename P0 = Void, typename P1 = Void,
		  typename P2 = Void, typename P3 = Void>
struct Signature
{
	typedef R							Ret;
	typedef P0							Arg0;
	typedef P1							Arg1;
	typedef P2							Arg2;
	typedef P3							Arg3;
	typedef typename Traits<Ret>::IVal	IRet;
	typedef typename Traits<Arg0>::IVal	IArg0;
	typedef typename Traits<Arg1>::IVal	IArg1;
	typedef typename Traits<Arg2>::IVal	IArg2;
	typedef typename Traits<Arg3>::IVal	IArg3;

	typedef Tuple4<	const Arg0&,	const Arg1&,	const Arg2&,	const Arg3&>	Args;
	typedef Tuple4<	const IArg0&,	const IArg1&,	const IArg2&,	const IArg3&>	IArgs;
	typedef Tuple4<	ExprP<Arg0>,	ExprP<Arg1>,	ExprP<Arg2>,	ExprP<Arg3> >	ArgExprs;
};

typedef vector<const ExprBase*> BaseArgExprs;

/*--------------------------------------------------------------------*//*!
 * \brief Type-independent operations for function objects.
 *
 *//*--------------------------------------------------------------------*/
class FuncBase
{
public:
	virtual				~FuncBase				(void)					{}
	virtual string		getName					(void)					const = 0;
	//! Name of extension that this function requires, or empty.
	virtual string		getRequiredExtension	(void)					const { return ""; }
	virtual Interval	getInputRange			(const bool is16bit)	const {DE_UNREF(is16bit); return Interval(true, -TCU_INFINITY, TCU_INFINITY); }
	virtual void		print					(ostream&,
												 const BaseArgExprs&)	const = 0;
	//! Index of output parameter, or -1 if none of the parameters is output.
	virtual int			getOutParamIndex		(void)					const { return -1; }

	virtual SpirVCaseT	getSpirvCase			(void)					const { return SPIRV_CASETYPE_NONE; }

	void				printDefinition			(ostream& os)			const
	{
		doPrintDefinition(os);
	}

	void				getUsedFuncs			(FuncSet& dst) const
	{
		this->doGetUsedFuncs(dst);
	}

protected:
	virtual void		doPrintDefinition		(ostream& os)			const = 0;
	virtual void		doGetUsedFuncs			(FuncSet& dst)			const = 0;
};

typedef Tuple4<string, string, string, string> ParamNames;

/*--------------------------------------------------------------------*//*!
 * \brief Function objects.
 *
 * Each Func object represents a GLSL function. It can be applied to interval
 * arguments, and it returns the an interval that is a conservative
 * approximation of the image of the GLSL function over the argument
 * intervals. That is, it is given a set of possible arguments and it returns
 * the set of possible values.
 *
 *//*--------------------------------------------------------------------*/
template <typename Sig_>
class Func : public FuncBase
{
public:
	typedef Sig_										Sig;
	typedef typename Sig::Ret							Ret;
	typedef typename Sig::Arg0							Arg0;
	typedef typename Sig::Arg1							Arg1;
	typedef typename Sig::Arg2							Arg2;
	typedef typename Sig::Arg3							Arg3;
	typedef typename Sig::IRet							IRet;
	typedef typename Sig::IArg0							IArg0;
	typedef typename Sig::IArg1							IArg1;
	typedef typename Sig::IArg2							IArg2;
	typedef typename Sig::IArg3							IArg3;
	typedef typename Sig::Args							Args;
	typedef typename Sig::IArgs							IArgs;
	typedef typename Sig::ArgExprs						ArgExprs;

	void				print			(ostream&			os,
										 const BaseArgExprs& args)				const
	{
		this->doPrint(os, args);
	}

	IRet				apply			(const EvalContext&	ctx,
										 const IArg0&		arg0 = IArg0(),
										 const IArg1&		arg1 = IArg1(),
										 const IArg2&		arg2 = IArg2(),
										 const IArg3&		arg3 = IArg3())		const
	{
		return this->applyArgs(ctx, IArgs(arg0, arg1, arg2, arg3));
	}

	IRet				fail			(const EvalContext&	ctx,
										 const IArg0&		arg0 = IArg0(),
										 const IArg1&		arg1 = IArg1(),
										 const IArg2&		arg2 = IArg2(),
										 const IArg3&		arg3 = IArg3())		const
	{
		return this->doFail(ctx, IArgs(arg0, arg1, arg2, arg3));
	}
	IRet				applyArgs		(const EvalContext&	ctx,
										 const IArgs&		args)				const
	{
		return this->doApply(ctx, args);
	}
	ExprP<Ret>			operator()		(const ExprP<Arg0>&		arg0 = voidP(),
										 const ExprP<Arg1>&		arg1 = voidP(),
										 const ExprP<Arg2>&		arg2 = voidP(),
										 const ExprP<Arg3>&		arg3 = voidP())		const;

	const ParamNames&	getParamNames	(void)									const
	{
		return this->doGetParamNames();
	}

protected:
	virtual IRet		doApply			(const EvalContext&,
										 const IArgs&)							const = 0;
	virtual IRet		doFail			(const EvalContext&	ctx,
										 const IArgs&		args)				const
	{
		return this->doApply(ctx, args);
	}
	virtual void		doPrint			(ostream& os, const BaseArgExprs& args)	const
	{
		os << getName() << "(";

		if (isTypeValid<Arg0>())
			os << *args[0];

		if (isTypeValid<Arg1>())
			os << ", " << *args[1];

		if (isTypeValid<Arg2>())
			os << ", " << *args[2];

		if (isTypeValid<Arg3>())
			os << ", " << *args[3];

		os << ")";
	}

	virtual const ParamNames&	doGetParamNames	(void)							const
	{
		static ParamNames	names	("a", "b", "c", "d");
		return names;
	}
};

template <typename Sig>
class Apply : public Expr<typename Sig::Ret>
{
public:
	typedef typename Sig::Ret				Ret;
	typedef typename Sig::Arg0				Arg0;
	typedef typename Sig::Arg1				Arg1;
	typedef typename Sig::Arg2				Arg2;
	typedef typename Sig::Arg3				Arg3;
	typedef typename Expr<Ret>::Val			Val;
	typedef typename Expr<Ret>::IVal		IVal;
	typedef Func<Sig>						ApplyFunc;
	typedef typename ApplyFunc::ArgExprs	ArgExprs;

						Apply	(const ApplyFunc&		func,
								 const ExprP<Arg0>&		arg0 = voidP(),
								 const ExprP<Arg1>&		arg1 = voidP(),
								 const ExprP<Arg2>&		arg2 = voidP(),
								 const ExprP<Arg3>&		arg3 = voidP())
							: m_func	(func),
							  m_args	(arg0, arg1, arg2, arg3) {}

						Apply	(const ApplyFunc&	func,
								 const ArgExprs&	args)
							: m_func	(func),
							  m_args	(args) {}
protected:
	void				doPrintExpr			(ostream& os) const
	{
		BaseArgExprs	args;
		args.push_back(m_args.a.get());
		args.push_back(m_args.b.get());
		args.push_back(m_args.c.get());
		args.push_back(m_args.d.get());
		m_func.print(os, args);
	}

	IVal				doEvaluate		(const EvalContext& ctx) const
	{
		return m_func.apply(ctx,
							m_args.a->evaluate(ctx), m_args.b->evaluate(ctx),
							m_args.c->evaluate(ctx), m_args.d->evaluate(ctx));
	}

	void				doGetUsedFuncs	(FuncSet& dst) const
	{
		m_func.getUsedFuncs(dst);
		m_args.a->getUsedFuncs(dst);
		m_args.b->getUsedFuncs(dst);
		m_args.c->getUsedFuncs(dst);
		m_args.d->getUsedFuncs(dst);
	}

	const ApplyFunc&	m_func;
	ArgExprs			m_args;
};

template<typename T>
class Alternatives : public Func<Signature<T, T, T> >
{
public:
	typedef typename	Alternatives::Sig		Sig;

protected:
	typedef typename	Alternatives::IRet		IRet;
	typedef typename	Alternatives::IArgs		IArgs;

	virtual string		getName				(void) const			{ return "alternatives"; }
	virtual void		doPrintDefinition	(std::ostream&) const	{}
	void				doGetUsedFuncs		(FuncSet&) const		{}

	virtual IRet		doApply				(const EvalContext&, const IArgs& args) const
	{
		return unionIVal<T>(args.a, args.b);
	}

	virtual void		doPrint				(ostream& os, const BaseArgExprs& args)	const
	{
		os << "{" << *args[0] << " | " << *args[1] << "}";
	}
};

template <typename Sig>
ExprP<typename Sig::Ret> createApply (const Func<Sig>&						func,
									  const typename Func<Sig>::ArgExprs&	args)
{
	return exprP(new Apply<Sig>(func, args));
}

template <typename Sig>
ExprP<typename Sig::Ret> createApply (
	const Func<Sig>&			func,
	const ExprP<typename Sig::Arg0>&	arg0 = voidP(),
	const ExprP<typename Sig::Arg1>&	arg1 = voidP(),
	const ExprP<typename Sig::Arg2>&	arg2 = voidP(),
	const ExprP<typename Sig::Arg3>&	arg3 = voidP())
{
	return exprP(new Apply<Sig>(func, arg0, arg1, arg2, arg3));
}

template <typename Sig>
ExprP<typename Sig::Ret> Func<Sig>::operator() (const ExprP<typename Sig::Arg0>& arg0,
												const ExprP<typename Sig::Arg1>& arg1,
												const ExprP<typename Sig::Arg2>& arg2,
												const ExprP<typename Sig::Arg3>& arg3) const
{
	return createApply(*this, arg0, arg1, arg2, arg3);
}

template <typename F>
ExprP<typename F::Ret> app (const ExprP<typename F::Arg0>& arg0 = voidP(),
							const ExprP<typename F::Arg1>& arg1 = voidP(),
							const ExprP<typename F::Arg2>& arg2 = voidP(),
							const ExprP<typename F::Arg3>& arg3 = voidP())
{
	return createApply(instance<F>(), arg0, arg1, arg2, arg3);
}

template <typename F>
typename F::IRet call (const EvalContext&			ctx,
					   const typename F::IArg0&		arg0 = Void(),
					   const typename F::IArg1&		arg1 = Void(),
					   const typename F::IArg2&		arg2 = Void(),
					   const typename F::IArg3&		arg3 = Void())
{
	return instance<F>().apply(ctx, arg0, arg1, arg2, arg3);
}

template <typename T>
ExprP<T> alternatives (const ExprP<T>& arg0,
					   const ExprP<T>& arg1)
{
	return createApply<typename Alternatives<T>::Sig>(instance<Alternatives<T> >(), arg0, arg1);
}

template <typename Sig>
class ApplyVar : public Apply<Sig>
{
public:
	typedef typename Sig::Ret				Ret;
	typedef typename Sig::Arg0				Arg0;
	typedef typename Sig::Arg1				Arg1;
	typedef typename Sig::Arg2				Arg2;
	typedef typename Sig::Arg3				Arg3;
	typedef typename Expr<Ret>::Val			Val;
	typedef typename Expr<Ret>::IVal		IVal;
	typedef Func<Sig>						ApplyFunc;
	typedef typename ApplyFunc::ArgExprs	ArgExprs;

						ApplyVar	(const ApplyFunc&			func,
									 const VariableP<Arg0>&		arg0,
									 const VariableP<Arg1>&		arg1,
									 const VariableP<Arg2>&		arg2,
									 const VariableP<Arg3>&		arg3)
							: Apply<Sig> (func, arg0, arg1, arg2, arg3) {}
protected:
	IVal				doEvaluate		(const EvalContext& ctx) const
	{
		const Variable<Arg0>&	var0 = static_cast<const Variable<Arg0>&>(*this->m_args.a);
		const Variable<Arg1>&	var1 = static_cast<const Variable<Arg1>&>(*this->m_args.b);
		const Variable<Arg2>&	var2 = static_cast<const Variable<Arg2>&>(*this->m_args.c);
		const Variable<Arg3>&	var3 = static_cast<const Variable<Arg3>&>(*this->m_args.d);
		return this->m_func.apply(ctx,
								  ctx.env.lookup(var0), ctx.env.lookup(var1),
								  ctx.env.lookup(var2), ctx.env.lookup(var3));
	}

	IVal				doFails		(const EvalContext& ctx) const
	{
		const Variable<Arg0>&	var0 = static_cast<const Variable<Arg0>&>(*this->m_args.a);
		const Variable<Arg1>&	var1 = static_cast<const Variable<Arg1>&>(*this->m_args.b);
		const Variable<Arg2>&	var2 = static_cast<const Variable<Arg2>&>(*this->m_args.c);
		const Variable<Arg3>&	var3 = static_cast<const Variable<Arg3>&>(*this->m_args.d);
		return this->m_func.fail(ctx,
								  ctx.env.lookup(var0), ctx.env.lookup(var1),
								  ctx.env.lookup(var2), ctx.env.lookup(var3));
	}
};

template <typename Sig>
ExprP<typename Sig::Ret> applyVar (const Func<Sig>&						func,
								   const VariableP<typename Sig::Arg0>&	arg0,
								   const VariableP<typename Sig::Arg1>&	arg1,
								   const VariableP<typename Sig::Arg2>&	arg2,
								   const VariableP<typename Sig::Arg3>&	arg3)
{
	return exprP(new ApplyVar<Sig>(func, arg0, arg1, arg2, arg3));
}

template <typename Sig_>
class DerivedFunc : public Func<Sig_>
{
public:
	typedef typename DerivedFunc::ArgExprs		ArgExprs;
	typedef typename DerivedFunc::IRet			IRet;
	typedef typename DerivedFunc::IArgs			IArgs;
	typedef typename DerivedFunc::Ret			Ret;
	typedef typename DerivedFunc::Arg0			Arg0;
	typedef typename DerivedFunc::Arg1			Arg1;
	typedef typename DerivedFunc::Arg2			Arg2;
	typedef typename DerivedFunc::Arg3			Arg3;
	typedef typename DerivedFunc::IArg0			IArg0;
	typedef typename DerivedFunc::IArg1			IArg1;
	typedef typename DerivedFunc::IArg2			IArg2;
	typedef typename DerivedFunc::IArg3			IArg3;

protected:
	void						doPrintDefinition	(ostream& os) const
	{
		const ParamNames&	paramNames	= this->getParamNames();

		initialize();

		os << dataTypeNameOf<Ret>() << " " << this->getName()
			<< "(";
		if (isTypeValid<Arg0>())
			os << dataTypeNameOf<Arg0>() << " " << paramNames.a;
		if (isTypeValid<Arg1>())
			os << ", " << dataTypeNameOf<Arg1>() << " " << paramNames.b;
		if (isTypeValid<Arg2>())
			os << ", " << dataTypeNameOf<Arg2>() << " " << paramNames.c;
		if (isTypeValid<Arg3>())
			os << ", " << dataTypeNameOf<Arg3>() << " " << paramNames.d;
		os << ")\n{\n";

		for (size_t ndx = 0; ndx < m_body.size(); ++ndx)
			os << *m_body[ndx];
		os << "return " << *m_ret << ";\n";
		os << "}\n";
	}

	IRet						doApply			(const EvalContext&	ctx,
												 const IArgs&		args) const
	{
		Environment	funEnv;
		IArgs&		mutArgs		= const_cast<IArgs&>(args);
		IRet		ret;

		initialize();

		funEnv.bind(*m_var0, args.a);
		funEnv.bind(*m_var1, args.b);
		funEnv.bind(*m_var2, args.c);
		funEnv.bind(*m_var3, args.d);

		{
			EvalContext	funCtx(ctx.format, ctx.floatPrecision, funEnv, ctx.callDepth);

			for (size_t ndx = 0; ndx < m_body.size(); ++ndx)
				m_body[ndx]->execute(funCtx);

			ret = m_ret->evaluate(funCtx);
		}

		// \todo [lauri] Store references instead of values in environment
		const_cast<IArg0&>(mutArgs.a) = funEnv.lookup(*m_var0);
		const_cast<IArg1&>(mutArgs.b) = funEnv.lookup(*m_var1);
		const_cast<IArg2&>(mutArgs.c) = funEnv.lookup(*m_var2);
		const_cast<IArg3&>(mutArgs.d) = funEnv.lookup(*m_var3);

		return ret;
	}

	void						doGetUsedFuncs	(FuncSet& dst) const
	{
		initialize();
		if (dst.insert(this).second)
		{
			for (size_t ndx = 0; ndx < m_body.size(); ++ndx)
				m_body[ndx]->getUsedFuncs(dst);
			m_ret->getUsedFuncs(dst);
		}
	}

	virtual ExprP<Ret>			doExpand		(ExpandContext& ctx, const ArgExprs& args_) const = 0;

	// These are transparently initialized when first needed. They cannot be
	// initialized in the constructor because they depend on the doExpand
	// method of the subclass.

	mutable VariableP<Arg0>		m_var0;
	mutable VariableP<Arg1>		m_var1;
	mutable VariableP<Arg2>		m_var2;
	mutable VariableP<Arg3>		m_var3;
	mutable vector<StatementP>	m_body;
	mutable ExprP<Ret>			m_ret;

private:

	void				initialize		(void)	const
	{
		if (!m_ret)
		{
			const ParamNames&	paramNames	= this->getParamNames();
			Counter				symCounter;
			ExpandContext		ctx			(symCounter);
			ArgExprs			args;

			args.a	= m_var0 = variable<Arg0>(paramNames.a);
			args.b	= m_var1 = variable<Arg1>(paramNames.b);
			args.c	= m_var2 = variable<Arg2>(paramNames.c);
			args.d	= m_var3 = variable<Arg3>(paramNames.d);

			m_ret	= this->doExpand(ctx, args);
			m_body	= ctx.getStatements();
		}
	}
};

template <typename Sig>
class PrimitiveFunc : public Func<Sig>
{
public:
	typedef typename PrimitiveFunc::Ret			Ret;
	typedef typename PrimitiveFunc::ArgExprs	ArgExprs;

protected:
	void	doPrintDefinition	(ostream&) const	{}
	void	doGetUsedFuncs		(FuncSet&) const	{}
};

template <typename T>
class Cond : public PrimitiveFunc<Signature<T, bool, T, T> >
{
public:
	typedef typename Cond::IArgs	IArgs;
	typedef typename Cond::IRet		IRet;

	string	getName	(void) const
	{
		return "_cond";
	}

protected:

	void	doPrint	(ostream& os, const BaseArgExprs& args) const
	{
		os << "(" << *args[0] << " ? " << *args[1] << " : " << *args[2] << ")";
	}

	IRet	doApply	(const EvalContext&, const IArgs& iargs)const
	{
		IRet	ret;

		if (iargs.a.contains(true))
			ret = unionIVal<T>(ret, iargs.b);

		if (iargs.a.contains(false))
			ret = unionIVal<T>(ret, iargs.c);

		return ret;
	}
};

template <typename T>
class CompareOperator : public PrimitiveFunc<Signature<bool, T, T> >
{
public:
	typedef typename CompareOperator::IArgs	IArgs;
	typedef typename CompareOperator::IArg0	IArg0;
	typedef typename CompareOperator::IArg1	IArg1;
	typedef typename CompareOperator::IRet	IRet;

protected:
	void			doPrint	(ostream& os, const BaseArgExprs& args) const
	{
		os << "(" << *args[0] << getSymbol() << *args[1] << ")";
	}

	Interval		doApply	(const EvalContext&, const IArgs& iargs) const
	{
		const IArg0&	arg0 = iargs.a;
		const IArg1&	arg1 = iargs.b;
		IRet	ret;

		if (canSucceed(arg0, arg1))
			ret |= true;
		if (canFail(arg0, arg1))
			ret |= false;

		return ret;
	}

	virtual string	getSymbol	(void) const = 0;
	virtual bool	canSucceed	(const IArg0&, const IArg1&) const = 0;
	virtual bool	canFail		(const IArg0&, const IArg1&) const = 0;
};

template <typename T>
class LessThan : public CompareOperator<T>
{
public:
	string	getName		(void) const									{ return "lessThan"; }

protected:
	string	getSymbol	(void) const									{ return "<";		}

	bool	canSucceed	(const Interval& a, const Interval& b) const
	{
		return (a.lo() < b.hi());
	}

	bool	canFail		(const Interval& a, const Interval& b) const
	{
		return !(a.hi() < b.lo());
	}
};

template <typename T>
ExprP<bool> operator< (const ExprP<T>& a, const ExprP<T>& b)
{
	return app<LessThan<T> >(a, b);
}

template <typename T>
ExprP<T> cond (const ExprP<bool>&	test,
			   const ExprP<T>&		consequent,
			   const ExprP<T>&		alternative)
{
	return app<Cond<T> >(test, consequent, alternative);
}

/*--------------------------------------------------------------------*//*!
 *
 * @}
 *
 *//*--------------------------------------------------------------------*/
//Proper parameters for template T
//	Signature<float, float>		32bit tests
//	Signature<float, deFloat16>	16bit tests
//	Signature<double, double>	64bit tests
template< class T>
class FloatFunc1 : public PrimitiveFunc<T>
{
protected:
		Interval			doApply			(const EvalContext& ctx, const typename Signature<typename T::Ret, typename T::Arg0>::IArgs& iargs) const
	{
		return this->applyMonotone(ctx, iargs.a);
	}

	Interval			applyMonotone	(const EvalContext& ctx, const Interval& iarg0) const
	{
		Interval ret;

		TCU_INTERVAL_APPLY_MONOTONE1(ret, arg0, iarg0, val,
									 TCU_SET_INTERVAL(val, point,
													  point = this->applyPoint(ctx, arg0)));

		ret |= innerExtrema(ctx, iarg0);
		ret &= (this->getCodomain(ctx) | TCU_NAN);

		return ctx.format.convert(ret);
	}

	virtual Interval	innerExtrema	(const EvalContext&, const Interval&) const
	{
		return Interval(); // empty interval, i.e. no extrema
	}

	virtual Interval	applyPoint		(const EvalContext& ctx, double arg0) const
	{
		const double	exact	= this->applyExact(arg0);
		const double	prec	= this->precision(ctx, exact, arg0);

		return exact + Interval(-prec, prec);
	}

	virtual double		applyExact		(double) const
	{
		TCU_THROW(InternalError, "Cannot apply");
	}

	virtual Interval	getCodomain		(const EvalContext&) const
	{
		return Interval::unbounded(true);
	}

	virtual double		precision		(const EvalContext& ctx, double, double) const = 0;
};

/*Proper parameters for template T
	Signature<double, double>	64bit tests
	Signature<float, float>		32bit tests
	Signature<float, deFloat16>	16bit tests*/
template <class T>
class CFloatFunc1 : public FloatFunc1<T>
{
public:
						CFloatFunc1	(const string& name, tcu::DoubleFunc1& func)
							: m_name(name), m_func(func) {}

	string				getName		(void) const		{ return m_name; }

protected:
	double				applyExact	(double x) const	{ return m_func(x); }

	const string		m_name;
	tcu::DoubleFunc1&	m_func;
};

//<Signature<float, deFloat16, deFloat16> >
//<Signature<float, float, float> >
//<Signature<double, double, double> >
template <class T>
class FloatFunc2 : public PrimitiveFunc<T>
{
protected:
	Interval			doApply			(const EvalContext&	ctx, const typename Signature<typename T::Ret, typename T::Arg0, typename T::Arg1>::IArgs& iargs) const
	{
		return this->applyMonotone(ctx, iargs.a, iargs.b);
	}

	Interval			applyMonotone	(const EvalContext&	ctx,
										 const Interval&	xi,
										 const Interval&	yi) const
	{
		Interval reti;

		TCU_INTERVAL_APPLY_MONOTONE2(reti, x, xi, y, yi, ret,
									 TCU_SET_INTERVAL(ret, point,
													  point = this->applyPoint(ctx, x, y)));
		reti |= innerExtrema(ctx, xi, yi);
		reti &= (this->getCodomain(ctx) | TCU_NAN);

		return ctx.format.convert(reti);
	}

	virtual Interval	innerExtrema	(const EvalContext&,
										 const Interval&,
										 const Interval&) const
	{
		return Interval(); // empty interval, i.e. no extrema
	}

	virtual Interval	applyPoint		(const EvalContext&	ctx,
										 double				x,
										 double				y) const
	{
		const double exact	= this->applyExact(x, y);
		const double prec	= this->precision(ctx, exact, x, y);

		return exact + Interval(-prec, prec);
	}

	virtual double		applyExact		(double, double) const
	{
		TCU_THROW(InternalError, "Cannot apply");
	}

	virtual Interval	getCodomain		(const EvalContext&) const
	{
		return Interval::unbounded(true);
	}

	virtual double		precision		(const EvalContext&	ctx,
										 double				ret,
										 double				x,
										 double				y) const = 0;
};

template <class T>
class CFloatFunc2 : public FloatFunc2<T>
{
public:
						CFloatFunc2	(const string&		name,
									 tcu::DoubleFunc2&	func)
							: m_name(name)
							, m_func(func)
	{
	}

	string				getName		(void) const						{ return m_name; }

protected:
	double				applyExact	(double x, double y) const			{ return m_func(x, y); }

	const string		m_name;
	tcu::DoubleFunc2&	m_func;
};

template <class T>
class InfixOperator : public FloatFunc2<T>
{
protected:
	virtual string	getSymbol		(void) const = 0;

	void			doPrint			(ostream& os, const BaseArgExprs& args) const
	{
		os << "(" << *args[0] << " " << getSymbol() << " " << *args[1] << ")";
	}

	Interval		applyPoint		(const EvalContext&	ctx,
									 double				x,
									 double				y) const
	{
		const double exact	= this->applyExact(x, y);

		// Allow either representable number on both sides of the exact value,
		// but require exactly representable values to be preserved.
		return ctx.format.roundOut(exact, !deIsInf(x) && !deIsInf(y));
	}

	double			precision		(const EvalContext&, double, double, double) const
	{
		return 0.0;
	}
};

class InfixOperator16Bit : public FloatFunc2 <Signature<float, deFloat16, deFloat16> >
{
protected:
	virtual string	getSymbol		(void) const = 0;

	void			doPrint			(ostream& os, const BaseArgExprs& args) const
	{
		os << "(" << *args[0] << " " << getSymbol() << " " << *args[1] << ")";
	}

	Interval		applyPoint		(const EvalContext&	ctx,
									 double				x,
									 double				y) const
	{
		const double exact	= this->applyExact(x, y);

		// Allow either representable number on both sides of the exact value,
		// but require exactly representable values to be preserved.
		return ctx.format.roundOut(exact, !deIsInf(x) && !deIsInf(y));
	}

	double			precision		(const EvalContext&, double, double, double) const
	{
		return 0.0;
	}
};

template <class T>
class FloatFunc3 : public PrimitiveFunc<T>
{
protected:
	Interval			doApply			(const EvalContext&	ctx, const typename Signature<typename T::Ret, typename T::Arg0, typename T::Arg1, typename T::Arg2>::IArgs& iargs) const
	{
		return this->applyMonotone(ctx, iargs.a, iargs.b, iargs.c);
	}

	Interval			applyMonotone	(const EvalContext&	ctx,
										 const Interval&	xi,
										 const Interval&	yi,
										 const Interval&	zi) const
	{
		Interval reti;
		TCU_INTERVAL_APPLY_MONOTONE3(reti, x, xi, y, yi, z, zi, ret,
									 TCU_SET_INTERVAL(ret, point,
													  point = this->applyPoint(ctx, x, y, z)));
		return ctx.format.convert(reti);
	}

	virtual Interval	applyPoint		(const EvalContext&	ctx,
										 double				x,
										 double				y,
										 double				z) const
	{
		const double exact	= this->applyExact(x, y, z);
		const double prec	= this->precision(ctx, exact, x, y, z);
		return exact + Interval(-prec, prec);
	}

	virtual double		applyExact		(double, double, double) const
	{
		TCU_THROW(InternalError, "Cannot apply");
	}

	virtual double		precision		(const EvalContext&	ctx,
										 double				result,
										 double				x,
										 double				y,
										 double				z) const = 0;
};

// We define syntactic sugar functions for expression constructors. Since
// these have the same names as ordinary mathematical operations (sin, log
// etc.), it's better to give them a dedicated namespace.
namespace Functions
{

using namespace tcu;

template <class T>
class Comparison : public InfixOperator < T >
{
public:
	string		getName			(void) const	{ return "comparison"; }
	string		getSymbol		(void) const	{ return ""; }

	SpirVCaseT	getSpirvCase	() const		{ return SPIRV_CASETYPE_COMPARE; }

	Interval	doApply			(const EvalContext&						ctx,
								 const typename Comparison<T>::IArgs&	iargs) const
	{
		DE_UNREF(ctx);
		if (iargs.a.hasNaN() || iargs.b.hasNaN())
		{
			return TCU_NAN; // one of the floats is NaN: block analysis
		}

		int operationFlag = 1;
		int result = 0;
		const double a = iargs.a.midpoint();
		const double b = iargs.b.midpoint();

		for (int i = 0; i<2; ++i)
		{
			if (a == b)
				result += operationFlag;
			operationFlag = operationFlag << 1;

			if (a > b)
				result += operationFlag;
			operationFlag = operationFlag << 1;

			if (a < b)
				result += operationFlag;
			operationFlag = operationFlag << 1;

			if (a >= b)
				result += operationFlag;
			operationFlag = operationFlag << 1;

			if (a <= b)
				result += operationFlag;
			operationFlag = operationFlag << 1;
		}
		return result;
	}
};

template <class T>
class Add : public InfixOperator < T >
{
public:
	string		getName		(void) const						{ return "add"; }
	string		getSymbol	(void) const						{ return "+"; }

	Interval	doApply		(const EvalContext&	ctx,
							 const typename Signature<typename T::Ret, typename T::Arg0, typename T::Arg1>::IArgs& iargs) const
	{
		// Fast-path for common case
		if (iargs.a.isOrdinary(ctx.format.getMaxValue()) && iargs.b.isOrdinary(ctx.format.getMaxValue()))
		{
			Interval ret;
			TCU_SET_INTERVAL_BOUNDS(ret, sum,
									sum = iargs.a.lo() + iargs.b.lo(),
									sum = iargs.a.hi() + iargs.b.hi());
			return ctx.format.convert(ctx.format.roundOut(ret, true));
		}
		return this->applyMonotone(ctx, iargs.a, iargs.b);
	}

protected:
	double		applyExact	(double x, double y) const			{ return x + y; }
};

template<class T>
class Mul : public InfixOperator<T>
{
public:
	string		getName		(void) const									{ return "mul"; }
	string		getSymbol	(void) const									{ return "*"; }

	Interval	doApply		(const EvalContext&	ctx, const typename Signature<typename T::Ret, typename T::Arg0, typename T::Arg1>::IArgs& iargs) const
	{
		Interval a = iargs.a;
		Interval b = iargs.b;

		// Fast-path for common case
		if (a.isOrdinary(ctx.format.getMaxValue()) && b.isOrdinary(ctx.format.getMaxValue()))
		{
			Interval ret;
			if (a.hi() < 0)
			{
				a = -a;
				b = -b;
			}
			if (a.lo() >= 0 && b.lo() >= 0)
			{
				TCU_SET_INTERVAL_BOUNDS(ret, prod,
										prod = a.lo() * b.lo(),
										prod = a.hi() * b.hi());
				return ctx.format.convert(ctx.format.roundOut(ret, true));
			}
			if (a.lo() >= 0 && b.hi() <= 0)
			{
				TCU_SET_INTERVAL_BOUNDS(ret, prod,
										prod = a.hi() * b.lo(),
										prod = a.lo() * b.hi());
				return ctx.format.convert(ctx.format.roundOut(ret, true));
			}
		}
		return this->applyMonotone(ctx, iargs.a, iargs.b);
	}

protected:
	double		applyExact	(double x, double y) const						{ return x * y; }

	Interval	innerExtrema(const EvalContext&, const Interval& xi, const Interval& yi) const
	{
		if (((xi.contains(-TCU_INFINITY) || xi.contains(TCU_INFINITY)) && yi.contains(0.0)) ||
			((yi.contains(-TCU_INFINITY) || yi.contains(TCU_INFINITY)) && xi.contains(0.0)))
			return Interval(TCU_NAN);

		return Interval();
	}
};

template<class T>
class Sub : public InfixOperator <T>
{
public:
	string		getName		(void) const				{ return "sub"; }
	string		getSymbol	(void) const				{ return "-"; }

	Interval	doApply		(const EvalContext&	ctx, const typename Signature<typename T::Ret, typename T::Arg0, typename T::Arg1>::IArgs& iargs) const
	{
		// Fast-path for common case
		if (iargs.a.isOrdinary(ctx.format.getMaxValue()) && iargs.b.isOrdinary(ctx.format.getMaxValue()))
		{
			Interval ret;

			TCU_SET_INTERVAL_BOUNDS(ret, diff,
									diff = iargs.a.lo() - iargs.b.hi(),
									diff = iargs.a.hi() - iargs.b.lo());
			return ctx.format.convert(ctx.format.roundOut(ret, true));

		}
		else
		{
			return this->applyMonotone(ctx, iargs.a, iargs.b);
		}
	}

protected:
	double		applyExact	(double x, double y) const	{ return x - y; }
};

template <class T>
class Negate : public FloatFunc1<T>
{
public:
	string	getName		(void) const									{ return "_negate"; }
	void	doPrint		(ostream& os, const BaseArgExprs& args) const	{ os << "-" << *args[0]; }

protected:
	double	precision	(const EvalContext&, double, double) const		{ return 0.0; }
	double	applyExact	(double x) const								{ return -x; }
};

template <class T>
class Div : public InfixOperator<T>
{
public:
	string		getName			(void) const						{ return "div"; }

protected:
	string		getSymbol		(void) const						{ return "/"; }

	Interval	innerExtrema	(const EvalContext&,
								 const Interval&		nom,
								 const Interval&		den) const
	{
		Interval ret;

		if (den.contains(0.0))
		{
			if (nom.contains(0.0))
				ret |= TCU_NAN;

			if (nom.lo() < 0.0 || nom.hi() > 0.0)
				ret |= Interval::unbounded();
		}

		return ret;
	}

	double		applyExact		(double x, double y) const { return x / y; }

	Interval	applyPoint		(const EvalContext&	ctx, double x, double y) const
	{
		Interval ret = FloatFunc2<T>::applyPoint(ctx, x, y);

		if (!deIsInf(x) && !deIsInf(y) && y != 0.0)
		{
			const Interval dst = ctx.format.convert(ret);
			if (dst.contains(-TCU_INFINITY)) ret |= -ctx.format.getMaxValue();
			if (dst.contains(+TCU_INFINITY)) ret |= +ctx.format.getMaxValue();
		}

		return ret;
	}

	double		precision		(const EvalContext& ctx, double ret, double, double den) const
	{
		const FloatFormat&	fmt		= ctx.format;