glsBuiltinPrecisionTests.cpp revision d6148171f88da1301f053e2e0236afc69416137c
1/*-------------------------------------------------------------------------
2 * drawElements Quality Program OpenGL (ES) Module
3 * -----------------------------------------------
4 *
5 * Copyright 2014 The Android Open Source Project
6 *
7 * Licensed under the Apache License, Version 2.0 (the "License");
8 * you may not use this file except in compliance with the License.
9 * You may obtain a copy of the License at
10 *
11 *      http://www.apache.org/licenses/LICENSE-2.0
12 *
13 * Unless required by applicable law or agreed to in writing, software
14 * distributed under the License is distributed on an "AS IS" BASIS,
15 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
16 * See the License for the specific language governing permissions and
17 * limitations under the License.
18 *
19 *//*!
20 * \file
21 * \brief Precision and range tests for GLSL builtins and types.
22 *
23 *//*--------------------------------------------------------------------*/
24
25#include "glsBuiltinPrecisionTests.hpp"
26
27#include "deMath.h"
28#include "deMemory.h"
29#include "deDefs.hpp"
30#include "deRandom.hpp"
31#include "deSTLUtil.hpp"
32#include "deStringUtil.hpp"
33#include "deUniquePtr.hpp"
34#include "deSharedPtr.hpp"
35#include "deArrayUtil.hpp"
36
37#include "tcuCommandLine.hpp"
38#include "tcuFloatFormat.hpp"
39#include "tcuInterval.hpp"
40#include "tcuTestCase.hpp"
41#include "tcuTestLog.hpp"
42#include "tcuVector.hpp"
43#include "tcuMatrix.hpp"
44#include "tcuResultCollector.hpp"
45
46#include "gluContextInfo.hpp"
47#include "gluVarType.hpp"
48#include "gluRenderContext.hpp"
49#include "glwDefs.hpp"
50
51#include "glsShaderExecUtil.hpp"
52
53#include <cmath>
54#include <string>
55#include <sstream>
56#include <iostream>
57#include <map>
58#include <utility>
59
60// Uncomment this to get evaluation trace dumps to std::cerr
61// #define GLS_ENABLE_TRACE
62
63// set this to true to dump even passing results
64#define GLS_LOG_ALL_RESULTS false
65
66namespace deqp
67{
68namespace gls
69{
70namespace BuiltinPrecisionTests
71{
72
73using std::string;
74using std::map;
75using std::ostream;
76using std::ostringstream;
77using std::pair;
78using std::vector;
79using std::set;
80
81using de::MovePtr;
82using de::Random;
83using de::SharedPtr;
84using de::UniquePtr;
85using tcu::Interval;
86using tcu::FloatFormat;
87using tcu::MessageBuilder;
88using tcu::TestCase;
89using tcu::TestLog;
90using tcu::Vector;
91using tcu::Matrix;
92namespace matrix = tcu::matrix;
93using glu::Precision;
94using glu::RenderContext;
95using glu::VarType;
96using glu::DataType;
97using glu::ShaderType;
98using glu::ContextInfo;
99using gls::ShaderExecUtil::Symbol;
100
101typedef TestCase::IterateResult IterateResult;
102
103using namespace glw;
104using namespace tcu;
105
106/*--------------------------------------------------------------------*//*!
107 * \brief Generic singleton creator.
108 *
109 * instance<T>() returns a reference to a unique default-constructed instance
110 * of T. This is mainly used for our GLSL function implementations: each
111 * function is implemented by an object, and each of the objects has a
112 * distinct class. It would be extremely toilsome to maintain a separate
113 * context object that contained individual instances of the function classes,
114 * so we have to resort to global singleton instances.
115 *
116 *//*--------------------------------------------------------------------*/
117template <typename T>
118const T& instance (void)
119{
120	static const T s_instance = T();
121	return s_instance;
122}
123
124/*--------------------------------------------------------------------*//*!
125 * \brief Dummy placeholder type for unused template parameters.
126 *
127 * In the precision tests we are dealing with functions of different arities.
128 * To minimize code duplication, we only define templates with the maximum
129 * number of arguments, currently four. If a function's arity is less than the
130 * maximum, Void us used as the type for unused arguments.
131 *
132 * Although Voids are not used at run-time, they still must be compilable, so
133 * they must support all operations that other types do.
134 *
135 *//*--------------------------------------------------------------------*/
136struct Void
137{
138	typedef	Void		Element;
139	enum
140	{
141		SIZE = 0,
142	};
143
144	template <typename T>
145	explicit			Void			(const T&)		{}
146						Void			(void)			{}
147						operator double	(void)	const	{ return TCU_NAN; }
148
149	// These are used to make Voids usable as containers in container-generic code.
150	Void&				operator[]		(int)			{ return *this; }
151	const Void&			operator[]		(int)	const	{ return *this; }
152};
153
154ostream& operator<< (ostream& os, Void) { return os << "()"; }
155
156//! Returns true for all other types except Void
157template <typename T>	bool isTypeValid		(void)	{ return true;	}
158template <>				bool isTypeValid<Void>	(void)	{ return false;	}
159
160//! Utility function for getting the name of a data type.
161//! This is used in vector and matrix constructors.
162template <typename T>
163const char* dataTypeNameOf (void)
164{
165	return glu::getDataTypeName(glu::dataTypeOf<T>());
166}
167
168template <>
169const char* dataTypeNameOf<Void> (void)
170{
171	DE_ASSERT(!"Impossible");
172	return DE_NULL;
173}
174
175//! A hack to get Void support for VarType.
176template <typename T>
177VarType getVarTypeOf (Precision prec = glu::PRECISION_LAST)
178{
179	return glu::varTypeOf<T>(prec);
180}
181
182template <>
183VarType getVarTypeOf<Void> (Precision)
184{
185	DE_ASSERT(!"Impossible");
186	return VarType();
187}
188
189/*--------------------------------------------------------------------*//*!
190 * \brief Type traits for generalized interval types.
191 *
192 * We are trying to compute sets of acceptable values not only for
193 * float-valued expressions but also for compound values: vectors and
194 * matrices. We approximate a set of vectors as a vector of intervals and
195 * likewise for matrices.
196 *
197 * We now need generalized operations for each type and its interval
198 * approximation. These are given in the type Traits<T>.
199 *
200 * The type Traits<T>::IVal is the approximation of T: it is `Interval` for
201 * scalar types, and a vector or matrix of intervals for container types.
202 *
203 * To allow template inference to take place, there are function wrappers for
204 * the actual operations in Traits<T>. Hence we can just use:
205 *
206 * makeIVal(someFloat)
207 *
208 * instead of:
209 *
210 * Traits<float>::doMakeIVal(value)
211 *
212 *//*--------------------------------------------------------------------*/
213
214template <typename T> struct Traits;
215
216//! Create container from elementwise singleton values.
217template <typename T>
218typename Traits<T>::IVal makeIVal (const T& value)
219{
220	return Traits<T>::doMakeIVal(value);
221}
222
223//! Elementwise union of intervals.
224template <typename T>
225typename Traits<T>::IVal unionIVal (const typename Traits<T>::IVal& a,
226									const typename Traits<T>::IVal& b)
227{
228	return Traits<T>::doUnion(a, b);
229}
230
231//! Returns true iff every element of `ival` contains the corresponding element of `value`.
232template <typename T>
233bool contains (const typename Traits<T>::IVal& ival, const T& value)
234{
235	return Traits<T>::doContains(ival, value);
236}
237
238//! Print out an interval with the precision of `fmt`.
239template <typename T>
240void printIVal (const FloatFormat& fmt, const typename Traits<T>::IVal& ival, ostream& os)
241{
242	Traits<T>::doPrintIVal(fmt, ival, os);
243}
244
245template <typename T>
246string intervalToString (const FloatFormat& fmt, const typename Traits<T>::IVal& ival)
247{
248	ostringstream oss;
249	printIVal<T>(fmt, ival, oss);
250	return oss.str();
251}
252
253//! Print out a value with the precision of `fmt`.
254template <typename T>
255void printValue (const FloatFormat& fmt, const T& value, ostream& os)
256{
257	Traits<T>::doPrintValue(fmt, value, os);
258}
259
260template <typename T>
261string valueToString (const FloatFormat& fmt, const T& val)
262{
263	ostringstream oss;
264	printValue(fmt, val, oss);
265	return oss.str();
266}
267
268//! Approximate `value` elementwise to the float precision defined in `fmt`.
269//! The resulting interval might not be a singleton if rounding in both
270//! directions is allowed.
271template <typename T>
272typename Traits<T>::IVal round (const FloatFormat& fmt, const T& value)
273{
274	return Traits<T>::doRound(fmt, value);
275}
276
277template <typename T>
278typename Traits<T>::IVal convert (const FloatFormat&				fmt,
279								  const typename Traits<T>::IVal&	value)
280{
281	return Traits<T>::doConvert(fmt, value);
282}
283
284//! Common traits for scalar types.
285template <typename T>
286struct ScalarTraits
287{
288	typedef 			Interval		IVal;
289
290	static Interval		doMakeIVal		(const T& value)
291	{
292		// Thankfully all scalar types have a well-defined conversion to `double`,
293		// hence Interval can represent their ranges without problems.
294		return Interval(double(value));
295	}
296
297	static Interval		doUnion			(const Interval& a, const Interval& b)
298	{
299		return a | b;
300	}
301
302	static bool			doContains		(const Interval& a, T value)
303	{
304		return a.contains(double(value));
305	}
306
307	static Interval		doConvert		(const FloatFormat& fmt, const IVal& ival)
308	{
309		return fmt.convert(ival);
310	}
311
312	static Interval		doRound			(const FloatFormat& fmt, T value)
313	{
314		return fmt.roundOut(double(value), false);
315	}
316};
317
318template<>
319struct Traits<float> : ScalarTraits<float>
320{
321	static void			doPrintIVal		(const FloatFormat&	fmt,
322										 const Interval&	ival,
323										 ostream&			os)
324	{
325		os << fmt.intervalToHex(ival);
326	}
327
328	static void			doPrintValue	(const FloatFormat&	fmt,
329										 const float&		value,
330										 ostream&			os)
331	{
332		os << fmt.floatToHex(value);
333	}
334};
335
336template<>
337struct Traits<bool> : ScalarTraits<bool>
338{
339	static void			doPrintValue	(const FloatFormat&,
340										 const float&		value,
341										 ostream&			os)
342	{
343		os << (value ? "true" : "false");
344	}
345
346	static void			doPrintIVal		(const FloatFormat&,
347										 const Interval&	ival,
348										 ostream&			os)
349	{
350		os << "{";
351		if (ival.contains(false))
352			os << "false";
353		if (ival.contains(false) && ival.contains(true))
354			os << ", ";
355		if (ival.contains(true))
356			os << "true";
357		os << "}";
358	}
359};
360
361template<>
362struct Traits<int> : ScalarTraits<int>
363{
364	static void			doPrintValue 	(const FloatFormat&,
365										 const int&			value,
366										 ostream&			os)
367	{
368		os << value;
369	}
370
371	static void			doPrintIVal		(const FloatFormat&,
372										 const Interval&	ival,
373										 ostream&			os)
374	{
375		os << "[" << int(ival.lo()) << ", " << int(ival.hi()) << "]";
376	}
377};
378
379//! Common traits for containers, i.e. vectors and matrices.
380//! T is the container type itself, I is the same type with interval elements.
381template <typename T, typename I>
382struct ContainerTraits
383{
384	typedef typename	T::Element		Element;
385	typedef				I				IVal;
386
387	static IVal			doMakeIVal		(const T& value)
388	{
389		IVal ret;
390
391		for (int ndx = 0; ndx < T::SIZE; ++ndx)
392			ret[ndx] = makeIVal(value[ndx]);
393
394		return ret;
395	}
396
397	static IVal			doUnion			(const IVal& a, const IVal& b)
398	{
399		IVal ret;
400
401		for (int ndx = 0; ndx < T::SIZE; ++ndx)
402			ret[ndx] = unionIVal<Element>(a[ndx], b[ndx]);
403
404		return ret;
405	}
406
407	static bool			doContains		(const IVal& ival, const T& value)
408	{
409		for (int ndx = 0; ndx < T::SIZE; ++ndx)
410			if (!contains(ival[ndx], value[ndx]))
411				return false;
412
413		return true;
414	}
415
416	static void			doPrintIVal		(const FloatFormat& fmt, const IVal ival, ostream& os)
417	{
418		os << "(";
419
420		for (int ndx = 0; ndx < T::SIZE; ++ndx)
421		{
422			if (ndx > 0)
423				os << ", ";
424
425			printIVal<Element>(fmt, ival[ndx], os);
426		}
427
428		os << ")";
429	}
430
431	static void			doPrintValue	(const FloatFormat& fmt, const T& value, ostream& os)
432	{
433		os << dataTypeNameOf<T>() << "(";
434
435		for (int ndx = 0; ndx < T::SIZE; ++ndx)
436		{
437			if (ndx > 0)
438				os << ", ";
439
440			printValue<Element>(fmt, value[ndx], os);
441		}
442
443		os << ")";
444	}
445
446	static IVal			doConvert		(const FloatFormat& fmt, const IVal& value)
447	{
448		IVal ret;
449
450		for (int ndx = 0; ndx < T::SIZE; ++ndx)
451			ret[ndx] = convert<Element>(fmt, value[ndx]);
452
453		return ret;
454	}
455
456	static IVal			doRound			(const FloatFormat& fmt, T value)
457	{
458		IVal ret;
459
460		for (int ndx = 0; ndx < T::SIZE; ++ndx)
461			ret[ndx] = round(fmt, value[ndx]);
462
463		return ret;
464	}
465};
466
467template <typename T, int Size>
468struct Traits<Vector<T, Size> > :
469	ContainerTraits<Vector<T, Size>, Vector<typename Traits<T>::IVal, Size> >
470{
471};
472
473template <typename T, int Rows, int Cols>
474struct Traits<Matrix<T, Rows, Cols> > :
475	ContainerTraits<Matrix<T, Rows, Cols>, Matrix<typename Traits<T>::IVal, Rows, Cols> >
476{
477};
478
479//! Void traits. These are just dummies, but technically valid: a Void is a
480//! unit type with a single possible value.
481template<>
482struct Traits<Void>
483{
484	typedef		Void			IVal;
485
486	static Void	doMakeIVal		(const Void& value) 					{ return value; }
487	static Void	doUnion			(const Void&, const Void&)				{ return Void(); }
488	static bool	doContains		(const Void&, Void) 					{ return true; }
489	static Void	doRound			(const FloatFormat&, const Void& value)	{ return value; }
490	static Void	doConvert		(const FloatFormat&, const Void& value)	{ return value; }
491
492	static void	doPrintValue 	(const FloatFormat&, const Void&, ostream& os)
493	{
494		os << "()";
495	}
496
497	static void	doPrintIVal		(const FloatFormat&, const Void&, ostream& os)
498	{
499		os << "()";
500	}
501};
502
503//! This is needed for container-generic operations.
504//! We want a scalar type T to be its own "one-element vector".
505template <typename T, int Size> struct ContainerOf	{ typedef Vector<T, Size>	Container; };
506
507template <typename T>			struct ContainerOf<T, 1>		{ typedef T		Container; };
508template <int Size> 			struct ContainerOf<Void, Size>	{ typedef Void	Container; };
509
510// This is a kludge that is only needed to get the ExprP::operator[] syntactic sugar to work.
511template <typename T>	struct ElementOf		{ typedef	typename T::Element	Element; };
512template <>				struct ElementOf<float>	{ typedef	void				Element; };
513template <>				struct ElementOf<bool>	{ typedef	void				Element; };
514template <>				struct ElementOf<int>	{ typedef	void				Element; };
515
516/*--------------------------------------------------------------------*//*!
517 *
518 * \name Abstract syntax for expressions and statements.
519 *
520 * We represent GLSL programs as syntax objects: an Expr<T> represents an
521 * expression whose GLSL type corresponds to the C++ type T, and a Statement
522 * represents a statement.
523 *
524 * To ease memory management, we use shared pointers to refer to expressions
525 * and statements. ExprP<T> is a shared pointer to an Expr<T>, and StatementP
526 * is a shared pointer to a Statement.
527 *
528 * \{
529 *
530 *//*--------------------------------------------------------------------*/
531
532class ExprBase;
533class ExpandContext;
534class Statement;
535class StatementP;
536class FuncBase;
537template <typename T> class ExprP;
538template <typename T> class Variable;
539template <typename T> class VariableP;
540template <typename T> class DefaultSampling;
541
542typedef set<const FuncBase*> FuncSet;
543
544template <typename T>
545VariableP<T>	variable			(const string& name);
546StatementP		compoundStatement	(const vector<StatementP>& statements);
547
548/*--------------------------------------------------------------------*//*!
549 * \brief A variable environment.
550 *
551 * An Environment object maintains the mapping between variables of the
552 * abstract syntax tree and their values.
553 *
554 * \todo [2014-03-28 lauri] At least run-time type safety.
555 *
556 *//*--------------------------------------------------------------------*/
557class Environment
558{
559public:
560	template<typename T>
561	void						bind	(const Variable<T>&					variable,
562										 const typename Traits<T>::IVal&	value)
563	{
564		deUint8* const data = new deUint8[sizeof(value)];
565
566		deMemcpy(data, &value, sizeof(value));
567		de::insert(m_map, variable.getName(), SharedPtr<deUint8>(data));
568	}
569
570	template<typename T>
571	typename Traits<T>::IVal&	lookup	(const Variable<T>& variable) const
572	{
573		deUint8* const data = de::lookup(m_map, variable.getName()).get();
574
575		return *reinterpret_cast<typename Traits<T>::IVal*>(data);
576	}
577
578private:
579	map<string, SharedPtr<deUint8, de::ArrayDeleter<deUint8> > >	m_map;
580};
581
582/*--------------------------------------------------------------------*//*!
583 * \brief Evaluation context.
584 *
585 * The evaluation context contains everything that separates one execution of
586 * an expression from the next. Currently this means the desired floating
587 * point precision and the current variable environment.
588 *
589 *//*--------------------------------------------------------------------*/
590struct EvalContext
591{
592	EvalContext (const FloatFormat&	format_,
593				 Precision			floatPrecision_,
594				 Environment&		env_,
595				 int				callDepth_ = 0)
596		: format			(format_)
597		, floatPrecision	(floatPrecision_)
598		, env				(env_)
599		, callDepth			(callDepth_) {}
600
601	FloatFormat		format;
602	Precision		floatPrecision;
603	Environment&	env;
604	int				callDepth;
605};
606
607/*--------------------------------------------------------------------*//*!
608 * \brief Simple incremental counter.
609 *
610 * This is used to make sure that different ExpandContexts will not produce
611 * overlapping temporary names.
612 *
613 *//*--------------------------------------------------------------------*/
614class Counter
615{
616public:
617			Counter		(int count = 0) : m_count(count) {}
618	int		operator()	(void) { return m_count++; }
619
620private:
621	int		m_count;
622};
623
624class ExpandContext
625{
626public:
627						ExpandContext	(Counter& symCounter) : m_symCounter(symCounter) {}
628						ExpandContext	(const ExpandContext& parent)
629							: m_symCounter(parent.m_symCounter) {}
630
631	template<typename T>
632	VariableP<T>		genSym			(const string& baseName)
633	{
634		return variable<T>(baseName + de::toString(m_symCounter()));
635	}
636
637	void				addStatement	(const StatementP& stmt)
638	{
639		m_statements.push_back(stmt);
640	}
641
642	vector<StatementP>	getStatements	(void) const
643	{
644		return m_statements;
645	}
646private:
647	Counter&			m_symCounter;
648	vector<StatementP>	m_statements;
649};
650
651/*--------------------------------------------------------------------*//*!
652 * \brief A statement or declaration.
653 *
654 * Statements have no values. Instead, they are executed for their side
655 * effects only: the execute() method should modify at least one variable in
656 * the environment.
657 *
658 * As a bit of a kludge, a Statement object can also represent a declaration:
659 * when it is evaluated, it can add a variable binding to the environment
660 * instead of modifying a current one.
661 *
662 *//*--------------------------------------------------------------------*/
663class Statement
664{
665public:
666	virtual	~Statement		(void) 							{								 }
667	//! Execute the statement, modifying the environment of `ctx`
668	void	execute			(EvalContext&	ctx)	const	{ this->doExecute(ctx);			 }
669	void	print			(ostream&		os)		const	{ this->doPrint(os);			 }
670	//! Add the functions used in this statement to `dst`.
671	void	getUsedFuncs	(FuncSet& dst)			const	{ this->doGetUsedFuncs(dst);	 }
672
673protected:
674	virtual void	doPrint			(ostream& os)			const	= 0;
675	virtual void	doExecute		(EvalContext& ctx)		const	= 0;
676	virtual void	doGetUsedFuncs	(FuncSet& dst)			const	= 0;
677};
678
679ostream& operator<<(ostream& os, const Statement& stmt)
680{
681	stmt.print(os);
682	return os;
683}
684
685/*--------------------------------------------------------------------*//*!
686 * \brief Smart pointer for statements (and declarations)
687 *
688 *//*--------------------------------------------------------------------*/
689class StatementP : public SharedPtr<const Statement>
690{
691public:
692	typedef		SharedPtr<const Statement>	Super;
693
694				StatementP			(void) {}
695	explicit	StatementP			(const Statement* ptr)	: Super(ptr) {}
696				StatementP			(const Super& ptr)		: Super(ptr) {}
697};
698
699/*--------------------------------------------------------------------*//*!
700 * \brief
701 *
702 * A statement that modifies a variable or a declaration that binds a variable.
703 *
704 *//*--------------------------------------------------------------------*/
705template <typename T>
706class VariableStatement : public Statement
707{
708public:
709					VariableStatement	(const VariableP<T>& variable, const ExprP<T>& value,
710										 bool isDeclaration)
711						: m_variable		(variable)
712						, m_value			(value)
713						, m_isDeclaration	(isDeclaration) {}
714
715protected:
716	void			doPrint				(ostream& os)							const
717	{
718		if (m_isDeclaration)
719			os << glu::declare(getVarTypeOf<T>(), m_variable->getName());
720		else
721			os << m_variable->getName();
722
723		os << " = " << *m_value << ";\n";
724	}
725
726	void			doExecute			(EvalContext& ctx)						const
727	{
728		if (m_isDeclaration)
729			ctx.env.bind(*m_variable, m_value->evaluate(ctx));
730		else
731			ctx.env.lookup(*m_variable) = m_value->evaluate(ctx);
732	}
733
734	void			doGetUsedFuncs		(FuncSet& dst)							const
735	{
736		m_value->getUsedFuncs(dst);
737	}
738
739	VariableP<T>	m_variable;
740	ExprP<T>		m_value;
741	bool			m_isDeclaration;
742};
743
744template <typename T>
745StatementP variableStatement (const VariableP<T>&	variable,
746							  const ExprP<T>&		value,
747							  bool					isDeclaration)
748{
749	return StatementP(new VariableStatement<T>(variable, value, isDeclaration));
750}
751
752template <typename T>
753StatementP variableDeclaration (const VariableP<T>& variable, const ExprP<T>& definiens)
754{
755	return variableStatement(variable, definiens, true);
756}
757
758template <typename T>
759StatementP variableAssignment (const VariableP<T>& variable, const ExprP<T>& value)
760{
761	return variableStatement(variable, value, false);
762}
763
764/*--------------------------------------------------------------------*//*!
765 * \brief A compound statement, i.e. a block.
766 *
767 * A compound statement is executed by executing its constituent statements in
768 * sequence.
769 *
770 *//*--------------------------------------------------------------------*/
771class CompoundStatement : public Statement
772{
773public:
774						CompoundStatement	(const vector<StatementP>& statements)
775							: m_statements	(statements) {}
776
777protected:
778	void				doPrint				(ostream&		os)						const
779	{
780		os << "{\n";
781
782		for (size_t ndx = 0; ndx < m_statements.size(); ++ndx)
783			os << *m_statements[ndx];
784
785		os << "}\n";
786	}
787
788	void				doExecute			(EvalContext&	ctx)					const
789	{
790		for (size_t ndx = 0; ndx < m_statements.size(); ++ndx)
791			m_statements[ndx]->execute(ctx);
792	}
793
794	void				doGetUsedFuncs		(FuncSet& dst)							const
795	{
796		for (size_t ndx = 0; ndx < m_statements.size(); ++ndx)
797			m_statements[ndx]->getUsedFuncs(dst);
798	}
799
800	vector<StatementP>	m_statements;
801};
802
803StatementP compoundStatement(const vector<StatementP>& statements)
804{
805	return StatementP(new CompoundStatement(statements));
806}
807
808//! Common base class for all expressions regardless of their type.
809class ExprBase
810{
811public:
812	virtual				~ExprBase		(void)									{}
813	void				printExpr		(ostream& os) const { this->doPrintExpr(os); }
814
815	//! Output the functions that this expression refers to
816	void				getUsedFuncs	(FuncSet& dst) const
817	{
818		this->doGetUsedFuncs(dst);
819	}
820
821protected:
822	virtual void		doPrintExpr		(ostream&)	const	{}
823	virtual void		doGetUsedFuncs	(FuncSet&)	const	{}
824};
825
826//! Type-specific operations for an expression representing type T.
827template <typename T>
828class Expr : public ExprBase
829{
830public:
831	typedef 			T				Val;
832	typedef typename 	Traits<T>::IVal	IVal;
833
834	IVal				evaluate		(const EvalContext&	ctx) const;
835
836protected:
837	virtual IVal		doEvaluate		(const EvalContext&	ctx) const = 0;
838};
839
840//! Evaluate an expression with the given context, optionally tracing the calls to stderr.
841template <typename T>
842typename Traits<T>::IVal Expr<T>::evaluate (const EvalContext& ctx) const
843{
844#ifdef GLS_ENABLE_TRACE
845	static const FloatFormat	highpFmt	(-126, 127, 23, true,
846											 tcu::MAYBE,
847											 tcu::YES,
848											 tcu::MAYBE);
849	EvalContext					newCtx		(ctx.format, ctx.floatPrecision,
850											 ctx.env, ctx.callDepth + 1);
851	const IVal					ret			= this->doEvaluate(newCtx);
852
853	if (isTypeValid<T>())
854	{
855		std::cerr << string(ctx.callDepth, ' ');
856		this->printExpr(std::cerr);
857		std::cerr << " -> " << intervalToString<T>(highpFmt, ret) << std::endl;
858	}
859	return ret;
860#else
861	return this->doEvaluate(ctx);
862#endif
863}
864
865template <typename T>
866class ExprPBase : public SharedPtr<const Expr<T> >
867{
868public:
869};
870
871ostream& operator<< (ostream& os, const ExprBase& expr)
872{
873	expr.printExpr(os);
874	return os;
875}
876
877/*--------------------------------------------------------------------*//*!
878 * \brief Shared pointer to an expression of a container type.
879 *
880 * Container types (i.e. vectors and matrices) support the subscription
881 * operator. This class provides a bit of syntactic sugar to allow us to use
882 * the C++ subscription operator to create a subscription expression.
883 *//*--------------------------------------------------------------------*/
884template <typename T>
885class ContainerExprPBase : public ExprPBase<T>
886{
887public:
888	ExprP<typename T::Element>	operator[]	(int i) const;
889};
890
891template <typename T>
892class ExprP : public ExprPBase<T> {};
893
894// We treat Voids as containers since the dummy parameters in generalized
895// vector functions are represented as Voids.
896template <>
897class ExprP<Void> : public ContainerExprPBase<Void> {};
898
899template <typename T, int Size>
900class ExprP<Vector<T, Size> > : public ContainerExprPBase<Vector<T, Size> > {};
901
902template <typename T, int Rows, int Cols>
903class ExprP<Matrix<T, Rows, Cols> > : public ContainerExprPBase<Matrix<T, Rows, Cols> > {};
904
905template <typename T> ExprP<T> exprP (void)
906{
907	return ExprP<T>();
908}
909
910template <typename T>
911ExprP<T> exprP (const SharedPtr<const Expr<T> >& ptr)
912{
913	ExprP<T> ret;
914	static_cast<SharedPtr<const Expr<T> >&>(ret) = ptr;
915	return ret;
916}
917
918template <typename T>
919ExprP<T> exprP (const Expr<T>* ptr)
920{
921	return exprP(SharedPtr<const Expr<T> >(ptr));
922}
923
924/*--------------------------------------------------------------------*//*!
925 * \brief A shared pointer to a variable expression.
926 *
927 * This is just a narrowing of ExprP for the operations that require a variable
928 * instead of an arbitrary expression.
929 *
930 *//*--------------------------------------------------------------------*/
931template <typename T>
932class VariableP : public SharedPtr<const Variable<T> >
933{
934public:
935	typedef		SharedPtr<const Variable<T> >	Super;
936	explicit	VariableP	(const Variable<T>* ptr) : Super(ptr) {}
937				VariableP	(void) {}
938				VariableP	(const Super& ptr) : Super(ptr) {}
939
940	operator	ExprP<T>	(void) const { return exprP(SharedPtr<const Expr<T> >(*this)); }
941};
942
943/*--------------------------------------------------------------------*//*!
944 * \name Syntactic sugar operators for expressions.
945 *
946 * @{
947 *
948 * These operators allow the use of C++ syntax to construct GLSL expressions
949 * containing operators: e.g. "a+b" creates an addition expression with
950 * operands a and b, and so on.
951 *
952 *//*--------------------------------------------------------------------*/
953ExprP<float>						operator-(const ExprP<float>&						arg0);
954ExprP<float>						operator+(const ExprP<float>&						arg0,
955											  const ExprP<float>&						arg1);
956ExprP<float>						operator-(const ExprP<float>&						arg0,
957											  const ExprP<float>&						arg1);
958ExprP<float> 						operator*(const ExprP<float>&						arg0,
959											  const ExprP<float>&						arg1);
960ExprP<float> 						operator/(const ExprP<float>&						arg0,
961											  const ExprP<float>&						arg1);
962template<int Size>
963ExprP<Vector<float, Size> >			operator-(const ExprP<Vector<float, Size> >&		arg0);
964template<int Size>
965ExprP<Vector<float, Size> >			operator*(const ExprP<Vector<float, Size> >&		arg0,
966											  const ExprP<float>&						arg1);
967template<int Size>
968ExprP<Vector<float, Size> >			operator*(const ExprP<Vector<float, Size> >&		arg0,
969											  const ExprP<Vector<float, Size> >&		arg1);
970template<int Size>
971ExprP<Vector<float, Size> >			operator-(const ExprP<Vector<float, Size> >&		arg0,
972											  const ExprP<Vector<float, Size> >&		arg1);
973template<int Left, int Mid, int Right>
974ExprP<Matrix<float, Left, Right> >	operator* (const ExprP<Matrix<float, Left, Mid> >&	left,
975											   const ExprP<Matrix<float, Mid, Right> >&	right);
976template<int Rows, int Cols>
977ExprP<Vector<float, Rows> >			operator* (const ExprP<Vector<float, Cols> >&		left,
978											   const ExprP<Matrix<float, Rows, Cols> >&	right);
979template<int Rows, int Cols>
980ExprP<Vector<float, Cols> >			operator* (const ExprP<Matrix<float, Rows, Cols> >&	left,
981											   const ExprP<Vector<float, Rows> >&		right);
982template<int Rows, int Cols>
983ExprP<Matrix<float, Rows, Cols> >	operator* (const ExprP<Matrix<float, Rows, Cols> >&	left,
984											   const ExprP<float>&						right);
985template<int Rows, int Cols>
986ExprP<Matrix<float, Rows, Cols> > 	operator+ (const ExprP<Matrix<float, Rows, Cols> >&	left,
987											   const ExprP<Matrix<float, Rows, Cols> >&	right);
988template<int Rows, int Cols>
989ExprP<Matrix<float, Rows, Cols> > 	operator- (const ExprP<Matrix<float, Rows, Cols> >&	mat);
990
991//! @}
992
993/*--------------------------------------------------------------------*//*!
994 * \brief Variable expression.
995 *
996 * A variable is evaluated by looking up its range of possible values from an
997 * environment.
998 *//*--------------------------------------------------------------------*/
999template <typename T>
1000class Variable : public Expr<T>
1001{
1002public:
1003	typedef typename Expr<T>::IVal IVal;
1004
1005					Variable	(const string& name) : m_name (name) {}
1006	string			getName		(void)							const { return m_name; }
1007
1008protected:
1009	void			doPrintExpr	(ostream& os)					const { os << m_name; }
1010	IVal			doEvaluate	(const EvalContext& ctx)		const
1011	{
1012		return ctx.env.lookup<T>(*this);
1013	}
1014
1015private:
1016	string	m_name;
1017};
1018
1019template <typename T>
1020VariableP<T> variable (const string& name)
1021{
1022	return VariableP<T>(new Variable<T>(name));
1023}
1024
1025template <typename T>
1026VariableP<T> bindExpression (const string& name, ExpandContext& ctx, const ExprP<T>& expr)
1027{
1028	VariableP<T> var = ctx.genSym<T>(name);
1029	ctx.addStatement(variableDeclaration(var, expr));
1030	return var;
1031}
1032
1033/*--------------------------------------------------------------------*//*!
1034 * \brief Constant expression.
1035 *
1036 * A constant is evaluated by rounding it to a set of possible values allowed
1037 * by the current floating point precision.
1038 *//*--------------------------------------------------------------------*/
1039template <typename T>
1040class Constant : public Expr<T>
1041{
1042public:
1043	typedef typename Expr<T>::IVal IVal;
1044
1045			Constant		(const T& value) : m_value(value) {}
1046
1047protected:
1048	void	doPrintExpr		(ostream& os) const 		{ os << m_value; }
1049	IVal	doEvaluate		(const EvalContext&) const	{ return makeIVal(m_value); }
1050
1051private:
1052	T		m_value;
1053};
1054
1055template <typename T>
1056ExprP<T> constant (const T& value)
1057{
1058	return exprP(new Constant<T>(value));
1059}
1060
1061//! Return a reference to a singleton void constant.
1062const ExprP<Void>& voidP (void)
1063{
1064	static const ExprP<Void> singleton = constant(Void());
1065
1066	return singleton;
1067}
1068
1069/*--------------------------------------------------------------------*//*!
1070 * \brief Four-element tuple.
1071 *
1072 * This is used for various things where we need one thing for each possible
1073 * function parameter. Currently the maximum supported number of parameters is
1074 * four.
1075 *//*--------------------------------------------------------------------*/
1076template <typename T0 = Void, typename T1 = Void, typename T2 = Void, typename T3 = Void>
1077struct Tuple4
1078{
1079	explicit Tuple4 (const T0& e0 = T0(),
1080					 const T1& e1 = T1(),
1081					 const T2& e2 = T2(),
1082					 const T3& e3 = T3())
1083		: a	(e0)
1084		, b	(e1)
1085		, c	(e2)
1086		, d	(e3)
1087	{
1088	}
1089
1090	T0 a;
1091	T1 b;
1092	T2 c;
1093	T3 d;
1094};
1095
1096/*--------------------------------------------------------------------*//*!
1097 * \brief Function signature.
1098 *
1099 * This is a purely compile-time structure used to bundle all types in a
1100 * function signature together. This makes passing the signature around in
1101 * templates easier, since we only need to take and pass a single Sig instead
1102 * of a bunch of parameter types and a return type.
1103 *
1104 *//*--------------------------------------------------------------------*/
1105template <typename R,
1106		  typename P0 = Void, typename P1 = Void,
1107		  typename P2 = Void, typename P3 = Void>
1108struct Signature
1109{
1110	typedef R							Ret;
1111	typedef P0							Arg0;
1112	typedef P1							Arg1;
1113	typedef P2							Arg2;
1114	typedef P3							Arg3;
1115	typedef typename Traits<Ret>::IVal	IRet;
1116	typedef typename Traits<Arg0>::IVal	IArg0;
1117	typedef typename Traits<Arg1>::IVal	IArg1;
1118	typedef typename Traits<Arg2>::IVal	IArg2;
1119	typedef typename Traits<Arg3>::IVal	IArg3;
1120
1121	typedef Tuple4<	const Arg0&,	const Arg1&,	const Arg2&,	const Arg3&>	Args;
1122	typedef Tuple4<	const IArg0&,	const IArg1&,	const IArg2&,	const IArg3&> 	IArgs;
1123	typedef Tuple4<	ExprP<Arg0>,	ExprP<Arg1>,	ExprP<Arg2>,	ExprP<Arg3> >	ArgExprs;
1124};
1125
1126typedef vector<const ExprBase*> BaseArgExprs;
1127
1128/*--------------------------------------------------------------------*//*!
1129 * \brief Type-independent operations for function objects.
1130 *
1131 *//*--------------------------------------------------------------------*/
1132class FuncBase
1133{
1134public:
1135	virtual			~FuncBase				(void)					{}
1136	virtual string	getName					(void) 					const = 0;
1137	//! Name of extension that this function requires, or empty.
1138	virtual string	getRequiredExtension	(void) 					const { return ""; }
1139	virtual void	print					(ostream&,
1140											 const BaseArgExprs&)	const = 0;
1141	//! Index of output parameter, or -1 if none of the parameters is output.
1142	virtual int		getOutParamIndex		(void)					const { return -1; }
1143
1144	void			printDefinition			(ostream& os)			const
1145	{
1146		doPrintDefinition(os);
1147	}
1148
1149	void				getUsedFuncs		(FuncSet& dst) const
1150	{
1151		this->doGetUsedFuncs(dst);
1152	}
1153
1154protected:
1155	virtual void	doPrintDefinition		(ostream& os)			const = 0;
1156	virtual void	doGetUsedFuncs			(FuncSet& dst)			const = 0;
1157};
1158
1159typedef Tuple4<string, string, string, string> ParamNames;
1160
1161/*--------------------------------------------------------------------*//*!
1162 * \brief Function objects.
1163 *
1164 * Each Func object represents a GLSL function. It can be applied to interval
1165 * arguments, and it returns the an interval that is a conservative
1166 * approximation of the image of the GLSL function over the argument
1167 * intervals. That is, it is given a set of possible arguments and it returns
1168 * the set of possible values.
1169 *
1170 *//*--------------------------------------------------------------------*/
1171template <typename Sig_>
1172class Func : public FuncBase
1173{
1174public:
1175	typedef Sig_										Sig;
1176	typedef typename Sig::Ret							Ret;
1177	typedef typename Sig::Arg0							Arg0;
1178	typedef typename Sig::Arg1							Arg1;
1179	typedef typename Sig::Arg2							Arg2;
1180	typedef typename Sig::Arg3							Arg3;
1181	typedef typename Sig::IRet							IRet;
1182	typedef typename Sig::IArg0							IArg0;
1183	typedef typename Sig::IArg1							IArg1;
1184	typedef typename Sig::IArg2							IArg2;
1185	typedef typename Sig::IArg3							IArg3;
1186	typedef typename Sig::Args							Args;
1187	typedef typename Sig::IArgs							IArgs;
1188	typedef typename Sig::ArgExprs						ArgExprs;
1189
1190	void				print			(ostream&			os,
1191										 const BaseArgExprs& args)				const
1192	{
1193		this->doPrint(os, args);
1194	}
1195
1196	IRet				apply			(const EvalContext&	ctx,
1197										 const IArg0&		arg0 = IArg0(),
1198										 const IArg1&		arg1 = IArg1(),
1199										 const IArg2&		arg2 = IArg2(),
1200										 const IArg3&		arg3 = IArg3())		const
1201	{
1202		return this->applyArgs(ctx, IArgs(arg0, arg1, arg2, arg3));
1203	}
1204	IRet				applyArgs		(const EvalContext&	ctx,
1205										 const IArgs&		args)				const
1206	{
1207		return this->doApply(ctx, args);
1208	}
1209	ExprP<Ret>			operator()		(const ExprP<Arg0>&		arg0 = voidP(),
1210										 const ExprP<Arg1>&		arg1 = voidP(),
1211										 const ExprP<Arg2>&		arg2 = voidP(),
1212										 const ExprP<Arg3>&		arg3 = voidP())		const;
1213
1214	const ParamNames&	getParamNames	(void)									const
1215	{
1216		return this->doGetParamNames();
1217	}
1218
1219protected:
1220	virtual IRet		doApply			(const EvalContext&,
1221										 const IArgs&)							const = 0;
1222	virtual void		doPrint			(ostream& os, const BaseArgExprs& args)	const
1223	{
1224		os << getName() << "(";
1225
1226		if (isTypeValid<Arg0>())
1227			os << *args[0];
1228
1229		if (isTypeValid<Arg1>())
1230			os << ", " << *args[1];
1231
1232		if (isTypeValid<Arg2>())
1233			os << ", " << *args[2];
1234
1235		if (isTypeValid<Arg3>())
1236			os << ", " << *args[3];
1237
1238		os << ")";
1239	}
1240
1241	virtual const ParamNames&	doGetParamNames	(void)							const
1242	{
1243		static ParamNames	names	("a", "b", "c", "d");
1244		return names;
1245	}
1246};
1247
1248template <typename Sig>
1249class Apply : public Expr<typename Sig::Ret>
1250{
1251public:
1252	typedef typename Sig::Ret				Ret;
1253	typedef typename Sig::Arg0				Arg0;
1254	typedef typename Sig::Arg1				Arg1;
1255	typedef typename Sig::Arg2				Arg2;
1256	typedef typename Sig::Arg3				Arg3;
1257	typedef typename Expr<Ret>::Val			Val;
1258	typedef typename Expr<Ret>::IVal		IVal;
1259	typedef Func<Sig>						ApplyFunc;
1260	typedef typename ApplyFunc::ArgExprs	ArgExprs;
1261
1262						Apply	(const ApplyFunc&		func,
1263								 const ExprP<Arg0>&		arg0 = voidP(),
1264								 const ExprP<Arg1>&		arg1 = voidP(),
1265								 const ExprP<Arg2>&		arg2 = voidP(),
1266								 const ExprP<Arg3>&		arg3 = voidP())
1267							: m_func	(func),
1268							  m_args	(arg0, arg1, arg2, arg3) {}
1269
1270						Apply	(const ApplyFunc&	func,
1271								 const ArgExprs&	args)
1272							: m_func	(func),
1273							  m_args	(args) {}
1274protected:
1275	void				doPrintExpr			(ostream& os) const
1276	{
1277		BaseArgExprs	args;
1278		args.push_back(m_args.a.get());
1279		args.push_back(m_args.b.get());
1280		args.push_back(m_args.c.get());
1281		args.push_back(m_args.d.get());
1282		m_func.print(os, args);
1283	}
1284
1285	IVal				doEvaluate		(const EvalContext& ctx) const
1286	{
1287		return m_func.apply(ctx,
1288							m_args.a->evaluate(ctx), m_args.b->evaluate(ctx),
1289							m_args.c->evaluate(ctx), m_args.d->evaluate(ctx));
1290	}
1291
1292	void				doGetUsedFuncs	(FuncSet& dst) const
1293	{
1294		m_func.getUsedFuncs(dst);
1295		m_args.a->getUsedFuncs(dst);
1296		m_args.b->getUsedFuncs(dst);
1297		m_args.c->getUsedFuncs(dst);
1298		m_args.d->getUsedFuncs(dst);
1299	}
1300
1301	const ApplyFunc&	m_func;
1302	ArgExprs			m_args;
1303};
1304
1305template <typename Sig>
1306ExprP<typename Sig::Ret> createApply (const Func<Sig>&						func,
1307									  const typename Func<Sig>::ArgExprs&	args)
1308{
1309	return exprP(new Apply<Sig>(func, args));
1310}
1311
1312template <typename Sig>
1313ExprP<typename Sig::Ret> createApply (
1314	const Func<Sig>&			func,
1315	const ExprP<typename Sig::Arg0>&	arg0 = voidP(),
1316	const ExprP<typename Sig::Arg1>&	arg1 = voidP(),
1317	const ExprP<typename Sig::Arg2>&	arg2 = voidP(),
1318	const ExprP<typename Sig::Arg3>&	arg3 = voidP())
1319{
1320	return exprP(new Apply<Sig>(func, arg0, arg1, arg2, arg3));
1321}
1322
1323template <typename Sig>
1324ExprP<typename Sig::Ret> Func<Sig>::operator() (const ExprP<typename Sig::Arg0>& arg0,
1325												const ExprP<typename Sig::Arg1>& arg1,
1326												const ExprP<typename Sig::Arg2>& arg2,
1327												const ExprP<typename Sig::Arg3>& arg3) const
1328{
1329	return createApply(*this, arg0, arg1, arg2, arg3);
1330}
1331
1332template <typename F>
1333ExprP<typename F::Ret> app (const ExprP<typename F::Arg0>& arg0 = voidP(),
1334							const ExprP<typename F::Arg1>& arg1 = voidP(),
1335							const ExprP<typename F::Arg2>& arg2 = voidP(),
1336							const ExprP<typename F::Arg3>& arg3 = voidP())
1337{
1338	return createApply(instance<F>(), arg0, arg1, arg2, arg3);
1339}
1340
1341template <typename F>
1342typename F::IRet call (const EvalContext&			ctx,
1343					   const typename F::IArg0&		arg0 = Void(),
1344					   const typename F::IArg1&		arg1 = Void(),
1345					   const typename F::IArg2&		arg2 = Void(),
1346					   const typename F::IArg3&		arg3 = Void())
1347{
1348	return instance<F>().apply(ctx, arg0, arg1, arg2, arg3);
1349}
1350
1351template <typename Sig>
1352class ApplyVar : public Apply<Sig>
1353{
1354public:
1355	typedef typename Sig::Ret				Ret;
1356	typedef typename Sig::Arg0				Arg0;
1357	typedef typename Sig::Arg1				Arg1;
1358	typedef typename Sig::Arg2				Arg2;
1359	typedef typename Sig::Arg3				Arg3;
1360	typedef typename Expr<Ret>::Val			Val;
1361	typedef typename Expr<Ret>::IVal		IVal;
1362	typedef Func<Sig>						ApplyFunc;
1363	typedef typename ApplyFunc::ArgExprs	ArgExprs;
1364
1365						ApplyVar	(const ApplyFunc&			func,
1366									 const VariableP<Arg0>&		arg0,
1367									 const VariableP<Arg1>&		arg1,
1368									 const VariableP<Arg2>&		arg2,
1369									 const VariableP<Arg3>&		arg3)
1370							: Apply<Sig> (func, arg0, arg1, arg2, arg3) {}
1371protected:
1372	IVal				doEvaluate		(const EvalContext& ctx) const
1373	{
1374		const Variable<Arg0>&	var0 = static_cast<const Variable<Arg0>&>(*this->m_args.a);
1375		const Variable<Arg1>&	var1 = static_cast<const Variable<Arg1>&>(*this->m_args.b);
1376		const Variable<Arg2>&	var2 = static_cast<const Variable<Arg2>&>(*this->m_args.c);
1377		const Variable<Arg3>&	var3 = static_cast<const Variable<Arg3>&>(*this->m_args.d);
1378		return this->m_func.apply(ctx,
1379								  ctx.env.lookup(var0), ctx.env.lookup(var1),
1380								  ctx.env.lookup(var2), ctx.env.lookup(var3));
1381	}
1382};
1383
1384template <typename Sig>
1385ExprP<typename Sig::Ret> applyVar (const Func<Sig>&						func,
1386								   const VariableP<typename Sig::Arg0>&	arg0,
1387								   const VariableP<typename Sig::Arg1>&	arg1,
1388								   const VariableP<typename Sig::Arg2>&	arg2,
1389								   const VariableP<typename Sig::Arg3>&	arg3)
1390{
1391	return exprP(new ApplyVar<Sig>(func, arg0, arg1, arg2, arg3));
1392}
1393
1394template <typename Sig_>
1395class DerivedFunc : public Func<Sig_>
1396{
1397public:
1398	typedef typename DerivedFunc::ArgExprs		ArgExprs;
1399	typedef typename DerivedFunc::IRet			IRet;
1400	typedef typename DerivedFunc::IArgs			IArgs;
1401	typedef typename DerivedFunc::Ret			Ret;
1402	typedef typename DerivedFunc::Arg0			Arg0;
1403	typedef typename DerivedFunc::Arg1			Arg1;
1404	typedef typename DerivedFunc::Arg2			Arg2;
1405	typedef typename DerivedFunc::Arg3			Arg3;
1406	typedef typename DerivedFunc::IArg0			IArg0;
1407	typedef typename DerivedFunc::IArg1			IArg1;
1408	typedef typename DerivedFunc::IArg2			IArg2;
1409	typedef typename DerivedFunc::IArg3			IArg3;
1410
1411protected:
1412	void						doPrintDefinition	(ostream& os) const
1413	{
1414		const ParamNames&	paramNames	= this->getParamNames();
1415
1416		initialize();
1417
1418		os << dataTypeNameOf<Ret>() << " " << this->getName()
1419			<< "(";
1420		if (isTypeValid<Arg0>())
1421			os << dataTypeNameOf<Arg0>() << " " << paramNames.a;
1422		if (isTypeValid<Arg1>())
1423			os << ", " << dataTypeNameOf<Arg1>() << " " << paramNames.b;
1424		if (isTypeValid<Arg2>())
1425			os << ", " << dataTypeNameOf<Arg2>() << " " << paramNames.c;
1426		if (isTypeValid<Arg3>())
1427			os << ", " << dataTypeNameOf<Arg3>() << " " << paramNames.d;
1428		os << ")\n{\n";
1429
1430		for (size_t ndx = 0; ndx < m_body.size(); ++ndx)
1431			os << *m_body[ndx];
1432		os << "return " << *m_ret << ";\n";
1433		os << "}\n";
1434	}
1435
1436	IRet						doApply			(const EvalContext&	ctx,
1437												 const IArgs&		args) const
1438	{
1439		Environment	funEnv;
1440		IArgs&		mutArgs		= const_cast<IArgs&>(args);
1441		IRet		ret;
1442
1443		initialize();
1444
1445		funEnv.bind(*m_var0, args.a);
1446		funEnv.bind(*m_var1, args.b);
1447		funEnv.bind(*m_var2, args.c);
1448		funEnv.bind(*m_var3, args.d);
1449
1450		{
1451			EvalContext	funCtx(ctx.format, ctx.floatPrecision, funEnv, ctx.callDepth);
1452
1453			for (size_t ndx = 0; ndx < m_body.size(); ++ndx)
1454				m_body[ndx]->execute(funCtx);
1455
1456			ret = m_ret->evaluate(funCtx);
1457		}
1458
1459		// \todo [lauri] Store references instead of values in environment
1460		const_cast<IArg0&>(mutArgs.a) = funEnv.lookup(*m_var0);
1461		const_cast<IArg1&>(mutArgs.b) = funEnv.lookup(*m_var1);
1462		const_cast<IArg2&>(mutArgs.c) = funEnv.lookup(*m_var2);
1463		const_cast<IArg3&>(mutArgs.d) = funEnv.lookup(*m_var3);
1464
1465		return ret;
1466	}
1467
1468	void						doGetUsedFuncs	(FuncSet& dst) const
1469	{
1470		initialize();
1471		if (dst.insert(this).second)
1472		{
1473			for (size_t ndx = 0; ndx < m_body.size(); ++ndx)
1474				m_body[ndx]->getUsedFuncs(dst);
1475			m_ret->getUsedFuncs(dst);
1476		}
1477	}
1478
1479	virtual ExprP<Ret>			doExpand		(ExpandContext& ctx, const ArgExprs& args_) const = 0;
1480
1481	// These are transparently initialized when first needed. They cannot be
1482	// initialized in the constructor because they depend on the doExpand
1483	// method of the subclass.
1484
1485	mutable VariableP<Arg0>		m_var0;
1486	mutable VariableP<Arg1>		m_var1;
1487	mutable VariableP<Arg2>		m_var2;
1488	mutable VariableP<Arg3>		m_var3;
1489	mutable vector<StatementP>	m_body;
1490	mutable ExprP<Ret>			m_ret;
1491
1492private:
1493
1494	void				initialize		(void)	const
1495	{
1496		if (!m_ret)
1497		{
1498			const ParamNames&	paramNames	= this->getParamNames();
1499			Counter				symCounter;
1500			ExpandContext		ctx			(symCounter);
1501			ArgExprs			args;
1502
1503			args.a	= m_var0 = variable<Arg0>(paramNames.a);
1504			args.b	= m_var1 = variable<Arg1>(paramNames.b);
1505			args.c	= m_var2 = variable<Arg2>(paramNames.c);
1506			args.d	= m_var3 = variable<Arg3>(paramNames.d);
1507
1508			m_ret	= this->doExpand(ctx, args);
1509			m_body	= ctx.getStatements();
1510		}
1511	}
1512};
1513
1514template <typename Sig>
1515class PrimitiveFunc : public Func<Sig>
1516{
1517public:
1518	typedef typename PrimitiveFunc::Ret			Ret;
1519	typedef typename PrimitiveFunc::ArgExprs	ArgExprs;
1520
1521protected:
1522	void	doPrintDefinition	(ostream&) const	{}
1523	void	doGetUsedFuncs		(FuncSet&) const	{}
1524};
1525
1526template <typename T>
1527class Cond : public PrimitiveFunc<Signature<T, bool, T, T> >
1528{
1529public:
1530	typedef typename Cond::IArgs	IArgs;
1531	typedef typename Cond::IRet		IRet;
1532
1533	string	getName	(void) const
1534	{
1535		return "_cond";
1536	}
1537
1538protected:
1539
1540	void	doPrint	(ostream& os, const BaseArgExprs& args) const
1541	{
1542		os << "(" << *args[0] << " ? " << *args[1] << " : " << *args[2] << ")";
1543	}
1544
1545	IRet	doApply	(const EvalContext&, const IArgs& iargs)const
1546	{
1547		IRet	ret;
1548
1549		if (iargs.a.contains(true))
1550			ret = unionIVal<T>(ret, iargs.b);
1551
1552		if (iargs.a.contains(false))
1553			ret = unionIVal<T>(ret, iargs.c);
1554
1555		return ret;
1556	}
1557};
1558
1559template <typename T>
1560class CompareOperator : public PrimitiveFunc<Signature<bool, T, T> >
1561{
1562public:
1563	typedef typename CompareOperator::IArgs	IArgs;
1564	typedef typename CompareOperator::IArg0	IArg0;
1565	typedef typename CompareOperator::IArg1	IArg1;
1566	typedef typename CompareOperator::IRet	IRet;
1567
1568protected:
1569	void			doPrint	(ostream& os, const BaseArgExprs& args) const
1570	{
1571		os << "(" << *args[0] << getSymbol() << *args[1] << ")";
1572	}
1573
1574	Interval		doApply	(const EvalContext&, const IArgs& iargs) const
1575	{
1576		const IArg0&	arg0 = iargs.a;
1577		const IArg1&	arg1 = iargs.b;
1578		IRet	ret;
1579
1580		if (canSucceed(arg0, arg1))
1581			ret |= true;
1582		if (canFail(arg0, arg1))
1583			ret |= false;
1584
1585		return ret;
1586	}
1587
1588	virtual string	getSymbol	(void) const = 0;
1589	virtual bool	canSucceed	(const IArg0&, const IArg1&) const = 0;
1590	virtual bool	canFail		(const IArg0&, const IArg1&) const = 0;
1591};
1592
1593template <typename T>
1594class LessThan : public CompareOperator<T>
1595{
1596public:
1597	string	getName		(void) const									{ return "lessThan"; }
1598
1599protected:
1600	string	getSymbol	(void) const									{ return "<";		}
1601
1602	bool	canSucceed	(const Interval& a, const Interval& b) const
1603	{
1604		return (a.lo() < b.hi());
1605	}
1606
1607	bool	canFail		(const Interval& a, const Interval& b) const
1608	{
1609		return !(a.hi() < b.lo());
1610	}
1611};
1612
1613template <typename T>
1614ExprP<bool> operator< (const ExprP<T>& a, const ExprP<T>& b)
1615{
1616	return app<LessThan<T> >(a, b);
1617}
1618
1619template <typename T>
1620ExprP<T> cond (const ExprP<bool>&	test,
1621			   const ExprP<T>&		consequent,
1622			   const ExprP<T>&		alternative)
1623{
1624	return app<Cond<T> >(test, consequent, alternative);
1625}
1626
1627/*--------------------------------------------------------------------*//*!
1628 *
1629 * @}
1630 *
1631 *//*--------------------------------------------------------------------*/
1632
1633class FloatFunc1 : public PrimitiveFunc<Signature<float, float> >
1634{
1635protected:
1636	Interval			doApply			(const EvalContext& ctx, const IArgs& iargs) const
1637	{
1638		return this->applyMonotone(ctx, iargs.a);
1639	}
1640
1641	Interval			applyMonotone	(const EvalContext& ctx, const Interval& iarg0) const
1642	{
1643		Interval ret;
1644
1645		TCU_INTERVAL_APPLY_MONOTONE1(ret, arg0, iarg0, val,
1646									 TCU_SET_INTERVAL(val, point,
1647													  point = this->applyPoint(ctx, arg0)));
1648
1649		ret |= innerExtrema(ctx, iarg0);
1650		ret &= (this->getCodomain() | TCU_NAN);
1651
1652		return ctx.format.convert(ret);
1653	}
1654
1655	virtual Interval	innerExtrema	(const EvalContext&, const Interval&) const
1656	{
1657		return Interval(); // empty interval, i.e. no extrema
1658	}
1659
1660	virtual Interval	applyPoint		(const EvalContext& ctx, double arg0) const
1661	{
1662		const double	exact	= this->applyExact(arg0);
1663		const double	prec	= this->precision(ctx, exact, arg0);
1664
1665		return exact + Interval(-prec, prec);
1666	}
1667
1668	virtual double		applyExact		(double) const
1669	{
1670		TCU_THROW(InternalError, "Cannot apply");
1671	}
1672
1673	virtual Interval	getCodomain		(void) const
1674	{
1675		return Interval::unbounded(true);
1676	}
1677
1678	virtual double		precision		(const EvalContext& ctx, double, double) const = 0;
1679};
1680
1681class CFloatFunc1 : public FloatFunc1
1682{
1683public:
1684			CFloatFunc1	(const string& name, DoubleFunc1& func)
1685				: m_name(name), m_func(func) {}
1686
1687	string			getName		(void) const		{ return m_name; }
1688
1689protected:
1690	double			applyExact	(double x) const	{ return m_func(x); }
1691
1692	const string	m_name;
1693	DoubleFunc1&	m_func;
1694};
1695
1696class FloatFunc2 : public PrimitiveFunc<Signature<float, float, float> >
1697{
1698protected:
1699	Interval			doApply			(const EvalContext&	ctx, const IArgs& iargs) const
1700	{
1701		return this->applyMonotone(ctx, iargs.a, iargs.b);
1702	}
1703
1704	Interval			applyMonotone	(const EvalContext&	ctx,
1705										 const Interval&	xi,
1706										 const Interval&	yi) const
1707	{
1708		Interval reti;
1709
1710		TCU_INTERVAL_APPLY_MONOTONE2(reti, x, xi, y, yi, ret,
1711									 TCU_SET_INTERVAL(ret, point,
1712													  point = this->applyPoint(ctx, x, y)));
1713		reti |= innerExtrema(ctx, xi, yi);
1714		reti &= (this->getCodomain() | TCU_NAN);
1715
1716		return ctx.format.convert(reti);
1717	}
1718
1719	virtual Interval	innerExtrema	(const EvalContext&,
1720										 const Interval&,
1721										 const Interval&) const
1722	{
1723		return Interval(); // empty interval, i.e. no extrema
1724	}
1725
1726	virtual Interval	applyPoint		(const EvalContext&	ctx,
1727										 double 			x,
1728										 double 			y) const
1729	{
1730		const double exact	= this->applyExact(x, y);
1731		const double prec	= this->precision(ctx, exact, x, y);
1732
1733		return exact + Interval(-prec, prec);
1734	}
1735
1736	virtual double		applyExact		(double, double) const
1737	{
1738		TCU_THROW(InternalError, "Cannot apply");
1739	}
1740
1741	virtual Interval	getCodomain		(void) const
1742	{
1743		return Interval::unbounded(true);
1744	}
1745
1746	virtual double		precision		(const EvalContext&	ctx,
1747										 double				ret,
1748										 double				x,
1749										 double				y) const = 0;
1750};
1751
1752class CFloatFunc2 : public FloatFunc2
1753{
1754public:
1755					CFloatFunc2	(const string&	name,
1756								 DoubleFunc2&	func)
1757						: m_name(name)
1758						, m_func(func)
1759	{
1760	}
1761
1762	string			getName		(void) const						{ return m_name; }
1763
1764protected:
1765	double			applyExact	(double x, double y) const			{ return m_func(x, y); }
1766
1767	const string	m_name;
1768	DoubleFunc2&	m_func;
1769};
1770
1771class InfixOperator : public FloatFunc2
1772{
1773protected:
1774	virtual string	getSymbol		(void) const = 0;
1775
1776	void			doPrint			(ostream& os, const BaseArgExprs& args) const
1777	{
1778		os << "(" << *args[0] << " " << getSymbol() << " " << *args[1] << ")";
1779	}
1780
1781	Interval		applyPoint		(const EvalContext&	ctx,
1782									 double 			x,
1783									 double 			y) const
1784	{
1785		const double exact	= this->applyExact(x, y);
1786
1787		// Allow either representable number on both sides of the exact value,
1788		// but require exactly representable values to be preserved.
1789		return ctx.format.roundOut(exact, !deIsInf(x) && !deIsInf(y));
1790	}
1791
1792	double			precision		(const EvalContext&, double, double, double) const
1793	{
1794		return 0.0;
1795	}
1796};
1797
1798class FloatFunc3 : public PrimitiveFunc<Signature<float, float, float, float> >
1799{
1800protected:
1801	Interval			doApply			(const EvalContext&	ctx, const IArgs& iargs) const
1802	{
1803		return this->applyMonotone(ctx, iargs.a, iargs.b, iargs.c);
1804	}
1805
1806	Interval			applyMonotone	(const EvalContext&	ctx,
1807										 const Interval&	xi,
1808										 const Interval&	yi,
1809										 const Interval&	zi) const
1810	{
1811		Interval reti;
1812		TCU_INTERVAL_APPLY_MONOTONE3(reti, x, xi, y, yi, z, zi, ret,
1813									 TCU_SET_INTERVAL(ret, point,
1814													  point = this->applyPoint(ctx, x, y, z)));
1815		return ctx.format.convert(reti);
1816	}
1817
1818	virtual Interval	applyPoint		(const EvalContext&	ctx,
1819										 double 			x,
1820										 double 			y,
1821										 double 			z) const
1822	{
1823		const double exact	= this->applyExact(x, y, z);
1824		const double prec	= this->precision(ctx, exact, x, y, z);
1825		return exact + Interval(-prec, prec);
1826	}
1827
1828	virtual double		applyExact		(double, double, double) const
1829	{
1830		TCU_THROW(InternalError, "Cannot apply");
1831	}
1832
1833	virtual double		precision		(const EvalContext&	ctx,
1834										 double				result,
1835										 double				x,
1836										 double				y,
1837										 double				z) const = 0;
1838};
1839
1840// We define syntactic sugar functions for expression constructors. Since
1841// these have the same names as ordinary mathematical operations (sin, log
1842// etc.), it's better to give them a dedicated namespace.
1843namespace Functions
1844{
1845
1846using namespace tcu;
1847
1848class Add : public InfixOperator
1849{
1850public:
1851	string		getName		(void) const 						{ return "add"; }
1852	string		getSymbol	(void) const 						{ return "+"; }
1853
1854	Interval	doApply		(const EvalContext&	ctx,
1855							 const IArgs&		iargs) const
1856	{
1857		// Fast-path for common case
1858		if (iargs.a.isOrdinary() && iargs.b.isOrdinary())
1859		{
1860			Interval ret;
1861			TCU_SET_INTERVAL_BOUNDS(ret, sum,
1862									sum = iargs.a.lo() + iargs.b.lo(),
1863									sum = iargs.a.hi() + iargs.b.hi());
1864			return ctx.format.convert(ctx.format.roundOut(ret, true));
1865		}
1866		return this->applyMonotone(ctx, iargs.a, iargs.b);
1867	}
1868
1869protected:
1870	double		applyExact	(double x, double y) const 			{ return x + y; }
1871};
1872
1873class Mul : public InfixOperator
1874{
1875public:
1876	string		getName		(void) const 									{ return "mul"; }
1877	string		getSymbol	(void) const 									{ return "*"; }
1878
1879	Interval	doApply		(const EvalContext&	ctx, const IArgs& iargs) const
1880	{
1881		Interval a = iargs.a;
1882		Interval b = iargs.b;
1883
1884		// Fast-path for common case
1885		if (a.isOrdinary() && b.isOrdinary())
1886		{
1887			Interval ret;
1888			if (a.hi() < 0)
1889			{
1890				a = -a;
1891				b = -b;
1892			}
1893			if (a.lo() >= 0 && b.lo() >= 0)
1894			{
1895				TCU_SET_INTERVAL_BOUNDS(ret, prod,
1896										prod = iargs.a.lo() * iargs.b.lo(),
1897										prod = iargs.a.hi() * iargs.b.hi());
1898				return ctx.format.convert(ctx.format.roundOut(ret, true));
1899			}
1900			if (a.lo() >= 0 && b.hi() <= 0)
1901			{
1902				TCU_SET_INTERVAL_BOUNDS(ret, prod,
1903										prod = iargs.a.hi() * iargs.b.lo(),
1904										prod = iargs.a.lo() * iargs.b.hi());
1905				return ctx.format.convert(ctx.format.roundOut(ret, true));
1906			}
1907		}
1908		return this->applyMonotone(ctx, iargs.a, iargs.b);
1909	}
1910
1911protected:
1912	double		applyExact	(double x, double y) const						{ return x * y; }
1913
1914	Interval	innerExtrema(const EvalContext&, const Interval& xi, const Interval& yi) const
1915	{
1916		if (((xi.contains(-TCU_INFINITY) || xi.contains(TCU_INFINITY)) && yi.contains(0.0)) ||
1917			((yi.contains(-TCU_INFINITY) || yi.contains(TCU_INFINITY)) && xi.contains(0.0)))
1918			return Interval(TCU_NAN);
1919
1920		return Interval();
1921	}
1922};
1923
1924class Sub : public InfixOperator
1925{
1926public:
1927	string		getName		(void) const 				{ return "sub"; }
1928	string		getSymbol	(void) const 				{ return "-"; }
1929
1930	Interval	doApply		(const EvalContext&	ctx, const IArgs& iargs) const
1931	{
1932		// Fast-path for common case
1933		if (iargs.a.isOrdinary() && iargs.b.isOrdinary())
1934		{
1935			Interval ret;
1936
1937			TCU_SET_INTERVAL_BOUNDS(ret, diff,
1938									diff = iargs.a.lo() - iargs.b.hi(),
1939									diff = iargs.a.hi() - iargs.b.lo());
1940			return ctx.format.convert(ctx.format.roundOut(ret, true));
1941
1942		}
1943		else
1944		{
1945			return this->applyMonotone(ctx, iargs.a, iargs.b);
1946		}
1947	}
1948
1949protected:
1950	double		applyExact	(double x, double y) const	{ return x - y; }
1951};
1952
1953class Negate : public FloatFunc1
1954{
1955public:
1956	string	getName		(void) const									{ return "_negate"; }
1957	void	doPrint		(ostream& os, const BaseArgExprs& args) const	{ os << "-" << *args[0]; }
1958
1959protected:
1960	double	precision	(const EvalContext&, double, double) const		{ return 0.0; }
1961	double	applyExact	(double x) const								{ return -x; }
1962};
1963
1964class Div : public InfixOperator
1965{
1966public:
1967	string		getName			(void) const 						{ return "div"; }
1968
1969protected:
1970	string		getSymbol		(void) const 						{ return "/"; }
1971
1972	Interval	innerExtrema	(const EvalContext&,
1973								 const Interval&		nom,
1974								 const Interval&		den) const
1975	{
1976		Interval ret;
1977
1978		if (den.contains(0.0))
1979		{
1980			if (nom.contains(0.0))
1981				ret |= TCU_NAN;
1982
1983			if (nom.lo() < 0.0 || nom.hi() > 0.0)
1984				ret |= Interval::unbounded();
1985		}
1986
1987		return ret;
1988	}
1989
1990protected:
1991
1992	double		applyExact	(double x, double y) const { return x / y; }
1993
1994	Interval	applyPoint	(const EvalContext&	ctx, double x, double y) const
1995	{
1996		return FloatFunc2::applyPoint(ctx, x, y);
1997	}
1998
1999	double		precision	(const EvalContext& ctx, double ret, double, double den) const
2000	{
2001		const FloatFormat&	fmt		= ctx.format;
2002
2003		// \todo [2014-03-05 lauri] Check that the limits in GLSL 3.10 are actually correct.
2004		// For now, we assume that division's precision is 2.5 ULP when the value is within
2005		// [2^MINEXP, 2^MAXEXP-1]
2006
2007		if (den == 0.0)
2008			return 0.0; // Result must be exactly inf
2009		else if (de::inBounds(deAbs(den),
2010							  deLdExp(1.0, fmt.getMinExp()),
2011							  deLdExp(1.0, fmt.getMaxExp() - 1)))
2012			return fmt.ulp(ret, 2.5);
2013		else
2014			return TCU_INFINITY; // Can be any number, but must be a number.
2015	}
2016};
2017
2018class InverseSqrt : public FloatFunc1
2019{
2020public:
2021	string		getName		(void) const							{ return "inversesqrt"; }
2022
2023protected:
2024	double		applyExact	(double x) const						{ return 1.0 / deSqrt(x); }
2025
2026	double		precision	(const EvalContext& ctx, double ret, double x) const
2027	{
2028		return x <= 0 ? TCU_NAN : ctx.format.ulp(ret, 2.0);
2029	}
2030
2031	Interval	getCodomain	(void) const
2032	{
2033		return Interval(0.0, TCU_INFINITY);
2034	}
2035};
2036
2037class ExpFunc : public CFloatFunc1
2038{
2039public:
2040				ExpFunc		(const string& name, DoubleFunc1& func)
2041					: CFloatFunc1(name, func) {}
2042protected:
2043	double		precision	(const EvalContext& ctx, double ret, double x) const
2044	{
2045		switch (ctx.floatPrecision)
2046		{
2047			case glu::PRECISION_HIGHP:
2048				return ctx.format.ulp(ret, 3.0 + 2.0 * deAbs(x));
2049			case glu::PRECISION_MEDIUMP:
2050				return ctx.format.ulp(ret, 2.0 + 2.0 * deAbs(x));
2051			case glu::PRECISION_LOWP:
2052				return ctx.format.ulp(ret, 2.0);
2053			default:
2054				DE_ASSERT(!"Impossible");
2055		}
2056		return 0;
2057	}
2058
2059	Interval	getCodomain	(void) const
2060	{
2061		return Interval(0.0, TCU_INFINITY);
2062	}
2063};
2064
2065class Exp2	: public ExpFunc	{ public: Exp2 (void)	: ExpFunc("exp2", deExp2) {} };
2066class Exp	: public ExpFunc	{ public: Exp (void)	: ExpFunc("exp", deExp) {} };
2067
2068ExprP<float> exp2	(const ExprP<float>& x)	{ return app<Exp2>(x); }
2069ExprP<float> exp	(const ExprP<float>& x)	{ return app<Exp>(x); }
2070
2071class LogFunc : public CFloatFunc1
2072{
2073public:
2074				LogFunc		(const string& name, DoubleFunc1& func)
2075					: CFloatFunc1(name, func) {}
2076
2077protected:
2078	double		precision	(const EvalContext& ctx, double ret, double x) const
2079	{
2080		if (x <= 0)
2081			return TCU_NAN;
2082
2083		switch (ctx.floatPrecision)
2084		{
2085			case glu::PRECISION_HIGHP:
2086				return (0.5 <= x && x <= 2.0) ? deLdExp(1.0, -21) : ctx.format.ulp(ret, 3.0);
2087			case glu::PRECISION_MEDIUMP:
2088				return (0.5 <= x && x <= 2.0) ? deLdExp(1.0, -7) : ctx.format.ulp(ret, 2.0);
2089			case glu::PRECISION_LOWP:
2090				return ctx.format.ulp(ret, 2.0);
2091			default:
2092				DE_ASSERT(!"Impossible");
2093		}
2094
2095		return 0;
2096	}
2097};
2098
2099class Log2	: public LogFunc		{ public: Log2	(void) : LogFunc("log2", deLog2) {} };
2100class Log	: public LogFunc		{ public: Log	(void) : LogFunc("log", deLog) {} };
2101
2102ExprP<float> log2	(const ExprP<float>& x)	{ return app<Log2>(x); }
2103ExprP<float> log	(const ExprP<float>& x)	{ return app<Log>(x); }
2104
2105#define DEFINE_CONSTRUCTOR1(CLASS, TRET, NAME, T0) \
2106ExprP<TRET> NAME (const ExprP<T0>& arg0) { return app<CLASS>(arg0); }
2107
2108#define DEFINE_DERIVED1(CLASS, TRET, NAME, T0, ARG0, EXPANSION)			\
2109class CLASS : public DerivedFunc<Signature<TRET, T0> >		 			\
2110{																		\
2111public:																	\
2112	string			getName		(void) const		{ return #NAME; }	\
2113																		\
2114protected:																\
2115	ExprP<TRET>		doExpand		(ExpandContext&,					\
2116									 const CLASS::ArgExprs& args_) const \
2117	{																	\
2118		const ExprP<float>& ARG0 = args_.a;								\
2119		return EXPANSION;												\
2120	}																	\
2121};																		\
2122DEFINE_CONSTRUCTOR1(CLASS, TRET, NAME, T0)
2123
2124#define DEFINE_DERIVED_FLOAT1(CLASS, NAME, ARG0, EXPANSION) \
2125	DEFINE_DERIVED1(CLASS, float, NAME, float, ARG0, EXPANSION)
2126
2127#define DEFINE_CONSTRUCTOR2(CLASS, TRET, NAME, T0, T1)				\
2128ExprP<TRET> NAME (const ExprP<T0>& arg0, const ExprP<T1>& arg1)		\
2129{																	\
2130	return app<CLASS>(arg0, arg1);									\
2131}
2132
2133#define DEFINE_DERIVED2(CLASS, TRET, NAME, T0, Arg0, T1, Arg1, EXPANSION) \
2134class CLASS : public DerivedFunc<Signature<TRET, T0, T1> >		 		\
2135{																		\
2136public:																	\
2137	string			getName		(void) const		{ return #NAME; }	\
2138																		\
2139protected:																\
2140	ExprP<TRET>		doExpand	(ExpandContext&, const ArgExprs& args_) const \
2141	{																	\
2142		const ExprP<T0>& Arg0 = args_.a;								\
2143		const ExprP<T1>& Arg1 = args_.b;								\
2144		return EXPANSION;												\
2145	}																	\
2146};																		\
2147DEFINE_CONSTRUCTOR2(CLASS, TRET, NAME, T0, T1)
2148
2149#define DEFINE_DERIVED_FLOAT2(CLASS, NAME, Arg0, Arg1, EXPANSION)		\
2150	DEFINE_DERIVED2(CLASS, float, NAME, float, Arg0, float, Arg1, EXPANSION)
2151
2152#define DEFINE_CONSTRUCTOR3(CLASS, TRET, NAME, T0, T1, T2)				\
2153ExprP<TRET> NAME (const ExprP<T0>& arg0, const ExprP<T1>& arg1, const ExprP<T2>& arg2) \
2154{																		\
2155	return app<CLASS>(arg0, arg1, arg2);								\
2156}
2157
2158#define DEFINE_DERIVED3(CLASS, TRET, NAME, T0, ARG0, T1, ARG1, T2, ARG2, EXPANSION) \
2159class CLASS : public DerivedFunc<Signature<TRET, T0, T1, T2> >					\
2160{																				\
2161public:																			\
2162	string			getName		(void) const	{ return #NAME; }				\
2163																				\
2164protected:																		\
2165	ExprP<TRET>		doExpand	(ExpandContext&, const ArgExprs& args_) const	\
2166	{																			\
2167		const ExprP<T0>& ARG0 = args_.a;										\
2168		const ExprP<T1>& ARG1 = args_.b;										\
2169		const ExprP<T2>& ARG2 = args_.c;										\
2170		return EXPANSION;														\
2171	}																			\
2172};																				\
2173DEFINE_CONSTRUCTOR3(CLASS, TRET, NAME, T0, T1, T2)
2174
2175#define DEFINE_DERIVED_FLOAT3(CLASS, NAME, ARG0, ARG1, ARG2, EXPANSION)			\
2176	DEFINE_DERIVED3(CLASS, float, NAME, float, ARG0, float, ARG1, float, ARG2, EXPANSION)
2177
2178#define DEFINE_CONSTRUCTOR4(CLASS, TRET, NAME, T0, T1, T2, T3)			\
2179ExprP<TRET> NAME (const ExprP<T0>& arg0, const ExprP<T1>& arg1,			\
2180				  const ExprP<T2>& arg2, const ExprP<T3>& arg3)			\
2181{																		\
2182	return app<CLASS>(arg0, arg1, arg2, arg3);							\
2183}
2184
2185DEFINE_DERIVED_FLOAT1(Sqrt,		sqrt,		x, 		constant(1.0f) / app<InverseSqrt>(x));
2186DEFINE_DERIVED_FLOAT2(Pow,		pow,		x,	y,	exp2(y * log2(x)));
2187DEFINE_DERIVED_FLOAT1(Radians,	radians,	d, 		(constant(DE_PI) / constant(180.0f)) * d);
2188DEFINE_DERIVED_FLOAT1(Degrees,	degrees,	r,	 	(constant(180.0f) / constant(DE_PI)) * r);
2189
2190class TrigFunc : public CFloatFunc1
2191{
2192public:
2193					TrigFunc		(const string&		name,
2194									 DoubleFunc1&		func,
2195									 const Interval&	loEx,
2196									 const Interval&	hiEx)
2197						: CFloatFunc1	(name, func)
2198						, m_loExtremum	(loEx)
2199						, m_hiExtremum	(hiEx) {}
2200
2201protected:
2202	Interval		innerExtrema	(const EvalContext&, const Interval& angle) const
2203	{
2204		const double		lo		= angle.lo();
2205		const double		hi		= angle.hi();
2206		const int			loSlope	= doGetSlope(lo);
2207		const int			hiSlope	= doGetSlope(hi);
2208
2209		// Detect the high and low values the function can take between the
2210		// interval endpoints.
2211		if (angle.length() >= 2.0 * DE_PI_DOUBLE)
2212		{
2213			// The interval is longer than a full cycle, so it must get all possible values.
2214			return m_hiExtremum | m_loExtremum;
2215		}
2216		else if (loSlope == 1 && hiSlope == -1)
2217		{
2218			// The slope can change from positive to negative only at the maximum value.
2219			return m_hiExtremum;
2220		}
2221		else if (loSlope == -1 && hiSlope == 1)
2222		{
2223			// The slope can change from negative to positive only at the maximum value.
2224			return m_loExtremum;
2225		}
2226		else if (loSlope == hiSlope &&
2227				 deIntSign(applyExact(hi) - applyExact(lo)) * loSlope == -1)
2228		{
2229			// The slope has changed twice between the endpoints, so both extrema are included.
2230			return m_hiExtremum | m_loExtremum;
2231		}
2232
2233		return Interval();
2234	}
2235
2236	Interval	getCodomain			(void) const
2237	{
2238		// Ensure that result is always within [-1, 1], or NaN (for +-inf)
2239		return Interval(-1.0, 1.0) | TCU_NAN;
2240	}
2241
2242	double		precision			(const EvalContext& ctx, double ret, double arg) const
2243	{
2244		if (ctx.floatPrecision == glu::PRECISION_HIGHP)
2245		{
2246			// Use precision from OpenCL fast relaxed math
2247			if (-DE_PI_DOUBLE <= arg && arg <= DE_PI_DOUBLE)
2248			{
2249				return deLdExp(1.0, -11);
2250			}
2251			else
2252			{
2253				// "larger otherwise", let's pick |x| * 2^-12 , which is slightly over
2254				// 2^-11 at x == pi.
2255				return deLdExp(deAbs(arg), -12);
2256			}
2257		}
2258		else
2259		{
2260			// from OpenCL half-float extension specification
2261			return ctx.format.ulp(ret, 2.0);
2262		}
2263	}
2264
2265	virtual int		doGetSlope		(double angle) const = 0;
2266
2267	Interval		m_loExtremum;
2268	Interval		m_hiExtremum;
2269};
2270
2271class Sin : public TrigFunc
2272{
2273public:
2274				Sin			(void) : TrigFunc("sin", deSin, -1.0, 1.0) {}
2275
2276protected:
2277	int			doGetSlope	(double angle) const { return deIntSign(deCos(angle)); }
2278};
2279
2280ExprP<float> sin (const ExprP<float>& x) { return app<Sin>(x); }
2281
2282class Cos : public TrigFunc
2283{
2284public:
2285				Cos			(void) : TrigFunc("cos", deCos, -1.0, 1.0) {}
2286
2287protected:
2288	int			doGetSlope	(double angle) const { return -deIntSign(deSin(angle)); }
2289};
2290
2291ExprP<float> cos (const ExprP<float>& x) { return app<Cos>(x); }
2292
2293DEFINE_DERIVED_FLOAT1(Tan, tan, x, sin(x) * (constant(1.0f) / cos(x)));
2294
2295class ArcTrigFunc : public CFloatFunc1
2296{
2297public:
2298					ArcTrigFunc	(const string&		name,
2299								 DoubleFunc1&		func,
2300								 double				precisionULPs,
2301								 const Interval&	domain,
2302								 const Interval&	codomain)
2303						: CFloatFunc1		(name, func)
2304						, m_precision		(precisionULPs)
2305						, m_domain			(domain)
2306						, m_codomain		(codomain) {}
2307
2308protected:
2309	double			precision	(const EvalContext& ctx, double ret, double x) const
2310	{
2311		if (!m_domain.contains(x))
2312			return TCU_NAN;
2313
2314		if (ctx.floatPrecision == glu::PRECISION_HIGHP)
2315		{
2316			// Use OpenCL's precision
2317			return ctx.format.ulp(ret, m_precision);
2318		}
2319		else
2320		{
2321			// Use OpenCL half-float spec
2322			return ctx.format.ulp(ret, 2.0);
2323		}
2324	}
2325
2326	// We could implement getCodomain with m_codomain, but choose not to,
2327	// because it seems too strict with trascendental constants like pi.
2328
2329	const double	m_precision;
2330	const Interval	m_domain;
2331	const Interval	m_codomain;
2332};
2333
2334class ASin : public ArcTrigFunc
2335{
2336public:
2337	ASin (void) : ArcTrigFunc("asin", deAsin, 4.0,
2338							  Interval(-1.0, 1.0),
2339							  Interval(-DE_PI_DOUBLE * 0.5, DE_PI_DOUBLE * 0.5)) {}
2340};
2341
2342class ACos : public ArcTrigFunc
2343{
2344public:
2345	ACos (void) : ArcTrigFunc("acos", deAcos, 4.0,
2346							  Interval(-1.0, 1.0),
2347							  Interval(0.0, DE_PI_DOUBLE)) {}
2348};
2349
2350class ATan : public ArcTrigFunc
2351{
2352public:
2353	ATan (void) : ArcTrigFunc("atan", deAtanOver, 5.0,
2354							  Interval::unbounded(),
2355							  Interval(-DE_PI_DOUBLE * 0.5, DE_PI_DOUBLE * 0.5)) {}
2356};
2357
2358class ATan2 : public CFloatFunc2
2359{
2360public:
2361				ATan2			(void) : CFloatFunc2 ("atan", deAtan2) {}
2362
2363protected:
2364	Interval	innerExtrema	(const EvalContext&,
2365								 const Interval&		yi,
2366								 const Interval& 		xi) const
2367	{
2368		Interval ret;
2369
2370		if (yi.contains(0.0))
2371		{
2372			if (xi.contains(0.0))
2373				ret |= TCU_NAN;
2374			if (xi.intersects(Interval(-TCU_INFINITY, 0.0)))
2375				ret |= Interval(-DE_PI_DOUBLE, DE_PI_DOUBLE);
2376		}
2377
2378		return ret;
2379	}
2380
2381	double		precision		(const EvalContext& ctx, double ret, double, double) const
2382	{
2383		if (ctx.floatPrecision == glu::PRECISION_HIGHP)
2384			return ctx.format.ulp(ret, 6.0);
2385		else
2386			return ctx.format.ulp(ret, 2.0);
2387	}
2388
2389	// Codomain could be [-pi, pi], but that would probably be too strict.
2390};
2391
2392DEFINE_DERIVED_FLOAT1(Sinh, sinh, x, (exp(x) - exp(-x)) / constant(2.0f));
2393DEFINE_DERIVED_FLOAT1(Cosh, cosh, x, (exp(x) + exp(-x)) / constant(2.0f));
2394DEFINE_DERIVED_FLOAT1(Tanh, tanh, x, sinh(x) / cosh(x));
2395
2396// These are not defined as derived forms in the GLSL ES spec, but
2397// that gives us a reasonable precision.
2398DEFINE_DERIVED_FLOAT1(ASinh, asinh, x, log(x + sqrt(x * x + constant(1.0f))));
2399DEFINE_DERIVED_FLOAT1(ACosh, acosh, x, log(x + sqrt((x + constant(1.0f)) *
2400													(x - constant(1.0f)))));
2401DEFINE_DERIVED_FLOAT1(ATanh, atanh, x, constant(0.5f) * log((constant(1.0f) + x) /
2402															(constant(1.0f) - x)));
2403
2404template <typename T>
2405class GetComponent : public PrimitiveFunc<Signature<typename T::Element, T, int> >
2406{
2407public:
2408	typedef		typename GetComponent::IRet	IRet;
2409
2410	string		getName		(void) const { return "_getComponent"; }
2411
2412	void		print		(ostream&				os,
2413							 const BaseArgExprs&	args) const
2414	{
2415		os << *args[0] << "[" << *args[1] << "]";
2416	}
2417
2418protected:
2419	IRet		doApply 	(const EvalContext&,
2420							 const typename GetComponent::IArgs& iargs) const
2421	{
2422		IRet ret;
2423
2424		for (int compNdx = 0; compNdx < T::SIZE; ++compNdx)
2425		{
2426			if (iargs.b.contains(compNdx))
2427				ret = unionIVal<typename T::Element>(ret, iargs.a[compNdx]);
2428		}
2429
2430		return ret;
2431	}
2432
2433};
2434
2435template <typename T>
2436ExprP<typename T::Element> getComponent (const ExprP<T>& container, int ndx)
2437{
2438	DE_ASSERT(0 <= ndx && ndx < T::SIZE);
2439	return app<GetComponent<T> >(container, constant(ndx));
2440}
2441
2442template <typename T>	string	vecNamePrefix			(void);
2443template <>				string	vecNamePrefix<float>	(void) { return ""; }
2444template <>				string	vecNamePrefix<int>		(void) { return "i"; }
2445template <>				string	vecNamePrefix<bool>		(void) { return "b"; }
2446
2447template <typename T, int Size>
2448string vecName (void) { return vecNamePrefix<T>() + "vec" + de::toString(Size); }
2449
2450template <typename T, int Size> class GenVec;
2451
2452template <typename T>
2453class GenVec<T, 1> : public DerivedFunc<Signature<T, T> >
2454{
2455public:
2456	typedef typename GenVec<T, 1>::ArgExprs ArgExprs;
2457
2458	string		getName		(void) const
2459	{
2460		return "_" + vecName<T, 1>();
2461	}
2462
2463protected:
2464
2465	ExprP<T>	doExpand	(ExpandContext&, const ArgExprs& args) const { return args.a; }
2466};
2467
2468template <typename T>
2469class GenVec<T, 2> : public PrimitiveFunc<Signature<Vector<T, 2>, T, T> >
2470{
2471public:
2472	typedef typename GenVec::IRet	IRet;
2473	typedef typename GenVec::IArgs	IArgs;
2474
2475	string		getName		(void) const
2476	{
2477		return vecName<T, 2>();
2478	}
2479
2480protected:
2481	IRet		doApply		(const EvalContext&, const IArgs& iargs) const
2482	{
2483		return IRet(iargs.a, iargs.b);
2484	}
2485};
2486
2487template <typename T>
2488class GenVec<T, 3> : public PrimitiveFunc<Signature<Vector<T, 3>, T, T, T> >
2489{
2490public:
2491	typedef typename GenVec::IRet	IRet;
2492	typedef typename GenVec::IArgs	IArgs;
2493
2494	string	getName		(void) const
2495	{
2496		return vecName<T, 3>();
2497	}
2498
2499protected:
2500	IRet	doApply		(const EvalContext&, const IArgs& iargs) const
2501	{
2502		return IRet(iargs.a, iargs.b, iargs.c);
2503	}
2504};
2505
2506template <typename T>
2507class GenVec<T, 4> : public PrimitiveFunc<Signature<Vector<T, 4>, T, T, T, T> >
2508{
2509public:
2510	typedef typename GenVec::IRet	IRet;
2511	typedef typename GenVec::IArgs	IArgs;
2512
2513	string		getName		(void) const { return vecName<T, 4>(); }
2514
2515protected:
2516	IRet		doApply		(const EvalContext&, const IArgs& iargs) const
2517	{
2518		return IRet(iargs.a, iargs.b, iargs.c, iargs.d);
2519	}
2520};
2521
2522
2523
2524template <typename T, int Rows, int Columns>
2525class GenMat;
2526
2527template <typename T, int Rows>
2528class GenMat<T, Rows, 2> : public PrimitiveFunc<
2529	Signature<Matrix<T, Rows, 2>, Vector<T, Rows>, Vector<T, Rows> > >
2530{
2531public:
2532	typedef typename GenMat::Ret	Ret;
2533	typedef typename GenMat::IRet	IRet;
2534	typedef typename GenMat::IArgs	IArgs;
2535
2536	string		getName		(void) const
2537	{
2538		return dataTypeNameOf<Ret>();
2539	}
2540
2541protected:
2542
2543	IRet		doApply		(const EvalContext&, const IArgs& iargs) const
2544	{
2545		IRet	ret;
2546		ret[0] = iargs.a;
2547		ret[1] = iargs.b;
2548		return ret;
2549	}
2550};
2551
2552template <typename T, int Rows>
2553class GenMat<T, Rows, 3> : public PrimitiveFunc<
2554	Signature<Matrix<T, Rows, 3>, Vector<T, Rows>, Vector<T, Rows>, Vector<T, Rows> > >
2555{
2556public:
2557	typedef typename GenMat::Ret	Ret;
2558	typedef typename GenMat::IRet	IRet;
2559	typedef typename GenMat::IArgs	IArgs;
2560
2561	string	getName	(void) const
2562	{
2563		return dataTypeNameOf<Ret>();
2564	}
2565
2566protected:
2567
2568	IRet	doApply	(const EvalContext&, const IArgs& iargs) const
2569	{
2570		IRet	ret;
2571		ret[0] = iargs.a;
2572		ret[1] = iargs.b;
2573		ret[2] = iargs.c;
2574		return ret;
2575	}
2576};
2577
2578template <typename T, int Rows>
2579class GenMat<T, Rows, 4> : public PrimitiveFunc<
2580	Signature<Matrix<T, Rows, 4>,
2581			  Vector<T, Rows>, Vector<T, Rows>, Vector<T, Rows>, Vector<T, Rows> > >
2582{
2583public:
2584	typedef typename GenMat::Ret	Ret;
2585	typedef typename GenMat::IRet	IRet;
2586	typedef typename GenMat::IArgs	IArgs;
2587
2588	string	getName	(void) const
2589	{
2590		return dataTypeNameOf<Ret>();
2591	}
2592
2593protected:
2594	IRet	doApply	(const EvalContext&, const IArgs& iargs) const
2595	{
2596		IRet	ret;
2597		ret[0] = iargs.a;
2598		ret[1] = iargs.b;
2599		ret[2] = iargs.c;
2600		ret[3] = iargs.d;
2601		return ret;
2602	}
2603};
2604
2605template <typename T, int Rows>
2606ExprP<Matrix<T, Rows, 2> > mat2 (const ExprP<Vector<T, Rows> >& arg0,
2607								 const ExprP<Vector<T, Rows> >& arg1)
2608{
2609	return app<GenMat<T, Rows, 2> >(arg0, arg1);
2610}
2611
2612template <typename T, int Rows>
2613ExprP<Matrix<T, Rows, 3> > mat3 (const ExprP<Vector<T, Rows> >& arg0,
2614								 const ExprP<Vector<T, Rows> >& arg1,
2615								 const ExprP<Vector<T, Rows> >& arg2)
2616{
2617	return app<GenMat<T, Rows, 3> >(arg0, arg1, arg2);
2618}
2619
2620template <typename T, int Rows>
2621ExprP<Matrix<T, Rows, 4> > mat4 (const ExprP<Vector<T, Rows> >& arg0,
2622								 const ExprP<Vector<T, Rows> >& arg1,
2623								 const ExprP<Vector<T, Rows> >& arg2,
2624								 const ExprP<Vector<T, Rows> >& arg3)
2625{
2626	return app<GenMat<T, Rows, 4> >(arg0, arg1, arg2, arg3);
2627}
2628
2629
2630template <int Rows, int Cols>
2631class MatNeg : public PrimitiveFunc<Signature<Matrix<float, Rows, Cols>,
2632											  Matrix<float, Rows, Cols> > >
2633{
2634public:
2635	typedef typename MatNeg::IRet		IRet;
2636	typedef typename MatNeg::IArgs		IArgs;
2637
2638	string	getName	(void) const
2639	{
2640		return "_matNeg";
2641	}
2642
2643protected:
2644	void	doPrint	(ostream& os, const BaseArgExprs& args) const
2645	{
2646		os << "-(" << *args[0] << ")";
2647	}
2648
2649	IRet	doApply	(const EvalContext&, const IArgs& iargs)			const
2650	{
2651		IRet	ret;
2652
2653		for (int col = 0; col < Cols; ++col)
2654		{
2655			for (int row = 0; row < Rows; ++row)
2656				ret[col][row] = -iargs.a[col][row];
2657		}
2658
2659		return ret;
2660	}
2661};
2662
2663template <typename T, typename Sig>
2664class CompWiseFunc : public PrimitiveFunc<Sig>
2665{
2666public:
2667	typedef Func<Signature<T, T, T> >	ScalarFunc;
2668
2669	string				getName			(void)									const
2670	{
2671		return doGetScalarFunc().getName();
2672	}
2673protected:
2674	void				doPrint			(ostream&				os,
2675										 const BaseArgExprs&	args)			const
2676	{
2677		doGetScalarFunc().print(os, args);
2678	}
2679
2680	virtual
2681	const ScalarFunc&	doGetScalarFunc	(void)									const = 0;
2682};
2683
2684template <int Rows, int Cols>
2685class CompMatFuncBase : public CompWiseFunc<float, Signature<Matrix<float, Rows, Cols>,
2686															 Matrix<float, Rows, Cols>,
2687															 Matrix<float, Rows, Cols> > >
2688{
2689public:
2690	typedef typename CompMatFuncBase::IRet		IRet;
2691	typedef typename CompMatFuncBase::IArgs		IArgs;
2692
2693protected:
2694
2695	IRet	doApply	(const EvalContext& ctx, const IArgs& iargs) const
2696	{
2697		IRet			ret;
2698
2699		for (int col = 0; col < Cols; ++col)
2700		{
2701			for (int row = 0; row < Rows; ++row)
2702				ret[col][row] = this->doGetScalarFunc().apply(ctx,
2703															  iargs.a[col][row],
2704															  iargs.b[col][row]);
2705		}
2706
2707		return ret;
2708	}
2709};
2710
2711template <typename F, int Rows, int Cols>
2712class CompMatFunc : public CompMatFuncBase<Rows, Cols>
2713{
2714protected:
2715	const typename CompMatFunc::ScalarFunc&	doGetScalarFunc	(void) const
2716	{
2717		return instance<F>();
2718	}
2719};
2720
2721class ScalarMatrixCompMult : public Mul
2722{
2723public:
2724	string	getName	(void) const
2725	{
2726		return "matrixCompMult";
2727	}
2728
2729	void	doPrint	(ostream& os, const BaseArgExprs& args) const
2730	{
2731		Func<Sig>::doPrint(os, args);
2732	}
2733};
2734
2735template <int Rows, int Cols>
2736class MatrixCompMult : public CompMatFunc<ScalarMatrixCompMult, Rows, Cols>
2737{
2738};
2739
2740template <int Rows, int Cols>
2741class ScalarMatFuncBase : public CompWiseFunc<float, Signature<Matrix<float, Rows, Cols>,
2742															   Matrix<float, Rows, Cols>,
2743															   float> >
2744{
2745public:
2746	typedef typename ScalarMatFuncBase::IRet	IRet;
2747	typedef typename ScalarMatFuncBase::IArgs	IArgs;
2748
2749protected:
2750
2751	IRet	doApply	(const EvalContext& ctx, const IArgs& iargs) const
2752	{
2753		IRet	ret;
2754
2755		for (int col = 0; col < Cols; ++col)
2756		{
2757			for (int row = 0; row < Rows; ++row)
2758				ret[col][row] = this->doGetScalarFunc().apply(ctx, iargs.a[col][row], iargs.b);
2759		}
2760
2761		return ret;
2762	}
2763};
2764
2765template <typename F, int Rows, int Cols>
2766class ScalarMatFunc : public ScalarMatFuncBase<Rows, Cols>
2767{
2768protected:
2769	const typename ScalarMatFunc::ScalarFunc&	doGetScalarFunc	(void)	const
2770	{
2771		return instance<F>();
2772	}
2773};
2774
2775template<typename T, int Size> struct GenXType;
2776
2777template<typename T>
2778struct GenXType<T, 1>
2779{
2780	static ExprP<T>	genXType	(const ExprP<T>& x) { return x; }
2781};
2782
2783template<typename T>
2784struct GenXType<T, 2>
2785{
2786	static ExprP<Vector<T, 2> >	genXType	(const ExprP<T>& x)
2787	{
2788		return app<GenVec<T, 2> >(x, x);
2789	}
2790};
2791
2792template<typename T>
2793struct GenXType<T, 3>
2794{
2795	static ExprP<Vector<T, 3> >	genXType	(const ExprP<T>& x)
2796	{
2797		return app<GenVec<T, 3> >(x, x, x);
2798	}
2799};
2800
2801template<typename T>
2802struct GenXType<T, 4>
2803{
2804	static ExprP<Vector<T, 4> >	genXType	(const ExprP<T>& x)
2805	{
2806		return app<GenVec<T, 4> >(x, x, x, x);
2807	}
2808};
2809
2810//! Returns an expression of vector of size `Size` (or scalar if Size == 1),
2811//! with each element initialized with the expression `x`.
2812template<typename T, int Size>
2813ExprP<typename ContainerOf<T, Size>::Container> genXType (const ExprP<T>& x)
2814{
2815	return GenXType<T, Size>::genXType(x);
2816}
2817
2818typedef GenVec<float, 2> FloatVec2;
2819DEFINE_CONSTRUCTOR2(FloatVec2, Vec2, vec2, float, float)
2820
2821typedef GenVec<float, 3> FloatVec3;
2822DEFINE_CONSTRUCTOR3(FloatVec3, Vec3, vec3, float, float, float)
2823
2824typedef GenVec<float, 4> FloatVec4;
2825DEFINE_CONSTRUCTOR4(FloatVec4, Vec4, vec4, float, float, float, float)
2826
2827template <int Size>
2828class Dot : public DerivedFunc<Signature<float, Vector<float, Size>, Vector<float, Size> > >
2829{
2830public:
2831	typedef typename Dot::ArgExprs ArgExprs;
2832
2833	string			getName		(void) const
2834	{
2835		return "dot";
2836	}
2837
2838protected:
2839	ExprP<float>	doExpand 	(ExpandContext&, const ArgExprs& args) const
2840	{
2841		ExprP<float> val = args.a[0] * args.b[0];
2842
2843		for (int ndx = 1; ndx < Size; ++ndx)
2844			val = val + args.a[ndx] * args.b[ndx];
2845
2846		return val;
2847	}
2848};
2849
2850template <>
2851class Dot<1> : public DerivedFunc<Signature<float, float, float> >
2852{
2853public:
2854	string			getName		(void) const
2855	{
2856		return "dot";
2857	}
2858
2859	ExprP<float>	doExpand 	(ExpandContext&, const ArgExprs& args) const
2860	{
2861		return args.a * args.b;
2862	}
2863};
2864
2865template <int Size>
2866ExprP<float> dot (const ExprP<Vector<float, Size> >& x, const ExprP<Vector<float, Size> >& y)
2867{
2868	return app<Dot<Size> >(x, y);
2869}
2870
2871ExprP<float> dot (const ExprP<float>& x, const ExprP<float>& y)
2872{
2873	return app<Dot<1> >(x, y);
2874}
2875
2876template <int Size>
2877class Length : public DerivedFunc<
2878	Signature<float, typename ContainerOf<float, Size>::Container> >
2879{
2880public:
2881	typedef typename Length::ArgExprs ArgExprs;
2882
2883	string			getName		(void) const
2884	{
2885		return "length";
2886	}
2887
2888protected:
2889	ExprP<float>	doExpand	(ExpandContext&, const ArgExprs& args) const
2890	{
2891		return sqrt(dot(args.a, args.a));
2892	}
2893};
2894
2895template <int Size>
2896ExprP<float> length (const ExprP<typename ContainerOf<float, Size>::Container>& x)
2897{
2898	return app<Length<Size> >(x);
2899}
2900
2901template <int Size>
2902class Distance : public DerivedFunc<
2903	Signature<float,
2904			  typename ContainerOf<float, Size>::Container,
2905			  typename ContainerOf<float, Size>::Container> >
2906{
2907public:
2908	typedef typename	Distance::Ret		Ret;
2909	typedef typename	Distance::ArgExprs	ArgExprs;
2910
2911	string		getName		(void) const
2912	{
2913		return "distance";
2914	}
2915
2916protected:
2917	ExprP<Ret>	doExpand 	(ExpandContext&, const ArgExprs& args) const
2918	{
2919		return length<Size>(args.a - args.b);
2920	}
2921};
2922
2923// cross
2924
2925class Cross : public DerivedFunc<Signature<Vec3, Vec3, Vec3> >
2926{
2927public:
2928	string			getName		(void) const
2929	{
2930		return "cross";
2931	}
2932
2933protected:
2934	ExprP<Vec3>		doExpand 	(ExpandContext&, const ArgExprs& x) const
2935	{
2936		return vec3(x.a[1] * x.b[2] - x.b[1] * x.a[2],
2937					x.a[2] * x.b[0] - x.b[2] * x.a[0],
2938					x.a[0] * x.b[1] - x.b[0] * x.a[1]);
2939	}
2940};
2941
2942DEFINE_CONSTRUCTOR2(Cross, Vec3, cross, Vec3, Vec3)
2943
2944template<int Size>
2945class Normalize : public DerivedFunc<
2946	Signature<typename ContainerOf<float, Size>::Container,
2947			  typename ContainerOf<float, Size>::Container> >
2948{
2949public:
2950	typedef typename	Normalize::Ret		Ret;
2951	typedef typename	Normalize::ArgExprs	ArgExprs;
2952
2953	string		getName		(void) const
2954	{
2955		return "normalize";
2956	}
2957
2958protected:
2959	ExprP<Ret>	doExpand	(ExpandContext&, const ArgExprs& args) const
2960	{
2961		return args.a / length<Size>(args.a);
2962	}
2963};
2964
2965template <int Size>
2966class FaceForward : public DerivedFunc<
2967	Signature<typename ContainerOf<float, Size>::Container,
2968			  typename ContainerOf<float, Size>::Container,
2969			  typename ContainerOf<float, Size>::Container,
2970			  typename ContainerOf<float, Size>::Container> >
2971{
2972public:
2973	typedef typename	FaceForward::Ret		Ret;
2974	typedef typename	FaceForward::ArgExprs	ArgExprs;
2975
2976	string		getName		(void) const
2977	{
2978		return "faceforward";
2979	}
2980
2981protected:
2982
2983
2984	ExprP<Ret>	doExpand	(ExpandContext&, const ArgExprs& args) const
2985	{
2986		return cond(dot(args.c, args.b) < constant(0.0f), args.a, -args.a);
2987	}
2988};
2989
2990template <int Size>
2991class Reflect : public DerivedFunc<
2992	Signature<typename ContainerOf<float, Size>::Container,
2993			  typename ContainerOf<float, Size>::Container,
2994			  typename ContainerOf<float, Size>::Container> >
2995{
2996public:
2997	typedef typename	Reflect::Ret		Ret;
2998	typedef typename	Reflect::ArgExprs	ArgExprs;
2999
3000	string		getName		(void) const
3001	{
3002		return "reflect";
3003	}
3004
3005protected:
3006	ExprP<Ret>	doExpand	(ExpandContext&, const ArgExprs& args) const
3007	{
3008		return args.a - (args.b * dot(args.b, args.a) * constant(2.0f));
3009	}
3010};
3011
3012template <int Size>
3013class Refract : public DerivedFunc<
3014	Signature<typename ContainerOf<float, Size>::Container,
3015			  typename ContainerOf<float, Size>::Container,
3016			  typename ContainerOf<float, Size>::Container,
3017			  float> >
3018{
3019public:
3020	typedef typename	Refract::Ret		Ret;
3021	typedef typename	Refract::Arg0		Arg0;
3022	typedef typename	Refract::Arg1		Arg1;
3023	typedef typename	Refract::ArgExprs	ArgExprs;
3024
3025	string		getName		(void) const
3026	{
3027		return "refract";
3028	}
3029
3030protected:
3031	ExprP<Ret>	doExpand	(ExpandContext&	ctx, const ArgExprs& args) const
3032	{
3033		const ExprP<Arg0>&	i		= args.a;
3034		const ExprP<Arg1>&	n		= args.b;
3035		const ExprP<float>&	eta		= args.c;
3036		const ExprP<float>	dotNI	= bindExpression("dotNI", ctx, dot(n, i));
3037		const ExprP<float>	k		= bindExpression("k", ctx, constant(1.0f) - eta * eta *
3038												 (constant(1.0f) - dotNI * dotNI));
3039
3040		return cond(k < constant(0.0f),
3041					genXType<float, Size>(constant(0.0f)),
3042					i * eta - n * (eta * dotNI + sqrt(k)));
3043	}
3044};
3045
3046class PreciseFunc1 : public CFloatFunc1
3047{
3048public:
3049			PreciseFunc1	(const string& name, DoubleFunc1& func) : CFloatFunc1(name, func) {}
3050protected:
3051	double	precision		(const EvalContext&, double, double) const	{ return 0.0; }
3052};
3053
3054class Abs : public PreciseFunc1
3055{
3056public:
3057	Abs (void) : PreciseFunc1("abs", deAbs) {}
3058};
3059
3060class Sign : public PreciseFunc1
3061{
3062public:
3063	Sign (void) : PreciseFunc1("sign", deSign) {}
3064};
3065
3066class Floor : public PreciseFunc1
3067{
3068public:
3069	Floor (void) : PreciseFunc1("floor", deFloor) {}
3070};
3071
3072class Trunc : public PreciseFunc1
3073{
3074public:
3075	Trunc (void) : PreciseFunc1("trunc", deTrunc) {}
3076};
3077
3078class Round : public FloatFunc1
3079{
3080public:
3081	string		getName		(void) const								{ return "round"; }
3082
3083protected:
3084	Interval	applyPoint	(const EvalContext&, double x) const
3085	{
3086		double			truncated	= 0.0;
3087		const double	fract		= deModf(x, &truncated);
3088		Interval		ret;
3089
3090		if (fabs(fract) <= 0.5)
3091			ret |= truncated;
3092		if (fabs(fract) >= 0.5)
3093			ret |= truncated + deSign(fract);
3094
3095		return ret;
3096	}
3097
3098	double		precision	(const EvalContext&, double, double) const	{ return 0.0; }
3099};
3100
3101class RoundEven : public PreciseFunc1
3102{
3103public:
3104	RoundEven (void) : PreciseFunc1("roundEven", deRoundEven) {}
3105};
3106
3107class Ceil : public PreciseFunc1
3108{
3109public:
3110	Ceil (void) : PreciseFunc1("ceil", deCeil) {}
3111};
3112
3113DEFINE_DERIVED_FLOAT1(Fract, fract, x, x - app<Floor>(x));
3114
3115class PreciseFunc2 : public CFloatFunc2
3116{
3117public:
3118			PreciseFunc2	(const string& name, DoubleFunc2& func) : CFloatFunc2(name, func) {}
3119protected:
3120	double	precision		(const EvalContext&, double, double, double) const { return 0.0; }
3121};
3122
3123DEFINE_DERIVED_FLOAT2(Mod, mod, x, y, x - y * app<Floor>(x / y));
3124
3125class Modf : public PrimitiveFunc<Signature<float, float, float> >
3126{
3127public:
3128	string	getName				(void) const
3129	{
3130		return "modf";
3131	}
3132
3133protected:
3134	IRet	doApply				(const EvalContext&, const IArgs& iargs) const
3135	{
3136		Interval	fracIV;
3137		Interval&	wholeIV		= const_cast<Interval&>(iargs.b);
3138		double		intPart		= 0;
3139
3140		TCU_INTERVAL_APPLY_MONOTONE1(fracIV, x, iargs.a, frac, frac = deModf(x, &intPart));
3141		TCU_INTERVAL_APPLY_MONOTONE1(wholeIV, x, iargs.a, whole,
3142									 deModf(x, &intPart); whole = intPart);
3143		return fracIV;
3144	}
3145
3146	int		getOutParamIndex	(void) const
3147	{
3148		return 1;
3149	}
3150};
3151
3152class Min : public PreciseFunc2 { public: Min (void) : PreciseFunc2("min", deMin) {} };
3153class Max : public PreciseFunc2 { public: Max (void) : PreciseFunc2("max", deMax) {} };
3154
3155class Clamp : public FloatFunc3
3156{
3157public:
3158	string	getName		(void) const { return "clamp"; }
3159
3160	double	applyExact	(double x, double minVal, double maxVal) const
3161	{
3162		return de::min(de::max(x, minVal), maxVal);
3163	}
3164
3165	double	precision	(const EvalContext&, double, double, double minVal, double maxVal) const
3166	{
3167		return minVal > maxVal ? TCU_NAN : 0.0;
3168	}
3169};
3170
3171ExprP<float> clamp(const ExprP<float>& x, const ExprP<float>& minVal, const ExprP<float>& maxVal)
3172{
3173	return app<Clamp>(x, minVal, maxVal);
3174}
3175
3176DEFINE_DERIVED_FLOAT3(Mix, mix, x, y, a, (x * (constant(1.0f) - a)) + y * a);
3177
3178static double step (double edge, double x)
3179{
3180	return x < edge ? 0.0 : 1.0;
3181}
3182
3183class Step : public PreciseFunc2 { public: Step (void) : PreciseFunc2("step", step) {} };
3184
3185class SmoothStep : public DerivedFunc<Signature<float, float, float, float> >
3186{
3187public:
3188	string		getName		(void) const
3189	{
3190		return "smoothstep";
3191	}
3192
3193protected:
3194
3195	ExprP<Ret>	doExpand 	(ExpandContext& ctx, const ArgExprs& args) const
3196	{
3197		const ExprP<float>&		edge0	= args.a;
3198		const ExprP<float>&		edge1	= args.b;
3199		const ExprP<float>&		x		= args.c;
3200		const ExprP<float>		tExpr	= clamp((x - edge0) / (edge1 - edge0),
3201											constant(0.0f), constant(1.0f));
3202		const ExprP<float>		t		= bindExpression("t", ctx, tExpr);
3203
3204		return (t * t * (constant(3.0f) - constant(2.0f) * t));
3205	}
3206};
3207
3208class FrExp : public PrimitiveFunc<Signature<float, float, int> >
3209{
3210public:
3211	string	getName			(void) const
3212	{
3213		return "frexp";
3214	}
3215
3216protected:
3217	IRet	doApply			(const EvalContext&, const IArgs& iargs) const
3218	{
3219		IRet			ret;
3220		const IArg0&	x			= iargs.a;
3221		IArg1&			exponent	= const_cast<IArg1&>(iargs.b);
3222
3223		if (x.hasNaN() || x.contains(TCU_INFINITY) || x.contains(-TCU_INFINITY))
3224		{
3225			// GLSL (in contrast to IEEE) says that result of applying frexp
3226			// to infinity is undefined
3227			ret = Interval::unbounded() | TCU_NAN;
3228			exponent = Interval(-deLdExp(1.0, 31), deLdExp(1.0, 31)-1);
3229		}
3230		else if (!x.empty())
3231		{
3232			int				loExp	= 0;
3233			const double	loFrac	= deFrExp(x.lo(), &loExp);
3234			int				hiExp	= 0;
3235			const double	hiFrac	= deFrExp(x.hi(), &hiExp);
3236
3237			if (deSign(loFrac) != deSign(hiFrac))
3238			{
3239				exponent = Interval(-TCU_INFINITY, de::max(loExp, hiExp));
3240				ret = Interval();
3241				if (deSign(loFrac) < 0)
3242					ret |= Interval(-1.0 + DBL_EPSILON*0.5, 0.0);
3243				if (deSign(hiFrac) > 0)
3244					ret |= Interval(0.0, 1.0 - DBL_EPSILON*0.5);
3245			}
3246			else
3247			{
3248				exponent = Interval(loExp, hiExp);
3249				if (loExp == hiExp)
3250					ret = Interval(loFrac, hiFrac);
3251				else
3252					ret = deSign(loFrac) * Interval(0.5, 1.0 - DBL_EPSILON*0.5);
3253			}
3254		}
3255
3256		return ret;
3257	}
3258
3259	int	getOutParamIndex	(void) const
3260	{
3261		return 1;
3262	}
3263};
3264
3265class LdExp : public PrimitiveFunc<Signature<float, float, int> >
3266{
3267public:
3268	string		getName			(void) const
3269	{
3270		return "ldexp";
3271	}
3272
3273protected:
3274	Interval	doApply			(const EvalContext& ctx, const IArgs& iargs) const
3275	{
3276		Interval	ret = call<Exp2>(ctx, iargs.b);
3277		// Khronos bug 11180 consensus: if exp2(exponent) cannot be represented,
3278		// the result is undefined.
3279
3280		if (ret.contains(TCU_INFINITY) | ret.contains(-TCU_INFINITY))
3281			ret |= TCU_NAN;
3282
3283		return call<Mul>(ctx, iargs.a, ret);
3284	}
3285};
3286
3287template<int Rows, int Columns>
3288class Transpose : public PrimitiveFunc<Signature<Matrix<float, Rows, Columns>,
3289												 Matrix<float, Columns, Rows> > >
3290{
3291public:
3292	typedef typename Transpose::IRet	IRet;
3293	typedef typename Transpose::IArgs	IArgs;
3294
3295	string		getName		(void) const
3296	{
3297		return "transpose";
3298	}
3299
3300protected:
3301	IRet		doApply		(const EvalContext&, const IArgs& iargs) const
3302	{
3303		IRet ret;
3304
3305		for (int rowNdx = 0; rowNdx < Rows; ++rowNdx)
3306		{
3307			for (int colNdx = 0; colNdx < Columns; ++colNdx)
3308				ret(rowNdx, colNdx) = iargs.a(colNdx, rowNdx);
3309		}
3310
3311		return ret;
3312	}
3313};
3314
3315template<typename Ret, typename Arg0, typename Arg1>
3316class MulFunc : public PrimitiveFunc<Signature<Ret, Arg0, Arg1> >
3317{
3318public:
3319	string	getName	(void) const 									{ return "mul"; }
3320
3321protected:
3322	void	doPrint	(ostream& os, const BaseArgExprs& args) const
3323	{
3324		os << "(" << *args[0] << " * " << *args[1] << ")";
3325	}
3326};
3327
3328template<int LeftRows, int Middle, int RightCols>
3329class MatMul : public MulFunc<Matrix<float, LeftRows, RightCols>,
3330							  Matrix<float, LeftRows, Middle>,
3331							  Matrix<float, Middle, RightCols> >
3332{
3333protected:
3334	typedef typename MatMul::IRet	IRet;
3335	typedef typename MatMul::IArgs	IArgs;
3336	typedef typename MatMul::IArg0	IArg0;
3337	typedef typename MatMul::IArg1	IArg1;
3338
3339	IRet	doApply	(const EvalContext&	ctx, const IArgs& iargs) const
3340	{
3341		const IArg0&	left	= iargs.a;
3342		const IArg1&	right	= iargs.b;
3343		IRet			ret;
3344
3345		for (int row = 0; row < LeftRows; ++row)
3346		{
3347			for (int col = 0; col < RightCols; ++col)
3348			{
3349				Interval	element	(0.0);
3350
3351				for (int ndx = 0; ndx < Middle; ++ndx)
3352					element = call<Add>(ctx, element,
3353										call<Mul>(ctx, left[ndx][row], right[col][ndx]));
3354
3355				ret[col][row] = element;
3356			}
3357		}
3358
3359		return ret;
3360	}
3361};
3362
3363template<int Rows, int Cols>
3364class VecMatMul : public MulFunc<Vector<float, Cols>,
3365								 Vector<float, Rows>,
3366								 Matrix<float, Rows, Cols> >
3367{
3368public:
3369	typedef typename VecMatMul::IRet	IRet;
3370	typedef typename VecMatMul::IArgs	IArgs;
3371	typedef typename VecMatMul::IArg0	IArg0;
3372	typedef typename VecMatMul::IArg1	IArg1;
3373
3374protected:
3375	IRet	doApply	(const EvalContext& ctx, const IArgs& iargs) const
3376	{
3377		const IArg0&	left	= iargs.a;
3378		const IArg1&	right	= iargs.b;
3379		IRet			ret;
3380
3381		for (int col = 0; col < Cols; ++col)
3382		{
3383			Interval	element	(0.0);
3384
3385			for (int row = 0; row < Rows; ++row)
3386				element = call<Add>(ctx, element, call<Mul>(ctx, left[row], right[col][row]));
3387
3388			ret[col] = element;
3389		}
3390
3391		return ret;
3392	}
3393};
3394
3395template<int Rows, int Cols>
3396class MatVecMul : public MulFunc<Vector<float, Rows>,
3397								 Matrix<float, Rows, Cols>,
3398								 Vector<float, Cols> >
3399{
3400public:
3401	typedef typename MatVecMul::IRet	IRet;
3402	typedef typename MatVecMul::IArgs	IArgs;
3403	typedef typename MatVecMul::IArg0	IArg0;
3404	typedef typename MatVecMul::IArg1	IArg1;
3405
3406protected:
3407	IRet	doApply	(const EvalContext& ctx, const IArgs& iargs) const
3408	{
3409		const IArg0&	left	= iargs.a;
3410		const IArg1&	right	= iargs.b;
3411
3412		return call<VecMatMul<Cols, Rows> >(ctx, right,
3413											call<Transpose<Rows, Cols> >(ctx, left));
3414	}
3415};
3416
3417template<int Rows, int Cols>
3418class OuterProduct : public PrimitiveFunc<Signature<Matrix<float, Rows, Cols>,
3419													Vector<float, Rows>,
3420													Vector<float, Cols> > >
3421{
3422public:
3423	typedef typename OuterProduct::IRet		IRet;
3424	typedef typename OuterProduct::IArgs	IArgs;
3425
3426	string	getName	(void) const
3427	{
3428		return "outerProduct";
3429	}
3430
3431protected:
3432	IRet	doApply	(const EvalContext& ctx, const IArgs& iargs) const
3433	{
3434		IRet	ret;
3435
3436		for (int row = 0; row < Rows; ++row)
3437		{
3438			for (int col = 0; col < Cols; ++col)
3439				ret[col][row] = call<Mul>(ctx, iargs.a[row], iargs.b[col]);
3440		}
3441
3442		return ret;
3443	}
3444};
3445
3446template<int Rows, int Cols>
3447ExprP<Matrix<float, Rows, Cols> > outerProduct (const ExprP<Vector<float, Rows> >& left,
3448												const ExprP<Vector<float, Cols> >& right)
3449{
3450	return app<OuterProduct<Rows, Cols> >(left, right);
3451}
3452
3453template<int Size>
3454class DeterminantBase : public DerivedFunc<Signature<float, Matrix<float, Size, Size> > >
3455{
3456public:
3457	string	getName	(void) const { return "determinant"; }
3458};
3459
3460template<int Size>
3461class Determinant;
3462
3463template<int Size>
3464ExprP<float> determinant (ExprP<Matrix<float, Size, Size> > mat)
3465{
3466	return app<Determinant<Size> >(mat);
3467}
3468
3469template<>
3470class Determinant<2> : public DeterminantBase<2>
3471{
3472protected:
3473	ExprP<Ret>	doExpand (ExpandContext&, const ArgExprs& args)	const
3474	{
3475		ExprP<Mat2>	mat	= args.a;
3476
3477		return mat[0][0] * mat[1][1] - mat[1][0] * mat[0][1];
3478	}
3479};
3480
3481template<>
3482class Determinant<3> : public DeterminantBase<3>
3483{
3484protected:
3485	ExprP<Ret> doExpand (ExpandContext&, const ArgExprs& args) const
3486	{
3487		ExprP<Mat3>	mat	= args.a;
3488
3489		return (mat[0][0] * (mat[1][1] * mat[2][2] - mat[1][2] * mat[2][1]) +
3490				mat[0][1] * (mat[1][2] * mat[2][0] - mat[1][0] * mat[2][2]) +
3491				mat[0][2] * (mat[1][0] * mat[2][1] - mat[1][1] * mat[2][0]));
3492	}
3493};
3494
3495template<>
3496class Determinant<4> : public DeterminantBase<4>
3497{
3498protected:
3499	 ExprP<Ret>	doExpand	(ExpandContext& ctx, const ArgExprs& args) const
3500	{
3501		ExprP<Mat4>	mat	= args.a;
3502		ExprP<Mat3>	minors[4];
3503
3504		for (int ndx = 0; ndx < 4; ++ndx)
3505		{
3506			ExprP<Vec4>		minorColumns[3];
3507			ExprP<Vec3>		columns[3];
3508
3509			for (int col = 0; col < 3; ++col)
3510				minorColumns[col] = mat[col < ndx ? col : col + 1];
3511
3512			for (int col = 0; col < 3; ++col)
3513				columns[col] = vec3(minorColumns[0][col+1],
3514									minorColumns[1][col+1],
3515									minorColumns[2][col+1]);
3516
3517			minors[ndx] = bindExpression("minor", ctx,
3518										 mat3(columns[0], columns[1], columns[2]));
3519		}
3520
3521		return (mat[0][0] * determinant(minors[0]) -
3522				mat[1][0] * determinant(minors[1]) +
3523				mat[2][0] * determinant(minors[2]) -
3524				mat[3][0] * determinant(minors[3]));
3525	}
3526};
3527
3528template<int Size> class Inverse;
3529
3530template <int Size>
3531ExprP<Matrix<float, Size, Size> > inverse (ExprP<Matrix<float, Size, Size> > mat)
3532{
3533	return app<Inverse<Size> >(mat);
3534}
3535
3536template<>
3537class Inverse<2> : public DerivedFunc<Signature<Mat2, Mat2> >
3538{
3539public:
3540	string		getName	(void) const
3541	{
3542		return "inverse";
3543	}
3544
3545protected:
3546	ExprP<Ret>	doExpand (ExpandContext& ctx, const ArgExprs& args) const
3547	{
3548		ExprP<Mat2>		mat = args.a;
3549		ExprP<float>	det	= bindExpression("det", ctx, determinant(mat));
3550
3551		return mat2(vec2(mat[1][1] / det, -mat[0][1] / det),
3552					vec2(-mat[1][0] / det, mat[0][0] / det));
3553	}
3554};
3555
3556template<>
3557class Inverse<3> : public DerivedFunc<Signature<Mat3, Mat3> >
3558{
3559public:
3560	string		getName		(void) const
3561	{
3562		return "inverse";
3563	}
3564
3565protected:
3566	ExprP<Ret>	doExpand 	(ExpandContext& ctx, const ArgExprs& args)			const
3567	{
3568		ExprP<Mat3>		mat		= args.a;
3569		ExprP<Mat2>		invA	= bindExpression("invA", ctx,
3570												 inverse(mat2(vec2(mat[0][0], mat[0][1]),
3571															  vec2(mat[1][0], mat[1][1]))));
3572
3573		ExprP<Vec2>		matB	= bindExpression("matB", ctx, vec2(mat[2][0], mat[2][1]));
3574		ExprP<Vec2>		matC	= bindExpression("matC", ctx, vec2(mat[0][2], mat[1][2]));
3575		ExprP<float>	matD	= bindExpression("matD", ctx, mat[2][2]);
3576
3577		ExprP<float>	schur	= bindExpression("schur", ctx,
3578												 constant(1.0f) /
3579												 (matD - dot(matC * invA, matB)));
3580
3581		ExprP<Vec2>		t1		= invA * matB;
3582		ExprP<Vec2>		t2		= t1 * schur;
3583		ExprP<Mat2>		t3		= outerProduct(t2, matC);
3584		ExprP<Mat2>		t4		= t3 * invA;
3585		ExprP<Mat2>		t5		= invA + t4;
3586		ExprP<Mat2>		blockA	= bindExpression("blockA", ctx, t5);
3587		ExprP<Vec2>		blockB	= bindExpression("blockB", ctx,
3588												 (invA * matB) * -schur);
3589		ExprP<Vec2>		blockC	= bindExpression("blockC", ctx,
3590												 (matC * invA) * -schur);
3591
3592		return mat3(vec3(blockA[0][0], blockA[0][1], blockC[0]),
3593					vec3(blockA[1][0], blockA[1][1], blockC[1]),
3594					vec3(blockB[0], blockB[1], schur));
3595	}
3596};
3597
3598template<>
3599class Inverse<4> : public DerivedFunc<Signature<Mat4, Mat4> >
3600{
3601public:
3602	string		getName		(void) const { return "inverse"; }
3603
3604protected:
3605	ExprP<Ret>			doExpand 			(ExpandContext&		ctx,
3606											 const ArgExprs&	args)			const
3607	{
3608		ExprP<Mat4>	mat		= args.a;
3609		ExprP<Mat2>	invA	= bindExpression("invA", ctx,
3610											 inverse(mat2(vec2(mat[0][0], mat[0][1]),
3611														  vec2(mat[1][0], mat[1][1]))));
3612		ExprP<Mat2>	matB	= bindExpression("matB", ctx,
3613											 mat2(vec2(mat[2][0], mat[2][1]),
3614												  vec2(mat[3][0], mat[3][1])));
3615		ExprP<Mat2>	matC	= bindExpression("matC", ctx,
3616											 mat2(vec2(mat[0][2], mat[0][3]),
3617												  vec2(mat[1][2], mat[1][3])));
3618		ExprP<Mat2>	matD	= bindExpression("matD", ctx,
3619											 mat2(vec2(mat[2][2], mat[2][3]),
3620												  vec2(mat[3][2], mat[3][3])));
3621		ExprP<Mat2>	schur	= bindExpression("schur", ctx,
3622											 inverse(matD + -(matC * invA * matB)));
3623		ExprP<Mat2>	blockA	= bindExpression("blockA", ctx,
3624											 invA + (invA * matB * schur * matC * invA));
3625		ExprP<Mat2>	blockB	= bindExpression("blockB", ctx,
3626											 (-invA) * matB * schur);
3627		ExprP<Mat2>	blockC	= bindExpression("blockC", ctx,
3628											 (-schur) * matC * invA);
3629
3630		return mat4(vec4(blockA[0][0], blockA[0][1], blockC[0][0], blockC[0][1]),
3631					vec4(blockA[1][0], blockA[1][1], blockC[1][0], blockC[1][1]),
3632					vec4(blockB[0][0], blockB[0][1], schur[0][0], schur[0][1]),
3633					vec4(blockB[1][0], blockB[1][1], schur[1][0], schur[1][1]));
3634	}
3635};
3636
3637class Fma : public DerivedFunc<Signature<float, float, float, float> >
3638{
3639public:
3640	string			getName					(void) const
3641	{
3642		return "fma";
3643	}
3644
3645	string			getRequiredExtension	(void) const
3646	{
3647		return "GL_EXT_gpu_shader5";
3648	}
3649
3650protected:
3651	ExprP<float>	doExpand 				(ExpandContext&, const ArgExprs& x) const
3652	{
3653		return x.a * x.b + x.c;
3654	}
3655};
3656
3657} // Functions
3658
3659using namespace Functions;
3660
3661template <typename T>
3662ExprP<typename T::Element> ContainerExprPBase<T>::operator[] (int i) const
3663{
3664	return Functions::getComponent(exprP<T>(*this), i);
3665}
3666
3667ExprP<float> operator+ (const ExprP<float>& arg0, const ExprP<float>& arg1)
3668{
3669	return app<Add>(arg0, arg1);
3670}
3671
3672ExprP<float> operator- (const ExprP<float>& arg0, const ExprP<float>& arg1)
3673{
3674	return app<Sub>(arg0, arg1);
3675}
3676
3677ExprP<float> operator- (const ExprP<float>& arg0)
3678{
3679	return app<Negate>(arg0);
3680}
3681
3682ExprP<float> operator* (const ExprP<float>& arg0, const ExprP<float>& arg1)
3683{
3684	return app<Mul>(arg0, arg1);
3685}
3686
3687ExprP<float> operator/ (const ExprP<float>& arg0, const ExprP<float>& arg1)
3688{
3689	return app<Div>(arg0, arg1);
3690}
3691
3692template <typename Sig_, int Size>
3693class GenFunc : public PrimitiveFunc<Signature<
3694	typename ContainerOf<typename Sig_::Ret, Size>::Container,
3695	typename ContainerOf<typename Sig_::Arg0, Size>::Container,
3696	typename ContainerOf<typename Sig_::Arg1, Size>::Container,
3697	typename ContainerOf<typename Sig_::Arg2, Size>::Container,
3698	typename ContainerOf<typename Sig_::Arg3, Size>::Container> >
3699{
3700public:
3701	typedef typename GenFunc::IArgs		IArgs;
3702	typedef typename GenFunc::IRet		IRet;
3703
3704			GenFunc					(const Func<Sig_>&	scalarFunc) : m_func (scalarFunc) {}
3705
3706	string	getName					(void) const
3707	{
3708		return m_func.getName();
3709	}
3710
3711	int		getOutParamIndex		(void) const
3712	{
3713		return m_func.getOutParamIndex();
3714	}
3715
3716	string	getRequiredExtension	(void) const
3717	{
3718		return m_func.getRequiredExtension();
3719	}
3720
3721protected:
3722	void	doPrint					(ostream& os, const BaseArgExprs& args) const
3723	{
3724		m_func.print(os, args);
3725	}
3726
3727	IRet	doApply					(const EvalContext& ctx, const IArgs& iargs) const
3728	{
3729		IRet ret;
3730
3731		for (int ndx = 0; ndx < Size; ++ndx)
3732		{
3733			ret[ndx] =
3734				m_func.apply(ctx, iargs.a[ndx], iargs.b[ndx], iargs.c[ndx], iargs.d[ndx]);
3735		}
3736
3737		return ret;
3738	}
3739
3740	void	doGetUsedFuncs			(FuncSet& dst) const
3741	{
3742		m_func.getUsedFuncs(dst);
3743	}
3744
3745	const Func<Sig_>&	m_func;
3746};
3747
3748template <typename F, int Size>
3749class VectorizedFunc : public GenFunc<typename F::Sig, Size>
3750{
3751public:
3752	VectorizedFunc	(void) : GenFunc<typename F::Sig, Size>(instance<F>()) {}
3753};
3754
3755
3756
3757template <typename Sig_, int Size>
3758class FixedGenFunc : public PrimitiveFunc <Signature<
3759	typename ContainerOf<typename Sig_::Ret, Size>::Container,
3760	typename ContainerOf<typename Sig_::Arg0, Size>::Container,
3761	typename Sig_::Arg1,
3762	typename ContainerOf<typename Sig_::Arg2, Size>::Container,
3763	typename ContainerOf<typename Sig_::Arg3, Size>::Container> >
3764{
3765public:
3766	typedef typename FixedGenFunc::IArgs		IArgs;
3767	typedef typename FixedGenFunc::IRet			IRet;
3768
3769	string						getName			(void) const
3770	{
3771		return this->doGetScalarFunc().getName();
3772	}
3773
3774protected:
3775	void						doPrint			(ostream& os, const BaseArgExprs& args) const
3776	{
3777		this->doGetScalarFunc().print(os, args);
3778	}
3779
3780	IRet						doApply			(const EvalContext& ctx,
3781												 const IArgs&		iargs) const
3782	{
3783		IRet				ret;
3784		const Func<Sig_>&	func	= this->doGetScalarFunc();
3785
3786		for (int ndx = 0; ndx < Size; ++ndx)
3787			ret[ndx] = func.apply(ctx, iargs.a[ndx], iargs.b, iargs.c[ndx], iargs.d[ndx]);
3788
3789		return ret;
3790	}
3791
3792	virtual const Func<Sig_>&	doGetScalarFunc	(void) const = 0;
3793};
3794
3795template <typename F, int Size>
3796class FixedVecFunc : public FixedGenFunc<typename F::Sig, Size>
3797{
3798protected:
3799	const Func<typename F::Sig>& doGetScalarFunc	(void) const { return instance<F>(); }
3800};
3801
3802template<typename Sig>
3803struct GenFuncs
3804{
3805	GenFuncs (const Func<Sig>&			func_,
3806			  const GenFunc<Sig, 2>&	func2_,
3807			  const GenFunc<Sig, 3>&	func3_,
3808			  const GenFunc<Sig, 4>&	func4_)
3809		: func	(func_)
3810		, func2	(func2_)
3811		, func3	(func3_)
3812		, func4	(func4_)
3813	{}
3814
3815	const Func<Sig>&		func;
3816	const GenFunc<Sig, 2>&	func2;
3817	const GenFunc<Sig, 3>&	func3;
3818	const GenFunc<Sig, 4>&	func4;
3819};
3820
3821template<typename F>
3822GenFuncs<typename F::Sig> makeVectorizedFuncs (void)
3823{
3824	return GenFuncs<typename F::Sig>(instance<F>(),
3825									 instance<VectorizedFunc<F, 2> >(),
3826									 instance<VectorizedFunc<F, 3> >(),
3827									 instance<VectorizedFunc<F, 4> >());
3828}
3829
3830template<int Size>
3831ExprP<Vector<float, Size> > operator*(const ExprP<Vector<float, Size> >& arg0,
3832									  const ExprP<Vector<float, Size> >& arg1)
3833{
3834	return app<VectorizedFunc<Mul, Size> >(arg0, arg1);
3835}
3836
3837template<int Size>
3838ExprP<Vector<float, Size> > operator*(const ExprP<Vector<float, Size> >&	arg0,
3839									  const ExprP<float>&					arg1)
3840{
3841	return app<FixedVecFunc<Mul, Size> >(arg0, arg1);
3842}
3843
3844template<int Size>
3845ExprP<Vector<float, Size> > operator/(const ExprP<Vector<float, Size> >&	arg0,
3846									  const ExprP<float>&					arg1)
3847{
3848	return app<FixedVecFunc<Div, Size> >(arg0, arg1);
3849}
3850
3851template<int Size>
3852ExprP<Vector<float, Size> > operator-(const ExprP<Vector<float, Size> >& arg0)
3853{
3854	return app<VectorizedFunc<Negate, Size> >(arg0);
3855}
3856
3857template<int Size>
3858ExprP<Vector<float, Size> > operator-(const ExprP<Vector<float, Size> >& arg0,
3859									  const ExprP<Vector<float, Size> >& arg1)
3860{
3861	return app<VectorizedFunc<Sub, Size> >(arg0, arg1);
3862}
3863
3864template<int LeftRows, int Middle, int RightCols>
3865ExprP<Matrix<float, LeftRows, RightCols> >
3866operator* (const ExprP<Matrix<float, LeftRows, Middle> >&	left,
3867		   const ExprP<Matrix<float, Middle, RightCols> >&	right)
3868{
3869	return app<MatMul<LeftRows, Middle, RightCols> >(left, right);
3870}
3871
3872template<int Rows, int Cols>
3873ExprP<Vector<float, Rows> > operator* (const ExprP<Vector<float, Cols> >&		left,
3874									   const ExprP<Matrix<float, Rows, Cols> >&	right)
3875{
3876	return app<VecMatMul<Rows, Cols> >(left, right);
3877}
3878
3879template<int Rows, int Cols>
3880ExprP<Vector<float, Cols> > operator* (const ExprP<Matrix<float, Rows, Cols> >&	left,
3881									   const ExprP<Vector<float, Rows> >&		right)
3882{
3883	return app<MatVecMul<Rows, Cols> >(left, right);
3884}
3885
3886template<int Rows, int Cols>
3887ExprP<Matrix<float, Rows, Cols> > operator* (const ExprP<Matrix<float, Rows, Cols> >&	left,
3888											 const ExprP<float>&						right)
3889{
3890	return app<ScalarMatFunc<Mul, Rows, Cols> >(left, right);
3891}
3892
3893template<int Rows, int Cols>
3894ExprP<Matrix<float, Rows, Cols> > operator+ (const ExprP<Matrix<float, Rows, Cols> >&	left,
3895											 const ExprP<Matrix<float, Rows, Cols> >&	right)
3896{
3897	return app<CompMatFunc<Add, Rows, Cols> >(left, right);
3898}
3899
3900template<int Rows, int Cols>
3901ExprP<Matrix<float, Rows, Cols> > operator- (const ExprP<Matrix<float, Rows, Cols> >&	mat)
3902{
3903	return app<MatNeg<Rows, Cols> >(mat);
3904}
3905
3906template <typename T>
3907class Sampling
3908{
3909public:
3910	virtual void	genFixeds	(const FloatFormat&, vector<T>&)			const {}
3911	virtual T		genRandom	(const FloatFormat&, Precision, Random&)	const { return T(); }
3912	virtual double	getWeight	(void)										const { return 0.0; }
3913};
3914
3915template <>
3916class DefaultSampling<Void> : public Sampling<Void>
3917{
3918public:
3919	void	genFixeds	(const FloatFormat&, vector<Void>& dst) const { dst.push_back(Void()); }
3920};
3921
3922template <>
3923class DefaultSampling<bool> : public Sampling<bool>
3924{
3925public:
3926	void	genFixeds	(const FloatFormat&, vector<bool>& dst) const
3927	{
3928		dst.push_back(true);
3929		dst.push_back(false);
3930	}
3931};
3932
3933template <>
3934class DefaultSampling<int> : public Sampling<int>
3935{
3936public:
3937	int		genRandom	(const FloatFormat&, Precision prec, Random& rnd) const
3938	{
3939		const int	exp		= rnd.getInt(0, getNumBits(prec)-2);
3940		const int	sign	= rnd.getBool() ? -1 : 1;
3941
3942		return sign * rnd.getInt(0, 1L << exp);
3943	}
3944
3945	void	genFixeds	(const FloatFormat&, vector<int>& dst) const
3946	{
3947		dst.push_back(0);
3948		dst.push_back(-1);
3949		dst.push_back(1);
3950	}
3951	double	getWeight	(void) const { return 1.0; }
3952
3953private:
3954	static inline int getNumBits (Precision prec)
3955	{
3956		switch (prec)
3957		{
3958			case glu::PRECISION_LOWP:		return 8;
3959			case glu::PRECISION_MEDIUMP:	return 16;
3960			case glu::PRECISION_HIGHP:		return 32;
3961			default:
3962				DE_ASSERT(false);
3963				return 0;
3964		}
3965	}
3966};
3967
3968template <>
3969class DefaultSampling<float> : public Sampling<float>
3970{
3971public:
3972	float	genRandom	(const FloatFormat& format, Precision prec, Random& rnd) const;
3973	void	genFixeds	(const FloatFormat& format, vector<float>& dst) const;
3974	double	getWeight	(void) const { return 1.0; }
3975};
3976
3977//! Generate a random float from a reasonable general-purpose distribution.
3978float DefaultSampling<float>::genRandom (const FloatFormat& format,
3979										 Precision,
3980										 Random&			rnd) const
3981{
3982	const int		minExp			= format.getMinExp();
3983	const int		maxExp			= format.getMaxExp();
3984	const bool		haveSubnormal	= format.hasSubnormal() != tcu::NO;
3985
3986	// Choose exponent so that the cumulative distribution is cubic.
3987	// This makes the probability distribution quadratic, with the peak centered on zero.
3988	const double	minRoot			= deCbrt(minExp - 0.5 - (haveSubnormal ? 1.0 : 0.0));
3989	const double	maxRoot			= deCbrt(maxExp + 0.5);
3990	const int		fractionBits	= format.getFractionBits();
3991	const int		exp				= int(deRoundEven(dePow(rnd.getDouble(minRoot, maxRoot),
3992															3.0)));
3993	float			base			= 0.0f; // integral power of two
3994	float			quantum			= 0.0f; // smallest representable difference in the binade
3995	float			significand		= 0.0f; // Significand.
3996
3997	DE_ASSERT(fractionBits < std::numeric_limits<float>::digits);
3998
3999	// Generate some occasional special numbers
4000	switch (rnd.getInt(0, 64))
4001	{
4002		case 0: 	return 0;
4003		case 1:		return TCU_INFINITY;
4004		case 2:		return -TCU_INFINITY;
4005		case 3:		return TCU_NAN;
4006		default:	break;
4007	}
4008
4009	if (exp >= minExp)
4010	{
4011		// Normal number
4012		base = deFloatLdExp(1.0f, exp);
4013		quantum = deFloatLdExp(1.0f, exp - fractionBits);
4014	}
4015	else
4016	{
4017		// Subnormal
4018		base = 0.0f;
4019		quantum = deFloatLdExp(1.0f, minExp - fractionBits);
4020	}
4021
4022	switch (rnd.getInt(0, 16))
4023	{
4024		case 0: // The highest number in this binade, significand is all bits one.
4025			significand = base - quantum;
4026			break;
4027		case 1: // Significand is one.
4028			significand = quantum;
4029			break;
4030		case 2: // Significand is zero.
4031			significand = 0.0;
4032			break;
4033		default: // Random (evenly distributed) significand.
4034		{
4035			deUint64 intFraction = rnd.getUint64() & ((1 << fractionBits) - 1);
4036			significand = float(intFraction) * quantum;
4037		}
4038	}
4039
4040	// Produce positive numbers more often than negative.
4041	return (rnd.getInt(0,3) == 0 ? -1.0f : 1.0f) * (base + significand);
4042}
4043
4044//! Generate a standard set of floats that should always be tested.
4045void DefaultSampling<float>::genFixeds (const FloatFormat& format, vector<float>& dst) const
4046{
4047	const int			minExp			= format.getMinExp();
4048	const int			maxExp			= format.getMaxExp();
4049	const int			fractionBits	= format.getFractionBits();
4050	const float			minQuantum		= deFloatLdExp(1.0f, minExp - fractionBits);
4051	const float			minNormalized	= deFloatLdExp(1.0f, minExp);
4052	const float			maxQuantum		= deFloatLdExp(1.0f, maxExp - fractionBits);
4053
4054	// NaN
4055	dst.push_back(TCU_NAN);
4056	// Zero
4057	dst.push_back(0.0f);
4058
4059	for (int sign = -1; sign <= 1; sign += 2)
4060	{
4061		// Smallest subnormal
4062		dst.push_back(sign * minQuantum);
4063
4064		// Largest subnormal
4065		dst.push_back(sign * (minNormalized - minQuantum));
4066
4067		// Smallest normalized
4068		dst.push_back(sign * minNormalized);
4069
4070		// Next smallest normalized
4071		dst.push_back(sign * (minNormalized + minQuantum));
4072
4073		dst.push_back(sign * 0.5f);
4074		dst.push_back(sign * 1.0f);
4075		dst.push_back(sign * 2.0f);
4076
4077		// Largest number
4078		dst.push_back(sign * (deFloatLdExp(1.0f, maxExp) +
4079							  (deFloatLdExp(1.0f, maxExp) - maxQuantum)));
4080
4081		dst.push_back(sign * TCU_INFINITY);
4082	}
4083}
4084
4085template <typename T, int Size>
4086class DefaultSampling<Vector<T, Size> > : public Sampling<Vector<T, Size> >
4087{
4088public:
4089	typedef Vector<T, Size>		Value;
4090
4091	Value	genRandom	(const FloatFormat& fmt, Precision prec, Random& rnd) const
4092	{
4093		Value ret;
4094
4095		for (int ndx = 0; ndx < Size; ++ndx)
4096			ret[ndx] = instance<DefaultSampling<T> >().genRandom(fmt, prec, rnd);
4097
4098		return ret;
4099	}
4100
4101	void	genFixeds	(const FloatFormat& fmt, vector<Value>& dst) const
4102	{
4103		vector<T> scalars;
4104
4105		instance<DefaultSampling<T> >().genFixeds(fmt, scalars);
4106
4107		for (size_t scalarNdx = 0; scalarNdx < scalars.size(); ++scalarNdx)
4108			dst.push_back(Value(scalars[scalarNdx]));
4109	}
4110
4111	double	getWeight	(void) const
4112	{
4113		return dePow(instance<DefaultSampling<T> >().getWeight(), Size);
4114	}
4115};
4116
4117template <typename T, int Rows, int Columns>
4118class DefaultSampling<Matrix<T, Rows, Columns> > : public Sampling<Matrix<T, Rows, Columns> >
4119{
4120public:
4121	typedef Matrix<T, Rows, Columns>		Value;
4122
4123	Value	genRandom	(const FloatFormat& fmt, Precision prec, Random& rnd) const
4124	{
4125		Value ret;
4126
4127		for (int rowNdx = 0; rowNdx < Rows; ++rowNdx)
4128			for (int colNdx = 0; colNdx < Columns; ++colNdx)
4129				ret(rowNdx, colNdx) = instance<DefaultSampling<T> >().genRandom(fmt, prec, rnd);
4130
4131		return ret;
4132	}
4133
4134	void	genFixeds	(const FloatFormat& fmt, vector<Value>& dst) const
4135	{
4136		vector<T> scalars;
4137
4138		instance<DefaultSampling<T> >().genFixeds(fmt, scalars);
4139
4140		for (size_t scalarNdx = 0; scalarNdx < scalars.size(); ++scalarNdx)
4141			dst.push_back(Value(scalars[scalarNdx]));
4142
4143		if (Columns == Rows)
4144		{
4145			Value	mat	(0.0);
4146			T		x	= T(1.0f);
4147			mat[0][0] = x;
4148			for (int ndx = 0; ndx < Columns; ++ndx)
4149			{
4150				mat[Columns-1-ndx][ndx] = x;
4151				x *= T(2.0f);
4152			}
4153			dst.push_back(mat);
4154		}
4155	}
4156
4157	double	getWeight	(void) const
4158	{
4159		return dePow(instance<DefaultSampling<T> >().getWeight(), Rows * Columns);
4160	}
4161};
4162
4163struct Context
4164{
4165	Context		(const string&		name_,
4166				 TestContext&		testContext_,
4167				 RenderContext&		renderContext_,
4168				 const FloatFormat&	floatFormat_,
4169				 const FloatFormat&	highpFormat_,
4170				 Precision			precision_,
4171				 ShaderType			shaderType_,
4172				 size_t				numRandoms_)
4173		: name				(name_)
4174		, testContext		(testContext_)
4175		, renderContext		(renderContext_)
4176		, floatFormat		(floatFormat_)
4177		, highpFormat		(highpFormat_)
4178		, precision			(precision_)
4179		, shaderType		(shaderType_)
4180		, numRandoms		(numRandoms_) {}
4181
4182	string				name;
4183	TestContext&		testContext;
4184	RenderContext&		renderContext;
4185	FloatFormat			floatFormat;
4186	FloatFormat			highpFormat;
4187	Precision			precision;
4188	ShaderType			shaderType;
4189	size_t				numRandoms;
4190};
4191
4192template<typename In0_ = Void, typename In1_ = Void, typename In2_ = Void, typename In3_ = Void>
4193struct InTypes
4194{
4195	typedef	In0_	In0;
4196	typedef	In1_	In1;
4197	typedef	In2_	In2;
4198	typedef	In3_	In3;
4199};
4200
4201template <typename In>
4202int numInputs (void)
4203{
4204	return (!isTypeValid<typename In::In0>() ? 0 :
4205			!isTypeValid<typename In::In1>() ? 1 :
4206			!isTypeValid<typename In::In2>() ? 2 :
4207			!isTypeValid<typename In::In3>() ? 3 :
4208			4);
4209}
4210
4211template<typename Out0_, typename Out1_ = Void>
4212struct OutTypes
4213{
4214	typedef	Out0_	Out0;
4215	typedef	Out1_	Out1;
4216};
4217
4218template <typename Out>
4219int numOutputs (void)
4220{
4221	return (!isTypeValid<typename Out::Out0>() ? 0 :
4222			!isTypeValid<typename Out::Out1>() ? 1 :
4223			2);
4224}
4225
4226template<typename In>
4227struct Inputs
4228{
4229	vector<typename In::In0>	in0;
4230	vector<typename In::In1>	in1;
4231	vector<typename In::In2>	in2;
4232	vector<typename In::In3>	in3;
4233};
4234
4235template<typename Out>
4236struct Outputs
4237{
4238	Outputs	(size_t size) : out0(size), out1(size) {}
4239
4240	vector<typename Out::Out0>	out0;
4241	vector<typename Out::Out1>	out1;
4242};
4243
4244template<typename In, typename Out>
4245struct Variables
4246{
4247	VariableP<typename In::In0>		in0;
4248	VariableP<typename In::In1>		in1;
4249	VariableP<typename In::In2>		in2;
4250	VariableP<typename In::In3>		in3;
4251	VariableP<typename Out::Out0>	out0;
4252	VariableP<typename Out::Out1>	out1;
4253};
4254
4255template<typename In>
4256struct Samplings
4257{
4258	Samplings	(const Sampling<typename In::In0>&	in0_,
4259				 const Sampling<typename In::In1>&	in1_,
4260				 const Sampling<typename In::In2>&	in2_,
4261				 const Sampling<typename In::In3>&	in3_)
4262		: in0 (in0_), in1 (in1_), in2 (in2_), in3 (in3_) {}
4263
4264	const Sampling<typename In::In0>&	in0;
4265	const Sampling<typename In::In1>&	in1;
4266	const Sampling<typename In::In2>&	in2;
4267	const Sampling<typename In::In3>&	in3;
4268};
4269
4270template<typename In>
4271struct DefaultSamplings : Samplings<In>
4272{
4273	DefaultSamplings	(void)
4274		: Samplings<In>(instance<DefaultSampling<typename In::In0> >(),
4275						instance<DefaultSampling<typename In::In1> >(),
4276						instance<DefaultSampling<typename In::In2> >(),
4277						instance<DefaultSampling<typename In::In3> >()) {}
4278};
4279
4280class PrecisionCase : public TestCase
4281{
4282public:
4283	IterateResult		iterate			(void);
4284
4285protected:
4286						PrecisionCase	(const Context&		context,
4287										 const string&		name,
4288										 const string&		extension	= "")
4289							: TestCase		(context.testContext,
4290											 name.c_str(),
4291											 name.c_str())
4292							, m_ctx			(context)
4293							, m_status		()
4294							, m_rnd			(0xdeadbeefu +
4295											 context.testContext.getCommandLine().getBaseSeed())
4296							, m_extension	(extension)
4297	{
4298	}
4299
4300	RenderContext&		getRenderContext(void) const 			{ return m_ctx.renderContext; }
4301
4302	const FloatFormat&	getFormat		(void) const 			{ return m_ctx.floatFormat; }
4303
4304	TestLog&			log				(void) const 			{ return m_testCtx.getLog(); }
4305
4306	virtual void		runTest			(void) = 0;
4307
4308	template <typename In, typename Out>
4309	void				testStatement	(const Variables<In, Out>&	variables,
4310										 const Inputs<In>&			inputs,
4311										 const Statement&			stmt);
4312
4313	template<typename T>
4314	Symbol				makeSymbol		(const Variable<T>& variable)
4315	{
4316		return Symbol(variable.getName(), getVarTypeOf<T>(m_ctx.precision));
4317	}
4318
4319	Context				m_ctx;
4320	ResultCollector		m_status;
4321	Random				m_rnd;
4322	const string		m_extension;
4323};
4324
4325IterateResult PrecisionCase::iterate (void)
4326{
4327	runTest();
4328	m_status.setTestContextResult(m_testCtx);
4329	return STOP;
4330}
4331
4332template <typename In, typename Out>
4333void PrecisionCase::testStatement (const Variables<In, Out>&	variables,
4334								   const Inputs<In>&			inputs,
4335								   const Statement&				stmt)
4336{
4337	using namespace ShaderExecUtil;
4338
4339	typedef typename 	In::In0		In0;
4340	typedef typename 	In::In1		In1;
4341	typedef typename 	In::In2		In2;
4342	typedef typename 	In::In3		In3;
4343	typedef typename 	Out::Out0	Out0;
4344	typedef typename 	Out::Out1	Out1;
4345
4346	const FloatFormat&	fmt			= getFormat();
4347	const int			inCount		= numInputs<In>();
4348	const int			outCount	= numOutputs<Out>();
4349	const size_t		numValues	= (inCount > 0) ? inputs.in0.size() : 1;
4350	Outputs<Out>		outputs		(numValues);
4351	ShaderSpec			spec;
4352	const FloatFormat	highpFmt	= m_ctx.highpFormat;
4353	const int			maxMsgs		= 100;
4354	int					numErrors	= 0;
4355	Environment			env; 		// Hoisted out of the inner loop for optimization.
4356
4357	switch (inCount)
4358	{
4359		case 4: DE_ASSERT(inputs.in3.size() == numValues);
4360		case 3: DE_ASSERT(inputs.in2.size() == numValues);
4361		case 2: DE_ASSERT(inputs.in1.size() == numValues);
4362		case 1: DE_ASSERT(inputs.in0.size() == numValues);
4363		default: break;
4364	}
4365
4366	// Print out the statement and its definitions
4367	log() << TestLog::Message << "Statement: " << stmt << TestLog::EndMessage;
4368	{
4369		ostringstream	oss;
4370		FuncSet			funcs;
4371
4372		stmt.getUsedFuncs(funcs);
4373		for (FuncSet::const_iterator it = funcs.begin(); it != funcs.end(); ++it)
4374		{
4375			(*it)->printDefinition(oss);
4376		}
4377		if (!funcs.empty())
4378			log() << TestLog::Message << "Reference definitions:\n" << oss.str()
4379				  << TestLog::EndMessage;
4380	}
4381
4382	// Initialize ShaderSpec from precision, variables and statement.
4383	{
4384		ostringstream os;
4385		os << "precision " << glu::getPrecisionName(m_ctx.precision) << " float;\n";
4386		spec.globalDeclarations = os.str();
4387	}
4388
4389	spec.version = getContextTypeGLSLVersion(getRenderContext().getType());
4390
4391	if (!m_extension.empty())
4392		spec.globalDeclarations = "#extension " + m_extension + " : require\n";
4393
4394	spec.inputs.resize(inCount);
4395
4396	switch (inCount)
4397	{
4398		case 4: spec.inputs[3] = makeSymbol(*variables.in3);
4399		case 3:	spec.inputs[2] = makeSymbol(*variables.in2);
4400		case 2:	spec.inputs[1] = makeSymbol(*variables.in1);
4401		case 1:	spec.inputs[0] = makeSymbol(*variables.in0);
4402		default: break;
4403	}
4404
4405	spec.outputs.resize(outCount);
4406
4407	switch (outCount)
4408	{
4409		case 2:	spec.outputs[1] = makeSymbol(*variables.out1);
4410		case 1:	spec.outputs[0] = makeSymbol(*variables.out0);
4411		default: break;
4412	}
4413
4414	spec.source = de::toString(stmt);
4415
4416	// Run the shader with inputs.
4417	{
4418		UniquePtr<ShaderExecutor>	executor		(createExecutor(getRenderContext(),
4419																	m_ctx.shaderType,
4420																	spec));
4421		const void*					inputArr[]		=
4422		{
4423			&inputs.in0.front(), &inputs.in1.front(), &inputs.in2.front(), &inputs.in3.front(),
4424		};
4425		void*						outputArr[]		=
4426		{
4427			&outputs.out0.front(), &outputs.out1.front(),
4428		};
4429
4430		executor->log(log());
4431		if (!executor->isOk())
4432			TCU_FAIL("Shader compilation failed");
4433
4434		executor->useProgram();
4435		executor->execute(int(numValues), inputArr, outputArr);
4436	}
4437
4438	// Initialize environment with dummy values so we don't need to bind in inner loop.
4439	{
4440		const typename Traits<In0>::IVal		in0;
4441		const typename Traits<In1>::IVal		in1;
4442		const typename Traits<In2>::IVal		in2;
4443		const typename Traits<In3>::IVal		in3;
4444		const typename Traits<Out0>::IVal		reference0;
4445		const typename Traits<Out1>::IVal		reference1;
4446
4447		env.bind(*variables.in0, in0);
4448		env.bind(*variables.in1, in1);
4449		env.bind(*variables.in2, in2);
4450		env.bind(*variables.in3, in3);
4451		env.bind(*variables.out0, reference0);
4452		env.bind(*variables.out1, reference1);
4453	}
4454
4455	// For each input tuple, compute output reference interval and compare
4456	// shader output to the reference.
4457	for (size_t valueNdx = 0; valueNdx < numValues; valueNdx++)
4458	{
4459		bool						result		= true;
4460		typename Traits<Out0>::IVal	reference0;
4461		typename Traits<Out1>::IVal	reference1;
4462
4463		env.lookup(*variables.in0) = convert<In0>(fmt, round(fmt, inputs.in0[valueNdx]));
4464		env.lookup(*variables.in1) = convert<In1>(fmt, round(fmt, inputs.in1[valueNdx]));
4465		env.lookup(*variables.in2) = convert<In2>(fmt, round(fmt, inputs.in2[valueNdx]));
4466		env.lookup(*variables.in3) = convert<In3>(fmt, round(fmt, inputs.in3[valueNdx]));
4467
4468		{
4469			EvalContext	ctx (fmt, m_ctx.precision, env);
4470			stmt.execute(ctx);
4471		}
4472
4473		switch (outCount)
4474		{
4475			case 2:
4476				reference1 = convert<Out1>(highpFmt, env.lookup(*variables.out1));
4477				if (!m_status.check(contains(reference1, outputs.out1[valueNdx]),
4478									"Shader output 1 is outside acceptable range"))
4479					result = false;
4480			case 1:
4481				reference0 = convert<Out0>(highpFmt, env.lookup(*variables.out0));
4482				if (!m_status.check(contains(reference0, outputs.out0[valueNdx]),
4483									"Shader output 0 is outside acceptable range"))
4484					result = false;
4485			default: break;
4486		}
4487
4488		if (!result)
4489			++numErrors;
4490
4491		if ((!result && numErrors <= maxMsgs) || GLS_LOG_ALL_RESULTS)
4492		{
4493			MessageBuilder	builder	= log().message();
4494
4495			builder << (result ? "Passed" : "Failed") << " sample:\n";
4496
4497			if (inCount > 0)
4498			{
4499				builder << "\t" << variables.in0->getName() << " = "
4500						<< valueToString(highpFmt, inputs.in0[valueNdx]) << "\n";
4501			}
4502
4503			if (inCount > 1)
4504			{
4505				builder << "\t" << variables.in1->getName() << " = "
4506						<< valueToString(highpFmt, inputs.in1[valueNdx]) << "\n";
4507			}
4508
4509			if (inCount > 2)
4510			{
4511				builder << "\t" << variables.in2->getName() << " = "
4512						<< valueToString(highpFmt, inputs.in2[valueNdx]) << "\n";
4513			}
4514
4515			if (inCount > 3)
4516			{
4517				builder << "\t" << variables.in3->getName() << " = "
4518						<< valueToString(highpFmt, inputs.in3[valueNdx]) << "\n";
4519			}
4520
4521			if (outCount > 0)
4522			{
4523				builder << "\t" << variables.out0->getName() << " = "
4524						<< valueToString(highpFmt, outputs.out0[valueNdx]) << "\n"
4525						<< "\tExpected range: "
4526						<< intervalToString<typename Out::Out0>(highpFmt, reference0) << "\n";
4527			}
4528
4529			if (outCount > 1)
4530			{
4531				builder << "\t" << variables.out1->getName() << " = "
4532						<< valueToString(highpFmt, outputs.out1[valueNdx]) << "\n"
4533						<< "\tExpected range: "
4534						<< intervalToString<typename Out::Out1>(highpFmt, reference1) << "\n";
4535			}
4536
4537			builder << TestLog::EndMessage;
4538		}
4539	}
4540
4541	if (numErrors > maxMsgs)
4542	{
4543		log() << TestLog::Message << "(Skipped " << (numErrors - maxMsgs) << " messages.)"
4544			  << TestLog::EndMessage;
4545	}
4546
4547	if (numErrors == 0)
4548	{
4549		log() << TestLog::Message << "All " << numValues << " inputs passed."
4550			  << TestLog::EndMessage;
4551	}
4552	else
4553	{
4554		log() << TestLog::Message << numErrors << "/" << numValues << " inputs failed."
4555			  << TestLog::EndMessage;
4556	}
4557}
4558
4559
4560
4561template <typename T>
4562struct InputLess
4563{
4564	bool operator() (const T& val1, const T& val2) const
4565	{
4566		return val1 < val2;
4567	}
4568};
4569
4570template <typename T>
4571bool inputLess (const T& val1, const T& val2)
4572{
4573	return InputLess<T>()(val1, val2);
4574}
4575
4576template <>
4577struct InputLess<float>
4578{
4579	bool operator() (const float& val1, const float& val2) const
4580	{
4581		if (deIsNaN(val1))
4582			return false;
4583		if (deIsNaN(val2))
4584			return true;
4585		return val1 < val2;
4586	}
4587};
4588
4589template <typename T, int Size>
4590struct InputLess<Vector<T, Size> >
4591{
4592	bool operator() (const Vector<T, Size>& vec1, const Vector<T, Size>& vec2) const
4593	{
4594		for (int ndx = 0; ndx < Size; ++ndx)
4595		{
4596			if (inputLess(vec1[ndx], vec2[ndx]))
4597				return true;
4598			if (inputLess(vec2[ndx], vec1[ndx]))
4599				return false;
4600		}
4601
4602		return false;
4603	}
4604};
4605
4606template <typename T, int Rows, int Cols>
4607struct InputLess<Matrix<T, Rows, Cols> >
4608{
4609	bool operator() (const Matrix<T, Rows, Cols>& mat1,
4610					 const Matrix<T, Rows, Cols>& mat2) const
4611	{
4612		for (int col = 0; col < Cols; ++col)
4613		{
4614			if (inputLess(mat1[col], mat2[col]))
4615				return true;
4616			if (inputLess(mat2[col], mat1[col]))
4617				return false;
4618		}
4619
4620		return false;
4621	}
4622};
4623
4624template <typename In>
4625struct InTuple :
4626	public Tuple4<typename In::In0, typename In::In1, typename In::In2, typename In::In3>
4627{
4628	InTuple	(const typename In::In0& in0,
4629			 const typename In::In1& in1,
4630			 const typename In::In2& in2,
4631			 const typename In::In3& in3)
4632		: Tuple4<typename In::In0, typename In::In1, typename In::In2, typename In::In3>
4633		  (in0, in1, in2, in3) {}
4634};
4635
4636template <typename In>
4637struct InputLess<InTuple<In> >
4638{
4639	bool operator() (const InTuple<In>& in1, const InTuple<In>& in2) const
4640	{
4641		if (inputLess(in1.a, in2.a))
4642			return true;
4643		if (inputLess(in2.a, in1.a))
4644			return false;
4645		if (inputLess(in1.b, in2.b))
4646			return true;
4647		if (inputLess(in2.b, in1.b))
4648			return false;
4649		if (inputLess(in1.c, in2.c))
4650			return true;
4651		if (inputLess(in2.c, in1.c))
4652			return false;
4653		if (inputLess(in1.d, in2.d))
4654			return true;
4655		return false;
4656	};
4657};
4658
4659template<typename In>
4660Inputs<In> generateInputs (const Samplings<In>&	samplings,
4661						   const FloatFormat&	floatFormat,
4662						   Precision			intPrecision,
4663						   size_t				numSamples,
4664						   Random&				rnd)
4665{
4666	Inputs<In>									ret;
4667	Inputs<In>									fixedInputs;
4668	set<InTuple<In>, InputLess<InTuple<In> > >	seenInputs;
4669
4670	samplings.in0.genFixeds(floatFormat, fixedInputs.in0);
4671	samplings.in1.genFixeds(floatFormat, fixedInputs.in1);
4672	samplings.in2.genFixeds(floatFormat, fixedInputs.in2);
4673	samplings.in3.genFixeds(floatFormat, fixedInputs.in3);
4674
4675	for (size_t ndx0 = 0; ndx0 < fixedInputs.in0.size(); ++ndx0)
4676	{
4677		for (size_t ndx1 = 0; ndx1 < fixedInputs.in1.size(); ++ndx1)
4678		{
4679			for (size_t ndx2 = 0; ndx2 < fixedInputs.in2.size(); ++ndx2)
4680			{
4681				for (size_t ndx3 = 0; ndx3 < fixedInputs.in3.size(); ++ndx3)
4682				{
4683					const InTuple<In>	tuple	(fixedInputs.in0[ndx0],
4684												 fixedInputs.in1[ndx1],
4685												 fixedInputs.in2[ndx2],
4686												 fixedInputs.in3[ndx3]);
4687
4688					seenInputs.insert(tuple);
4689					ret.in0.push_back(tuple.a);
4690					ret.in1.push_back(tuple.b);
4691					ret.in2.push_back(tuple.c);
4692					ret.in3.push_back(tuple.d);
4693				}
4694			}
4695		}
4696	}
4697
4698	for (size_t ndx = 0; ndx < numSamples; ++ndx)
4699	{
4700		const typename In::In0	in0		= samplings.in0.genRandom(floatFormat, intPrecision, rnd);
4701		const typename In::In1	in1		= samplings.in1.genRandom(floatFormat, intPrecision, rnd);
4702		const typename In::In2	in2		= samplings.in2.genRandom(floatFormat, intPrecision, rnd);
4703		const typename In::In3	in3		= samplings.in3.genRandom(floatFormat, intPrecision, rnd);
4704		const InTuple<In>		tuple	(in0, in1, in2, in3);
4705
4706		if (de::contains(seenInputs, tuple))
4707			continue;
4708
4709		seenInputs.insert(tuple);
4710		ret.in0.push_back(in0);
4711		ret.in1.push_back(in1);
4712		ret.in2.push_back(in2);
4713		ret.in3.push_back(in3);
4714	}
4715
4716	return ret;
4717}
4718
4719class FuncCaseBase : public PrecisionCase
4720{
4721public:
4722	IterateResult	iterate			(void);
4723
4724protected:
4725					FuncCaseBase	(const Context&		context,
4726									 const string&		name,
4727									 const FuncBase&	func)
4728						: PrecisionCase	(context, name, func.getRequiredExtension()) {}
4729};
4730
4731IterateResult FuncCaseBase::iterate (void)
4732{
4733	MovePtr<ContextInfo>	info	(ContextInfo::create(getRenderContext()));
4734
4735	if (!m_extension.empty() && !info->isExtensionSupported(m_extension.c_str()))
4736		throw NotSupportedError("Unsupported extension: " + m_extension);
4737
4738	runTest();
4739
4740	m_status.setTestContextResult(m_testCtx);
4741	return STOP;
4742}
4743
4744template <typename Sig>
4745class FuncCase : public FuncCaseBase
4746{
4747public:
4748	typedef Func<Sig>						CaseFunc;
4749	typedef typename Sig::Ret				Ret;
4750	typedef typename Sig::Arg0				Arg0;
4751	typedef typename Sig::Arg1				Arg1;
4752	typedef typename Sig::Arg2				Arg2;
4753	typedef typename Sig::Arg3				Arg3;
4754	typedef InTypes<Arg0, Arg1, Arg2, Arg3>	In;
4755	typedef OutTypes<Ret>					Out;
4756
4757					FuncCase	(const Context&		context,
4758								 const string&		name,
4759								 const CaseFunc&	func)
4760						: FuncCaseBase	(context, name, func)
4761						, m_func		(func) {}
4762
4763protected:
4764	void				runTest		(void);
4765
4766	virtual const Samplings<In>&	getSamplings	(void)
4767	{
4768		return instance<DefaultSamplings<In> >();
4769	}
4770
4771private:
4772	const CaseFunc&			m_func;
4773};
4774
4775template <typename Sig>
4776void FuncCase<Sig>::runTest (void)
4777{
4778	const Inputs<In>	inputs	(generateInputs(getSamplings(),
4779												m_ctx.floatFormat,
4780												m_ctx.precision,
4781												m_ctx.numRandoms,
4782												m_rnd));
4783	Variables<In, Out>	variables;
4784
4785	variables.out0	= variable<Ret>("out0");
4786	variables.out1	= variable<Void>("out1");
4787	variables.in0	= variable<Arg0>("in0");
4788	variables.in1	= variable<Arg1>("in1");
4789	variables.in2	= variable<Arg2>("in2");
4790	variables.in3	= variable<Arg3>("in3");
4791
4792	{
4793		ExprP<Ret>	expr	= applyVar(m_func,
4794									   variables.in0, variables.in1,
4795									   variables.in2, variables.in3);
4796		StatementP	stmt	= variableAssignment(variables.out0, expr);
4797
4798		this->testStatement(variables, inputs, *stmt);
4799	}
4800}
4801
4802template <typename Sig>
4803class InOutFuncCase : public FuncCaseBase
4804{
4805public:
4806	typedef Func<Sig>						CaseFunc;
4807	typedef typename Sig::Ret				Ret;
4808	typedef typename Sig::Arg0				Arg0;
4809	typedef typename Sig::Arg1				Arg1;
4810	typedef typename Sig::Arg2				Arg2;
4811	typedef typename Sig::Arg3				Arg3;
4812	typedef InTypes<Arg0, Arg2, Arg3>		In;
4813	typedef OutTypes<Ret, Arg1>				Out;
4814
4815					InOutFuncCase	(const Context&		context,
4816									 const string&		name,
4817									 const CaseFunc&	func)
4818						: FuncCaseBase	(context, name, func)
4819						, m_func		(func) {}
4820
4821protected:
4822	void				runTest		(void);
4823
4824	virtual const Samplings<In>&	getSamplings	(void)
4825	{
4826		return instance<DefaultSamplings<In> >();
4827	}
4828
4829private:
4830	const CaseFunc&			m_func;
4831};
4832
4833template <typename Sig>
4834void InOutFuncCase<Sig>::runTest (void)
4835{
4836	const Inputs<In>	inputs	(generateInputs(getSamplings(),
4837												m_ctx.floatFormat,
4838												m_ctx.precision,
4839												m_ctx.numRandoms,
4840												m_rnd));
4841	Variables<In, Out>	variables;
4842
4843	variables.out0	= variable<Ret>("out0");
4844	variables.out1	= variable<Arg1>("out1");
4845	variables.in0	= variable<Arg0>("in0");
4846	variables.in1	= variable<Arg2>("in1");
4847	variables.in2	= variable<Arg3>("in2");
4848	variables.in3	= variable<Void>("in3");
4849
4850	{
4851		ExprP<Ret>	expr	= applyVar(m_func,
4852									   variables.in0, variables.out1,
4853									   variables.in1, variables.in2);
4854		StatementP	stmt	= variableAssignment(variables.out0, expr);
4855
4856		this->testStatement(variables, inputs, *stmt);
4857	}
4858}
4859
4860template <typename Sig>
4861PrecisionCase* createFuncCase (const Context&	context,
4862							   const string&	name,
4863							   const Func<Sig>&	func)
4864{
4865	switch (func.getOutParamIndex())
4866	{
4867		case -1:
4868			return new FuncCase<Sig>(context, name, func);
4869		case 1:
4870			return new InOutFuncCase<Sig>(context, name, func);
4871		default:
4872			DE_ASSERT(!"Impossible");
4873	}
4874	return DE_NULL;
4875}
4876
4877class CaseFactory
4878{
4879public:
4880	virtual						~CaseFactory	(void) {}
4881	virtual MovePtr<TestNode>	createCase		(const Context& ctx) const = 0;
4882	virtual string				getName			(void) const = 0;
4883	virtual string				getDesc			(void) const = 0;
4884};
4885
4886class FuncCaseFactory : public CaseFactory
4887{
4888public:
4889	virtual const FuncBase&	getFunc		(void) const = 0;
4890
4891	string					getName		(void) const
4892	{
4893		return de::toLower(getFunc().getName());
4894	}
4895
4896	string					getDesc		(void) const
4897	{
4898		return "Function '" + getFunc().getName() + "'";
4899	}
4900};
4901
4902template <typename Sig>
4903class GenFuncCaseFactory : public CaseFactory
4904{
4905public:
4906
4907						GenFuncCaseFactory	(const GenFuncs<Sig>&	funcs,
4908											 const string&			name)
4909							: m_funcs	(funcs)
4910							, m_name	(de::toLower(name)) {}
4911
4912	MovePtr<TestNode>	createCase			(const Context& ctx) const
4913	{
4914		TestCaseGroup*	group = new TestCaseGroup(ctx.testContext,
4915												  ctx.name.c_str(), ctx.name.c_str());
4916
4917		group->addChild(createFuncCase(ctx, "scalar", m_funcs.func));
4918		group->addChild(createFuncCase(ctx, "vec2", m_funcs.func2));
4919		group->addChild(createFuncCase(ctx, "vec3", m_funcs.func3));
4920		group->addChild(createFuncCase(ctx, "vec4", m_funcs.func4));
4921
4922		return MovePtr<TestNode>(group);
4923	}
4924
4925	string				getName				(void) const
4926	{
4927		return m_name;
4928	}
4929
4930	string				getDesc				(void) const
4931	{
4932		return "Function '" + m_funcs.func.getName() + "'";
4933	}
4934
4935private:
4936	const GenFuncs<Sig>	m_funcs;
4937	string				m_name;
4938};
4939
4940template <template <int> class GenF>
4941class TemplateFuncCaseFactory : public FuncCaseFactory
4942{
4943public:
4944	MovePtr<TestNode>	createCase		(const Context& ctx) const
4945	{
4946		TestCaseGroup*	group = new TestCaseGroup(ctx.testContext,
4947							  ctx.name.c_str(), ctx.name.c_str());
4948		group->addChild(createFuncCase(ctx, "scalar", instance<GenF<1> >()));
4949		group->addChild(createFuncCase(ctx, "vec2", instance<GenF<2> >()));
4950		group->addChild(createFuncCase(ctx, "vec3", instance<GenF<3> >()));
4951		group->addChild(createFuncCase(ctx, "vec4", instance<GenF<4> >()));
4952
4953		return MovePtr<TestNode>(group);
4954	}
4955
4956	const FuncBase&		getFunc			(void) const { return instance<GenF<1> >(); }
4957};
4958
4959template <template <int> class GenF>
4960class SquareMatrixFuncCaseFactory : public FuncCaseFactory
4961{
4962public:
4963	MovePtr<TestNode>	createCase		(const Context& ctx) const
4964	{
4965		TestCaseGroup*	group = new TestCaseGroup(ctx.testContext,
4966							  ctx.name.c_str(), ctx.name.c_str());
4967		group->addChild(createFuncCase(ctx, "mat2", instance<GenF<2> >()));
4968#if 0
4969		// disabled until we get reasonable results
4970		group->addChild(createFuncCase(ctx, "mat3", instance<GenF<3> >()));
4971		group->addChild(createFuncCase(ctx, "mat4", instance<GenF<4> >()));
4972#endif
4973
4974		return MovePtr<TestNode>(group);
4975	}
4976
4977	const FuncBase&		getFunc			(void) const { return instance<GenF<2> >(); }
4978};
4979
4980template <template <int, int> class GenF>
4981class MatrixFuncCaseFactory : public FuncCaseFactory
4982{
4983public:
4984	MovePtr<TestNode>	createCase		(const Context& ctx) const
4985	{
4986		TestCaseGroup*	const group = new TestCaseGroup(ctx.testContext,
4987														ctx.name.c_str(), ctx.name.c_str());
4988
4989		this->addCase<2, 2>(ctx, group);
4990		this->addCase<3, 2>(ctx, group);
4991		this->addCase<4, 2>(ctx, group);
4992		this->addCase<2, 3>(ctx, group);
4993		this->addCase<3, 3>(ctx, group);
4994		this->addCase<4, 3>(ctx, group);
4995		this->addCase<2, 4>(ctx, group);
4996		this->addCase<3, 4>(ctx, group);
4997		this->addCase<4, 4>(ctx, group);
4998
4999		return MovePtr<TestNode>(group);
5000	}
5001
5002	const FuncBase&		getFunc			(void) const { return instance<GenF<2,2> >(); }
5003
5004private:
5005	template <int Rows, int Cols>
5006	void				addCase			(const Context& ctx, TestCaseGroup* group) const
5007	{
5008		const char*	const name = dataTypeNameOf<Matrix<float, Rows, Cols> >();
5009
5010		group->addChild(createFuncCase(ctx, name, instance<GenF<Rows, Cols> >()));
5011	}
5012};
5013
5014template <typename Sig>
5015class SimpleFuncCaseFactory : public CaseFactory
5016{
5017public:
5018						SimpleFuncCaseFactory	(const Func<Sig>& func) : m_func(func) {}
5019
5020	MovePtr<TestNode>	createCase		(const Context& ctx) const
5021	{
5022		return MovePtr<TestNode>(createFuncCase(ctx, ctx.name.c_str(), m_func));
5023	}
5024
5025	string				getName					(void) const
5026	{
5027		return de::toLower(m_func.getName());
5028	}
5029
5030	string				getDesc					(void) const
5031	{
5032		return "Function '" + getName() + "'";
5033	}
5034
5035private:
5036	const Func<Sig>&	m_func;
5037};
5038
5039template <typename F>
5040SharedPtr<SimpleFuncCaseFactory<typename F::Sig> > createSimpleFuncCaseFactory (void)
5041{
5042	return SharedPtr<SimpleFuncCaseFactory<typename F::Sig> >(
5043		new SimpleFuncCaseFactory<typename F::Sig>(instance<F>()));
5044}
5045
5046class BuiltinFuncs : public CaseFactories
5047{
5048public:
5049	const vector<const CaseFactory*>	getFactories	(void) const
5050	{
5051		vector<const CaseFactory*> ret;
5052
5053		for (size_t ndx = 0; ndx < m_factories.size(); ++ndx)
5054			ret.push_back(m_factories[ndx].get());
5055
5056		return ret;
5057	}
5058
5059	void								addFactory		(SharedPtr<const CaseFactory> fact)
5060	{
5061		m_factories.push_back(fact);
5062	}
5063
5064private:
5065	vector<SharedPtr<const CaseFactory> >			m_factories;
5066};
5067
5068template <typename F>
5069void addScalarFactory(BuiltinFuncs& funcs, string name = "")
5070{
5071	if (name.empty())
5072		name = instance<F>().getName();
5073
5074	funcs.addFactory(SharedPtr<const CaseFactory>(new GenFuncCaseFactory<typename F::Sig>(
5075													  makeVectorizedFuncs<F>(), name)));
5076}
5077
5078MovePtr<const CaseFactories> createES3BuiltinCases (void)
5079{
5080	MovePtr<BuiltinFuncs>	funcs	(new BuiltinFuncs());
5081
5082	addScalarFactory<Add>(*funcs);
5083	addScalarFactory<Sub>(*funcs);
5084	addScalarFactory<Mul>(*funcs);
5085	addScalarFactory<Div>(*funcs);
5086
5087	addScalarFactory<Radians>(*funcs);
5088	addScalarFactory<Degrees>(*funcs);
5089	addScalarFactory<Sin>(*funcs);
5090	addScalarFactory<Cos>(*funcs);
5091	addScalarFactory<Tan>(*funcs);
5092	addScalarFactory<ASin>(*funcs);
5093	addScalarFactory<ACos>(*funcs);
5094	addScalarFactory<ATan2>(*funcs, "atan2");
5095	addScalarFactory<ATan>(*funcs);
5096	addScalarFactory<Sinh>(*funcs);
5097	addScalarFactory<Cosh>(*funcs);
5098	addScalarFactory<Tanh>(*funcs);
5099	addScalarFactory<ASinh>(*funcs);
5100	addScalarFactory<ACosh>(*funcs);
5101	addScalarFactory<ATanh>(*funcs);
5102
5103	addScalarFactory<Pow>(*funcs);
5104	addScalarFactory<Exp>(*funcs);
5105	addScalarFactory<Log>(*funcs);
5106	addScalarFactory<Exp2>(*funcs);
5107	addScalarFactory<Log2>(*funcs);
5108	addScalarFactory<Sqrt>(*funcs);
5109	addScalarFactory<InverseSqrt>(*funcs);
5110
5111	addScalarFactory<Abs>(*funcs);
5112	addScalarFactory<Sign>(*funcs);
5113	addScalarFactory<Floor>(*funcs);
5114	addScalarFactory<Trunc>(*funcs);
5115	addScalarFactory<Round>(*funcs);
5116	addScalarFactory<RoundEven>(*funcs);
5117	addScalarFactory<Ceil>(*funcs);
5118	addScalarFactory<Fract>(*funcs);
5119	addScalarFactory<Mod>(*funcs);
5120	funcs->addFactory(createSimpleFuncCaseFactory<Modf>());
5121	addScalarFactory<Min>(*funcs);
5122	addScalarFactory<Max>(*funcs);
5123	addScalarFactory<Clamp>(*funcs);
5124	addScalarFactory<Mix>(*funcs);
5125	addScalarFactory<Step>(*funcs);
5126	addScalarFactory<SmoothStep>(*funcs);
5127
5128	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Length>()));
5129	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Distance>()));
5130	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Dot>()));
5131	funcs->addFactory(createSimpleFuncCaseFactory<Cross>());
5132	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Normalize>()));
5133	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<FaceForward>()));
5134	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Reflect>()));
5135	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Refract>()));
5136
5137
5138	funcs->addFactory(SharedPtr<const CaseFactory>(new MatrixFuncCaseFactory<MatrixCompMult>()));
5139	funcs->addFactory(SharedPtr<const CaseFactory>(new MatrixFuncCaseFactory<OuterProduct>()));
5140	funcs->addFactory(SharedPtr<const CaseFactory>(new MatrixFuncCaseFactory<Transpose>()));
5141	funcs->addFactory(SharedPtr<const CaseFactory>(new SquareMatrixFuncCaseFactory<Determinant>()));
5142	funcs->addFactory(SharedPtr<const CaseFactory>(new SquareMatrixFuncCaseFactory<Inverse>()));
5143
5144	return MovePtr<const CaseFactories>(funcs.release());
5145}
5146
5147MovePtr<const CaseFactories> createES31BuiltinCases (void)
5148{
5149	MovePtr<BuiltinFuncs>	funcs	(new BuiltinFuncs());
5150
5151	addScalarFactory<FrExp>(*funcs);
5152	addScalarFactory<LdExp>(*funcs);
5153	addScalarFactory<Fma>(*funcs);
5154
5155	return MovePtr<const CaseFactories>(funcs.release());
5156}
5157
5158struct PrecisionTestContext
5159{
5160	PrecisionTestContext	(TestContext&				testCtx_,
5161							 RenderContext&				renderCtx_,
5162							 const FloatFormat&			highp_,
5163							 const FloatFormat&			mediump_,
5164							 const FloatFormat&			lowp_,
5165							 const vector<ShaderType>&	shaderTypes_,
5166							 int						numRandoms_)
5167		: testCtx		(testCtx_)
5168		, renderCtx		(renderCtx_)
5169		, shaderTypes	(shaderTypes_)
5170		, numRandoms	(numRandoms_)
5171	{
5172		formats[glu::PRECISION_HIGHP]	= &highp_;
5173		formats[glu::PRECISION_MEDIUMP]	= &mediump_;
5174		formats[glu::PRECISION_LOWP]	= &lowp_;
5175	}
5176
5177	TestContext&			testCtx;
5178	RenderContext&			renderCtx;
5179	const FloatFormat*		formats[glu::PRECISION_LAST];
5180	vector<ShaderType>		shaderTypes;
5181	int						numRandoms;
5182};
5183
5184TestCaseGroup* createFuncGroup (const PrecisionTestContext&	ctx,
5185								const CaseFactory&			factory)
5186{
5187	TestCaseGroup* const 	group	= new TestCaseGroup(ctx.testCtx,
5188														factory.getName().c_str(),
5189														factory.getDesc().c_str());
5190
5191	for (int precNdx = 0; precNdx < glu::PRECISION_LAST; ++precNdx)
5192	{
5193		const Precision		precision	= Precision(precNdx);
5194		const string		precName	(glu::getPrecisionName(precision));
5195		const FloatFormat&	fmt			= *de::getSizedArrayElement<glu::PRECISION_LAST>(ctx.formats, precNdx);
5196		const FloatFormat&	highpFmt	= *de::getSizedArrayElement<glu::PRECISION_LAST>(ctx.formats,
5197																						 glu::PRECISION_HIGHP);
5198
5199		for (size_t shaderNdx = 0; shaderNdx < ctx.shaderTypes.size(); ++shaderNdx)
5200		{
5201			const ShaderType	shaderType	= ctx.shaderTypes[shaderNdx];
5202			const string 		shaderName	(glu::getShaderTypeName(shaderType));
5203			const string		name		= precName + "_" + shaderName;
5204			const Context		caseCtx		(name, ctx.testCtx, ctx.renderCtx, fmt, highpFmt,
5205											 precision, shaderType, ctx.numRandoms);
5206
5207			group->addChild(factory.createCase(caseCtx).release());
5208		}
5209	}
5210
5211	return group;
5212}
5213
5214void addBuiltinPrecisionTests (TestContext&					testCtx,
5215							   RenderContext&				renderCtx,
5216							   const CaseFactories&			cases,
5217							   const vector<ShaderType>&	shaderTypes,
5218							   TestCaseGroup&				dstGroup)
5219{
5220	const FloatFormat				highp		(-126, 127, 23, true,
5221												 tcu::MAYBE,	// subnormals
5222												 tcu::YES,		// infinities
5223												 tcu::MAYBE);	// NaN
5224	// \todo [2014-04-01 lauri] Check these once Khronos bug 11840 is resolved.
5225	const FloatFormat				mediump		(-13, 13, 9, false);
5226	// A fixed-point format is just a floating point format with a fixed
5227	// exponent and support for subnormals.
5228	const FloatFormat				lowp		(0, 0, 7, false, tcu::YES);
5229	const PrecisionTestContext		ctx			(testCtx, renderCtx, highp, mediump, lowp,
5230												 shaderTypes, 16384);
5231
5232	for (size_t ndx = 0; ndx < cases.getFactories().size(); ++ndx)
5233		dstGroup.addChild(createFuncGroup(ctx, *cases.getFactories()[ndx]));
5234}
5235
5236} // BuiltinPrecisionTests
5237} // gls
5238} // deqp
5239