glsBuiltinPrecisionTests.cpp revision 1cc61b7d03cad727bbddd00cea8d78f4f6cc9047
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_FATAL("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_FATAL("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 != 0.0f ? "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, de::ArrayDeleter<deUint8>()));
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> >	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 T>
1306class Alternatives : public Func<Signature<T, T, T> >
1307{
1308public:
1309	typedef typename	Alternatives::Sig		Sig;
1310
1311protected:
1312	typedef typename	Alternatives::IRet		IRet;
1313	typedef typename	Alternatives::IArgs		IArgs;
1314
1315	virtual string		getName				(void) const			{ return "alternatives"; }
1316	virtual void		doPrintDefinition	(std::ostream&) const	{}
1317	void				doGetUsedFuncs		(FuncSet&) const		{}
1318
1319	virtual IRet		doApply				(const EvalContext&, const IArgs& args) const
1320	{
1321		return unionIVal<T>(args.a, args.b);
1322	}
1323
1324	virtual void		doPrint				(ostream& os, const BaseArgExprs& args)	const
1325	{
1326		os << "{" << *args[0] << " | " << *args[1] << "}";
1327	}
1328};
1329
1330template <typename Sig>
1331ExprP<typename Sig::Ret> createApply (const Func<Sig>&						func,
1332									  const typename Func<Sig>::ArgExprs&	args)
1333{
1334	return exprP(new Apply<Sig>(func, args));
1335}
1336
1337template <typename Sig>
1338ExprP<typename Sig::Ret> createApply (
1339	const Func<Sig>&			func,
1340	const ExprP<typename Sig::Arg0>&	arg0 = voidP(),
1341	const ExprP<typename Sig::Arg1>&	arg1 = voidP(),
1342	const ExprP<typename Sig::Arg2>&	arg2 = voidP(),
1343	const ExprP<typename Sig::Arg3>&	arg3 = voidP())
1344{
1345	return exprP(new Apply<Sig>(func, arg0, arg1, arg2, arg3));
1346}
1347
1348template <typename Sig>
1349ExprP<typename Sig::Ret> Func<Sig>::operator() (const ExprP<typename Sig::Arg0>& arg0,
1350												const ExprP<typename Sig::Arg1>& arg1,
1351												const ExprP<typename Sig::Arg2>& arg2,
1352												const ExprP<typename Sig::Arg3>& arg3) const
1353{
1354	return createApply(*this, arg0, arg1, arg2, arg3);
1355}
1356
1357template <typename F>
1358ExprP<typename F::Ret> app (const ExprP<typename F::Arg0>& arg0 = voidP(),
1359							const ExprP<typename F::Arg1>& arg1 = voidP(),
1360							const ExprP<typename F::Arg2>& arg2 = voidP(),
1361							const ExprP<typename F::Arg3>& arg3 = voidP())
1362{
1363	return createApply(instance<F>(), arg0, arg1, arg2, arg3);
1364}
1365
1366template <typename F>
1367typename F::IRet call (const EvalContext&			ctx,
1368					   const typename F::IArg0&		arg0 = Void(),
1369					   const typename F::IArg1&		arg1 = Void(),
1370					   const typename F::IArg2&		arg2 = Void(),
1371					   const typename F::IArg3&		arg3 = Void())
1372{
1373	return instance<F>().apply(ctx, arg0, arg1, arg2, arg3);
1374}
1375
1376template <typename T>
1377ExprP<T> alternatives (const ExprP<T>& arg0,
1378					   const ExprP<T>& arg1)
1379{
1380	return createApply<typename Alternatives<T>::Sig>(instance<Alternatives<T> >(), arg0, arg1);
1381}
1382
1383template <typename Sig>
1384class ApplyVar : public Apply<Sig>
1385{
1386public:
1387	typedef typename Sig::Ret				Ret;
1388	typedef typename Sig::Arg0				Arg0;
1389	typedef typename Sig::Arg1				Arg1;
1390	typedef typename Sig::Arg2				Arg2;
1391	typedef typename Sig::Arg3				Arg3;
1392	typedef typename Expr<Ret>::Val			Val;
1393	typedef typename Expr<Ret>::IVal		IVal;
1394	typedef Func<Sig>						ApplyFunc;
1395	typedef typename ApplyFunc::ArgExprs	ArgExprs;
1396
1397						ApplyVar	(const ApplyFunc&			func,
1398									 const VariableP<Arg0>&		arg0,
1399									 const VariableP<Arg1>&		arg1,
1400									 const VariableP<Arg2>&		arg2,
1401									 const VariableP<Arg3>&		arg3)
1402							: Apply<Sig> (func, arg0, arg1, arg2, arg3) {}
1403protected:
1404	IVal				doEvaluate		(const EvalContext& ctx) const
1405	{
1406		const Variable<Arg0>&	var0 = static_cast<const Variable<Arg0>&>(*this->m_args.a);
1407		const Variable<Arg1>&	var1 = static_cast<const Variable<Arg1>&>(*this->m_args.b);
1408		const Variable<Arg2>&	var2 = static_cast<const Variable<Arg2>&>(*this->m_args.c);
1409		const Variable<Arg3>&	var3 = static_cast<const Variable<Arg3>&>(*this->m_args.d);
1410		return this->m_func.apply(ctx,
1411								  ctx.env.lookup(var0), ctx.env.lookup(var1),
1412								  ctx.env.lookup(var2), ctx.env.lookup(var3));
1413	}
1414};
1415
1416template <typename Sig>
1417ExprP<typename Sig::Ret> applyVar (const Func<Sig>&						func,
1418								   const VariableP<typename Sig::Arg0>&	arg0,
1419								   const VariableP<typename Sig::Arg1>&	arg1,
1420								   const VariableP<typename Sig::Arg2>&	arg2,
1421								   const VariableP<typename Sig::Arg3>&	arg3)
1422{
1423	return exprP(new ApplyVar<Sig>(func, arg0, arg1, arg2, arg3));
1424}
1425
1426template <typename Sig_>
1427class DerivedFunc : public Func<Sig_>
1428{
1429public:
1430	typedef typename DerivedFunc::ArgExprs		ArgExprs;
1431	typedef typename DerivedFunc::IRet			IRet;
1432	typedef typename DerivedFunc::IArgs			IArgs;
1433	typedef typename DerivedFunc::Ret			Ret;
1434	typedef typename DerivedFunc::Arg0			Arg0;
1435	typedef typename DerivedFunc::Arg1			Arg1;
1436	typedef typename DerivedFunc::Arg2			Arg2;
1437	typedef typename DerivedFunc::Arg3			Arg3;
1438	typedef typename DerivedFunc::IArg0			IArg0;
1439	typedef typename DerivedFunc::IArg1			IArg1;
1440	typedef typename DerivedFunc::IArg2			IArg2;
1441	typedef typename DerivedFunc::IArg3			IArg3;
1442
1443protected:
1444	void						doPrintDefinition	(ostream& os) const
1445	{
1446		const ParamNames&	paramNames	= this->getParamNames();
1447
1448		initialize();
1449
1450		os << dataTypeNameOf<Ret>() << " " << this->getName()
1451			<< "(";
1452		if (isTypeValid<Arg0>())
1453			os << dataTypeNameOf<Arg0>() << " " << paramNames.a;
1454		if (isTypeValid<Arg1>())
1455			os << ", " << dataTypeNameOf<Arg1>() << " " << paramNames.b;
1456		if (isTypeValid<Arg2>())
1457			os << ", " << dataTypeNameOf<Arg2>() << " " << paramNames.c;
1458		if (isTypeValid<Arg3>())
1459			os << ", " << dataTypeNameOf<Arg3>() << " " << paramNames.d;
1460		os << ")\n{\n";
1461
1462		for (size_t ndx = 0; ndx < m_body.size(); ++ndx)
1463			os << *m_body[ndx];
1464		os << "return " << *m_ret << ";\n";
1465		os << "}\n";
1466	}
1467
1468	IRet						doApply			(const EvalContext&	ctx,
1469												 const IArgs&		args) const
1470	{
1471		Environment	funEnv;
1472		IArgs&		mutArgs		= const_cast<IArgs&>(args);
1473		IRet		ret;
1474
1475		initialize();
1476
1477		funEnv.bind(*m_var0, args.a);
1478		funEnv.bind(*m_var1, args.b);
1479		funEnv.bind(*m_var2, args.c);
1480		funEnv.bind(*m_var3, args.d);
1481
1482		{
1483			EvalContext	funCtx(ctx.format, ctx.floatPrecision, funEnv, ctx.callDepth);
1484
1485			for (size_t ndx = 0; ndx < m_body.size(); ++ndx)
1486				m_body[ndx]->execute(funCtx);
1487
1488			ret = m_ret->evaluate(funCtx);
1489		}
1490
1491		// \todo [lauri] Store references instead of values in environment
1492		const_cast<IArg0&>(mutArgs.a) = funEnv.lookup(*m_var0);
1493		const_cast<IArg1&>(mutArgs.b) = funEnv.lookup(*m_var1);
1494		const_cast<IArg2&>(mutArgs.c) = funEnv.lookup(*m_var2);
1495		const_cast<IArg3&>(mutArgs.d) = funEnv.lookup(*m_var3);
1496
1497		return ret;
1498	}
1499
1500	void						doGetUsedFuncs	(FuncSet& dst) const
1501	{
1502		initialize();
1503		if (dst.insert(this).second)
1504		{
1505			for (size_t ndx = 0; ndx < m_body.size(); ++ndx)
1506				m_body[ndx]->getUsedFuncs(dst);
1507			m_ret->getUsedFuncs(dst);
1508		}
1509	}
1510
1511	virtual ExprP<Ret>			doExpand		(ExpandContext& ctx, const ArgExprs& args_) const = 0;
1512
1513	// These are transparently initialized when first needed. They cannot be
1514	// initialized in the constructor because they depend on the doExpand
1515	// method of the subclass.
1516
1517	mutable VariableP<Arg0>		m_var0;
1518	mutable VariableP<Arg1>		m_var1;
1519	mutable VariableP<Arg2>		m_var2;
1520	mutable VariableP<Arg3>		m_var3;
1521	mutable vector<StatementP>	m_body;
1522	mutable ExprP<Ret>			m_ret;
1523
1524private:
1525
1526	void				initialize		(void)	const
1527	{
1528		if (!m_ret)
1529		{
1530			const ParamNames&	paramNames	= this->getParamNames();
1531			Counter				symCounter;
1532			ExpandContext		ctx			(symCounter);
1533			ArgExprs			args;
1534
1535			args.a	= m_var0 = variable<Arg0>(paramNames.a);
1536			args.b	= m_var1 = variable<Arg1>(paramNames.b);
1537			args.c	= m_var2 = variable<Arg2>(paramNames.c);
1538			args.d	= m_var3 = variable<Arg3>(paramNames.d);
1539
1540			m_ret	= this->doExpand(ctx, args);
1541			m_body	= ctx.getStatements();
1542		}
1543	}
1544};
1545
1546template <typename Sig>
1547class PrimitiveFunc : public Func<Sig>
1548{
1549public:
1550	typedef typename PrimitiveFunc::Ret			Ret;
1551	typedef typename PrimitiveFunc::ArgExprs	ArgExprs;
1552
1553protected:
1554	void	doPrintDefinition	(ostream&) const	{}
1555	void	doGetUsedFuncs		(FuncSet&) const	{}
1556};
1557
1558template <typename T>
1559class Cond : public PrimitiveFunc<Signature<T, bool, T, T> >
1560{
1561public:
1562	typedef typename Cond::IArgs	IArgs;
1563	typedef typename Cond::IRet		IRet;
1564
1565	string	getName	(void) const
1566	{
1567		return "_cond";
1568	}
1569
1570protected:
1571
1572	void	doPrint	(ostream& os, const BaseArgExprs& args) const
1573	{
1574		os << "(" << *args[0] << " ? " << *args[1] << " : " << *args[2] << ")";
1575	}
1576
1577	IRet	doApply	(const EvalContext&, const IArgs& iargs)const
1578	{
1579		IRet	ret;
1580
1581		if (iargs.a.contains(true))
1582			ret = unionIVal<T>(ret, iargs.b);
1583
1584		if (iargs.a.contains(false))
1585			ret = unionIVal<T>(ret, iargs.c);
1586
1587		return ret;
1588	}
1589};
1590
1591template <typename T>
1592class CompareOperator : public PrimitiveFunc<Signature<bool, T, T> >
1593{
1594public:
1595	typedef typename CompareOperator::IArgs	IArgs;
1596	typedef typename CompareOperator::IArg0	IArg0;
1597	typedef typename CompareOperator::IArg1	IArg1;
1598	typedef typename CompareOperator::IRet	IRet;
1599
1600protected:
1601	void			doPrint	(ostream& os, const BaseArgExprs& args) const
1602	{
1603		os << "(" << *args[0] << getSymbol() << *args[1] << ")";
1604	}
1605
1606	Interval		doApply	(const EvalContext&, const IArgs& iargs) const
1607	{
1608		const IArg0&	arg0 = iargs.a;
1609		const IArg1&	arg1 = iargs.b;
1610		IRet	ret;
1611
1612		if (canSucceed(arg0, arg1))
1613			ret |= true;
1614		if (canFail(arg0, arg1))
1615			ret |= false;
1616
1617		return ret;
1618	}
1619
1620	virtual string	getSymbol	(void) const = 0;
1621	virtual bool	canSucceed	(const IArg0&, const IArg1&) const = 0;
1622	virtual bool	canFail		(const IArg0&, const IArg1&) const = 0;
1623};
1624
1625template <typename T>
1626class LessThan : public CompareOperator<T>
1627{
1628public:
1629	string	getName		(void) const									{ return "lessThan"; }
1630
1631protected:
1632	string	getSymbol	(void) const									{ return "<";		}
1633
1634	bool	canSucceed	(const Interval& a, const Interval& b) const
1635	{
1636		return (a.lo() < b.hi());
1637	}
1638
1639	bool	canFail		(const Interval& a, const Interval& b) const
1640	{
1641		return !(a.hi() < b.lo());
1642	}
1643};
1644
1645template <typename T>
1646ExprP<bool> operator< (const ExprP<T>& a, const ExprP<T>& b)
1647{
1648	return app<LessThan<T> >(a, b);
1649}
1650
1651template <typename T>
1652ExprP<T> cond (const ExprP<bool>&	test,
1653			   const ExprP<T>&		consequent,
1654			   const ExprP<T>&		alternative)
1655{
1656	return app<Cond<T> >(test, consequent, alternative);
1657}
1658
1659/*--------------------------------------------------------------------*//*!
1660 *
1661 * @}
1662 *
1663 *//*--------------------------------------------------------------------*/
1664
1665class FloatFunc1 : public PrimitiveFunc<Signature<float, float> >
1666{
1667protected:
1668	Interval			doApply			(const EvalContext& ctx, const IArgs& iargs) const
1669	{
1670		return this->applyMonotone(ctx, iargs.a);
1671	}
1672
1673	Interval			applyMonotone	(const EvalContext& ctx, const Interval& iarg0) const
1674	{
1675		Interval ret;
1676
1677		TCU_INTERVAL_APPLY_MONOTONE1(ret, arg0, iarg0, val,
1678									 TCU_SET_INTERVAL(val, point,
1679													  point = this->applyPoint(ctx, arg0)));
1680
1681		ret |= innerExtrema(ctx, iarg0);
1682		ret &= (this->getCodomain() | TCU_NAN);
1683
1684		return ctx.format.convert(ret);
1685	}
1686
1687	virtual Interval	innerExtrema	(const EvalContext&, const Interval&) const
1688	{
1689		return Interval(); // empty interval, i.e. no extrema
1690	}
1691
1692	virtual Interval	applyPoint		(const EvalContext& ctx, double arg0) const
1693	{
1694		const double	exact	= this->applyExact(arg0);
1695		const double	prec	= this->precision(ctx, exact, arg0);
1696
1697		return exact + Interval(-prec, prec);
1698	}
1699
1700	virtual double		applyExact		(double) const
1701	{
1702		TCU_THROW(InternalError, "Cannot apply");
1703	}
1704
1705	virtual Interval	getCodomain		(void) const
1706	{
1707		return Interval::unbounded(true);
1708	}
1709
1710	virtual double		precision		(const EvalContext& ctx, double, double) const = 0;
1711};
1712
1713class CFloatFunc1 : public FloatFunc1
1714{
1715public:
1716			CFloatFunc1	(const string& name, DoubleFunc1& func)
1717				: m_name(name), m_func(func) {}
1718
1719	string			getName		(void) const		{ return m_name; }
1720
1721protected:
1722	double			applyExact	(double x) const	{ return m_func(x); }
1723
1724	const string	m_name;
1725	DoubleFunc1&	m_func;
1726};
1727
1728class FloatFunc2 : public PrimitiveFunc<Signature<float, float, float> >
1729{
1730protected:
1731	Interval			doApply			(const EvalContext&	ctx, const IArgs& iargs) const
1732	{
1733		return this->applyMonotone(ctx, iargs.a, iargs.b);
1734	}
1735
1736	Interval			applyMonotone	(const EvalContext&	ctx,
1737										 const Interval&	xi,
1738										 const Interval&	yi) const
1739	{
1740		Interval reti;
1741
1742		TCU_INTERVAL_APPLY_MONOTONE2(reti, x, xi, y, yi, ret,
1743									 TCU_SET_INTERVAL(ret, point,
1744													  point = this->applyPoint(ctx, x, y)));
1745		reti |= innerExtrema(ctx, xi, yi);
1746		reti &= (this->getCodomain() | TCU_NAN);
1747
1748		return ctx.format.convert(reti);
1749	}
1750
1751	virtual Interval	innerExtrema	(const EvalContext&,
1752										 const Interval&,
1753										 const Interval&) const
1754	{
1755		return Interval(); // empty interval, i.e. no extrema
1756	}
1757
1758	virtual Interval	applyPoint		(const EvalContext&	ctx,
1759										 double 			x,
1760										 double 			y) const
1761	{
1762		const double exact	= this->applyExact(x, y);
1763		const double prec	= this->precision(ctx, exact, x, y);
1764
1765		return exact + Interval(-prec, prec);
1766	}
1767
1768	virtual double		applyExact		(double, double) const
1769	{
1770		TCU_THROW(InternalError, "Cannot apply");
1771	}
1772
1773	virtual Interval	getCodomain		(void) const
1774	{
1775		return Interval::unbounded(true);
1776	}
1777
1778	virtual double		precision		(const EvalContext&	ctx,
1779										 double				ret,
1780										 double				x,
1781										 double				y) const = 0;
1782};
1783
1784class CFloatFunc2 : public FloatFunc2
1785{
1786public:
1787					CFloatFunc2	(const string&	name,
1788								 DoubleFunc2&	func)
1789						: m_name(name)
1790						, m_func(func)
1791	{
1792	}
1793
1794	string			getName		(void) const						{ return m_name; }
1795
1796protected:
1797	double			applyExact	(double x, double y) const			{ return m_func(x, y); }
1798
1799	const string	m_name;
1800	DoubleFunc2&	m_func;
1801};
1802
1803class InfixOperator : public FloatFunc2
1804{
1805protected:
1806	virtual string	getSymbol		(void) const = 0;
1807
1808	void			doPrint			(ostream& os, const BaseArgExprs& args) const
1809	{
1810		os << "(" << *args[0] << " " << getSymbol() << " " << *args[1] << ")";
1811	}
1812
1813	Interval		applyPoint		(const EvalContext&	ctx,
1814									 double 			x,
1815									 double 			y) const
1816	{
1817		const double exact	= this->applyExact(x, y);
1818
1819		// Allow either representable number on both sides of the exact value,
1820		// but require exactly representable values to be preserved.
1821		return ctx.format.roundOut(exact, !deIsInf(x) && !deIsInf(y));
1822	}
1823
1824	double			precision		(const EvalContext&, double, double, double) const
1825	{
1826		return 0.0;
1827	}
1828};
1829
1830class FloatFunc3 : public PrimitiveFunc<Signature<float, float, float, float> >
1831{
1832protected:
1833	Interval			doApply			(const EvalContext&	ctx, const IArgs& iargs) const
1834	{
1835		return this->applyMonotone(ctx, iargs.a, iargs.b, iargs.c);
1836	}
1837
1838	Interval			applyMonotone	(const EvalContext&	ctx,
1839										 const Interval&	xi,
1840										 const Interval&	yi,
1841										 const Interval&	zi) const
1842	{
1843		Interval reti;
1844		TCU_INTERVAL_APPLY_MONOTONE3(reti, x, xi, y, yi, z, zi, ret,
1845									 TCU_SET_INTERVAL(ret, point,
1846													  point = this->applyPoint(ctx, x, y, z)));
1847		return ctx.format.convert(reti);
1848	}
1849
1850	virtual Interval	applyPoint		(const EvalContext&	ctx,
1851										 double 			x,
1852										 double 			y,
1853										 double 			z) const
1854	{
1855		const double exact	= this->applyExact(x, y, z);
1856		const double prec	= this->precision(ctx, exact, x, y, z);
1857		return exact + Interval(-prec, prec);
1858	}
1859
1860	virtual double		applyExact		(double, double, double) const
1861	{
1862		TCU_THROW(InternalError, "Cannot apply");
1863	}
1864
1865	virtual double		precision		(const EvalContext&	ctx,
1866										 double				result,
1867										 double				x,
1868										 double				y,
1869										 double				z) const = 0;
1870};
1871
1872// We define syntactic sugar functions for expression constructors. Since
1873// these have the same names as ordinary mathematical operations (sin, log
1874// etc.), it's better to give them a dedicated namespace.
1875namespace Functions
1876{
1877
1878using namespace tcu;
1879
1880class Add : public InfixOperator
1881{
1882public:
1883	string		getName		(void) const 						{ return "add"; }
1884	string		getSymbol	(void) const 						{ return "+"; }
1885
1886	Interval	doApply		(const EvalContext&	ctx,
1887							 const IArgs&		iargs) const
1888	{
1889		// Fast-path for common case
1890		if (iargs.a.isOrdinary() && iargs.b.isOrdinary())
1891		{
1892			Interval ret;
1893			TCU_SET_INTERVAL_BOUNDS(ret, sum,
1894									sum = iargs.a.lo() + iargs.b.lo(),
1895									sum = iargs.a.hi() + iargs.b.hi());
1896			return ctx.format.convert(ctx.format.roundOut(ret, true));
1897		}
1898		return this->applyMonotone(ctx, iargs.a, iargs.b);
1899	}
1900
1901protected:
1902	double		applyExact	(double x, double y) const 			{ return x + y; }
1903};
1904
1905class Mul : public InfixOperator
1906{
1907public:
1908	string		getName		(void) const 									{ return "mul"; }
1909	string		getSymbol	(void) const 									{ return "*"; }
1910
1911	Interval	doApply		(const EvalContext&	ctx, const IArgs& iargs) const
1912	{
1913		Interval a = iargs.a;
1914		Interval b = iargs.b;
1915
1916		// Fast-path for common case
1917		if (a.isOrdinary() && b.isOrdinary())
1918		{
1919			Interval ret;
1920			if (a.hi() < 0)
1921			{
1922				a = -a;
1923				b = -b;
1924			}
1925			if (a.lo() >= 0 && b.lo() >= 0)
1926			{
1927				TCU_SET_INTERVAL_BOUNDS(ret, prod,
1928										prod = iargs.a.lo() * iargs.b.lo(),
1929										prod = iargs.a.hi() * iargs.b.hi());
1930				return ctx.format.convert(ctx.format.roundOut(ret, true));
1931			}
1932			if (a.lo() >= 0 && b.hi() <= 0)
1933			{
1934				TCU_SET_INTERVAL_BOUNDS(ret, prod,
1935										prod = iargs.a.hi() * iargs.b.lo(),
1936										prod = iargs.a.lo() * iargs.b.hi());
1937				return ctx.format.convert(ctx.format.roundOut(ret, true));
1938			}
1939		}
1940		return this->applyMonotone(ctx, iargs.a, iargs.b);
1941	}
1942
1943protected:
1944	double		applyExact	(double x, double y) const						{ return x * y; }
1945
1946	Interval	innerExtrema(const EvalContext&, const Interval& xi, const Interval& yi) const
1947	{
1948		if (((xi.contains(-TCU_INFINITY) || xi.contains(TCU_INFINITY)) && yi.contains(0.0)) ||
1949			((yi.contains(-TCU_INFINITY) || yi.contains(TCU_INFINITY)) && xi.contains(0.0)))
1950			return Interval(TCU_NAN);
1951
1952		return Interval();
1953	}
1954};
1955
1956class Sub : public InfixOperator
1957{
1958public:
1959	string		getName		(void) const 				{ return "sub"; }
1960	string		getSymbol	(void) const 				{ return "-"; }
1961
1962	Interval	doApply		(const EvalContext&	ctx, const IArgs& iargs) const
1963	{
1964		// Fast-path for common case
1965		if (iargs.a.isOrdinary() && iargs.b.isOrdinary())
1966		{
1967			Interval ret;
1968
1969			TCU_SET_INTERVAL_BOUNDS(ret, diff,
1970									diff = iargs.a.lo() - iargs.b.hi(),
1971									diff = iargs.a.hi() - iargs.b.lo());
1972			return ctx.format.convert(ctx.format.roundOut(ret, true));
1973
1974		}
1975		else
1976		{
1977			return this->applyMonotone(ctx, iargs.a, iargs.b);
1978		}
1979	}
1980
1981protected:
1982	double		applyExact	(double x, double y) const	{ return x - y; }
1983};
1984
1985class Negate : public FloatFunc1
1986{
1987public:
1988	string	getName		(void) const									{ return "_negate"; }
1989	void	doPrint		(ostream& os, const BaseArgExprs& args) const	{ os << "-" << *args[0]; }
1990
1991protected:
1992	double	precision	(const EvalContext&, double, double) const		{ return 0.0; }
1993	double	applyExact	(double x) const								{ return -x; }
1994};
1995
1996class Div : public InfixOperator
1997{
1998public:
1999	string		getName			(void) const 						{ return "div"; }
2000
2001protected:
2002	string		getSymbol		(void) const 						{ return "/"; }
2003
2004	Interval	innerExtrema	(const EvalContext&,
2005								 const Interval&		nom,
2006								 const Interval&		den) const
2007	{
2008		Interval ret;
2009
2010		if (den.contains(0.0))
2011		{
2012			if (nom.contains(0.0))
2013				ret |= TCU_NAN;
2014
2015			if (nom.lo() < 0.0 || nom.hi() > 0.0)
2016				ret |= Interval::unbounded();
2017		}
2018
2019		return ret;
2020	}
2021
2022	double		applyExact		(double x, double y) const { return x / y; }
2023
2024	Interval	applyPoint		(const EvalContext&	ctx, double x, double y) const
2025	{
2026		Interval ret = FloatFunc2::applyPoint(ctx, x, y);
2027
2028		if (!deIsInf(x) && !deIsInf(y) && y != 0.0)
2029		{
2030			const Interval dst = ctx.format.convert(ret);
2031			if (dst.contains(-TCU_INFINITY)) ret |= -ctx.format.getMaxValue();
2032			if (dst.contains(+TCU_INFINITY)) ret |= +ctx.format.getMaxValue();
2033		}
2034
2035		return ret;
2036	}
2037
2038	double		precision		(const EvalContext& ctx, double ret, double, double den) const
2039	{
2040		const FloatFormat&	fmt		= ctx.format;
2041
2042		// \todo [2014-03-05 lauri] Check that the limits in GLSL 3.10 are actually correct.
2043		// For now, we assume that division's precision is 2.5 ULP when the value is within
2044		// [2^MINEXP, 2^MAXEXP-1]
2045
2046		if (den == 0.0)
2047			return 0.0; // Result must be exactly inf
2048		else if (de::inBounds(deAbs(den),
2049							  deLdExp(1.0, fmt.getMinExp()),
2050							  deLdExp(1.0, fmt.getMaxExp() - 1)))
2051			return fmt.ulp(ret, 2.5);
2052		else
2053			return TCU_INFINITY; // Can be any number, but must be a number.
2054	}
2055};
2056
2057class InverseSqrt : public FloatFunc1
2058{
2059public:
2060	string		getName		(void) const							{ return "inversesqrt"; }
2061
2062protected:
2063	double		applyExact	(double x) const						{ return 1.0 / deSqrt(x); }
2064
2065	double		precision	(const EvalContext& ctx, double ret, double x) const
2066	{
2067		return x <= 0 ? TCU_NAN : ctx.format.ulp(ret, 2.0);
2068	}
2069
2070	Interval	getCodomain	(void) const
2071	{
2072		return Interval(0.0, TCU_INFINITY);
2073	}
2074};
2075
2076class ExpFunc : public CFloatFunc1
2077{
2078public:
2079				ExpFunc		(const string& name, DoubleFunc1& func)
2080					: CFloatFunc1(name, func) {}
2081protected:
2082	double		precision	(const EvalContext& ctx, double ret, double x) const
2083	{
2084		switch (ctx.floatPrecision)
2085		{
2086			case glu::PRECISION_HIGHP:
2087				return ctx.format.ulp(ret, 3.0 + 2.0 * deAbs(x));
2088			case glu::PRECISION_MEDIUMP:
2089				return ctx.format.ulp(ret, 2.0 + 2.0 * deAbs(x));
2090			case glu::PRECISION_LOWP:
2091				return ctx.format.ulp(ret, 2.0);
2092			default:
2093				DE_FATAL("Impossible");
2094		}
2095		return 0;
2096	}
2097
2098	Interval	getCodomain	(void) const
2099	{
2100		return Interval(0.0, TCU_INFINITY);
2101	}
2102};
2103
2104class Exp2	: public ExpFunc	{ public: Exp2 (void)	: ExpFunc("exp2", deExp2) {} };
2105class Exp	: public ExpFunc	{ public: Exp (void)	: ExpFunc("exp", deExp) {} };
2106
2107ExprP<float> exp2	(const ExprP<float>& x)	{ return app<Exp2>(x); }
2108ExprP<float> exp	(const ExprP<float>& x)	{ return app<Exp>(x); }
2109
2110class LogFunc : public CFloatFunc1
2111{
2112public:
2113				LogFunc		(const string& name, DoubleFunc1& func)
2114					: CFloatFunc1(name, func) {}
2115
2116protected:
2117	double		precision	(const EvalContext& ctx, double ret, double x) const
2118	{
2119		if (x <= 0)
2120			return TCU_NAN;
2121
2122		switch (ctx.floatPrecision)
2123		{
2124			case glu::PRECISION_HIGHP:
2125				return (0.5 <= x && x <= 2.0) ? deLdExp(1.0, -21) : ctx.format.ulp(ret, 3.0);
2126			case glu::PRECISION_MEDIUMP:
2127				return (0.5 <= x && x <= 2.0) ? deLdExp(1.0, -7) : ctx.format.ulp(ret, 2.0);
2128			case glu::PRECISION_LOWP:
2129				return ctx.format.ulp(ret, 2.0);
2130			default:
2131				DE_FATAL("Impossible");
2132		}
2133
2134		return 0;
2135	}
2136};
2137
2138class Log2	: public LogFunc		{ public: Log2	(void) : LogFunc("log2", deLog2) {} };
2139class Log	: public LogFunc		{ public: Log	(void) : LogFunc("log", deLog) {} };
2140
2141ExprP<float> log2	(const ExprP<float>& x)	{ return app<Log2>(x); }
2142ExprP<float> log	(const ExprP<float>& x)	{ return app<Log>(x); }
2143
2144#define DEFINE_CONSTRUCTOR1(CLASS, TRET, NAME, T0) \
2145ExprP<TRET> NAME (const ExprP<T0>& arg0) { return app<CLASS>(arg0); }
2146
2147#define DEFINE_DERIVED1(CLASS, TRET, NAME, T0, ARG0, EXPANSION)			\
2148class CLASS : public DerivedFunc<Signature<TRET, T0> >		 			\
2149{																		\
2150public:																	\
2151	string			getName		(void) const		{ return #NAME; }	\
2152																		\
2153protected:																\
2154	ExprP<TRET>		doExpand		(ExpandContext&,					\
2155									 const CLASS::ArgExprs& args_) const \
2156	{																	\
2157		const ExprP<float>& ARG0 = args_.a;								\
2158		return EXPANSION;												\
2159	}																	\
2160};																		\
2161DEFINE_CONSTRUCTOR1(CLASS, TRET, NAME, T0)
2162
2163#define DEFINE_DERIVED_FLOAT1(CLASS, NAME, ARG0, EXPANSION) \
2164	DEFINE_DERIVED1(CLASS, float, NAME, float, ARG0, EXPANSION)
2165
2166#define DEFINE_CONSTRUCTOR2(CLASS, TRET, NAME, T0, T1)				\
2167ExprP<TRET> NAME (const ExprP<T0>& arg0, const ExprP<T1>& arg1)		\
2168{																	\
2169	return app<CLASS>(arg0, arg1);									\
2170}
2171
2172#define DEFINE_DERIVED2(CLASS, TRET, NAME, T0, Arg0, T1, Arg1, EXPANSION) \
2173class CLASS : public DerivedFunc<Signature<TRET, T0, T1> >		 		\
2174{																		\
2175public:																	\
2176	string			getName		(void) const		{ return #NAME; }	\
2177																		\
2178protected:																\
2179	ExprP<TRET>		doExpand	(ExpandContext&, const ArgExprs& args_) const \
2180	{																	\
2181		const ExprP<T0>& Arg0 = args_.a;								\
2182		const ExprP<T1>& Arg1 = args_.b;								\
2183		return EXPANSION;												\
2184	}																	\
2185};																		\
2186DEFINE_CONSTRUCTOR2(CLASS, TRET, NAME, T0, T1)
2187
2188#define DEFINE_DERIVED_FLOAT2(CLASS, NAME, Arg0, Arg1, EXPANSION)		\
2189	DEFINE_DERIVED2(CLASS, float, NAME, float, Arg0, float, Arg1, EXPANSION)
2190
2191#define DEFINE_CONSTRUCTOR3(CLASS, TRET, NAME, T0, T1, T2)				\
2192ExprP<TRET> NAME (const ExprP<T0>& arg0, const ExprP<T1>& arg1, const ExprP<T2>& arg2) \
2193{																		\
2194	return app<CLASS>(arg0, arg1, arg2);								\
2195}
2196
2197#define DEFINE_DERIVED3(CLASS, TRET, NAME, T0, ARG0, T1, ARG1, T2, ARG2, EXPANSION) \
2198class CLASS : public DerivedFunc<Signature<TRET, T0, T1, T2> >					\
2199{																				\
2200public:																			\
2201	string			getName		(void) const	{ return #NAME; }				\
2202																				\
2203protected:																		\
2204	ExprP<TRET>		doExpand	(ExpandContext&, const ArgExprs& args_) const	\
2205	{																			\
2206		const ExprP<T0>& ARG0 = args_.a;										\
2207		const ExprP<T1>& ARG1 = args_.b;										\
2208		const ExprP<T2>& ARG2 = args_.c;										\
2209		return EXPANSION;														\
2210	}																			\
2211};																				\
2212DEFINE_CONSTRUCTOR3(CLASS, TRET, NAME, T0, T1, T2)
2213
2214#define DEFINE_DERIVED_FLOAT3(CLASS, NAME, ARG0, ARG1, ARG2, EXPANSION)			\
2215	DEFINE_DERIVED3(CLASS, float, NAME, float, ARG0, float, ARG1, float, ARG2, EXPANSION)
2216
2217#define DEFINE_CONSTRUCTOR4(CLASS, TRET, NAME, T0, T1, T2, T3)			\
2218ExprP<TRET> NAME (const ExprP<T0>& arg0, const ExprP<T1>& arg1,			\
2219				  const ExprP<T2>& arg2, const ExprP<T3>& arg3)			\
2220{																		\
2221	return app<CLASS>(arg0, arg1, arg2, arg3);							\
2222}
2223
2224DEFINE_DERIVED_FLOAT1(Sqrt,		sqrt,		x, 		constant(1.0f) / app<InverseSqrt>(x));
2225DEFINE_DERIVED_FLOAT2(Pow,		pow,		x,	y,	exp2(y * log2(x)));
2226DEFINE_DERIVED_FLOAT1(Radians,	radians,	d, 		(constant(DE_PI) / constant(180.0f)) * d);
2227DEFINE_DERIVED_FLOAT1(Degrees,	degrees,	r,	 	(constant(180.0f) / constant(DE_PI)) * r);
2228
2229class TrigFunc : public CFloatFunc1
2230{
2231public:
2232					TrigFunc		(const string&		name,
2233									 DoubleFunc1&		func,
2234									 const Interval&	loEx,
2235									 const Interval&	hiEx)
2236						: CFloatFunc1	(name, func)
2237						, m_loExtremum	(loEx)
2238						, m_hiExtremum	(hiEx) {}
2239
2240protected:
2241	Interval		innerExtrema	(const EvalContext&, const Interval& angle) const
2242	{
2243		const double		lo		= angle.lo();
2244		const double		hi		= angle.hi();
2245		const int			loSlope	= doGetSlope(lo);
2246		const int			hiSlope	= doGetSlope(hi);
2247
2248		// Detect the high and low values the function can take between the
2249		// interval endpoints.
2250		if (angle.length() >= 2.0 * DE_PI_DOUBLE)
2251		{
2252			// The interval is longer than a full cycle, so it must get all possible values.
2253			return m_hiExtremum | m_loExtremum;
2254		}
2255		else if (loSlope == 1 && hiSlope == -1)
2256		{
2257			// The slope can change from positive to negative only at the maximum value.
2258			return m_hiExtremum;
2259		}
2260		else if (loSlope == -1 && hiSlope == 1)
2261		{
2262			// The slope can change from negative to positive only at the maximum value.
2263			return m_loExtremum;
2264		}
2265		else if (loSlope == hiSlope &&
2266				 deIntSign(applyExact(hi) - applyExact(lo)) * loSlope == -1)
2267		{
2268			// The slope has changed twice between the endpoints, so both extrema are included.
2269			return m_hiExtremum | m_loExtremum;
2270		}
2271
2272		return Interval();
2273	}
2274
2275	Interval	getCodomain			(void) const
2276	{
2277		// Ensure that result is always within [-1, 1], or NaN (for +-inf)
2278		return Interval(-1.0, 1.0) | TCU_NAN;
2279	}
2280
2281	double		precision			(const EvalContext& ctx, double ret, double arg) const
2282	{
2283		if (ctx.floatPrecision == glu::PRECISION_HIGHP)
2284		{
2285			// Use precision from OpenCL fast relaxed math
2286			if (-DE_PI_DOUBLE <= arg && arg <= DE_PI_DOUBLE)
2287			{
2288				return deLdExp(1.0, -11);
2289			}
2290			else
2291			{
2292				// "larger otherwise", let's pick |x| * 2^-12 , which is slightly over
2293				// 2^-11 at x == pi.
2294				return deLdExp(deAbs(arg), -12);
2295			}
2296		}
2297		else if (ctx.floatPrecision == glu::PRECISION_MEDIUMP)
2298		{
2299			if (-DE_PI_DOUBLE <= arg && arg <= DE_PI_DOUBLE)
2300			{
2301				// from OpenCL half-float extension specification
2302				return ctx.format.ulp(ret, 2.0);
2303			}
2304			else
2305			{
2306				// |x| * 2^-10, slightly larger than 2 ULP at x == pi
2307				return deLdExp(deAbs(DE_PI_DOUBLE), -10);
2308			}
2309		}
2310		else
2311		{
2312			DE_ASSERT(ctx.floatPrecision == glu::PRECISION_LOWP);
2313
2314			// from OpenCL half-float extension specification
2315			return ctx.format.ulp(ret, 2.0);
2316		}
2317	}
2318
2319	virtual int		doGetSlope		(double angle) const = 0;
2320
2321	Interval		m_loExtremum;
2322	Interval		m_hiExtremum;
2323};
2324
2325class Sin : public TrigFunc
2326{
2327public:
2328				Sin			(void) : TrigFunc("sin", deSin, -1.0, 1.0) {}
2329
2330protected:
2331	int			doGetSlope	(double angle) const { return deIntSign(deCos(angle)); }
2332};
2333
2334ExprP<float> sin (const ExprP<float>& x) { return app<Sin>(x); }
2335
2336class Cos : public TrigFunc
2337{
2338public:
2339				Cos			(void) : TrigFunc("cos", deCos, -1.0, 1.0) {}
2340
2341protected:
2342	int			doGetSlope	(double angle) const { return -deIntSign(deSin(angle)); }
2343};
2344
2345ExprP<float> cos (const ExprP<float>& x) { return app<Cos>(x); }
2346
2347DEFINE_DERIVED_FLOAT1(Tan, tan, x, sin(x) * (constant(1.0f) / cos(x)));
2348
2349class ArcTrigFunc : public CFloatFunc1
2350{
2351public:
2352					ArcTrigFunc	(const string&		name,
2353								 DoubleFunc1&		func,
2354								 double				precisionULPs,
2355								 const Interval&	domain,
2356								 const Interval&	codomain)
2357						: CFloatFunc1		(name, func)
2358						, m_precision		(precisionULPs)
2359						, m_domain			(domain)
2360						, m_codomain		(codomain) {}
2361
2362protected:
2363	double			precision	(const EvalContext& ctx, double ret, double x) const
2364	{
2365		if (!m_domain.contains(x))
2366			return TCU_NAN;
2367
2368		if (ctx.floatPrecision == glu::PRECISION_HIGHP)
2369		{
2370			// Use OpenCL's fast relaxed math precision
2371			return ctx.format.ulp(ret, m_precision);
2372		}
2373		else
2374		{
2375			// Use OpenCL half-float spec
2376			return ctx.format.ulp(ret, 2.0);
2377		}
2378	}
2379
2380	// We could implement getCodomain with m_codomain, but choose not to,
2381	// because it seems too strict with trascendental constants like pi.
2382
2383	const double	m_precision;
2384	const Interval	m_domain;
2385	const Interval	m_codomain;
2386};
2387
2388class ASin : public ArcTrigFunc
2389{
2390public:
2391	ASin (void) : ArcTrigFunc("asin", deAsin, 4096.0,
2392							  Interval(-1.0, 1.0),
2393							  Interval(-DE_PI_DOUBLE * 0.5, DE_PI_DOUBLE * 0.5)) {}
2394};
2395
2396class ACos : public ArcTrigFunc
2397{
2398public:
2399	ACos (void) : ArcTrigFunc("acos", deAcos, 4096.0,
2400							  Interval(-1.0, 1.0),
2401							  Interval(0.0, DE_PI_DOUBLE)) {}
2402};
2403
2404class ATan : public ArcTrigFunc
2405{
2406public:
2407	ATan (void) : ArcTrigFunc("atan", deAtanOver, 4096.0,
2408							  Interval::unbounded(),
2409							  Interval(-DE_PI_DOUBLE * 0.5, DE_PI_DOUBLE * 0.5)) {}
2410};
2411
2412class ATan2 : public CFloatFunc2
2413{
2414public:
2415				ATan2			(void) : CFloatFunc2 ("atan", deAtan2) {}
2416
2417protected:
2418	Interval	innerExtrema	(const EvalContext&		ctx,
2419								 const Interval&		yi,
2420								 const Interval& 		xi) const
2421	{
2422		Interval ret;
2423
2424		if (yi.contains(0.0))
2425		{
2426			if (xi.contains(0.0))
2427				ret |= TCU_NAN;
2428			if (xi.intersects(Interval(-TCU_INFINITY, 0.0)))
2429				ret |= Interval(-DE_PI_DOUBLE, DE_PI_DOUBLE);
2430		}
2431
2432		if (ctx.format.hasInf() != YES && (!yi.isFinite() || !xi.isFinite()))
2433		{
2434			// Infinities may not be supported, allow anything, including NaN
2435			ret |= TCU_NAN;
2436		}
2437
2438		return ret;
2439	}
2440
2441	double		precision		(const EvalContext& ctx, double ret, double, double) const
2442	{
2443		if (ctx.floatPrecision == glu::PRECISION_HIGHP)
2444			return ctx.format.ulp(ret, 4096.0);
2445		else
2446			return ctx.format.ulp(ret, 2.0);
2447	}
2448
2449	// Codomain could be [-pi, pi], but that would probably be too strict.
2450};
2451
2452DEFINE_DERIVED_FLOAT1(Sinh, sinh, x, (exp(x) - exp(-x)) / constant(2.0f));
2453DEFINE_DERIVED_FLOAT1(Cosh, cosh, x, (exp(x) + exp(-x)) / constant(2.0f));
2454DEFINE_DERIVED_FLOAT1(Tanh, tanh, x, sinh(x) / cosh(x));
2455
2456// These are not defined as derived forms in the GLSL ES spec, but
2457// that gives us a reasonable precision.
2458DEFINE_DERIVED_FLOAT1(ASinh, asinh, x, log(x + sqrt(x * x + constant(1.0f))));
2459DEFINE_DERIVED_FLOAT1(ACosh, acosh, x, log(x + sqrt((x + constant(1.0f)) *
2460													(x - constant(1.0f)))));
2461DEFINE_DERIVED_FLOAT1(ATanh, atanh, x, constant(0.5f) * log((constant(1.0f) + x) /
2462															(constant(1.0f) - x)));
2463
2464template <typename T>
2465class GetComponent : public PrimitiveFunc<Signature<typename T::Element, T, int> >
2466{
2467public:
2468	typedef		typename GetComponent::IRet	IRet;
2469
2470	string		getName		(void) const { return "_getComponent"; }
2471
2472	void		print		(ostream&				os,
2473							 const BaseArgExprs&	args) const
2474	{
2475		os << *args[0] << "[" << *args[1] << "]";
2476	}
2477
2478protected:
2479	IRet		doApply 	(const EvalContext&,
2480							 const typename GetComponent::IArgs& iargs) const
2481	{
2482		IRet ret;
2483
2484		for (int compNdx = 0; compNdx < T::SIZE; ++compNdx)
2485		{
2486			if (iargs.b.contains(compNdx))
2487				ret = unionIVal<typename T::Element>(ret, iargs.a[compNdx]);
2488		}
2489
2490		return ret;
2491	}
2492
2493};
2494
2495template <typename T>
2496ExprP<typename T::Element> getComponent (const ExprP<T>& container, int ndx)
2497{
2498	DE_ASSERT(0 <= ndx && ndx < T::SIZE);
2499	return app<GetComponent<T> >(container, constant(ndx));
2500}
2501
2502template <typename T>	string	vecNamePrefix			(void);
2503template <>				string	vecNamePrefix<float>	(void) { return ""; }
2504template <>				string	vecNamePrefix<int>		(void) { return "i"; }
2505template <>				string	vecNamePrefix<bool>		(void) { return "b"; }
2506
2507template <typename T, int Size>
2508string vecName (void) { return vecNamePrefix<T>() + "vec" + de::toString(Size); }
2509
2510template <typename T, int Size> class GenVec;
2511
2512template <typename T>
2513class GenVec<T, 1> : public DerivedFunc<Signature<T, T> >
2514{
2515public:
2516	typedef typename GenVec<T, 1>::ArgExprs ArgExprs;
2517
2518	string		getName		(void) const
2519	{
2520		return "_" + vecName<T, 1>();
2521	}
2522
2523protected:
2524
2525	ExprP<T>	doExpand	(ExpandContext&, const ArgExprs& args) const { return args.a; }
2526};
2527
2528template <typename T>
2529class GenVec<T, 2> : public PrimitiveFunc<Signature<Vector<T, 2>, T, T> >
2530{
2531public:
2532	typedef typename GenVec::IRet	IRet;
2533	typedef typename GenVec::IArgs	IArgs;
2534
2535	string		getName		(void) const
2536	{
2537		return vecName<T, 2>();
2538	}
2539
2540protected:
2541	IRet		doApply		(const EvalContext&, const IArgs& iargs) const
2542	{
2543		return IRet(iargs.a, iargs.b);
2544	}
2545};
2546
2547template <typename T>
2548class GenVec<T, 3> : public PrimitiveFunc<Signature<Vector<T, 3>, T, T, T> >
2549{
2550public:
2551	typedef typename GenVec::IRet	IRet;
2552	typedef typename GenVec::IArgs	IArgs;
2553
2554	string	getName		(void) const
2555	{
2556		return vecName<T, 3>();
2557	}
2558
2559protected:
2560	IRet	doApply		(const EvalContext&, const IArgs& iargs) const
2561	{
2562		return IRet(iargs.a, iargs.b, iargs.c);
2563	}
2564};
2565
2566template <typename T>
2567class GenVec<T, 4> : public PrimitiveFunc<Signature<Vector<T, 4>, T, T, T, T> >
2568{
2569public:
2570	typedef typename GenVec::IRet	IRet;
2571	typedef typename GenVec::IArgs	IArgs;
2572
2573	string		getName		(void) const { return vecName<T, 4>(); }
2574
2575protected:
2576	IRet		doApply		(const EvalContext&, const IArgs& iargs) const
2577	{
2578		return IRet(iargs.a, iargs.b, iargs.c, iargs.d);
2579	}
2580};
2581
2582
2583
2584template <typename T, int Rows, int Columns>
2585class GenMat;
2586
2587template <typename T, int Rows>
2588class GenMat<T, Rows, 2> : public PrimitiveFunc<
2589	Signature<Matrix<T, Rows, 2>, Vector<T, Rows>, Vector<T, Rows> > >
2590{
2591public:
2592	typedef typename GenMat::Ret	Ret;
2593	typedef typename GenMat::IRet	IRet;
2594	typedef typename GenMat::IArgs	IArgs;
2595
2596	string		getName		(void) const
2597	{
2598		return dataTypeNameOf<Ret>();
2599	}
2600
2601protected:
2602
2603	IRet		doApply		(const EvalContext&, const IArgs& iargs) const
2604	{
2605		IRet	ret;
2606		ret[0] = iargs.a;
2607		ret[1] = iargs.b;
2608		return ret;
2609	}
2610};
2611
2612template <typename T, int Rows>
2613class GenMat<T, Rows, 3> : public PrimitiveFunc<
2614	Signature<Matrix<T, Rows, 3>, Vector<T, Rows>, Vector<T, Rows>, Vector<T, Rows> > >
2615{
2616public:
2617	typedef typename GenMat::Ret	Ret;
2618	typedef typename GenMat::IRet	IRet;
2619	typedef typename GenMat::IArgs	IArgs;
2620
2621	string	getName	(void) const
2622	{
2623		return dataTypeNameOf<Ret>();
2624	}
2625
2626protected:
2627
2628	IRet	doApply	(const EvalContext&, const IArgs& iargs) const
2629	{
2630		IRet	ret;
2631		ret[0] = iargs.a;
2632		ret[1] = iargs.b;
2633		ret[2] = iargs.c;
2634		return ret;
2635	}
2636};
2637
2638template <typename T, int Rows>
2639class GenMat<T, Rows, 4> : public PrimitiveFunc<
2640	Signature<Matrix<T, Rows, 4>,
2641			  Vector<T, Rows>, Vector<T, Rows>, Vector<T, Rows>, Vector<T, Rows> > >
2642{
2643public:
2644	typedef typename GenMat::Ret	Ret;
2645	typedef typename GenMat::IRet	IRet;
2646	typedef typename GenMat::IArgs	IArgs;
2647
2648	string	getName	(void) const
2649	{
2650		return dataTypeNameOf<Ret>();
2651	}
2652
2653protected:
2654	IRet	doApply	(const EvalContext&, const IArgs& iargs) const
2655	{
2656		IRet	ret;
2657		ret[0] = iargs.a;
2658		ret[1] = iargs.b;
2659		ret[2] = iargs.c;
2660		ret[3] = iargs.d;
2661		return ret;
2662	}
2663};
2664
2665template <typename T, int Rows>
2666ExprP<Matrix<T, Rows, 2> > mat2 (const ExprP<Vector<T, Rows> >& arg0,
2667								 const ExprP<Vector<T, Rows> >& arg1)
2668{
2669	return app<GenMat<T, Rows, 2> >(arg0, arg1);
2670}
2671
2672template <typename T, int Rows>
2673ExprP<Matrix<T, Rows, 3> > mat3 (const ExprP<Vector<T, Rows> >& arg0,
2674								 const ExprP<Vector<T, Rows> >& arg1,
2675								 const ExprP<Vector<T, Rows> >& arg2)
2676{
2677	return app<GenMat<T, Rows, 3> >(arg0, arg1, arg2);
2678}
2679
2680template <typename T, int Rows>
2681ExprP<Matrix<T, Rows, 4> > mat4 (const ExprP<Vector<T, Rows> >& arg0,
2682								 const ExprP<Vector<T, Rows> >& arg1,
2683								 const ExprP<Vector<T, Rows> >& arg2,
2684								 const ExprP<Vector<T, Rows> >& arg3)
2685{
2686	return app<GenMat<T, Rows, 4> >(arg0, arg1, arg2, arg3);
2687}
2688
2689
2690template <int Rows, int Cols>
2691class MatNeg : public PrimitiveFunc<Signature<Matrix<float, Rows, Cols>,
2692											  Matrix<float, Rows, Cols> > >
2693{
2694public:
2695	typedef typename MatNeg::IRet		IRet;
2696	typedef typename MatNeg::IArgs		IArgs;
2697
2698	string	getName	(void) const
2699	{
2700		return "_matNeg";
2701	}
2702
2703protected:
2704	void	doPrint	(ostream& os, const BaseArgExprs& args) const
2705	{
2706		os << "-(" << *args[0] << ")";
2707	}
2708
2709	IRet	doApply	(const EvalContext&, const IArgs& iargs)			const
2710	{
2711		IRet	ret;
2712
2713		for (int col = 0; col < Cols; ++col)
2714		{
2715			for (int row = 0; row < Rows; ++row)
2716				ret[col][row] = -iargs.a[col][row];
2717		}
2718
2719		return ret;
2720	}
2721};
2722
2723template <typename T, typename Sig>
2724class CompWiseFunc : public PrimitiveFunc<Sig>
2725{
2726public:
2727	typedef Func<Signature<T, T, T> >	ScalarFunc;
2728
2729	string				getName			(void)									const
2730	{
2731		return doGetScalarFunc().getName();
2732	}
2733protected:
2734	void				doPrint			(ostream&				os,
2735										 const BaseArgExprs&	args)			const
2736	{
2737		doGetScalarFunc().print(os, args);
2738	}
2739
2740	virtual
2741	const ScalarFunc&	doGetScalarFunc	(void)									const = 0;
2742};
2743
2744template <int Rows, int Cols>
2745class CompMatFuncBase : public CompWiseFunc<float, Signature<Matrix<float, Rows, Cols>,
2746															 Matrix<float, Rows, Cols>,
2747															 Matrix<float, Rows, Cols> > >
2748{
2749public:
2750	typedef typename CompMatFuncBase::IRet		IRet;
2751	typedef typename CompMatFuncBase::IArgs		IArgs;
2752
2753protected:
2754
2755	IRet	doApply	(const EvalContext& ctx, const IArgs& iargs) const
2756	{
2757		IRet			ret;
2758
2759		for (int col = 0; col < Cols; ++col)
2760		{
2761			for (int row = 0; row < Rows; ++row)
2762				ret[col][row] = this->doGetScalarFunc().apply(ctx,
2763															  iargs.a[col][row],
2764															  iargs.b[col][row]);
2765		}
2766
2767		return ret;
2768	}
2769};
2770
2771template <typename F, int Rows, int Cols>
2772class CompMatFunc : public CompMatFuncBase<Rows, Cols>
2773{
2774protected:
2775	const typename CompMatFunc::ScalarFunc&	doGetScalarFunc	(void) const
2776	{
2777		return instance<F>();
2778	}
2779};
2780
2781class ScalarMatrixCompMult : public Mul
2782{
2783public:
2784	string	getName	(void) const
2785	{
2786		return "matrixCompMult";
2787	}
2788
2789	void	doPrint	(ostream& os, const BaseArgExprs& args) const
2790	{
2791		Func<Sig>::doPrint(os, args);
2792	}
2793};
2794
2795template <int Rows, int Cols>
2796class MatrixCompMult : public CompMatFunc<ScalarMatrixCompMult, Rows, Cols>
2797{
2798};
2799
2800template <int Rows, int Cols>
2801class ScalarMatFuncBase : public CompWiseFunc<float, Signature<Matrix<float, Rows, Cols>,
2802															   Matrix<float, Rows, Cols>,
2803															   float> >
2804{
2805public:
2806	typedef typename ScalarMatFuncBase::IRet	IRet;
2807	typedef typename ScalarMatFuncBase::IArgs	IArgs;
2808
2809protected:
2810
2811	IRet	doApply	(const EvalContext& ctx, const IArgs& iargs) const
2812	{
2813		IRet	ret;
2814
2815		for (int col = 0; col < Cols; ++col)
2816		{
2817			for (int row = 0; row < Rows; ++row)
2818				ret[col][row] = this->doGetScalarFunc().apply(ctx, iargs.a[col][row], iargs.b);
2819		}
2820
2821		return ret;
2822	}
2823};
2824
2825template <typename F, int Rows, int Cols>
2826class ScalarMatFunc : public ScalarMatFuncBase<Rows, Cols>
2827{
2828protected:
2829	const typename ScalarMatFunc::ScalarFunc&	doGetScalarFunc	(void)	const
2830	{
2831		return instance<F>();
2832	}
2833};
2834
2835template<typename T, int Size> struct GenXType;
2836
2837template<typename T>
2838struct GenXType<T, 1>
2839{
2840	static ExprP<T>	genXType	(const ExprP<T>& x) { return x; }
2841};
2842
2843template<typename T>
2844struct GenXType<T, 2>
2845{
2846	static ExprP<Vector<T, 2> >	genXType	(const ExprP<T>& x)
2847	{
2848		return app<GenVec<T, 2> >(x, x);
2849	}
2850};
2851
2852template<typename T>
2853struct GenXType<T, 3>
2854{
2855	static ExprP<Vector<T, 3> >	genXType	(const ExprP<T>& x)
2856	{
2857		return app<GenVec<T, 3> >(x, x, x);
2858	}
2859};
2860
2861template<typename T>
2862struct GenXType<T, 4>
2863{
2864	static ExprP<Vector<T, 4> >	genXType	(const ExprP<T>& x)
2865	{
2866		return app<GenVec<T, 4> >(x, x, x, x);
2867	}
2868};
2869
2870//! Returns an expression of vector of size `Size` (or scalar if Size == 1),
2871//! with each element initialized with the expression `x`.
2872template<typename T, int Size>
2873ExprP<typename ContainerOf<T, Size>::Container> genXType (const ExprP<T>& x)
2874{
2875	return GenXType<T, Size>::genXType(x);
2876}
2877
2878typedef GenVec<float, 2> FloatVec2;
2879DEFINE_CONSTRUCTOR2(FloatVec2, Vec2, vec2, float, float)
2880
2881typedef GenVec<float, 3> FloatVec3;
2882DEFINE_CONSTRUCTOR3(FloatVec3, Vec3, vec3, float, float, float)
2883
2884typedef GenVec<float, 4> FloatVec4;
2885DEFINE_CONSTRUCTOR4(FloatVec4, Vec4, vec4, float, float, float, float)
2886
2887template <int Size>
2888class Dot : public DerivedFunc<Signature<float, Vector<float, Size>, Vector<float, Size> > >
2889{
2890public:
2891	typedef typename Dot::ArgExprs ArgExprs;
2892
2893	string			getName		(void) const
2894	{
2895		return "dot";
2896	}
2897
2898protected:
2899	ExprP<float>	doExpand 	(ExpandContext&, const ArgExprs& args) const
2900	{
2901		ExprP<float> val = args.a[0] * args.b[0];
2902
2903		for (int ndx = 1; ndx < Size; ++ndx)
2904			val = val + args.a[ndx] * args.b[ndx];
2905
2906		return val;
2907	}
2908};
2909
2910template <>
2911class Dot<1> : public DerivedFunc<Signature<float, float, float> >
2912{
2913public:
2914	string			getName		(void) const
2915	{
2916		return "dot";
2917	}
2918
2919	ExprP<float>	doExpand 	(ExpandContext&, const ArgExprs& args) const
2920	{
2921		return args.a * args.b;
2922	}
2923};
2924
2925template <int Size>
2926ExprP<float> dot (const ExprP<Vector<float, Size> >& x, const ExprP<Vector<float, Size> >& y)
2927{
2928	return app<Dot<Size> >(x, y);
2929}
2930
2931ExprP<float> dot (const ExprP<float>& x, const ExprP<float>& y)
2932{
2933	return app<Dot<1> >(x, y);
2934}
2935
2936template <int Size>
2937class Length : public DerivedFunc<
2938	Signature<float, typename ContainerOf<float, Size>::Container> >
2939{
2940public:
2941	typedef typename Length::ArgExprs ArgExprs;
2942
2943	string			getName		(void) const
2944	{
2945		return "length";
2946	}
2947
2948protected:
2949	ExprP<float>	doExpand	(ExpandContext&, const ArgExprs& args) const
2950	{
2951		return sqrt(dot(args.a, args.a));
2952	}
2953};
2954
2955template <int Size>
2956ExprP<float> length (const ExprP<typename ContainerOf<float, Size>::Container>& x)
2957{
2958	return app<Length<Size> >(x);
2959}
2960
2961template <int Size>
2962class Distance : public DerivedFunc<
2963	Signature<float,
2964			  typename ContainerOf<float, Size>::Container,
2965			  typename ContainerOf<float, Size>::Container> >
2966{
2967public:
2968	typedef typename	Distance::Ret		Ret;
2969	typedef typename	Distance::ArgExprs	ArgExprs;
2970
2971	string		getName		(void) const
2972	{
2973		return "distance";
2974	}
2975
2976protected:
2977	ExprP<Ret>	doExpand 	(ExpandContext&, const ArgExprs& args) const
2978	{
2979		return length<Size>(args.a - args.b);
2980	}
2981};
2982
2983// cross
2984
2985class Cross : public DerivedFunc<Signature<Vec3, Vec3, Vec3> >
2986{
2987public:
2988	string			getName		(void) const
2989	{
2990		return "cross";
2991	}
2992
2993protected:
2994	ExprP<Vec3>		doExpand 	(ExpandContext&, const ArgExprs& x) const
2995	{
2996		return vec3(x.a[1] * x.b[2] - x.b[1] * x.a[2],
2997					x.a[2] * x.b[0] - x.b[2] * x.a[0],
2998					x.a[0] * x.b[1] - x.b[0] * x.a[1]);
2999	}
3000};
3001
3002DEFINE_CONSTRUCTOR2(Cross, Vec3, cross, Vec3, Vec3)
3003
3004template<int Size>
3005class Normalize : public DerivedFunc<
3006	Signature<typename ContainerOf<float, Size>::Container,
3007			  typename ContainerOf<float, Size>::Container> >
3008{
3009public:
3010	typedef typename	Normalize::Ret		Ret;
3011	typedef typename	Normalize::ArgExprs	ArgExprs;
3012
3013	string		getName		(void) const
3014	{
3015		return "normalize";
3016	}
3017
3018protected:
3019	ExprP<Ret>	doExpand	(ExpandContext&, const ArgExprs& args) const
3020	{
3021		return args.a / length<Size>(args.a);
3022	}
3023};
3024
3025template <int Size>
3026class FaceForward : public DerivedFunc<
3027	Signature<typename ContainerOf<float, Size>::Container,
3028			  typename ContainerOf<float, Size>::Container,
3029			  typename ContainerOf<float, Size>::Container,
3030			  typename ContainerOf<float, Size>::Container> >
3031{
3032public:
3033	typedef typename	FaceForward::Ret		Ret;
3034	typedef typename	FaceForward::ArgExprs	ArgExprs;
3035
3036	string		getName		(void) const
3037	{
3038		return "faceforward";
3039	}
3040
3041protected:
3042
3043
3044	ExprP<Ret>	doExpand	(ExpandContext&, const ArgExprs& args) const
3045	{
3046		return cond(dot(args.c, args.b) < constant(0.0f), args.a, -args.a);
3047	}
3048};
3049
3050template <int Size>
3051class Reflect : public DerivedFunc<
3052	Signature<typename ContainerOf<float, Size>::Container,
3053			  typename ContainerOf<float, Size>::Container,
3054			  typename ContainerOf<float, Size>::Container> >
3055{
3056public:
3057	typedef typename	Reflect::Ret		Ret;
3058	typedef typename	Reflect::Arg0		Arg0;
3059	typedef typename	Reflect::Arg1		Arg1;
3060	typedef typename	Reflect::ArgExprs	ArgExprs;
3061
3062	string		getName		(void) const
3063	{
3064		return "reflect";
3065	}
3066
3067protected:
3068	ExprP<Ret>	doExpand	(ExpandContext& ctx, const ArgExprs& args) const
3069	{
3070		const ExprP<Arg0>&	i		= args.a;
3071		const ExprP<Arg1>&	n		= args.b;
3072		const ExprP<float>	dotNI	= bindExpression("dotNI", ctx, dot(n, i));
3073
3074		return i - alternatives((n * dotNI) * constant(2.0f),
3075								n * (dotNI * constant(2.0f)));
3076	}
3077};
3078
3079template <int Size>
3080class Refract : public DerivedFunc<
3081	Signature<typename ContainerOf<float, Size>::Container,
3082			  typename ContainerOf<float, Size>::Container,
3083			  typename ContainerOf<float, Size>::Container,
3084			  float> >
3085{
3086public:
3087	typedef typename	Refract::Ret		Ret;
3088	typedef typename	Refract::Arg0		Arg0;
3089	typedef typename	Refract::Arg1		Arg1;
3090	typedef typename	Refract::ArgExprs	ArgExprs;
3091
3092	string		getName		(void) const
3093	{
3094		return "refract";
3095	}
3096
3097protected:
3098	ExprP<Ret>	doExpand	(ExpandContext&	ctx, const ArgExprs& args) const
3099	{
3100		const ExprP<Arg0>&	i		= args.a;
3101		const ExprP<Arg1>&	n		= args.b;
3102		const ExprP<float>&	eta		= args.c;
3103		const ExprP<float>	dotNI	= bindExpression("dotNI", ctx, dot(n, i));
3104		const ExprP<float>	k		= bindExpression("k", ctx, constant(1.0f) - eta * eta *
3105												 (constant(1.0f) - dotNI * dotNI));
3106
3107		return cond(k < constant(0.0f),
3108					genXType<float, Size>(constant(0.0f)),
3109					i * eta - n * (eta * dotNI + sqrt(k)));
3110	}
3111};
3112
3113class PreciseFunc1 : public CFloatFunc1
3114{
3115public:
3116			PreciseFunc1	(const string& name, DoubleFunc1& func) : CFloatFunc1(name, func) {}
3117protected:
3118	double	precision		(const EvalContext&, double, double) const	{ return 0.0; }
3119};
3120
3121class Abs : public PreciseFunc1
3122{
3123public:
3124	Abs (void) : PreciseFunc1("abs", deAbs) {}
3125};
3126
3127class Sign : public PreciseFunc1
3128{
3129public:
3130	Sign (void) : PreciseFunc1("sign", deSign) {}
3131};
3132
3133class Floor : public PreciseFunc1
3134{
3135public:
3136	Floor (void) : PreciseFunc1("floor", deFloor) {}
3137};
3138
3139class Trunc : public PreciseFunc1
3140{
3141public:
3142	Trunc (void) : PreciseFunc1("trunc", deTrunc) {}
3143};
3144
3145class Round : public FloatFunc1
3146{
3147public:
3148	string		getName		(void) const								{ return "round"; }
3149
3150protected:
3151	Interval	applyPoint	(const EvalContext&, double x) const
3152	{
3153		double			truncated	= 0.0;
3154		const double	fract		= deModf(x, &truncated);
3155		Interval		ret;
3156
3157		if (fabs(fract) <= 0.5)
3158			ret |= truncated;
3159		if (fabs(fract) >= 0.5)
3160			ret |= truncated + deSign(fract);
3161
3162		return ret;
3163	}
3164
3165	double		precision	(const EvalContext&, double, double) const	{ return 0.0; }
3166};
3167
3168class RoundEven : public PreciseFunc1
3169{
3170public:
3171	RoundEven (void) : PreciseFunc1("roundEven", deRoundEven) {}
3172};
3173
3174class Ceil : public PreciseFunc1
3175{
3176public:
3177	Ceil (void) : PreciseFunc1("ceil", deCeil) {}
3178};
3179
3180DEFINE_DERIVED_FLOAT1(Fract, fract, x, x - app<Floor>(x));
3181
3182class PreciseFunc2 : public CFloatFunc2
3183{
3184public:
3185			PreciseFunc2	(const string& name, DoubleFunc2& func) : CFloatFunc2(name, func) {}
3186protected:
3187	double	precision		(const EvalContext&, double, double, double) const { return 0.0; }
3188};
3189
3190DEFINE_DERIVED_FLOAT2(Mod, mod, x, y, x - y * app<Floor>(x / y));
3191
3192class Modf : public PrimitiveFunc<Signature<float, float, float> >
3193{
3194public:
3195	string	getName				(void) const
3196	{
3197		return "modf";
3198	}
3199
3200protected:
3201	IRet	doApply				(const EvalContext& ctx, const IArgs& iargs) const
3202	{
3203		Interval	fracIV;
3204		Interval&	wholeIV		= const_cast<Interval&>(iargs.b);
3205		double		intPart		= 0;
3206
3207		TCU_INTERVAL_APPLY_MONOTONE1(fracIV, x, iargs.a, frac, frac = deModf(x, &intPart));
3208		TCU_INTERVAL_APPLY_MONOTONE1(wholeIV, x, iargs.a, whole,
3209									 deModf(x, &intPart); whole = intPart);
3210
3211		if ((ctx.format.hasInf() != YES) && !iargs.a.isFinite())
3212		{
3213			// Behavior on modf(Inf) not well-defined, allow anything as a fractional part
3214			fracIV |= TCU_NAN;
3215		}
3216
3217		return fracIV;
3218	}
3219
3220	int		getOutParamIndex	(void) const
3221	{
3222		return 1;
3223	}
3224};
3225
3226class Min : public PreciseFunc2 { public: Min (void) : PreciseFunc2("min", deMin) {} };
3227class Max : public PreciseFunc2 { public: Max (void) : PreciseFunc2("max", deMax) {} };
3228
3229class Clamp : public FloatFunc3
3230{
3231public:
3232	string	getName		(void) const { return "clamp"; }
3233
3234	double	applyExact	(double x, double minVal, double maxVal) const
3235	{
3236		return de::min(de::max(x, minVal), maxVal);
3237	}
3238
3239	double	precision	(const EvalContext&, double, double, double minVal, double maxVal) const
3240	{
3241		return minVal > maxVal ? TCU_NAN : 0.0;
3242	}
3243};
3244
3245ExprP<float> clamp(const ExprP<float>& x, const ExprP<float>& minVal, const ExprP<float>& maxVal)
3246{
3247	return app<Clamp>(x, minVal, maxVal);
3248}
3249
3250DEFINE_DERIVED_FLOAT3(Mix, mix, x, y, a, alternatives((x * (constant(1.0f) - a)) + y * a,
3251													  x + (y - x) * a));
3252
3253static double step (double edge, double x)
3254{
3255	return x < edge ? 0.0 : 1.0;
3256}
3257
3258class Step : public PreciseFunc2 { public: Step (void) : PreciseFunc2("step", step) {} };
3259
3260class SmoothStep : public DerivedFunc<Signature<float, float, float, float> >
3261{
3262public:
3263	string		getName		(void) const
3264	{
3265		return "smoothstep";
3266	}
3267
3268protected:
3269
3270	ExprP<Ret>	doExpand 	(ExpandContext& ctx, const ArgExprs& args) const
3271	{
3272		const ExprP<float>&		edge0	= args.a;
3273		const ExprP<float>&		edge1	= args.b;
3274		const ExprP<float>&		x		= args.c;
3275		const ExprP<float>		tExpr	= clamp((x - edge0) / (edge1 - edge0),
3276											constant(0.0f), constant(1.0f));
3277		const ExprP<float>		t		= bindExpression("t", ctx, tExpr);
3278
3279		return (t * t * (constant(3.0f) - constant(2.0f) * t));
3280	}
3281};
3282
3283class FrExp : public PrimitiveFunc<Signature<float, float, int> >
3284{
3285public:
3286	string	getName			(void) const
3287	{
3288		return "frexp";
3289	}
3290
3291protected:
3292	IRet	doApply			(const EvalContext&, const IArgs& iargs) const
3293	{
3294		IRet			ret;
3295		const IArg0&	x			= iargs.a;
3296		IArg1&			exponent	= const_cast<IArg1&>(iargs.b);
3297
3298		if (x.hasNaN() || x.contains(TCU_INFINITY) || x.contains(-TCU_INFINITY))
3299		{
3300			// GLSL (in contrast to IEEE) says that result of applying frexp
3301			// to infinity is undefined
3302			ret = Interval::unbounded() | TCU_NAN;
3303			exponent = Interval(-deLdExp(1.0, 31), deLdExp(1.0, 31)-1);
3304		}
3305		else if (!x.empty())
3306		{
3307			int				loExp	= 0;
3308			const double	loFrac	= deFrExp(x.lo(), &loExp);
3309			int				hiExp	= 0;
3310			const double	hiFrac	= deFrExp(x.hi(), &hiExp);
3311
3312			if (deSign(loFrac) != deSign(hiFrac))
3313			{
3314				exponent = Interval(-TCU_INFINITY, de::max(loExp, hiExp));
3315				ret = Interval();
3316				if (deSign(loFrac) < 0)
3317					ret |= Interval(-1.0 + DBL_EPSILON*0.5, 0.0);
3318				if (deSign(hiFrac) > 0)
3319					ret |= Interval(0.0, 1.0 - DBL_EPSILON*0.5);
3320			}
3321			else
3322			{
3323				exponent = Interval(loExp, hiExp);
3324				if (loExp == hiExp)
3325					ret = Interval(loFrac, hiFrac);
3326				else
3327					ret = deSign(loFrac) * Interval(0.5, 1.0 - DBL_EPSILON*0.5);
3328			}
3329		}
3330
3331		return ret;
3332	}
3333
3334	int	getOutParamIndex	(void) const
3335	{
3336		return 1;
3337	}
3338};
3339
3340class LdExp : public PrimitiveFunc<Signature<float, float, int> >
3341{
3342public:
3343	string		getName			(void) const
3344	{
3345		return "ldexp";
3346	}
3347
3348protected:
3349	Interval	doApply			(const EvalContext& ctx, const IArgs& iargs) const
3350	{
3351		Interval	ret = call<Exp2>(ctx, iargs.b);
3352		// Khronos bug 11180 consensus: if exp2(exponent) cannot be represented,
3353		// the result is undefined.
3354
3355		if (ret.contains(TCU_INFINITY) | ret.contains(-TCU_INFINITY))
3356			ret |= TCU_NAN;
3357
3358		return call<Mul>(ctx, iargs.a, ret);
3359	}
3360};
3361
3362template<int Rows, int Columns>
3363class Transpose : public PrimitiveFunc<Signature<Matrix<float, Rows, Columns>,
3364												 Matrix<float, Columns, Rows> > >
3365{
3366public:
3367	typedef typename Transpose::IRet	IRet;
3368	typedef typename Transpose::IArgs	IArgs;
3369
3370	string		getName		(void) const
3371	{
3372		return "transpose";
3373	}
3374
3375protected:
3376	IRet		doApply		(const EvalContext&, const IArgs& iargs) const
3377	{
3378		IRet ret;
3379
3380		for (int rowNdx = 0; rowNdx < Rows; ++rowNdx)
3381		{
3382			for (int colNdx = 0; colNdx < Columns; ++colNdx)
3383				ret(rowNdx, colNdx) = iargs.a(colNdx, rowNdx);
3384		}
3385
3386		return ret;
3387	}
3388};
3389
3390template<typename Ret, typename Arg0, typename Arg1>
3391class MulFunc : public PrimitiveFunc<Signature<Ret, Arg0, Arg1> >
3392{
3393public:
3394	string	getName	(void) const 									{ return "mul"; }
3395
3396protected:
3397	void	doPrint	(ostream& os, const BaseArgExprs& args) const
3398	{
3399		os << "(" << *args[0] << " * " << *args[1] << ")";
3400	}
3401};
3402
3403template<int LeftRows, int Middle, int RightCols>
3404class MatMul : public MulFunc<Matrix<float, LeftRows, RightCols>,
3405							  Matrix<float, LeftRows, Middle>,
3406							  Matrix<float, Middle, RightCols> >
3407{
3408protected:
3409	typedef typename MatMul::IRet	IRet;
3410	typedef typename MatMul::IArgs	IArgs;
3411	typedef typename MatMul::IArg0	IArg0;
3412	typedef typename MatMul::IArg1	IArg1;
3413
3414	IRet	doApply	(const EvalContext&	ctx, const IArgs& iargs) const
3415	{
3416		const IArg0&	left	= iargs.a;
3417		const IArg1&	right	= iargs.b;
3418		IRet			ret;
3419
3420		for (int row = 0; row < LeftRows; ++row)
3421		{
3422			for (int col = 0; col < RightCols; ++col)
3423			{
3424				Interval	element	(0.0);
3425
3426				for (int ndx = 0; ndx < Middle; ++ndx)
3427					element = call<Add>(ctx, element,
3428										call<Mul>(ctx, left[ndx][row], right[col][ndx]));
3429
3430				ret[col][row] = element;
3431			}
3432		}
3433
3434		return ret;
3435	}
3436};
3437
3438template<int Rows, int Cols>
3439class VecMatMul : public MulFunc<Vector<float, Cols>,
3440								 Vector<float, Rows>,
3441								 Matrix<float, Rows, Cols> >
3442{
3443public:
3444	typedef typename VecMatMul::IRet	IRet;
3445	typedef typename VecMatMul::IArgs	IArgs;
3446	typedef typename VecMatMul::IArg0	IArg0;
3447	typedef typename VecMatMul::IArg1	IArg1;
3448
3449protected:
3450	IRet	doApply	(const EvalContext& ctx, const IArgs& iargs) const
3451	{
3452		const IArg0&	left	= iargs.a;
3453		const IArg1&	right	= iargs.b;
3454		IRet			ret;
3455
3456		for (int col = 0; col < Cols; ++col)
3457		{
3458			Interval	element	(0.0);
3459
3460			for (int row = 0; row < Rows; ++row)
3461				element = call<Add>(ctx, element, call<Mul>(ctx, left[row], right[col][row]));
3462
3463			ret[col] = element;
3464		}
3465
3466		return ret;
3467	}
3468};
3469
3470template<int Rows, int Cols>
3471class MatVecMul : public MulFunc<Vector<float, Rows>,
3472								 Matrix<float, Rows, Cols>,
3473								 Vector<float, Cols> >
3474{
3475public:
3476	typedef typename MatVecMul::IRet	IRet;
3477	typedef typename MatVecMul::IArgs	IArgs;
3478	typedef typename MatVecMul::IArg0	IArg0;
3479	typedef typename MatVecMul::IArg1	IArg1;
3480
3481protected:
3482	IRet	doApply	(const EvalContext& ctx, const IArgs& iargs) const
3483	{
3484		const IArg0&	left	= iargs.a;
3485		const IArg1&	right	= iargs.b;
3486
3487		return call<VecMatMul<Cols, Rows> >(ctx, right,
3488											call<Transpose<Rows, Cols> >(ctx, left));
3489	}
3490};
3491
3492template<int Rows, int Cols>
3493class OuterProduct : public PrimitiveFunc<Signature<Matrix<float, Rows, Cols>,
3494													Vector<float, Rows>,
3495													Vector<float, Cols> > >
3496{
3497public:
3498	typedef typename OuterProduct::IRet		IRet;
3499	typedef typename OuterProduct::IArgs	IArgs;
3500
3501	string	getName	(void) const
3502	{
3503		return "outerProduct";
3504	}
3505
3506protected:
3507	IRet	doApply	(const EvalContext& ctx, const IArgs& iargs) const
3508	{
3509		IRet	ret;
3510
3511		for (int row = 0; row < Rows; ++row)
3512		{
3513			for (int col = 0; col < Cols; ++col)
3514				ret[col][row] = call<Mul>(ctx, iargs.a[row], iargs.b[col]);
3515		}
3516
3517		return ret;
3518	}
3519};
3520
3521template<int Rows, int Cols>
3522ExprP<Matrix<float, Rows, Cols> > outerProduct (const ExprP<Vector<float, Rows> >& left,
3523												const ExprP<Vector<float, Cols> >& right)
3524{
3525	return app<OuterProduct<Rows, Cols> >(left, right);
3526}
3527
3528template<int Size>
3529class DeterminantBase : public DerivedFunc<Signature<float, Matrix<float, Size, Size> > >
3530{
3531public:
3532	string	getName	(void) const { return "determinant"; }
3533};
3534
3535template<int Size>
3536class Determinant;
3537
3538template<int Size>
3539ExprP<float> determinant (ExprP<Matrix<float, Size, Size> > mat)
3540{
3541	return app<Determinant<Size> >(mat);
3542}
3543
3544template<>
3545class Determinant<2> : public DeterminantBase<2>
3546{
3547protected:
3548	ExprP<Ret>	doExpand (ExpandContext&, const ArgExprs& args)	const
3549	{
3550		ExprP<Mat2>	mat	= args.a;
3551
3552		return mat[0][0] * mat[1][1] - mat[1][0] * mat[0][1];
3553	}
3554};
3555
3556template<>
3557class Determinant<3> : public DeterminantBase<3>
3558{
3559protected:
3560	ExprP<Ret> doExpand (ExpandContext&, const ArgExprs& args) const
3561	{
3562		ExprP<Mat3>	mat	= args.a;
3563
3564		return (mat[0][0] * (mat[1][1] * mat[2][2] - mat[1][2] * mat[2][1]) +
3565				mat[0][1] * (mat[1][2] * mat[2][0] - mat[1][0] * mat[2][2]) +
3566				mat[0][2] * (mat[1][0] * mat[2][1] - mat[1][1] * mat[2][0]));
3567	}
3568};
3569
3570template<>
3571class Determinant<4> : public DeterminantBase<4>
3572{
3573protected:
3574	 ExprP<Ret>	doExpand	(ExpandContext& ctx, const ArgExprs& args) const
3575	{
3576		ExprP<Mat4>	mat	= args.a;
3577		ExprP<Mat3>	minors[4];
3578
3579		for (int ndx = 0; ndx < 4; ++ndx)
3580		{
3581			ExprP<Vec4>		minorColumns[3];
3582			ExprP<Vec3>		columns[3];
3583
3584			for (int col = 0; col < 3; ++col)
3585				minorColumns[col] = mat[col < ndx ? col : col + 1];
3586
3587			for (int col = 0; col < 3; ++col)
3588				columns[col] = vec3(minorColumns[0][col+1],
3589									minorColumns[1][col+1],
3590									minorColumns[2][col+1]);
3591
3592			minors[ndx] = bindExpression("minor", ctx,
3593										 mat3(columns[0], columns[1], columns[2]));
3594		}
3595
3596		return (mat[0][0] * determinant(minors[0]) -
3597				mat[1][0] * determinant(minors[1]) +
3598				mat[2][0] * determinant(minors[2]) -
3599				mat[3][0] * determinant(minors[3]));
3600	}
3601};
3602
3603template<int Size> class Inverse;
3604
3605template <int Size>
3606ExprP<Matrix<float, Size, Size> > inverse (ExprP<Matrix<float, Size, Size> > mat)
3607{
3608	return app<Inverse<Size> >(mat);
3609}
3610
3611template<>
3612class Inverse<2> : public DerivedFunc<Signature<Mat2, Mat2> >
3613{
3614public:
3615	string		getName	(void) const
3616	{
3617		return "inverse";
3618	}
3619
3620protected:
3621	ExprP<Ret>	doExpand (ExpandContext& ctx, const ArgExprs& args) const
3622	{
3623		ExprP<Mat2>		mat = args.a;
3624		ExprP<float>	det	= bindExpression("det", ctx, determinant(mat));
3625
3626		return mat2(vec2(mat[1][1] / det, -mat[0][1] / det),
3627					vec2(-mat[1][0] / det, mat[0][0] / det));
3628	}
3629};
3630
3631template<>
3632class Inverse<3> : public DerivedFunc<Signature<Mat3, Mat3> >
3633{
3634public:
3635	string		getName		(void) const
3636	{
3637		return "inverse";
3638	}
3639
3640protected:
3641	ExprP<Ret>	doExpand 	(ExpandContext& ctx, const ArgExprs& args)			const
3642	{
3643		ExprP<Mat3>		mat		= args.a;
3644		ExprP<Mat2>		invA	= bindExpression("invA", ctx,
3645												 inverse(mat2(vec2(mat[0][0], mat[0][1]),
3646															  vec2(mat[1][0], mat[1][1]))));
3647
3648		ExprP<Vec2>		matB	= bindExpression("matB", ctx, vec2(mat[2][0], mat[2][1]));
3649		ExprP<Vec2>		matC	= bindExpression("matC", ctx, vec2(mat[0][2], mat[1][2]));
3650		ExprP<float>	matD	= bindExpression("matD", ctx, mat[2][2]);
3651
3652		ExprP<float>	schur	= bindExpression("schur", ctx,
3653												 constant(1.0f) /
3654												 (matD - dot(matC * invA, matB)));
3655
3656		ExprP<Vec2>		t1		= invA * matB;
3657		ExprP<Vec2>		t2		= t1 * schur;
3658		ExprP<Mat2>		t3		= outerProduct(t2, matC);
3659		ExprP<Mat2>		t4		= t3 * invA;
3660		ExprP<Mat2>		t5		= invA + t4;
3661		ExprP<Mat2>		blockA	= bindExpression("blockA", ctx, t5);
3662		ExprP<Vec2>		blockB	= bindExpression("blockB", ctx,
3663												 (invA * matB) * -schur);
3664		ExprP<Vec2>		blockC	= bindExpression("blockC", ctx,
3665												 (matC * invA) * -schur);
3666
3667		return mat3(vec3(blockA[0][0], blockA[0][1], blockC[0]),
3668					vec3(blockA[1][0], blockA[1][1], blockC[1]),
3669					vec3(blockB[0], blockB[1], schur));
3670	}
3671};
3672
3673template<>
3674class Inverse<4> : public DerivedFunc<Signature<Mat4, Mat4> >
3675{
3676public:
3677	string		getName		(void) const { return "inverse"; }
3678
3679protected:
3680	ExprP<Ret>			doExpand 			(ExpandContext&		ctx,
3681											 const ArgExprs&	args)			const
3682	{
3683		ExprP<Mat4>	mat		= args.a;
3684		ExprP<Mat2>	invA	= bindExpression("invA", ctx,
3685											 inverse(mat2(vec2(mat[0][0], mat[0][1]),
3686														  vec2(mat[1][0], mat[1][1]))));
3687		ExprP<Mat2>	matB	= bindExpression("matB", ctx,
3688											 mat2(vec2(mat[2][0], mat[2][1]),
3689												  vec2(mat[3][0], mat[3][1])));
3690		ExprP<Mat2>	matC	= bindExpression("matC", ctx,
3691											 mat2(vec2(mat[0][2], mat[0][3]),
3692												  vec2(mat[1][2], mat[1][3])));
3693		ExprP<Mat2>	matD	= bindExpression("matD", ctx,
3694											 mat2(vec2(mat[2][2], mat[2][3]),
3695												  vec2(mat[3][2], mat[3][3])));
3696		ExprP<Mat2>	schur	= bindExpression("schur", ctx,
3697											 inverse(matD + -(matC * invA * matB)));
3698		ExprP<Mat2>	blockA	= bindExpression("blockA", ctx,
3699											 invA + (invA * matB * schur * matC * invA));
3700		ExprP<Mat2>	blockB	= bindExpression("blockB", ctx,
3701											 (-invA) * matB * schur);
3702		ExprP<Mat2>	blockC	= bindExpression("blockC", ctx,
3703											 (-schur) * matC * invA);
3704
3705		return mat4(vec4(blockA[0][0], blockA[0][1], blockC[0][0], blockC[0][1]),
3706					vec4(blockA[1][0], blockA[1][1], blockC[1][0], blockC[1][1]),
3707					vec4(blockB[0][0], blockB[0][1], schur[0][0], schur[0][1]),
3708					vec4(blockB[1][0], blockB[1][1], schur[1][0], schur[1][1]));
3709	}
3710};
3711
3712class Fma : public DerivedFunc<Signature<float, float, float, float> >
3713{
3714public:
3715	string			getName					(void) const
3716	{
3717		return "fma";
3718	}
3719
3720	string			getRequiredExtension	(void) const
3721	{
3722		return "GL_EXT_gpu_shader5";
3723	}
3724
3725protected:
3726	ExprP<float>	doExpand 				(ExpandContext&, const ArgExprs& x) const
3727	{
3728		return x.a * x.b + x.c;
3729	}
3730};
3731
3732} // Functions
3733
3734using namespace Functions;
3735
3736template <typename T>
3737ExprP<typename T::Element> ContainerExprPBase<T>::operator[] (int i) const
3738{
3739	return Functions::getComponent(exprP<T>(*this), i);
3740}
3741
3742ExprP<float> operator+ (const ExprP<float>& arg0, const ExprP<float>& arg1)
3743{
3744	return app<Add>(arg0, arg1);
3745}
3746
3747ExprP<float> operator- (const ExprP<float>& arg0, const ExprP<float>& arg1)
3748{
3749	return app<Sub>(arg0, arg1);
3750}
3751
3752ExprP<float> operator- (const ExprP<float>& arg0)
3753{
3754	return app<Negate>(arg0);
3755}
3756
3757ExprP<float> operator* (const ExprP<float>& arg0, const ExprP<float>& arg1)
3758{
3759	return app<Mul>(arg0, arg1);
3760}
3761
3762ExprP<float> operator/ (const ExprP<float>& arg0, const ExprP<float>& arg1)
3763{
3764	return app<Div>(arg0, arg1);
3765}
3766
3767template <typename Sig_, int Size>
3768class GenFunc : public PrimitiveFunc<Signature<
3769	typename ContainerOf<typename Sig_::Ret, Size>::Container,
3770	typename ContainerOf<typename Sig_::Arg0, Size>::Container,
3771	typename ContainerOf<typename Sig_::Arg1, Size>::Container,
3772	typename ContainerOf<typename Sig_::Arg2, Size>::Container,
3773	typename ContainerOf<typename Sig_::Arg3, Size>::Container> >
3774{
3775public:
3776	typedef typename GenFunc::IArgs		IArgs;
3777	typedef typename GenFunc::IRet		IRet;
3778
3779			GenFunc					(const Func<Sig_>&	scalarFunc) : m_func (scalarFunc) {}
3780
3781	string	getName					(void) const
3782	{
3783		return m_func.getName();
3784	}
3785
3786	int		getOutParamIndex		(void) const
3787	{
3788		return m_func.getOutParamIndex();
3789	}
3790
3791	string	getRequiredExtension	(void) const
3792	{
3793		return m_func.getRequiredExtension();
3794	}
3795
3796protected:
3797	void	doPrint					(ostream& os, const BaseArgExprs& args) const
3798	{
3799		m_func.print(os, args);
3800	}
3801
3802	IRet	doApply					(const EvalContext& ctx, const IArgs& iargs) const
3803	{
3804		IRet ret;
3805
3806		for (int ndx = 0; ndx < Size; ++ndx)
3807		{
3808			ret[ndx] =
3809				m_func.apply(ctx, iargs.a[ndx], iargs.b[ndx], iargs.c[ndx], iargs.d[ndx]);
3810		}
3811
3812		return ret;
3813	}
3814
3815	void	doGetUsedFuncs			(FuncSet& dst) const
3816	{
3817		m_func.getUsedFuncs(dst);
3818	}
3819
3820	const Func<Sig_>&	m_func;
3821};
3822
3823template <typename F, int Size>
3824class VectorizedFunc : public GenFunc<typename F::Sig, Size>
3825{
3826public:
3827	VectorizedFunc	(void) : GenFunc<typename F::Sig, Size>(instance<F>()) {}
3828};
3829
3830
3831
3832template <typename Sig_, int Size>
3833class FixedGenFunc : public PrimitiveFunc <Signature<
3834	typename ContainerOf<typename Sig_::Ret, Size>::Container,
3835	typename ContainerOf<typename Sig_::Arg0, Size>::Container,
3836	typename Sig_::Arg1,
3837	typename ContainerOf<typename Sig_::Arg2, Size>::Container,
3838	typename ContainerOf<typename Sig_::Arg3, Size>::Container> >
3839{
3840public:
3841	typedef typename FixedGenFunc::IArgs		IArgs;
3842	typedef typename FixedGenFunc::IRet			IRet;
3843
3844	string						getName			(void) const
3845	{
3846		return this->doGetScalarFunc().getName();
3847	}
3848
3849protected:
3850	void						doPrint			(ostream& os, const BaseArgExprs& args) const
3851	{
3852		this->doGetScalarFunc().print(os, args);
3853	}
3854
3855	IRet						doApply			(const EvalContext& ctx,
3856												 const IArgs&		iargs) const
3857	{
3858		IRet				ret;
3859		const Func<Sig_>&	func	= this->doGetScalarFunc();
3860
3861		for (int ndx = 0; ndx < Size; ++ndx)
3862			ret[ndx] = func.apply(ctx, iargs.a[ndx], iargs.b, iargs.c[ndx], iargs.d[ndx]);
3863
3864		return ret;
3865	}
3866
3867	virtual const Func<Sig_>&	doGetScalarFunc	(void) const = 0;
3868};
3869
3870template <typename F, int Size>
3871class FixedVecFunc : public FixedGenFunc<typename F::Sig, Size>
3872{
3873protected:
3874	const Func<typename F::Sig>& doGetScalarFunc	(void) const { return instance<F>(); }
3875};
3876
3877template<typename Sig>
3878struct GenFuncs
3879{
3880	GenFuncs (const Func<Sig>&			func_,
3881			  const GenFunc<Sig, 2>&	func2_,
3882			  const GenFunc<Sig, 3>&	func3_,
3883			  const GenFunc<Sig, 4>&	func4_)
3884		: func	(func_)
3885		, func2	(func2_)
3886		, func3	(func3_)
3887		, func4	(func4_)
3888	{}
3889
3890	const Func<Sig>&		func;
3891	const GenFunc<Sig, 2>&	func2;
3892	const GenFunc<Sig, 3>&	func3;
3893	const GenFunc<Sig, 4>&	func4;
3894};
3895
3896template<typename F>
3897GenFuncs<typename F::Sig> makeVectorizedFuncs (void)
3898{
3899	return GenFuncs<typename F::Sig>(instance<F>(),
3900									 instance<VectorizedFunc<F, 2> >(),
3901									 instance<VectorizedFunc<F, 3> >(),
3902									 instance<VectorizedFunc<F, 4> >());
3903}
3904
3905template<int Size>
3906ExprP<Vector<float, Size> > operator*(const ExprP<Vector<float, Size> >& arg0,
3907									  const ExprP<Vector<float, Size> >& arg1)
3908{
3909	return app<VectorizedFunc<Mul, Size> >(arg0, arg1);
3910}
3911
3912template<int Size>
3913ExprP<Vector<float, Size> > operator*(const ExprP<Vector<float, Size> >&	arg0,
3914									  const ExprP<float>&					arg1)
3915{
3916	return app<FixedVecFunc<Mul, Size> >(arg0, arg1);
3917}
3918
3919template<int Size>
3920ExprP<Vector<float, Size> > operator/(const ExprP<Vector<float, Size> >&	arg0,
3921									  const ExprP<float>&					arg1)
3922{
3923	return app<FixedVecFunc<Div, Size> >(arg0, arg1);
3924}
3925
3926template<int Size>
3927ExprP<Vector<float, Size> > operator-(const ExprP<Vector<float, Size> >& arg0)
3928{
3929	return app<VectorizedFunc<Negate, Size> >(arg0);
3930}
3931
3932template<int Size>
3933ExprP<Vector<float, Size> > operator-(const ExprP<Vector<float, Size> >& arg0,
3934									  const ExprP<Vector<float, Size> >& arg1)
3935{
3936	return app<VectorizedFunc<Sub, Size> >(arg0, arg1);
3937}
3938
3939template<int LeftRows, int Middle, int RightCols>
3940ExprP<Matrix<float, LeftRows, RightCols> >
3941operator* (const ExprP<Matrix<float, LeftRows, Middle> >&	left,
3942		   const ExprP<Matrix<float, Middle, RightCols> >&	right)
3943{
3944	return app<MatMul<LeftRows, Middle, RightCols> >(left, right);
3945}
3946
3947template<int Rows, int Cols>
3948ExprP<Vector<float, Rows> > operator* (const ExprP<Vector<float, Cols> >&		left,
3949									   const ExprP<Matrix<float, Rows, Cols> >&	right)
3950{
3951	return app<VecMatMul<Rows, Cols> >(left, right);
3952}
3953
3954template<int Rows, int Cols>
3955ExprP<Vector<float, Cols> > operator* (const ExprP<Matrix<float, Rows, Cols> >&	left,
3956									   const ExprP<Vector<float, Rows> >&		right)
3957{
3958	return app<MatVecMul<Rows, Cols> >(left, right);
3959}
3960
3961template<int Rows, int Cols>
3962ExprP<Matrix<float, Rows, Cols> > operator* (const ExprP<Matrix<float, Rows, Cols> >&	left,
3963											 const ExprP<float>&						right)
3964{
3965	return app<ScalarMatFunc<Mul, Rows, Cols> >(left, right);
3966}
3967
3968template<int Rows, int Cols>
3969ExprP<Matrix<float, Rows, Cols> > operator+ (const ExprP<Matrix<float, Rows, Cols> >&	left,
3970											 const ExprP<Matrix<float, Rows, Cols> >&	right)
3971{
3972	return app<CompMatFunc<Add, Rows, Cols> >(left, right);
3973}
3974
3975template<int Rows, int Cols>
3976ExprP<Matrix<float, Rows, Cols> > operator- (const ExprP<Matrix<float, Rows, Cols> >&	mat)
3977{
3978	return app<MatNeg<Rows, Cols> >(mat);
3979}
3980
3981template <typename T>
3982class Sampling
3983{
3984public:
3985	virtual void	genFixeds	(const FloatFormat&, vector<T>&)			const {}
3986	virtual T		genRandom	(const FloatFormat&, Precision, Random&)	const { return T(); }
3987	virtual double	getWeight	(void)										const { return 0.0; }
3988};
3989
3990template <>
3991class DefaultSampling<Void> : public Sampling<Void>
3992{
3993public:
3994	void	genFixeds	(const FloatFormat&, vector<Void>& dst) const { dst.push_back(Void()); }
3995};
3996
3997template <>
3998class DefaultSampling<bool> : public Sampling<bool>
3999{
4000public:
4001	void	genFixeds	(const FloatFormat&, vector<bool>& dst) const
4002	{
4003		dst.push_back(true);
4004		dst.push_back(false);
4005	}
4006};
4007
4008template <>
4009class DefaultSampling<int> : public Sampling<int>
4010{
4011public:
4012	int		genRandom	(const FloatFormat&, Precision prec, Random& rnd) const
4013	{
4014		const int	exp		= rnd.getInt(0, getNumBits(prec)-2);
4015		const int	sign	= rnd.getBool() ? -1 : 1;
4016
4017		return sign * rnd.getInt(0, (deInt32)1 << exp);
4018	}
4019
4020	void	genFixeds	(const FloatFormat&, vector<int>& dst) const
4021	{
4022		dst.push_back(0);
4023		dst.push_back(-1);
4024		dst.push_back(1);
4025	}
4026	double	getWeight	(void) const { return 1.0; }
4027
4028private:
4029	static inline int getNumBits (Precision prec)
4030	{
4031		switch (prec)
4032		{
4033			case glu::PRECISION_LOWP:		return 8;
4034			case glu::PRECISION_MEDIUMP:	return 16;
4035			case glu::PRECISION_HIGHP:		return 32;
4036			default:
4037				DE_ASSERT(false);
4038				return 0;
4039		}
4040	}
4041};
4042
4043template <>
4044class DefaultSampling<float> : public Sampling<float>
4045{
4046public:
4047	float	genRandom	(const FloatFormat& format, Precision prec, Random& rnd) const;
4048	void	genFixeds	(const FloatFormat& format, vector<float>& dst) const;
4049	double	getWeight	(void) const { return 1.0; }
4050};
4051
4052//! Generate a random float from a reasonable general-purpose distribution.
4053float DefaultSampling<float>::genRandom (const FloatFormat& format,
4054										 Precision,
4055										 Random&			rnd) const
4056{
4057	const int		minExp			= format.getMinExp();
4058	const int		maxExp			= format.getMaxExp();
4059	const bool		haveSubnormal	= format.hasSubnormal() != tcu::NO;
4060
4061	// Choose exponent so that the cumulative distribution is cubic.
4062	// This makes the probability distribution quadratic, with the peak centered on zero.
4063	const double	minRoot			= deCbrt(minExp - 0.5 - (haveSubnormal ? 1.0 : 0.0));
4064	const double	maxRoot			= deCbrt(maxExp + 0.5);
4065	const int		fractionBits	= format.getFractionBits();
4066	const int		exp				= int(deRoundEven(dePow(rnd.getDouble(minRoot, maxRoot),
4067															3.0)));
4068	float			base			= 0.0f; // integral power of two
4069	float			quantum			= 0.0f; // smallest representable difference in the binade
4070	float			significand		= 0.0f; // Significand.
4071
4072	DE_ASSERT(fractionBits < std::numeric_limits<float>::digits);
4073
4074	// Generate some occasional special numbers
4075	switch (rnd.getInt(0, 64))
4076	{
4077		case 0: 	return 0;
4078		case 1:		return TCU_INFINITY;
4079		case 2:		return -TCU_INFINITY;
4080		case 3:		return TCU_NAN;
4081		default:	break;
4082	}
4083
4084	if (exp >= minExp)
4085	{
4086		// Normal number
4087		base = deFloatLdExp(1.0f, exp);
4088		quantum = deFloatLdExp(1.0f, exp - fractionBits);
4089	}
4090	else
4091	{
4092		// Subnormal
4093		base = 0.0f;
4094		quantum = deFloatLdExp(1.0f, minExp - fractionBits);
4095	}
4096
4097	switch (rnd.getInt(0, 16))
4098	{
4099		case 0: // The highest number in this binade, significand is all bits one.
4100			significand = base - quantum;
4101			break;
4102		case 1: // Significand is one.
4103			significand = quantum;
4104			break;
4105		case 2: // Significand is zero.
4106			significand = 0.0;
4107			break;
4108		default: // Random (evenly distributed) significand.
4109		{
4110			deUint64 intFraction = rnd.getUint64() & ((1 << fractionBits) - 1);
4111			significand = float(intFraction) * quantum;
4112		}
4113	}
4114
4115	// Produce positive numbers more often than negative.
4116	return (rnd.getInt(0,3) == 0 ? -1.0f : 1.0f) * (base + significand);
4117}
4118
4119//! Generate a standard set of floats that should always be tested.
4120void DefaultSampling<float>::genFixeds (const FloatFormat& format, vector<float>& dst) const
4121{
4122	const int			minExp			= format.getMinExp();
4123	const int			maxExp			= format.getMaxExp();
4124	const int			fractionBits	= format.getFractionBits();
4125	const float			minQuantum		= deFloatLdExp(1.0f, minExp - fractionBits);
4126	const float			minNormalized	= deFloatLdExp(1.0f, minExp);
4127	const float			maxQuantum		= deFloatLdExp(1.0f, maxExp - fractionBits);
4128
4129	// NaN
4130	dst.push_back(TCU_NAN);
4131	// Zero
4132	dst.push_back(0.0f);
4133
4134	for (int sign = -1; sign <= 1; sign += 2)
4135	{
4136		// Smallest subnormal
4137		dst.push_back((float)sign * minQuantum);
4138
4139		// Largest subnormal
4140		dst.push_back((float)sign * (minNormalized - minQuantum));
4141
4142		// Smallest normalized
4143		dst.push_back((float)sign * minNormalized);
4144
4145		// Next smallest normalized
4146		dst.push_back((float)sign * (minNormalized + minQuantum));
4147
4148		dst.push_back((float)sign * 0.5f);
4149		dst.push_back((float)sign * 1.0f);
4150		dst.push_back((float)sign * 2.0f);
4151
4152		// Largest number
4153		dst.push_back((float)sign * (deFloatLdExp(1.0f, maxExp) +
4154									(deFloatLdExp(1.0f, maxExp) - maxQuantum)));
4155
4156		dst.push_back((float)sign * TCU_INFINITY);
4157	}
4158}
4159
4160template <typename T, int Size>
4161class DefaultSampling<Vector<T, Size> > : public Sampling<Vector<T, Size> >
4162{
4163public:
4164	typedef Vector<T, Size>		Value;
4165
4166	Value	genRandom	(const FloatFormat& fmt, Precision prec, Random& rnd) const
4167	{
4168		Value ret;
4169
4170		for (int ndx = 0; ndx < Size; ++ndx)
4171			ret[ndx] = instance<DefaultSampling<T> >().genRandom(fmt, prec, rnd);
4172
4173		return ret;
4174	}
4175
4176	void	genFixeds	(const FloatFormat& fmt, vector<Value>& dst) const
4177	{
4178		vector<T> scalars;
4179
4180		instance<DefaultSampling<T> >().genFixeds(fmt, scalars);
4181
4182		for (size_t scalarNdx = 0; scalarNdx < scalars.size(); ++scalarNdx)
4183			dst.push_back(Value(scalars[scalarNdx]));
4184	}
4185
4186	double	getWeight	(void) const
4187	{
4188		return dePow(instance<DefaultSampling<T> >().getWeight(), Size);
4189	}
4190};
4191
4192template <typename T, int Rows, int Columns>
4193class DefaultSampling<Matrix<T, Rows, Columns> > : public Sampling<Matrix<T, Rows, Columns> >
4194{
4195public:
4196	typedef Matrix<T, Rows, Columns>		Value;
4197
4198	Value	genRandom	(const FloatFormat& fmt, Precision prec, Random& rnd) const
4199	{
4200		Value ret;
4201
4202		for (int rowNdx = 0; rowNdx < Rows; ++rowNdx)
4203			for (int colNdx = 0; colNdx < Columns; ++colNdx)
4204				ret(rowNdx, colNdx) = instance<DefaultSampling<T> >().genRandom(fmt, prec, rnd);
4205
4206		return ret;
4207	}
4208
4209	void	genFixeds	(const FloatFormat& fmt, vector<Value>& dst) const
4210	{
4211		vector<T> scalars;
4212
4213		instance<DefaultSampling<T> >().genFixeds(fmt, scalars);
4214
4215		for (size_t scalarNdx = 0; scalarNdx < scalars.size(); ++scalarNdx)
4216			dst.push_back(Value(scalars[scalarNdx]));
4217
4218		if (Columns == Rows)
4219		{
4220			Value	mat	(0.0);
4221			T		x	= T(1.0f);
4222			mat[0][0] = x;
4223			for (int ndx = 0; ndx < Columns; ++ndx)
4224			{
4225				mat[Columns-1-ndx][ndx] = x;
4226				x *= T(2.0f);
4227			}
4228			dst.push_back(mat);
4229		}
4230	}
4231
4232	double	getWeight	(void) const
4233	{
4234		return dePow(instance<DefaultSampling<T> >().getWeight(), Rows * Columns);
4235	}
4236};
4237
4238struct Context
4239{
4240	Context		(const string&		name_,
4241				 TestContext&		testContext_,
4242				 RenderContext&		renderContext_,
4243				 const FloatFormat&	floatFormat_,
4244				 const FloatFormat&	highpFormat_,
4245				 Precision			precision_,
4246				 ShaderType			shaderType_,
4247				 size_t				numRandoms_)
4248		: name				(name_)
4249		, testContext		(testContext_)
4250		, renderContext		(renderContext_)
4251		, floatFormat		(floatFormat_)
4252		, highpFormat		(highpFormat_)
4253		, precision			(precision_)
4254		, shaderType		(shaderType_)
4255		, numRandoms		(numRandoms_) {}
4256
4257	string				name;
4258	TestContext&		testContext;
4259	RenderContext&		renderContext;
4260	FloatFormat			floatFormat;
4261	FloatFormat			highpFormat;
4262	Precision			precision;
4263	ShaderType			shaderType;
4264	size_t				numRandoms;
4265};
4266
4267template<typename In0_ = Void, typename In1_ = Void, typename In2_ = Void, typename In3_ = Void>
4268struct InTypes
4269{
4270	typedef	In0_	In0;
4271	typedef	In1_	In1;
4272	typedef	In2_	In2;
4273	typedef	In3_	In3;
4274};
4275
4276template <typename In>
4277int numInputs (void)
4278{
4279	return (!isTypeValid<typename In::In0>() ? 0 :
4280			!isTypeValid<typename In::In1>() ? 1 :
4281			!isTypeValid<typename In::In2>() ? 2 :
4282			!isTypeValid<typename In::In3>() ? 3 :
4283			4);
4284}
4285
4286template<typename Out0_, typename Out1_ = Void>
4287struct OutTypes
4288{
4289	typedef	Out0_	Out0;
4290	typedef	Out1_	Out1;
4291};
4292
4293template <typename Out>
4294int numOutputs (void)
4295{
4296	return (!isTypeValid<typename Out::Out0>() ? 0 :
4297			!isTypeValid<typename Out::Out1>() ? 1 :
4298			2);
4299}
4300
4301template<typename In>
4302struct Inputs
4303{
4304	vector<typename In::In0>	in0;
4305	vector<typename In::In1>	in1;
4306	vector<typename In::In2>	in2;
4307	vector<typename In::In3>	in3;
4308};
4309
4310template<typename Out>
4311struct Outputs
4312{
4313	Outputs	(size_t size) : out0(size), out1(size) {}
4314
4315	vector<typename Out::Out0>	out0;
4316	vector<typename Out::Out1>	out1;
4317};
4318
4319template<typename In, typename Out>
4320struct Variables
4321{
4322	VariableP<typename In::In0>		in0;
4323	VariableP<typename In::In1>		in1;
4324	VariableP<typename In::In2>		in2;
4325	VariableP<typename In::In3>		in3;
4326	VariableP<typename Out::Out0>	out0;
4327	VariableP<typename Out::Out1>	out1;
4328};
4329
4330template<typename In>
4331struct Samplings
4332{
4333	Samplings	(const Sampling<typename In::In0>&	in0_,
4334				 const Sampling<typename In::In1>&	in1_,
4335				 const Sampling<typename In::In2>&	in2_,
4336				 const Sampling<typename In::In3>&	in3_)
4337		: in0 (in0_), in1 (in1_), in2 (in2_), in3 (in3_) {}
4338
4339	const Sampling<typename In::In0>&	in0;
4340	const Sampling<typename In::In1>&	in1;
4341	const Sampling<typename In::In2>&	in2;
4342	const Sampling<typename In::In3>&	in3;
4343};
4344
4345template<typename In>
4346struct DefaultSamplings : Samplings<In>
4347{
4348	DefaultSamplings	(void)
4349		: Samplings<In>(instance<DefaultSampling<typename In::In0> >(),
4350						instance<DefaultSampling<typename In::In1> >(),
4351						instance<DefaultSampling<typename In::In2> >(),
4352						instance<DefaultSampling<typename In::In3> >()) {}
4353};
4354
4355class PrecisionCase : public TestCase
4356{
4357public:
4358	IterateResult		iterate			(void);
4359
4360protected:
4361						PrecisionCase	(const Context&		context,
4362										 const string&		name,
4363										 const string&		extension	= "")
4364							: TestCase		(context.testContext,
4365											 name.c_str(),
4366											 name.c_str())
4367							, m_ctx			(context)
4368							, m_status		()
4369							, m_rnd			(0xdeadbeefu +
4370											 context.testContext.getCommandLine().getBaseSeed())
4371							, m_extension	(extension)
4372	{
4373	}
4374
4375	RenderContext&		getRenderContext(void) const 			{ return m_ctx.renderContext; }
4376
4377	const FloatFormat&	getFormat		(void) const 			{ return m_ctx.floatFormat; }
4378
4379	TestLog&			log				(void) const 			{ return m_testCtx.getLog(); }
4380
4381	virtual void		runTest			(void) = 0;
4382
4383	template <typename In, typename Out>
4384	void				testStatement	(const Variables<In, Out>&	variables,
4385										 const Inputs<In>&			inputs,
4386										 const Statement&			stmt);
4387
4388	template<typename T>
4389	Symbol				makeSymbol		(const Variable<T>& variable)
4390	{
4391		return Symbol(variable.getName(), getVarTypeOf<T>(m_ctx.precision));
4392	}
4393
4394	Context				m_ctx;
4395	ResultCollector		m_status;
4396	Random				m_rnd;
4397	const string		m_extension;
4398};
4399
4400IterateResult PrecisionCase::iterate (void)
4401{
4402	runTest();
4403	m_status.setTestContextResult(m_testCtx);
4404	return STOP;
4405}
4406
4407template <typename In, typename Out>
4408void PrecisionCase::testStatement (const Variables<In, Out>&	variables,
4409								   const Inputs<In>&			inputs,
4410								   const Statement&				stmt)
4411{
4412	using namespace ShaderExecUtil;
4413
4414	typedef typename 	In::In0		In0;
4415	typedef typename 	In::In1		In1;
4416	typedef typename 	In::In2		In2;
4417	typedef typename 	In::In3		In3;
4418	typedef typename 	Out::Out0	Out0;
4419	typedef typename 	Out::Out1	Out1;
4420
4421	const FloatFormat&	fmt			= getFormat();
4422	const int			inCount		= numInputs<In>();
4423	const int			outCount	= numOutputs<Out>();
4424	const size_t		numValues	= (inCount > 0) ? inputs.in0.size() : 1;
4425	Outputs<Out>		outputs		(numValues);
4426	ShaderSpec			spec;
4427	const FloatFormat	highpFmt	= m_ctx.highpFormat;
4428	const int			maxMsgs		= 100;
4429	int					numErrors	= 0;
4430	Environment			env; 		// Hoisted out of the inner loop for optimization.
4431
4432	switch (inCount)
4433	{
4434		case 4: DE_ASSERT(inputs.in3.size() == numValues);
4435		case 3: DE_ASSERT(inputs.in2.size() == numValues);
4436		case 2: DE_ASSERT(inputs.in1.size() == numValues);
4437		case 1: DE_ASSERT(inputs.in0.size() == numValues);
4438		default: break;
4439	}
4440
4441	// Print out the statement and its definitions
4442	log() << TestLog::Message << "Statement: " << stmt << TestLog::EndMessage;
4443	{
4444		ostringstream	oss;
4445		FuncSet			funcs;
4446
4447		stmt.getUsedFuncs(funcs);
4448		for (FuncSet::const_iterator it = funcs.begin(); it != funcs.end(); ++it)
4449		{
4450			(*it)->printDefinition(oss);
4451		}
4452		if (!funcs.empty())
4453			log() << TestLog::Message << "Reference definitions:\n" << oss.str()
4454				  << TestLog::EndMessage;
4455	}
4456
4457	// Initialize ShaderSpec from precision, variables and statement.
4458	{
4459		ostringstream os;
4460		os << "precision " << glu::getPrecisionName(m_ctx.precision) << " float;\n";
4461		spec.globalDeclarations = os.str();
4462	}
4463
4464	spec.version = getContextTypeGLSLVersion(getRenderContext().getType());
4465
4466	if (!m_extension.empty())
4467		spec.globalDeclarations = "#extension " + m_extension + " : require\n";
4468
4469	spec.inputs.resize(inCount);
4470
4471	switch (inCount)
4472	{
4473		case 4: spec.inputs[3] = makeSymbol(*variables.in3);
4474		case 3:	spec.inputs[2] = makeSymbol(*variables.in2);
4475		case 2:	spec.inputs[1] = makeSymbol(*variables.in1);
4476		case 1:	spec.inputs[0] = makeSymbol(*variables.in0);
4477		default: break;
4478	}
4479
4480	spec.outputs.resize(outCount);
4481
4482	switch (outCount)
4483	{
4484		case 2:	spec.outputs[1] = makeSymbol(*variables.out1);
4485		case 1:	spec.outputs[0] = makeSymbol(*variables.out0);
4486		default: break;
4487	}
4488
4489	spec.source = de::toString(stmt);
4490
4491	// Run the shader with inputs.
4492	{
4493		UniquePtr<ShaderExecutor>	executor		(createExecutor(getRenderContext(),
4494																	m_ctx.shaderType,
4495																	spec));
4496		const void*					inputArr[]		=
4497		{
4498			&inputs.in0.front(), &inputs.in1.front(), &inputs.in2.front(), &inputs.in3.front(),
4499		};
4500		void*						outputArr[]		=
4501		{
4502			&outputs.out0.front(), &outputs.out1.front(),
4503		};
4504
4505		executor->log(log());
4506		if (!executor->isOk())
4507			TCU_FAIL("Shader compilation failed");
4508
4509		executor->useProgram();
4510		executor->execute(int(numValues), inputArr, outputArr);
4511	}
4512
4513	// Initialize environment with dummy values so we don't need to bind in inner loop.
4514	{
4515		const typename Traits<In0>::IVal		in0;
4516		const typename Traits<In1>::IVal		in1;
4517		const typename Traits<In2>::IVal		in2;
4518		const typename Traits<In3>::IVal		in3;
4519		const typename Traits<Out0>::IVal		reference0;
4520		const typename Traits<Out1>::IVal		reference1;
4521
4522		env.bind(*variables.in0, in0);
4523		env.bind(*variables.in1, in1);
4524		env.bind(*variables.in2, in2);
4525		env.bind(*variables.in3, in3);
4526		env.bind(*variables.out0, reference0);
4527		env.bind(*variables.out1, reference1);
4528	}
4529
4530	// For each input tuple, compute output reference interval and compare
4531	// shader output to the reference.
4532	for (size_t valueNdx = 0; valueNdx < numValues; valueNdx++)
4533	{
4534		bool						result		= true;
4535		typename Traits<Out0>::IVal	reference0;
4536		typename Traits<Out1>::IVal	reference1;
4537
4538		env.lookup(*variables.in0) = convert<In0>(fmt, round(fmt, inputs.in0[valueNdx]));
4539		env.lookup(*variables.in1) = convert<In1>(fmt, round(fmt, inputs.in1[valueNdx]));
4540		env.lookup(*variables.in2) = convert<In2>(fmt, round(fmt, inputs.in2[valueNdx]));
4541		env.lookup(*variables.in3) = convert<In3>(fmt, round(fmt, inputs.in3[valueNdx]));
4542
4543		{
4544			EvalContext	ctx (fmt, m_ctx.precision, env);
4545			stmt.execute(ctx);
4546		}
4547
4548		switch (outCount)
4549		{
4550			case 2:
4551				reference1 = convert<Out1>(highpFmt, env.lookup(*variables.out1));
4552				if (!m_status.check(contains(reference1, outputs.out1[valueNdx]),
4553									"Shader output 1 is outside acceptable range"))
4554					result = false;
4555			case 1:
4556				reference0 = convert<Out0>(highpFmt, env.lookup(*variables.out0));
4557				if (!m_status.check(contains(reference0, outputs.out0[valueNdx]),
4558									"Shader output 0 is outside acceptable range"))
4559					result = false;
4560			default: break;
4561		}
4562
4563		if (!result)
4564			++numErrors;
4565
4566		if ((!result && numErrors <= maxMsgs) || GLS_LOG_ALL_RESULTS)
4567		{
4568			MessageBuilder	builder	= log().message();
4569
4570			builder << (result ? "Passed" : "Failed") << " sample:\n";
4571
4572			if (inCount > 0)
4573			{
4574				builder << "\t" << variables.in0->getName() << " = "
4575						<< valueToString(highpFmt, inputs.in0[valueNdx]) << "\n";
4576			}
4577
4578			if (inCount > 1)
4579			{
4580				builder << "\t" << variables.in1->getName() << " = "
4581						<< valueToString(highpFmt, inputs.in1[valueNdx]) << "\n";
4582			}
4583
4584			if (inCount > 2)
4585			{
4586				builder << "\t" << variables.in2->getName() << " = "
4587						<< valueToString(highpFmt, inputs.in2[valueNdx]) << "\n";
4588			}
4589
4590			if (inCount > 3)
4591			{
4592				builder << "\t" << variables.in3->getName() << " = "
4593						<< valueToString(highpFmt, inputs.in3[valueNdx]) << "\n";
4594			}
4595
4596			if (outCount > 0)
4597			{
4598				builder << "\t" << variables.out0->getName() << " = "
4599						<< valueToString(highpFmt, outputs.out0[valueNdx]) << "\n"
4600						<< "\tExpected range: "
4601						<< intervalToString<typename Out::Out0>(highpFmt, reference0) << "\n";
4602			}
4603
4604			if (outCount > 1)
4605			{
4606				builder << "\t" << variables.out1->getName() << " = "
4607						<< valueToString(highpFmt, outputs.out1[valueNdx]) << "\n"
4608						<< "\tExpected range: "
4609						<< intervalToString<typename Out::Out1>(highpFmt, reference1) << "\n";
4610			}
4611
4612			builder << TestLog::EndMessage;
4613		}
4614	}
4615
4616	if (numErrors > maxMsgs)
4617	{
4618		log() << TestLog::Message << "(Skipped " << (numErrors - maxMsgs) << " messages.)"
4619			  << TestLog::EndMessage;
4620	}
4621
4622	if (numErrors == 0)
4623	{
4624		log() << TestLog::Message << "All " << numValues << " inputs passed."
4625			  << TestLog::EndMessage;
4626	}
4627	else
4628	{
4629		log() << TestLog::Message << numErrors << "/" << numValues << " inputs failed."
4630			  << TestLog::EndMessage;
4631	}
4632}
4633
4634
4635
4636template <typename T>
4637struct InputLess
4638{
4639	bool operator() (const T& val1, const T& val2) const
4640	{
4641		return val1 < val2;
4642	}
4643};
4644
4645template <typename T>
4646bool inputLess (const T& val1, const T& val2)
4647{
4648	return InputLess<T>()(val1, val2);
4649}
4650
4651template <>
4652struct InputLess<float>
4653{
4654	bool operator() (const float& val1, const float& val2) const
4655	{
4656		if (deIsNaN(val1))
4657			return false;
4658		if (deIsNaN(val2))
4659			return true;
4660		return val1 < val2;
4661	}
4662};
4663
4664template <typename T, int Size>
4665struct InputLess<Vector<T, Size> >
4666{
4667	bool operator() (const Vector<T, Size>& vec1, const Vector<T, Size>& vec2) const
4668	{
4669		for (int ndx = 0; ndx < Size; ++ndx)
4670		{
4671			if (inputLess(vec1[ndx], vec2[ndx]))
4672				return true;
4673			if (inputLess(vec2[ndx], vec1[ndx]))
4674				return false;
4675		}
4676
4677		return false;
4678	}
4679};
4680
4681template <typename T, int Rows, int Cols>
4682struct InputLess<Matrix<T, Rows, Cols> >
4683{
4684	bool operator() (const Matrix<T, Rows, Cols>& mat1,
4685					 const Matrix<T, Rows, Cols>& mat2) const
4686	{
4687		for (int col = 0; col < Cols; ++col)
4688		{
4689			if (inputLess(mat1[col], mat2[col]))
4690				return true;
4691			if (inputLess(mat2[col], mat1[col]))
4692				return false;
4693		}
4694
4695		return false;
4696	}
4697};
4698
4699template <typename In>
4700struct InTuple :
4701	public Tuple4<typename In::In0, typename In::In1, typename In::In2, typename In::In3>
4702{
4703	InTuple	(const typename In::In0& in0,
4704			 const typename In::In1& in1,
4705			 const typename In::In2& in2,
4706			 const typename In::In3& in3)
4707		: Tuple4<typename In::In0, typename In::In1, typename In::In2, typename In::In3>
4708		  (in0, in1, in2, in3) {}
4709};
4710
4711template <typename In>
4712struct InputLess<InTuple<In> >
4713{
4714	bool operator() (const InTuple<In>& in1, const InTuple<In>& in2) const
4715	{
4716		if (inputLess(in1.a, in2.a))
4717			return true;
4718		if (inputLess(in2.a, in1.a))
4719			return false;
4720		if (inputLess(in1.b, in2.b))
4721			return true;
4722		if (inputLess(in2.b, in1.b))
4723			return false;
4724		if (inputLess(in1.c, in2.c))
4725			return true;
4726		if (inputLess(in2.c, in1.c))
4727			return false;
4728		if (inputLess(in1.d, in2.d))
4729			return true;
4730		return false;
4731	};
4732};
4733
4734template<typename In>
4735Inputs<In> generateInputs (const Samplings<In>&	samplings,
4736						   const FloatFormat&	floatFormat,
4737						   Precision			intPrecision,
4738						   size_t				numSamples,
4739						   Random&				rnd)
4740{
4741	Inputs<In>									ret;
4742	Inputs<In>									fixedInputs;
4743	set<InTuple<In>, InputLess<InTuple<In> > >	seenInputs;
4744
4745	samplings.in0.genFixeds(floatFormat, fixedInputs.in0);
4746	samplings.in1.genFixeds(floatFormat, fixedInputs.in1);
4747	samplings.in2.genFixeds(floatFormat, fixedInputs.in2);
4748	samplings.in3.genFixeds(floatFormat, fixedInputs.in3);
4749
4750	for (size_t ndx0 = 0; ndx0 < fixedInputs.in0.size(); ++ndx0)
4751	{
4752		for (size_t ndx1 = 0; ndx1 < fixedInputs.in1.size(); ++ndx1)
4753		{
4754			for (size_t ndx2 = 0; ndx2 < fixedInputs.in2.size(); ++ndx2)
4755			{
4756				for (size_t ndx3 = 0; ndx3 < fixedInputs.in3.size(); ++ndx3)
4757				{
4758					const InTuple<In>	tuple	(fixedInputs.in0[ndx0],
4759												 fixedInputs.in1[ndx1],
4760												 fixedInputs.in2[ndx2],
4761												 fixedInputs.in3[ndx3]);
4762
4763					seenInputs.insert(tuple);
4764					ret.in0.push_back(tuple.a);
4765					ret.in1.push_back(tuple.b);
4766					ret.in2.push_back(tuple.c);
4767					ret.in3.push_back(tuple.d);
4768				}
4769			}
4770		}
4771	}
4772
4773	for (size_t ndx = 0; ndx < numSamples; ++ndx)
4774	{
4775		const typename In::In0	in0		= samplings.in0.genRandom(floatFormat, intPrecision, rnd);
4776		const typename In::In1	in1		= samplings.in1.genRandom(floatFormat, intPrecision, rnd);
4777		const typename In::In2	in2		= samplings.in2.genRandom(floatFormat, intPrecision, rnd);
4778		const typename In::In3	in3		= samplings.in3.genRandom(floatFormat, intPrecision, rnd);
4779		const InTuple<In>		tuple	(in0, in1, in2, in3);
4780
4781		if (de::contains(seenInputs, tuple))
4782			continue;
4783
4784		seenInputs.insert(tuple);
4785		ret.in0.push_back(in0);
4786		ret.in1.push_back(in1);
4787		ret.in2.push_back(in2);
4788		ret.in3.push_back(in3);
4789	}
4790
4791	return ret;
4792}
4793
4794class FuncCaseBase : public PrecisionCase
4795{
4796public:
4797	IterateResult	iterate			(void);
4798
4799protected:
4800					FuncCaseBase	(const Context&		context,
4801									 const string&		name,
4802									 const FuncBase&	func)
4803						: PrecisionCase	(context, name, func.getRequiredExtension()) {}
4804};
4805
4806IterateResult FuncCaseBase::iterate (void)
4807{
4808	MovePtr<ContextInfo>	info	(ContextInfo::create(getRenderContext()));
4809
4810	if (!m_extension.empty() && !info->isExtensionSupported(m_extension.c_str()))
4811		throw NotSupportedError("Unsupported extension: " + m_extension);
4812
4813	runTest();
4814
4815	m_status.setTestContextResult(m_testCtx);
4816	return STOP;
4817}
4818
4819template <typename Sig>
4820class FuncCase : public FuncCaseBase
4821{
4822public:
4823	typedef Func<Sig>						CaseFunc;
4824	typedef typename Sig::Ret				Ret;
4825	typedef typename Sig::Arg0				Arg0;
4826	typedef typename Sig::Arg1				Arg1;
4827	typedef typename Sig::Arg2				Arg2;
4828	typedef typename Sig::Arg3				Arg3;
4829	typedef InTypes<Arg0, Arg1, Arg2, Arg3>	In;
4830	typedef OutTypes<Ret>					Out;
4831
4832					FuncCase	(const Context&		context,
4833								 const string&		name,
4834								 const CaseFunc&	func)
4835						: FuncCaseBase	(context, name, func)
4836						, m_func		(func) {}
4837
4838protected:
4839	void				runTest		(void);
4840
4841	virtual const Samplings<In>&	getSamplings	(void)
4842	{
4843		return instance<DefaultSamplings<In> >();
4844	}
4845
4846private:
4847	const CaseFunc&			m_func;
4848};
4849
4850template <typename Sig>
4851void FuncCase<Sig>::runTest (void)
4852{
4853	const Inputs<In>	inputs	(generateInputs(getSamplings(),
4854												m_ctx.floatFormat,
4855												m_ctx.precision,
4856												m_ctx.numRandoms,
4857												m_rnd));
4858	Variables<In, Out>	variables;
4859
4860	variables.out0	= variable<Ret>("out0");
4861	variables.out1	= variable<Void>("out1");
4862	variables.in0	= variable<Arg0>("in0");
4863	variables.in1	= variable<Arg1>("in1");
4864	variables.in2	= variable<Arg2>("in2");
4865	variables.in3	= variable<Arg3>("in3");
4866
4867	{
4868		ExprP<Ret>	expr	= applyVar(m_func,
4869									   variables.in0, variables.in1,
4870									   variables.in2, variables.in3);
4871		StatementP	stmt	= variableAssignment(variables.out0, expr);
4872
4873		this->testStatement(variables, inputs, *stmt);
4874	}
4875}
4876
4877template <typename Sig>
4878class InOutFuncCase : public FuncCaseBase
4879{
4880public:
4881	typedef Func<Sig>						CaseFunc;
4882	typedef typename Sig::Ret				Ret;
4883	typedef typename Sig::Arg0				Arg0;
4884	typedef typename Sig::Arg1				Arg1;
4885	typedef typename Sig::Arg2				Arg2;
4886	typedef typename Sig::Arg3				Arg3;
4887	typedef InTypes<Arg0, Arg2, Arg3>		In;
4888	typedef OutTypes<Ret, Arg1>				Out;
4889
4890					InOutFuncCase	(const Context&		context,
4891									 const string&		name,
4892									 const CaseFunc&	func)
4893						: FuncCaseBase	(context, name, func)
4894						, m_func		(func) {}
4895
4896protected:
4897	void				runTest		(void);
4898
4899	virtual const Samplings<In>&	getSamplings	(void)
4900	{
4901		return instance<DefaultSamplings<In> >();
4902	}
4903
4904private:
4905	const CaseFunc&			m_func;
4906};
4907
4908template <typename Sig>
4909void InOutFuncCase<Sig>::runTest (void)
4910{
4911	const Inputs<In>	inputs	(generateInputs(getSamplings(),
4912												m_ctx.floatFormat,
4913												m_ctx.precision,
4914												m_ctx.numRandoms,
4915												m_rnd));
4916	Variables<In, Out>	variables;
4917
4918	variables.out0	= variable<Ret>("out0");
4919	variables.out1	= variable<Arg1>("out1");
4920	variables.in0	= variable<Arg0>("in0");
4921	variables.in1	= variable<Arg2>("in1");
4922	variables.in2	= variable<Arg3>("in2");
4923	variables.in3	= variable<Void>("in3");
4924
4925	{
4926		ExprP<Ret>	expr	= applyVar(m_func,
4927									   variables.in0, variables.out1,
4928									   variables.in1, variables.in2);
4929		StatementP	stmt	= variableAssignment(variables.out0, expr);
4930
4931		this->testStatement(variables, inputs, *stmt);
4932	}
4933}
4934
4935template <typename Sig>
4936PrecisionCase* createFuncCase (const Context&	context,
4937							   const string&	name,
4938							   const Func<Sig>&	func)
4939{
4940	switch (func.getOutParamIndex())
4941	{
4942		case -1:
4943			return new FuncCase<Sig>(context, name, func);
4944		case 1:
4945			return new InOutFuncCase<Sig>(context, name, func);
4946		default:
4947			DE_FATAL("Impossible");
4948	}
4949	return DE_NULL;
4950}
4951
4952class CaseFactory
4953{
4954public:
4955	virtual						~CaseFactory	(void) {}
4956	virtual MovePtr<TestNode>	createCase		(const Context& ctx) const = 0;
4957	virtual string				getName			(void) const = 0;
4958	virtual string				getDesc			(void) const = 0;
4959};
4960
4961class FuncCaseFactory : public CaseFactory
4962{
4963public:
4964	virtual const FuncBase&	getFunc		(void) const = 0;
4965
4966	string					getName		(void) const
4967	{
4968		return de::toLower(getFunc().getName());
4969	}
4970
4971	string					getDesc		(void) const
4972	{
4973		return "Function '" + getFunc().getName() + "'";
4974	}
4975};
4976
4977template <typename Sig>
4978class GenFuncCaseFactory : public CaseFactory
4979{
4980public:
4981
4982						GenFuncCaseFactory	(const GenFuncs<Sig>&	funcs,
4983											 const string&			name)
4984							: m_funcs	(funcs)
4985							, m_name	(de::toLower(name)) {}
4986
4987	MovePtr<TestNode>	createCase			(const Context& ctx) const
4988	{
4989		TestCaseGroup*	group = new TestCaseGroup(ctx.testContext,
4990												  ctx.name.c_str(), ctx.name.c_str());
4991
4992		group->addChild(createFuncCase(ctx, "scalar", m_funcs.func));
4993		group->addChild(createFuncCase(ctx, "vec2", m_funcs.func2));
4994		group->addChild(createFuncCase(ctx, "vec3", m_funcs.func3));
4995		group->addChild(createFuncCase(ctx, "vec4", m_funcs.func4));
4996
4997		return MovePtr<TestNode>(group);
4998	}
4999
5000	string				getName				(void) const
5001	{
5002		return m_name;
5003	}
5004
5005	string				getDesc				(void) const
5006	{
5007		return "Function '" + m_funcs.func.getName() + "'";
5008	}
5009
5010private:
5011	const GenFuncs<Sig>	m_funcs;
5012	string				m_name;
5013};
5014
5015template <template <int> class GenF>
5016class TemplateFuncCaseFactory : public FuncCaseFactory
5017{
5018public:
5019	MovePtr<TestNode>	createCase		(const Context& ctx) const
5020	{
5021		TestCaseGroup*	group = new TestCaseGroup(ctx.testContext,
5022							  ctx.name.c_str(), ctx.name.c_str());
5023		group->addChild(createFuncCase(ctx, "scalar", instance<GenF<1> >()));
5024		group->addChild(createFuncCase(ctx, "vec2", instance<GenF<2> >()));
5025		group->addChild(createFuncCase(ctx, "vec3", instance<GenF<3> >()));
5026		group->addChild(createFuncCase(ctx, "vec4", instance<GenF<4> >()));
5027
5028		return MovePtr<TestNode>(group);
5029	}
5030
5031	const FuncBase&		getFunc			(void) const { return instance<GenF<1> >(); }
5032};
5033
5034template <template <int> class GenF>
5035class SquareMatrixFuncCaseFactory : public FuncCaseFactory
5036{
5037public:
5038	MovePtr<TestNode>	createCase		(const Context& ctx) const
5039	{
5040		TestCaseGroup*	group = new TestCaseGroup(ctx.testContext,
5041							  ctx.name.c_str(), ctx.name.c_str());
5042		group->addChild(createFuncCase(ctx, "mat2", instance<GenF<2> >()));
5043#if 0
5044		// disabled until we get reasonable results
5045		group->addChild(createFuncCase(ctx, "mat3", instance<GenF<3> >()));
5046		group->addChild(createFuncCase(ctx, "mat4", instance<GenF<4> >()));
5047#endif
5048
5049		return MovePtr<TestNode>(group);
5050	}
5051
5052	const FuncBase&		getFunc			(void) const { return instance<GenF<2> >(); }
5053};
5054
5055template <template <int, int> class GenF>
5056class MatrixFuncCaseFactory : public FuncCaseFactory
5057{
5058public:
5059	MovePtr<TestNode>	createCase		(const Context& ctx) const
5060	{
5061		TestCaseGroup*	const group = new TestCaseGroup(ctx.testContext,
5062														ctx.name.c_str(), ctx.name.c_str());
5063
5064		this->addCase<2, 2>(ctx, group);
5065		this->addCase<3, 2>(ctx, group);
5066		this->addCase<4, 2>(ctx, group);
5067		this->addCase<2, 3>(ctx, group);
5068		this->addCase<3, 3>(ctx, group);
5069		this->addCase<4, 3>(ctx, group);
5070		this->addCase<2, 4>(ctx, group);
5071		this->addCase<3, 4>(ctx, group);
5072		this->addCase<4, 4>(ctx, group);
5073
5074		return MovePtr<TestNode>(group);
5075	}
5076
5077	const FuncBase&		getFunc			(void) const { return instance<GenF<2,2> >(); }
5078
5079private:
5080	template <int Rows, int Cols>
5081	void				addCase			(const Context& ctx, TestCaseGroup* group) const
5082	{
5083		const char*	const name = dataTypeNameOf<Matrix<float, Rows, Cols> >();
5084
5085		group->addChild(createFuncCase(ctx, name, instance<GenF<Rows, Cols> >()));
5086	}
5087};
5088
5089template <typename Sig>
5090class SimpleFuncCaseFactory : public CaseFactory
5091{
5092public:
5093						SimpleFuncCaseFactory	(const Func<Sig>& func) : m_func(func) {}
5094
5095	MovePtr<TestNode>	createCase		(const Context& ctx) const
5096	{
5097		return MovePtr<TestNode>(createFuncCase(ctx, ctx.name.c_str(), m_func));
5098	}
5099
5100	string				getName					(void) const
5101	{
5102		return de::toLower(m_func.getName());
5103	}
5104
5105	string				getDesc					(void) const
5106	{
5107		return "Function '" + getName() + "'";
5108	}
5109
5110private:
5111	const Func<Sig>&	m_func;
5112};
5113
5114template <typename F>
5115SharedPtr<SimpleFuncCaseFactory<typename F::Sig> > createSimpleFuncCaseFactory (void)
5116{
5117	return SharedPtr<SimpleFuncCaseFactory<typename F::Sig> >(
5118		new SimpleFuncCaseFactory<typename F::Sig>(instance<F>()));
5119}
5120
5121class BuiltinFuncs : public CaseFactories
5122{
5123public:
5124	const vector<const CaseFactory*>	getFactories	(void) const
5125	{
5126		vector<const CaseFactory*> ret;
5127
5128		for (size_t ndx = 0; ndx < m_factories.size(); ++ndx)
5129			ret.push_back(m_factories[ndx].get());
5130
5131		return ret;
5132	}
5133
5134	void								addFactory		(SharedPtr<const CaseFactory> fact)
5135	{
5136		m_factories.push_back(fact);
5137	}
5138
5139private:
5140	vector<SharedPtr<const CaseFactory> >			m_factories;
5141};
5142
5143template <typename F>
5144void addScalarFactory(BuiltinFuncs& funcs, string name = "")
5145{
5146	if (name.empty())
5147		name = instance<F>().getName();
5148
5149	funcs.addFactory(SharedPtr<const CaseFactory>(new GenFuncCaseFactory<typename F::Sig>(
5150													  makeVectorizedFuncs<F>(), name)));
5151}
5152
5153MovePtr<const CaseFactories> createES3BuiltinCases (void)
5154{
5155	MovePtr<BuiltinFuncs>	funcs	(new BuiltinFuncs());
5156
5157	addScalarFactory<Add>(*funcs);
5158	addScalarFactory<Sub>(*funcs);
5159	addScalarFactory<Mul>(*funcs);
5160	addScalarFactory<Div>(*funcs);
5161
5162	addScalarFactory<Radians>(*funcs);
5163	addScalarFactory<Degrees>(*funcs);
5164	addScalarFactory<Sin>(*funcs);
5165	addScalarFactory<Cos>(*funcs);
5166	addScalarFactory<Tan>(*funcs);
5167	addScalarFactory<ASin>(*funcs);
5168	addScalarFactory<ACos>(*funcs);
5169	addScalarFactory<ATan2>(*funcs, "atan2");
5170	addScalarFactory<ATan>(*funcs);
5171	addScalarFactory<Sinh>(*funcs);
5172	addScalarFactory<Cosh>(*funcs);
5173	addScalarFactory<Tanh>(*funcs);
5174	addScalarFactory<ASinh>(*funcs);
5175	addScalarFactory<ACosh>(*funcs);
5176	addScalarFactory<ATanh>(*funcs);
5177
5178	addScalarFactory<Pow>(*funcs);
5179	addScalarFactory<Exp>(*funcs);
5180	addScalarFactory<Log>(*funcs);
5181	addScalarFactory<Exp2>(*funcs);
5182	addScalarFactory<Log2>(*funcs);
5183	addScalarFactory<Sqrt>(*funcs);
5184	addScalarFactory<InverseSqrt>(*funcs);
5185
5186	addScalarFactory<Abs>(*funcs);
5187	addScalarFactory<Sign>(*funcs);
5188	addScalarFactory<Floor>(*funcs);
5189	addScalarFactory<Trunc>(*funcs);
5190	addScalarFactory<Round>(*funcs);
5191	addScalarFactory<RoundEven>(*funcs);
5192	addScalarFactory<Ceil>(*funcs);
5193	addScalarFactory<Fract>(*funcs);
5194	addScalarFactory<Mod>(*funcs);
5195	funcs->addFactory(createSimpleFuncCaseFactory<Modf>());
5196	addScalarFactory<Min>(*funcs);
5197	addScalarFactory<Max>(*funcs);
5198	addScalarFactory<Clamp>(*funcs);
5199	addScalarFactory<Mix>(*funcs);
5200	addScalarFactory<Step>(*funcs);
5201	addScalarFactory<SmoothStep>(*funcs);
5202
5203	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Length>()));
5204	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Distance>()));
5205	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Dot>()));
5206	funcs->addFactory(createSimpleFuncCaseFactory<Cross>());
5207	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Normalize>()));
5208	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<FaceForward>()));
5209	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Reflect>()));
5210	funcs->addFactory(SharedPtr<const CaseFactory>(new TemplateFuncCaseFactory<Refract>()));
5211
5212
5213	funcs->addFactory(SharedPtr<const CaseFactory>(new MatrixFuncCaseFactory<MatrixCompMult>()));
5214	funcs->addFactory(SharedPtr<const CaseFactory>(new MatrixFuncCaseFactory<OuterProduct>()));
5215	funcs->addFactory(SharedPtr<const CaseFactory>(new MatrixFuncCaseFactory<Transpose>()));
5216	funcs->addFactory(SharedPtr<const CaseFactory>(new SquareMatrixFuncCaseFactory<Determinant>()));
5217	funcs->addFactory(SharedPtr<const CaseFactory>(new SquareMatrixFuncCaseFactory<Inverse>()));
5218
5219	return MovePtr<const CaseFactories>(funcs.release());
5220}
5221
5222MovePtr<const CaseFactories> createES31BuiltinCases (void)
5223{
5224	MovePtr<BuiltinFuncs>	funcs	(new BuiltinFuncs());
5225
5226	addScalarFactory<FrExp>(*funcs);
5227	addScalarFactory<LdExp>(*funcs);
5228	addScalarFactory<Fma>(*funcs);
5229
5230	return MovePtr<const CaseFactories>(funcs.release());
5231}
5232
5233struct PrecisionTestContext
5234{
5235	PrecisionTestContext	(TestContext&				testCtx_,
5236							 RenderContext&				renderCtx_,
5237							 const FloatFormat&			highp_,
5238							 const FloatFormat&			mediump_,
5239							 const FloatFormat&			lowp_,
5240							 const vector<ShaderType>&	shaderTypes_,
5241							 int						numRandoms_)
5242		: testCtx		(testCtx_)
5243		, renderCtx		(renderCtx_)
5244		, shaderTypes	(shaderTypes_)
5245		, numRandoms	(numRandoms_)
5246	{
5247		formats[glu::PRECISION_HIGHP]	= &highp_;
5248		formats[glu::PRECISION_MEDIUMP]	= &mediump_;
5249		formats[glu::PRECISION_LOWP]	= &lowp_;
5250	}
5251
5252	TestContext&			testCtx;
5253	RenderContext&			renderCtx;
5254	const FloatFormat*		formats[glu::PRECISION_LAST];
5255	vector<ShaderType>		shaderTypes;
5256	int						numRandoms;
5257};
5258
5259TestCaseGroup* createFuncGroup (const PrecisionTestContext&	ctx,
5260								const CaseFactory&			factory)
5261{
5262	TestCaseGroup* const 	group	= new TestCaseGroup(ctx.testCtx,
5263														factory.getName().c_str(),
5264														factory.getDesc().c_str());
5265
5266	for (int precNdx = 0; precNdx < glu::PRECISION_LAST; ++precNdx)
5267	{
5268		const Precision		precision	= Precision(precNdx);
5269		const string		precName	(glu::getPrecisionName(precision));
5270		const FloatFormat&	fmt			= *de::getSizedArrayElement<glu::PRECISION_LAST>(ctx.formats, precNdx);
5271		const FloatFormat&	highpFmt	= *de::getSizedArrayElement<glu::PRECISION_LAST>(ctx.formats,
5272																						 glu::PRECISION_HIGHP);
5273
5274		for (size_t shaderNdx = 0; shaderNdx < ctx.shaderTypes.size(); ++shaderNdx)
5275		{
5276			const ShaderType	shaderType	= ctx.shaderTypes[shaderNdx];
5277			const string 		shaderName	(glu::getShaderTypeName(shaderType));
5278			const string		name		= precName + "_" + shaderName;
5279			const Context		caseCtx		(name, ctx.testCtx, ctx.renderCtx, fmt, highpFmt,
5280											 precision, shaderType, ctx.numRandoms);
5281
5282			group->addChild(factory.createCase(caseCtx).release());
5283		}
5284	}
5285
5286	return group;
5287}
5288
5289void addBuiltinPrecisionTests (TestContext&					testCtx,
5290							   RenderContext&				renderCtx,
5291							   const CaseFactories&			cases,
5292							   const vector<ShaderType>&	shaderTypes,
5293							   TestCaseGroup&				dstGroup)
5294{
5295	const int						userRandoms	= testCtx.getCommandLine().getTestIterationCount();
5296	const int						defRandoms	= 16384;
5297	const int						numRandoms	= userRandoms > 0 ? userRandoms : defRandoms;
5298	const FloatFormat				highp		(-126, 127, 23, true,
5299												 tcu::MAYBE,	// subnormals
5300												 tcu::YES,		// infinities
5301												 tcu::MAYBE);	// NaN
5302	// \todo [2014-04-01 lauri] Check these once Khronos bug 11840 is resolved.
5303	const FloatFormat				mediump		(-13, 13, 9, false);
5304	// A fixed-point format is just a floating point format with a fixed
5305	// exponent and support for subnormals.
5306	const FloatFormat				lowp		(0, 0, 7, false, tcu::YES);
5307	const PrecisionTestContext		ctx			(testCtx, renderCtx, highp, mediump, lowp,
5308												 shaderTypes, numRandoms);
5309
5310	for (size_t ndx = 0; ndx < cases.getFactories().size(); ++ndx)
5311		dstGroup.addChild(createFuncGroup(ctx, *cases.getFactories()[ndx]));
5312}
5313
5314} // BuiltinPrecisionTests
5315} // gls
5316} // deqp
5317