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