1/*-------------------------------------------------------------------------
2 * drawElements Quality Program OpenGL ES 3.0 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 Shader derivate function tests.
22 *
23 * \todo [2013-06-25 pyry] Missing features:
24 *  - lines and points
25 *  - projected coordinates
26 *  - continous non-trivial functions (sin, exp)
27 *  - non-continous functions (step)
28 *//*--------------------------------------------------------------------*/
29
30#include "es3fShaderDerivateTests.hpp"
31#include "gluShaderProgram.hpp"
32#include "gluRenderContext.hpp"
33#include "gluDrawUtil.hpp"
34#include "gluPixelTransfer.hpp"
35#include "gluShaderUtil.hpp"
36#include "gluStrUtil.hpp"
37#include "gluTextureUtil.hpp"
38#include "gluTexture.hpp"
39#include "tcuStringTemplate.hpp"
40#include "tcuRenderTarget.hpp"
41#include "tcuSurface.hpp"
42#include "tcuTestLog.hpp"
43#include "tcuVectorUtil.hpp"
44#include "tcuTextureUtil.hpp"
45#include "tcuRGBA.hpp"
46#include "tcuFloat.hpp"
47#include "tcuInterval.hpp"
48#include "deRandom.hpp"
49#include "deUniquePtr.hpp"
50#include "deString.h"
51#include "glwEnums.hpp"
52#include "glwFunctions.hpp"
53#include "glsShaderRenderCase.hpp" // gls::setupDefaultUniforms()
54
55#include <sstream>
56
57namespace deqp
58{
59namespace gles3
60{
61namespace Functional
62{
63
64using std::vector;
65using std::string;
66using std::map;
67using tcu::TestLog;
68using std::ostringstream;
69
70enum
71{
72	VIEWPORT_WIDTH		= 167,
73	VIEWPORT_HEIGHT		= 103,
74	FBO_WIDTH			= 99,
75	FBO_HEIGHT			= 133,
76	MAX_FAILED_MESSAGES	= 10
77};
78
79enum DerivateFunc
80{
81	DERIVATE_DFDX	= 0,
82	DERIVATE_DFDY,
83	DERIVATE_FWIDTH,
84
85	DERIVATE_LAST
86};
87
88enum SurfaceType
89{
90	SURFACETYPE_DEFAULT_FRAMEBUFFER = 0,
91	SURFACETYPE_UNORM_FBO,
92	SURFACETYPE_FLOAT_FBO,	// \note Uses RGBA32UI fbo actually, since FP rendertargets are not in core spec.
93
94	SURFACETYPE_LAST
95};
96
97// Utilities
98
99namespace
100{
101
102class AutoFbo
103{
104public:
105	AutoFbo (const glw::Functions& gl)
106		: m_gl	(gl)
107		, m_fbo	(0)
108	{
109	}
110
111	~AutoFbo (void)
112	{
113		if (m_fbo)
114			m_gl.deleteFramebuffers(1, &m_fbo);
115	}
116
117	void gen (void)
118	{
119		DE_ASSERT(!m_fbo);
120		m_gl.genFramebuffers(1, &m_fbo);
121	}
122
123	deUint32 operator* (void) const { return m_fbo; }
124
125private:
126	const glw::Functions&	m_gl;
127	deUint32				m_fbo;
128};
129
130class AutoRbo
131{
132public:
133	AutoRbo (const glw::Functions& gl)
134		: m_gl	(gl)
135		, m_rbo	(0)
136	{
137	}
138
139	~AutoRbo (void)
140	{
141		if (m_rbo)
142			m_gl.deleteRenderbuffers(1, &m_rbo);
143	}
144
145	void gen (void)
146	{
147		DE_ASSERT(!m_rbo);
148		m_gl.genRenderbuffers(1, &m_rbo);
149	}
150
151	deUint32 operator* (void) const { return m_rbo; }
152
153private:
154	const glw::Functions&	m_gl;
155	deUint32				m_rbo;
156};
157
158} // anonymous
159
160static const char* getDerivateFuncName (DerivateFunc func)
161{
162	switch (func)
163	{
164		case DERIVATE_DFDX:		return "dFdx";
165		case DERIVATE_DFDY:		return "dFdy";
166		case DERIVATE_FWIDTH:	return "fwidth";
167		default:
168			DE_ASSERT(false);
169			return DE_NULL;
170	}
171}
172
173static const char* getDerivateFuncCaseName (DerivateFunc func)
174{
175	switch (func)
176	{
177		case DERIVATE_DFDX:		return "dfdx";
178		case DERIVATE_DFDY:		return "dfdy";
179		case DERIVATE_FWIDTH:	return "fwidth";
180		default:
181			DE_ASSERT(false);
182			return DE_NULL;
183	}
184}
185
186static inline tcu::BVec4 getDerivateMask (glu::DataType type)
187{
188	switch (type)
189	{
190		case glu::TYPE_FLOAT:		return tcu::BVec4(true, false, false, false);
191		case glu::TYPE_FLOAT_VEC2:	return tcu::BVec4(true, true, false, false);
192		case glu::TYPE_FLOAT_VEC3:	return tcu::BVec4(true, true, true, false);
193		case glu::TYPE_FLOAT_VEC4:	return tcu::BVec4(true, true, true, true);
194		default:
195			DE_ASSERT(false);
196			return tcu::BVec4(true);
197	}
198}
199
200static inline tcu::Vec4 readDerivate (const tcu::ConstPixelBufferAccess& surface, const tcu::Vec4& derivScale, const tcu::Vec4& derivBias, int x, int y)
201{
202	return (surface.getPixel(x, y) - derivBias) / derivScale;
203}
204
205static inline tcu::UVec4 getCompExpBits (const tcu::Vec4& v)
206{
207	return tcu::UVec4(tcu::Float32(v[0]).exponentBits(),
208					  tcu::Float32(v[1]).exponentBits(),
209					  tcu::Float32(v[2]).exponentBits(),
210					  tcu::Float32(v[3]).exponentBits());
211}
212
213float computeFloatingPointError (const float value, const int numAccurateBits)
214{
215	const int		numGarbageBits	= 23-numAccurateBits;
216	const deUint32	mask			= (1u<<numGarbageBits)-1u;
217	const int		exp				= tcu::Float32(value).exponent();
218
219	return tcu::Float32::construct(+1, exp, (1u<<23) | mask).asFloat() - tcu::Float32::construct(+1, exp, 1u<<23).asFloat();
220}
221
222static int getNumMantissaBits (const glu::Precision precision)
223{
224	switch (precision)
225	{
226		case glu::PRECISION_HIGHP:		return 23;
227		case glu::PRECISION_MEDIUMP:	return 10;
228		case glu::PRECISION_LOWP:		return 6;
229		default:
230			DE_ASSERT(false);
231			return 0;
232	}
233}
234
235static int getMinExponent (const glu::Precision precision)
236{
237	switch (precision)
238	{
239		case glu::PRECISION_HIGHP:		return -126;
240		case glu::PRECISION_MEDIUMP:	return -14;
241		case glu::PRECISION_LOWP:		return -8;
242		default:
243			DE_ASSERT(false);
244			return 0;
245	}
246}
247
248static float getSingleULPForExponent (int exp, int numMantissaBits)
249{
250	if (numMantissaBits > 0)
251	{
252		DE_ASSERT(numMantissaBits <= 23);
253
254		const int ulpBitNdx = 23-numMantissaBits;
255		return tcu::Float32::construct(+1, exp, (1<<23) | (1 << ulpBitNdx)).asFloat() - tcu::Float32::construct(+1, exp, (1<<23)).asFloat();
256	}
257	else
258	{
259		DE_ASSERT(numMantissaBits == 0);
260		return tcu::Float32::construct(+1, exp, (1<<23)).asFloat();
261	}
262}
263
264static float getSingleULPForValue (float value, int numMantissaBits)
265{
266	const int exp = tcu::Float32(value).exponent();
267	return getSingleULPForExponent(exp, numMantissaBits);
268}
269
270static float convertFloatFlushToZeroRtn (float value, int minExponent, int numAccurateBits)
271{
272	if (value == 0.0f)
273	{
274		return 0.0f;
275	}
276	else
277	{
278		const tcu::Float32	inputFloat			= tcu::Float32(value);
279		const int			numTruncatedBits	= 23-numAccurateBits;
280		const deUint32		truncMask			= (1u<<numTruncatedBits)-1u;
281
282		if (value > 0.0f)
283		{
284			if (value > 0.0f && tcu::Float32(value).exponent() < minExponent)
285			{
286				// flush to zero if possible
287				return 0.0f;
288			}
289			else
290			{
291				// just mask away non-representable bits
292				return tcu::Float32::construct(+1, inputFloat.exponent(), inputFloat.mantissa() & ~truncMask).asFloat();
293			}
294		}
295		else
296		{
297			if (inputFloat.mantissa() & truncMask)
298			{
299				// decrement one ulp if truncated bits are non-zero (i.e. if value is not representable)
300				return tcu::Float32::construct(-1, inputFloat.exponent(), inputFloat.mantissa() & ~truncMask).asFloat() - getSingleULPForExponent(inputFloat.exponent(), numAccurateBits);
301			}
302			else
303			{
304				// value is representable, no need to do anything
305				return value;
306			}
307		}
308	}
309}
310
311static float convertFloatFlushToZeroRtp (float value, int minExponent, int numAccurateBits)
312{
313	return -convertFloatFlushToZeroRtn(-value, minExponent, numAccurateBits);
314}
315
316static float addErrorUlp (float value, float numUlps, int numMantissaBits)
317{
318	return value + numUlps * getSingleULPForValue(value, numMantissaBits);
319}
320
321enum
322{
323	INTERPOLATION_LOST_BITS = 3, // number mantissa of bits allowed to be lost in varying interpolation
324};
325
326static inline tcu::Vec4 getDerivateThreshold (const glu::Precision precision, const tcu::Vec4& valueMin, const tcu::Vec4& valueMax, const tcu::Vec4& expectedDerivate)
327{
328	const int			baseBits		= getNumMantissaBits(precision);
329	const tcu::UVec4	derivExp		= getCompExpBits(expectedDerivate);
330	const tcu::UVec4	maxValueExp		= max(getCompExpBits(valueMin), getCompExpBits(valueMax));
331	const tcu::UVec4	numBitsLost		= maxValueExp - min(maxValueExp, derivExp);
332	const tcu::IVec4	numAccurateBits	= max(baseBits - numBitsLost.asInt() - (int)INTERPOLATION_LOST_BITS, tcu::IVec4(0));
333
334	return tcu::Vec4(computeFloatingPointError(expectedDerivate[0], numAccurateBits[0]),
335					 computeFloatingPointError(expectedDerivate[1], numAccurateBits[1]),
336					 computeFloatingPointError(expectedDerivate[2], numAccurateBits[2]),
337					 computeFloatingPointError(expectedDerivate[3], numAccurateBits[3]));
338}
339
340namespace
341{
342
343struct LogVecComps
344{
345	const tcu::Vec4&	v;
346	int					numComps;
347
348	LogVecComps (const tcu::Vec4& v_, int numComps_)
349		: v			(v_)
350		, numComps	(numComps_)
351	{
352	}
353};
354
355std::ostream& operator<< (std::ostream& str, const LogVecComps& v)
356{
357	DE_ASSERT(de::inRange(v.numComps, 1, 4));
358	if (v.numComps == 1)		return str << v.v[0];
359	else if (v.numComps == 2)	return str << v.v.toWidth<2>();
360	else if (v.numComps == 3)	return str << v.v.toWidth<3>();
361	else						return str << v.v;
362}
363
364} // anonymous
365
366enum VerificationLogging
367{
368	LOG_ALL = 0,
369	LOG_NOTHING
370};
371
372static bool verifyConstantDerivate (tcu::TestLog&						log,
373									const tcu::ConstPixelBufferAccess&	result,
374									const tcu::PixelBufferAccess&		errorMask,
375									glu::DataType						dataType,
376									const tcu::Vec4&					reference,
377									const tcu::Vec4&					threshold,
378									const tcu::Vec4&					scale,
379									const tcu::Vec4&					bias,
380									VerificationLogging					logPolicy = LOG_ALL)
381{
382	const int			numComps		= glu::getDataTypeFloatScalars(dataType);
383	const tcu::BVec4	mask			= tcu::logicalNot(getDerivateMask(dataType));
384	int					numFailedPixels	= 0;
385
386	if (logPolicy == LOG_ALL)
387		log << TestLog::Message << "Expecting " << LogVecComps(reference, numComps) << " with threshold " << LogVecComps(threshold, numComps) << TestLog::EndMessage;
388
389	for (int y = 0; y < result.getHeight(); y++)
390	{
391		for (int x = 0; x < result.getWidth(); x++)
392		{
393			const tcu::Vec4		resDerivate		= readDerivate(result, scale, bias, x, y);
394			const bool			isOk			= tcu::allEqual(tcu::logicalOr(tcu::lessThanEqual(tcu::abs(reference - resDerivate), threshold), mask), tcu::BVec4(true));
395
396			if (!isOk)
397			{
398				if (numFailedPixels < MAX_FAILED_MESSAGES && logPolicy == LOG_ALL)
399					log << TestLog::Message << "FAIL: got " << LogVecComps(resDerivate, numComps)
400											<< ", diff = " << LogVecComps(tcu::abs(reference - resDerivate), numComps)
401											<< ", at x = " << x << ", y = " << y
402						<< TestLog::EndMessage;
403				numFailedPixels += 1;
404				errorMask.setPixel(tcu::RGBA::red().toVec(), x, y);
405			}
406		}
407	}
408
409	if (numFailedPixels >= MAX_FAILED_MESSAGES && logPolicy == LOG_ALL)
410		log << TestLog::Message << "..." << TestLog::EndMessage;
411
412	if (numFailedPixels > 0 && logPolicy == LOG_ALL)
413		log << TestLog::Message << "FAIL: found " << numFailedPixels << " failed pixels" << TestLog::EndMessage;
414
415	return numFailedPixels == 0;
416}
417
418struct Linear2DFunctionEvaluator
419{
420	tcu::Matrix<float, 4, 3> matrix;
421
422	//      .-----.
423	//      | s_x |
424	//  M x | s_y |
425	//      | 1.0 |
426	//      '-----'
427	tcu::Vec4 evaluateAt (float screenX, float screenY) const;
428};
429
430tcu::Vec4 Linear2DFunctionEvaluator::evaluateAt (float screenX, float screenY) const
431{
432	const tcu::Vec3 position(screenX, screenY, 1.0f);
433	return matrix * position;
434}
435
436static bool reverifyConstantDerivateWithFlushRelaxations (tcu::TestLog&							log,
437														  const tcu::ConstPixelBufferAccess&	result,
438														  const tcu::PixelBufferAccess&			errorMask,
439														  glu::DataType							dataType,
440														  glu::Precision						precision,
441														  const tcu::Vec4&						derivScale,
442														  const tcu::Vec4&						derivBias,
443														  const tcu::Vec4&						surfaceThreshold,
444														  DerivateFunc							derivateFunc,
445														  const Linear2DFunctionEvaluator&		function)
446{
447	DE_ASSERT(result.getWidth() == errorMask.getWidth());
448	DE_ASSERT(result.getHeight() == errorMask.getHeight());
449	DE_ASSERT(derivateFunc == DERIVATE_DFDX || derivateFunc == DERIVATE_DFDY);
450
451	const tcu::IVec4	red						(255, 0, 0, 255);
452	const tcu::IVec4	green					(0, 255, 0, 255);
453	const float			divisionErrorUlps		= 2.5f;
454
455	const int			numComponents			= glu::getDataTypeFloatScalars(dataType);
456	const int			numBits					= getNumMantissaBits(precision);
457	const int			minExponent				= getMinExponent(precision);
458
459	const int			numVaryingSampleBits	= numBits - INTERPOLATION_LOST_BITS;
460	int					numFailedPixels			= 0;
461
462	tcu::clear(errorMask, green);
463
464	// search for failed pixels
465	for (int y = 0; y < result.getHeight(); ++y)
466	for (int x = 0; x < result.getWidth(); ++x)
467	{
468		//                 flushToZero?(f2z?(functionValueCurrent) - f2z?(functionValueBefore))
469		// flushToZero? ( ------------------------------------------------------------------------ +- 2.5 ULP )
470		//                                                  dx
471
472		const tcu::Vec4	resultDerivative		= readDerivate(result, derivScale, derivBias, x, y);
473
474		// sample at the front of the back pixel and the back of the front pixel to cover the whole area of
475		// legal sample positions. In general case this is NOT OK, but we know that the target funtion is
476		// (mostly*) linear which allows us to take the sample points at arbitrary points. This gets us the
477		// maximum difference possible in exponents which are used in error bound calculations.
478		// * non-linearity may happen around zero or with very high function values due to subnorms not
479		//   behaving well.
480		const tcu::Vec4	functionValueForward	= (derivateFunc == DERIVATE_DFDX)
481													? (function.evaluateAt((float)x + 2.0f, (float)y + 0.5f))
482													: (function.evaluateAt((float)x + 0.5f, (float)y + 2.0f));
483		const tcu::Vec4	functionValueBackward	= (derivateFunc == DERIVATE_DFDX)
484													? (function.evaluateAt((float)x - 1.0f, (float)y + 0.5f))
485													: (function.evaluateAt((float)x + 0.5f, (float)y - 1.0f));
486
487		bool	anyComponentFailed				= false;
488
489		// check components separately
490		for (int c = 0; c < numComponents; ++c)
491		{
492			// Simulate interpolation. Add allowed interpolation error and round to target precision. Allow one half ULP (i.e. correct rounding)
493			const tcu::Interval	forwardComponent		(convertFloatFlushToZeroRtn(addErrorUlp((float)functionValueForward[c],  -0.5f, numVaryingSampleBits), minExponent, numBits),
494														 convertFloatFlushToZeroRtp(addErrorUlp((float)functionValueForward[c],  +0.5f, numVaryingSampleBits), minExponent, numBits));
495			const tcu::Interval	backwardComponent		(convertFloatFlushToZeroRtn(addErrorUlp((float)functionValueBackward[c], -0.5f, numVaryingSampleBits), minExponent, numBits),
496														 convertFloatFlushToZeroRtp(addErrorUlp((float)functionValueBackward[c], +0.5f, numVaryingSampleBits), minExponent, numBits));
497			const int			maxValueExp				= de::max(de::max(tcu::Float32(forwardComponent.lo()).exponent(),   tcu::Float32(forwardComponent.hi()).exponent()),
498																  de::max(tcu::Float32(backwardComponent.lo()).exponent(),  tcu::Float32(backwardComponent.hi()).exponent()));
499
500			// subtraction in numerator will likely cause a cancellation of the most
501			// significant bits. Apply error bounds.
502
503			const tcu::Interval	numerator				(forwardComponent - backwardComponent);
504			const int			numeratorLoExp			= tcu::Float32(numerator.lo()).exponent();
505			const int			numeratorHiExp			= tcu::Float32(numerator.hi()).exponent();
506			const int			numeratorLoBitsLost		= de::max(0, maxValueExp - numeratorLoExp); //!< must clamp to zero since if forward and backward components have different
507			const int			numeratorHiBitsLost		= de::max(0, maxValueExp - numeratorHiExp); //!< sign, numerator might have larger exponent than its operands.
508			const int			numeratorLoBits			= de::max(0, numBits - numeratorLoBitsLost);
509			const int			numeratorHiBits			= de::max(0, numBits - numeratorHiBitsLost);
510
511			const tcu::Interval	numeratorRange			(convertFloatFlushToZeroRtn((float)numerator.lo(), minExponent, numeratorLoBits),
512														 convertFloatFlushToZeroRtp((float)numerator.hi(), minExponent, numeratorHiBits));
513
514			const tcu::Interval	divisionRange			= numeratorRange / 3.0f; // legal sample area is anywhere within this and neighboring pixels (i.e. size = 3)
515			const tcu::Interval	divisionResultRange		(convertFloatFlushToZeroRtn(addErrorUlp((float)divisionRange.lo(), -divisionErrorUlps, numBits), minExponent, numBits),
516														 convertFloatFlushToZeroRtp(addErrorUlp((float)divisionRange.hi(), +divisionErrorUlps, numBits), minExponent, numBits));
517			const tcu::Interval	finalResultRange		(divisionResultRange.lo() - surfaceThreshold[c], divisionResultRange.hi() + surfaceThreshold[c]);
518
519			if (resultDerivative[c] >= finalResultRange.lo() && resultDerivative[c] <= finalResultRange.hi())
520			{
521				// value ok
522			}
523			else
524			{
525				if (numFailedPixels < MAX_FAILED_MESSAGES)
526					log << tcu::TestLog::Message
527						<< "Error in pixel at " << x << ", " << y << " with component " << c << " (channel " << ("rgba"[c]) << ")\n"
528						<< "\tGot pixel value " << result.getPixelInt(x, y) << "\n"
529						<< "\t\tdFd" << ((derivateFunc == DERIVATE_DFDX) ? ('x') : ('y')) << " ~= " << resultDerivative[c] << "\n"
530						<< "\t\tdifference to a valid range: "
531							<< ((resultDerivative[c] < finalResultRange.lo()) ? ("-") : ("+"))
532							<< ((resultDerivative[c] < finalResultRange.lo()) ? (finalResultRange.lo() - resultDerivative[c]) : (resultDerivative[c] - finalResultRange.hi()))
533							<< "\n"
534						<< "\tDerivative value range:\n"
535						<< "\t\tMin: " << finalResultRange.lo() << "\n"
536						<< "\t\tMax: " << finalResultRange.hi() << "\n"
537						<< tcu::TestLog::EndMessage;
538
539				++numFailedPixels;
540				anyComponentFailed = true;
541			}
542		}
543
544		if (anyComponentFailed)
545			errorMask.setPixel(red, x, y);
546	}
547
548	if (numFailedPixels >= MAX_FAILED_MESSAGES)
549		log << TestLog::Message << "..." << TestLog::EndMessage;
550
551	if (numFailedPixels > 0)
552		log << TestLog::Message << "FAIL: found " << numFailedPixels << " failed pixels" << TestLog::EndMessage;
553
554	return numFailedPixels == 0;
555}
556
557// TriangleDerivateCase
558
559class TriangleDerivateCase : public TestCase
560{
561public:
562						TriangleDerivateCase	(Context& context, const char* name, const char* description);
563						~TriangleDerivateCase	(void);
564
565	IterateResult		iterate					(void);
566
567protected:
568	virtual void		setupRenderState		(deUint32 program) { DE_UNREF(program); }
569	virtual bool		verify					(const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask) = DE_NULL;
570
571	tcu::IVec2			getViewportSize			(void) const;
572	tcu::Vec4			getSurfaceThreshold		(void) const;
573
574	glu::DataType		m_dataType;
575	glu::Precision		m_precision;
576
577	glu::DataType		m_coordDataType;
578	glu::Precision		m_coordPrecision;
579
580	std::string			m_fragmentSrc;
581
582	tcu::Vec4			m_coordMin;
583	tcu::Vec4			m_coordMax;
584	tcu::Vec4			m_derivScale;
585	tcu::Vec4			m_derivBias;
586
587	SurfaceType			m_surfaceType;
588	int					m_numSamples;
589	deUint32			m_hint;
590};
591
592TriangleDerivateCase::TriangleDerivateCase (Context& context, const char* name, const char* description)
593	: TestCase			(context, name, description)
594	, m_dataType		(glu::TYPE_LAST)
595	, m_precision		(glu::PRECISION_LAST)
596	, m_coordDataType	(glu::TYPE_LAST)
597	, m_coordPrecision	(glu::PRECISION_LAST)
598	, m_surfaceType		(SURFACETYPE_DEFAULT_FRAMEBUFFER)
599	, m_numSamples		(0)
600	, m_hint			(GL_DONT_CARE)
601{
602	DE_ASSERT(m_surfaceType != SURFACETYPE_DEFAULT_FRAMEBUFFER || m_numSamples == 0);
603}
604
605TriangleDerivateCase::~TriangleDerivateCase (void)
606{
607	TriangleDerivateCase::deinit();
608}
609
610static std::string genVertexSource (glu::DataType coordType, glu::Precision precision)
611{
612	DE_ASSERT(glu::isDataTypeFloatOrVec(coordType));
613
614	const char* vertexTmpl =
615		"#version 300 es\n"
616		"in highp vec4 a_position;\n"
617		"in ${PRECISION} ${DATATYPE} a_coord;\n"
618		"out ${PRECISION} ${DATATYPE} v_coord;\n"
619		"void main (void)\n"
620		"{\n"
621		"	gl_Position = a_position;\n"
622		"	v_coord = a_coord;\n"
623		"}\n";
624
625	map<string, string> vertexParams;
626
627	vertexParams["PRECISION"]	= glu::getPrecisionName(precision);
628	vertexParams["DATATYPE"]	= glu::getDataTypeName(coordType);
629
630	return tcu::StringTemplate(vertexTmpl).specialize(vertexParams);
631}
632
633inline tcu::IVec2 TriangleDerivateCase::getViewportSize (void) const
634{
635	if (m_surfaceType == SURFACETYPE_DEFAULT_FRAMEBUFFER)
636	{
637		const int	width	= de::min<int>(m_context.getRenderTarget().getWidth(),	VIEWPORT_WIDTH);
638		const int	height	= de::min<int>(m_context.getRenderTarget().getHeight(),	VIEWPORT_HEIGHT);
639		return tcu::IVec2(width, height);
640	}
641	else
642		return tcu::IVec2(FBO_WIDTH, FBO_HEIGHT);
643}
644
645TriangleDerivateCase::IterateResult TriangleDerivateCase::iterate (void)
646{
647	const glw::Functions&		gl				= m_context.getRenderContext().getFunctions();
648	const glu::ShaderProgram	program			(m_context.getRenderContext(), glu::makeVtxFragSources(genVertexSource(m_coordDataType, m_coordPrecision), m_fragmentSrc));
649	de::Random					rnd				(deStringHash(getName()) ^ 0xbbc24);
650	const bool					useFbo			= m_surfaceType != SURFACETYPE_DEFAULT_FRAMEBUFFER;
651	const deUint32				fboFormat		= m_surfaceType == SURFACETYPE_FLOAT_FBO ? GL_RGBA32UI : GL_RGBA8;
652	const tcu::IVec2			viewportSize	= getViewportSize();
653	const int					viewportX		= useFbo ? 0 : rnd.getInt(0, m_context.getRenderTarget().getWidth()		- viewportSize.x());
654	const int					viewportY		= useFbo ? 0 : rnd.getInt(0, m_context.getRenderTarget().getHeight()	- viewportSize.y());
655	AutoFbo						fbo				(gl);
656	AutoRbo						rbo				(gl);
657	tcu::TextureLevel			result;
658
659	m_testCtx.getLog() << program;
660
661	if (!program.isOk())
662		TCU_FAIL("Compile failed");
663
664	if (useFbo)
665	{
666		m_testCtx.getLog() << TestLog::Message
667						   << "Rendering to FBO, format = " << glu::getTextureFormatStr(fboFormat)
668						   << ", samples = " << m_numSamples
669						   << TestLog::EndMessage;
670
671		fbo.gen();
672		rbo.gen();
673
674		gl.bindRenderbuffer(GL_RENDERBUFFER, *rbo);
675		gl.renderbufferStorageMultisample(GL_RENDERBUFFER, m_numSamples, fboFormat, viewportSize.x(), viewportSize.y());
676		gl.bindFramebuffer(GL_FRAMEBUFFER, *fbo);
677		gl.framebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, *rbo);
678		TCU_CHECK(gl.checkFramebufferStatus(GL_FRAMEBUFFER) == GL_FRAMEBUFFER_COMPLETE);
679	}
680	else
681	{
682		const tcu::PixelFormat pixelFormat = m_context.getRenderTarget().getPixelFormat();
683
684		m_testCtx.getLog()
685			<< TestLog::Message
686			<< "Rendering to default framebuffer\n"
687			<< "\tColor depth: R=" << pixelFormat.redBits << ", G=" << pixelFormat.greenBits << ", B=" << pixelFormat.blueBits << ", A=" << pixelFormat.alphaBits
688			<< TestLog::EndMessage;
689	}
690
691	m_testCtx.getLog() << TestLog::Message << "in: " << m_coordMin << " -> " << m_coordMax << "\n"
692										   << "v_coord.x = in.x * x\n"
693										   << "v_coord.y = in.y * y\n"
694										   << "v_coord.z = in.z * (x+y)/2\n"
695										   << "v_coord.w = in.w * (1 - (x+y)/2)\n"
696					   << TestLog::EndMessage
697					   << TestLog::Message << "u_scale: " << m_derivScale << ", u_bias: " << m_derivBias << " (displayed values have scale/bias removed)" << TestLog::EndMessage
698					   << TestLog::Message << "Viewport: " << viewportSize.x() << "x" << viewportSize.y() << TestLog::EndMessage
699					   << TestLog::Message << "GL_FRAGMENT_SHADER_DERIVATE_HINT: " << glu::getHintModeStr(m_hint) << TestLog::EndMessage;
700
701	// Draw
702	{
703		const float positions[] =
704		{
705			-1.0f, -1.0f, 0.0f, 1.0f,
706			-1.0f,  1.0f, 0.0f, 1.0f,
707			 1.0f, -1.0f, 0.0f, 1.0f,
708			 1.0f,  1.0f, 0.0f, 1.0f
709		};
710		const float coords[] =
711		{
712			m_coordMin.x(), m_coordMin.y(), m_coordMin.z(),							m_coordMax.w(),
713			m_coordMin.x(), m_coordMax.y(), (m_coordMin.z()+m_coordMax.z())*0.5f,	(m_coordMin.w()+m_coordMax.w())*0.5f,
714			m_coordMax.x(), m_coordMin.y(), (m_coordMin.z()+m_coordMax.z())*0.5f,	(m_coordMin.w()+m_coordMax.w())*0.5f,
715			m_coordMax.x(), m_coordMax.y(), m_coordMax.z(),							m_coordMin.w()
716		};
717		const glu::VertexArrayBinding vertexArrays[] =
718		{
719			glu::va::Float("a_position",	4, 4, 0, &positions[0]),
720			glu::va::Float("a_coord",		4, 4, 0, &coords[0])
721		};
722		const deUint16 indices[] = { 0, 2, 1, 2, 3, 1 };
723
724		gl.clearColor(0.125f, 0.25f, 0.5f, 1.0f);
725		gl.clear(GL_COLOR_BUFFER_BIT|GL_DEPTH_BUFFER_BIT|GL_STENCIL_BUFFER_BIT);
726		gl.disable(GL_DITHER);
727
728		gl.useProgram(program.getProgram());
729
730		{
731			const int	scaleLoc	= gl.getUniformLocation(program.getProgram(), "u_scale");
732			const int	biasLoc		= gl.getUniformLocation(program.getProgram(), "u_bias");
733
734			switch (m_dataType)
735			{
736				case glu::TYPE_FLOAT:
737					gl.uniform1f(scaleLoc, m_derivScale.x());
738					gl.uniform1f(biasLoc, m_derivBias.x());
739					break;
740
741				case glu::TYPE_FLOAT_VEC2:
742					gl.uniform2fv(scaleLoc, 1, m_derivScale.getPtr());
743					gl.uniform2fv(biasLoc, 1, m_derivBias.getPtr());
744					break;
745
746				case glu::TYPE_FLOAT_VEC3:
747					gl.uniform3fv(scaleLoc, 1, m_derivScale.getPtr());
748					gl.uniform3fv(biasLoc, 1, m_derivBias.getPtr());
749					break;
750
751				case glu::TYPE_FLOAT_VEC4:
752					gl.uniform4fv(scaleLoc, 1, m_derivScale.getPtr());
753					gl.uniform4fv(biasLoc, 1, m_derivBias.getPtr());
754					break;
755
756				default:
757					DE_ASSERT(false);
758			}
759		}
760
761		gls::setupDefaultUniforms(m_context.getRenderContext(), program.getProgram());
762		setupRenderState(program.getProgram());
763
764		gl.hint(GL_FRAGMENT_SHADER_DERIVATIVE_HINT, m_hint);
765		GLU_EXPECT_NO_ERROR(gl.getError(), "Setup program state");
766
767		gl.viewport(viewportX, viewportY, viewportSize.x(), viewportSize.y());
768		glu::draw(m_context.getRenderContext(), program.getProgram(), DE_LENGTH_OF_ARRAY(vertexArrays), &vertexArrays[0],
769				  glu::pr::Triangles(DE_LENGTH_OF_ARRAY(indices), &indices[0]));
770		GLU_EXPECT_NO_ERROR(gl.getError(), "Draw");
771	}
772
773	// Read back results
774	{
775		const bool		isMSAA		= useFbo && m_numSamples > 0;
776		AutoFbo			resFbo		(gl);
777		AutoRbo			resRbo		(gl);
778
779		// Resolve if necessary
780		if (isMSAA)
781		{
782			resFbo.gen();
783			resRbo.gen();
784
785			gl.bindRenderbuffer(GL_RENDERBUFFER, *resRbo);
786			gl.renderbufferStorageMultisample(GL_RENDERBUFFER, 0, fboFormat, viewportSize.x(), viewportSize.y());
787			gl.bindFramebuffer(GL_DRAW_FRAMEBUFFER, *resFbo);
788			gl.framebufferRenderbuffer(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, *resRbo);
789			TCU_CHECK(gl.checkFramebufferStatus(GL_FRAMEBUFFER) == GL_FRAMEBUFFER_COMPLETE);
790
791			gl.blitFramebuffer(0, 0, viewportSize.x(), viewportSize.y(), 0, 0, viewportSize.x(), viewportSize.y(), GL_COLOR_BUFFER_BIT, GL_NEAREST);
792			GLU_EXPECT_NO_ERROR(gl.getError(), "Resolve blit");
793
794			gl.bindFramebuffer(GL_READ_FRAMEBUFFER, *resFbo);
795		}
796
797		switch (m_surfaceType)
798		{
799			case SURFACETYPE_DEFAULT_FRAMEBUFFER:
800			case SURFACETYPE_UNORM_FBO:
801				result.setStorage(tcu::TextureFormat(tcu::TextureFormat::RGBA, tcu::TextureFormat::UNORM_INT8), viewportSize.x(), viewportSize.y());
802				glu::readPixels(m_context.getRenderContext(), viewportX, viewportY, result);
803				break;
804
805			case SURFACETYPE_FLOAT_FBO:
806			{
807				const tcu::TextureFormat	dataFormat		(tcu::TextureFormat::RGBA, tcu::TextureFormat::FLOAT);
808				const tcu::TextureFormat	transferFormat	(tcu::TextureFormat::RGBA, tcu::TextureFormat::UNSIGNED_INT32);
809
810				result.setStorage(dataFormat, viewportSize.x(), viewportSize.y());
811				glu::readPixels(m_context.getRenderContext(), viewportX, viewportY,
812								tcu::PixelBufferAccess(transferFormat, result.getWidth(), result.getHeight(), result.getDepth(), result.getAccess().getDataPtr()));
813				break;
814			}
815
816			default:
817				DE_ASSERT(false);
818		}
819
820		GLU_EXPECT_NO_ERROR(gl.getError(), "Read pixels");
821	}
822
823	// Verify
824	{
825		tcu::Surface errorMask(result.getWidth(), result.getHeight());
826		tcu::clear(errorMask.getAccess(), tcu::RGBA::green().toVec());
827
828		const bool isOk = verify(result.getAccess(), errorMask.getAccess());
829
830		m_testCtx.getLog() << TestLog::ImageSet("Result", "Result images")
831						   << TestLog::Image("Rendered", "Rendered image", result);
832
833		if (!isOk)
834			m_testCtx.getLog() << TestLog::Image("ErrorMask", "Error mask", errorMask);
835
836		m_testCtx.getLog() << TestLog::EndImageSet;
837
838		m_testCtx.setTestResult(isOk ? QP_TEST_RESULT_PASS	: QP_TEST_RESULT_FAIL,
839								isOk ? "Pass"				: "Image comparison failed");
840	}
841
842	return STOP;
843}
844
845tcu::Vec4 TriangleDerivateCase::getSurfaceThreshold (void) const
846{
847	switch (m_surfaceType)
848	{
849		case SURFACETYPE_DEFAULT_FRAMEBUFFER:
850		{
851			const tcu::PixelFormat	pixelFormat		= m_context.getRenderTarget().getPixelFormat();
852			const tcu::IVec4		channelBits		(pixelFormat.redBits, pixelFormat.greenBits, pixelFormat.blueBits, pixelFormat.alphaBits);
853			const tcu::IVec4		intThreshold	= tcu::IVec4(1) << (8 - channelBits);
854			const tcu::Vec4			normThreshold	= intThreshold.asFloat() / 255.0f;
855
856			return normThreshold;
857		}
858
859		case SURFACETYPE_UNORM_FBO:				return tcu::IVec4(1).asFloat() / 255.0f;
860		case SURFACETYPE_FLOAT_FBO:				return tcu::Vec4(0.0f);
861		default:
862			DE_ASSERT(false);
863			return tcu::Vec4(0.0f);
864	}
865}
866
867// ConstantDerivateCase
868
869class ConstantDerivateCase : public TriangleDerivateCase
870{
871public:
872						ConstantDerivateCase		(Context& context, const char* name, const char* description, DerivateFunc func, glu::DataType type);
873						~ConstantDerivateCase		(void) {}
874
875	void				init						(void);
876
877protected:
878	bool				verify						(const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask);
879
880private:
881	DerivateFunc		m_func;
882};
883
884ConstantDerivateCase::ConstantDerivateCase (Context& context, const char* name, const char* description, DerivateFunc func, glu::DataType type)
885	: TriangleDerivateCase	(context, name, description)
886	, m_func				(func)
887{
888	m_dataType			= type;
889	m_precision			= glu::PRECISION_HIGHP;
890	m_coordDataType		= m_dataType;
891	m_coordPrecision	= m_precision;
892}
893
894void ConstantDerivateCase::init (void)
895{
896	const char* fragmentTmpl =
897		"#version 300 es\n"
898		"layout(location = 0) out mediump vec4 o_color;\n"
899		"uniform ${PRECISION} ${DATATYPE} u_scale;\n"
900		"uniform ${PRECISION} ${DATATYPE} u_bias;\n"
901		"void main (void)\n"
902		"{\n"
903		"	${PRECISION} ${DATATYPE} res = ${FUNC}(${VALUE}) * u_scale + u_bias;\n"
904		"	o_color = ${CAST_TO_OUTPUT};\n"
905		"}\n";
906	map<string, string> fragmentParams;
907	fragmentParams["PRECISION"]			= glu::getPrecisionName(m_precision);
908	fragmentParams["DATATYPE"]			= glu::getDataTypeName(m_dataType);
909	fragmentParams["FUNC"]				= getDerivateFuncName(m_func);
910	fragmentParams["VALUE"]				= m_dataType == glu::TYPE_FLOAT_VEC4 ? "vec4(1.0, 7.2, -1e5, 0.0)" :
911										  m_dataType == glu::TYPE_FLOAT_VEC3 ? "vec3(1e2, 8.0, 0.01)" :
912										  m_dataType == glu::TYPE_FLOAT_VEC2 ? "vec2(-0.0, 2.7)" :
913										  /* TYPE_FLOAT */					   "7.7";
914	fragmentParams["CAST_TO_OUTPUT"]	= m_dataType == glu::TYPE_FLOAT_VEC4 ? "res" :
915										  m_dataType == glu::TYPE_FLOAT_VEC3 ? "vec4(res, 1.0)" :
916										  m_dataType == glu::TYPE_FLOAT_VEC2 ? "vec4(res, 0.0, 1.0)" :
917										  /* TYPE_FLOAT */					   "vec4(res, 0.0, 0.0, 1.0)";
918
919	m_fragmentSrc = tcu::StringTemplate(fragmentTmpl).specialize(fragmentParams);
920
921	m_derivScale	= tcu::Vec4(1e3f, 1e3f, 1e3f, 1e3f);
922	m_derivBias		= tcu::Vec4(0.5f, 0.5f, 0.5f, 0.5f);
923}
924
925bool ConstantDerivateCase::verify (const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask)
926{
927	const tcu::Vec4 reference	(0.0f); // Derivate of constant argument should always be 0
928	const tcu::Vec4	threshold	= getSurfaceThreshold() / abs(m_derivScale);
929
930	return verifyConstantDerivate(m_testCtx.getLog(), result, errorMask, m_dataType,
931								  reference, threshold, m_derivScale, m_derivBias);
932}
933
934// LinearDerivateCase
935
936class LinearDerivateCase : public TriangleDerivateCase
937{
938public:
939						LinearDerivateCase		(Context& context, const char* name, const char* description, DerivateFunc func, glu::DataType type, glu::Precision precision, deUint32 hint, SurfaceType surfaceType, int numSamples, const char* fragmentSrcTmpl);
940						~LinearDerivateCase		(void) {}
941
942	void				init					(void);
943
944protected:
945	bool				verify					(const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask);
946
947private:
948	DerivateFunc		m_func;
949	std::string			m_fragmentTmpl;
950};
951
952LinearDerivateCase::LinearDerivateCase (Context& context, const char* name, const char* description, DerivateFunc func, glu::DataType type, glu::Precision precision, deUint32 hint, SurfaceType surfaceType, int numSamples, const char* fragmentSrcTmpl)
953	: TriangleDerivateCase	(context, name, description)
954	, m_func				(func)
955	, m_fragmentTmpl		(fragmentSrcTmpl)
956{
957	m_dataType			= type;
958	m_precision			= precision;
959	m_coordDataType		= m_dataType;
960	m_coordPrecision	= m_precision;
961	m_hint				= hint;
962	m_surfaceType		= surfaceType;
963	m_numSamples		= numSamples;
964}
965
966void LinearDerivateCase::init (void)
967{
968	const tcu::IVec2	viewportSize	= getViewportSize();
969	const float			w				= float(viewportSize.x());
970	const float			h				= float(viewportSize.y());
971	const bool			packToInt		= m_surfaceType == SURFACETYPE_FLOAT_FBO;
972	map<string, string>	fragmentParams;
973
974	fragmentParams["OUTPUT_TYPE"]		= glu::getDataTypeName(packToInt ? glu::TYPE_UINT_VEC4 : glu::TYPE_FLOAT_VEC4);
975	fragmentParams["OUTPUT_PREC"]		= glu::getPrecisionName(packToInt ? glu::PRECISION_HIGHP : m_precision);
976	fragmentParams["PRECISION"]			= glu::getPrecisionName(m_precision);
977	fragmentParams["DATATYPE"]			= glu::getDataTypeName(m_dataType);
978	fragmentParams["FUNC"]				= getDerivateFuncName(m_func);
979
980	if (packToInt)
981	{
982		fragmentParams["CAST_TO_OUTPUT"]	= m_dataType == glu::TYPE_FLOAT_VEC4 ? "floatBitsToUint(res)" :
983											  m_dataType == glu::TYPE_FLOAT_VEC3 ? "floatBitsToUint(vec4(res, 1.0))" :
984											  m_dataType == glu::TYPE_FLOAT_VEC2 ? "floatBitsToUint(vec4(res, 0.0, 1.0))" :
985											  /* TYPE_FLOAT */					   "floatBitsToUint(vec4(res, 0.0, 0.0, 1.0))";
986	}
987	else
988	{
989		fragmentParams["CAST_TO_OUTPUT"]	= m_dataType == glu::TYPE_FLOAT_VEC4 ? "res" :
990											  m_dataType == glu::TYPE_FLOAT_VEC3 ? "vec4(res, 1.0)" :
991											  m_dataType == glu::TYPE_FLOAT_VEC2 ? "vec4(res, 0.0, 1.0)" :
992											  /* TYPE_FLOAT */					   "vec4(res, 0.0, 0.0, 1.0)";
993	}
994
995	m_fragmentSrc = tcu::StringTemplate(m_fragmentTmpl.c_str()).specialize(fragmentParams);
996
997	switch (m_precision)
998	{
999		case glu::PRECISION_HIGHP:
1000			m_coordMin = tcu::Vec4(-97.f, 0.2f, 71.f, 74.f);
1001			m_coordMax = tcu::Vec4(-13.2f, -77.f, 44.f, 76.f);
1002			break;
1003
1004		case glu::PRECISION_MEDIUMP:
1005			m_coordMin = tcu::Vec4(-37.0f, 47.f, -7.f, 0.0f);
1006			m_coordMax = tcu::Vec4(-1.0f, 12.f, 7.f, 19.f);
1007			break;
1008
1009		case glu::PRECISION_LOWP:
1010			m_coordMin = tcu::Vec4(0.0f, -1.0f, 0.0f, 1.0f);
1011			m_coordMax = tcu::Vec4(1.0f, 1.0f, -1.0f, -1.0f);
1012			break;
1013
1014		default:
1015			DE_ASSERT(false);
1016	}
1017
1018	if (m_surfaceType == SURFACETYPE_FLOAT_FBO)
1019	{
1020		// No scale or bias used for accuracy.
1021		m_derivScale	= tcu::Vec4(1.0f);
1022		m_derivBias		= tcu::Vec4(0.0f);
1023	}
1024	else
1025	{
1026		// Compute scale - bias that normalizes to 0..1 range.
1027		const tcu::Vec4 dx = (m_coordMax - m_coordMin) / tcu::Vec4(w, w, w*0.5f, -w*0.5f);
1028		const tcu::Vec4 dy = (m_coordMax - m_coordMin) / tcu::Vec4(h, h, h*0.5f, -h*0.5f);
1029
1030		switch (m_func)
1031		{
1032			case DERIVATE_DFDX:
1033				m_derivScale = 0.5f / dx;
1034				break;
1035
1036			case DERIVATE_DFDY:
1037				m_derivScale = 0.5f / dy;
1038				break;
1039
1040			case DERIVATE_FWIDTH:
1041				m_derivScale = 0.5f / (tcu::abs(dx) + tcu::abs(dy));
1042				break;
1043
1044			default:
1045				DE_ASSERT(false);
1046		}
1047
1048		m_derivBias = tcu::Vec4(0.0f, 0.0f, 0.0f, 0.0f);
1049	}
1050}
1051
1052bool LinearDerivateCase::verify (const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask)
1053{
1054	const tcu::Vec4		xScale				= tcu::Vec4(1.0f, 0.0f, 0.5f, -0.5f);
1055	const tcu::Vec4		yScale				= tcu::Vec4(0.0f, 1.0f, 0.5f, -0.5f);
1056	const tcu::Vec4		surfaceThreshold	= getSurfaceThreshold() / abs(m_derivScale);
1057
1058	if (m_func == DERIVATE_DFDX || m_func == DERIVATE_DFDY)
1059	{
1060		const bool			isX			= m_func == DERIVATE_DFDX;
1061		const float			div			= isX ? float(result.getWidth()) : float(result.getHeight());
1062		const tcu::Vec4		scale		= isX ? xScale : yScale;
1063		const tcu::Vec4		reference	= ((m_coordMax - m_coordMin) / div) * scale;
1064		const tcu::Vec4		opThreshold	= getDerivateThreshold(m_precision, m_coordMin*scale, m_coordMax*scale, reference);
1065		const tcu::Vec4		threshold	= max(surfaceThreshold, opThreshold);
1066		const int			numComps	= glu::getDataTypeFloatScalars(m_dataType);
1067
1068		m_testCtx.getLog()
1069			<< tcu::TestLog::Message
1070			<< "Verifying result image.\n"
1071			<< "\tValid derivative is " << LogVecComps(reference, numComps) << " with threshold " << LogVecComps(threshold, numComps)
1072			<< tcu::TestLog::EndMessage;
1073
1074		// short circuit if result is strictly within the normal value error bounds.
1075		// This improves performance significantly.
1076		if (verifyConstantDerivate(m_testCtx.getLog(), result, errorMask, m_dataType,
1077								   reference, threshold, m_derivScale, m_derivBias,
1078								   LOG_NOTHING))
1079		{
1080			m_testCtx.getLog()
1081				<< tcu::TestLog::Message
1082				<< "No incorrect derivatives found, result valid."
1083				<< tcu::TestLog::EndMessage;
1084
1085			return true;
1086		}
1087
1088		// some pixels exceed error bounds calculated for normal values. Verify that these
1089		// potentially invalid pixels are in fact valid due to (for example) subnorm flushing.
1090
1091		m_testCtx.getLog()
1092			<< tcu::TestLog::Message
1093			<< "Initial verification failed, verifying image by calculating accurate error bounds for each result pixel.\n"
1094			<< "\tVerifying each result derivative is within its range of legal result values."
1095			<< tcu::TestLog::EndMessage;
1096
1097		{
1098			const tcu::IVec2			viewportSize	= getViewportSize();
1099			const float					w				= float(viewportSize.x());
1100			const float					h				= float(viewportSize.y());
1101			const tcu::Vec4				valueRamp		= (m_coordMax - m_coordMin);
1102			Linear2DFunctionEvaluator	function;
1103
1104			function.matrix.setRow(0, tcu::Vec3(valueRamp.x() / w, 0.0f, m_coordMin.x()));
1105			function.matrix.setRow(1, tcu::Vec3(0.0f, valueRamp.y() / h, m_coordMin.y()));
1106			function.matrix.setRow(2, tcu::Vec3(valueRamp.z() / w, valueRamp.z() / h, m_coordMin.z() + m_coordMin.z()) / 2.0f);
1107			function.matrix.setRow(3, tcu::Vec3(-valueRamp.w() / w, -valueRamp.w() / h, m_coordMax.w() + m_coordMax.w()) / 2.0f);
1108
1109			return reverifyConstantDerivateWithFlushRelaxations(m_testCtx.getLog(), result, errorMask,
1110																m_dataType, m_precision, m_derivScale,
1111																m_derivBias, surfaceThreshold, m_func,
1112																function);
1113		}
1114	}
1115	else
1116	{
1117		DE_ASSERT(m_func == DERIVATE_FWIDTH);
1118		const float			w			= float(result.getWidth());
1119		const float			h			= float(result.getHeight());
1120
1121		const tcu::Vec4		dx			= ((m_coordMax - m_coordMin) / w) * xScale;
1122		const tcu::Vec4		dy			= ((m_coordMax - m_coordMin) / h) * yScale;
1123		const tcu::Vec4		reference	= tcu::abs(dx) + tcu::abs(dy);
1124		const tcu::Vec4		dxThreshold	= getDerivateThreshold(m_precision, m_coordMin*xScale, m_coordMax*xScale, dx);
1125		const tcu::Vec4		dyThreshold	= getDerivateThreshold(m_precision, m_coordMin*yScale, m_coordMax*yScale, dy);
1126		const tcu::Vec4		threshold	= max(surfaceThreshold, max(dxThreshold, dyThreshold));
1127
1128		return verifyConstantDerivate(m_testCtx.getLog(), result, errorMask, m_dataType,
1129									  reference, threshold, m_derivScale, m_derivBias);
1130	}
1131}
1132
1133// TextureDerivateCase
1134
1135class TextureDerivateCase : public TriangleDerivateCase
1136{
1137public:
1138						TextureDerivateCase		(Context& context, const char* name, const char* description, DerivateFunc func, glu::DataType type, glu::Precision precision, deUint32 hint, SurfaceType surfaceType, int numSamples);
1139						~TextureDerivateCase	(void);
1140
1141	void				init					(void);
1142	void				deinit					(void);
1143
1144protected:
1145	void				setupRenderState		(deUint32 program);
1146	bool				verify					(const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask);
1147
1148private:
1149	DerivateFunc		m_func;
1150
1151	tcu::Vec4			m_texValueMin;
1152	tcu::Vec4			m_texValueMax;
1153	glu::Texture2D*		m_texture;
1154};
1155
1156TextureDerivateCase::TextureDerivateCase (Context& context, const char* name, const char* description, DerivateFunc func, glu::DataType type, glu::Precision precision, deUint32 hint, SurfaceType surfaceType, int numSamples)
1157	: TriangleDerivateCase	(context, name, description)
1158	, m_func				(func)
1159	, m_texture				(DE_NULL)
1160{
1161	m_dataType			= type;
1162	m_precision			= precision;
1163	m_coordDataType		= glu::TYPE_FLOAT_VEC2;
1164	m_coordPrecision	= glu::PRECISION_HIGHP;
1165	m_hint				= hint;
1166	m_surfaceType		= surfaceType;
1167	m_numSamples		= numSamples;
1168}
1169
1170TextureDerivateCase::~TextureDerivateCase (void)
1171{
1172	delete m_texture;
1173}
1174
1175void TextureDerivateCase::init (void)
1176{
1177	// Generate shader
1178	{
1179		const char* fragmentTmpl =
1180			"#version 300 es\n"
1181			"in highp vec2 v_coord;\n"
1182			"layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1183			"uniform ${PRECISION} sampler2D u_sampler;\n"
1184			"uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1185			"uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1186			"void main (void)\n"
1187			"{\n"
1188			"	${PRECISION} vec4 tex = texture(u_sampler, v_coord);\n"
1189			"	${PRECISION} ${DATATYPE} res = ${FUNC}(tex${SWIZZLE}) * u_scale + u_bias;\n"
1190			"	o_color = ${CAST_TO_OUTPUT};\n"
1191			"}\n";
1192
1193		const bool			packToInt		= m_surfaceType == SURFACETYPE_FLOAT_FBO;
1194		map<string, string> fragmentParams;
1195
1196		fragmentParams["OUTPUT_TYPE"]		= glu::getDataTypeName(packToInt ? glu::TYPE_UINT_VEC4 : glu::TYPE_FLOAT_VEC4);
1197		fragmentParams["OUTPUT_PREC"]		= glu::getPrecisionName(packToInt ? glu::PRECISION_HIGHP : m_precision);
1198		fragmentParams["PRECISION"]			= glu::getPrecisionName(m_precision);
1199		fragmentParams["DATATYPE"]			= glu::getDataTypeName(m_dataType);
1200		fragmentParams["FUNC"]				= getDerivateFuncName(m_func);
1201		fragmentParams["SWIZZLE"]			= m_dataType == glu::TYPE_FLOAT_VEC4 ? "" :
1202											  m_dataType == glu::TYPE_FLOAT_VEC3 ? ".xyz" :
1203											  m_dataType == glu::TYPE_FLOAT_VEC2 ? ".xy" :
1204											  /* TYPE_FLOAT */					   ".x";
1205
1206		if (packToInt)
1207		{
1208			fragmentParams["CAST_TO_OUTPUT"]	= m_dataType == glu::TYPE_FLOAT_VEC4 ? "floatBitsToUint(res)" :
1209												  m_dataType == glu::TYPE_FLOAT_VEC3 ? "floatBitsToUint(vec4(res, 1.0))" :
1210												  m_dataType == glu::TYPE_FLOAT_VEC2 ? "floatBitsToUint(vec4(res, 0.0, 1.0))" :
1211												  /* TYPE_FLOAT */					   "floatBitsToUint(vec4(res, 0.0, 0.0, 1.0))";
1212		}
1213		else
1214		{
1215			fragmentParams["CAST_TO_OUTPUT"]	= m_dataType == glu::TYPE_FLOAT_VEC4 ? "res" :
1216												  m_dataType == glu::TYPE_FLOAT_VEC3 ? "vec4(res, 1.0)" :
1217												  m_dataType == glu::TYPE_FLOAT_VEC2 ? "vec4(res, 0.0, 1.0)" :
1218												  /* TYPE_FLOAT */					   "vec4(res, 0.0, 0.0, 1.0)";
1219		}
1220
1221		m_fragmentSrc = tcu::StringTemplate(fragmentTmpl).specialize(fragmentParams);
1222	}
1223
1224	// Texture size matches viewport and nearest sampling is used. Thus texture sampling
1225	// is equal to just interpolating the texture value range.
1226
1227	// Determine value range for texture.
1228
1229	switch (m_precision)
1230	{
1231		case glu::PRECISION_HIGHP:
1232			m_texValueMin = tcu::Vec4(-97.f, 0.2f, 71.f, 74.f);
1233			m_texValueMax = tcu::Vec4(-13.2f, -77.f, 44.f, 76.f);
1234			break;
1235
1236		case glu::PRECISION_MEDIUMP:
1237			m_texValueMin = tcu::Vec4(-37.0f, 47.f, -7.f, 0.0f);
1238			m_texValueMax = tcu::Vec4(-1.0f, 12.f, 7.f, 19.f);
1239			break;
1240
1241		case glu::PRECISION_LOWP:
1242			m_texValueMin = tcu::Vec4(0.0f, -1.0f, 0.0f, 1.0f);
1243			m_texValueMax = tcu::Vec4(1.0f, 1.0f, -1.0f, -1.0f);
1244			break;
1245
1246		default:
1247			DE_ASSERT(false);
1248	}
1249
1250	// Lowp and mediump cases use RGBA16F format, while highp uses RGBA32F.
1251	{
1252		const tcu::IVec2 viewportSize = getViewportSize();
1253		DE_ASSERT(!m_texture);
1254		m_texture = new glu::Texture2D(m_context.getRenderContext(), m_precision == glu::PRECISION_HIGHP ? GL_RGBA32F : GL_RGBA16F, viewportSize.x(), viewportSize.y());
1255		m_texture->getRefTexture().allocLevel(0);
1256	}
1257
1258	// Texture coordinates
1259	m_coordMin = tcu::Vec4(0.0f);
1260	m_coordMax = tcu::Vec4(1.0f);
1261
1262	// Fill with gradients.
1263	{
1264		const tcu::PixelBufferAccess level0 = m_texture->getRefTexture().getLevel(0);
1265		for (int y = 0; y < level0.getHeight(); y++)
1266		{
1267			for (int x = 0; x < level0.getWidth(); x++)
1268			{
1269				const float		xf		= (float(x)+0.5f) / float(level0.getWidth());
1270				const float		yf		= (float(y)+0.5f) / float(level0.getHeight());
1271				const tcu::Vec4	s		= tcu::Vec4(xf, yf, (xf+yf)/2.0f, 1.0f - (xf+yf)/2.0f);
1272
1273				level0.setPixel(m_texValueMin + (m_texValueMax - m_texValueMin)*s, x, y);
1274			}
1275		}
1276	}
1277
1278	m_texture->upload();
1279
1280	if (m_surfaceType == SURFACETYPE_FLOAT_FBO)
1281	{
1282		// No scale or bias used for accuracy.
1283		m_derivScale	= tcu::Vec4(1.0f);
1284		m_derivBias		= tcu::Vec4(0.0f);
1285	}
1286	else
1287	{
1288		// Compute scale - bias that normalizes to 0..1 range.
1289		const tcu::IVec2	viewportSize	= getViewportSize();
1290		const float			w				= float(viewportSize.x());
1291		const float			h				= float(viewportSize.y());
1292		const tcu::Vec4		dx				= (m_texValueMax - m_texValueMin) / tcu::Vec4(w, w, w*0.5f, -w*0.5f);
1293		const tcu::Vec4		dy				= (m_texValueMax - m_texValueMin) / tcu::Vec4(h, h, h*0.5f, -h*0.5f);
1294
1295		switch (m_func)
1296		{
1297			case DERIVATE_DFDX:
1298				m_derivScale = 0.5f / dx;
1299				break;
1300
1301			case DERIVATE_DFDY:
1302				m_derivScale = 0.5f / dy;
1303				break;
1304
1305			case DERIVATE_FWIDTH:
1306				m_derivScale = 0.5f / (tcu::abs(dx) + tcu::abs(dy));
1307				break;
1308
1309			default:
1310				DE_ASSERT(false);
1311		}
1312
1313		m_derivBias = tcu::Vec4(0.0f, 0.0f, 0.0f, 0.0f);
1314	}
1315}
1316
1317void TextureDerivateCase::deinit (void)
1318{
1319	delete m_texture;
1320	m_texture = DE_NULL;
1321}
1322
1323void TextureDerivateCase::setupRenderState (deUint32 program)
1324{
1325	const glw::Functions&	gl			= m_context.getRenderContext().getFunctions();
1326	const int				texUnit		= 1;
1327
1328	gl.activeTexture	(GL_TEXTURE0+texUnit);
1329	gl.bindTexture		(GL_TEXTURE_2D, m_texture->getGLTexture());
1330	gl.texParameteri	(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER,	GL_NEAREST);
1331	gl.texParameteri	(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER,	GL_NEAREST);
1332	gl.texParameteri	(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S,		GL_CLAMP_TO_EDGE);
1333	gl.texParameteri	(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T,		GL_CLAMP_TO_EDGE);
1334
1335	gl.uniform1i		(gl.getUniformLocation(program, "u_sampler"), texUnit);
1336}
1337
1338bool TextureDerivateCase::verify (const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask)
1339{
1340	// \note Edges are ignored in comparison
1341	if (result.getWidth() < 2 || result.getHeight() < 2)
1342		throw tcu::NotSupportedError("Too small viewport");
1343
1344	tcu::ConstPixelBufferAccess	compareArea			= tcu::getSubregion(result, 1, 1, result.getWidth()-2, result.getHeight()-2);
1345	tcu::PixelBufferAccess		maskArea			= tcu::getSubregion(errorMask, 1, 1, errorMask.getWidth()-2, errorMask.getHeight()-2);
1346	const tcu::Vec4				xScale				= tcu::Vec4(1.0f, 0.0f, 0.5f, -0.5f);
1347	const tcu::Vec4				yScale				= tcu::Vec4(0.0f, 1.0f, 0.5f, -0.5f);
1348	const float					w					= float(result.getWidth());
1349	const float					h					= float(result.getHeight());
1350
1351	const tcu::Vec4				surfaceThreshold	= getSurfaceThreshold() / abs(m_derivScale);
1352
1353	if (m_func == DERIVATE_DFDX || m_func == DERIVATE_DFDY)
1354	{
1355		const bool			isX			= m_func == DERIVATE_DFDX;
1356		const float			div			= isX ? w : h;
1357		const tcu::Vec4		scale		= isX ? xScale : yScale;
1358		const tcu::Vec4		reference	= ((m_texValueMax - m_texValueMin) / div) * scale;
1359		const tcu::Vec4		opThreshold	= getDerivateThreshold(m_precision, m_texValueMin*scale, m_texValueMax*scale, reference);
1360		const tcu::Vec4		threshold	= max(surfaceThreshold, opThreshold);
1361		const int			numComps	= glu::getDataTypeFloatScalars(m_dataType);
1362
1363		m_testCtx.getLog()
1364			<< tcu::TestLog::Message
1365			<< "Verifying result image.\n"
1366			<< "\tValid derivative is " << LogVecComps(reference, numComps) << " with threshold " << LogVecComps(threshold, numComps)
1367			<< tcu::TestLog::EndMessage;
1368
1369		// short circuit if result is strictly within the normal value error bounds.
1370		// This improves performance significantly.
1371		if (verifyConstantDerivate(m_testCtx.getLog(), compareArea, maskArea, m_dataType,
1372								   reference, threshold, m_derivScale, m_derivBias,
1373								   LOG_NOTHING))
1374		{
1375			m_testCtx.getLog()
1376				<< tcu::TestLog::Message
1377				<< "No incorrect derivatives found, result valid."
1378				<< tcu::TestLog::EndMessage;
1379
1380			return true;
1381		}
1382
1383		// some pixels exceed error bounds calculated for normal values. Verify that these
1384		// potentially invalid pixels are in fact valid due to (for example) subnorm flushing.
1385
1386		m_testCtx.getLog()
1387			<< tcu::TestLog::Message
1388			<< "Initial verification failed, verifying image by calculating accurate error bounds for each result pixel.\n"
1389			<< "\tVerifying each result derivative is within its range of legal result values."
1390			<< tcu::TestLog::EndMessage;
1391
1392		{
1393			const tcu::Vec4				valueRamp		= (m_texValueMax - m_texValueMin);
1394			Linear2DFunctionEvaluator	function;
1395
1396			function.matrix.setRow(0, tcu::Vec3(valueRamp.x() / w, 0.0f, m_texValueMin.x()));
1397			function.matrix.setRow(1, tcu::Vec3(0.0f, valueRamp.y() / h, m_texValueMin.y()));
1398			function.matrix.setRow(2, tcu::Vec3(valueRamp.z() / w, valueRamp.z() / h, m_texValueMin.z() + m_texValueMin.z()) / 2.0f);
1399			function.matrix.setRow(3, tcu::Vec3(-valueRamp.w() / w, -valueRamp.w() / h, m_texValueMax.w() + m_texValueMax.w()) / 2.0f);
1400
1401			return reverifyConstantDerivateWithFlushRelaxations(m_testCtx.getLog(), compareArea, maskArea,
1402																m_dataType, m_precision, m_derivScale,
1403																m_derivBias, surfaceThreshold, m_func,
1404																function);
1405		}
1406	}
1407	else
1408	{
1409		DE_ASSERT(m_func == DERIVATE_FWIDTH);
1410		const tcu::Vec4	dx			= ((m_texValueMax - m_texValueMin) / w) * xScale;
1411		const tcu::Vec4	dy			= ((m_texValueMax - m_texValueMin) / h) * yScale;
1412		const tcu::Vec4	reference	= tcu::abs(dx) + tcu::abs(dy);
1413		const tcu::Vec4	dxThreshold	= getDerivateThreshold(m_precision, m_texValueMin*xScale, m_texValueMax*xScale, dx);
1414		const tcu::Vec4	dyThreshold	= getDerivateThreshold(m_precision, m_texValueMin*yScale, m_texValueMax*yScale, dy);
1415		const tcu::Vec4	threshold	= max(surfaceThreshold, max(dxThreshold, dyThreshold));
1416
1417		return verifyConstantDerivate(m_testCtx.getLog(), compareArea, maskArea, m_dataType,
1418									  reference, threshold, m_derivScale, m_derivBias);
1419	}
1420}
1421
1422ShaderDerivateTests::ShaderDerivateTests (Context& context)
1423	: TestCaseGroup(context, "derivate", "Derivate Function Tests")
1424{
1425}
1426
1427ShaderDerivateTests::~ShaderDerivateTests (void)
1428{
1429}
1430
1431struct FunctionSpec
1432{
1433	std::string		name;
1434	DerivateFunc	function;
1435	glu::DataType	dataType;
1436	glu::Precision	precision;
1437
1438	FunctionSpec (const std::string& name_, DerivateFunc function_, glu::DataType dataType_, glu::Precision precision_)
1439		: name		(name_)
1440		, function	(function_)
1441		, dataType	(dataType_)
1442		, precision	(precision_)
1443	{
1444	}
1445};
1446
1447void ShaderDerivateTests::init (void)
1448{
1449	static const struct
1450	{
1451		const char*		name;
1452		const char*		description;
1453		const char*		source;
1454	} s_linearDerivateCases[] =
1455	{
1456		{
1457			"linear",
1458			"Basic derivate of linearly interpolated argument",
1459
1460			"#version 300 es\n"
1461			"in ${PRECISION} ${DATATYPE} v_coord;\n"
1462			"layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1463			"uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1464			"uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1465			"void main (void)\n"
1466			"{\n"
1467			"	${PRECISION} ${DATATYPE} res = ${FUNC}(v_coord) * u_scale + u_bias;\n"
1468			"	o_color = ${CAST_TO_OUTPUT};\n"
1469			"}\n"
1470		},
1471		{
1472			"in_function",
1473			"Derivate of linear function argument",
1474
1475			"#version 300 es\n"
1476			"in ${PRECISION} ${DATATYPE} v_coord;\n"
1477			"layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1478			"uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1479			"uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1480			"\n"
1481			"${PRECISION} ${DATATYPE} computeRes (${PRECISION} ${DATATYPE} value)\n"
1482			"{\n"
1483			"	return ${FUNC}(v_coord) * u_scale + u_bias;\n"
1484			"}\n"
1485			"\n"
1486			"void main (void)\n"
1487			"{\n"
1488			"	${PRECISION} ${DATATYPE} res = computeRes(v_coord);\n"
1489			"	o_color = ${CAST_TO_OUTPUT};\n"
1490			"}\n"
1491		},
1492		{
1493			"static_if",
1494			"Derivate of linearly interpolated value in static if",
1495
1496			"#version 300 es\n"
1497			"in ${PRECISION} ${DATATYPE} v_coord;\n"
1498			"layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1499			"uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1500			"uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1501			"void main (void)\n"
1502			"{\n"
1503			"	${PRECISION} ${DATATYPE} res;\n"
1504			"	if (false)\n"
1505			"		res = ${FUNC}(-v_coord) * u_scale + u_bias;\n"
1506			"	else\n"
1507			"		res = ${FUNC}(v_coord) * u_scale + u_bias;\n"
1508			"	o_color = ${CAST_TO_OUTPUT};\n"
1509			"}\n"
1510		},
1511		{
1512			"static_loop",
1513			"Derivate of linearly interpolated value in static loop",
1514
1515			"#version 300 es\n"
1516			"in ${PRECISION} ${DATATYPE} v_coord;\n"
1517			"layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1518			"uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1519			"uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1520			"void main (void)\n"
1521			"{\n"
1522			"	${PRECISION} ${DATATYPE} res = ${DATATYPE}(0.0);\n"
1523			"	for (int i = 0; i < 2; i++)\n"
1524			"		res += ${FUNC}(v_coord * float(i));\n"
1525			"	res = res * u_scale + u_bias;\n"
1526			"	o_color = ${CAST_TO_OUTPUT};\n"
1527			"}\n"
1528		},
1529		{
1530			"static_switch",
1531			"Derivate of linearly interpolated value in static switch",
1532
1533			"#version 300 es\n"
1534			"in ${PRECISION} ${DATATYPE} v_coord;\n"
1535			"layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1536			"uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1537			"uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1538			"void main (void)\n"
1539			"{\n"
1540			"	${PRECISION} ${DATATYPE} res;\n"
1541			"	switch (1)\n"
1542			"	{\n"
1543			"		case 0:	res = ${FUNC}(-v_coord) * u_scale + u_bias;	break;\n"
1544			"		case 1:	res = ${FUNC}(v_coord) * u_scale + u_bias;	break;\n"
1545			"	}\n"
1546			"	o_color = ${CAST_TO_OUTPUT};\n"
1547			"}\n"
1548		},
1549		{
1550			"uniform_if",
1551			"Derivate of linearly interpolated value in uniform if",
1552
1553			"#version 300 es\n"
1554			"in ${PRECISION} ${DATATYPE} v_coord;\n"
1555			"layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1556			"uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1557			"uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1558			"uniform bool ub_true;\n"
1559			"void main (void)\n"
1560			"{\n"
1561			"	${PRECISION} ${DATATYPE} res;\n"
1562			"	if (ub_true)"
1563			"		res = ${FUNC}(v_coord) * u_scale + u_bias;\n"
1564			"	else\n"
1565			"		res = ${FUNC}(-v_coord) * u_scale + u_bias;\n"
1566			"	o_color = ${CAST_TO_OUTPUT};\n"
1567			"}\n"
1568		},
1569		{
1570			"uniform_loop",
1571			"Derivate of linearly interpolated value in uniform loop",
1572
1573			"#version 300 es\n"
1574			"in ${PRECISION} ${DATATYPE} v_coord;\n"
1575			"layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1576			"uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1577			"uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1578			"uniform int ui_two;\n"
1579			"void main (void)\n"
1580			"{\n"
1581			"	${PRECISION} ${DATATYPE} res = ${DATATYPE}(0.0);\n"
1582			"	for (int i = 0; i < ui_two; i++)\n"
1583			"		res += ${FUNC}(v_coord * float(i));\n"
1584			"	res = res * u_scale + u_bias;\n"
1585			"	o_color = ${CAST_TO_OUTPUT};\n"
1586			"}\n"
1587		},
1588		{
1589			"uniform_switch",
1590			"Derivate of linearly interpolated value in uniform switch",
1591
1592			"#version 300 es\n"
1593			"in ${PRECISION} ${DATATYPE} v_coord;\n"
1594			"layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1595			"uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1596			"uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1597			"uniform int ui_one;\n"
1598			"void main (void)\n"
1599			"{\n"
1600			"	${PRECISION} ${DATATYPE} res;\n"
1601			"	switch (ui_one)\n"
1602			"	{\n"
1603			"		case 0:	res = ${FUNC}(-v_coord) * u_scale + u_bias;	break;\n"
1604			"		case 1:	res = ${FUNC}(v_coord) * u_scale + u_bias;	break;\n"
1605			"	}\n"
1606			"	o_color = ${CAST_TO_OUTPUT};\n"
1607			"}\n"
1608		},
1609	};
1610
1611	static const struct
1612	{
1613		const char*		name;
1614		SurfaceType		surfaceType;
1615		int				numSamples;
1616	} s_fboConfigs[] =
1617	{
1618		{ "fbo",		SURFACETYPE_DEFAULT_FRAMEBUFFER,	0 },
1619		{ "fbo_msaa2",	SURFACETYPE_UNORM_FBO,				2 },
1620		{ "fbo_msaa4",	SURFACETYPE_UNORM_FBO,				4 },
1621		{ "fbo_float",	SURFACETYPE_FLOAT_FBO,				0 },
1622	};
1623
1624	static const struct
1625	{
1626		const char*		name;
1627		deUint32		hint;
1628	} s_hints[] =
1629	{
1630		{ "fastest",	GL_FASTEST	},
1631		{ "nicest",		GL_NICEST	},
1632	};
1633
1634	static const struct
1635	{
1636		const char*		name;
1637		SurfaceType		surfaceType;
1638		int				numSamples;
1639	} s_hintFboConfigs[] =
1640	{
1641		{ "default",		SURFACETYPE_DEFAULT_FRAMEBUFFER,	0 },
1642		{ "fbo_msaa4",		SURFACETYPE_UNORM_FBO,				4 },
1643		{ "fbo_float",		SURFACETYPE_FLOAT_FBO,				0 }
1644	};
1645
1646	static const struct
1647	{
1648		const char*		name;
1649		SurfaceType		surfaceType;
1650		int				numSamples;
1651		deUint32		hint;
1652	} s_textureConfigs[] =
1653	{
1654		{ "basic",			SURFACETYPE_DEFAULT_FRAMEBUFFER,	0,	GL_DONT_CARE	},
1655		{ "msaa4",			SURFACETYPE_UNORM_FBO,				4,	GL_DONT_CARE	},
1656		{ "float_fastest",	SURFACETYPE_FLOAT_FBO,				0,	GL_FASTEST		},
1657		{ "float_nicest",	SURFACETYPE_FLOAT_FBO,				0,	GL_NICEST		},
1658	};
1659
1660	// .dfdx, .dfdy, .fwidth
1661	for (int funcNdx = 0; funcNdx < DERIVATE_LAST; funcNdx++)
1662	{
1663		const DerivateFunc			function		= DerivateFunc(funcNdx);
1664		tcu::TestCaseGroup* const	functionGroup	= new tcu::TestCaseGroup(m_testCtx, getDerivateFuncCaseName(function), getDerivateFuncName(function));
1665		addChild(functionGroup);
1666
1667		// .constant - no precision variants, checks that derivate of constant arguments is 0
1668		{
1669			tcu::TestCaseGroup* const constantGroup = new tcu::TestCaseGroup(m_testCtx, "constant", "Derivate of constant argument");
1670			functionGroup->addChild(constantGroup);
1671
1672			for (int vecSize = 1; vecSize <= 4; vecSize++)
1673			{
1674				const glu::DataType dataType = vecSize > 1 ? glu::getDataTypeFloatVec(vecSize) : glu::TYPE_FLOAT;
1675				constantGroup->addChild(new ConstantDerivateCase(m_context, glu::getDataTypeName(dataType), "", function, dataType));
1676			}
1677		}
1678
1679		// Cases based on LinearDerivateCase
1680		for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(s_linearDerivateCases); caseNdx++)
1681		{
1682			tcu::TestCaseGroup* const linearCaseGroup	= new tcu::TestCaseGroup(m_testCtx, s_linearDerivateCases[caseNdx].name, s_linearDerivateCases[caseNdx].description);
1683			const char*			source					= s_linearDerivateCases[caseNdx].source;
1684			functionGroup->addChild(linearCaseGroup);
1685
1686			for (int vecSize = 1; vecSize <= 4; vecSize++)
1687			{
1688				for (int precNdx = 0; precNdx < glu::PRECISION_LAST; precNdx++)
1689				{
1690					const glu::DataType		dataType		= vecSize > 1 ? glu::getDataTypeFloatVec(vecSize) : glu::TYPE_FLOAT;
1691					const glu::Precision	precision		= glu::Precision(precNdx);
1692					const SurfaceType		surfaceType		= SURFACETYPE_DEFAULT_FRAMEBUFFER;
1693					const int				numSamples		= 0;
1694					const deUint32			hint			= GL_DONT_CARE;
1695					ostringstream			caseName;
1696
1697					if (caseNdx != 0 && precision == glu::PRECISION_LOWP)
1698						continue; // Skip as lowp doesn't actually produce any bits when rendered to default FB.
1699
1700					caseName << glu::getDataTypeName(dataType) << "_" << glu::getPrecisionName(precision);
1701
1702					linearCaseGroup->addChild(new LinearDerivateCase(m_context, caseName.str().c_str(), "", function, dataType, precision, hint, surfaceType, numSamples, source));
1703				}
1704			}
1705		}
1706
1707		// Fbo cases
1708		for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(s_fboConfigs); caseNdx++)
1709		{
1710			tcu::TestCaseGroup*	const	fboGroup		= new tcu::TestCaseGroup(m_testCtx, s_fboConfigs[caseNdx].name, "Derivate usage when rendering into FBO");
1711			const char*					source			= s_linearDerivateCases[0].source; // use source from .linear group
1712			const SurfaceType			surfaceType		= s_fboConfigs[caseNdx].surfaceType;
1713			const int					numSamples		= s_fboConfigs[caseNdx].numSamples;
1714			functionGroup->addChild(fboGroup);
1715
1716			for (int vecSize = 1; vecSize <= 4; vecSize++)
1717			{
1718				for (int precNdx = 0; precNdx < glu::PRECISION_LAST; precNdx++)
1719				{
1720					const glu::DataType		dataType		= vecSize > 1 ? glu::getDataTypeFloatVec(vecSize) : glu::TYPE_FLOAT;
1721					const glu::Precision	precision		= glu::Precision(precNdx);
1722					const deUint32			hint			= GL_DONT_CARE;
1723					ostringstream			caseName;
1724
1725					if (surfaceType != SURFACETYPE_FLOAT_FBO && precision == glu::PRECISION_LOWP)
1726						continue; // Skip as lowp doesn't actually produce any bits when rendered to U8 RT.
1727
1728					caseName << glu::getDataTypeName(dataType) << "_" << glu::getPrecisionName(precision);
1729
1730					fboGroup->addChild(new LinearDerivateCase(m_context, caseName.str().c_str(), "", function, dataType, precision, hint, surfaceType, numSamples, source));
1731				}
1732			}
1733		}
1734
1735		// .fastest, .nicest
1736		for (int hintCaseNdx = 0; hintCaseNdx < DE_LENGTH_OF_ARRAY(s_hints); hintCaseNdx++)
1737		{
1738			tcu::TestCaseGroup* const	hintGroup		= new tcu::TestCaseGroup(m_testCtx, s_hints[hintCaseNdx].name, "Shader derivate hints");
1739			const char*					source			= s_linearDerivateCases[0].source; // use source from .linear group
1740			const deUint32				hint			= s_hints[hintCaseNdx].hint;
1741			functionGroup->addChild(hintGroup);
1742
1743			for (int fboCaseNdx = 0; fboCaseNdx < DE_LENGTH_OF_ARRAY(s_hintFboConfigs); fboCaseNdx++)
1744			{
1745				tcu::TestCaseGroup*	const	fboGroup		= new tcu::TestCaseGroup(m_testCtx, s_hintFboConfigs[fboCaseNdx].name, "");
1746				const SurfaceType			surfaceType		= s_hintFboConfigs[fboCaseNdx].surfaceType;
1747				const int					numSamples		= s_hintFboConfigs[fboCaseNdx].numSamples;
1748				hintGroup->addChild(fboGroup);
1749
1750				for (int vecSize = 1; vecSize <= 4; vecSize++)
1751				{
1752					for (int precNdx = 0; precNdx < glu::PRECISION_LAST; precNdx++)
1753					{
1754						const glu::DataType		dataType		= vecSize > 1 ? glu::getDataTypeFloatVec(vecSize) : glu::TYPE_FLOAT;
1755						const glu::Precision	precision		= glu::Precision(precNdx);
1756						ostringstream			caseName;
1757
1758						if (surfaceType != SURFACETYPE_FLOAT_FBO && precision == glu::PRECISION_LOWP)
1759							continue; // Skip as lowp doesn't actually produce any bits when rendered to U8 RT.
1760
1761						caseName << glu::getDataTypeName(dataType) << "_" << glu::getPrecisionName(precision);
1762
1763						fboGroup->addChild(new LinearDerivateCase(m_context, caseName.str().c_str(), "", function, dataType, precision, hint, surfaceType, numSamples, source));
1764					}
1765				}
1766			}
1767		}
1768
1769		// .texture
1770		{
1771			tcu::TestCaseGroup* const textureGroup = new tcu::TestCaseGroup(m_testCtx, "texture", "Derivate of texture lookup result");
1772			functionGroup->addChild(textureGroup);
1773
1774			for (int texCaseNdx = 0; texCaseNdx < DE_LENGTH_OF_ARRAY(s_textureConfigs); texCaseNdx++)
1775			{
1776				tcu::TestCaseGroup*	const	caseGroup		= new tcu::TestCaseGroup(m_testCtx, s_textureConfigs[texCaseNdx].name, "");
1777				const SurfaceType			surfaceType		= s_textureConfigs[texCaseNdx].surfaceType;
1778				const int					numSamples		= s_textureConfigs[texCaseNdx].numSamples;
1779				const deUint32				hint			= s_textureConfigs[texCaseNdx].hint;
1780				textureGroup->addChild(caseGroup);
1781
1782				for (int vecSize = 1; vecSize <= 4; vecSize++)
1783				{
1784					for (int precNdx = 0; precNdx < glu::PRECISION_LAST; precNdx++)
1785					{
1786						const glu::DataType		dataType		= vecSize > 1 ? glu::getDataTypeFloatVec(vecSize) : glu::TYPE_FLOAT;
1787						const glu::Precision	precision		= glu::Precision(precNdx);
1788						ostringstream			caseName;
1789
1790						if (surfaceType != SURFACETYPE_FLOAT_FBO && precision == glu::PRECISION_LOWP)
1791							continue; // Skip as lowp doesn't actually produce any bits when rendered to U8 RT.
1792
1793						caseName << glu::getDataTypeName(dataType) << "_" << glu::getPrecisionName(precision);
1794
1795						caseGroup->addChild(new TextureDerivateCase(m_context, caseName.str().c_str(), "", function, dataType, precision, hint, surfaceType, numSamples));
1796					}
1797				}
1798			}
1799		}
1800	}
1801}
1802
1803} // Functional
1804} // gles3
1805} // deqp
1806