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
57 namespace deqp
58 {
59 namespace gles3
60 {
61 namespace Functional
62 {
63
64 using std::vector;
65 using std::string;
66 using std::map;
67 using tcu::TestLog;
68 using std::ostringstream;
69
70 enum
71 {
72 VIEWPORT_WIDTH = 167,
73 VIEWPORT_HEIGHT = 103,
74 FBO_WIDTH = 99,
75 FBO_HEIGHT = 133,
76 MAX_FAILED_MESSAGES = 10
77 };
78
79 enum DerivateFunc
80 {
81 DERIVATE_DFDX = 0,
82 DERIVATE_DFDY,
83 DERIVATE_FWIDTH,
84
85 DERIVATE_LAST
86 };
87
88 enum 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
99 namespace
100 {
101
102 class AutoFbo
103 {
104 public:
AutoFbo(const glw::Functions & gl)105 AutoFbo (const glw::Functions& gl)
106 : m_gl (gl)
107 , m_fbo (0)
108 {
109 }
110
~AutoFbo(void)111 ~AutoFbo (void)
112 {
113 if (m_fbo)
114 m_gl.deleteFramebuffers(1, &m_fbo);
115 }
116
gen(void)117 void gen (void)
118 {
119 DE_ASSERT(!m_fbo);
120 m_gl.genFramebuffers(1, &m_fbo);
121 }
122
operator *(void) const123 deUint32 operator* (void) const { return m_fbo; }
124
125 private:
126 const glw::Functions& m_gl;
127 deUint32 m_fbo;
128 };
129
130 class AutoRbo
131 {
132 public:
AutoRbo(const glw::Functions & gl)133 AutoRbo (const glw::Functions& gl)
134 : m_gl (gl)
135 , m_rbo (0)
136 {
137 }
138
~AutoRbo(void)139 ~AutoRbo (void)
140 {
141 if (m_rbo)
142 m_gl.deleteRenderbuffers(1, &m_rbo);
143 }
144
gen(void)145 void gen (void)
146 {
147 DE_ASSERT(!m_rbo);
148 m_gl.genRenderbuffers(1, &m_rbo);
149 }
150
operator *(void) const151 deUint32 operator* (void) const { return m_rbo; }
152
153 private:
154 const glw::Functions& m_gl;
155 deUint32 m_rbo;
156 };
157
158 } // anonymous
159
getDerivateFuncName(DerivateFunc func)160 static 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
getDerivateFuncCaseName(DerivateFunc func)173 static 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
getDerivateMask(glu::DataType type)186 static 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
readDerivate(const tcu::ConstPixelBufferAccess & surface,const tcu::Vec4 & derivScale,const tcu::Vec4 & derivBias,int x,int y)200 static 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
getCompExpBits(const tcu::Vec4 & v)205 static 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
computeFloatingPointError(const float value,const int numAccurateBits)213 float 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() < -3) ? -3 : 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
getNumMantissaBits(const glu::Precision precision)222 static 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
getMinExponent(const glu::Precision precision)235 static 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
getSingleULPForExponent(int exp,int numMantissaBits)248 static 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
getSingleULPForValue(float value,int numMantissaBits)264 static float getSingleULPForValue (float value, int numMantissaBits)
265 {
266 const int exp = tcu::Float32(value).exponent();
267 return getSingleULPForExponent(exp, numMantissaBits);
268 }
269
convertFloatFlushToZeroRtn(float value,int minExponent,int numAccurateBits)270 static 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
convertFloatFlushToZeroRtp(float value,int minExponent,int numAccurateBits)311 static float convertFloatFlushToZeroRtp (float value, int minExponent, int numAccurateBits)
312 {
313 return -convertFloatFlushToZeroRtn(-value, minExponent, numAccurateBits);
314 }
315
addErrorUlp(float value,float numUlps,int numMantissaBits)316 static float addErrorUlp (float value, float numUlps, int numMantissaBits)
317 {
318 return value + numUlps * getSingleULPForValue(value, numMantissaBits);
319 }
320
321 enum
322 {
323 INTERPOLATION_LOST_BITS = 3, // number mantissa of bits allowed to be lost in varying interpolation
324 };
325
getInterpolationLostBitsWarning(const glu::Precision precision)326 static int getInterpolationLostBitsWarning (const glu::Precision precision)
327 {
328 // number mantissa of bits allowed to be lost in varying interpolation
329 switch (precision)
330 {
331 case glu::PRECISION_HIGHP: return 9;
332 case glu::PRECISION_MEDIUMP: return 3;
333 case glu::PRECISION_LOWP: return 3;
334 default:
335 DE_ASSERT(false);
336 return 0;
337 }
338 }
339
getDerivateThreshold(const glu::Precision precision,const tcu::Vec4 & valueMin,const tcu::Vec4 & valueMax,const tcu::Vec4 & expectedDerivate)340 static inline tcu::Vec4 getDerivateThreshold (const glu::Precision precision, const tcu::Vec4& valueMin, const tcu::Vec4& valueMax, const tcu::Vec4& expectedDerivate)
341 {
342 const int baseBits = getNumMantissaBits(precision);
343 const tcu::UVec4 derivExp = getCompExpBits(expectedDerivate);
344 const tcu::UVec4 maxValueExp = max(getCompExpBits(valueMin), getCompExpBits(valueMax));
345 const tcu::UVec4 numBitsLost = maxValueExp - min(maxValueExp, derivExp);
346 const tcu::IVec4 numAccurateBits = max(baseBits - numBitsLost.asInt() - (int)INTERPOLATION_LOST_BITS, tcu::IVec4(0));
347
348 return tcu::Vec4(computeFloatingPointError(expectedDerivate[0], numAccurateBits[0]),
349 computeFloatingPointError(expectedDerivate[1], numAccurateBits[1]),
350 computeFloatingPointError(expectedDerivate[2], numAccurateBits[2]),
351 computeFloatingPointError(expectedDerivate[3], numAccurateBits[3]));
352 }
353
getDerivateThresholdWarning(const glu::Precision precision,const tcu::Vec4 & valueMin,const tcu::Vec4 & valueMax,const tcu::Vec4 & expectedDerivate)354 static inline tcu::Vec4 getDerivateThresholdWarning (const glu::Precision precision, const tcu::Vec4& valueMin, const tcu::Vec4& valueMax, const tcu::Vec4& expectedDerivate)
355 {
356 const int baseBits = getNumMantissaBits(precision);
357 const tcu::UVec4 derivExp = getCompExpBits(expectedDerivate);
358 const tcu::UVec4 maxValueExp = max(getCompExpBits(valueMin), getCompExpBits(valueMax));
359 const tcu::UVec4 numBitsLost = maxValueExp - min(maxValueExp, derivExp);
360 const tcu::IVec4 numAccurateBits = max(baseBits - numBitsLost.asInt() - getInterpolationLostBitsWarning(precision), tcu::IVec4(0));
361
362 return tcu::Vec4(computeFloatingPointError(expectedDerivate[0], numAccurateBits[0]),
363 computeFloatingPointError(expectedDerivate[1], numAccurateBits[1]),
364 computeFloatingPointError(expectedDerivate[2], numAccurateBits[2]),
365 computeFloatingPointError(expectedDerivate[3], numAccurateBits[3]));
366 }
367
368
369 namespace
370 {
371
372 struct LogVecComps
373 {
374 const tcu::Vec4& v;
375 int numComps;
376
LogVecCompsdeqp::gles3::Functional::__anon53c5d24a0411::LogVecComps377 LogVecComps (const tcu::Vec4& v_, int numComps_)
378 : v (v_)
379 , numComps (numComps_)
380 {
381 }
382 };
383
operator <<(std::ostream & str,const LogVecComps & v)384 std::ostream& operator<< (std::ostream& str, const LogVecComps& v)
385 {
386 DE_ASSERT(de::inRange(v.numComps, 1, 4));
387 if (v.numComps == 1) return str << v.v[0];
388 else if (v.numComps == 2) return str << v.v.toWidth<2>();
389 else if (v.numComps == 3) return str << v.v.toWidth<3>();
390 else return str << v.v;
391 }
392
393 } // anonymous
394
395 enum VerificationLogging
396 {
397 LOG_ALL = 0,
398 LOG_NOTHING
399 };
400
verifyConstantDerivate(tcu::TestLog & log,const tcu::ConstPixelBufferAccess & result,const tcu::PixelBufferAccess & errorMask,glu::DataType dataType,const tcu::Vec4 & reference,const tcu::Vec4 & threshold,const tcu::Vec4 & scale,const tcu::Vec4 & bias,VerificationLogging logPolicy=LOG_ALL)401 static qpTestResult verifyConstantDerivate (tcu::TestLog& log,
402 const tcu::ConstPixelBufferAccess& result,
403 const tcu::PixelBufferAccess& errorMask,
404 glu::DataType dataType,
405 const tcu::Vec4& reference,
406 const tcu::Vec4& threshold,
407 const tcu::Vec4& scale,
408 const tcu::Vec4& bias,
409 VerificationLogging logPolicy = LOG_ALL)
410 {
411 const int numComps = glu::getDataTypeFloatScalars(dataType);
412 const tcu::BVec4 mask = tcu::logicalNot(getDerivateMask(dataType));
413 int numFailedPixels = 0;
414
415 if (logPolicy == LOG_ALL)
416 log << TestLog::Message << "Expecting " << LogVecComps(reference, numComps) << " with threshold " << LogVecComps(threshold, numComps) << TestLog::EndMessage;
417
418 for (int y = 0; y < result.getHeight(); y++)
419 {
420 for (int x = 0; x < result.getWidth(); x++)
421 {
422 const tcu::Vec4 resDerivate = readDerivate(result, scale, bias, x, y);
423 const bool isOk = tcu::allEqual(tcu::logicalOr(tcu::lessThanEqual(tcu::abs(reference - resDerivate), threshold), mask), tcu::BVec4(true));
424
425 if (!isOk)
426 {
427 if (numFailedPixels < MAX_FAILED_MESSAGES && logPolicy == LOG_ALL)
428 log << TestLog::Message << "FAIL: got " << LogVecComps(resDerivate, numComps)
429 << ", diff = " << LogVecComps(tcu::abs(reference - resDerivate), numComps)
430 << ", at x = " << x << ", y = " << y
431 << TestLog::EndMessage;
432 numFailedPixels += 1;
433 errorMask.setPixel(tcu::RGBA::red().toVec(), x, y);
434 }
435 }
436 }
437
438 if (numFailedPixels >= MAX_FAILED_MESSAGES && logPolicy == LOG_ALL)
439 log << TestLog::Message << "..." << TestLog::EndMessage;
440
441 if (numFailedPixels > 0 && logPolicy == LOG_ALL)
442 log << TestLog::Message << "FAIL: found " << numFailedPixels << " failed pixels" << TestLog::EndMessage;
443
444 return (numFailedPixels == 0) ? QP_TEST_RESULT_PASS : QP_TEST_RESULT_FAIL;
445 }
446
447 struct Linear2DFunctionEvaluator
448 {
449 tcu::Matrix<float, 4, 3> matrix;
450
451 // .-----.
452 // | s_x |
453 // M x | s_y |
454 // | 1.0 |
455 // '-----'
456 tcu::Vec4 evaluateAt (float screenX, float screenY) const;
457 };
458
evaluateAt(float screenX,float screenY) const459 tcu::Vec4 Linear2DFunctionEvaluator::evaluateAt (float screenX, float screenY) const
460 {
461 const tcu::Vec3 position(screenX, screenY, 1.0f);
462 return matrix * position;
463 }
464
reverifyConstantDerivateWithFlushRelaxations(tcu::TestLog & log,const tcu::ConstPixelBufferAccess & result,const tcu::PixelBufferAccess & errorMask,glu::DataType dataType,glu::Precision precision,const tcu::Vec4 & derivScale,const tcu::Vec4 & derivBias,const tcu::Vec4 & surfaceThreshold,DerivateFunc derivateFunc,const Linear2DFunctionEvaluator & function)465 static qpTestResult reverifyConstantDerivateWithFlushRelaxations (tcu::TestLog& log,
466 const tcu::ConstPixelBufferAccess& result,
467 const tcu::PixelBufferAccess& errorMask,
468 glu::DataType dataType,
469 glu::Precision precision,
470 const tcu::Vec4& derivScale,
471 const tcu::Vec4& derivBias,
472 const tcu::Vec4& surfaceThreshold,
473 DerivateFunc derivateFunc,
474 const Linear2DFunctionEvaluator& function)
475 {
476 DE_ASSERT(result.getWidth() == errorMask.getWidth());
477 DE_ASSERT(result.getHeight() == errorMask.getHeight());
478 DE_ASSERT(derivateFunc == DERIVATE_DFDX || derivateFunc == DERIVATE_DFDY);
479
480 const tcu::IVec4 red (255, 0, 0, 255);
481 const tcu::IVec4 green (0, 255, 0, 255);
482 const float divisionErrorUlps = 2.5f;
483
484 const int numComponents = glu::getDataTypeFloatScalars(dataType);
485 const int numBits = getNumMantissaBits(precision);
486 const int minExponent = getMinExponent(precision);
487
488 const int numVaryingSampleBits = numBits - INTERPOLATION_LOST_BITS;
489 int numFailedPixels = 0;
490
491 tcu::clear(errorMask, green);
492
493 // search for failed pixels
494 for (int y = 0; y < result.getHeight(); ++y)
495 for (int x = 0; x < result.getWidth(); ++x)
496 {
497 // flushToZero?(f2z?(functionValueCurrent) - f2z?(functionValueBefore))
498 // flushToZero? ( ------------------------------------------------------------------------ +- 2.5 ULP )
499 // dx
500
501 const tcu::Vec4 resultDerivative = readDerivate(result, derivScale, derivBias, x, y);
502
503 // sample at the front of the back pixel and the back of the front pixel to cover the whole area of
504 // legal sample positions. In general case this is NOT OK, but we know that the target funtion is
505 // (mostly*) linear which allows us to take the sample points at arbitrary points. This gets us the
506 // maximum difference possible in exponents which are used in error bound calculations.
507 // * non-linearity may happen around zero or with very high function values due to subnorms not
508 // behaving well.
509 const tcu::Vec4 functionValueForward = (derivateFunc == DERIVATE_DFDX)
510 ? (function.evaluateAt((float)x + 2.0f, (float)y + 0.5f))
511 : (function.evaluateAt((float)x + 0.5f, (float)y + 2.0f));
512 const tcu::Vec4 functionValueBackward = (derivateFunc == DERIVATE_DFDX)
513 ? (function.evaluateAt((float)x - 1.0f, (float)y + 0.5f))
514 : (function.evaluateAt((float)x + 0.5f, (float)y - 1.0f));
515
516 bool anyComponentFailed = false;
517
518 // check components separately
519 for (int c = 0; c < numComponents; ++c)
520 {
521 // Simulate interpolation. Add allowed interpolation error and round to target precision. Allow one half ULP (i.e. correct rounding)
522 const tcu::Interval forwardComponent (convertFloatFlushToZeroRtn(addErrorUlp((float)functionValueForward[c], -0.5f, numVaryingSampleBits), minExponent, numBits),
523 convertFloatFlushToZeroRtp(addErrorUlp((float)functionValueForward[c], +0.5f, numVaryingSampleBits), minExponent, numBits));
524 const tcu::Interval backwardComponent (convertFloatFlushToZeroRtn(addErrorUlp((float)functionValueBackward[c], -0.5f, numVaryingSampleBits), minExponent, numBits),
525 convertFloatFlushToZeroRtp(addErrorUlp((float)functionValueBackward[c], +0.5f, numVaryingSampleBits), minExponent, numBits));
526 const int maxValueExp = de::max(de::max(tcu::Float32(forwardComponent.lo()).exponent(), tcu::Float32(forwardComponent.hi()).exponent()),
527 de::max(tcu::Float32(backwardComponent.lo()).exponent(), tcu::Float32(backwardComponent.hi()).exponent()));
528
529 // subtraction in numerator will likely cause a cancellation of the most
530 // significant bits. Apply error bounds.
531
532 const tcu::Interval numerator (forwardComponent - backwardComponent);
533 const int numeratorLoExp = tcu::Float32(numerator.lo()).exponent();
534 const int numeratorHiExp = tcu::Float32(numerator.hi()).exponent();
535 const int numeratorLoBitsLost = de::max(0, maxValueExp - numeratorLoExp); //!< must clamp to zero since if forward and backward components have different
536 const int numeratorHiBitsLost = de::max(0, maxValueExp - numeratorHiExp); //!< sign, numerator might have larger exponent than its operands.
537 const int numeratorLoBits = de::max(0, numBits - numeratorLoBitsLost);
538 const int numeratorHiBits = de::max(0, numBits - numeratorHiBitsLost);
539
540 const tcu::Interval numeratorRange (convertFloatFlushToZeroRtn((float)numerator.lo(), minExponent, numeratorLoBits),
541 convertFloatFlushToZeroRtp((float)numerator.hi(), minExponent, numeratorHiBits));
542
543 const tcu::Interval divisionRange = numeratorRange / 3.0f; // legal sample area is anywhere within this and neighboring pixels (i.e. size = 3)
544 const tcu::Interval divisionResultRange (convertFloatFlushToZeroRtn(addErrorUlp((float)divisionRange.lo(), -divisionErrorUlps, numBits), minExponent, numBits),
545 convertFloatFlushToZeroRtp(addErrorUlp((float)divisionRange.hi(), +divisionErrorUlps, numBits), minExponent, numBits));
546 const tcu::Interval finalResultRange (divisionResultRange.lo() - surfaceThreshold[c], divisionResultRange.hi() + surfaceThreshold[c]);
547
548 if (resultDerivative[c] >= finalResultRange.lo() && resultDerivative[c] <= finalResultRange.hi())
549 {
550 // value ok
551 }
552 else
553 {
554 if (numFailedPixels < MAX_FAILED_MESSAGES)
555 log << tcu::TestLog::Message
556 << "Error in pixel at " << x << ", " << y << " with component " << c << " (channel " << ("rgba"[c]) << ")\n"
557 << "\tGot pixel value " << result.getPixelInt(x, y) << "\n"
558 << "\t\tdFd" << ((derivateFunc == DERIVATE_DFDX) ? ('x') : ('y')) << " ~= " << resultDerivative[c] << "\n"
559 << "\t\tdifference to a valid range: "
560 << ((resultDerivative[c] < finalResultRange.lo()) ? ("-") : ("+"))
561 << ((resultDerivative[c] < finalResultRange.lo()) ? (finalResultRange.lo() - resultDerivative[c]) : (resultDerivative[c] - finalResultRange.hi()))
562 << "\n"
563 << "\tDerivative value range:\n"
564 << "\t\tMin: " << finalResultRange.lo() << "\n"
565 << "\t\tMax: " << finalResultRange.hi() << "\n"
566 << tcu::TestLog::EndMessage;
567
568 ++numFailedPixels;
569 anyComponentFailed = true;
570 }
571 }
572
573 if (anyComponentFailed)
574 errorMask.setPixel(red, x, y);
575 }
576
577 if (numFailedPixels >= MAX_FAILED_MESSAGES)
578 log << TestLog::Message << "..." << TestLog::EndMessage;
579
580 if (numFailedPixels > 0)
581 log << TestLog::Message << "FAIL: found " << numFailedPixels << " failed pixels" << TestLog::EndMessage;
582
583 return (numFailedPixels == 0) ? QP_TEST_RESULT_PASS : QP_TEST_RESULT_FAIL;
584 }
585
586 // TriangleDerivateCase
587
588 class TriangleDerivateCase : public TestCase
589 {
590 public:
591 TriangleDerivateCase (Context& context, const char* name, const char* description);
592 ~TriangleDerivateCase (void);
593
594 IterateResult iterate (void);
595
596 protected:
setupRenderState(deUint32 program)597 virtual void setupRenderState (deUint32 program) { DE_UNREF(program); }
598 virtual qpTestResult verify (const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask) = DE_NULL;
599
600 tcu::IVec2 getViewportSize (void) const;
601 tcu::Vec4 getSurfaceThreshold (void) const;
602
603 glu::DataType m_dataType;
604 glu::Precision m_precision;
605
606 glu::DataType m_coordDataType;
607 glu::Precision m_coordPrecision;
608
609 std::string m_fragmentSrc;
610
611 tcu::Vec4 m_coordMin;
612 tcu::Vec4 m_coordMax;
613 tcu::Vec4 m_derivScale;
614 tcu::Vec4 m_derivBias;
615
616 SurfaceType m_surfaceType;
617 int m_numSamples;
618 deUint32 m_hint;
619 };
620
TriangleDerivateCase(Context & context,const char * name,const char * description)621 TriangleDerivateCase::TriangleDerivateCase (Context& context, const char* name, const char* description)
622 : TestCase (context, name, description)
623 , m_dataType (glu::TYPE_LAST)
624 , m_precision (glu::PRECISION_LAST)
625 , m_coordDataType (glu::TYPE_LAST)
626 , m_coordPrecision (glu::PRECISION_LAST)
627 , m_surfaceType (SURFACETYPE_DEFAULT_FRAMEBUFFER)
628 , m_numSamples (0)
629 , m_hint (GL_DONT_CARE)
630 {
631 DE_ASSERT(m_surfaceType != SURFACETYPE_DEFAULT_FRAMEBUFFER || m_numSamples == 0);
632 }
633
~TriangleDerivateCase(void)634 TriangleDerivateCase::~TriangleDerivateCase (void)
635 {
636 TriangleDerivateCase::deinit();
637 }
638
genVertexSource(glu::DataType coordType,glu::Precision precision)639 static std::string genVertexSource (glu::DataType coordType, glu::Precision precision)
640 {
641 DE_ASSERT(glu::isDataTypeFloatOrVec(coordType));
642
643 const char* vertexTmpl =
644 "#version 300 es\n"
645 "in highp vec4 a_position;\n"
646 "in ${PRECISION} ${DATATYPE} a_coord;\n"
647 "out ${PRECISION} ${DATATYPE} v_coord;\n"
648 "void main (void)\n"
649 "{\n"
650 " gl_Position = a_position;\n"
651 " v_coord = a_coord;\n"
652 "}\n";
653
654 map<string, string> vertexParams;
655
656 vertexParams["PRECISION"] = glu::getPrecisionName(precision);
657 vertexParams["DATATYPE"] = glu::getDataTypeName(coordType);
658
659 return tcu::StringTemplate(vertexTmpl).specialize(vertexParams);
660 }
661
getViewportSize(void) const662 inline tcu::IVec2 TriangleDerivateCase::getViewportSize (void) const
663 {
664 if (m_surfaceType == SURFACETYPE_DEFAULT_FRAMEBUFFER)
665 {
666 const int width = de::min<int>(m_context.getRenderTarget().getWidth(), VIEWPORT_WIDTH);
667 const int height = de::min<int>(m_context.getRenderTarget().getHeight(), VIEWPORT_HEIGHT);
668 return tcu::IVec2(width, height);
669 }
670 else
671 return tcu::IVec2(FBO_WIDTH, FBO_HEIGHT);
672 }
673
iterate(void)674 TriangleDerivateCase::IterateResult TriangleDerivateCase::iterate (void)
675 {
676 const glw::Functions& gl = m_context.getRenderContext().getFunctions();
677 const glu::ShaderProgram program (m_context.getRenderContext(), glu::makeVtxFragSources(genVertexSource(m_coordDataType, m_coordPrecision), m_fragmentSrc));
678 de::Random rnd (deStringHash(getName()) ^ 0xbbc24);
679 const bool useFbo = m_surfaceType != SURFACETYPE_DEFAULT_FRAMEBUFFER;
680 const deUint32 fboFormat = m_surfaceType == SURFACETYPE_FLOAT_FBO ? GL_RGBA32UI : GL_RGBA8;
681 const tcu::IVec2 viewportSize = getViewportSize();
682 const int viewportX = useFbo ? 0 : rnd.getInt(0, m_context.getRenderTarget().getWidth() - viewportSize.x());
683 const int viewportY = useFbo ? 0 : rnd.getInt(0, m_context.getRenderTarget().getHeight() - viewportSize.y());
684 AutoFbo fbo (gl);
685 AutoRbo rbo (gl);
686 tcu::TextureLevel result;
687
688 m_testCtx.getLog() << program;
689
690 if (!program.isOk())
691 TCU_FAIL("Compile failed");
692
693 if (useFbo)
694 {
695 m_testCtx.getLog() << TestLog::Message
696 << "Rendering to FBO, format = " << glu::getTextureFormatStr(fboFormat)
697 << ", samples = " << m_numSamples
698 << TestLog::EndMessage;
699
700 fbo.gen();
701 rbo.gen();
702
703 gl.bindRenderbuffer(GL_RENDERBUFFER, *rbo);
704 gl.renderbufferStorageMultisample(GL_RENDERBUFFER, m_numSamples, fboFormat, viewportSize.x(), viewportSize.y());
705 gl.bindFramebuffer(GL_FRAMEBUFFER, *fbo);
706 gl.framebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, *rbo);
707 TCU_CHECK(gl.checkFramebufferStatus(GL_FRAMEBUFFER) == GL_FRAMEBUFFER_COMPLETE);
708 }
709 else
710 {
711 const tcu::PixelFormat pixelFormat = m_context.getRenderTarget().getPixelFormat();
712
713 m_testCtx.getLog()
714 << TestLog::Message
715 << "Rendering to default framebuffer\n"
716 << "\tColor depth: R=" << pixelFormat.redBits << ", G=" << pixelFormat.greenBits << ", B=" << pixelFormat.blueBits << ", A=" << pixelFormat.alphaBits
717 << TestLog::EndMessage;
718 }
719
720 m_testCtx.getLog() << TestLog::Message << "in: " << m_coordMin << " -> " << m_coordMax << "\n"
721 << "v_coord.x = in.x * x\n"
722 << "v_coord.y = in.y * y\n"
723 << "v_coord.z = in.z * (x+y)/2\n"
724 << "v_coord.w = in.w * (1 - (x+y)/2)\n"
725 << TestLog::EndMessage
726 << TestLog::Message << "u_scale: " << m_derivScale << ", u_bias: " << m_derivBias << " (displayed values have scale/bias removed)" << TestLog::EndMessage
727 << TestLog::Message << "Viewport: " << viewportSize.x() << "x" << viewportSize.y() << TestLog::EndMessage
728 << TestLog::Message << "GL_FRAGMENT_SHADER_DERIVATE_HINT: " << glu::getHintModeStr(m_hint) << TestLog::EndMessage;
729
730 // Draw
731 {
732 const float positions[] =
733 {
734 -1.0f, -1.0f, 0.0f, 1.0f,
735 -1.0f, 1.0f, 0.0f, 1.0f,
736 1.0f, -1.0f, 0.0f, 1.0f,
737 1.0f, 1.0f, 0.0f, 1.0f
738 };
739 const float coords[] =
740 {
741 m_coordMin.x(), m_coordMin.y(), m_coordMin.z(), m_coordMax.w(),
742 m_coordMin.x(), m_coordMax.y(), (m_coordMin.z()+m_coordMax.z())*0.5f, (m_coordMin.w()+m_coordMax.w())*0.5f,
743 m_coordMax.x(), m_coordMin.y(), (m_coordMin.z()+m_coordMax.z())*0.5f, (m_coordMin.w()+m_coordMax.w())*0.5f,
744 m_coordMax.x(), m_coordMax.y(), m_coordMax.z(), m_coordMin.w()
745 };
746 const glu::VertexArrayBinding vertexArrays[] =
747 {
748 glu::va::Float("a_position", 4, 4, 0, &positions[0]),
749 glu::va::Float("a_coord", 4, 4, 0, &coords[0])
750 };
751 const deUint16 indices[] = { 0, 2, 1, 2, 3, 1 };
752
753 gl.clearColor(0.125f, 0.25f, 0.5f, 1.0f);
754 gl.clear(GL_COLOR_BUFFER_BIT|GL_DEPTH_BUFFER_BIT|GL_STENCIL_BUFFER_BIT);
755 gl.disable(GL_DITHER);
756
757 gl.useProgram(program.getProgram());
758
759 {
760 const int scaleLoc = gl.getUniformLocation(program.getProgram(), "u_scale");
761 const int biasLoc = gl.getUniformLocation(program.getProgram(), "u_bias");
762
763 switch (m_dataType)
764 {
765 case glu::TYPE_FLOAT:
766 gl.uniform1f(scaleLoc, m_derivScale.x());
767 gl.uniform1f(biasLoc, m_derivBias.x());
768 break;
769
770 case glu::TYPE_FLOAT_VEC2:
771 gl.uniform2fv(scaleLoc, 1, m_derivScale.getPtr());
772 gl.uniform2fv(biasLoc, 1, m_derivBias.getPtr());
773 break;
774
775 case glu::TYPE_FLOAT_VEC3:
776 gl.uniform3fv(scaleLoc, 1, m_derivScale.getPtr());
777 gl.uniform3fv(biasLoc, 1, m_derivBias.getPtr());
778 break;
779
780 case glu::TYPE_FLOAT_VEC4:
781 gl.uniform4fv(scaleLoc, 1, m_derivScale.getPtr());
782 gl.uniform4fv(biasLoc, 1, m_derivBias.getPtr());
783 break;
784
785 default:
786 DE_ASSERT(false);
787 }
788 }
789
790 gls::setupDefaultUniforms(m_context.getRenderContext(), program.getProgram());
791 setupRenderState(program.getProgram());
792
793 gl.hint(GL_FRAGMENT_SHADER_DERIVATIVE_HINT, m_hint);
794 GLU_EXPECT_NO_ERROR(gl.getError(), "Setup program state");
795
796 gl.viewport(viewportX, viewportY, viewportSize.x(), viewportSize.y());
797 glu::draw(m_context.getRenderContext(), program.getProgram(), DE_LENGTH_OF_ARRAY(vertexArrays), &vertexArrays[0],
798 glu::pr::Triangles(DE_LENGTH_OF_ARRAY(indices), &indices[0]));
799 GLU_EXPECT_NO_ERROR(gl.getError(), "Draw");
800 }
801
802 // Read back results
803 {
804 const bool isMSAA = useFbo && m_numSamples > 0;
805 AutoFbo resFbo (gl);
806 AutoRbo resRbo (gl);
807
808 // Resolve if necessary
809 if (isMSAA)
810 {
811 resFbo.gen();
812 resRbo.gen();
813
814 gl.bindRenderbuffer(GL_RENDERBUFFER, *resRbo);
815 gl.renderbufferStorageMultisample(GL_RENDERBUFFER, 0, fboFormat, viewportSize.x(), viewportSize.y());
816 gl.bindFramebuffer(GL_DRAW_FRAMEBUFFER, *resFbo);
817 gl.framebufferRenderbuffer(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, *resRbo);
818 TCU_CHECK(gl.checkFramebufferStatus(GL_FRAMEBUFFER) == GL_FRAMEBUFFER_COMPLETE);
819
820 gl.blitFramebuffer(0, 0, viewportSize.x(), viewportSize.y(), 0, 0, viewportSize.x(), viewportSize.y(), GL_COLOR_BUFFER_BIT, GL_NEAREST);
821 GLU_EXPECT_NO_ERROR(gl.getError(), "Resolve blit");
822
823 gl.bindFramebuffer(GL_READ_FRAMEBUFFER, *resFbo);
824 }
825
826 switch (m_surfaceType)
827 {
828 case SURFACETYPE_DEFAULT_FRAMEBUFFER:
829 case SURFACETYPE_UNORM_FBO:
830 result.setStorage(tcu::TextureFormat(tcu::TextureFormat::RGBA, tcu::TextureFormat::UNORM_INT8), viewportSize.x(), viewportSize.y());
831 glu::readPixels(m_context.getRenderContext(), viewportX, viewportY, result);
832 break;
833
834 case SURFACETYPE_FLOAT_FBO:
835 {
836 const tcu::TextureFormat dataFormat (tcu::TextureFormat::RGBA, tcu::TextureFormat::FLOAT);
837 const tcu::TextureFormat transferFormat (tcu::TextureFormat::RGBA, tcu::TextureFormat::UNSIGNED_INT32);
838
839 result.setStorage(dataFormat, viewportSize.x(), viewportSize.y());
840 glu::readPixels(m_context.getRenderContext(), viewportX, viewportY,
841 tcu::PixelBufferAccess(transferFormat, result.getWidth(), result.getHeight(), result.getDepth(), result.getAccess().getDataPtr()));
842 break;
843 }
844
845 default:
846 DE_ASSERT(false);
847 }
848
849 GLU_EXPECT_NO_ERROR(gl.getError(), "Read pixels");
850 }
851
852 // Verify
853 {
854 tcu::Surface errorMask(result.getWidth(), result.getHeight());
855 tcu::clear(errorMask.getAccess(), tcu::RGBA::green().toVec());
856
857 const qpTestResult testResult = verify(result.getAccess(), errorMask.getAccess());
858 const char* failStr = "Fail";
859
860 m_testCtx.getLog() << TestLog::ImageSet("Result", "Result images")
861 << TestLog::Image("Rendered", "Rendered image", result);
862
863 if (testResult != QP_TEST_RESULT_PASS)
864 m_testCtx.getLog() << TestLog::Image("ErrorMask", "Error mask", errorMask);
865
866 m_testCtx.getLog() << TestLog::EndImageSet;
867
868 if (testResult == QP_TEST_RESULT_PASS)
869 failStr = "Pass";
870 else if (testResult == QP_TEST_RESULT_QUALITY_WARNING)
871 failStr = "QualityWarning";
872
873 m_testCtx.setTestResult(testResult, failStr);
874
875 }
876
877 return STOP;
878 }
879
getSurfaceThreshold(void) const880 tcu::Vec4 TriangleDerivateCase::getSurfaceThreshold (void) const
881 {
882 switch (m_surfaceType)
883 {
884 case SURFACETYPE_DEFAULT_FRAMEBUFFER:
885 {
886 const tcu::PixelFormat pixelFormat = m_context.getRenderTarget().getPixelFormat();
887 const tcu::IVec4 channelBits (pixelFormat.redBits, pixelFormat.greenBits, pixelFormat.blueBits, pixelFormat.alphaBits);
888 const tcu::IVec4 intThreshold = tcu::IVec4(1) << (8 - channelBits);
889 const tcu::Vec4 normThreshold = intThreshold.asFloat() / 255.0f;
890
891 return normThreshold;
892 }
893
894 case SURFACETYPE_UNORM_FBO: return tcu::IVec4(1).asFloat() / 255.0f;
895 case SURFACETYPE_FLOAT_FBO: return tcu::Vec4(0.0f);
896 default:
897 DE_ASSERT(false);
898 return tcu::Vec4(0.0f);
899 }
900 }
901
902 // ConstantDerivateCase
903
904 class ConstantDerivateCase : public TriangleDerivateCase
905 {
906 public:
907 ConstantDerivateCase (Context& context, const char* name, const char* description, DerivateFunc func, glu::DataType type);
~ConstantDerivateCase(void)908 ~ConstantDerivateCase (void) {}
909
910 void init (void);
911
912 protected:
913 qpTestResult verify (const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask);
914
915 private:
916 DerivateFunc m_func;
917 };
918
ConstantDerivateCase(Context & context,const char * name,const char * description,DerivateFunc func,glu::DataType type)919 ConstantDerivateCase::ConstantDerivateCase (Context& context, const char* name, const char* description, DerivateFunc func, glu::DataType type)
920 : TriangleDerivateCase (context, name, description)
921 , m_func (func)
922 {
923 m_dataType = type;
924 m_precision = glu::PRECISION_HIGHP;
925 m_coordDataType = m_dataType;
926 m_coordPrecision = m_precision;
927 }
928
init(void)929 void ConstantDerivateCase::init (void)
930 {
931 const char* fragmentTmpl =
932 "#version 300 es\n"
933 "layout(location = 0) out mediump vec4 o_color;\n"
934 "uniform ${PRECISION} ${DATATYPE} u_scale;\n"
935 "uniform ${PRECISION} ${DATATYPE} u_bias;\n"
936 "void main (void)\n"
937 "{\n"
938 " ${PRECISION} ${DATATYPE} res = ${FUNC}(${VALUE}) * u_scale + u_bias;\n"
939 " o_color = ${CAST_TO_OUTPUT};\n"
940 "}\n";
941 map<string, string> fragmentParams;
942 fragmentParams["PRECISION"] = glu::getPrecisionName(m_precision);
943 fragmentParams["DATATYPE"] = glu::getDataTypeName(m_dataType);
944 fragmentParams["FUNC"] = getDerivateFuncName(m_func);
945 fragmentParams["VALUE"] = m_dataType == glu::TYPE_FLOAT_VEC4 ? "vec4(1.0, 7.2, -1e5, 0.0)" :
946 m_dataType == glu::TYPE_FLOAT_VEC3 ? "vec3(1e2, 8.0, 0.01)" :
947 m_dataType == glu::TYPE_FLOAT_VEC2 ? "vec2(-0.0, 2.7)" :
948 /* TYPE_FLOAT */ "7.7";
949 fragmentParams["CAST_TO_OUTPUT"] = m_dataType == glu::TYPE_FLOAT_VEC4 ? "res" :
950 m_dataType == glu::TYPE_FLOAT_VEC3 ? "vec4(res, 1.0)" :
951 m_dataType == glu::TYPE_FLOAT_VEC2 ? "vec4(res, 0.0, 1.0)" :
952 /* TYPE_FLOAT */ "vec4(res, 0.0, 0.0, 1.0)";
953
954 m_fragmentSrc = tcu::StringTemplate(fragmentTmpl).specialize(fragmentParams);
955
956 m_derivScale = tcu::Vec4(1e3f, 1e3f, 1e3f, 1e3f);
957 m_derivBias = tcu::Vec4(0.5f, 0.5f, 0.5f, 0.5f);
958 }
959
verify(const tcu::ConstPixelBufferAccess & result,const tcu::PixelBufferAccess & errorMask)960 qpTestResult ConstantDerivateCase::verify (const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask)
961 {
962 const tcu::Vec4 reference (0.0f); // Derivate of constant argument should always be 0
963 const tcu::Vec4 threshold = getSurfaceThreshold() / abs(m_derivScale);
964
965 return verifyConstantDerivate(m_testCtx.getLog(), result, errorMask, m_dataType,
966 reference, threshold, m_derivScale, m_derivBias);
967 }
968
969 // LinearDerivateCase
970
971 class LinearDerivateCase : public TriangleDerivateCase
972 {
973 public:
974 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);
~LinearDerivateCase(void)975 ~LinearDerivateCase (void) {}
976
977 void init (void);
978
979 protected:
980 qpTestResult verify (const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask);
981
982 private:
983 DerivateFunc m_func;
984 std::string m_fragmentTmpl;
985 };
986
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)987 LinearDerivateCase::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)
988 : TriangleDerivateCase (context, name, description)
989 , m_func (func)
990 , m_fragmentTmpl (fragmentSrcTmpl)
991 {
992 m_dataType = type;
993 m_precision = precision;
994 m_coordDataType = m_dataType;
995 m_coordPrecision = m_precision;
996 m_hint = hint;
997 m_surfaceType = surfaceType;
998 m_numSamples = numSamples;
999 }
1000
init(void)1001 void LinearDerivateCase::init (void)
1002 {
1003 const tcu::IVec2 viewportSize = getViewportSize();
1004 const float w = float(viewportSize.x());
1005 const float h = float(viewportSize.y());
1006 const bool packToInt = m_surfaceType == SURFACETYPE_FLOAT_FBO;
1007 map<string, string> fragmentParams;
1008
1009 fragmentParams["OUTPUT_TYPE"] = glu::getDataTypeName(packToInt ? glu::TYPE_UINT_VEC4 : glu::TYPE_FLOAT_VEC4);
1010 fragmentParams["OUTPUT_PREC"] = glu::getPrecisionName(packToInt ? glu::PRECISION_HIGHP : m_precision);
1011 fragmentParams["PRECISION"] = glu::getPrecisionName(m_precision);
1012 fragmentParams["DATATYPE"] = glu::getDataTypeName(m_dataType);
1013 fragmentParams["FUNC"] = getDerivateFuncName(m_func);
1014
1015 if (packToInt)
1016 {
1017 fragmentParams["CAST_TO_OUTPUT"] = m_dataType == glu::TYPE_FLOAT_VEC4 ? "floatBitsToUint(res)" :
1018 m_dataType == glu::TYPE_FLOAT_VEC3 ? "floatBitsToUint(vec4(res, 1.0))" :
1019 m_dataType == glu::TYPE_FLOAT_VEC2 ? "floatBitsToUint(vec4(res, 0.0, 1.0))" :
1020 /* TYPE_FLOAT */ "floatBitsToUint(vec4(res, 0.0, 0.0, 1.0))";
1021 }
1022 else
1023 {
1024 fragmentParams["CAST_TO_OUTPUT"] = m_dataType == glu::TYPE_FLOAT_VEC4 ? "res" :
1025 m_dataType == glu::TYPE_FLOAT_VEC3 ? "vec4(res, 1.0)" :
1026 m_dataType == glu::TYPE_FLOAT_VEC2 ? "vec4(res, 0.0, 1.0)" :
1027 /* TYPE_FLOAT */ "vec4(res, 0.0, 0.0, 1.0)";
1028 }
1029
1030 m_fragmentSrc = tcu::StringTemplate(m_fragmentTmpl.c_str()).specialize(fragmentParams);
1031
1032 switch (m_precision)
1033 {
1034 case glu::PRECISION_HIGHP:
1035 m_coordMin = tcu::Vec4(-97.f, 0.2f, 71.f, 74.f);
1036 m_coordMax = tcu::Vec4(-13.2f, -77.f, 44.f, 76.f);
1037 break;
1038
1039 case glu::PRECISION_MEDIUMP:
1040 m_coordMin = tcu::Vec4(-37.0f, 47.f, -7.f, 0.0f);
1041 m_coordMax = tcu::Vec4(-1.0f, 12.f, 7.f, 19.f);
1042 break;
1043
1044 case glu::PRECISION_LOWP:
1045 m_coordMin = tcu::Vec4(0.0f, -1.0f, 0.0f, 1.0f);
1046 m_coordMax = tcu::Vec4(1.0f, 1.0f, -1.0f, -1.0f);
1047 break;
1048
1049 default:
1050 DE_ASSERT(false);
1051 }
1052
1053 if (m_surfaceType == SURFACETYPE_FLOAT_FBO)
1054 {
1055 // No scale or bias used for accuracy.
1056 m_derivScale = tcu::Vec4(1.0f);
1057 m_derivBias = tcu::Vec4(0.0f);
1058 }
1059 else
1060 {
1061 // Compute scale - bias that normalizes to 0..1 range.
1062 const tcu::Vec4 dx = (m_coordMax - m_coordMin) / tcu::Vec4(w, w, w*0.5f, -w*0.5f);
1063 const tcu::Vec4 dy = (m_coordMax - m_coordMin) / tcu::Vec4(h, h, h*0.5f, -h*0.5f);
1064
1065 switch (m_func)
1066 {
1067 case DERIVATE_DFDX:
1068 m_derivScale = 0.5f / dx;
1069 break;
1070
1071 case DERIVATE_DFDY:
1072 m_derivScale = 0.5f / dy;
1073 break;
1074
1075 case DERIVATE_FWIDTH:
1076 m_derivScale = 0.5f / (tcu::abs(dx) + tcu::abs(dy));
1077 break;
1078
1079 default:
1080 DE_ASSERT(false);
1081 }
1082
1083 m_derivBias = tcu::Vec4(0.0f, 0.0f, 0.0f, 0.0f);
1084 }
1085 }
1086
verify(const tcu::ConstPixelBufferAccess & result,const tcu::PixelBufferAccess & errorMask)1087 qpTestResult LinearDerivateCase::verify (const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask)
1088 {
1089 const tcu::Vec4 xScale = tcu::Vec4(1.0f, 0.0f, 0.5f, -0.5f);
1090 const tcu::Vec4 yScale = tcu::Vec4(0.0f, 1.0f, 0.5f, -0.5f);
1091 const tcu::Vec4 surfaceThreshold = getSurfaceThreshold() / abs(m_derivScale);
1092
1093 if (m_func == DERIVATE_DFDX || m_func == DERIVATE_DFDY)
1094 {
1095 const bool isX = m_func == DERIVATE_DFDX;
1096 const float div = isX ? float(result.getWidth()) : float(result.getHeight());
1097 const tcu::Vec4 scale = isX ? xScale : yScale;
1098 tcu::Vec4 reference = ((m_coordMax - m_coordMin) / div);
1099 const tcu::Vec4 opThreshold = getDerivateThreshold(m_precision, m_coordMin, m_coordMax, reference);
1100 const tcu::Vec4 opThresholdW = getDerivateThresholdWarning(m_precision, m_coordMin, m_coordMax, reference);
1101 const tcu::Vec4 threshold = max(surfaceThreshold, opThreshold);
1102 const tcu::Vec4 thresholdW = max(surfaceThreshold, opThresholdW);
1103 const int numComps = glu::getDataTypeFloatScalars(m_dataType);
1104
1105 /* adjust the reference value for the correct dfdx or dfdy sample adjacency */
1106 reference = reference * scale;
1107
1108 m_testCtx.getLog()
1109 << tcu::TestLog::Message
1110 << "Verifying result image.\n"
1111 << "\tValid derivative is " << LogVecComps(reference, numComps) << " with threshold " << LogVecComps(threshold, numComps)
1112 << tcu::TestLog::EndMessage;
1113
1114 // short circuit if result is strictly within the normal value error bounds.
1115 // This improves performance significantly.
1116 if (verifyConstantDerivate(m_testCtx.getLog(), result, errorMask, m_dataType,
1117 reference, threshold, m_derivScale, m_derivBias,
1118 LOG_NOTHING) == QP_TEST_RESULT_PASS)
1119 {
1120 m_testCtx.getLog()
1121 << tcu::TestLog::Message
1122 << "No incorrect derivatives found, result valid."
1123 << tcu::TestLog::EndMessage;
1124
1125 return QP_TEST_RESULT_PASS;
1126 }
1127
1128 // Check with relaxed threshold value
1129 if (verifyConstantDerivate(m_testCtx.getLog(), result, errorMask, m_dataType,
1130 reference, thresholdW, m_derivScale, m_derivBias,
1131 LOG_NOTHING) == QP_TEST_RESULT_PASS)
1132 {
1133 m_testCtx.getLog()
1134 << tcu::TestLog::Message
1135 << "No incorrect derivatives found, result valid with quality warning."
1136 << tcu::TestLog::EndMessage;
1137
1138 return QP_TEST_RESULT_QUALITY_WARNING;
1139 }
1140
1141 // some pixels exceed error bounds calculated for normal values. Verify that these
1142 // potentially invalid pixels are in fact valid due to (for example) subnorm flushing.
1143
1144 m_testCtx.getLog()
1145 << tcu::TestLog::Message
1146 << "Initial verification failed, verifying image by calculating accurate error bounds for each result pixel.\n"
1147 << "\tVerifying each result derivative is within its range of legal result values."
1148 << tcu::TestLog::EndMessage;
1149
1150 {
1151 const tcu::IVec2 viewportSize = getViewportSize();
1152 const float w = float(viewportSize.x());
1153 const float h = float(viewportSize.y());
1154 const tcu::Vec4 valueRamp = (m_coordMax - m_coordMin);
1155 Linear2DFunctionEvaluator function;
1156
1157 function.matrix.setRow(0, tcu::Vec3(valueRamp.x() / w, 0.0f, m_coordMin.x()));
1158 function.matrix.setRow(1, tcu::Vec3(0.0f, valueRamp.y() / h, m_coordMin.y()));
1159 function.matrix.setRow(2, tcu::Vec3(valueRamp.z() / w, valueRamp.z() / h, m_coordMin.z() + m_coordMin.z()) / 2.0f);
1160 function.matrix.setRow(3, tcu::Vec3(-valueRamp.w() / w, -valueRamp.w() / h, m_coordMax.w() + m_coordMax.w()) / 2.0f);
1161
1162 return reverifyConstantDerivateWithFlushRelaxations(m_testCtx.getLog(), result, errorMask,
1163 m_dataType, m_precision, m_derivScale,
1164 m_derivBias, surfaceThreshold, m_func,
1165 function);
1166 }
1167 }
1168 else
1169 {
1170 DE_ASSERT(m_func == DERIVATE_FWIDTH);
1171 const float w = float(result.getWidth());
1172 const float h = float(result.getHeight());
1173
1174 const tcu::Vec4 dx = ((m_coordMax - m_coordMin) / w) * xScale;
1175 const tcu::Vec4 dy = ((m_coordMax - m_coordMin) / h) * yScale;
1176 const tcu::Vec4 reference = tcu::abs(dx) + tcu::abs(dy);
1177 const tcu::Vec4 dxThreshold = getDerivateThreshold(m_precision, m_coordMin*xScale, m_coordMax*xScale, dx);
1178 const tcu::Vec4 dyThreshold = getDerivateThreshold(m_precision, m_coordMin*yScale, m_coordMax*yScale, dy);
1179 const tcu::Vec4 dxThresholdW = getDerivateThresholdWarning(m_precision, m_coordMin*xScale, m_coordMax*xScale, dx);
1180 const tcu::Vec4 dyThresholdW = getDerivateThresholdWarning(m_precision, m_coordMin*yScale, m_coordMax*yScale, dy);
1181 const tcu::Vec4 threshold = max(surfaceThreshold, max(dxThreshold, dyThreshold));
1182 const tcu::Vec4 thresholdW = max(surfaceThreshold, max(dxThresholdW, dyThresholdW));
1183 qpTestResult testResult = QP_TEST_RESULT_FAIL;
1184
1185 testResult = verifyConstantDerivate(m_testCtx.getLog(), result, errorMask, m_dataType,
1186 reference, threshold, m_derivScale, m_derivBias);
1187
1188 // return if result is pass
1189 if (testResult == QP_TEST_RESULT_PASS)
1190 return testResult;
1191
1192 // re-check with relaxed threshold
1193 testResult = verifyConstantDerivate(m_testCtx.getLog(), result, errorMask, m_dataType,
1194 reference, thresholdW, m_derivScale, m_derivBias);
1195
1196 // if with relaxed threshold test is passing then mark the result with quality warning.
1197 if (testResult == QP_TEST_RESULT_PASS)
1198 testResult = QP_TEST_RESULT_QUALITY_WARNING;
1199
1200 return testResult;
1201 }
1202 }
1203
1204 // TextureDerivateCase
1205
1206 class TextureDerivateCase : public TriangleDerivateCase
1207 {
1208 public:
1209 TextureDerivateCase (Context& context, const char* name, const char* description, DerivateFunc func, glu::DataType type, glu::Precision precision, deUint32 hint, SurfaceType surfaceType, int numSamples);
1210 ~TextureDerivateCase (void);
1211
1212 void init (void);
1213 void deinit (void);
1214
1215 protected:
1216 void setupRenderState (deUint32 program);
1217 qpTestResult verify (const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask);
1218
1219 private:
1220 DerivateFunc m_func;
1221
1222 tcu::Vec4 m_texValueMin;
1223 tcu::Vec4 m_texValueMax;
1224 glu::Texture2D* m_texture;
1225 };
1226
TextureDerivateCase(Context & context,const char * name,const char * description,DerivateFunc func,glu::DataType type,glu::Precision precision,deUint32 hint,SurfaceType surfaceType,int numSamples)1227 TextureDerivateCase::TextureDerivateCase (Context& context, const char* name, const char* description, DerivateFunc func, glu::DataType type, glu::Precision precision, deUint32 hint, SurfaceType surfaceType, int numSamples)
1228 : TriangleDerivateCase (context, name, description)
1229 , m_func (func)
1230 , m_texture (DE_NULL)
1231 {
1232 m_dataType = type;
1233 m_precision = precision;
1234 m_coordDataType = glu::TYPE_FLOAT_VEC2;
1235 m_coordPrecision = glu::PRECISION_HIGHP;
1236 m_hint = hint;
1237 m_surfaceType = surfaceType;
1238 m_numSamples = numSamples;
1239 }
1240
~TextureDerivateCase(void)1241 TextureDerivateCase::~TextureDerivateCase (void)
1242 {
1243 delete m_texture;
1244 }
1245
init(void)1246 void TextureDerivateCase::init (void)
1247 {
1248 // Generate shader
1249 {
1250 const char* fragmentTmpl =
1251 "#version 300 es\n"
1252 "in highp vec2 v_coord;\n"
1253 "layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1254 "uniform ${PRECISION} sampler2D u_sampler;\n"
1255 "uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1256 "uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1257 "void main (void)\n"
1258 "{\n"
1259 " ${PRECISION} vec4 tex = texture(u_sampler, v_coord);\n"
1260 " ${PRECISION} ${DATATYPE} res = ${FUNC}(tex${SWIZZLE}) * u_scale + u_bias;\n"
1261 " o_color = ${CAST_TO_OUTPUT};\n"
1262 "}\n";
1263
1264 const bool packToInt = m_surfaceType == SURFACETYPE_FLOAT_FBO;
1265 map<string, string> fragmentParams;
1266
1267 fragmentParams["OUTPUT_TYPE"] = glu::getDataTypeName(packToInt ? glu::TYPE_UINT_VEC4 : glu::TYPE_FLOAT_VEC4);
1268 fragmentParams["OUTPUT_PREC"] = glu::getPrecisionName(packToInt ? glu::PRECISION_HIGHP : m_precision);
1269 fragmentParams["PRECISION"] = glu::getPrecisionName(m_precision);
1270 fragmentParams["DATATYPE"] = glu::getDataTypeName(m_dataType);
1271 fragmentParams["FUNC"] = getDerivateFuncName(m_func);
1272 fragmentParams["SWIZZLE"] = m_dataType == glu::TYPE_FLOAT_VEC4 ? "" :
1273 m_dataType == glu::TYPE_FLOAT_VEC3 ? ".xyz" :
1274 m_dataType == glu::TYPE_FLOAT_VEC2 ? ".xy" :
1275 /* TYPE_FLOAT */ ".x";
1276
1277 if (packToInt)
1278 {
1279 fragmentParams["CAST_TO_OUTPUT"] = m_dataType == glu::TYPE_FLOAT_VEC4 ? "floatBitsToUint(res)" :
1280 m_dataType == glu::TYPE_FLOAT_VEC3 ? "floatBitsToUint(vec4(res, 1.0))" :
1281 m_dataType == glu::TYPE_FLOAT_VEC2 ? "floatBitsToUint(vec4(res, 0.0, 1.0))" :
1282 /* TYPE_FLOAT */ "floatBitsToUint(vec4(res, 0.0, 0.0, 1.0))";
1283 }
1284 else
1285 {
1286 fragmentParams["CAST_TO_OUTPUT"] = m_dataType == glu::TYPE_FLOAT_VEC4 ? "res" :
1287 m_dataType == glu::TYPE_FLOAT_VEC3 ? "vec4(res, 1.0)" :
1288 m_dataType == glu::TYPE_FLOAT_VEC2 ? "vec4(res, 0.0, 1.0)" :
1289 /* TYPE_FLOAT */ "vec4(res, 0.0, 0.0, 1.0)";
1290 }
1291
1292 m_fragmentSrc = tcu::StringTemplate(fragmentTmpl).specialize(fragmentParams);
1293 }
1294
1295 // Texture size matches viewport and nearest sampling is used. Thus texture sampling
1296 // is equal to just interpolating the texture value range.
1297
1298 // Determine value range for texture.
1299
1300 switch (m_precision)
1301 {
1302 case glu::PRECISION_HIGHP:
1303 m_texValueMin = tcu::Vec4(-97.f, 0.2f, 71.f, 74.f);
1304 m_texValueMax = tcu::Vec4(-13.2f, -77.f, 44.f, 76.f);
1305 break;
1306
1307 case glu::PRECISION_MEDIUMP:
1308 m_texValueMin = tcu::Vec4(-37.0f, 47.f, -7.f, 0.0f);
1309 m_texValueMax = tcu::Vec4(-1.0f, 12.f, 7.f, 19.f);
1310 break;
1311
1312 case glu::PRECISION_LOWP:
1313 m_texValueMin = tcu::Vec4(0.0f, -1.0f, 0.0f, 1.0f);
1314 m_texValueMax = tcu::Vec4(1.0f, 1.0f, -1.0f, -1.0f);
1315 break;
1316
1317 default:
1318 DE_ASSERT(false);
1319 }
1320
1321 // Lowp and mediump cases use RGBA16F format, while highp uses RGBA32F.
1322 {
1323 const tcu::IVec2 viewportSize = getViewportSize();
1324 DE_ASSERT(!m_texture);
1325 m_texture = new glu::Texture2D(m_context.getRenderContext(), m_precision == glu::PRECISION_HIGHP ? GL_RGBA32F : GL_RGBA16F, viewportSize.x(), viewportSize.y());
1326 m_texture->getRefTexture().allocLevel(0);
1327 }
1328
1329 // Texture coordinates
1330 m_coordMin = tcu::Vec4(0.0f);
1331 m_coordMax = tcu::Vec4(1.0f);
1332
1333 // Fill with gradients.
1334 {
1335 const tcu::PixelBufferAccess level0 = m_texture->getRefTexture().getLevel(0);
1336 for (int y = 0; y < level0.getHeight(); y++)
1337 {
1338 for (int x = 0; x < level0.getWidth(); x++)
1339 {
1340 const float xf = (float(x)+0.5f) / float(level0.getWidth());
1341 const float yf = (float(y)+0.5f) / float(level0.getHeight());
1342 const tcu::Vec4 s = tcu::Vec4(xf, yf, (xf+yf)/2.0f, 1.0f - (xf+yf)/2.0f);
1343
1344 level0.setPixel(m_texValueMin + (m_texValueMax - m_texValueMin)*s, x, y);
1345 }
1346 }
1347 }
1348
1349 m_texture->upload();
1350
1351 if (m_surfaceType == SURFACETYPE_FLOAT_FBO)
1352 {
1353 // No scale or bias used for accuracy.
1354 m_derivScale = tcu::Vec4(1.0f);
1355 m_derivBias = tcu::Vec4(0.0f);
1356 }
1357 else
1358 {
1359 // Compute scale - bias that normalizes to 0..1 range.
1360 const tcu::IVec2 viewportSize = getViewportSize();
1361 const float w = float(viewportSize.x());
1362 const float h = float(viewportSize.y());
1363 const tcu::Vec4 dx = (m_texValueMax - m_texValueMin) / tcu::Vec4(w, w, w*0.5f, -w*0.5f);
1364 const tcu::Vec4 dy = (m_texValueMax - m_texValueMin) / tcu::Vec4(h, h, h*0.5f, -h*0.5f);
1365
1366 switch (m_func)
1367 {
1368 case DERIVATE_DFDX:
1369 m_derivScale = 0.5f / dx;
1370 break;
1371
1372 case DERIVATE_DFDY:
1373 m_derivScale = 0.5f / dy;
1374 break;
1375
1376 case DERIVATE_FWIDTH:
1377 m_derivScale = 0.5f / (tcu::abs(dx) + tcu::abs(dy));
1378 break;
1379
1380 default:
1381 DE_ASSERT(false);
1382 }
1383
1384 m_derivBias = tcu::Vec4(0.0f, 0.0f, 0.0f, 0.0f);
1385 }
1386 }
1387
deinit(void)1388 void TextureDerivateCase::deinit (void)
1389 {
1390 delete m_texture;
1391 m_texture = DE_NULL;
1392 }
1393
setupRenderState(deUint32 program)1394 void TextureDerivateCase::setupRenderState (deUint32 program)
1395 {
1396 const glw::Functions& gl = m_context.getRenderContext().getFunctions();
1397 const int texUnit = 1;
1398
1399 gl.activeTexture (GL_TEXTURE0+texUnit);
1400 gl.bindTexture (GL_TEXTURE_2D, m_texture->getGLTexture());
1401 gl.texParameteri (GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1402 gl.texParameteri (GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1403 gl.texParameteri (GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1404 gl.texParameteri (GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1405
1406 gl.uniform1i (gl.getUniformLocation(program, "u_sampler"), texUnit);
1407 }
1408
verify(const tcu::ConstPixelBufferAccess & result,const tcu::PixelBufferAccess & errorMask)1409 qpTestResult TextureDerivateCase::verify (const tcu::ConstPixelBufferAccess& result, const tcu::PixelBufferAccess& errorMask)
1410 {
1411 // \note Edges are ignored in comparison
1412 if (result.getWidth() < 2 || result.getHeight() < 2)
1413 throw tcu::NotSupportedError("Too small viewport");
1414
1415 tcu::ConstPixelBufferAccess compareArea = tcu::getSubregion(result, 1, 1, result.getWidth()-2, result.getHeight()-2);
1416 tcu::PixelBufferAccess maskArea = tcu::getSubregion(errorMask, 1, 1, errorMask.getWidth()-2, errorMask.getHeight()-2);
1417 const tcu::Vec4 xScale = tcu::Vec4(1.0f, 0.0f, 0.5f, -0.5f);
1418 const tcu::Vec4 yScale = tcu::Vec4(0.0f, 1.0f, 0.5f, -0.5f);
1419 const float w = float(result.getWidth());
1420 const float h = float(result.getHeight());
1421
1422 const tcu::Vec4 surfaceThreshold = getSurfaceThreshold() / abs(m_derivScale);
1423
1424 if (m_func == DERIVATE_DFDX || m_func == DERIVATE_DFDY)
1425 {
1426 const bool isX = m_func == DERIVATE_DFDX;
1427 const float div = isX ? w : h;
1428 const tcu::Vec4 scale = isX ? xScale : yScale;
1429 tcu::Vec4 reference = ((m_texValueMax - m_texValueMin) / div);
1430 const tcu::Vec4 opThreshold = getDerivateThreshold(m_precision, m_texValueMin, m_texValueMax, reference);
1431 const tcu::Vec4 opThresholdW = getDerivateThresholdWarning(m_precision, m_texValueMin, m_texValueMax, reference);
1432 const tcu::Vec4 threshold = max(surfaceThreshold, opThreshold);
1433 const tcu::Vec4 thresholdW = max(surfaceThreshold, opThresholdW);
1434 const int numComps = glu::getDataTypeFloatScalars(m_dataType);
1435
1436 /* adjust the reference value for the correct dfdx or dfdy sample adjacency */
1437 reference = reference * scale;
1438
1439 m_testCtx.getLog()
1440 << tcu::TestLog::Message
1441 << "Verifying result image.\n"
1442 << "\tValid derivative is " << LogVecComps(reference, numComps) << " with threshold " << LogVecComps(threshold, numComps)
1443 << tcu::TestLog::EndMessage;
1444
1445 // short circuit if result is strictly within the normal value error bounds.
1446 // This improves performance significantly.
1447 if (verifyConstantDerivate(m_testCtx.getLog(), compareArea, maskArea, m_dataType,
1448 reference, threshold, m_derivScale, m_derivBias,
1449 LOG_NOTHING) == QP_TEST_RESULT_PASS)
1450 {
1451 m_testCtx.getLog()
1452 << tcu::TestLog::Message
1453 << "No incorrect derivatives found, result valid."
1454 << tcu::TestLog::EndMessage;
1455
1456 return QP_TEST_RESULT_PASS;
1457 }
1458
1459 m_testCtx.getLog()
1460 << tcu::TestLog::Message
1461 << "Verifying result image.\n"
1462 << "\tValid derivative is " << LogVecComps(reference, numComps) << " with Warning threshold " << LogVecComps(thresholdW, numComps)
1463 << tcu::TestLog::EndMessage;
1464
1465 // Re-check with relaxed threshold
1466 if (verifyConstantDerivate(m_testCtx.getLog(), compareArea, maskArea, m_dataType,
1467 reference, thresholdW, m_derivScale, m_derivBias,
1468 LOG_NOTHING) == QP_TEST_RESULT_PASS)
1469 {
1470 m_testCtx.getLog()
1471 << tcu::TestLog::Message
1472 << "No incorrect derivatives found, result valid with quality warning."
1473 << tcu::TestLog::EndMessage;
1474
1475 return QP_TEST_RESULT_QUALITY_WARNING;
1476 }
1477
1478
1479 // some pixels exceed error bounds calculated for normal values. Verify that these
1480 // potentially invalid pixels are in fact valid due to (for example) subnorm flushing.
1481
1482 m_testCtx.getLog()
1483 << tcu::TestLog::Message
1484 << "Initial verification failed, verifying image by calculating accurate error bounds for each result pixel.\n"
1485 << "\tVerifying each result derivative is within its range of legal result values."
1486 << tcu::TestLog::EndMessage;
1487
1488 {
1489 const tcu::Vec4 valueRamp = (m_texValueMax - m_texValueMin);
1490 Linear2DFunctionEvaluator function;
1491
1492 function.matrix.setRow(0, tcu::Vec3(valueRamp.x() / w, 0.0f, m_texValueMin.x()));
1493 function.matrix.setRow(1, tcu::Vec3(0.0f, valueRamp.y() / h, m_texValueMin.y()));
1494 function.matrix.setRow(2, tcu::Vec3(valueRamp.z() / w, valueRamp.z() / h, m_texValueMin.z() + m_texValueMin.z()) / 2.0f);
1495 function.matrix.setRow(3, tcu::Vec3(-valueRamp.w() / w, -valueRamp.w() / h, m_texValueMax.w() + m_texValueMax.w()) / 2.0f);
1496
1497 return reverifyConstantDerivateWithFlushRelaxations(m_testCtx.getLog(), compareArea, maskArea,
1498 m_dataType, m_precision, m_derivScale,
1499 m_derivBias, surfaceThreshold, m_func,
1500 function);
1501 }
1502 }
1503 else
1504 {
1505 DE_ASSERT(m_func == DERIVATE_FWIDTH);
1506 const tcu::Vec4 dx = ((m_texValueMax - m_texValueMin) / w) * xScale;
1507 const tcu::Vec4 dy = ((m_texValueMax - m_texValueMin) / h) * yScale;
1508 const tcu::Vec4 reference = tcu::abs(dx) + tcu::abs(dy);
1509 const tcu::Vec4 dxThreshold = getDerivateThreshold(m_precision, m_texValueMin*xScale, m_texValueMax*xScale, dx);
1510 const tcu::Vec4 dyThreshold = getDerivateThreshold(m_precision, m_texValueMin*yScale, m_texValueMax*yScale, dy);
1511 const tcu::Vec4 dxThresholdW = getDerivateThresholdWarning(m_precision, m_texValueMin*xScale, m_texValueMax*xScale, dx);
1512 const tcu::Vec4 dyThresholdW = getDerivateThresholdWarning(m_precision, m_texValueMin*yScale, m_texValueMax*yScale, dy);
1513 const tcu::Vec4 threshold = max(surfaceThreshold, max(dxThreshold, dyThreshold));
1514 const tcu::Vec4 thresholdW = max(surfaceThreshold, max(dxThresholdW, dyThresholdW));
1515 qpTestResult testResult = QP_TEST_RESULT_FAIL;
1516
1517 testResult = verifyConstantDerivate(m_testCtx.getLog(), compareArea, maskArea, m_dataType,
1518 reference, threshold, m_derivScale, m_derivBias);
1519
1520 if (testResult == QP_TEST_RESULT_PASS)
1521 return testResult;
1522
1523 // Re-Check with relaxed threshold
1524 testResult = verifyConstantDerivate(m_testCtx.getLog(), compareArea, maskArea, m_dataType,
1525 reference, thresholdW, m_derivScale, m_derivBias);
1526
1527 // If test is passing with relaxed threshold then mark quality warning
1528 if (testResult == QP_TEST_RESULT_PASS)
1529 testResult = QP_TEST_RESULT_QUALITY_WARNING;
1530
1531 return testResult;
1532 }
1533 }
1534
ShaderDerivateTests(Context & context)1535 ShaderDerivateTests::ShaderDerivateTests (Context& context)
1536 : TestCaseGroup(context, "derivate", "Derivate Function Tests")
1537 {
1538 }
1539
~ShaderDerivateTests(void)1540 ShaderDerivateTests::~ShaderDerivateTests (void)
1541 {
1542 }
1543
1544 struct FunctionSpec
1545 {
1546 std::string name;
1547 DerivateFunc function;
1548 glu::DataType dataType;
1549 glu::Precision precision;
1550
FunctionSpecdeqp::gles3::Functional::FunctionSpec1551 FunctionSpec (const std::string& name_, DerivateFunc function_, glu::DataType dataType_, glu::Precision precision_)
1552 : name (name_)
1553 , function (function_)
1554 , dataType (dataType_)
1555 , precision (precision_)
1556 {
1557 }
1558 };
1559
init(void)1560 void ShaderDerivateTests::init (void)
1561 {
1562 static const struct
1563 {
1564 const char* name;
1565 const char* description;
1566 const char* source;
1567 } s_linearDerivateCases[] =
1568 {
1569 {
1570 "linear",
1571 "Basic derivate of linearly interpolated argument",
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 "void main (void)\n"
1579 "{\n"
1580 " ${PRECISION} ${DATATYPE} res = ${FUNC}(v_coord) * u_scale + u_bias;\n"
1581 " o_color = ${CAST_TO_OUTPUT};\n"
1582 "}\n"
1583 },
1584 {
1585 "in_function",
1586 "Derivate of linear function argument",
1587
1588 "#version 300 es\n"
1589 "in ${PRECISION} ${DATATYPE} v_coord;\n"
1590 "layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1591 "uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1592 "uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1593 "\n"
1594 "${PRECISION} ${DATATYPE} computeRes (${PRECISION} ${DATATYPE} value)\n"
1595 "{\n"
1596 " return ${FUNC}(v_coord) * u_scale + u_bias;\n"
1597 "}\n"
1598 "\n"
1599 "void main (void)\n"
1600 "{\n"
1601 " ${PRECISION} ${DATATYPE} res = computeRes(v_coord);\n"
1602 " o_color = ${CAST_TO_OUTPUT};\n"
1603 "}\n"
1604 },
1605 {
1606 "static_if",
1607 "Derivate of linearly interpolated value in static if",
1608
1609 "#version 300 es\n"
1610 "in ${PRECISION} ${DATATYPE} v_coord;\n"
1611 "layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1612 "uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1613 "uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1614 "void main (void)\n"
1615 "{\n"
1616 " ${PRECISION} ${DATATYPE} res;\n"
1617 " if (false)\n"
1618 " res = ${FUNC}(-v_coord) * u_scale + u_bias;\n"
1619 " else\n"
1620 " res = ${FUNC}(v_coord) * u_scale + u_bias;\n"
1621 " o_color = ${CAST_TO_OUTPUT};\n"
1622 "}\n"
1623 },
1624 {
1625 "static_loop",
1626 "Derivate of linearly interpolated value in static loop",
1627
1628 "#version 300 es\n"
1629 "in ${PRECISION} ${DATATYPE} v_coord;\n"
1630 "layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1631 "uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1632 "uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1633 "void main (void)\n"
1634 "{\n"
1635 " ${PRECISION} ${DATATYPE} res = ${DATATYPE}(0.0);\n"
1636 " for (int i = 0; i < 2; i++)\n"
1637 " res += ${FUNC}(v_coord * float(i));\n"
1638 " res = res * u_scale + u_bias;\n"
1639 " o_color = ${CAST_TO_OUTPUT};\n"
1640 "}\n"
1641 },
1642 {
1643 "static_switch",
1644 "Derivate of linearly interpolated value in static switch",
1645
1646 "#version 300 es\n"
1647 "in ${PRECISION} ${DATATYPE} v_coord;\n"
1648 "layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1649 "uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1650 "uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1651 "void main (void)\n"
1652 "{\n"
1653 " ${PRECISION} ${DATATYPE} res;\n"
1654 " switch (1)\n"
1655 " {\n"
1656 " case 0: res = ${FUNC}(-v_coord) * u_scale + u_bias; break;\n"
1657 " case 1: res = ${FUNC}(v_coord) * u_scale + u_bias; break;\n"
1658 " }\n"
1659 " o_color = ${CAST_TO_OUTPUT};\n"
1660 "}\n"
1661 },
1662 {
1663 "uniform_if",
1664 "Derivate of linearly interpolated value in uniform if",
1665
1666 "#version 300 es\n"
1667 "in ${PRECISION} ${DATATYPE} v_coord;\n"
1668 "layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1669 "uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1670 "uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1671 "uniform bool ub_true;\n"
1672 "void main (void)\n"
1673 "{\n"
1674 " ${PRECISION} ${DATATYPE} res;\n"
1675 " if (ub_true)"
1676 " res = ${FUNC}(v_coord) * u_scale + u_bias;\n"
1677 " else\n"
1678 " res = ${FUNC}(-v_coord) * u_scale + u_bias;\n"
1679 " o_color = ${CAST_TO_OUTPUT};\n"
1680 "}\n"
1681 },
1682 {
1683 "uniform_loop",
1684 "Derivate of linearly interpolated value in uniform loop",
1685
1686 "#version 300 es\n"
1687 "in ${PRECISION} ${DATATYPE} v_coord;\n"
1688 "layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1689 "uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1690 "uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1691 "uniform int ui_two;\n"
1692 "void main (void)\n"
1693 "{\n"
1694 " ${PRECISION} ${DATATYPE} res = ${DATATYPE}(0.0);\n"
1695 " for (int i = 0; i < ui_two; i++)\n"
1696 " res += ${FUNC}(v_coord * float(i));\n"
1697 " res = res * u_scale + u_bias;\n"
1698 " o_color = ${CAST_TO_OUTPUT};\n"
1699 "}\n"
1700 },
1701 {
1702 "uniform_switch",
1703 "Derivate of linearly interpolated value in uniform switch",
1704
1705 "#version 300 es\n"
1706 "in ${PRECISION} ${DATATYPE} v_coord;\n"
1707 "layout(location = 0) out ${OUTPUT_PREC} ${OUTPUT_TYPE} o_color;\n"
1708 "uniform ${PRECISION} ${DATATYPE} u_scale;\n"
1709 "uniform ${PRECISION} ${DATATYPE} u_bias;\n"
1710 "uniform int ui_one;\n"
1711 "void main (void)\n"
1712 "{\n"
1713 " ${PRECISION} ${DATATYPE} res;\n"
1714 " switch (ui_one)\n"
1715 " {\n"
1716 " case 0: res = ${FUNC}(-v_coord) * u_scale + u_bias; break;\n"
1717 " case 1: res = ${FUNC}(v_coord) * u_scale + u_bias; break;\n"
1718 " }\n"
1719 " o_color = ${CAST_TO_OUTPUT};\n"
1720 "}\n"
1721 },
1722 };
1723
1724 static const struct
1725 {
1726 const char* name;
1727 SurfaceType surfaceType;
1728 int numSamples;
1729 } s_fboConfigs[] =
1730 {
1731 { "fbo", SURFACETYPE_DEFAULT_FRAMEBUFFER, 0 },
1732 { "fbo_msaa2", SURFACETYPE_UNORM_FBO, 2 },
1733 { "fbo_msaa4", SURFACETYPE_UNORM_FBO, 4 },
1734 { "fbo_float", SURFACETYPE_FLOAT_FBO, 0 },
1735 };
1736
1737 static const struct
1738 {
1739 const char* name;
1740 deUint32 hint;
1741 } s_hints[] =
1742 {
1743 { "fastest", GL_FASTEST },
1744 { "nicest", GL_NICEST },
1745 };
1746
1747 static const struct
1748 {
1749 const char* name;
1750 SurfaceType surfaceType;
1751 int numSamples;
1752 } s_hintFboConfigs[] =
1753 {
1754 { "default", SURFACETYPE_DEFAULT_FRAMEBUFFER, 0 },
1755 { "fbo_msaa4", SURFACETYPE_UNORM_FBO, 4 },
1756 { "fbo_float", SURFACETYPE_FLOAT_FBO, 0 }
1757 };
1758
1759 static const struct
1760 {
1761 const char* name;
1762 SurfaceType surfaceType;
1763 int numSamples;
1764 deUint32 hint;
1765 } s_textureConfigs[] =
1766 {
1767 { "basic", SURFACETYPE_DEFAULT_FRAMEBUFFER, 0, GL_DONT_CARE },
1768 { "msaa4", SURFACETYPE_UNORM_FBO, 4, GL_DONT_CARE },
1769 { "float_fastest", SURFACETYPE_FLOAT_FBO, 0, GL_FASTEST },
1770 { "float_nicest", SURFACETYPE_FLOAT_FBO, 0, GL_NICEST },
1771 };
1772
1773 // .dfdx, .dfdy, .fwidth
1774 for (int funcNdx = 0; funcNdx < DERIVATE_LAST; funcNdx++)
1775 {
1776 const DerivateFunc function = DerivateFunc(funcNdx);
1777 tcu::TestCaseGroup* const functionGroup = new tcu::TestCaseGroup(m_testCtx, getDerivateFuncCaseName(function), getDerivateFuncName(function));
1778 addChild(functionGroup);
1779
1780 // .constant - no precision variants, checks that derivate of constant arguments is 0
1781 {
1782 tcu::TestCaseGroup* const constantGroup = new tcu::TestCaseGroup(m_testCtx, "constant", "Derivate of constant argument");
1783 functionGroup->addChild(constantGroup);
1784
1785 for (int vecSize = 1; vecSize <= 4; vecSize++)
1786 {
1787 const glu::DataType dataType = vecSize > 1 ? glu::getDataTypeFloatVec(vecSize) : glu::TYPE_FLOAT;
1788 constantGroup->addChild(new ConstantDerivateCase(m_context, glu::getDataTypeName(dataType), "", function, dataType));
1789 }
1790 }
1791
1792 // Cases based on LinearDerivateCase
1793 for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(s_linearDerivateCases); caseNdx++)
1794 {
1795 tcu::TestCaseGroup* const linearCaseGroup = new tcu::TestCaseGroup(m_testCtx, s_linearDerivateCases[caseNdx].name, s_linearDerivateCases[caseNdx].description);
1796 const char* source = s_linearDerivateCases[caseNdx].source;
1797 functionGroup->addChild(linearCaseGroup);
1798
1799 for (int vecSize = 1; vecSize <= 4; vecSize++)
1800 {
1801 for (int precNdx = 0; precNdx < glu::PRECISION_LAST; precNdx++)
1802 {
1803 const glu::DataType dataType = vecSize > 1 ? glu::getDataTypeFloatVec(vecSize) : glu::TYPE_FLOAT;
1804 const glu::Precision precision = glu::Precision(precNdx);
1805 const SurfaceType surfaceType = SURFACETYPE_DEFAULT_FRAMEBUFFER;
1806 const int numSamples = 0;
1807 const deUint32 hint = GL_DONT_CARE;
1808 ostringstream caseName;
1809
1810 if (caseNdx != 0 && precision == glu::PRECISION_LOWP)
1811 continue; // Skip as lowp doesn't actually produce any bits when rendered to default FB.
1812
1813 caseName << glu::getDataTypeName(dataType) << "_" << glu::getPrecisionName(precision);
1814
1815 linearCaseGroup->addChild(new LinearDerivateCase(m_context, caseName.str().c_str(), "", function, dataType, precision, hint, surfaceType, numSamples, source));
1816 }
1817 }
1818 }
1819
1820 // Fbo cases
1821 for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(s_fboConfigs); caseNdx++)
1822 {
1823 tcu::TestCaseGroup* const fboGroup = new tcu::TestCaseGroup(m_testCtx, s_fboConfigs[caseNdx].name, "Derivate usage when rendering into FBO");
1824 const char* source = s_linearDerivateCases[0].source; // use source from .linear group
1825 const SurfaceType surfaceType = s_fboConfigs[caseNdx].surfaceType;
1826 const int numSamples = s_fboConfigs[caseNdx].numSamples;
1827 functionGroup->addChild(fboGroup);
1828
1829 for (int vecSize = 1; vecSize <= 4; vecSize++)
1830 {
1831 for (int precNdx = 0; precNdx < glu::PRECISION_LAST; precNdx++)
1832 {
1833 const glu::DataType dataType = vecSize > 1 ? glu::getDataTypeFloatVec(vecSize) : glu::TYPE_FLOAT;
1834 const glu::Precision precision = glu::Precision(precNdx);
1835 const deUint32 hint = GL_DONT_CARE;
1836 ostringstream caseName;
1837
1838 if (surfaceType != SURFACETYPE_FLOAT_FBO && precision == glu::PRECISION_LOWP)
1839 continue; // Skip as lowp doesn't actually produce any bits when rendered to U8 RT.
1840
1841 caseName << glu::getDataTypeName(dataType) << "_" << glu::getPrecisionName(precision);
1842
1843 fboGroup->addChild(new LinearDerivateCase(m_context, caseName.str().c_str(), "", function, dataType, precision, hint, surfaceType, numSamples, source));
1844 }
1845 }
1846 }
1847
1848 // .fastest, .nicest
1849 for (int hintCaseNdx = 0; hintCaseNdx < DE_LENGTH_OF_ARRAY(s_hints); hintCaseNdx++)
1850 {
1851 tcu::TestCaseGroup* const hintGroup = new tcu::TestCaseGroup(m_testCtx, s_hints[hintCaseNdx].name, "Shader derivate hints");
1852 const char* source = s_linearDerivateCases[0].source; // use source from .linear group
1853 const deUint32 hint = s_hints[hintCaseNdx].hint;
1854 functionGroup->addChild(hintGroup);
1855
1856 for (int fboCaseNdx = 0; fboCaseNdx < DE_LENGTH_OF_ARRAY(s_hintFboConfigs); fboCaseNdx++)
1857 {
1858 tcu::TestCaseGroup* const fboGroup = new tcu::TestCaseGroup(m_testCtx, s_hintFboConfigs[fboCaseNdx].name, "");
1859 const SurfaceType surfaceType = s_hintFboConfigs[fboCaseNdx].surfaceType;
1860 const int numSamples = s_hintFboConfigs[fboCaseNdx].numSamples;
1861 hintGroup->addChild(fboGroup);
1862
1863 for (int vecSize = 1; vecSize <= 4; vecSize++)
1864 {
1865 for (int precNdx = 0; precNdx < glu::PRECISION_LAST; precNdx++)
1866 {
1867 const glu::DataType dataType = vecSize > 1 ? glu::getDataTypeFloatVec(vecSize) : glu::TYPE_FLOAT;
1868 const glu::Precision precision = glu::Precision(precNdx);
1869 ostringstream caseName;
1870
1871 if (surfaceType != SURFACETYPE_FLOAT_FBO && precision == glu::PRECISION_LOWP)
1872 continue; // Skip as lowp doesn't actually produce any bits when rendered to U8 RT.
1873
1874 caseName << glu::getDataTypeName(dataType) << "_" << glu::getPrecisionName(precision);
1875
1876 fboGroup->addChild(new LinearDerivateCase(m_context, caseName.str().c_str(), "", function, dataType, precision, hint, surfaceType, numSamples, source));
1877 }
1878 }
1879 }
1880 }
1881
1882 // .texture
1883 {
1884 tcu::TestCaseGroup* const textureGroup = new tcu::TestCaseGroup(m_testCtx, "texture", "Derivate of texture lookup result");
1885 functionGroup->addChild(textureGroup);
1886
1887 for (int texCaseNdx = 0; texCaseNdx < DE_LENGTH_OF_ARRAY(s_textureConfigs); texCaseNdx++)
1888 {
1889 tcu::TestCaseGroup* const caseGroup = new tcu::TestCaseGroup(m_testCtx, s_textureConfigs[texCaseNdx].name, "");
1890 const SurfaceType surfaceType = s_textureConfigs[texCaseNdx].surfaceType;
1891 const int numSamples = s_textureConfigs[texCaseNdx].numSamples;
1892 const deUint32 hint = s_textureConfigs[texCaseNdx].hint;
1893 textureGroup->addChild(caseGroup);
1894
1895 for (int vecSize = 1; vecSize <= 4; vecSize++)
1896 {
1897 for (int precNdx = 0; precNdx < glu::PRECISION_LAST; precNdx++)
1898 {
1899 const glu::DataType dataType = vecSize > 1 ? glu::getDataTypeFloatVec(vecSize) : glu::TYPE_FLOAT;
1900 const glu::Precision precision = glu::Precision(precNdx);
1901 ostringstream caseName;
1902
1903 if (surfaceType != SURFACETYPE_FLOAT_FBO && precision == glu::PRECISION_LOWP)
1904 continue; // Skip as lowp doesn't actually produce any bits when rendered to U8 RT.
1905
1906 caseName << glu::getDataTypeName(dataType) << "_" << glu::getPrecisionName(precision);
1907
1908 caseGroup->addChild(new TextureDerivateCase(m_context, caseName.str().c_str(), "", function, dataType, precision, hint, surfaceType, numSamples));
1909 }
1910 }
1911 }
1912 }
1913 }
1914 }
1915
1916 } // Functional
1917 } // gles3
1918 } // deqp
1919