1 //
2 // Copyright (c) 2017 The Khronos Group Inc.
3 //
4 // Licensed under the Apache License, Version 2.0 (the "License");
5 // you may not use this file except in compliance with the License.
6 // You may obtain a copy of the License at
7 //
8 // http://www.apache.org/licenses/LICENSE-2.0
9 //
10 // Unless required by applicable law or agreed to in writing, software
11 // distributed under the License is distributed on an "AS IS" BASIS,
12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 // See the License for the specific language governing permissions and
14 // limitations under the License.
15 //
16
17 #include "common.h"
18 #include "function_list.h"
19 #include "test_functions.h"
20 #include "utility.h"
21
22 #include <cinttypes>
23 #include <climits>
24 #include <cstring>
25
26 namespace {
27
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)28 cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
29 {
30 BuildKernelInfo &info = *(BuildKernelInfo *)p;
31 auto generator = [](const std::string &kernel_name, const char *builtin,
32 cl_uint vector_size_index) {
33 return GetBinaryKernel(kernel_name, builtin, ParameterType::Double,
34 ParameterType::Int, ParameterType::Double,
35 ParameterType::Double, vector_size_index);
36 };
37 return BuildKernels(info, job_id, generator);
38 }
39
40
41 struct ComputeReferenceInfoD
42 {
43 const double *x;
44 const double *y;
45 double *r;
46 int *i;
47 long double (*f_ffpI)(long double, long double, int *);
48 cl_uint lim;
49 cl_uint count;
50 };
51
ReferenceD(cl_uint jid,cl_uint tid,void * userInfo)52 cl_int ReferenceD(cl_uint jid, cl_uint tid, void *userInfo)
53 {
54 ComputeReferenceInfoD *cri = (ComputeReferenceInfoD *)userInfo;
55 cl_uint lim = cri->lim;
56 cl_uint count = cri->count;
57 cl_uint off = jid * count;
58 const double *x = cri->x + off;
59 const double *y = cri->y + off;
60 double *r = cri->r + off;
61 int *i = cri->i + off;
62 long double (*f)(long double, long double, int *) = cri->f_ffpI;
63
64 if (off + count > lim) count = lim - off;
65
66 Force64BitFPUPrecision();
67
68 for (cl_uint j = 0; j < count; ++j)
69 r[j] = (double)f((long double)x[j], (long double)y[j], i + j);
70
71 return CL_SUCCESS;
72 }
73
74 } // anonymous namespace
75
TestFunc_DoubleI_Double_Double(const Func * f,MTdata d,bool relaxedMode)76 int TestFunc_DoubleI_Double_Double(const Func *f, MTdata d, bool relaxedMode)
77 {
78 int error;
79 Programs programs;
80 const unsigned thread_id = 0; // Test is currently not multithreaded.
81 KernelMatrix kernels;
82 float maxError = 0.0f;
83 int64_t maxError2 = 0;
84 int ftz = f->ftz || gForceFTZ;
85 double maxErrorVal = 0.0f;
86 double maxErrorVal2 = 0.0f;
87 uint64_t step = getTestStep(sizeof(double), BUFFER_SIZE);
88
89 logFunctionInfo(f->name, sizeof(cl_double), relaxedMode);
90
91 cl_uint threadCount = GetThreadCount();
92
93 Force64BitFPUPrecision();
94
95 int testingRemquo = !strcmp(f->name, "remquo");
96
97 // Init the kernels
98 BuildKernelInfo build_info{ 1, kernels, programs, f->nameInCode,
99 relaxedMode };
100 if ((error = ThreadPool_Do(BuildKernelFn,
101 gMaxVectorSizeIndex - gMinVectorSizeIndex,
102 &build_info)))
103 return error;
104
105 for (uint64_t i = 0; i < (1ULL << 32); i += step)
106 {
107 // Init input array
108 double *p = (double *)gIn;
109 double *p2 = (double *)gIn2;
110 for (size_t j = 0; j < BUFFER_SIZE / sizeof(double); j++)
111 {
112 p[j] = DoubleFromUInt32(genrand_int32(d));
113 p2[j] = DoubleFromUInt32(genrand_int32(d));
114 }
115
116 if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0,
117 BUFFER_SIZE, gIn, 0, NULL, NULL)))
118 {
119 vlog_error("\n*** Error %d in clEnqueueWriteBuffer ***\n", error);
120 return error;
121 }
122
123 if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer2, CL_FALSE, 0,
124 BUFFER_SIZE, gIn2, 0, NULL, NULL)))
125 {
126 vlog_error("\n*** Error %d in clEnqueueWriteBuffer2 ***\n", error);
127 return error;
128 }
129
130 // Write garbage into output arrays
131 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
132 {
133 uint32_t pattern = 0xffffdead;
134 if (gHostFill)
135 {
136 memset_pattern4(gOut[j], &pattern, BUFFER_SIZE);
137 if ((error = clEnqueueWriteBuffer(gQueue, gOutBuffer[j],
138 CL_FALSE, 0, BUFFER_SIZE,
139 gOut[j], 0, NULL, NULL)))
140 {
141 vlog_error(
142 "\n*** Error %d in clEnqueueWriteBuffer2(%d) ***\n",
143 error, j);
144 return error;
145 }
146
147 memset_pattern4(gOut2[j], &pattern, BUFFER_SIZE);
148 if ((error = clEnqueueWriteBuffer(gQueue, gOutBuffer2[j],
149 CL_FALSE, 0, BUFFER_SIZE,
150 gOut2[j], 0, NULL, NULL)))
151 {
152 vlog_error(
153 "\n*** Error %d in clEnqueueWriteBuffer2b(%d) ***\n",
154 error, j);
155 return error;
156 }
157 }
158 else
159 {
160 if ((error = clEnqueueFillBuffer(gQueue, gOutBuffer[j],
161 &pattern, sizeof(pattern), 0,
162 BUFFER_SIZE, 0, NULL, NULL)))
163 {
164 vlog_error("Error: clEnqueueFillBuffer 1 failed! err: %d\n",
165 error);
166 return error;
167 }
168
169 if ((error = clEnqueueFillBuffer(gQueue, gOutBuffer2[j],
170 &pattern, sizeof(pattern), 0,
171 BUFFER_SIZE, 0, NULL, NULL)))
172 {
173 vlog_error("Error: clEnqueueFillBuffer 2 failed! err: %d\n",
174 error);
175 return error;
176 }
177 }
178 }
179
180 // Run the kernels
181 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
182 {
183 size_t vectorSize = sizeof(cl_double) * sizeValues[j];
184 size_t localCount = (BUFFER_SIZE + vectorSize - 1)
185 / vectorSize; // BUFFER_SIZE / vectorSize rounded up
186 if ((error = clSetKernelArg(kernels[j][thread_id], 0,
187 sizeof(gOutBuffer[j]), &gOutBuffer[j])))
188 {
189 LogBuildError(programs[j]);
190 return error;
191 }
192 if ((error =
193 clSetKernelArg(kernels[j][thread_id], 1,
194 sizeof(gOutBuffer2[j]), &gOutBuffer2[j])))
195 {
196 LogBuildError(programs[j]);
197 return error;
198 }
199 if ((error = clSetKernelArg(kernels[j][thread_id], 2,
200 sizeof(gInBuffer), &gInBuffer)))
201 {
202 LogBuildError(programs[j]);
203 return error;
204 }
205 if ((error = clSetKernelArg(kernels[j][thread_id], 3,
206 sizeof(gInBuffer2), &gInBuffer2)))
207 {
208 LogBuildError(programs[j]);
209 return error;
210 }
211
212 if ((error = clEnqueueNDRangeKernel(gQueue, kernels[j][thread_id],
213 1, NULL, &localCount, NULL, 0,
214 NULL, NULL)))
215 {
216 vlog_error("FAILED -- could not execute kernel\n");
217 return error;
218 }
219 }
220
221 // Get that moving
222 if ((error = clFlush(gQueue))) vlog("clFlush failed\n");
223
224 // Calculate the correctly rounded reference result
225 double *s = (double *)gIn;
226 double *s2 = (double *)gIn2;
227
228 if (threadCount > 1)
229 {
230 ComputeReferenceInfoD cri;
231 cri.x = s;
232 cri.y = s2;
233 cri.r = (double *)gOut_Ref;
234 cri.i = (int *)gOut_Ref2;
235 cri.f_ffpI = f->dfunc.f_ffpI;
236 cri.lim = BUFFER_SIZE / sizeof(double);
237 cri.count = (cri.lim + threadCount - 1) / threadCount;
238 ThreadPool_Do(ReferenceD, threadCount, &cri);
239 }
240 else
241 {
242 double *r = (double *)gOut_Ref;
243 int *r2 = (int *)gOut_Ref2;
244 for (size_t j = 0; j < BUFFER_SIZE / sizeof(double); j++)
245 r[j] = (double)f->dfunc.f_ffpI(s[j], s2[j], r2 + j);
246 }
247
248 // Read the data back
249 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
250 {
251 if ((error =
252 clEnqueueReadBuffer(gQueue, gOutBuffer[j], CL_TRUE, 0,
253 BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
254 {
255 vlog_error("ReadArray failed %d\n", error);
256 return error;
257 }
258 if ((error =
259 clEnqueueReadBuffer(gQueue, gOutBuffer2[j], CL_TRUE, 0,
260 BUFFER_SIZE, gOut2[j], 0, NULL, NULL)))
261 {
262 vlog_error("ReadArray2 failed %d\n", error);
263 return error;
264 }
265 }
266
267 if (gSkipCorrectnessTesting) break;
268
269 // Verify data
270 uint64_t *t = (uint64_t *)gOut_Ref;
271 int32_t *t2 = (int32_t *)gOut_Ref2;
272 for (size_t j = 0; j < BUFFER_SIZE / sizeof(double); j++)
273 {
274 for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
275 {
276 uint64_t *q = (uint64_t *)gOut[k];
277 int32_t *q2 = (int32_t *)gOut2[k];
278
279 // Check for exact match to correctly rounded result
280 if (t[j] == q[j] && t2[j] == q2[j]) continue;
281
282 // Check for paired NaNs
283 if ((t[j] & 0x7fffffffffffffffUL) > 0x7ff0000000000000UL
284 && (q[j] & 0x7fffffffffffffffUL) > 0x7ff0000000000000UL
285 && t2[j] == q2[j])
286 continue;
287
288 double test = ((double *)q)[j];
289 int correct2 = INT_MIN;
290 long double correct = f->dfunc.f_ffpI(s[j], s2[j], &correct2);
291 float err = Bruteforce_Ulp_Error_Double(test, correct);
292 int64_t iErr;
293
294 // in case of remquo, we only care about the sign and last
295 // seven bits of integer as per the spec.
296 if (testingRemquo)
297 iErr = (long long)(q2[j] & 0x0000007f)
298 - (long long)(correct2 & 0x0000007f);
299 else
300 iErr = (long long)q2[j] - (long long)correct2;
301
302 // For remquo, if y = 0, x is infinite, or either is NaN
303 // then the standard either neglects to say what is returned
304 // in iptr or leaves it undefined or implementation defined.
305 int iptrUndefined = fabs(((double *)gIn)[j]) == INFINITY
306 || ((double *)gIn2)[j] == 0.0 || isnan(((double *)gIn2)[j])
307 || isnan(((double *)gIn)[j]);
308 if (iptrUndefined) iErr = 0;
309
310 int fail = !(fabsf(err) <= f->double_ulps && iErr == 0);
311 if ((ftz || relaxedMode) && fail)
312 {
313 // retry per section 6.5.3.2
314 if (IsDoubleResultSubnormal(correct, f->double_ulps))
315 {
316 fail = fail && !(test == 0.0f && iErr == 0);
317 if (!fail) err = 0.0f;
318 }
319
320 // retry per section 6.5.3.3
321 if (IsDoubleSubnormal(s[j]))
322 {
323 int correct3i, correct4i;
324 long double correct3 =
325 f->dfunc.f_ffpI(0.0, s2[j], &correct3i);
326 long double correct4 =
327 f->dfunc.f_ffpI(-0.0, s2[j], &correct4i);
328 float err2 =
329 Bruteforce_Ulp_Error_Double(test, correct3);
330 float err3 =
331 Bruteforce_Ulp_Error_Double(test, correct4);
332 int64_t iErr3 = (long long)q2[j] - (long long)correct3i;
333 int64_t iErr4 = (long long)q2[j] - (long long)correct4i;
334 fail = fail
335 && ((!(fabsf(err2) <= f->double_ulps && iErr3 == 0))
336 && (!(fabsf(err3) <= f->double_ulps
337 && iErr4 == 0)));
338 if (fabsf(err2) < fabsf(err)) err = err2;
339 if (fabsf(err3) < fabsf(err)) err = err3;
340 if (llabs(iErr3) < llabs(iErr)) iErr = iErr3;
341 if (llabs(iErr4) < llabs(iErr)) iErr = iErr4;
342
343 // retry per section 6.5.3.4
344 if (IsDoubleResultSubnormal(correct2, f->double_ulps)
345 || IsDoubleResultSubnormal(correct3,
346 f->double_ulps))
347 {
348 fail = fail
349 && !(test == 0.0f
350 && (iErr3 == 0 || iErr4 == 0));
351 if (!fail) err = 0.0f;
352 }
353
354 // try with both args as zero
355 if (IsDoubleSubnormal(s2[j]))
356 {
357 int correct7i, correct8i;
358 correct3 = f->dfunc.f_ffpI(0.0, 0.0, &correct3i);
359 correct4 = f->dfunc.f_ffpI(-0.0, 0.0, &correct4i);
360 long double correct7 =
361 f->dfunc.f_ffpI(0.0, -0.0, &correct7i);
362 long double correct8 =
363 f->dfunc.f_ffpI(-0.0, -0.0, &correct8i);
364 err2 = Bruteforce_Ulp_Error_Double(test, correct3);
365 err3 = Bruteforce_Ulp_Error_Double(test, correct4);
366 float err4 =
367 Bruteforce_Ulp_Error_Double(test, correct7);
368 float err5 =
369 Bruteforce_Ulp_Error_Double(test, correct8);
370 iErr3 = (long long)q2[j] - (long long)correct3i;
371 iErr4 = (long long)q2[j] - (long long)correct4i;
372 int64_t iErr7 =
373 (long long)q2[j] - (long long)correct7i;
374 int64_t iErr8 =
375 (long long)q2[j] - (long long)correct8i;
376 fail = fail
377 && ((!(fabsf(err2) <= f->double_ulps
378 && iErr3 == 0))
379 && (!(fabsf(err3) <= f->double_ulps
380 && iErr4 == 0))
381 && (!(fabsf(err4) <= f->double_ulps
382 && iErr7 == 0))
383 && (!(fabsf(err5) <= f->double_ulps
384 && iErr8 == 0)));
385 if (fabsf(err2) < fabsf(err)) err = err2;
386 if (fabsf(err3) < fabsf(err)) err = err3;
387 if (fabsf(err4) < fabsf(err)) err = err4;
388 if (fabsf(err5) < fabsf(err)) err = err5;
389 if (llabs(iErr3) < llabs(iErr)) iErr = iErr3;
390 if (llabs(iErr4) < llabs(iErr)) iErr = iErr4;
391 if (llabs(iErr7) < llabs(iErr)) iErr = iErr7;
392 if (llabs(iErr8) < llabs(iErr)) iErr = iErr8;
393
394 // retry per section 6.5.3.4
395 if (IsDoubleResultSubnormal(correct3,
396 f->double_ulps)
397 || IsDoubleResultSubnormal(correct4,
398 f->double_ulps)
399 || IsDoubleResultSubnormal(correct7,
400 f->double_ulps)
401 || IsDoubleResultSubnormal(correct8,
402 f->double_ulps))
403 {
404 fail = fail
405 && !(test == 0.0f
406 && (iErr3 == 0 || iErr4 == 0
407 || iErr7 == 0 || iErr8 == 0));
408 if (!fail) err = 0.0f;
409 }
410 }
411 }
412 else if (IsDoubleSubnormal(s2[j]))
413 {
414 int correct3i, correct4i;
415 long double correct3 =
416 f->dfunc.f_ffpI(s[j], 0.0, &correct3i);
417 long double correct4 =
418 f->dfunc.f_ffpI(s[j], -0.0, &correct4i);
419 float err2 =
420 Bruteforce_Ulp_Error_Double(test, correct3);
421 float err3 =
422 Bruteforce_Ulp_Error_Double(test, correct4);
423 int64_t iErr3 = (long long)q2[j] - (long long)correct3i;
424 int64_t iErr4 = (long long)q2[j] - (long long)correct4i;
425 fail = fail
426 && ((!(fabsf(err2) <= f->double_ulps && iErr3 == 0))
427 && (!(fabsf(err3) <= f->double_ulps
428 && iErr4 == 0)));
429 if (fabsf(err2) < fabsf(err)) err = err2;
430 if (fabsf(err3) < fabsf(err)) err = err3;
431 if (llabs(iErr3) < llabs(iErr)) iErr = iErr3;
432 if (llabs(iErr4) < llabs(iErr)) iErr = iErr4;
433
434 // retry per section 6.5.3.4
435 if (IsDoubleResultSubnormal(correct2, f->double_ulps)
436 || IsDoubleResultSubnormal(correct3,
437 f->double_ulps))
438 {
439 fail = fail
440 && !(test == 0.0f
441 && (iErr3 == 0 || iErr4 == 0));
442 if (!fail) err = 0.0f;
443 }
444 }
445 }
446 if (fabsf(err) > maxError)
447 {
448 maxError = fabsf(err);
449 maxErrorVal = s[j];
450 }
451 if (llabs(iErr) > maxError2)
452 {
453 maxError2 = llabs(iErr);
454 maxErrorVal2 = s[j];
455 }
456
457 if (fail)
458 {
459 vlog_error("\nERROR: %sD%s: {%f, %" PRId64
460 "} ulp error at {%.13la, "
461 "%.13la} ({ 0x%16.16" PRIx64 ", 0x%16.16" PRIx64
462 "}): *{%.13la, "
463 "%d} ({ 0x%16.16" PRIx64
464 ", 0x%8.8x}) vs. {%.13la, %d} ({ "
465 "0x%16.16" PRIx64 ", 0x%8.8x})\n",
466 f->name, sizeNames[k], err, iErr,
467 ((double *)gIn)[j], ((double *)gIn2)[j],
468 ((cl_ulong *)gIn)[j], ((cl_ulong *)gIn2)[j],
469 ((double *)gOut_Ref)[j], ((int *)gOut_Ref2)[j],
470 ((cl_ulong *)gOut_Ref)[j],
471 ((cl_uint *)gOut_Ref2)[j], test, q2[j],
472 ((cl_ulong *)q)[j], ((cl_uint *)q2)[j]);
473 return -1;
474 }
475 }
476 }
477
478 if (0 == (i & 0x0fffffff))
479 {
480 if (gVerboseBruteForce)
481 {
482 vlog("base:%14" PRIu64 " step:%10" PRIu64
483 " bufferSize:%10d \n",
484 i, step, BUFFER_SIZE);
485 }
486 else
487 {
488 vlog(".");
489 }
490 fflush(stdout);
491 }
492 }
493
494 if (!gSkipCorrectnessTesting)
495 {
496 if (gWimpyMode)
497 vlog("Wimp pass");
498 else
499 vlog("passed");
500
501 vlog("\t{%8.2f, %" PRId64 "} @ {%a, %a}", maxError, maxError2,
502 maxErrorVal, maxErrorVal2);
503 }
504
505 vlog("\n");
506
507 return CL_SUCCESS;
508 }
509