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 "function_list.h"
18 #include "test_functions.h"
19 #include "utility.h"
20 
21 #include <cstring>
22 
BuildKernel(const char * name,int vectorSize,cl_uint kernel_count,cl_kernel * k,cl_program * p,bool relaxedMode)23 static int BuildKernel(const char *name, int vectorSize, cl_uint kernel_count,
24                        cl_kernel *k, cl_program *p, bool relaxedMode)
25 {
26     const char *c[] = { "__kernel void math_kernel",
27                         sizeNames[vectorSize],
28                         "( __global float",
29                         sizeNames[vectorSize],
30                         "* out, __global float",
31                         sizeNames[vectorSize],
32                         "* in )\n"
33                         "{\n"
34                         "   size_t i = get_global_id(0);\n"
35                         "   out[i] = ",
36                         name,
37                         "( in[i] );\n"
38                         "}\n" };
39 
40     const char *c3[] = {
41         "__kernel void math_kernel",
42         sizeNames[vectorSize],
43         "( __global float* out, __global float* in)\n"
44         "{\n"
45         "   size_t i = get_global_id(0);\n"
46         "   if( i + 1 < get_global_size(0) )\n"
47         "   {\n"
48         "       float3 f0 = vload3( 0, in + 3 * i );\n"
49         "       f0 = ",
50         name,
51         "( f0 );\n"
52         "       vstore3( f0, 0, out + 3*i );\n"
53         "   }\n"
54         "   else\n"
55         "   {\n"
56         "       size_t parity = i & 1;   // Figure out how many elements are "
57         "left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two "
58         "buffer size \n"
59         "       float3 f0;\n"
60         "       switch( parity )\n"
61         "       {\n"
62         "           case 1:\n"
63         "               f0 = (float3)( in[3*i], NAN, NAN ); \n"
64         "               break;\n"
65         "           case 0:\n"
66         "               f0 = (float3)( in[3*i], in[3*i+1], NAN ); \n"
67         "               break;\n"
68         "       }\n"
69         "       f0 = ",
70         name,
71         "( f0 );\n"
72         "       switch( parity )\n"
73         "       {\n"
74         "           case 0:\n"
75         "               out[3*i+1] = f0.y; \n"
76         "               // fall through\n"
77         "           case 1:\n"
78         "               out[3*i] = f0.x; \n"
79         "               break;\n"
80         "       }\n"
81         "   }\n"
82         "}\n"
83     };
84 
85     const char **kern = c;
86     size_t kernSize = sizeof(c) / sizeof(c[0]);
87 
88     if (sizeValues[vectorSize] == 3)
89     {
90         kern = c3;
91         kernSize = sizeof(c3) / sizeof(c3[0]);
92     }
93 
94     char testName[32];
95     snprintf(testName, sizeof(testName) - 1, "math_kernel%s",
96              sizeNames[vectorSize]);
97 
98     return MakeKernels(kern, (cl_uint)kernSize, testName, kernel_count, k, p,
99                        relaxedMode);
100 }
101 
102 typedef struct BuildKernelInfo
103 {
104     cl_uint offset; // the first vector size to build
105     cl_uint kernel_count;
106     cl_kernel **kernels;
107     cl_program *programs;
108     const char *nameInCode;
109     bool relaxedMode; // Whether to build with -cl-fast-relaxed-math.
110 } BuildKernelInfo;
111 
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)112 static cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
113 {
114     BuildKernelInfo *info = (BuildKernelInfo *)p;
115     cl_uint i = info->offset + job_id;
116     return BuildKernel(info->nameInCode, i, info->kernel_count,
117                        info->kernels[i], info->programs + i, info->relaxedMode);
118 }
119 
120 // Thread specific data for a worker thread
121 typedef struct ThreadInfo
122 {
123     cl_mem inBuf; // input buffer for the thread
124     cl_mem outBuf[VECTOR_SIZE_COUNT]; // output buffers for the thread
125     float maxError; // max error value. Init to 0.
126     double maxErrorValue; // position of the max error value.  Init to 0.
127     cl_command_queue tQueue; // per thread command queue to improve performance
128 } ThreadInfo;
129 
130 typedef struct TestInfo
131 {
132     size_t subBufferSize; // Size of the sub-buffer in elements
133     const Func *f; // A pointer to the function info
134     cl_program programs[VECTOR_SIZE_COUNT]; // programs for various vector sizes
135     cl_kernel
136         *k[VECTOR_SIZE_COUNT]; // arrays of thread-specific kernels for each
137                                // worker thread:  k[vector_size][thread_id]
138     ThreadInfo *
139         tinfo; // An array of thread specific information for each worker thread
140     cl_uint threadCount; // Number of worker threads
141     cl_uint jobCount; // Number of jobs
142     cl_uint step; // step between each chunk and the next.
143     cl_uint scale; // stride between individual test values
144     float ulps; // max_allowed ulps
145     int ftz; // non-zero if running in flush to zero mode
146 
147     int isRangeLimited; // 1 if the function is only to be evaluated over a
148                         // range
149     float half_sin_cos_tan_limit;
150     bool relaxedMode; // True if test is running in relaxed mode, false
151                       // otherwise.
152 } TestInfo;
153 
154 static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data);
155 
TestFunc_Float_Float(const Func * f,MTdata d,bool relaxedMode)156 int TestFunc_Float_Float(const Func *f, MTdata d, bool relaxedMode)
157 {
158     TestInfo test_info;
159     cl_int error;
160     float maxError = 0.0f;
161     double maxErrorVal = 0.0;
162     int skipTestingRelaxed = (relaxedMode && strcmp(f->name, "tan") == 0);
163 
164     logFunctionInfo(f->name, sizeof(cl_float), relaxedMode);
165 
166     // Init test_info
167     memset(&test_info, 0, sizeof(test_info));
168     test_info.threadCount = GetThreadCount();
169     test_info.subBufferSize = BUFFER_SIZE
170         / (sizeof(cl_float) * RoundUpToNextPowerOfTwo(test_info.threadCount));
171     test_info.scale = getTestScale(sizeof(cl_float));
172 
173     test_info.step = (cl_uint)test_info.subBufferSize * test_info.scale;
174     if (test_info.step / test_info.subBufferSize != test_info.scale)
175     {
176         // there was overflow
177         test_info.jobCount = 1;
178     }
179     else
180     {
181         test_info.jobCount = (cl_uint)((1ULL << 32) / test_info.step);
182     }
183 
184     test_info.f = f;
185     test_info.ulps = gIsEmbedded ? f->float_embedded_ulps : f->float_ulps;
186     test_info.ftz =
187         f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
188     test_info.relaxedMode = relaxedMode;
189     // cl_kernels aren't thread safe, so we make one for each vector size for
190     // every thread
191     for (auto i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++)
192     {
193         size_t array_size = test_info.threadCount * sizeof(cl_kernel);
194         test_info.k[i] = (cl_kernel *)malloc(array_size);
195         if (NULL == test_info.k[i])
196         {
197             vlog_error("Error: Unable to allocate storage for kernels!\n");
198             error = CL_OUT_OF_HOST_MEMORY;
199             goto exit;
200         }
201         memset(test_info.k[i], 0, array_size);
202     }
203     test_info.tinfo =
204         (ThreadInfo *)malloc(test_info.threadCount * sizeof(*test_info.tinfo));
205     if (NULL == test_info.tinfo)
206     {
207         vlog_error(
208             "Error: Unable to allocate storage for thread specific data.\n");
209         error = CL_OUT_OF_HOST_MEMORY;
210         goto exit;
211     }
212     memset(test_info.tinfo, 0,
213            test_info.threadCount * sizeof(*test_info.tinfo));
214     for (cl_uint i = 0; i < test_info.threadCount; i++)
215     {
216         cl_buffer_region region = {
217             i * test_info.subBufferSize * sizeof(cl_float),
218             test_info.subBufferSize * sizeof(cl_float)
219         };
220         test_info.tinfo[i].inBuf =
221             clCreateSubBuffer(gInBuffer, CL_MEM_READ_ONLY,
222                               CL_BUFFER_CREATE_TYPE_REGION, &region, &error);
223         if (error || NULL == test_info.tinfo[i].inBuf)
224         {
225             vlog_error("Error: Unable to create sub-buffer of gInBuffer for "
226                        "region {%zd, %zd}\n",
227                        region.origin, region.size);
228             goto exit;
229         }
230 
231         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
232         {
233             test_info.tinfo[i].outBuf[j] = clCreateSubBuffer(
234                 gOutBuffer[j], CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION,
235                 &region, &error);
236             if (error || NULL == test_info.tinfo[i].outBuf[j])
237             {
238                 vlog_error("Error: Unable to create sub-buffer of "
239                            "gOutBuffer[%d] for region {%zd, %zd}\n",
240                            (int)j, region.origin, region.size);
241                 goto exit;
242             }
243         }
244         test_info.tinfo[i].tQueue =
245             clCreateCommandQueue(gContext, gDevice, 0, &error);
246         if (NULL == test_info.tinfo[i].tQueue || error)
247         {
248             vlog_error("clCreateCommandQueue failed. (%d)\n", error);
249             goto exit;
250         }
251     }
252 
253     // Check for special cases for unary float
254     test_info.isRangeLimited = 0;
255     test_info.half_sin_cos_tan_limit = 0;
256     if (0 == strcmp(f->name, "half_sin") || 0 == strcmp(f->name, "half_cos"))
257     {
258         test_info.isRangeLimited = 1;
259         test_info.half_sin_cos_tan_limit = 1.0f
260             + test_info.ulps
261                 * (FLT_EPSILON / 2.0f); // out of range results from finite
262                                         // inputs must be in [-1,1]
263     }
264     else if (0 == strcmp(f->name, "half_tan"))
265     {
266         test_info.isRangeLimited = 1;
267         test_info.half_sin_cos_tan_limit =
268             INFINITY; // out of range resut from finite inputs must be numeric
269     }
270 
271     // Init the kernels
272     {
273         BuildKernelInfo build_info = {
274             gMinVectorSizeIndex, test_info.threadCount, test_info.k,
275             test_info.programs,  f->nameInCode,         relaxedMode
276         };
277         if ((error = ThreadPool_Do(BuildKernelFn,
278                                    gMaxVectorSizeIndex - gMinVectorSizeIndex,
279                                    &build_info)))
280             goto exit;
281     }
282 
283     // Run the kernels
284     if (!gSkipCorrectnessTesting || skipTestingRelaxed)
285     {
286         error = ThreadPool_Do(Test, test_info.jobCount, &test_info);
287 
288         // Accumulate the arithmetic errors
289         for (cl_uint i = 0; i < test_info.threadCount; i++)
290         {
291             if (test_info.tinfo[i].maxError > maxError)
292             {
293                 maxError = test_info.tinfo[i].maxError;
294                 maxErrorVal = test_info.tinfo[i].maxErrorValue;
295             }
296         }
297 
298         if (error) goto exit;
299 
300         if (gWimpyMode)
301             vlog("Wimp pass");
302         else
303             vlog("passed");
304 
305         if (skipTestingRelaxed)
306         {
307             vlog(" (rlx skip correctness testing)\n");
308             goto exit;
309         }
310 
311         vlog("\t%8.2f @ %a", maxError, maxErrorVal);
312     }
313 
314     vlog("\n");
315 
316 exit:
317     // Release
318     for (auto i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++)
319     {
320         clReleaseProgram(test_info.programs[i]);
321         if (test_info.k[i])
322         {
323             for (cl_uint j = 0; j < test_info.threadCount; j++)
324                 clReleaseKernel(test_info.k[i][j]);
325 
326             free(test_info.k[i]);
327         }
328     }
329     if (test_info.tinfo)
330     {
331         for (cl_uint i = 0; i < test_info.threadCount; i++)
332         {
333             clReleaseMemObject(test_info.tinfo[i].inBuf);
334             for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
335                 clReleaseMemObject(test_info.tinfo[i].outBuf[j]);
336             clReleaseCommandQueue(test_info.tinfo[i].tQueue);
337         }
338 
339         free(test_info.tinfo);
340     }
341 
342     return error;
343 }
344 
Test(cl_uint job_id,cl_uint thread_id,void * data)345 static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
346 {
347     const TestInfo *job = (const TestInfo *)data;
348     size_t buffer_elements = job->subBufferSize;
349     size_t buffer_size = buffer_elements * sizeof(cl_float);
350     cl_uint scale = job->scale;
351     cl_uint base = job_id * (cl_uint)job->step;
352     ThreadInfo *tinfo = job->tinfo + thread_id;
353     fptr func = job->f->func;
354     const char *fname = job->f->name;
355     bool relaxedMode = job->relaxedMode;
356     float ulps = getAllowedUlpError(job->f, relaxedMode);
357     if (relaxedMode)
358     {
359         func = job->f->rfunc;
360     }
361 
362     cl_int error;
363 
364     int isRangeLimited = job->isRangeLimited;
365     float half_sin_cos_tan_limit = job->half_sin_cos_tan_limit;
366     int ftz = job->ftz;
367 
368     // start the map of the output arrays
369     cl_event e[VECTOR_SIZE_COUNT];
370     cl_uint *out[VECTOR_SIZE_COUNT];
371     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
372     {
373         out[j] = (cl_uint *)clEnqueueMapBuffer(
374             tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_WRITE, 0,
375             buffer_size, 0, NULL, e + j, &error);
376         if (error || NULL == out[j])
377         {
378             vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j,
379                        error);
380             return error;
381         }
382     }
383 
384     // Get that moving
385     if ((error = clFlush(tinfo->tQueue))) vlog("clFlush failed\n");
386 
387     // Write the new values to the input array
388     cl_uint *p = (cl_uint *)gIn + thread_id * buffer_elements;
389     for (size_t j = 0; j < buffer_elements; j++)
390     {
391         p[j] = base + j * scale;
392         if (relaxedMode)
393         {
394             float p_j = *(float *)&p[j];
395             if (strcmp(fname, "sin") == 0
396                 || strcmp(fname, "cos")
397                     == 0) // the domain of the function is [-pi,pi]
398             {
399                 if (fabs(p_j) > M_PI) ((float *)p)[j] = NAN;
400             }
401 
402             if (strcmp(fname, "reciprocal") == 0)
403             {
404                 const float l_limit = HEX_FLT(+, 1, 0, -, 126);
405                 const float u_limit = HEX_FLT(+, 1, 0, +, 126);
406 
407                 if (fabs(p_j) < l_limit
408                     || fabs(p_j) > u_limit) // the domain of the function is
409                                             // [2^-126,2^126]
410                     ((float *)p)[j] = NAN;
411             }
412         }
413     }
414 
415     if ((error = clEnqueueWriteBuffer(tinfo->tQueue, tinfo->inBuf, CL_FALSE, 0,
416                                       buffer_size, p, 0, NULL, NULL)))
417     {
418         vlog_error("Error: clEnqueueWriteBuffer failed! err: %d\n", error);
419         return error;
420     }
421 
422     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
423     {
424         // Wait for the map to finish
425         if ((error = clWaitForEvents(1, e + j)))
426         {
427             vlog_error("Error: clWaitForEvents failed! err: %d\n", error);
428             return error;
429         }
430         if ((error = clReleaseEvent(e[j])))
431         {
432             vlog_error("Error: clReleaseEvent failed! err: %d\n", error);
433             return error;
434         }
435 
436         // Fill the result buffer with garbage, so that old results don't carry
437         // over
438         uint32_t pattern = 0xffffdead;
439         memset_pattern4(out[j], &pattern, buffer_size);
440         if ((error = clEnqueueUnmapMemObject(tinfo->tQueue, tinfo->outBuf[j],
441                                              out[j], 0, NULL, NULL)))
442         {
443             vlog_error("Error: clEnqueueMapBuffer failed! err: %d\n", error);
444             return error;
445         }
446 
447         // run the kernel
448         size_t vectorCount =
449             (buffer_elements + sizeValues[j] - 1) / sizeValues[j];
450         cl_kernel kernel = job->k[j][thread_id]; // each worker thread has its
451                                                  // own copy of the cl_kernel
452         cl_program program = job->programs[j];
453 
454         if ((error = clSetKernelArg(kernel, 0, sizeof(tinfo->outBuf[j]),
455                                     &tinfo->outBuf[j])))
456         {
457             LogBuildError(program);
458             return error;
459         }
460         if ((error = clSetKernelArg(kernel, 1, sizeof(tinfo->inBuf),
461                                     &tinfo->inBuf)))
462         {
463             LogBuildError(program);
464             return error;
465         }
466 
467         if ((error = clEnqueueNDRangeKernel(tinfo->tQueue, kernel, 1, NULL,
468                                             &vectorCount, NULL, 0, NULL, NULL)))
469         {
470             vlog_error("FAILED -- could not execute kernel\n");
471             return error;
472         }
473     }
474 
475     // Get that moving
476     if ((error = clFlush(tinfo->tQueue))) vlog("clFlush 2 failed\n");
477 
478     if (gSkipCorrectnessTesting) return CL_SUCCESS;
479 
480     // Calculate the correctly rounded reference result
481     float *r = (float *)gOut_Ref + thread_id * buffer_elements;
482     float *s = (float *)p;
483     for (size_t j = 0; j < buffer_elements; j++) r[j] = (float)func.f_f(s[j]);
484 
485     // Read the data back -- no need to wait for the first N-1 buffers but wait
486     // for the last buffer. This is an in order queue.
487     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
488     {
489         cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
490         out[j] = (cl_uint *)clEnqueueMapBuffer(
491             tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
492             buffer_size, 0, NULL, NULL, &error);
493         if (error || NULL == out[j])
494         {
495             vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j,
496                        error);
497             return error;
498         }
499     }
500 
501     // Verify data
502     uint32_t *t = (uint32_t *)r;
503     for (size_t j = 0; j < buffer_elements; j++)
504     {
505         for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
506         {
507             uint32_t *q = out[k];
508 
509             // If we aren't getting the correctly rounded result
510             if (t[j] != q[j])
511             {
512                 float test = ((float *)q)[j];
513                 double correct = func.f_f(s[j]);
514                 float err = Ulp_Error(test, correct);
515                 float abs_error = Abs_Error(test, correct);
516                 int fail = 0;
517                 int use_abs_error = 0;
518 
519                 // it is possible for the output to not match the reference
520                 // result but for Ulp_Error to be zero, for example -1.#QNAN
521                 // vs. 1.#QNAN. In such cases there is no failure
522                 if (err == 0.0f)
523                 {
524                     fail = 0;
525                 }
526                 else if (relaxedMode)
527                 {
528                     if (strcmp(fname, "sin") == 0 || strcmp(fname, "cos") == 0)
529                     {
530                         fail = !(fabsf(abs_error) <= ulps);
531                         use_abs_error = 1;
532                     }
533                     if (strcmp(fname, "sinpi") == 0
534                         || strcmp(fname, "cospi") == 0)
535                     {
536                         if (s[j] >= -1.0 && s[j] <= 1.0)
537                         {
538                             fail = !(fabsf(abs_error) <= ulps);
539                             use_abs_error = 1;
540                         }
541                     }
542 
543                     if (strcmp(fname, "reciprocal") == 0)
544                     {
545                         fail = !(fabsf(err) <= ulps);
546                     }
547 
548                     if (strcmp(fname, "exp") == 0 || strcmp(fname, "exp2") == 0)
549                     {
550                         float exp_error = ulps;
551 
552                         if (!gIsEmbedded)
553                         {
554                             exp_error += floor(fabs(2 * s[j]));
555                         }
556 
557                         fail = !(fabsf(err) <= exp_error);
558                         ulps = exp_error;
559                     }
560                     if (strcmp(fname, "tan") == 0)
561                     {
562 
563                         if (!gFastRelaxedDerived)
564                         {
565                             fail = !(fabsf(err) <= ulps);
566                         }
567                         // Else fast math derived implementation does not
568                         // require ULP verification
569                     }
570                     if (strcmp(fname, "exp10") == 0)
571                     {
572                         if (!gFastRelaxedDerived)
573                         {
574                             fail = !(fabsf(err) <= ulps);
575                         }
576                         // Else fast math derived implementation does not
577                         // require ULP verification
578                     }
579                     if (strcmp(fname, "log") == 0 || strcmp(fname, "log2") == 0
580                         || strcmp(fname, "log10") == 0)
581                     {
582                         if (s[j] >= 0.5 && s[j] <= 2)
583                         {
584                             fail = !(fabsf(abs_error) <= ulps);
585                         }
586                         else
587                         {
588                             ulps = gIsEmbedded ? job->f->float_embedded_ulps
589                                                : job->f->float_ulps;
590                             fail = !(fabsf(err) <= ulps);
591                         }
592                     }
593 
594 
595                     // fast-relaxed implies finite-only
596                     if (IsFloatInfinity(correct) || IsFloatNaN(correct)
597                         || IsFloatInfinity(s[j]) || IsFloatNaN(s[j]))
598                     {
599                         fail = 0;
600                         err = 0;
601                     }
602                 }
603                 else
604                 {
605                     fail = !(fabsf(err) <= ulps);
606                 }
607 
608                 // half_sin/cos/tan are only valid between +-2**16, Inf, NaN
609                 if (isRangeLimited
610                     && fabsf(s[j]) > MAKE_HEX_FLOAT(0x1.0p16f, 0x1L, 16)
611                     && fabsf(s[j]) < INFINITY)
612                 {
613                     if (fabsf(test) <= half_sin_cos_tan_limit)
614                     {
615                         err = 0;
616                         fail = 0;
617                     }
618                 }
619 
620                 if (fail)
621                 {
622                     if (ftz)
623                     {
624                         typedef int (*CheckForSubnormal)(
625                             double, float); // If we are in fast relaxed math,
626                                             // we have a different calculation
627                                             // for the subnormal threshold.
628                         CheckForSubnormal isFloatResultSubnormalPtr;
629 
630                         if (relaxedMode)
631                         {
632                             isFloatResultSubnormalPtr =
633                                 &IsFloatResultSubnormalAbsError;
634                         }
635                         else
636                         {
637                             isFloatResultSubnormalPtr = &IsFloatResultSubnormal;
638                         }
639                         // retry per section 6.5.3.2
640                         if ((*isFloatResultSubnormalPtr)(correct, ulps))
641                         {
642                             fail = fail && (test != 0.0f);
643                             if (!fail) err = 0.0f;
644                         }
645 
646                         // retry per section 6.5.3.3
647                         if (IsFloatSubnormal(s[j]))
648                         {
649                             double correct2 = func.f_f(0.0);
650                             double correct3 = func.f_f(-0.0);
651                             float err2;
652                             float err3;
653                             if (use_abs_error)
654                             {
655                                 err2 = Abs_Error(test, correct2);
656                                 err3 = Abs_Error(test, correct3);
657                             }
658                             else
659                             {
660                                 err2 = Ulp_Error(test, correct2);
661                                 err3 = Ulp_Error(test, correct3);
662                             }
663                             fail = fail
664                                 && ((!(fabsf(err2) <= ulps))
665                                     && (!(fabsf(err3) <= ulps)));
666                             if (fabsf(err2) < fabsf(err)) err = err2;
667                             if (fabsf(err3) < fabsf(err)) err = err3;
668 
669                             // retry per section 6.5.3.4
670                             if ((*isFloatResultSubnormalPtr)(correct2, ulps)
671                                 || (*isFloatResultSubnormalPtr)(correct3, ulps))
672                             {
673                                 fail = fail && (test != 0.0f);
674                                 if (!fail) err = 0.0f;
675                             }
676                         }
677                     }
678                 }
679                 if (fabsf(err) > tinfo->maxError)
680                 {
681                     tinfo->maxError = fabsf(err);
682                     tinfo->maxErrorValue = s[j];
683                 }
684                 if (fail)
685                 {
686                     vlog_error("\nERROR: %s%s: %f ulp error at %a (0x%8.8x): "
687                                "*%a vs. %a\n",
688                                job->f->name, sizeNames[k], err, ((float *)s)[j],
689                                ((uint32_t *)s)[j], ((float *)t)[j], test);
690                     return -1;
691                 }
692             }
693         }
694     }
695 
696     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
697     {
698         if ((error = clEnqueueUnmapMemObject(tinfo->tQueue, tinfo->outBuf[j],
699                                              out[j], 0, NULL, NULL)))
700         {
701             vlog_error("Error: clEnqueueUnmapMemObject %d failed 2! err: %d\n",
702                        j, error);
703             return error;
704         }
705     }
706 
707     if ((error = clFlush(tinfo->tQueue))) vlog("clFlush 3 failed\n");
708 
709 
710     if (0 == (base & 0x0fffffff))
711     {
712         if (gVerboseBruteForce)
713         {
714             vlog("base:%14u step:%10u scale:%10u buf_elements:%10zd ulps:%5.3f "
715                  "ThreadCount:%2u\n",
716                  base, job->step, job->scale, buffer_elements, job->ulps,
717                  job->threadCount);
718         }
719         else
720         {
721             vlog(".");
722         }
723         fflush(stdout);
724     }
725 
726     return CL_SUCCESS;
727 }
728