xref: /aosp_15_r20/external/OpenCL-CTS/test_conformance/math_brute_force/macro_binary_double.cpp (revision 6467f958c7de8070b317fc65bcb0f6472e388d82)
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 <cstring>
24 
25 namespace {
26 
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)27 cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
28 {
29     BuildKernelInfo &info = *(BuildKernelInfo *)p;
30     auto generator = [](const std::string &kernel_name, const char *builtin,
31                         cl_uint vector_size_index) {
32         return GetBinaryKernel(kernel_name, builtin, ParameterType::Long,
33                                ParameterType::Double, ParameterType::Double,
34                                vector_size_index);
35     };
36     return BuildKernels(info, job_id, generator);
37 }
38 
39 // Thread specific data for a worker thread
40 struct ThreadInfo
41 {
42     // Input and output buffers for the thread
43     clMemWrapper inBuf;
44     clMemWrapper inBuf2;
45     Buffers outBuf;
46 
47     MTdataHolder d;
48 
49     // Per thread command queue to improve performance
50     clCommandQueueWrapper tQueue;
51 };
52 
53 struct TestInfo
54 {
55     size_t subBufferSize; // Size of the sub-buffer in elements
56     const Func *f; // A pointer to the function info
57 
58     // Programs for various vector sizes.
59     Programs programs;
60 
61     // Thread-specific kernels for each vector size:
62     // k[vector_size][thread_id]
63     KernelMatrix k;
64 
65     // Array of thread specific information
66     std::vector<ThreadInfo> tinfo;
67 
68     cl_uint threadCount; // Number of worker threads
69     cl_uint jobCount; // Number of jobs
70     cl_uint step; // step between each chunk and the next.
71     cl_uint scale; // stride between individual test values
72     int ftz; // non-zero if running in flush to zero mode
73     bool relaxedMode; // True if test is running in relaxed mode, false
74                       // otherwise.
75 };
76 
77 // A table of more difficult cases to get right
78 const double specialValues[] = {
79     -NAN,
80     -INFINITY,
81     -DBL_MAX,
82     MAKE_HEX_DOUBLE(-0x1.0000000000001p64, -0x10000000000001LL, 12),
83     MAKE_HEX_DOUBLE(-0x1.0p64, -0x1LL, 64),
84     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp63, -0x1fffffffffffffLL, 11),
85     MAKE_HEX_DOUBLE(-0x1.0000000000001p63, -0x10000000000001LL, 11),
86     MAKE_HEX_DOUBLE(-0x1.0p63, -0x1LL, 63),
87     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp62, -0x1fffffffffffffLL, 10),
88     MAKE_HEX_DOUBLE(-0x1.000002p32, -0x1000002LL, 8),
89     MAKE_HEX_DOUBLE(-0x1.0p32, -0x1LL, 32),
90     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp31, -0x1fffffffffffffLL, -21),
91     MAKE_HEX_DOUBLE(-0x1.0000000000001p31, -0x10000000000001LL, -21),
92     MAKE_HEX_DOUBLE(-0x1.0p31, -0x1LL, 31),
93     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp30, -0x1fffffffffffffLL, -22),
94     -1000.0,
95     -100.0,
96     -4.0,
97     -3.5,
98     -3.0,
99     MAKE_HEX_DOUBLE(-0x1.8000000000001p1, -0x18000000000001LL, -51),
100     -2.5,
101     MAKE_HEX_DOUBLE(-0x1.7ffffffffffffp1, -0x17ffffffffffffLL, -51),
102     -2.0,
103     MAKE_HEX_DOUBLE(-0x1.8000000000001p0, -0x18000000000001LL, -52),
104     -1.5,
105     MAKE_HEX_DOUBLE(-0x1.7ffffffffffffp0, -0x17ffffffffffffLL, -52),
106     MAKE_HEX_DOUBLE(-0x1.0000000000001p0, -0x10000000000001LL, -52),
107     -1.0,
108     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp-1, -0x1fffffffffffffLL, -53),
109     MAKE_HEX_DOUBLE(-0x1.0000000000001p-1, -0x10000000000001LL, -53),
110     -0.5,
111     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp-2, -0x1fffffffffffffLL, -54),
112     MAKE_HEX_DOUBLE(-0x1.0000000000001p-2, -0x10000000000001LL, -54),
113     -0.25,
114     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp-3, -0x1fffffffffffffLL, -55),
115     MAKE_HEX_DOUBLE(-0x1.0000000000001p-1022, -0x10000000000001LL, -1074),
116     -DBL_MIN,
117     MAKE_HEX_DOUBLE(-0x0.fffffffffffffp-1022, -0x0fffffffffffffLL, -1074),
118     MAKE_HEX_DOUBLE(-0x0.0000000000fffp-1022, -0x00000000000fffLL, -1074),
119     MAKE_HEX_DOUBLE(-0x0.00000000000fep-1022, -0x000000000000feLL, -1074),
120     MAKE_HEX_DOUBLE(-0x0.000000000000ep-1022, -0x0000000000000eLL, -1074),
121     MAKE_HEX_DOUBLE(-0x0.000000000000cp-1022, -0x0000000000000cLL, -1074),
122     MAKE_HEX_DOUBLE(-0x0.000000000000ap-1022, -0x0000000000000aLL, -1074),
123     MAKE_HEX_DOUBLE(-0x0.0000000000008p-1022, -0x00000000000008LL, -1074),
124     MAKE_HEX_DOUBLE(-0x0.0000000000007p-1022, -0x00000000000007LL, -1074),
125     MAKE_HEX_DOUBLE(-0x0.0000000000006p-1022, -0x00000000000006LL, -1074),
126     MAKE_HEX_DOUBLE(-0x0.0000000000005p-1022, -0x00000000000005LL, -1074),
127     MAKE_HEX_DOUBLE(-0x0.0000000000004p-1022, -0x00000000000004LL, -1074),
128     MAKE_HEX_DOUBLE(-0x0.0000000000003p-1022, -0x00000000000003LL, -1074),
129     MAKE_HEX_DOUBLE(-0x0.0000000000002p-1022, -0x00000000000002LL, -1074),
130     MAKE_HEX_DOUBLE(-0x0.0000000000001p-1022, -0x00000000000001LL, -1074),
131     -0.0,
132 
133     +NAN,
134     +INFINITY,
135     +DBL_MAX,
136     MAKE_HEX_DOUBLE(+0x1.0000000000001p64, +0x10000000000001LL, 12),
137     MAKE_HEX_DOUBLE(+0x1.0p64, +0x1LL, 64),
138     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp63, +0x1fffffffffffffLL, 11),
139     MAKE_HEX_DOUBLE(+0x1.0000000000001p63, +0x10000000000001LL, 11),
140     MAKE_HEX_DOUBLE(+0x1.0p63, +0x1LL, 63),
141     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp62, +0x1fffffffffffffLL, 10),
142     MAKE_HEX_DOUBLE(+0x1.000002p32, +0x1000002LL, 8),
143     MAKE_HEX_DOUBLE(+0x1.0p32, +0x1LL, 32),
144     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp31, +0x1fffffffffffffLL, -21),
145     MAKE_HEX_DOUBLE(+0x1.0000000000001p31, +0x10000000000001LL, -21),
146     MAKE_HEX_DOUBLE(+0x1.0p31, +0x1LL, 31),
147     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp30, +0x1fffffffffffffLL, -22),
148     +1000.0,
149     +100.0,
150     +4.0,
151     +3.5,
152     +3.0,
153     MAKE_HEX_DOUBLE(+0x1.8000000000001p1, +0x18000000000001LL, -51),
154     +2.5,
155     MAKE_HEX_DOUBLE(+0x1.7ffffffffffffp1, +0x17ffffffffffffLL, -51),
156     +2.0,
157     MAKE_HEX_DOUBLE(+0x1.8000000000001p0, +0x18000000000001LL, -52),
158     +1.5,
159     MAKE_HEX_DOUBLE(+0x1.7ffffffffffffp0, +0x17ffffffffffffLL, -52),
160     MAKE_HEX_DOUBLE(-0x1.0000000000001p0, -0x10000000000001LL, -52),
161     +1.0,
162     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp-1, +0x1fffffffffffffLL, -53),
163     MAKE_HEX_DOUBLE(+0x1.0000000000001p-1, +0x10000000000001LL, -53),
164     +0.5,
165     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp-2, +0x1fffffffffffffLL, -54),
166     MAKE_HEX_DOUBLE(+0x1.0000000000001p-2, +0x10000000000001LL, -54),
167     +0.25,
168     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp-3, +0x1fffffffffffffLL, -55),
169     MAKE_HEX_DOUBLE(+0x1.0000000000001p-1022, +0x10000000000001LL, -1074),
170     +DBL_MIN,
171     MAKE_HEX_DOUBLE(+0x0.fffffffffffffp-1022, +0x0fffffffffffffLL, -1074),
172     MAKE_HEX_DOUBLE(+0x0.0000000000fffp-1022, +0x00000000000fffLL, -1074),
173     MAKE_HEX_DOUBLE(+0x0.00000000000fep-1022, +0x000000000000feLL, -1074),
174     MAKE_HEX_DOUBLE(+0x0.000000000000ep-1022, +0x0000000000000eLL, -1074),
175     MAKE_HEX_DOUBLE(+0x0.000000000000cp-1022, +0x0000000000000cLL, -1074),
176     MAKE_HEX_DOUBLE(+0x0.000000000000ap-1022, +0x0000000000000aLL, -1074),
177     MAKE_HEX_DOUBLE(+0x0.0000000000008p-1022, +0x00000000000008LL, -1074),
178     MAKE_HEX_DOUBLE(+0x0.0000000000007p-1022, +0x00000000000007LL, -1074),
179     MAKE_HEX_DOUBLE(+0x0.0000000000006p-1022, +0x00000000000006LL, -1074),
180     MAKE_HEX_DOUBLE(+0x0.0000000000005p-1022, +0x00000000000005LL, -1074),
181     MAKE_HEX_DOUBLE(+0x0.0000000000004p-1022, +0x00000000000004LL, -1074),
182     MAKE_HEX_DOUBLE(+0x0.0000000000003p-1022, +0x00000000000003LL, -1074),
183     MAKE_HEX_DOUBLE(+0x0.0000000000002p-1022, +0x00000000000002LL, -1074),
184     MAKE_HEX_DOUBLE(+0x0.0000000000001p-1022, +0x00000000000001LL, -1074),
185     +0.0,
186 };
187 
188 constexpr size_t specialValuesCount =
189     sizeof(specialValues) / sizeof(specialValues[0]);
190 
Test(cl_uint job_id,cl_uint thread_id,void * data)191 cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
192 {
193     TestInfo *job = (TestInfo *)data;
194     size_t buffer_elements = job->subBufferSize;
195     size_t buffer_size = buffer_elements * sizeof(cl_double);
196     cl_uint base = job_id * (cl_uint)job->step;
197     ThreadInfo *tinfo = &(job->tinfo[thread_id]);
198     dptr dfunc = job->f->dfunc;
199     int ftz = job->ftz;
200     bool relaxedMode = job->relaxedMode;
201     MTdata d = tinfo->d;
202     cl_int error;
203     const char *name = job->f->name;
204     cl_long *t;
205     cl_long *r;
206     cl_double *s;
207     cl_double *s2;
208 
209     Force64BitFPUPrecision();
210 
211     cl_event e[VECTOR_SIZE_COUNT];
212     cl_long *out[VECTOR_SIZE_COUNT];
213     if (gHostFill)
214     {
215         // start the map of the output arrays
216         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
217         {
218             out[j] = (cl_long *)clEnqueueMapBuffer(
219                 tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_WRITE, 0,
220                 buffer_size, 0, NULL, e + j, &error);
221             if (error || NULL == out[j])
222             {
223                 vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j,
224                            error);
225                 return error;
226             }
227         }
228 
229         // Get that moving
230         if ((error = clFlush(tinfo->tQueue))) vlog("clFlush failed\n");
231     }
232 
233     // Init input array
234     double *p = (double *)gIn + thread_id * buffer_elements;
235     double *p2 = (double *)gIn2 + thread_id * buffer_elements;
236     cl_uint idx = 0;
237     int totalSpecialValueCount = specialValuesCount * specialValuesCount;
238     int lastSpecialJobIndex = (totalSpecialValueCount - 1) / buffer_elements;
239 
240     // Test edge cases
241     if (job_id <= (cl_uint)lastSpecialJobIndex)
242     {
243         uint32_t x, y;
244 
245         x = (job_id * buffer_elements) % specialValuesCount;
246         y = (job_id * buffer_elements) / specialValuesCount;
247 
248         for (; idx < buffer_elements; idx++)
249         {
250             p[idx] = specialValues[x];
251             p2[idx] = specialValues[y];
252             if (++x >= specialValuesCount)
253             {
254                 x = 0;
255                 y++;
256                 if (y >= specialValuesCount) break;
257             }
258         }
259     }
260 
261     // Init any remaining values
262     for (; idx < buffer_elements; idx++)
263     {
264         ((cl_ulong *)p)[idx] = genrand_int64(d);
265         ((cl_ulong *)p2)[idx] = genrand_int64(d);
266     }
267 
268     if ((error = clEnqueueWriteBuffer(tinfo->tQueue, tinfo->inBuf, CL_FALSE, 0,
269                                       buffer_size, p, 0, NULL, NULL)))
270     {
271         vlog_error("Error: clEnqueueWriteBuffer failed! err: %d\n", error);
272         return error;
273     }
274 
275     if ((error = clEnqueueWriteBuffer(tinfo->tQueue, tinfo->inBuf2, CL_FALSE, 0,
276                                       buffer_size, p2, 0, NULL, NULL)))
277     {
278         vlog_error("Error: clEnqueueWriteBuffer failed! err: %d\n", error);
279         return error;
280     }
281 
282     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
283     {
284         if (gHostFill)
285         {
286             // Wait for the map to finish
287             if ((error = clWaitForEvents(1, e + j)))
288             {
289                 vlog_error("Error: clWaitForEvents failed! err: %d\n", error);
290                 return error;
291             }
292             if ((error = clReleaseEvent(e[j])))
293             {
294                 vlog_error("Error: clReleaseEvent failed! err: %d\n", error);
295                 return error;
296             }
297         }
298 
299         // Fill the result buffer with garbage, so that old results don't carry
300         // over
301         uint32_t pattern = 0xffffdead;
302         if (gHostFill)
303         {
304             memset_pattern4(out[j], &pattern, buffer_size);
305             if ((error = clEnqueueUnmapMemObject(
306                      tinfo->tQueue, tinfo->outBuf[j], out[j], 0, NULL, NULL)))
307             {
308                 vlog_error("Error: clEnqueueUnmapMemObject failed! err: %d\n",
309                            error);
310                 return error;
311             }
312         }
313         else
314         {
315             if ((error = clEnqueueFillBuffer(tinfo->tQueue, tinfo->outBuf[j],
316                                              &pattern, sizeof(pattern), 0,
317                                              buffer_size, 0, NULL, NULL)))
318             {
319                 vlog_error("Error: clEnqueueFillBuffer failed! err: %d\n",
320                            error);
321                 return error;
322             }
323         }
324 
325         // Run the kernel
326         size_t vectorCount =
327             (buffer_elements + sizeValues[j] - 1) / sizeValues[j];
328         cl_kernel kernel = job->k[j][thread_id]; // each worker thread has its
329                                                  // own copy of the cl_kernel
330         cl_program program = job->programs[j];
331 
332         if ((error = clSetKernelArg(kernel, 0, sizeof(tinfo->outBuf[j]),
333                                     &tinfo->outBuf[j])))
334         {
335             LogBuildError(program);
336             return error;
337         }
338         if ((error = clSetKernelArg(kernel, 1, sizeof(tinfo->inBuf),
339                                     &tinfo->inBuf)))
340         {
341             LogBuildError(program);
342             return error;
343         }
344         if ((error = clSetKernelArg(kernel, 2, sizeof(tinfo->inBuf2),
345                                     &tinfo->inBuf2)))
346         {
347             LogBuildError(program);
348             return error;
349         }
350 
351         if ((error = clEnqueueNDRangeKernel(tinfo->tQueue, kernel, 1, NULL,
352                                             &vectorCount, NULL, 0, NULL, NULL)))
353         {
354             vlog_error("FAILED -- could not execute kernel\n");
355             return error;
356         }
357     }
358 
359     // Get that moving
360     if ((error = clFlush(tinfo->tQueue))) vlog("clFlush 2 failed\n");
361 
362     if (gSkipCorrectnessTesting) return CL_SUCCESS;
363 
364     // Calculate the correctly rounded reference result
365     r = (cl_long *)gOut_Ref + thread_id * buffer_elements;
366     s = (cl_double *)gIn + thread_id * buffer_elements;
367     s2 = (cl_double *)gIn2 + thread_id * buffer_elements;
368     for (size_t j = 0; j < buffer_elements; j++) r[j] = dfunc.i_ff(s[j], s2[j]);
369 
370     // Read the data back -- no need to wait for the first N-1 buffers but wait
371     // for the last buffer. This is an in order queue.
372     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
373     {
374         cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
375         out[j] = (cl_long *)clEnqueueMapBuffer(
376             tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
377             buffer_size, 0, NULL, NULL, &error);
378         if (error || NULL == out[j])
379         {
380             vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j,
381                        error);
382             return error;
383         }
384     }
385 
386     // Verify data
387     t = (cl_long *)r;
388     for (size_t j = 0; j < buffer_elements; j++)
389     {
390         cl_long *q = out[0];
391 
392         // If we aren't getting the correctly rounded result
393         if (gMinVectorSizeIndex == 0 && t[j] != q[j])
394         {
395             // If we aren't getting the correctly rounded result
396             if (ftz || relaxedMode)
397             {
398                 if (IsDoubleSubnormal(s[j]))
399                 {
400                     if (IsDoubleSubnormal(s2[j]))
401                     {
402                         int64_t correct = dfunc.i_ff(0.0f, 0.0f);
403                         int64_t correct2 = dfunc.i_ff(0.0f, -0.0f);
404                         int64_t correct3 = dfunc.i_ff(-0.0f, 0.0f);
405                         int64_t correct4 = dfunc.i_ff(-0.0f, -0.0f);
406 
407                         if (correct == q[j] || correct2 == q[j]
408                             || correct3 == q[j] || correct4 == q[j])
409                             continue;
410                     }
411                     else
412                     {
413                         int64_t correct = dfunc.i_ff(0.0f, s2[j]);
414                         int64_t correct2 = dfunc.i_ff(-0.0f, s2[j]);
415                         if (correct == q[j] || correct2 == q[j]) continue;
416                     }
417                 }
418                 else if (IsDoubleSubnormal(s2[j]))
419                 {
420                     int64_t correct = dfunc.i_ff(s[j], 0.0f);
421                     int64_t correct2 = dfunc.i_ff(s[j], -0.0f);
422                     if (correct == q[j] || correct2 == q[j]) continue;
423                 }
424             }
425 
426             cl_ulong err = t[j] - q[j];
427             if (q[j] > t[j]) err = q[j] - t[j];
428             vlog_error("\nERROR: %s: %" PRId64
429                        " ulp error at {%.13la, %.13la}: *%" PRId64 " "
430                        "vs. %" PRId64 "  (index: %zu)\n",
431                        name, err, ((double *)s)[j], ((double *)s2)[j], t[j],
432                        q[j], j);
433             return -1;
434         }
435 
436 
437         for (auto k = std::max(1U, gMinVectorSizeIndex);
438              k < gMaxVectorSizeIndex; k++)
439         {
440             q = (cl_long *)out[k];
441             // If we aren't getting the correctly rounded result
442             if (-t[j] != q[j])
443             {
444                 if (ftz || relaxedMode)
445                 {
446                     if (IsDoubleSubnormal(s[j]))
447                     {
448                         if (IsDoubleSubnormal(s2[j]))
449                         {
450                             int64_t correct = -dfunc.i_ff(0.0f, 0.0f);
451                             int64_t correct2 = -dfunc.i_ff(0.0f, -0.0f);
452                             int64_t correct3 = -dfunc.i_ff(-0.0f, 0.0f);
453                             int64_t correct4 = -dfunc.i_ff(-0.0f, -0.0f);
454 
455                             if (correct == q[j] || correct2 == q[j]
456                                 || correct3 == q[j] || correct4 == q[j])
457                                 continue;
458                         }
459                         else
460                         {
461                             int64_t correct = -dfunc.i_ff(0.0f, s2[j]);
462                             int64_t correct2 = -dfunc.i_ff(-0.0f, s2[j]);
463                             if (correct == q[j] || correct2 == q[j]) continue;
464                         }
465                     }
466                     else if (IsDoubleSubnormal(s2[j]))
467                     {
468                         int64_t correct = -dfunc.i_ff(s[j], 0.0f);
469                         int64_t correct2 = -dfunc.i_ff(s[j], -0.0f);
470                         if (correct == q[j] || correct2 == q[j]) continue;
471                     }
472                 }
473 
474                 cl_ulong err = -t[j] - q[j];
475                 if (q[j] > -t[j]) err = q[j] + t[j];
476                 vlog_error("\nERROR: %sD%s: %" PRId64 " ulp error at {%.13la, "
477                            "%.13la}: *%" PRId64 " vs. %" PRId64
478                            "  (index: %zu)\n",
479                            name, sizeNames[k], err, ((double *)s)[j],
480                            ((double *)s2)[j], -t[j], q[j], j);
481                 return -1;
482             }
483         }
484     }
485 
486     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
487     {
488         if ((error = clEnqueueUnmapMemObject(tinfo->tQueue, tinfo->outBuf[j],
489                                              out[j], 0, NULL, NULL)))
490         {
491             vlog_error("Error: clEnqueueUnmapMemObject %d failed 2! err: %d\n",
492                        j, error);
493             return error;
494         }
495     }
496 
497     if ((error = clFlush(tinfo->tQueue))) vlog("clFlush 3 failed\n");
498 
499 
500     if (0 == (base & 0x0fffffff))
501     {
502         if (gVerboseBruteForce)
503         {
504             vlog("base:%14u step:%10u scale:%10u buf_elements:%10zd "
505                  "ThreadCount:%2u\n",
506                  base, job->step, job->scale, buffer_elements,
507                  job->threadCount);
508         }
509         else
510         {
511             vlog(".");
512         }
513         fflush(stdout);
514     }
515 
516     return CL_SUCCESS;
517 }
518 
519 } // anonymous namespace
520 
TestMacro_Int_Double_Double(const Func * f,MTdata d,bool relaxedMode)521 int TestMacro_Int_Double_Double(const Func *f, MTdata d, bool relaxedMode)
522 {
523     TestInfo test_info{};
524     cl_int error;
525 
526     logFunctionInfo(f->name, sizeof(cl_double), relaxedMode);
527 
528     // Init test_info
529     test_info.threadCount = GetThreadCount();
530     test_info.subBufferSize = BUFFER_SIZE
531         / (sizeof(cl_double) * RoundUpToNextPowerOfTwo(test_info.threadCount));
532     test_info.scale = getTestScale(sizeof(cl_double));
533 
534     test_info.step = (cl_uint)test_info.subBufferSize * test_info.scale;
535     if (test_info.step / test_info.subBufferSize != test_info.scale)
536     {
537         // there was overflow
538         test_info.jobCount = 1;
539     }
540     else
541     {
542         test_info.jobCount = (cl_uint)((1ULL << 32) / test_info.step);
543     }
544 
545     test_info.f = f;
546     test_info.ftz = f->ftz || gForceFTZ;
547     test_info.relaxedMode = relaxedMode;
548 
549     test_info.tinfo.resize(test_info.threadCount);
550     for (cl_uint i = 0; i < test_info.threadCount; i++)
551     {
552         cl_buffer_region region = {
553             i * test_info.subBufferSize * sizeof(cl_double),
554             test_info.subBufferSize * sizeof(cl_double)
555         };
556         test_info.tinfo[i].inBuf =
557             clCreateSubBuffer(gInBuffer, CL_MEM_READ_ONLY,
558                               CL_BUFFER_CREATE_TYPE_REGION, &region, &error);
559         if (error || NULL == test_info.tinfo[i].inBuf)
560         {
561             vlog_error("Error: Unable to create sub-buffer of gInBuffer for "
562                        "region {%zd, %zd}\n",
563                        region.origin, region.size);
564             return error;
565         }
566         test_info.tinfo[i].inBuf2 =
567             clCreateSubBuffer(gInBuffer2, CL_MEM_READ_ONLY,
568                               CL_BUFFER_CREATE_TYPE_REGION, &region, &error);
569         if (error || NULL == test_info.tinfo[i].inBuf2)
570         {
571             vlog_error("Error: Unable to create sub-buffer of gInBuffer2 for "
572                        "region {%zd, %zd}\n",
573                        region.origin, region.size);
574             return error;
575         }
576 
577         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
578         {
579             test_info.tinfo[i].outBuf[j] = clCreateSubBuffer(
580                 gOutBuffer[j], CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION,
581                 &region, &error);
582             if (error || NULL == test_info.tinfo[i].outBuf[j])
583             {
584                 vlog_error("Error: Unable to create sub-buffer of "
585                            "gOutBuffer[%d] for region {%zd, %zd}\n",
586                            (int)j, region.origin, region.size);
587                 return error;
588             }
589         }
590         test_info.tinfo[i].tQueue =
591             clCreateCommandQueue(gContext, gDevice, 0, &error);
592         if (NULL == test_info.tinfo[i].tQueue || error)
593         {
594             vlog_error("clCreateCommandQueue failed. (%d)\n", error);
595             return error;
596         }
597 
598         test_info.tinfo[i].d = MTdataHolder(genrand_int32(d));
599     }
600 
601     // Init the kernels
602     BuildKernelInfo build_info{ test_info.threadCount, test_info.k,
603                                 test_info.programs, f->nameInCode,
604                                 relaxedMode };
605     if ((error = ThreadPool_Do(BuildKernelFn,
606                                gMaxVectorSizeIndex - gMinVectorSizeIndex,
607                                &build_info)))
608         return error;
609 
610     // Run the kernels
611     if (!gSkipCorrectnessTesting)
612     {
613         error = ThreadPool_Do(Test, test_info.jobCount, &test_info);
614         if (error) return error;
615 
616         if (gWimpyMode)
617             vlog("Wimp pass");
618         else
619             vlog("passed");
620     }
621 
622     vlog("\n");
623 
624     return CL_SUCCESS;
625 }
626