• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <climits>
23 #include <cstring>
24 
25 namespace {
26 
BuildKernel(const char * name,int vectorSize,cl_uint kernel_count,cl_kernel * k,cl_program * p,bool relaxedMode)27 int BuildKernel(const char *name, int vectorSize, cl_uint kernel_count,
28                 cl_kernel *k, cl_program *p, bool relaxedMode)
29 {
30     const char *c[] = { "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n",
31                         "__kernel void math_kernel",
32                         sizeNames[vectorSize],
33                         "( __global double",
34                         sizeNames[vectorSize],
35                         "* out, __global double",
36                         sizeNames[vectorSize],
37                         "* in1, __global int",
38                         sizeNames[vectorSize],
39                         "* in2 )\n"
40                         "{\n"
41                         "   size_t i = get_global_id(0);\n"
42                         "   out[i] = ",
43                         name,
44                         "( in1[i], in2[i] );\n"
45                         "}\n" };
46 
47     const char *c3[] = {
48         "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n",
49         "__kernel void math_kernel",
50         sizeNames[vectorSize],
51         "( __global double* out, __global double* in, __global int* in2)\n"
52         "{\n"
53         "   size_t i = get_global_id(0);\n"
54         "   if( i + 1 < get_global_size(0) )\n"
55         "   {\n"
56         "       double3 d0 = vload3( 0, in + 3 * i );\n"
57         "       int3 i0 = vload3( 0, in2 + 3 * i );\n"
58         "       d0 = ",
59         name,
60         "( d0, i0 );\n"
61         "       vstore3( d0, 0, out + 3*i );\n"
62         "   }\n"
63         "   else\n"
64         "   {\n"
65         "       size_t parity = i & 1;   // Figure out how many elements are "
66         "left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two "
67         "buffer size \n"
68         "       double3 d0;\n"
69         "       int3 i0;\n"
70         "       switch( parity )\n"
71         "       {\n"
72         "           case 1:\n"
73         "               d0 = (double3)( in[3*i], NAN, NAN ); \n"
74         "               i0 = (int3)( in2[3*i], 0xdead, 0xdead ); \n"
75         "               break;\n"
76         "           case 0:\n"
77         "               d0 = (double3)( in[3*i], in[3*i+1], NAN ); \n"
78         "               i0 = (int3)( in2[3*i], in2[3*i+1], 0xdead ); \n"
79         "               break;\n"
80         "       }\n"
81         "       d0 = ",
82         name,
83         "( d0, i0 );\n"
84         "       switch( parity )\n"
85         "       {\n"
86         "           case 0:\n"
87         "               out[3*i+1] = d0.y; \n"
88         "               // fall through\n"
89         "           case 1:\n"
90         "               out[3*i] = d0.x; \n"
91         "               break;\n"
92         "       }\n"
93         "   }\n"
94         "}\n"
95     };
96 
97     const char **kern = c;
98     size_t kernSize = sizeof(c) / sizeof(c[0]);
99 
100     if (sizeValues[vectorSize] == 3)
101     {
102         kern = c3;
103         kernSize = sizeof(c3) / sizeof(c3[0]);
104     }
105 
106     char testName[32];
107     snprintf(testName, sizeof(testName) - 1, "math_kernel%s",
108              sizeNames[vectorSize]);
109 
110     return MakeKernels(kern, (cl_uint)kernSize, testName, kernel_count, k, p,
111                        relaxedMode);
112 }
113 
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)114 cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
115 {
116     BuildKernelInfo *info = (BuildKernelInfo *)p;
117     cl_uint vectorSize = gMinVectorSizeIndex + job_id;
118     return BuildKernel(info->nameInCode, vectorSize, info->threadCount,
119                        info->kernels[vectorSize].data(),
120                        &(info->programs[vectorSize]), info->relaxedMode);
121 }
122 
123 // Thread specific data for a worker thread
124 struct ThreadInfo
125 {
126     // Input and output buffers for the thread
127     clMemWrapper inBuf;
128     clMemWrapper inBuf2;
129     Buffers outBuf;
130 
131     float maxError; // max error value. Init to 0.
132     double
133         maxErrorValue; // position of the max error value (param 1).  Init to 0.
134     cl_int maxErrorValue2; // position of the max error value (param 2).  Init
135                            // to 0.
136     MTdataHolder d;
137 
138     // Per thread command queue to improve performance
139     clCommandQueueWrapper tQueue;
140 };
141 
142 struct TestInfo
143 {
144     size_t subBufferSize; // Size of the sub-buffer in elements
145     const Func *f; // A pointer to the function info
146 
147     // Programs for various vector sizes.
148     Programs programs;
149 
150     // Thread-specific kernels for each vector size:
151     // k[vector_size][thread_id]
152     KernelMatrix k;
153 
154     // Array of thread specific information
155     std::vector<ThreadInfo> tinfo;
156 
157     cl_uint threadCount; // Number of worker threads
158     cl_uint jobCount; // Number of jobs
159     cl_uint step; // step between each chunk and the next.
160     cl_uint scale; // stride between individual test values
161     float ulps; // max_allowed ulps
162     int ftz; // non-zero if running in flush to zero mode
163     bool relaxedMode; // True if test is running in relaxed mode, false
164                       // otherwise.
165 
166     // no special values
167 };
168 
169 // A table of more difficult cases to get right
170 const double specialValues[] = {
171     -NAN,
172     -INFINITY,
173     -DBL_MAX,
174     MAKE_HEX_DOUBLE(-0x1.0000000000001p64, -0x10000000000001LL, 12),
175     MAKE_HEX_DOUBLE(-0x1.0p64, -0x1LL, 64),
176     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp63, -0x1fffffffffffffLL, 11),
177     MAKE_HEX_DOUBLE(-0x1.0000000000001p63, -0x10000000000001LL, 11),
178     MAKE_HEX_DOUBLE(-0x1.0p63, -0x1LL, 63),
179     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp62, -0x1fffffffffffffLL, 10),
180     MAKE_HEX_DOUBLE(-0x1.000002p32, -0x1000002LL, 8),
181     MAKE_HEX_DOUBLE(-0x1.0p32, -0x1LL, 32),
182     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp31, -0x1fffffffffffffLL, -21),
183     MAKE_HEX_DOUBLE(-0x1.0000000000001p31, -0x10000000000001LL, -21),
184     MAKE_HEX_DOUBLE(-0x1.0p31, -0x1LL, 31),
185     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp30, -0x1fffffffffffffLL, -22),
186     -1000.0,
187     -100.0,
188     -4.0,
189     -3.5,
190     -3.0,
191     MAKE_HEX_DOUBLE(-0x1.8000000000001p1, -0x18000000000001LL, -51),
192     -2.5,
193     MAKE_HEX_DOUBLE(-0x1.7ffffffffffffp1, -0x17ffffffffffffLL, -51),
194     -2.0,
195     MAKE_HEX_DOUBLE(-0x1.8000000000001p0, -0x18000000000001LL, -52),
196     -1.5,
197     MAKE_HEX_DOUBLE(-0x1.7ffffffffffffp0, -0x17ffffffffffffLL, -52),
198     MAKE_HEX_DOUBLE(-0x1.0000000000001p0, -0x10000000000001LL, -52),
199     -1.0,
200     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp-1, -0x1fffffffffffffLL, -53),
201     MAKE_HEX_DOUBLE(-0x1.0000000000001p-1, -0x10000000000001LL, -53),
202     -0.5,
203     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp-2, -0x1fffffffffffffLL, -54),
204     MAKE_HEX_DOUBLE(-0x1.0000000000001p-2, -0x10000000000001LL, -54),
205     -0.25,
206     MAKE_HEX_DOUBLE(-0x1.fffffffffffffp-3, -0x1fffffffffffffLL, -55),
207     MAKE_HEX_DOUBLE(-0x1.0000000000001p-1022, -0x10000000000001LL, -1074),
208     -DBL_MIN,
209     MAKE_HEX_DOUBLE(-0x0.fffffffffffffp-1022, -0x0fffffffffffffLL, -1074),
210     MAKE_HEX_DOUBLE(-0x0.0000000000fffp-1022, -0x00000000000fffLL, -1074),
211     MAKE_HEX_DOUBLE(-0x0.00000000000fep-1022, -0x000000000000feLL, -1074),
212     MAKE_HEX_DOUBLE(-0x0.000000000000ep-1022, -0x0000000000000eLL, -1074),
213     MAKE_HEX_DOUBLE(-0x0.000000000000cp-1022, -0x0000000000000cLL, -1074),
214     MAKE_HEX_DOUBLE(-0x0.000000000000ap-1022, -0x0000000000000aLL, -1074),
215     MAKE_HEX_DOUBLE(-0x0.0000000000008p-1022, -0x00000000000008LL, -1074),
216     MAKE_HEX_DOUBLE(-0x0.0000000000007p-1022, -0x00000000000007LL, -1074),
217     MAKE_HEX_DOUBLE(-0x0.0000000000006p-1022, -0x00000000000006LL, -1074),
218     MAKE_HEX_DOUBLE(-0x0.0000000000005p-1022, -0x00000000000005LL, -1074),
219     MAKE_HEX_DOUBLE(-0x0.0000000000004p-1022, -0x00000000000004LL, -1074),
220     MAKE_HEX_DOUBLE(-0x0.0000000000003p-1022, -0x00000000000003LL, -1074),
221     MAKE_HEX_DOUBLE(-0x0.0000000000002p-1022, -0x00000000000002LL, -1074),
222     MAKE_HEX_DOUBLE(-0x0.0000000000001p-1022, -0x00000000000001LL, -1074),
223     -0.0,
224 
225     +NAN,
226     +INFINITY,
227     +DBL_MAX,
228     MAKE_HEX_DOUBLE(+0x1.0000000000001p64, +0x10000000000001LL, 12),
229     MAKE_HEX_DOUBLE(+0x1.0p64, +0x1LL, 64),
230     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp63, +0x1fffffffffffffLL, 11),
231     MAKE_HEX_DOUBLE(+0x1.0000000000001p63, +0x10000000000001LL, 11),
232     MAKE_HEX_DOUBLE(+0x1.0p63, +0x1LL, 63),
233     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp62, +0x1fffffffffffffLL, 10),
234     MAKE_HEX_DOUBLE(+0x1.000002p32, +0x1000002LL, 8),
235     MAKE_HEX_DOUBLE(+0x1.0p32, +0x1LL, 32),
236     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp31, +0x1fffffffffffffLL, -21),
237     MAKE_HEX_DOUBLE(+0x1.0000000000001p31, +0x10000000000001LL, -21),
238     MAKE_HEX_DOUBLE(+0x1.0p31, +0x1LL, 31),
239     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp30, +0x1fffffffffffffLL, -22),
240     +1000.0,
241     +100.0,
242     +4.0,
243     +3.5,
244     +3.0,
245     MAKE_HEX_DOUBLE(+0x1.8000000000001p1, +0x18000000000001LL, -51),
246     +2.5,
247     MAKE_HEX_DOUBLE(+0x1.7ffffffffffffp1, +0x17ffffffffffffLL, -51),
248     +2.0,
249     MAKE_HEX_DOUBLE(+0x1.8000000000001p0, +0x18000000000001LL, -52),
250     +1.5,
251     MAKE_HEX_DOUBLE(+0x1.7ffffffffffffp0, +0x17ffffffffffffLL, -52),
252     MAKE_HEX_DOUBLE(-0x1.0000000000001p0, -0x10000000000001LL, -52),
253     +1.0,
254     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp-1, +0x1fffffffffffffLL, -53),
255     MAKE_HEX_DOUBLE(+0x1.0000000000001p-1, +0x10000000000001LL, -53),
256     +0.5,
257     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp-2, +0x1fffffffffffffLL, -54),
258     MAKE_HEX_DOUBLE(+0x1.0000000000001p-2, +0x10000000000001LL, -54),
259     +0.25,
260     MAKE_HEX_DOUBLE(+0x1.fffffffffffffp-3, +0x1fffffffffffffLL, -55),
261     MAKE_HEX_DOUBLE(+0x1.0000000000001p-1022, +0x10000000000001LL, -1074),
262     +DBL_MIN,
263     MAKE_HEX_DOUBLE(+0x0.fffffffffffffp-1022, +0x0fffffffffffffLL, -1074),
264     MAKE_HEX_DOUBLE(+0x0.0000000000fffp-1022, +0x00000000000fffLL, -1074),
265     MAKE_HEX_DOUBLE(+0x0.00000000000fep-1022, +0x000000000000feLL, -1074),
266     MAKE_HEX_DOUBLE(+0x0.000000000000ep-1022, +0x0000000000000eLL, -1074),
267     MAKE_HEX_DOUBLE(+0x0.000000000000cp-1022, +0x0000000000000cLL, -1074),
268     MAKE_HEX_DOUBLE(+0x0.000000000000ap-1022, +0x0000000000000aLL, -1074),
269     MAKE_HEX_DOUBLE(+0x0.0000000000008p-1022, +0x00000000000008LL, -1074),
270     MAKE_HEX_DOUBLE(+0x0.0000000000007p-1022, +0x00000000000007LL, -1074),
271     MAKE_HEX_DOUBLE(+0x0.0000000000006p-1022, +0x00000000000006LL, -1074),
272     MAKE_HEX_DOUBLE(+0x0.0000000000005p-1022, +0x00000000000005LL, -1074),
273     MAKE_HEX_DOUBLE(+0x0.0000000000004p-1022, +0x00000000000004LL, -1074),
274     MAKE_HEX_DOUBLE(+0x0.0000000000003p-1022, +0x00000000000003LL, -1074),
275     MAKE_HEX_DOUBLE(+0x0.0000000000002p-1022, +0x00000000000002LL, -1074),
276     MAKE_HEX_DOUBLE(+0x0.0000000000001p-1022, +0x00000000000001LL, -1074),
277     +0.0,
278 };
279 
280 constexpr size_t specialValuesCount =
281     sizeof(specialValues) / sizeof(specialValues[0]);
282 
283 const int specialValuesInt[] = {
284     0,       1,  2,  3,  1022,  1023,  1024,   INT_MIN,
285     INT_MAX, -1, -2, -3, -1022, -1023, -11024, -INT_MAX,
286 };
287 
288 constexpr size_t specialValuesIntCount =
289     sizeof(specialValuesInt) / sizeof(specialValuesInt[0]);
290 
Test(cl_uint job_id,cl_uint thread_id,void * data)291 cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
292 {
293     TestInfo *job = (TestInfo *)data;
294     size_t buffer_elements = job->subBufferSize;
295     size_t buffer_size = buffer_elements * sizeof(cl_double);
296     cl_uint base = job_id * (cl_uint)job->step;
297     ThreadInfo *tinfo = &(job->tinfo[thread_id]);
298     float ulps = job->ulps;
299     dptr func = job->f->dfunc;
300     int ftz = job->ftz;
301     bool relaxedMode = job->relaxedMode;
302     MTdata d = tinfo->d;
303     cl_int error;
304     const char *name = job->f->name;
305     cl_ulong *t;
306     cl_double *r;
307     cl_double *s;
308     cl_int *s2;
309 
310     Force64BitFPUPrecision();
311 
312     // start the map of the output arrays
313     cl_event e[VECTOR_SIZE_COUNT];
314     cl_ulong *out[VECTOR_SIZE_COUNT];
315     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
316     {
317         out[j] = (cl_ulong *)clEnqueueMapBuffer(
318             tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_WRITE, 0,
319             buffer_size, 0, NULL, e + j, &error);
320         if (error || NULL == out[j])
321         {
322             vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j,
323                        error);
324             return error;
325         }
326     }
327 
328     // Get that moving
329     if ((error = clFlush(tinfo->tQueue))) vlog("clFlush failed\n");
330 
331     // Init input array
332     cl_ulong *p = (cl_ulong *)gIn + thread_id * buffer_elements;
333     cl_int *p2 = (cl_int *)gIn2 + thread_id * buffer_elements;
334     size_t idx = 0;
335     int totalSpecialValueCount = specialValuesCount * specialValuesIntCount;
336     int lastSpecialJobIndex = (totalSpecialValueCount - 1) / buffer_elements;
337 
338     if (job_id <= (cl_uint)lastSpecialJobIndex)
339     { // test edge cases
340         cl_double *fp = (cl_double *)p;
341         cl_int *ip2 = (cl_int *)p2;
342         uint32_t x, y;
343 
344         x = (job_id * buffer_elements) % specialValuesCount;
345         y = (job_id * buffer_elements) / specialValuesCount;
346 
347         for (; idx < buffer_elements; idx++)
348         {
349             fp[idx] = specialValues[x];
350             ip2[idx] = specialValuesInt[y];
351             if (++x >= specialValuesCount)
352             {
353                 x = 0;
354                 y++;
355                 if (y >= specialValuesIntCount) break;
356             }
357         }
358     }
359 
360     // Init any remaining values.
361     for (; idx < buffer_elements; idx++)
362     {
363         p[idx] = DoubleFromUInt32(genrand_int32(d));
364         p2[idx] = genrand_int32(d);
365     }
366 
367     if ((error = clEnqueueWriteBuffer(tinfo->tQueue, tinfo->inBuf, CL_FALSE, 0,
368                                       buffer_size, p, 0, NULL, NULL)))
369     {
370         vlog_error("Error: clEnqueueWriteBuffer failed! err: %d\n", error);
371         goto exit;
372     }
373 
374     if ((error = clEnqueueWriteBuffer(tinfo->tQueue, tinfo->inBuf2, CL_FALSE, 0,
375                                       buffer_size / 2, p2, 0, NULL, NULL)))
376     {
377         vlog_error("Error: clEnqueueWriteBuffer failed! err: %d\n", error);
378         goto exit;
379     }
380 
381     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
382     {
383         // Wait for the map to finish
384         if ((error = clWaitForEvents(1, e + j)))
385         {
386             vlog_error("Error: clWaitForEvents failed! err: %d\n", error);
387             goto exit;
388         }
389         if ((error = clReleaseEvent(e[j])))
390         {
391             vlog_error("Error: clReleaseEvent failed! err: %d\n", error);
392             goto exit;
393         }
394 
395         // Fill the result buffer with garbage, so that old results don't carry
396         // over
397         uint32_t pattern = 0xffffdead;
398         memset_pattern4(out[j], &pattern, buffer_size);
399         if ((error = clEnqueueUnmapMemObject(tinfo->tQueue, tinfo->outBuf[j],
400                                              out[j], 0, NULL, NULL)))
401         {
402             vlog_error("Error: clEnqueueUnmapMemObject failed! err: %d\n",
403                        error);
404             goto exit;
405         }
406 
407         // run the kernel
408         size_t vectorCount =
409             (buffer_elements + sizeValues[j] - 1) / sizeValues[j];
410         cl_kernel kernel = job->k[j][thread_id]; // each worker thread has its
411                                                  // own copy of the cl_kernel
412         cl_program program = job->programs[j];
413 
414         if ((error = clSetKernelArg(kernel, 0, sizeof(tinfo->outBuf[j]),
415                                     &tinfo->outBuf[j])))
416         {
417             LogBuildError(program);
418             return error;
419         }
420         if ((error = clSetKernelArg(kernel, 1, sizeof(tinfo->inBuf),
421                                     &tinfo->inBuf)))
422         {
423             LogBuildError(program);
424             return error;
425         }
426         if ((error = clSetKernelArg(kernel, 2, sizeof(tinfo->inBuf2),
427                                     &tinfo->inBuf2)))
428         {
429             LogBuildError(program);
430             return error;
431         }
432 
433         if ((error = clEnqueueNDRangeKernel(tinfo->tQueue, kernel, 1, NULL,
434                                             &vectorCount, NULL, 0, NULL, NULL)))
435         {
436             vlog_error("FAILED -- could not execute kernel\n");
437             goto exit;
438         }
439     }
440 
441     // Get that moving
442     if ((error = clFlush(tinfo->tQueue))) vlog("clFlush 2 failed\n");
443 
444     if (gSkipCorrectnessTesting) return CL_SUCCESS;
445 
446     // Calculate the correctly rounded reference result
447     r = (cl_double *)gOut_Ref + thread_id * buffer_elements;
448     s = (cl_double *)gIn + thread_id * buffer_elements;
449     s2 = (cl_int *)gIn2 + thread_id * buffer_elements;
450     for (size_t j = 0; j < buffer_elements; j++)
451         r[j] = (cl_double)func.f_fi(s[j], s2[j]);
452 
453     // Read the data back -- no need to wait for the first N-1 buffers but wait
454     // for the last buffer. This is an in order queue.
455     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
456     {
457         cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
458         out[j] = (cl_ulong *)clEnqueueMapBuffer(
459             tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
460             buffer_size, 0, NULL, NULL, &error);
461         if (error || NULL == out[j])
462         {
463             vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j,
464                        error);
465             goto exit;
466         }
467     }
468 
469     // Verify data
470     t = (cl_ulong *)r;
471     for (size_t j = 0; j < buffer_elements; j++)
472     {
473         for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
474         {
475             cl_ulong *q = out[k];
476 
477             // If we aren't getting the correctly rounded result
478             if (t[j] != q[j])
479             {
480                 cl_double test = ((cl_double *)q)[j];
481                 long double correct = func.f_fi(s[j], s2[j]);
482                 float err = Bruteforce_Ulp_Error_Double(test, correct);
483                 int fail = !(fabsf(err) <= ulps);
484 
485                 if (fail && (ftz || relaxedMode))
486                 {
487                     // retry per section 6.5.3.2
488                     if (IsDoubleResultSubnormal(correct, ulps))
489                     {
490                         fail = fail && (test != 0.0f);
491                         if (!fail) err = 0.0f;
492                     }
493 
494                     // retry per section 6.5.3.3
495                     if (IsDoubleSubnormal(s[j]))
496                     {
497                         long double correct2 = func.f_fi(0.0, s2[j]);
498                         long double correct3 = func.f_fi(-0.0, s2[j]);
499                         float err2 =
500                             Bruteforce_Ulp_Error_Double(test, correct2);
501                         float err3 =
502                             Bruteforce_Ulp_Error_Double(test, correct3);
503                         fail = fail
504                             && ((!(fabsf(err2) <= ulps))
505                                 && (!(fabsf(err3) <= ulps)));
506                         if (fabsf(err2) < fabsf(err)) err = err2;
507                         if (fabsf(err3) < fabsf(err)) err = err3;
508 
509                         // retry per section 6.5.3.4
510                         if (IsDoubleResultSubnormal(correct2, ulps)
511                             || IsDoubleResultSubnormal(correct3, ulps))
512                         {
513                             fail = fail && (test != 0.0f);
514                             if (!fail) err = 0.0f;
515                         }
516                     }
517                 }
518 
519                 if (fabsf(err) > tinfo->maxError)
520                 {
521                     tinfo->maxError = fabsf(err);
522                     tinfo->maxErrorValue = s[j];
523                     tinfo->maxErrorValue2 = s2[j];
524                 }
525                 if (fail)
526                 {
527                     vlog_error("\nERROR: %s%s: %f ulp error at {%.13la, %d}: "
528                                "*%.13la vs. %.13la\n",
529                                name, sizeNames[k], err, s[j], s2[j], r[j],
530                                test);
531                     error = -1;
532                     goto exit;
533                 }
534             }
535         }
536     }
537 
538     for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
539     {
540         if ((error = clEnqueueUnmapMemObject(tinfo->tQueue, tinfo->outBuf[j],
541                                              out[j], 0, NULL, NULL)))
542         {
543             vlog_error("Error: clEnqueueUnmapMemObject %d failed 2! err: %d\n",
544                        j, error);
545             return error;
546         }
547     }
548 
549     if ((error = clFlush(tinfo->tQueue))) vlog("clFlush 3 failed\n");
550 
551 
552     if (0 == (base & 0x0fffffff))
553     {
554         if (gVerboseBruteForce)
555         {
556             vlog("base:%14u step:%10u scale:%10u buf_elements:%10zd ulps:%5.3f "
557                  "ThreadCount:%2u\n",
558                  base, job->step, job->scale, buffer_elements, job->ulps,
559                  job->threadCount);
560         }
561         else
562         {
563             vlog(".");
564         }
565         fflush(stdout);
566     }
567 
568 exit:
569     return error;
570 }
571 
572 } // anonymous namespace
573 
TestFunc_Double_Double_Int(const Func * f,MTdata d,bool relaxedMode)574 int TestFunc_Double_Double_Int(const Func *f, MTdata d, bool relaxedMode)
575 {
576     TestInfo test_info{};
577     cl_int error;
578     float maxError = 0.0f;
579     double maxErrorVal = 0.0;
580     cl_int maxErrorVal2 = 0;
581 
582     logFunctionInfo(f->name, sizeof(cl_double), relaxedMode);
583 
584     // Init test_info
585     test_info.threadCount = GetThreadCount();
586     test_info.subBufferSize = BUFFER_SIZE
587         / (sizeof(cl_double) * RoundUpToNextPowerOfTwo(test_info.threadCount));
588     test_info.scale = getTestScale(sizeof(cl_double));
589 
590     test_info.step = (cl_uint)test_info.subBufferSize * test_info.scale;
591     if (test_info.step / test_info.subBufferSize != test_info.scale)
592     {
593         // there was overflow
594         test_info.jobCount = 1;
595     }
596     else
597     {
598         test_info.jobCount = (cl_uint)((1ULL << 32) / test_info.step);
599     }
600 
601     test_info.f = f;
602     test_info.ulps = f->double_ulps;
603     test_info.ftz = f->ftz || gForceFTZ;
604     test_info.relaxedMode = relaxedMode;
605 
606     // cl_kernels aren't thread safe, so we make one for each vector size for
607     // every thread
608     for (auto i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++)
609     {
610         test_info.k[i].resize(test_info.threadCount, nullptr);
611     }
612 
613     test_info.tinfo.resize(test_info.threadCount);
614     for (cl_uint i = 0; i < test_info.threadCount; i++)
615     {
616         cl_buffer_region region = {
617             i * test_info.subBufferSize * sizeof(cl_double),
618             test_info.subBufferSize * sizeof(cl_double)
619         };
620         test_info.tinfo[i].inBuf =
621             clCreateSubBuffer(gInBuffer, CL_MEM_READ_ONLY,
622                               CL_BUFFER_CREATE_TYPE_REGION, &region, &error);
623         if (error || NULL == test_info.tinfo[i].inBuf)
624         {
625             vlog_error("Error: Unable to create sub-buffer of gInBuffer for "
626                        "region {%zd, %zd}\n",
627                        region.origin, region.size);
628             goto exit;
629         }
630         cl_buffer_region region2 = { i * test_info.subBufferSize
631                                          * sizeof(cl_int),
632                                      test_info.subBufferSize * sizeof(cl_int) };
633         test_info.tinfo[i].inBuf2 =
634             clCreateSubBuffer(gInBuffer2, CL_MEM_READ_ONLY,
635                               CL_BUFFER_CREATE_TYPE_REGION, &region2, &error);
636         if (error || NULL == test_info.tinfo[i].inBuf2)
637         {
638             vlog_error("Error: Unable to create sub-buffer of gInBuffer2 for "
639                        "region {%zd, %zd}\n",
640                        region.origin, region.size);
641             goto exit;
642         }
643 
644         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
645         {
646             test_info.tinfo[i].outBuf[j] = clCreateSubBuffer(
647                 gOutBuffer[j], CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION,
648                 &region, &error);
649             if (error || NULL == test_info.tinfo[i].outBuf[j])
650             {
651                 vlog_error("Error: Unable to create sub-buffer of "
652                            "gOutBuffer[%d] for region {%zd, %zd}\n",
653                            (int)j, region.origin, region.size);
654                 goto exit;
655             }
656         }
657         test_info.tinfo[i].tQueue =
658             clCreateCommandQueue(gContext, gDevice, 0, &error);
659         if (NULL == test_info.tinfo[i].tQueue || error)
660         {
661             vlog_error("clCreateCommandQueue failed. (%d)\n", error);
662             goto exit;
663         }
664 
665         test_info.tinfo[i].d = MTdataHolder(genrand_int32(d));
666     }
667 
668     // Init the kernels
669     {
670         BuildKernelInfo build_info{ test_info.threadCount, test_info.k,
671                                     test_info.programs, f->nameInCode,
672                                     relaxedMode };
673         if ((error = ThreadPool_Do(BuildKernelFn,
674                                    gMaxVectorSizeIndex - gMinVectorSizeIndex,
675                                    &build_info)))
676             goto exit;
677     }
678 
679     // Run the kernels
680     if (!gSkipCorrectnessTesting)
681     {
682         error = ThreadPool_Do(Test, test_info.jobCount, &test_info);
683 
684         // Accumulate the arithmetic errors
685         for (cl_uint i = 0; i < test_info.threadCount; i++)
686         {
687             if (test_info.tinfo[i].maxError > maxError)
688             {
689                 maxError = test_info.tinfo[i].maxError;
690                 maxErrorVal = test_info.tinfo[i].maxErrorValue;
691                 maxErrorVal2 = test_info.tinfo[i].maxErrorValue2;
692             }
693         }
694 
695         if (error) goto exit;
696 
697         if (gWimpyMode)
698             vlog("Wimp pass");
699         else
700             vlog("passed");
701 
702         vlog("\t%8.2f @ {%a, %d}", maxError, maxErrorVal, maxErrorVal2);
703     }
704 
705     vlog("\n");
706 
707 exit:
708     // Release
709     for (auto i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++)
710     {
711         for (auto &kernel : test_info.k[i])
712         {
713             clReleaseKernel(kernel);
714         }
715     }
716 
717     return error;
718 }
719