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 <climits>
22 #include <cstring>
23 
BuildKernel(const char * name,int vectorSize,cl_kernel * k,cl_program * p,bool relaxedMode)24 static int BuildKernel(const char *name, int vectorSize, cl_kernel *k,
25                        cl_program *p, bool relaxedMode)
26 {
27     const char *c[] = { "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n",
28                         "__kernel void math_kernel",
29                         sizeNames[vectorSize],
30                         "( __global double",
31                         sizeNames[vectorSize],
32                         "* out, __global int",
33                         sizeNames[vectorSize],
34                         "* out2, __global double",
35                         sizeNames[vectorSize],
36                         "* in1, __global double",
37                         sizeNames[vectorSize],
38                         "* in2 )\n"
39                         "{\n"
40                         "   size_t i = get_global_id(0);\n"
41                         "   out[i] = ",
42                         name,
43                         "( in1[i], in2[i], out2 + i );\n"
44                         "}\n" };
45 
46     const char *c3[] = {
47         "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n",
48         "__kernel void math_kernel",
49         sizeNames[vectorSize],
50         "( __global double* out, __global int* out2, __global double* in, "
51         "__global double* 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         "       double3 d1 = vload3( 0, in2 + 3 * i );\n"
58         "       int3 i0 = 0xdeaddead;\n"
59         "       d0 = ",
60         name,
61         "( d0, d1, &i0 );\n"
62         "       vstore3( d0, 0, out + 3*i );\n"
63         "       vstore3( i0, 0, out2 + 3*i );\n"
64         "   }\n"
65         "   else\n"
66         "   {\n"
67         "       size_t parity = i & 1;   // Figure out how many elements are "
68         "left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two "
69         "buffer size \n"
70         "       double3 d0;\n"
71         "       double3 d1;\n"
72         "       int3 i0 = 0xdeaddead;\n"
73         "       switch( parity )\n"
74         "       {\n"
75         "           case 1:\n"
76         "               d0 = (double3)( in[3*i], NAN, NAN ); \n"
77         "               d1 = (double3)( in2[3*i], NAN, NAN ); \n"
78         "               break;\n"
79         "           case 0:\n"
80         "               d0 = (double3)( in[3*i], in[3*i+1], NAN ); \n"
81         "               d1 = (double3)( in2[3*i], in2[3*i+1], NAN ); \n"
82         "               break;\n"
83         "       }\n"
84         "       d0 = ",
85         name,
86         "( d0, d1, &i0 );\n"
87         "       switch( parity )\n"
88         "       {\n"
89         "           case 0:\n"
90         "               out[3*i+1] = d0.y; \n"
91         "               out2[3*i+1] = i0.y; \n"
92         "               // fall through\n"
93         "           case 1:\n"
94         "               out[3*i] = d0.x; \n"
95         "               out2[3*i] = i0.x; \n"
96         "               break;\n"
97         "       }\n"
98         "   }\n"
99         "}\n"
100     };
101 
102     const char **kern = c;
103     size_t kernSize = sizeof(c) / sizeof(c[0]);
104 
105     if (sizeValues[vectorSize] == 3)
106     {
107         kern = c3;
108         kernSize = sizeof(c3) / sizeof(c3[0]);
109     }
110 
111     char testName[32];
112     snprintf(testName, sizeof(testName) - 1, "math_kernel%s",
113              sizeNames[vectorSize]);
114 
115     return MakeKernel(kern, (cl_uint)kernSize, testName, k, p, relaxedMode);
116 }
117 
118 typedef struct BuildKernelInfo
119 {
120     cl_uint offset; // the first vector size to build
121     cl_kernel *kernels;
122     cl_program *programs;
123     const char *nameInCode;
124     bool relaxedMode; // Whether to build with -cl-fast-relaxed-math.
125 } BuildKernelInfo;
126 
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)127 static cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
128 {
129     BuildKernelInfo *info = (BuildKernelInfo *)p;
130     cl_uint i = info->offset + job_id;
131     return BuildKernel(info->nameInCode, i, info->kernels + i,
132                        info->programs + i, info->relaxedMode);
133 }
134 
135 typedef struct ComputeReferenceInfoD_
136 {
137     const double *x;
138     const double *y;
139     double *r;
140     int *i;
141     long double (*f_ffpI)(long double, long double, int *);
142     cl_uint lim;
143     cl_uint count;
144 } ComputeReferenceInfoD;
145 
ReferenceD(cl_uint jid,cl_uint tid,void * userInfo)146 static cl_int ReferenceD(cl_uint jid, cl_uint tid, void *userInfo)
147 {
148     ComputeReferenceInfoD *cri = (ComputeReferenceInfoD *)userInfo;
149     cl_uint lim = cri->lim;
150     cl_uint count = cri->count;
151     cl_uint off = jid * count;
152     const double *x = cri->x + off;
153     const double *y = cri->y + off;
154     double *r = cri->r + off;
155     int *i = cri->i + off;
156     long double (*f)(long double, long double, int *) = cri->f_ffpI;
157 
158     if (off + count > lim) count = lim - off;
159 
160     Force64BitFPUPrecision();
161 
162     for (cl_uint j = 0; j < count; ++j)
163         r[j] = (double)f((long double)x[j], (long double)y[j], i + j);
164 
165     return CL_SUCCESS;
166 }
167 
TestFunc_DoubleI_Double_Double(const Func * f,MTdata d,bool relaxedMode)168 int TestFunc_DoubleI_Double_Double(const Func *f, MTdata d, bool relaxedMode)
169 {
170     int error;
171     cl_program programs[VECTOR_SIZE_COUNT];
172     cl_kernel kernels[VECTOR_SIZE_COUNT];
173     float maxError = 0.0f;
174     int64_t maxError2 = 0;
175     int ftz = f->ftz || gForceFTZ;
176     double maxErrorVal = 0.0f;
177     double maxErrorVal2 = 0.0f;
178     uint64_t step = getTestStep(sizeof(double), BUFFER_SIZE);
179 
180     logFunctionInfo(f->name, sizeof(cl_double), relaxedMode);
181 
182     cl_uint threadCount = GetThreadCount();
183 
184     Force64BitFPUPrecision();
185 
186     int testingRemquo = !strcmp(f->name, "remquo");
187 
188     // Init the kernels
189     {
190         BuildKernelInfo build_info = { gMinVectorSizeIndex, kernels, programs,
191                                        f->nameInCode, relaxedMode };
192         if ((error = ThreadPool_Do(BuildKernelFn,
193                                    gMaxVectorSizeIndex - gMinVectorSizeIndex,
194                                    &build_info)))
195             return error;
196     }
197 
198     for (uint64_t i = 0; i < (1ULL << 32); i += step)
199     {
200         // Init input array
201         double *p = (double *)gIn;
202         double *p2 = (double *)gIn2;
203         for (size_t j = 0; j < BUFFER_SIZE / sizeof(double); j++)
204         {
205             p[j] = DoubleFromUInt32(genrand_int32(d));
206             p2[j] = DoubleFromUInt32(genrand_int32(d));
207         }
208 
209         if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0,
210                                           BUFFER_SIZE, gIn, 0, NULL, NULL)))
211         {
212             vlog_error("\n*** Error %d in clEnqueueWriteBuffer ***\n", error);
213             return error;
214         }
215 
216         if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer2, CL_FALSE, 0,
217                                           BUFFER_SIZE, gIn2, 0, NULL, NULL)))
218         {
219             vlog_error("\n*** Error %d in clEnqueueWriteBuffer2 ***\n", error);
220             return error;
221         }
222 
223         // write garbage into output arrays
224         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
225         {
226             uint32_t pattern = 0xffffdead;
227             memset_pattern4(gOut[j], &pattern, BUFFER_SIZE);
228             if ((error =
229                      clEnqueueWriteBuffer(gQueue, gOutBuffer[j], CL_FALSE, 0,
230                                           BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
231             {
232                 vlog_error("\n*** Error %d in clEnqueueWriteBuffer2(%d) ***\n",
233                            error, j);
234                 goto exit;
235             }
236 
237             memset_pattern4(gOut2[j], &pattern, BUFFER_SIZE);
238             if ((error = clEnqueueWriteBuffer(gQueue, gOutBuffer2[j], CL_FALSE,
239                                               0, BUFFER_SIZE, gOut2[j], 0, NULL,
240                                               NULL)))
241             {
242                 vlog_error("\n*** Error %d in clEnqueueWriteBuffer2b(%d) ***\n",
243                            error, j);
244                 goto exit;
245             }
246         }
247 
248         // Run the kernels
249         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
250         {
251             size_t vectorSize = sizeof(cl_double) * sizeValues[j];
252             size_t localCount = (BUFFER_SIZE + vectorSize - 1)
253                 / vectorSize; // BUFFER_SIZE / vectorSize  rounded up
254             if ((error = clSetKernelArg(kernels[j], 0, sizeof(gOutBuffer[j]),
255                                         &gOutBuffer[j])))
256             {
257                 LogBuildError(programs[j]);
258                 goto exit;
259             }
260             if ((error = clSetKernelArg(kernels[j], 1, sizeof(gOutBuffer2[j]),
261                                         &gOutBuffer2[j])))
262             {
263                 LogBuildError(programs[j]);
264                 goto exit;
265             }
266             if ((error = clSetKernelArg(kernels[j], 2, sizeof(gInBuffer),
267                                         &gInBuffer)))
268             {
269                 LogBuildError(programs[j]);
270                 goto exit;
271             }
272             if ((error = clSetKernelArg(kernels[j], 3, sizeof(gInBuffer2),
273                                         &gInBuffer2)))
274             {
275                 LogBuildError(programs[j]);
276                 goto exit;
277             }
278 
279             if ((error =
280                      clEnqueueNDRangeKernel(gQueue, kernels[j], 1, NULL,
281                                             &localCount, NULL, 0, NULL, NULL)))
282             {
283                 vlog_error("FAILED -- could not execute kernel\n");
284                 goto exit;
285             }
286         }
287 
288         // Get that moving
289         if ((error = clFlush(gQueue))) vlog("clFlush failed\n");
290 
291         // Calculate the correctly rounded reference result
292         double *s = (double *)gIn;
293         double *s2 = (double *)gIn2;
294 
295         if (threadCount > 1)
296         {
297             ComputeReferenceInfoD cri;
298             cri.x = s;
299             cri.y = s2;
300             cri.r = (double *)gOut_Ref;
301             cri.i = (int *)gOut_Ref2;
302             cri.f_ffpI = f->dfunc.f_ffpI;
303             cri.lim = BUFFER_SIZE / sizeof(double);
304             cri.count = (cri.lim + threadCount - 1) / threadCount;
305             ThreadPool_Do(ReferenceD, threadCount, &cri);
306         }
307         else
308         {
309             double *r = (double *)gOut_Ref;
310             int *r2 = (int *)gOut_Ref2;
311             for (size_t j = 0; j < BUFFER_SIZE / sizeof(double); j++)
312                 r[j] = (double)f->dfunc.f_ffpI(s[j], s2[j], r2 + j);
313         }
314 
315         // Read the data back
316         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
317         {
318             if ((error =
319                      clEnqueueReadBuffer(gQueue, gOutBuffer[j], CL_TRUE, 0,
320                                          BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
321             {
322                 vlog_error("ReadArray failed %d\n", error);
323                 goto exit;
324             }
325             if ((error =
326                      clEnqueueReadBuffer(gQueue, gOutBuffer2[j], CL_TRUE, 0,
327                                          BUFFER_SIZE, gOut2[j], 0, NULL, NULL)))
328             {
329                 vlog_error("ReadArray2 failed %d\n", error);
330                 goto exit;
331             }
332         }
333 
334         if (gSkipCorrectnessTesting) break;
335 
336         // Verify data
337         uint64_t *t = (uint64_t *)gOut_Ref;
338         int32_t *t2 = (int32_t *)gOut_Ref2;
339         for (size_t j = 0; j < BUFFER_SIZE / sizeof(double); j++)
340         {
341             for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
342             {
343                 uint64_t *q = (uint64_t *)gOut[k];
344                 int32_t *q2 = (int32_t *)gOut2[k];
345 
346                 // Check for exact match to correctly rounded result
347                 if (t[j] == q[j] && t2[j] == q2[j]) continue;
348 
349                 // Check for paired NaNs
350                 if ((t[j] & 0x7fffffffffffffffUL) > 0x7ff0000000000000UL
351                     && (q[j] & 0x7fffffffffffffffUL) > 0x7ff0000000000000UL
352                     && t2[j] == q2[j])
353                     continue;
354 
355                 double test = ((double *)q)[j];
356                 int correct2 = INT_MIN;
357                 long double correct = f->dfunc.f_ffpI(s[j], s2[j], &correct2);
358                 float err = Bruteforce_Ulp_Error_Double(test, correct);
359                 int64_t iErr;
360 
361                 // in case of remquo, we only care about the sign and last
362                 // seven bits of integer as per the spec.
363                 if (testingRemquo)
364                     iErr = (long long)(q2[j] & 0x0000007f)
365                         - (long long)(correct2 & 0x0000007f);
366                 else
367                     iErr = (long long)q2[j] - (long long)correct2;
368 
369                 // For remquo, if y = 0, x is infinite, or either is NaN
370                 // then the standard either neglects to say what is returned
371                 // in iptr or leaves it undefined or implementation defined.
372                 int iptrUndefined = fabs(((double *)gIn)[j]) == INFINITY
373                     || ((double *)gIn2)[j] == 0.0 || isnan(((double *)gIn2)[j])
374                     || isnan(((double *)gIn)[j]);
375                 if (iptrUndefined) iErr = 0;
376 
377                 int fail = !(fabsf(err) <= f->double_ulps && iErr == 0);
378                 if (ftz && fail)
379                 {
380                     // retry per section 6.5.3.2
381                     if (IsDoubleResultSubnormal(correct, f->double_ulps))
382                     {
383                         fail = fail && !(test == 0.0f && iErr == 0);
384                         if (!fail) err = 0.0f;
385                     }
386 
387                     // retry per section 6.5.3.3
388                     if (IsDoubleSubnormal(s[j]))
389                     {
390                         int correct3i, correct4i;
391                         long double correct3 =
392                             f->dfunc.f_ffpI(0.0, s2[j], &correct3i);
393                         long double correct4 =
394                             f->dfunc.f_ffpI(-0.0, s2[j], &correct4i);
395                         float err2 =
396                             Bruteforce_Ulp_Error_Double(test, correct3);
397                         float err3 =
398                             Bruteforce_Ulp_Error_Double(test, correct4);
399                         int64_t iErr3 = (long long)q2[j] - (long long)correct3i;
400                         int64_t iErr4 = (long long)q2[j] - (long long)correct4i;
401                         fail = fail
402                             && ((!(fabsf(err2) <= f->double_ulps && iErr3 == 0))
403                                 && (!(fabsf(err3) <= f->double_ulps
404                                       && iErr4 == 0)));
405                         if (fabsf(err2) < fabsf(err)) err = err2;
406                         if (fabsf(err3) < fabsf(err)) err = err3;
407                         if (llabs(iErr3) < llabs(iErr)) iErr = iErr3;
408                         if (llabs(iErr4) < llabs(iErr)) iErr = iErr4;
409 
410                         // retry per section 6.5.3.4
411                         if (IsDoubleResultSubnormal(correct2, f->double_ulps)
412                             || IsDoubleResultSubnormal(correct3,
413                                                        f->double_ulps))
414                         {
415                             fail = fail
416                                 && !(test == 0.0f
417                                      && (iErr3 == 0 || iErr4 == 0));
418                             if (!fail) err = 0.0f;
419                         }
420 
421                         // try with both args as zero
422                         if (IsDoubleSubnormal(s2[j]))
423                         {
424                             int correct7i, correct8i;
425                             correct3 = f->dfunc.f_ffpI(0.0, 0.0, &correct3i);
426                             correct4 = f->dfunc.f_ffpI(-0.0, 0.0, &correct4i);
427                             long double correct7 =
428                                 f->dfunc.f_ffpI(0.0, -0.0, &correct7i);
429                             long double correct8 =
430                                 f->dfunc.f_ffpI(-0.0, -0.0, &correct8i);
431                             err2 = Bruteforce_Ulp_Error_Double(test, correct3);
432                             err3 = Bruteforce_Ulp_Error_Double(test, correct4);
433                             float err4 =
434                                 Bruteforce_Ulp_Error_Double(test, correct7);
435                             float err5 =
436                                 Bruteforce_Ulp_Error_Double(test, correct8);
437                             iErr3 = (long long)q2[j] - (long long)correct3i;
438                             iErr4 = (long long)q2[j] - (long long)correct4i;
439                             int64_t iErr7 =
440                                 (long long)q2[j] - (long long)correct7i;
441                             int64_t iErr8 =
442                                 (long long)q2[j] - (long long)correct8i;
443                             fail = fail
444                                 && ((!(fabsf(err2) <= f->double_ulps
445                                        && iErr3 == 0))
446                                     && (!(fabsf(err3) <= f->double_ulps
447                                           && iErr4 == 0))
448                                     && (!(fabsf(err4) <= f->double_ulps
449                                           && iErr7 == 0))
450                                     && (!(fabsf(err5) <= f->double_ulps
451                                           && iErr8 == 0)));
452                             if (fabsf(err2) < fabsf(err)) err = err2;
453                             if (fabsf(err3) < fabsf(err)) err = err3;
454                             if (fabsf(err4) < fabsf(err)) err = err4;
455                             if (fabsf(err5) < fabsf(err)) err = err5;
456                             if (llabs(iErr3) < llabs(iErr)) iErr = iErr3;
457                             if (llabs(iErr4) < llabs(iErr)) iErr = iErr4;
458                             if (llabs(iErr7) < llabs(iErr)) iErr = iErr7;
459                             if (llabs(iErr8) < llabs(iErr)) iErr = iErr8;
460 
461                             // retry per section 6.5.3.4
462                             if (IsDoubleResultSubnormal(correct3,
463                                                         f->double_ulps)
464                                 || IsDoubleResultSubnormal(correct4,
465                                                            f->double_ulps)
466                                 || IsDoubleResultSubnormal(correct7,
467                                                            f->double_ulps)
468                                 || IsDoubleResultSubnormal(correct8,
469                                                            f->double_ulps))
470                             {
471                                 fail = fail
472                                     && !(test == 0.0f
473                                          && (iErr3 == 0 || iErr4 == 0
474                                              || iErr7 == 0 || iErr8 == 0));
475                                 if (!fail) err = 0.0f;
476                             }
477                         }
478                     }
479                     else if (IsDoubleSubnormal(s2[j]))
480                     {
481                         int correct3i, correct4i;
482                         long double correct3 =
483                             f->dfunc.f_ffpI(s[j], 0.0, &correct3i);
484                         long double correct4 =
485                             f->dfunc.f_ffpI(s[j], -0.0, &correct4i);
486                         float err2 =
487                             Bruteforce_Ulp_Error_Double(test, correct3);
488                         float err3 =
489                             Bruteforce_Ulp_Error_Double(test, correct4);
490                         int64_t iErr3 = (long long)q2[j] - (long long)correct3i;
491                         int64_t iErr4 = (long long)q2[j] - (long long)correct4i;
492                         fail = fail
493                             && ((!(fabsf(err2) <= f->double_ulps && iErr3 == 0))
494                                 && (!(fabsf(err3) <= f->double_ulps
495                                       && iErr4 == 0)));
496                         if (fabsf(err2) < fabsf(err)) err = err2;
497                         if (fabsf(err3) < fabsf(err)) err = err3;
498                         if (llabs(iErr3) < llabs(iErr)) iErr = iErr3;
499                         if (llabs(iErr4) < llabs(iErr)) iErr = iErr4;
500 
501                         // retry per section 6.5.3.4
502                         if (IsDoubleResultSubnormal(correct2, f->double_ulps)
503                             || IsDoubleResultSubnormal(correct3,
504                                                        f->double_ulps))
505                         {
506                             fail = fail
507                                 && !(test == 0.0f
508                                      && (iErr3 == 0 || iErr4 == 0));
509                             if (!fail) err = 0.0f;
510                         }
511                     }
512                 }
513                 if (fabsf(err) > maxError)
514                 {
515                     maxError = fabsf(err);
516                     maxErrorVal = s[j];
517                 }
518                 if (llabs(iErr) > maxError2)
519                 {
520                     maxError2 = llabs(iErr);
521                     maxErrorVal2 = s[j];
522                 }
523 
524                 if (fail)
525                 {
526                     vlog_error(
527                         "\nERROR: %sD%s: {%f, %lld} ulp error at {%.13la, "
528                         "%.13la} ({ 0x%16.16llx, 0x%16.16llx}): *{%.13la, "
529                         "%d} ({ 0x%16.16llx, 0x%8.8x}) vs. {%.13la, %d} ({ "
530                         "0x%16.16llx, 0x%8.8x})\n",
531                         f->name, sizeNames[k], err, iErr, ((double *)gIn)[j],
532                         ((double *)gIn2)[j], ((cl_ulong *)gIn)[j],
533                         ((cl_ulong *)gIn2)[j], ((double *)gOut_Ref)[j],
534                         ((int *)gOut_Ref2)[j], ((cl_ulong *)gOut_Ref)[j],
535                         ((cl_uint *)gOut_Ref2)[j], test, q2[j],
536                         ((cl_ulong *)q)[j], ((cl_uint *)q2)[j]);
537                     error = -1;
538                     goto exit;
539                 }
540             }
541         }
542 
543         if (0 == (i & 0x0fffffff))
544         {
545             if (gVerboseBruteForce)
546             {
547                 vlog("base:%14u step:%10zu  bufferSize:%10zd \n", i, step,
548                      BUFFER_SIZE);
549             }
550             else
551             {
552                 vlog(".");
553             }
554             fflush(stdout);
555         }
556     }
557 
558     if (!gSkipCorrectnessTesting)
559     {
560         if (gWimpyMode)
561             vlog("Wimp pass");
562         else
563             vlog("passed");
564 
565         vlog("\t{%8.2f, %lld} @ {%a, %a}", maxError, maxError2, maxErrorVal,
566              maxErrorVal2);
567     }
568 
569     vlog("\n");
570 
571 exit:
572     // Release
573     for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
574     {
575         clReleaseKernel(kernels[k]);
576         clReleaseProgram(programs[k]);
577     }
578 
579     return error;
580 }
581