1 //
2 // Copyright (c) 2017 The Khronos Group Inc.
3 //
4 // Licensed under the Apache License, Version 2.0 (the "License");
5 // you may not use this file except in compliance with the License.
6 // You may obtain a copy of the License at
7 //
8 //    http://www.apache.org/licenses/LICENSE-2.0
9 //
10 // Unless required by applicable law or agreed to in writing, software
11 // distributed under the License is distributed on an "AS IS" BASIS,
12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 // See the License for the specific language governing permissions and
14 // limitations under the License.
15 //
16 
17 #include "common.h"
18 #include "function_list.h"
19 #include "test_functions.h"
20 #include "utility.h"
21 
22 #include <cinttypes>
23 #include <climits>
24 #include <cstring>
25 
26 namespace {
27 
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)28 cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
29 {
30     BuildKernelInfo &info = *(BuildKernelInfo *)p;
31     auto generator = [](const std::string &kernel_name, const char *builtin,
32                         cl_uint vector_size_index) {
33         return GetBinaryKernel(kernel_name, builtin, ParameterType::Float,
34                                ParameterType::Int, ParameterType::Float,
35                                ParameterType::Float, vector_size_index);
36     };
37     return BuildKernels(info, job_id, generator);
38 }
39 
40 struct ComputeReferenceInfoF
41 {
42     const float *x;
43     const float *y;
44     float *r;
45     int *i;
46     double (*f_ffpI)(double, double, int *);
47     cl_uint lim;
48     cl_uint count;
49 };
50 
ReferenceF(cl_uint jid,cl_uint tid,void * userInfo)51 cl_int ReferenceF(cl_uint jid, cl_uint tid, void *userInfo)
52 {
53     ComputeReferenceInfoF *cri = (ComputeReferenceInfoF *)userInfo;
54     cl_uint lim = cri->lim;
55     cl_uint count = cri->count;
56     cl_uint off = jid * count;
57     const float *x = cri->x + off;
58     const float *y = cri->y + off;
59     float *r = cri->r + off;
60     int *i = cri->i + off;
61     double (*f)(double, double, int *) = cri->f_ffpI;
62 
63     if (off + count > lim) count = lim - off;
64 
65     for (cl_uint j = 0; j < count; ++j)
66         r[j] = (float)f((double)x[j], (double)y[j], i + j);
67 
68     return CL_SUCCESS;
69 }
70 
71 } // anonymous namespace
72 
TestFunc_FloatI_Float_Float(const Func * f,MTdata d,bool relaxedMode)73 int TestFunc_FloatI_Float_Float(const Func *f, MTdata d, bool relaxedMode)
74 {
75     int error;
76 
77     logFunctionInfo(f->name, sizeof(cl_float), relaxedMode);
78 
79     Programs programs;
80     const unsigned thread_id = 0; // Test is currently not multithreaded.
81     KernelMatrix kernels;
82     float maxError = 0.0f;
83     int ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
84     int64_t maxError2 = 0;
85     float maxErrorVal = 0.0f;
86     float maxErrorVal2 = 0.0f;
87     uint64_t step = getTestStep(sizeof(float), BUFFER_SIZE);
88 
89     cl_uint threadCount = GetThreadCount();
90 
91     float float_ulps;
92     if (gIsEmbedded)
93         float_ulps = f->float_embedded_ulps;
94     else
95         float_ulps = f->float_ulps;
96 
97     int testingRemquo = !strcmp(f->name, "remquo");
98 
99     // Init the kernels
100     BuildKernelInfo build_info{ 1, kernels, programs, f->nameInCode,
101                                 relaxedMode };
102     if ((error = ThreadPool_Do(BuildKernelFn,
103                                gMaxVectorSizeIndex - gMinVectorSizeIndex,
104                                &build_info)))
105         return error;
106 
107     for (uint64_t i = 0; i < (1ULL << 32); i += step)
108     {
109         // Init input array
110         cl_uint *p = (cl_uint *)gIn;
111         cl_uint *p2 = (cl_uint *)gIn2;
112         for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
113         {
114             p[j] = genrand_int32(d);
115             p2[j] = genrand_int32(d);
116         }
117 
118         if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0,
119                                           BUFFER_SIZE, gIn, 0, NULL, NULL)))
120         {
121             vlog_error("\n*** Error %d in clEnqueueWriteBuffer ***\n", error);
122             return error;
123         }
124 
125         if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer2, CL_FALSE, 0,
126                                           BUFFER_SIZE, gIn2, 0, NULL, NULL)))
127         {
128             vlog_error("\n*** Error %d in clEnqueueWriteBuffer2 ***\n", error);
129             return error;
130         }
131 
132         // Write garbage into output arrays
133         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
134         {
135             uint32_t pattern = 0xffffdead;
136             if (gHostFill)
137             {
138                 memset_pattern4(gOut[j], &pattern, BUFFER_SIZE);
139                 if ((error = clEnqueueWriteBuffer(gQueue, gOutBuffer[j],
140                                                   CL_FALSE, 0, BUFFER_SIZE,
141                                                   gOut[j], 0, NULL, NULL)))
142                 {
143                     vlog_error(
144                         "\n*** Error %d in clEnqueueWriteBuffer2(%d) ***\n",
145                         error, j);
146                     return error;
147                 }
148 
149                 memset_pattern4(gOut2[j], &pattern, BUFFER_SIZE);
150                 if ((error = clEnqueueWriteBuffer(gQueue, gOutBuffer2[j],
151                                                   CL_FALSE, 0, BUFFER_SIZE,
152                                                   gOut2[j], 0, NULL, NULL)))
153                 {
154                     vlog_error(
155                         "\n*** Error %d in clEnqueueWriteBuffer2b(%d) ***\n",
156                         error, j);
157                     return error;
158                 }
159             }
160             else
161             {
162                 if ((error = clEnqueueFillBuffer(gQueue, gOutBuffer[j],
163                                                  &pattern, sizeof(pattern), 0,
164                                                  BUFFER_SIZE, 0, NULL, NULL)))
165                 {
166                     vlog_error("Error: clEnqueueFillBuffer 1 failed! err: %d\n",
167                                error);
168                     return error;
169                 }
170 
171                 if ((error = clEnqueueFillBuffer(gQueue, gOutBuffer2[j],
172                                                  &pattern, sizeof(pattern), 0,
173                                                  BUFFER_SIZE, 0, NULL, NULL)))
174                 {
175                     vlog_error("Error: clEnqueueFillBuffer 2 failed! err: %d\n",
176                                error);
177                     return error;
178                 }
179             }
180         }
181 
182         // Run the kernels
183         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
184         {
185             size_t vectorSize = sizeof(cl_float) * sizeValues[j];
186             size_t localCount = (BUFFER_SIZE + vectorSize - 1)
187                 / vectorSize; // BUFFER_SIZE / vectorSize  rounded up
188             if ((error = clSetKernelArg(kernels[j][thread_id], 0,
189                                         sizeof(gOutBuffer[j]), &gOutBuffer[j])))
190             {
191                 LogBuildError(programs[j]);
192                 return error;
193             }
194             if ((error =
195                      clSetKernelArg(kernels[j][thread_id], 1,
196                                     sizeof(gOutBuffer2[j]), &gOutBuffer2[j])))
197             {
198                 LogBuildError(programs[j]);
199                 return error;
200             }
201             if ((error = clSetKernelArg(kernels[j][thread_id], 2,
202                                         sizeof(gInBuffer), &gInBuffer)))
203             {
204                 LogBuildError(programs[j]);
205                 return error;
206             }
207             if ((error = clSetKernelArg(kernels[j][thread_id], 3,
208                                         sizeof(gInBuffer2), &gInBuffer2)))
209             {
210                 LogBuildError(programs[j]);
211                 return error;
212             }
213 
214             if ((error = clEnqueueNDRangeKernel(gQueue, kernels[j][thread_id],
215                                                 1, NULL, &localCount, NULL, 0,
216                                                 NULL, NULL)))
217             {
218                 vlog_error("FAILED -- could not execute kernel\n");
219                 return error;
220             }
221         }
222 
223         // Get that moving
224         if ((error = clFlush(gQueue))) vlog("clFlush failed\n");
225 
226         // Calculate the correctly rounded reference result
227         float *s = (float *)gIn;
228         float *s2 = (float *)gIn2;
229 
230         if (threadCount > 1)
231         {
232             ComputeReferenceInfoF cri;
233             cri.x = s;
234             cri.y = s2;
235             cri.r = (float *)gOut_Ref;
236             cri.i = (int *)gOut_Ref2;
237             cri.f_ffpI = f->func.f_ffpI;
238             cri.lim = BUFFER_SIZE / sizeof(float);
239             cri.count = (cri.lim + threadCount - 1) / threadCount;
240             ThreadPool_Do(ReferenceF, threadCount, &cri);
241         }
242         else
243         {
244             float *r = (float *)gOut_Ref;
245             int *r2 = (int *)gOut_Ref2;
246             for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
247                 r[j] = (float)f->func.f_ffpI(s[j], s2[j], r2 + j);
248         }
249 
250         // Read the data back
251         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
252         {
253             if ((error =
254                      clEnqueueReadBuffer(gQueue, gOutBuffer[j], CL_TRUE, 0,
255                                          BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
256             {
257                 vlog_error("ReadArray failed %d\n", error);
258                 return error;
259             }
260             if ((error =
261                      clEnqueueReadBuffer(gQueue, gOutBuffer2[j], CL_TRUE, 0,
262                                          BUFFER_SIZE, gOut2[j], 0, NULL, NULL)))
263             {
264                 vlog_error("ReadArray2 failed %d\n", error);
265                 return error;
266             }
267         }
268 
269         if (gSkipCorrectnessTesting) break;
270 
271         // Verify data
272         uint32_t *t = (uint32_t *)gOut_Ref;
273         int32_t *t2 = (int32_t *)gOut_Ref2;
274         for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
275         {
276             for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
277             {
278                 uint32_t *q = (uint32_t *)(gOut[k]);
279                 int32_t *q2 = (int32_t *)gOut2[k];
280 
281                 // Check for exact match to correctly rounded result
282                 if (t[j] == q[j] && t2[j] == q2[j]) continue;
283 
284                 // Check for paired NaNs
285                 if ((t[j] & 0x7fffffff) > 0x7f800000
286                     && (q[j] & 0x7fffffff) > 0x7f800000 && t2[j] == q2[j])
287                     continue;
288 
289                 float test = ((float *)q)[j];
290                 int correct2 = INT_MIN;
291                 double correct = f->func.f_ffpI(s[j], s2[j], &correct2);
292                 float err = Ulp_Error(test, correct);
293                 int64_t iErr;
294 
295                 // in case of remquo, we only care about the sign and last
296                 // seven bits of integer as per the spec.
297                 if (testingRemquo)
298                     iErr = (long long)(q2[j] & 0x0000007f)
299                         - (long long)(correct2 & 0x0000007f);
300                 else
301                     iErr = (long long)q2[j] - (long long)correct2;
302 
303                 // For remquo, if y = 0, x is infinite, or either is NaN
304                 // then the standard either neglects to say what is returned
305                 // in iptr or leaves it undefined or implementation defined.
306                 int iptrUndefined = fabs(((float *)gIn)[j]) == INFINITY
307                     || ((float *)gIn2)[j] == 0.0f || isnan(((float *)gIn2)[j])
308                     || isnan(((float *)gIn)[j]);
309                 if (iptrUndefined) iErr = 0;
310 
311                 int fail = !(fabsf(err) <= float_ulps && iErr == 0);
312                 if ((ftz || relaxedMode) && fail)
313                 {
314                     // retry per section 6.5.3.2
315                     if (IsFloatResultSubnormal(correct, float_ulps))
316                     {
317                         fail = fail && !(test == 0.0f && iErr == 0);
318                         if (!fail) err = 0.0f;
319                     }
320 
321                     // retry per section 6.5.3.3
322                     if (IsFloatSubnormal(s[j]))
323                     {
324                         int correct3i, correct4i;
325                         double correct3 =
326                             f->func.f_ffpI(0.0, s2[j], &correct3i);
327                         double correct4 =
328                             f->func.f_ffpI(-0.0, s2[j], &correct4i);
329                         float err2 = Ulp_Error(test, correct3);
330                         float err3 = Ulp_Error(test, correct4);
331                         int64_t iErr3 = (long long)q2[j] - (long long)correct3i;
332                         int64_t iErr4 = (long long)q2[j] - (long long)correct4i;
333                         fail = fail
334                             && ((!(fabsf(err2) <= float_ulps && iErr3 == 0))
335                                 && (!(fabsf(err3) <= float_ulps
336                                       && iErr4 == 0)));
337                         if (fabsf(err2) < fabsf(err)) err = err2;
338                         if (fabsf(err3) < fabsf(err)) err = err3;
339                         if (llabs(iErr3) < llabs(iErr)) iErr = iErr3;
340                         if (llabs(iErr4) < llabs(iErr)) iErr = iErr4;
341 
342                         // retry per section 6.5.3.4
343                         if (IsFloatResultSubnormal(correct2, float_ulps)
344                             || IsFloatResultSubnormal(correct3, float_ulps))
345                         {
346                             fail = fail
347                                 && !(test == 0.0f
348                                      && (iErr3 == 0 || iErr4 == 0));
349                             if (!fail) err = 0.0f;
350                         }
351 
352                         // try with both args as zero
353                         if (IsFloatSubnormal(s2[j]))
354                         {
355                             int correct7i, correct8i;
356                             correct3 = f->func.f_ffpI(0.0, 0.0, &correct3i);
357                             correct4 = f->func.f_ffpI(-0.0, 0.0, &correct4i);
358                             double correct7 =
359                                 f->func.f_ffpI(0.0, -0.0, &correct7i);
360                             double correct8 =
361                                 f->func.f_ffpI(-0.0, -0.0, &correct8i);
362                             err2 = Ulp_Error(test, correct3);
363                             err3 = Ulp_Error(test, correct4);
364                             float err4 = Ulp_Error(test, correct7);
365                             float err5 = Ulp_Error(test, correct8);
366                             iErr3 = (long long)q2[j] - (long long)correct3i;
367                             iErr4 = (long long)q2[j] - (long long)correct4i;
368                             int64_t iErr7 =
369                                 (long long)q2[j] - (long long)correct7i;
370                             int64_t iErr8 =
371                                 (long long)q2[j] - (long long)correct8i;
372                             fail = fail
373                                 && ((!(fabsf(err2) <= float_ulps && iErr3 == 0))
374                                     && (!(fabsf(err3) <= float_ulps
375                                           && iErr4 == 0))
376                                     && (!(fabsf(err4) <= float_ulps
377                                           && iErr7 == 0))
378                                     && (!(fabsf(err5) <= float_ulps
379                                           && iErr8 == 0)));
380                             if (fabsf(err2) < fabsf(err)) err = err2;
381                             if (fabsf(err3) < fabsf(err)) err = err3;
382                             if (fabsf(err4) < fabsf(err)) err = err4;
383                             if (fabsf(err5) < fabsf(err)) err = err5;
384                             if (llabs(iErr3) < llabs(iErr)) iErr = iErr3;
385                             if (llabs(iErr4) < llabs(iErr)) iErr = iErr4;
386                             if (llabs(iErr7) < llabs(iErr)) iErr = iErr7;
387                             if (llabs(iErr8) < llabs(iErr)) iErr = iErr8;
388 
389                             // retry per section 6.5.3.4
390                             if (IsFloatResultSubnormal(correct3, float_ulps)
391                                 || IsFloatResultSubnormal(correct4, float_ulps)
392                                 || IsFloatResultSubnormal(correct7, float_ulps)
393                                 || IsFloatResultSubnormal(correct8, float_ulps))
394                             {
395                                 fail = fail
396                                     && !(test == 0.0f
397                                          && (iErr3 == 0 || iErr4 == 0
398                                              || iErr7 == 0 || iErr8 == 0));
399                                 if (!fail) err = 0.0f;
400                             }
401                         }
402                     }
403                     else if (IsFloatSubnormal(s2[j]))
404                     {
405                         int correct3i, correct4i;
406                         double correct3 = f->func.f_ffpI(s[j], 0.0, &correct3i);
407                         double correct4 =
408                             f->func.f_ffpI(s[j], -0.0, &correct4i);
409                         float err2 = Ulp_Error(test, correct3);
410                         float err3 = Ulp_Error(test, correct4);
411                         int64_t iErr3 = (long long)q2[j] - (long long)correct3i;
412                         int64_t iErr4 = (long long)q2[j] - (long long)correct4i;
413                         fail = fail
414                             && ((!(fabsf(err2) <= float_ulps && iErr3 == 0))
415                                 && (!(fabsf(err3) <= float_ulps
416                                       && iErr4 == 0)));
417                         if (fabsf(err2) < fabsf(err)) err = err2;
418                         if (fabsf(err3) < fabsf(err)) err = err3;
419                         if (llabs(iErr3) < llabs(iErr)) iErr = iErr3;
420                         if (llabs(iErr4) < llabs(iErr)) iErr = iErr4;
421 
422                         // retry per section 6.5.3.4
423                         if (IsFloatResultSubnormal(correct2, float_ulps)
424                             || IsFloatResultSubnormal(correct3, float_ulps))
425                         {
426                             fail = fail
427                                 && !(test == 0.0f
428                                      && (iErr3 == 0 || iErr4 == 0));
429                             if (!fail) err = 0.0f;
430                         }
431                     }
432                 }
433                 if (fabsf(err) > maxError)
434                 {
435                     maxError = fabsf(err);
436                     maxErrorVal = s[j];
437                 }
438                 if (llabs(iErr) > maxError2)
439                 {
440                     maxError2 = llabs(iErr);
441                     maxErrorVal2 = s[j];
442                 }
443 
444                 if (fail)
445                 {
446                     vlog_error("\nERROR: %s%s: {%f, %" PRId64
447                                "} ulp error at {%a, %a} "
448                                "({0x%8.8x, 0x%8.8x}): *{%a, %d} ({0x%8.8x, "
449                                "0x%8.8x}) vs. {%a, %d} ({0x%8.8x, 0x%8.8x})\n",
450                                f->name, sizeNames[k], err, iErr,
451                                ((float *)gIn)[j], ((float *)gIn2)[j],
452                                ((cl_uint *)gIn)[j], ((cl_uint *)gIn2)[j],
453                                ((float *)gOut_Ref)[j], ((int *)gOut_Ref2)[j],
454                                ((cl_uint *)gOut_Ref)[j],
455                                ((cl_uint *)gOut_Ref2)[j], test, q2[j],
456                                ((cl_uint *)&test)[0], ((cl_uint *)q2)[j]);
457                     return -1;
458                 }
459             }
460         }
461 
462         if (0 == (i & 0x0fffffff))
463         {
464             if (gVerboseBruteForce)
465             {
466                 vlog("base:%14" PRIu64 " step:%10" PRIu64
467                      "  bufferSize:%10d \n",
468                      i, step, BUFFER_SIZE);
469             }
470             else
471             {
472                 vlog(".");
473             }
474             fflush(stdout);
475         }
476     }
477 
478     if (!gSkipCorrectnessTesting)
479     {
480         if (gWimpyMode)
481             vlog("Wimp pass");
482         else
483             vlog("passed");
484 
485         vlog("\t{%8.2f, %" PRId64 "} @ {%a, %a}", maxError, maxError2,
486              maxErrorVal, maxErrorVal2);
487     }
488 
489     vlog("\n");
490 
491     return CL_SUCCESS;
492 }
493