1 //
2 // Copyright (c) 2017 The Khronos Group Inc.
3 //
4 // Licensed under the Apache License, Version 2.0 (the "License");
5 // you may not use this file except in compliance with the License.
6 // You may obtain a copy of the License at
7 //
8 // http://www.apache.org/licenses/LICENSE-2.0
9 //
10 // Unless required by applicable law or agreed to in writing, software
11 // distributed under the License is distributed on an "AS IS" BASIS,
12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 // See the License for the specific language governing permissions and
14 // limitations under the License.
15 //
16
17 #include "function_list.h"
18 #include "test_functions.h"
19 #include "utility.h"
20
21 #include <cstring>
22
23 const float twoToMinus126 = MAKE_HEX_FLOAT(0x1p-126f, 1, -126);
24
BuildKernel(const char * name,int vectorSize,cl_uint kernel_count,cl_kernel * k,cl_program * p,bool relaxedMode)25 static int BuildKernel(const char *name, int vectorSize, cl_uint kernel_count,
26 cl_kernel *k, cl_program *p, bool relaxedMode)
27 {
28 const char *c[] = { "__kernel void math_kernel",
29 sizeNames[vectorSize],
30 "( __global float",
31 sizeNames[vectorSize],
32 "* out, __global float",
33 sizeNames[vectorSize],
34 "* in1, __global float",
35 sizeNames[vectorSize],
36 "* in2 )\n"
37 "{\n"
38 " size_t i = get_global_id(0);\n"
39 " out[i] = ",
40 name,
41 "( in1[i], in2[i] );\n"
42 "}\n" };
43
44 const char *c3[] = {
45 "__kernel void math_kernel",
46 sizeNames[vectorSize],
47 "( __global float* out, __global float* in, __global float* in2)\n"
48 "{\n"
49 " size_t i = get_global_id(0);\n"
50 " if( i + 1 < get_global_size(0) )\n"
51 " {\n"
52 " float3 f0 = vload3( 0, in + 3 * i );\n"
53 " float3 f1 = vload3( 0, in2 + 3 * i );\n"
54 " f0 = ",
55 name,
56 "( f0, f1 );\n"
57 " vstore3( f0, 0, out + 3*i );\n"
58 " }\n"
59 " else\n"
60 " {\n"
61 " size_t parity = i & 1; // Figure out how many elements are "
62 "left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two "
63 "buffer size \n"
64 " float3 f0;\n"
65 " float3 f1;\n"
66 " switch( parity )\n"
67 " {\n"
68 " case 1:\n"
69 " f0 = (float3)( in[3*i], NAN, NAN ); \n"
70 " f1 = (float3)( in2[3*i], NAN, NAN ); \n"
71 " break;\n"
72 " case 0:\n"
73 " f0 = (float3)( in[3*i], in[3*i+1], NAN ); \n"
74 " f1 = (float3)( in2[3*i], in2[3*i+1], NAN ); \n"
75 " break;\n"
76 " }\n"
77 " f0 = ",
78 name,
79 "( f0, f1 );\n"
80 " switch( parity )\n"
81 " {\n"
82 " case 0:\n"
83 " out[3*i+1] = f0.y; \n"
84 " // fall through\n"
85 " case 1:\n"
86 " out[3*i] = f0.x; \n"
87 " break;\n"
88 " }\n"
89 " }\n"
90 "}\n"
91 };
92
93 const char **kern = c;
94 size_t kernSize = sizeof(c) / sizeof(c[0]);
95
96 if (sizeValues[vectorSize] == 3)
97 {
98 kern = c3;
99 kernSize = sizeof(c3) / sizeof(c3[0]);
100 }
101
102 char testName[32];
103 snprintf(testName, sizeof(testName) - 1, "math_kernel%s",
104 sizeNames[vectorSize]);
105
106 return MakeKernels(kern, (cl_uint)kernSize, testName, kernel_count, k, p,
107 relaxedMode);
108 }
109
110 typedef struct BuildKernelInfo
111 {
112 cl_uint offset; // the first vector size to build
113 cl_uint kernel_count;
114 cl_kernel **kernels;
115 cl_program *programs;
116 const char *nameInCode;
117 bool relaxedMode; // Whether to build with -cl-fast-relaxed-math.
118 } BuildKernelInfo;
119
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)120 static cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
121 {
122 BuildKernelInfo *info = (BuildKernelInfo *)p;
123 cl_uint i = info->offset + job_id;
124 return BuildKernel(info->nameInCode, i, info->kernel_count,
125 info->kernels[i], info->programs + i, info->relaxedMode);
126 }
127
128 // Thread specific data for a worker thread
129 typedef struct ThreadInfo
130 {
131 cl_mem inBuf; // input buffer for the thread
132 cl_mem inBuf2; // input buffer for the thread
133 cl_mem outBuf[VECTOR_SIZE_COUNT]; // output buffers for the thread
134 float maxError; // max error value. Init to 0.
135 double
136 maxErrorValue; // position of the max error value (param 1). Init to 0.
137 double maxErrorValue2; // position of the max error value (param 2). Init
138 // to 0.
139 MTdata d;
140 cl_command_queue tQueue; // per thread command queue to improve performance
141 } ThreadInfo;
142
143 typedef struct TestInfo
144 {
145 size_t subBufferSize; // Size of the sub-buffer in elements
146 const Func *f; // A pointer to the function info
147 cl_program programs[VECTOR_SIZE_COUNT]; // programs for various vector sizes
148 cl_kernel
149 *k[VECTOR_SIZE_COUNT]; // arrays of thread-specific kernels for each
150 // worker thread: k[vector_size][thread_id]
151 ThreadInfo *
152 tinfo; // An array of thread specific information for each worker thread
153 cl_uint threadCount; // Number of worker threads
154 cl_uint jobCount; // Number of jobs
155 cl_uint step; // step between each chunk and the next.
156 cl_uint scale; // stride between individual test values
157 float ulps; // max_allowed ulps
158 int ftz; // non-zero if running in flush to zero mode
159
160 int isFDim;
161 int skipNanInf;
162 int isNextafter;
163 bool relaxedMode; // True if test is running in relaxed mode, false
164 // otherwise.
165 } TestInfo;
166
167 // A table of more difficult cases to get right
168 static const float specialValues[] = {
169 -NAN,
170 -INFINITY,
171 -FLT_MAX,
172 MAKE_HEX_FLOAT(-0x1.000002p64f, -0x1000002L, 40),
173 MAKE_HEX_FLOAT(-0x1.0p64f, -0x1L, 64),
174 MAKE_HEX_FLOAT(-0x1.fffffep63f, -0x1fffffeL, 39),
175 MAKE_HEX_FLOAT(-0x1.000002p63f, -0x1000002L, 39),
176 MAKE_HEX_FLOAT(-0x1.0p63f, -0x1L, 63),
177 MAKE_HEX_FLOAT(-0x1.fffffep62f, -0x1fffffeL, 38),
178 MAKE_HEX_FLOAT(-0x1.000002p32f, -0x1000002L, 8),
179 MAKE_HEX_FLOAT(-0x1.0p32f, -0x1L, 32),
180 MAKE_HEX_FLOAT(-0x1.fffffep31f, -0x1fffffeL, 7),
181 MAKE_HEX_FLOAT(-0x1.000002p31f, -0x1000002L, 7),
182 MAKE_HEX_FLOAT(-0x1.0p31f, -0x1L, 31),
183 MAKE_HEX_FLOAT(-0x1.fffffep30f, -0x1fffffeL, 6),
184 -1000.f,
185 -100.f,
186 -4.0f,
187 -3.5f,
188 -3.0f,
189 MAKE_HEX_FLOAT(-0x1.800002p1f, -0x1800002L, -23),
190 -2.5f,
191 MAKE_HEX_FLOAT(-0x1.7ffffep1f, -0x17ffffeL, -23),
192 -2.0f,
193 MAKE_HEX_FLOAT(-0x1.800002p0f, -0x1800002L, -24),
194 -1.5f,
195 MAKE_HEX_FLOAT(-0x1.7ffffep0f, -0x17ffffeL, -24),
196 MAKE_HEX_FLOAT(-0x1.000002p0f, -0x1000002L, -24),
197 -1.0f,
198 MAKE_HEX_FLOAT(-0x1.fffffep-1f, -0x1fffffeL, -25),
199 MAKE_HEX_FLOAT(-0x1.000002p-1f, -0x1000002L, -25),
200 -0.5f,
201 MAKE_HEX_FLOAT(-0x1.fffffep-2f, -0x1fffffeL, -26),
202 MAKE_HEX_FLOAT(-0x1.000002p-2f, -0x1000002L, -26),
203 -0.25f,
204 MAKE_HEX_FLOAT(-0x1.fffffep-3f, -0x1fffffeL, -27),
205 MAKE_HEX_FLOAT(-0x1.000002p-126f, -0x1000002L, -150),
206 -FLT_MIN,
207 MAKE_HEX_FLOAT(-0x0.fffffep-126f, -0x0fffffeL, -150),
208 MAKE_HEX_FLOAT(-0x0.000ffep-126f, -0x0000ffeL, -150),
209 MAKE_HEX_FLOAT(-0x0.0000fep-126f, -0x00000feL, -150),
210 MAKE_HEX_FLOAT(-0x0.00000ep-126f, -0x000000eL, -150),
211 MAKE_HEX_FLOAT(-0x0.00000cp-126f, -0x000000cL, -150),
212 MAKE_HEX_FLOAT(-0x0.00000ap-126f, -0x000000aL, -150),
213 MAKE_HEX_FLOAT(-0x0.000008p-126f, -0x0000008L, -150),
214 MAKE_HEX_FLOAT(-0x0.000006p-126f, -0x0000006L, -150),
215 MAKE_HEX_FLOAT(-0x0.000004p-126f, -0x0000004L, -150),
216 MAKE_HEX_FLOAT(-0x0.000002p-126f, -0x0000002L, -150),
217 -0.0f,
218
219 +NAN,
220 +INFINITY,
221 +FLT_MAX,
222 MAKE_HEX_FLOAT(+0x1.000002p64f, +0x1000002L, 40),
223 MAKE_HEX_FLOAT(+0x1.0p64f, +0x1L, 64),
224 MAKE_HEX_FLOAT(+0x1.fffffep63f, +0x1fffffeL, 39),
225 MAKE_HEX_FLOAT(+0x1.000002p63f, +0x1000002L, 39),
226 MAKE_HEX_FLOAT(+0x1.0p63f, +0x1L, 63),
227 MAKE_HEX_FLOAT(+0x1.fffffep62f, +0x1fffffeL, 38),
228 MAKE_HEX_FLOAT(+0x1.000002p32f, +0x1000002L, 8),
229 MAKE_HEX_FLOAT(+0x1.0p32f, +0x1L, 32),
230 MAKE_HEX_FLOAT(+0x1.fffffep31f, +0x1fffffeL, 7),
231 MAKE_HEX_FLOAT(+0x1.000002p31f, +0x1000002L, 7),
232 MAKE_HEX_FLOAT(+0x1.0p31f, +0x1L, 31),
233 MAKE_HEX_FLOAT(+0x1.fffffep30f, +0x1fffffeL, 6),
234 +1000.f,
235 +100.f,
236 +4.0f,
237 +3.5f,
238 +3.0f,
239 MAKE_HEX_FLOAT(+0x1.800002p1f, +0x1800002L, -23),
240 2.5f,
241 MAKE_HEX_FLOAT(+0x1.7ffffep1f, +0x17ffffeL, -23),
242 +2.0f,
243 MAKE_HEX_FLOAT(+0x1.800002p0f, +0x1800002L, -24),
244 1.5f,
245 MAKE_HEX_FLOAT(+0x1.7ffffep0f, +0x17ffffeL, -24),
246 MAKE_HEX_FLOAT(+0x1.000002p0f, +0x1000002L, -24),
247 +1.0f,
248 MAKE_HEX_FLOAT(+0x1.fffffep-1f, +0x1fffffeL, -25),
249 MAKE_HEX_FLOAT(+0x1.000002p-1f, +0x1000002L, -25),
250 +0.5f,
251 MAKE_HEX_FLOAT(+0x1.fffffep-2f, +0x1fffffeL, -26),
252 MAKE_HEX_FLOAT(+0x1.000002p-2f, +0x1000002L, -26),
253 +0.25f,
254 MAKE_HEX_FLOAT(+0x1.fffffep-3f, +0x1fffffeL, -27),
255 MAKE_HEX_FLOAT(0x1.000002p-126f, 0x1000002L, -150),
256 +FLT_MIN,
257 MAKE_HEX_FLOAT(+0x0.fffffep-126f, +0x0fffffeL, -150),
258 MAKE_HEX_FLOAT(+0x0.000ffep-126f, +0x0000ffeL, -150),
259 MAKE_HEX_FLOAT(+0x0.0000fep-126f, +0x00000feL, -150),
260 MAKE_HEX_FLOAT(+0x0.00000ep-126f, +0x000000eL, -150),
261 MAKE_HEX_FLOAT(+0x0.00000cp-126f, +0x000000cL, -150),
262 MAKE_HEX_FLOAT(+0x0.00000ap-126f, +0x000000aL, -150),
263 MAKE_HEX_FLOAT(+0x0.000008p-126f, +0x0000008L, -150),
264 MAKE_HEX_FLOAT(+0x0.000006p-126f, +0x0000006L, -150),
265 MAKE_HEX_FLOAT(+0x0.000004p-126f, +0x0000004L, -150),
266 MAKE_HEX_FLOAT(+0x0.000002p-126f, +0x0000002L, -150),
267 +0.0f,
268 };
269
270 static const size_t specialValuesCount =
271 sizeof(specialValues) / sizeof(specialValues[0]);
272
273 static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data);
274
TestFunc_Float_Float_Float(const Func * f,MTdata d,bool relaxedMode)275 int TestFunc_Float_Float_Float(const Func *f, MTdata d, bool relaxedMode)
276 {
277 TestInfo test_info;
278 cl_int error;
279 float maxError = 0.0f;
280 double maxErrorVal = 0.0;
281 double maxErrorVal2 = 0.0;
282
283 logFunctionInfo(f->name, sizeof(cl_float), relaxedMode);
284
285 // Init test_info
286 memset(&test_info, 0, sizeof(test_info));
287 test_info.threadCount = GetThreadCount();
288 test_info.subBufferSize = BUFFER_SIZE
289 / (sizeof(cl_float) * RoundUpToNextPowerOfTwo(test_info.threadCount));
290 test_info.scale = getTestScale(sizeof(cl_float));
291
292 test_info.step = (cl_uint)test_info.subBufferSize * test_info.scale;
293 if (test_info.step / test_info.subBufferSize != test_info.scale)
294 {
295 // there was overflow
296 test_info.jobCount = 1;
297 }
298 else
299 {
300 test_info.jobCount = (cl_uint)((1ULL << 32) / test_info.step);
301 }
302
303 test_info.f = f;
304 test_info.ulps = gIsEmbedded ? f->float_embedded_ulps : f->float_ulps;
305 test_info.ftz =
306 f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
307 test_info.relaxedMode = relaxedMode;
308 test_info.isFDim = 0 == strcmp("fdim", f->nameInCode);
309 test_info.skipNanInf = test_info.isFDim && !gInfNanSupport;
310 test_info.isNextafter = 0 == strcmp("nextafter", f->nameInCode);
311
312 // cl_kernels aren't thread safe, so we make one for each vector size for
313 // every thread
314 for (auto i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++)
315 {
316 size_t array_size = test_info.threadCount * sizeof(cl_kernel);
317 test_info.k[i] = (cl_kernel *)malloc(array_size);
318 if (NULL == test_info.k[i])
319 {
320 vlog_error("Error: Unable to allocate storage for kernels!\n");
321 error = CL_OUT_OF_HOST_MEMORY;
322 goto exit;
323 }
324 memset(test_info.k[i], 0, array_size);
325 }
326 test_info.tinfo =
327 (ThreadInfo *)malloc(test_info.threadCount * sizeof(*test_info.tinfo));
328 if (NULL == test_info.tinfo)
329 {
330 vlog_error(
331 "Error: Unable to allocate storage for thread specific data.\n");
332 error = CL_OUT_OF_HOST_MEMORY;
333 goto exit;
334 }
335 memset(test_info.tinfo, 0,
336 test_info.threadCount * sizeof(*test_info.tinfo));
337 for (cl_uint i = 0; i < test_info.threadCount; i++)
338 {
339 cl_buffer_region region = {
340 i * test_info.subBufferSize * sizeof(cl_float),
341 test_info.subBufferSize * sizeof(cl_float)
342 };
343 test_info.tinfo[i].inBuf =
344 clCreateSubBuffer(gInBuffer, CL_MEM_READ_ONLY,
345 CL_BUFFER_CREATE_TYPE_REGION, ®ion, &error);
346 if (error || NULL == test_info.tinfo[i].inBuf)
347 {
348 vlog_error("Error: Unable to create sub-buffer of gInBuffer for "
349 "region {%zd, %zd}\n",
350 region.origin, region.size);
351 goto exit;
352 }
353 test_info.tinfo[i].inBuf2 =
354 clCreateSubBuffer(gInBuffer2, CL_MEM_READ_ONLY,
355 CL_BUFFER_CREATE_TYPE_REGION, ®ion, &error);
356 if (error || NULL == test_info.tinfo[i].inBuf2)
357 {
358 vlog_error("Error: Unable to create sub-buffer of gInBuffer2 for "
359 "region {%zd, %zd}\n",
360 region.origin, region.size);
361 goto exit;
362 }
363
364 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
365 {
366 test_info.tinfo[i].outBuf[j] = clCreateSubBuffer(
367 gOutBuffer[j], CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION,
368 ®ion, &error);
369 if (error || NULL == test_info.tinfo[i].outBuf[j])
370 {
371 vlog_error("Error: Unable to create sub-buffer of "
372 "gOutBuffer[%d] for region {%zd, %zd}\n",
373 (int)j, region.origin, region.size);
374 goto exit;
375 }
376 }
377 test_info.tinfo[i].tQueue =
378 clCreateCommandQueue(gContext, gDevice, 0, &error);
379 if (NULL == test_info.tinfo[i].tQueue || error)
380 {
381 vlog_error("clCreateCommandQueue failed. (%d)\n", error);
382 goto exit;
383 }
384
385 test_info.tinfo[i].d = init_genrand(genrand_int32(d));
386 }
387
388 // Init the kernels
389 {
390 BuildKernelInfo build_info = {
391 gMinVectorSizeIndex, test_info.threadCount, test_info.k,
392 test_info.programs, f->nameInCode, relaxedMode
393 };
394 if ((error = ThreadPool_Do(BuildKernelFn,
395 gMaxVectorSizeIndex - gMinVectorSizeIndex,
396 &build_info)))
397 goto exit;
398 }
399
400 // Run the kernels
401 if (!gSkipCorrectnessTesting)
402 {
403 error = ThreadPool_Do(Test, test_info.jobCount, &test_info);
404
405 // Accumulate the arithmetic errors
406 for (cl_uint i = 0; i < test_info.threadCount; i++)
407 {
408 if (test_info.tinfo[i].maxError > maxError)
409 {
410 maxError = test_info.tinfo[i].maxError;
411 maxErrorVal = test_info.tinfo[i].maxErrorValue;
412 maxErrorVal2 = test_info.tinfo[i].maxErrorValue2;
413 }
414 }
415
416 if (error) goto exit;
417
418 if (gWimpyMode)
419 vlog("Wimp pass");
420 else
421 vlog("passed");
422
423 vlog("\t%8.2f @ {%a, %a}", maxError, maxErrorVal, maxErrorVal2);
424 }
425
426 vlog("\n");
427
428 exit:
429 // Release
430 for (auto i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++)
431 {
432 clReleaseProgram(test_info.programs[i]);
433 if (test_info.k[i])
434 {
435 for (cl_uint j = 0; j < test_info.threadCount; j++)
436 clReleaseKernel(test_info.k[i][j]);
437
438 free(test_info.k[i]);
439 }
440 }
441 if (test_info.tinfo)
442 {
443 for (cl_uint i = 0; i < test_info.threadCount; i++)
444 {
445 free_mtdata(test_info.tinfo[i].d);
446 clReleaseMemObject(test_info.tinfo[i].inBuf);
447 clReleaseMemObject(test_info.tinfo[i].inBuf2);
448 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
449 clReleaseMemObject(test_info.tinfo[i].outBuf[j]);
450 clReleaseCommandQueue(test_info.tinfo[i].tQueue);
451 }
452
453 free(test_info.tinfo);
454 }
455
456 return error;
457 }
458
Test(cl_uint job_id,cl_uint thread_id,void * data)459 static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
460 {
461 const TestInfo *job = (const TestInfo *)data;
462 size_t buffer_elements = job->subBufferSize;
463 size_t buffer_size = buffer_elements * sizeof(cl_float);
464 cl_uint base = job_id * (cl_uint)job->step;
465 ThreadInfo *tinfo = job->tinfo + thread_id;
466 fptr func = job->f->func;
467 int ftz = job->ftz;
468 bool relaxedMode = job->relaxedMode;
469 float ulps = getAllowedUlpError(job->f, relaxedMode);
470 MTdata d = tinfo->d;
471 cl_int error;
472 cl_uchar *overflow = (cl_uchar *)malloc(buffer_size);
473 const char *name = job->f->name;
474 int isFDim = job->isFDim;
475 int skipNanInf = job->skipNanInf;
476 int isNextafter = job->isNextafter;
477 cl_uint *t = 0;
478 cl_float *r = 0;
479 cl_float *s = 0;
480 cl_float *s2 = 0;
481 cl_int copysign_test = 0;
482 RoundingMode oldRoundMode;
483 int skipVerification = 0;
484
485 if (relaxedMode)
486 {
487 func = job->f->rfunc;
488 if (strcmp(name, "pow") == 0 && gFastRelaxedDerived)
489 {
490 ulps = INFINITY;
491 skipVerification = 1;
492 }
493 }
494
495 // start the map of the output arrays
496 cl_event e[VECTOR_SIZE_COUNT];
497 cl_uint *out[VECTOR_SIZE_COUNT];
498 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
499 {
500 out[j] = (cl_uint *)clEnqueueMapBuffer(
501 tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_WRITE, 0,
502 buffer_size, 0, NULL, e + j, &error);
503 if (error || NULL == out[j])
504 {
505 vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j,
506 error);
507 return error;
508 }
509 }
510
511 // Get that moving
512 if ((error = clFlush(tinfo->tQueue))) vlog("clFlush failed\n");
513
514 // Init input array
515 cl_uint *p = (cl_uint *)gIn + thread_id * buffer_elements;
516 cl_uint *p2 = (cl_uint *)gIn2 + thread_id * buffer_elements;
517 cl_uint idx = 0;
518 int totalSpecialValueCount = specialValuesCount * specialValuesCount;
519 int lastSpecialJobIndex = (totalSpecialValueCount - 1) / buffer_elements;
520
521 if (job_id <= (cl_uint)lastSpecialJobIndex)
522 { // test edge cases
523 float *fp = (float *)p;
524 float *fp2 = (float *)p2;
525 uint32_t x, y;
526
527 x = (job_id * buffer_elements) % specialValuesCount;
528 y = (job_id * buffer_elements) / specialValuesCount;
529
530 for (; idx < buffer_elements; idx++)
531 {
532 fp[idx] = specialValues[x];
533 fp2[idx] = specialValues[y];
534 ++x;
535 if (x >= specialValuesCount)
536 {
537 x = 0;
538 y++;
539 if (y >= specialValuesCount) break;
540 }
541 }
542 }
543
544 // Init any remaining values.
545 for (; idx < buffer_elements; idx++)
546 {
547 p[idx] = genrand_int32(d);
548 p2[idx] = genrand_int32(d);
549 }
550
551 if ((error = clEnqueueWriteBuffer(tinfo->tQueue, tinfo->inBuf, CL_FALSE, 0,
552 buffer_size, p, 0, NULL, NULL)))
553 {
554 vlog_error("Error: clEnqueueWriteBuffer failed! err: %d\n", error);
555 goto exit;
556 }
557
558 if ((error = clEnqueueWriteBuffer(tinfo->tQueue, tinfo->inBuf2, CL_FALSE, 0,
559 buffer_size, p2, 0, NULL, NULL)))
560 {
561 vlog_error("Error: clEnqueueWriteBuffer failed! err: %d\n", error);
562 goto exit;
563 }
564
565 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
566 {
567 // Wait for the map to finish
568 if ((error = clWaitForEvents(1, e + j)))
569 {
570 vlog_error("Error: clWaitForEvents failed! err: %d\n", error);
571 goto exit;
572 }
573 if ((error = clReleaseEvent(e[j])))
574 {
575 vlog_error("Error: clReleaseEvent failed! err: %d\n", error);
576 goto exit;
577 }
578
579 // Fill the result buffer with garbage, so that old results don't carry
580 // over
581 uint32_t pattern = 0xffffdead;
582 memset_pattern4(out[j], &pattern, buffer_size);
583 if ((error = clEnqueueUnmapMemObject(tinfo->tQueue, tinfo->outBuf[j],
584 out[j], 0, NULL, NULL)))
585 {
586 vlog_error("Error: clEnqueueMapBuffer failed! err: %d\n", error);
587 goto exit;
588 }
589
590 // run the kernel
591 size_t vectorCount =
592 (buffer_elements + sizeValues[j] - 1) / sizeValues[j];
593 cl_kernel kernel = job->k[j][thread_id]; // each worker thread has its
594 // own copy of the cl_kernel
595 cl_program program = job->programs[j];
596
597 if ((error = clSetKernelArg(kernel, 0, sizeof(tinfo->outBuf[j]),
598 &tinfo->outBuf[j])))
599 {
600 LogBuildError(program);
601 return error;
602 }
603 if ((error = clSetKernelArg(kernel, 1, sizeof(tinfo->inBuf),
604 &tinfo->inBuf)))
605 {
606 LogBuildError(program);
607 return error;
608 }
609 if ((error = clSetKernelArg(kernel, 2, sizeof(tinfo->inBuf2),
610 &tinfo->inBuf2)))
611 {
612 LogBuildError(program);
613 return error;
614 }
615
616 if ((error = clEnqueueNDRangeKernel(tinfo->tQueue, kernel, 1, NULL,
617 &vectorCount, NULL, 0, NULL, NULL)))
618 {
619 vlog_error("FAILED -- could not execute kernel\n");
620 goto exit;
621 }
622 }
623
624 // Get that moving
625 if ((error = clFlush(tinfo->tQueue))) vlog("clFlush 2 failed\n");
626
627 if (gSkipCorrectnessTesting)
628 {
629 if ((error = clFinish(tinfo->tQueue)))
630 {
631 vlog_error("Error: clFinish failed! err: %d\n", error);
632 goto exit;
633 }
634 free(overflow);
635 return CL_SUCCESS;
636 }
637
638 FPU_mode_type oldMode;
639 oldRoundMode = kRoundToNearestEven;
640 if (isFDim)
641 {
642 // Calculate the correctly rounded reference result
643 memset(&oldMode, 0, sizeof(oldMode));
644 if (ftz) ForceFTZ(&oldMode);
645
646 // Set the rounding mode to match the device
647 if (gIsInRTZMode) oldRoundMode = set_round(kRoundTowardZero, kfloat);
648 }
649
650 if (!strcmp(name, "copysign")) copysign_test = 1;
651
652 #define ref_func(s, s2) (copysign_test ? func.f_ff_f(s, s2) : func.f_ff(s, s2))
653
654 // Calculate the correctly rounded reference result
655 r = (float *)gOut_Ref + thread_id * buffer_elements;
656 s = (float *)gIn + thread_id * buffer_elements;
657 s2 = (float *)gIn2 + thread_id * buffer_elements;
658 if (skipNanInf)
659 {
660 for (size_t j = 0; j < buffer_elements; j++)
661 {
662 feclearexcept(FE_OVERFLOW);
663 r[j] = (float)ref_func(s[j], s2[j]);
664 overflow[j] =
665 FE_OVERFLOW == (FE_OVERFLOW & fetestexcept(FE_OVERFLOW));
666 }
667 }
668 else
669 {
670 for (size_t j = 0; j < buffer_elements; j++)
671 r[j] = (float)ref_func(s[j], s2[j]);
672 }
673
674 if (isFDim && ftz) RestoreFPState(&oldMode);
675
676 // Read the data back -- no need to wait for the first N-1 buffers but wait
677 // for the last buffer. This is an in order queue.
678 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
679 {
680 cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
681 out[j] = (cl_uint *)clEnqueueMapBuffer(
682 tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
683 buffer_size, 0, NULL, NULL, &error);
684 if (error || NULL == out[j])
685 {
686 vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j,
687 error);
688 goto exit;
689 }
690 }
691
692 if (!skipVerification)
693 {
694 // Verify data
695 t = (cl_uint *)r;
696 for (size_t j = 0; j < buffer_elements; j++)
697 {
698 for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
699 {
700 cl_uint *q = out[k];
701
702 // If we aren't getting the correctly rounded result
703 if (t[j] != q[j])
704 {
705 float test = ((float *)q)[j];
706 double correct = ref_func(s[j], s2[j]);
707
708 // Per section 10 paragraph 6, accept any result if an input
709 // or output is a infinity or NaN or overflow As per
710 // OpenCL 2.0 spec, section 5.8.4.3, enabling
711 // fast-relaxed-math mode also enables -cl-finite-math-only
712 // optimization. This optimization allows to assume that
713 // arguments and results are not NaNs or +/-INFs. Hence,
714 // accept any result if inputs or results are NaNs or INFs.
715 if (relaxedMode || skipNanInf)
716 {
717 if (skipNanInf && overflow[j]) continue;
718 // Note: no double rounding here. Reference functions
719 // calculate in single precision.
720 if (IsFloatInfinity(correct) || IsFloatNaN(correct)
721 || IsFloatInfinity(s2[j]) || IsFloatNaN(s2[j])
722 || IsFloatInfinity(s[j]) || IsFloatNaN(s[j]))
723 continue;
724 }
725
726 float err = Ulp_Error(test, correct);
727 int fail = !(fabsf(err) <= ulps);
728
729 if (fail && ftz)
730 {
731 // retry per section 6.5.3.2
732 if (IsFloatResultSubnormal(correct, ulps))
733 {
734 fail = fail && (test != 0.0f);
735 if (!fail) err = 0.0f;
736 }
737
738 // nextafter on FTZ platforms may return the smallest
739 // normal float (2^-126) given a denormal or a zero
740 // as the first argument. The rationale here is that
741 // nextafter flushes the argument to zero and then
742 // returns the next representable number in the
743 // direction of the second argument, and since
744 // denorms are considered as zero, the smallest
745 // normal number is the next representable number.
746 // In which case, it should have the same sign as the
747 // second argument.
748 if (isNextafter)
749 {
750 if (IsFloatSubnormal(s[j]) || s[j] == 0.0f)
751 {
752 float value = copysignf(twoToMinus126, s2[j]);
753 fail = fail && (test != value);
754 if (!fail) err = 0.0f;
755 }
756 }
757 else
758 {
759 // retry per section 6.5.3.3
760 if (IsFloatSubnormal(s[j]))
761 {
762 double correct2, correct3;
763 float err2, err3;
764
765 if (skipNanInf) feclearexcept(FE_OVERFLOW);
766
767 correct2 = ref_func(0.0, s2[j]);
768 correct3 = ref_func(-0.0, s2[j]);
769
770 // Per section 10 paragraph 6, accept any result
771 // if an input or output is a infinity or NaN or
772 // overflow As per OpenCL 2.0 spec,
773 // section 5.8.4.3, enabling fast-relaxed-math
774 // mode also enables -cl-finite-math-only
775 // optimization. This optimization allows to
776 // assume that arguments and results are not
777 // NaNs or +/-INFs. Hence, accept any result if
778 // inputs or results are NaNs or INFs.
779 if (relaxedMode || skipNanInf)
780 {
781 if (fetestexcept(FE_OVERFLOW) && skipNanInf)
782 continue;
783
784 // Note: no double rounding here. Reference
785 // functions calculate in single precision.
786 if (IsFloatInfinity(correct2)
787 || IsFloatNaN(correct2)
788 || IsFloatInfinity(correct3)
789 || IsFloatNaN(correct3))
790 continue;
791 }
792
793 err2 = Ulp_Error(test, correct2);
794 err3 = Ulp_Error(test, correct3);
795 fail = fail
796 && ((!(fabsf(err2) <= ulps))
797 && (!(fabsf(err3) <= ulps)));
798 if (fabsf(err2) < fabsf(err)) err = err2;
799 if (fabsf(err3) < fabsf(err)) err = err3;
800
801 // retry per section 6.5.3.4
802 if (IsFloatResultSubnormal(correct2, ulps)
803 || IsFloatResultSubnormal(correct3, ulps))
804 {
805 fail = fail && (test != 0.0f);
806 if (!fail) err = 0.0f;
807 }
808
809 // try with both args as zero
810 if (IsFloatSubnormal(s2[j]))
811 {
812 double correct4, correct5;
813 float err4, err5;
814
815 if (skipNanInf) feclearexcept(FE_OVERFLOW);
816
817 correct2 = ref_func(0.0, 0.0);
818 correct3 = ref_func(-0.0, 0.0);
819 correct4 = ref_func(0.0, -0.0);
820 correct5 = ref_func(-0.0, -0.0);
821
822 // Per section 10 paragraph 6, accept any
823 // result if an input or output is a
824 // infinity or NaN or overflow As per
825 // OpenCL 2.0 spec, section 5.8.4.3,
826 // enabling fast-relaxed-math mode also
827 // enables -cl-finite-math-only
828 // optimization. This optimization allows to
829 // assume that arguments and results are not
830 // NaNs or +/-INFs. Hence, accept any result
831 // if inputs or results are NaNs or INFs.
832 if (relaxedMode || skipNanInf)
833 {
834 if (fetestexcept(FE_OVERFLOW)
835 && skipNanInf)
836 continue;
837
838 // Note: no double rounding here.
839 // Reference functions calculate in
840 // single precision.
841 if (IsFloatInfinity(correct2)
842 || IsFloatNaN(correct2)
843 || IsFloatInfinity(correct3)
844 || IsFloatNaN(correct3)
845 || IsFloatInfinity(correct4)
846 || IsFloatNaN(correct4)
847 || IsFloatInfinity(correct5)
848 || IsFloatNaN(correct5))
849 continue;
850 }
851
852 err2 = Ulp_Error(test, correct2);
853 err3 = Ulp_Error(test, correct3);
854 err4 = Ulp_Error(test, correct4);
855 err5 = Ulp_Error(test, correct5);
856 fail = fail
857 && ((!(fabsf(err2) <= ulps))
858 && (!(fabsf(err3) <= ulps))
859 && (!(fabsf(err4) <= ulps))
860 && (!(fabsf(err5) <= ulps)));
861 if (fabsf(err2) < fabsf(err)) err = err2;
862 if (fabsf(err3) < fabsf(err)) err = err3;
863 if (fabsf(err4) < fabsf(err)) err = err4;
864 if (fabsf(err5) < fabsf(err)) err = err5;
865
866 // retry per section 6.5.3.4
867 if (IsFloatResultSubnormal(correct2, ulps)
868 || IsFloatResultSubnormal(correct3,
869 ulps)
870 || IsFloatResultSubnormal(correct4,
871 ulps)
872 || IsFloatResultSubnormal(correct5,
873 ulps))
874 {
875 fail = fail && (test != 0.0f);
876 if (!fail) err = 0.0f;
877 }
878 }
879 }
880 else if (IsFloatSubnormal(s2[j]))
881 {
882 double correct2, correct3;
883 float err2, err3;
884
885 if (skipNanInf) feclearexcept(FE_OVERFLOW);
886
887 correct2 = ref_func(s[j], 0.0);
888 correct3 = ref_func(s[j], -0.0);
889
890 // Per section 10 paragraph 6, accept any result
891 // if an input or output is a infinity or NaN or
892 // overflow As per OpenCL 2.0 spec,
893 // section 5.8.4.3, enabling fast-relaxed-math
894 // mode also enables -cl-finite-math-only
895 // optimization. This optimization allows to
896 // assume that arguments and results are not
897 // NaNs or +/-INFs. Hence, accept any result if
898 // inputs or results are NaNs or INFs.
899 if (relaxedMode || skipNanInf)
900 {
901 // Note: no double rounding here. Reference
902 // functions calculate in single precision.
903 if (overflow[j] && skipNanInf) continue;
904
905 if (IsFloatInfinity(correct2)
906 || IsFloatNaN(correct2)
907 || IsFloatInfinity(correct3)
908 || IsFloatNaN(correct3))
909 continue;
910 }
911
912 err2 = Ulp_Error(test, correct2);
913 err3 = Ulp_Error(test, correct3);
914 fail = fail
915 && ((!(fabsf(err2) <= ulps))
916 && (!(fabsf(err3) <= ulps)));
917 if (fabsf(err2) < fabsf(err)) err = err2;
918 if (fabsf(err3) < fabsf(err)) err = err3;
919
920 // retry per section 6.5.3.4
921 if (IsFloatResultSubnormal(correct2, ulps)
922 || IsFloatResultSubnormal(correct3, ulps))
923 {
924 fail = fail && (test != 0.0f);
925 if (!fail) err = 0.0f;
926 }
927 }
928 }
929 }
930
931 if (fabsf(err) > tinfo->maxError)
932 {
933 tinfo->maxError = fabsf(err);
934 tinfo->maxErrorValue = s[j];
935 tinfo->maxErrorValue2 = s2[j];
936 }
937 if (fail)
938 {
939 vlog_error(
940 "\nERROR: %s%s: %f ulp error at {%a (0x%x), %a "
941 "(0x%x)}: *%a vs. %a (0x%8.8x) at index: %d\n",
942 name, sizeNames[k], err, s[j], ((cl_uint *)s)[j],
943 s2[j], ((cl_uint *)s2)[j], r[j], test,
944 ((cl_uint *)&test)[0], j);
945 error = -1;
946 goto exit;
947 }
948 }
949 }
950 }
951 }
952
953 if (isFDim && gIsInRTZMode) (void)set_round(oldRoundMode, kfloat);
954
955 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
956 {
957 if ((error = clEnqueueUnmapMemObject(tinfo->tQueue, tinfo->outBuf[j],
958 out[j], 0, NULL, NULL)))
959 {
960 vlog_error("Error: clEnqueueUnmapMemObject %d failed 2! err: %d\n",
961 j, error);
962 return error;
963 }
964 }
965
966 if ((error = clFlush(tinfo->tQueue))) vlog("clFlush 3 failed\n");
967
968
969 if (0 == (base & 0x0fffffff))
970 {
971 if (gVerboseBruteForce)
972 {
973 vlog("base:%14u step:%10u scale:%10zu buf_elements:%10u ulps:%5.3f "
974 "ThreadCount:%2u\n",
975 base, job->step, job->scale, buffer_elements, job->ulps,
976 job->threadCount);
977 }
978 else
979 {
980 vlog(".");
981 }
982 fflush(stdout);
983 }
984
985 exit:
986 if (overflow) free(overflow);
987 return error;
988 }
989