1 //
2 // Copyright (c) 2017 The Khronos Group Inc.
3 //
4 // Licensed under the Apache License, Version 2.0 (the "License");
5 // you may not use this file except in compliance with the License.
6 // You may obtain a copy of the License at
7 //
8 // http://www.apache.org/licenses/LICENSE-2.0
9 //
10 // Unless required by applicable law or agreed to in writing, software
11 // distributed under the License is distributed on an "AS IS" BASIS,
12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 // See the License for the specific language governing permissions and
14 // limitations under the License.
15 //
16
17 #include "common.h"
18 #include "function_list.h"
19 #include "test_functions.h"
20 #include "utility.h"
21
22 #include <cinttypes>
23 #include <cstring>
24
25 namespace {
26
BuildKernel(const char * name,int vectorSize,cl_kernel * k,cl_program * p,bool relaxedMode)27 int BuildKernel(const char *name, int vectorSize, cl_kernel *k, cl_program *p,
28 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 "* out2, __global double",
38 sizeNames[vectorSize],
39 "* in )\n"
40 "{\n"
41 " size_t i = get_global_id(0);\n"
42 " out[i] = ",
43 name,
44 "( in[i], out2 + 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* out2, __global double* in)\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 f0 = vload3( 0, in + 3 * i );\n"
57 " double3 iout = NAN;\n"
58 " f0 = ",
59 name,
60 "( f0, &iout );\n"
61 " vstore3( f0, 0, out + 3*i );\n"
62 " vstore3( iout, 0, out2 + 3*i );\n"
63 " }\n"
64 " else\n"
65 " {\n"
66 " size_t parity = i & 1; // Figure out how many elements are "
67 "left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two "
68 "buffer size \n"
69 " double3 iout = NAN;\n"
70 " double3 f0;\n"
71 " switch( parity )\n"
72 " {\n"
73 " case 1:\n"
74 " f0 = (double3)( in[3*i], NAN, NAN ); \n"
75 " break;\n"
76 " case 0:\n"
77 " f0 = (double3)( in[3*i], in[3*i+1], NAN ); \n"
78 " break;\n"
79 " }\n"
80 " f0 = ",
81 name,
82 "( f0, &iout );\n"
83 " switch( parity )\n"
84 " {\n"
85 " case 0:\n"
86 " out[3*i+1] = f0.y; \n"
87 " out2[3*i+1] = iout.y; \n"
88 " // fall through\n"
89 " case 1:\n"
90 " out[3*i] = f0.x; \n"
91 " out2[3*i] = iout.x; \n"
92 " break;\n"
93 " }\n"
94 " }\n"
95 "}\n"
96 };
97
98 const char **kern = c;
99 size_t kernSize = sizeof(c) / sizeof(c[0]);
100
101 if (sizeValues[vectorSize] == 3)
102 {
103 kern = c3;
104 kernSize = sizeof(c3) / sizeof(c3[0]);
105 }
106
107 char testName[32];
108 snprintf(testName, sizeof(testName) - 1, "math_kernel%s",
109 sizeNames[vectorSize]);
110
111 return MakeKernel(kern, (cl_uint)kernSize, testName, k, p, relaxedMode);
112 }
113
114 struct BuildKernelInfo2
115 {
116 cl_kernel *kernels;
117 Programs &programs;
118 const char *nameInCode;
119 bool relaxedMode; // Whether to build with -cl-fast-relaxed-math.
120 };
121
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)122 cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
123 {
124 BuildKernelInfo2 *info = (BuildKernelInfo2 *)p;
125 cl_uint vectorSize = gMinVectorSizeIndex + job_id;
126 return BuildKernel(info->nameInCode, vectorSize, info->kernels + vectorSize,
127 &(info->programs[vectorSize]), info->relaxedMode);
128 }
129
130 } // anonymous namespace
131
TestFunc_Double2_Double(const Func * f,MTdata d,bool relaxedMode)132 int TestFunc_Double2_Double(const Func *f, MTdata d, bool relaxedMode)
133 {
134 int error;
135 Programs programs;
136 cl_kernel kernels[VECTOR_SIZE_COUNT];
137 float maxError0 = 0.0f;
138 float maxError1 = 0.0f;
139 int ftz = f->ftz || gForceFTZ;
140 double maxErrorVal0 = 0.0f;
141 double maxErrorVal1 = 0.0f;
142 uint64_t step = getTestStep(sizeof(cl_double), BUFFER_SIZE);
143 int scale =
144 (int)((1ULL << 32) / (16 * BUFFER_SIZE / sizeof(cl_double)) + 1);
145
146 logFunctionInfo(f->name, sizeof(cl_double), relaxedMode);
147
148 Force64BitFPUPrecision();
149
150 // Init the kernels
151 {
152 BuildKernelInfo2 build_info{ kernels, programs, f->nameInCode,
153 relaxedMode };
154 if ((error = ThreadPool_Do(BuildKernelFn,
155 gMaxVectorSizeIndex - gMinVectorSizeIndex,
156 &build_info)))
157 return error;
158 }
159
160 for (uint64_t i = 0; i < (1ULL << 32); i += step)
161 {
162 // Init input array
163 double *p = (double *)gIn;
164 if (gWimpyMode)
165 {
166 for (size_t j = 0; j < BUFFER_SIZE / sizeof(cl_double); j++)
167 p[j] = DoubleFromUInt32((uint32_t)i + j * scale);
168 }
169 else
170 {
171 for (size_t j = 0; j < BUFFER_SIZE / sizeof(cl_double); j++)
172 p[j] = DoubleFromUInt32((uint32_t)i + j);
173 }
174 if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0,
175 BUFFER_SIZE, gIn, 0, NULL, NULL)))
176 {
177 vlog_error("\n*** Error %d in clEnqueueWriteBuffer ***\n", error);
178 return error;
179 }
180
181 // write garbage into output arrays
182 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
183 {
184 uint32_t pattern = 0xffffdead;
185 memset_pattern4(gOut[j], &pattern, BUFFER_SIZE);
186 if ((error =
187 clEnqueueWriteBuffer(gQueue, gOutBuffer[j], CL_FALSE, 0,
188 BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
189 {
190 vlog_error("\n*** Error %d in clEnqueueWriteBuffer2(%d) ***\n",
191 error, j);
192 goto exit;
193 }
194
195 memset_pattern4(gOut2[j], &pattern, BUFFER_SIZE);
196 if ((error = clEnqueueWriteBuffer(gQueue, gOutBuffer2[j], CL_FALSE,
197 0, BUFFER_SIZE, gOut2[j], 0, NULL,
198 NULL)))
199 {
200 vlog_error("\n*** Error %d in clEnqueueWriteBuffer2b(%d) ***\n",
201 error, j);
202 goto exit;
203 }
204 }
205
206 // Run the kernels
207 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
208 {
209 size_t vectorSize = sizeValues[j] * sizeof(cl_double);
210 size_t localCount = (BUFFER_SIZE + vectorSize - 1) / vectorSize;
211 if ((error = clSetKernelArg(kernels[j], 0, sizeof(gOutBuffer[j]),
212 &gOutBuffer[j])))
213 {
214 LogBuildError(programs[j]);
215 goto exit;
216 }
217 if ((error = clSetKernelArg(kernels[j], 1, sizeof(gOutBuffer2[j]),
218 &gOutBuffer2[j])))
219 {
220 LogBuildError(programs[j]);
221 goto exit;
222 }
223 if ((error = clSetKernelArg(kernels[j], 2, sizeof(gInBuffer),
224 &gInBuffer)))
225 {
226 LogBuildError(programs[j]);
227 goto exit;
228 }
229
230 if ((error =
231 clEnqueueNDRangeKernel(gQueue, kernels[j], 1, NULL,
232 &localCount, NULL, 0, NULL, NULL)))
233 {
234 vlog_error("FAILED -- could not execute kernel\n");
235 goto exit;
236 }
237 }
238
239 // Get that moving
240 if ((error = clFlush(gQueue))) vlog("clFlush failed\n");
241
242 // Calculate the correctly rounded reference result
243 double *r = (double *)gOut_Ref;
244 double *r2 = (double *)gOut_Ref2;
245 double *s = (double *)gIn;
246 for (size_t j = 0; j < BUFFER_SIZE / sizeof(cl_double); j++)
247 {
248 long double dd;
249 r[j] = (double)f->dfunc.f_fpf(s[j], &dd);
250 r2[j] = (double)dd;
251 }
252
253 // Read the data back
254 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
255 {
256 if ((error =
257 clEnqueueReadBuffer(gQueue, gOutBuffer[j], CL_TRUE, 0,
258 BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
259 {
260 vlog_error("ReadArray failed %d\n", error);
261 goto exit;
262 }
263 if ((error =
264 clEnqueueReadBuffer(gQueue, gOutBuffer2[j], CL_TRUE, 0,
265 BUFFER_SIZE, gOut2[j], 0, NULL, NULL)))
266 {
267 vlog_error("ReadArray2 failed %d\n", error);
268 goto exit;
269 }
270 }
271
272 if (gSkipCorrectnessTesting) break;
273
274 // Verify data
275 uint64_t *t = (uint64_t *)gOut_Ref;
276 uint64_t *t2 = (uint64_t *)gOut_Ref2;
277 for (size_t j = 0; j < BUFFER_SIZE / sizeof(double); j++)
278 {
279 for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
280 {
281 uint64_t *q = (uint64_t *)(gOut[k]);
282 uint64_t *q2 = (uint64_t *)(gOut2[k]);
283
284 // If we aren't getting the correctly rounded result
285 if (t[j] != q[j] || t2[j] != q2[j])
286 {
287 double test = ((double *)q)[j];
288 double test2 = ((double *)q2)[j];
289 long double correct2;
290 long double correct = f->dfunc.f_fpf(s[j], &correct2);
291 float err = Bruteforce_Ulp_Error_Double(test, correct);
292 float err2 = Bruteforce_Ulp_Error_Double(test2, correct2);
293 int fail = !(fabsf(err) <= f->double_ulps
294 && fabsf(err2) <= f->double_ulps);
295 if (ftz || relaxedMode)
296 {
297 // retry per section 6.5.3.2
298 if (IsDoubleResultSubnormal(correct, f->double_ulps))
299 {
300 if (IsDoubleResultSubnormal(correct2,
301 f->double_ulps))
302 {
303 fail = fail && !(test == 0.0f && test2 == 0.0f);
304 if (!fail)
305 {
306 err = 0.0f;
307 err2 = 0.0f;
308 }
309 }
310 else
311 {
312 fail = fail
313 && !(test == 0.0f
314 && fabsf(err2) <= f->double_ulps);
315 if (!fail) err = 0.0f;
316 }
317 }
318 else if (IsDoubleResultSubnormal(correct2,
319 f->double_ulps))
320 {
321 fail = fail
322 && !(test2 == 0.0f
323 && fabsf(err) <= f->double_ulps);
324 if (!fail) err2 = 0.0f;
325 }
326
327 // retry per section 6.5.3.3
328 if (IsDoubleSubnormal(s[j]))
329 {
330 long double correct2p, correct2n;
331 long double correctp =
332 f->dfunc.f_fpf(0.0, &correct2p);
333 long double correctn =
334 f->dfunc.f_fpf(-0.0, &correct2n);
335 float errp =
336 Bruteforce_Ulp_Error_Double(test, correctp);
337 float err2p =
338 Bruteforce_Ulp_Error_Double(test, correct2p);
339 float errn =
340 Bruteforce_Ulp_Error_Double(test, correctn);
341 float err2n =
342 Bruteforce_Ulp_Error_Double(test, correct2n);
343 fail = fail
344 && ((!(fabsf(errp) <= f->double_ulps))
345 && (!(fabsf(err2p) <= f->double_ulps))
346 && ((!(fabsf(errn) <= f->double_ulps))
347 && (!(fabsf(err2n)
348 <= f->double_ulps))));
349 if (fabsf(errp) < fabsf(err)) err = errp;
350 if (fabsf(errn) < fabsf(err)) err = errn;
351 if (fabsf(err2p) < fabsf(err2)) err2 = err2p;
352 if (fabsf(err2n) < fabsf(err2)) err2 = err2n;
353
354 // retry per section 6.5.3.4
355 if (IsDoubleResultSubnormal(correctp,
356 f->double_ulps)
357 || IsDoubleResultSubnormal(correctn,
358 f->double_ulps))
359 {
360 if (IsDoubleResultSubnormal(correct2p,
361 f->double_ulps)
362 || IsDoubleResultSubnormal(correct2n,
363 f->double_ulps))
364 {
365 fail = fail
366 && !(test == 0.0f && test2 == 0.0f);
367 if (!fail) err = err2 = 0.0f;
368 }
369 else
370 {
371 fail = fail
372 && !(test == 0.0f
373 && fabsf(err2) <= f->double_ulps);
374 if (!fail) err = 0.0f;
375 }
376 }
377 else if (IsDoubleResultSubnormal(correct2p,
378 f->double_ulps)
379 || IsDoubleResultSubnormal(correct2n,
380 f->double_ulps))
381 {
382 fail = fail
383 && !(test2 == 0.0f
384 && (fabsf(err) <= f->double_ulps));
385 if (!fail) err2 = 0.0f;
386 }
387 }
388 }
389 if (fabsf(err) > maxError0)
390 {
391 maxError0 = fabsf(err);
392 maxErrorVal0 = s[j];
393 }
394 if (fabsf(err2) > maxError1)
395 {
396 maxError1 = fabsf(err2);
397 maxErrorVal1 = s[j];
398 }
399 if (fail)
400 {
401 vlog_error(
402 "\nERROR: %sD%s: {%f, %f} ulp error at %.13la: "
403 "*{%.13la, %.13la} vs. {%.13la, %.13la}\n",
404 f->name, sizeNames[k], err, err2,
405 ((double *)gIn)[j], ((double *)gOut_Ref)[j],
406 ((double *)gOut_Ref2)[j], test, test2);
407 error = -1;
408 goto exit;
409 }
410 }
411 }
412 }
413
414 if (0 == (i & 0x0fffffff))
415 {
416 if (gVerboseBruteForce)
417 {
418 vlog("base:%14" PRIu64 " step:%10" PRIu64
419 " bufferSize:%10d \n",
420 i, step, BUFFER_SIZE);
421 }
422 else
423 {
424 vlog(".");
425 }
426 fflush(stdout);
427 }
428 }
429
430 if (!gSkipCorrectnessTesting)
431 {
432 if (gWimpyMode)
433 vlog("Wimp pass");
434 else
435 vlog("passed");
436
437 vlog("\t{%8.2f, %8.2f} @ {%a, %a}", maxError0, maxError1, maxErrorVal0,
438 maxErrorVal1);
439 }
440
441 vlog("\n");
442
443 exit:
444 // Release
445 for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
446 {
447 clReleaseKernel(kernels[k]);
448 }
449
450 return error;
451 }
452