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[] = { "__kernel void math_kernel",
31 sizeNames[vectorSize],
32 "( __global float",
33 sizeNames[vectorSize],
34 "* out, __global float",
35 sizeNames[vectorSize],
36 "* out2, __global float",
37 sizeNames[vectorSize],
38 "* in )\n"
39 "{\n"
40 " size_t i = get_global_id(0);\n"
41 " out[i] = ",
42 name,
43 "( in[i], out2 + i );\n"
44 "}\n" };
45
46 const char *c3[] = {
47 "__kernel void math_kernel",
48 sizeNames[vectorSize],
49 "( __global float* out, __global float* out2, __global float* in)\n"
50 "{\n"
51 " size_t i = get_global_id(0);\n"
52 " if( i + 1 < get_global_size(0) )\n"
53 " {\n"
54 " float3 f0 = vload3( 0, in + 3 * i );\n"
55 " float3 iout = NAN;\n"
56 " f0 = ",
57 name,
58 "( f0, &iout );\n"
59 " vstore3( f0, 0, out + 3*i );\n"
60 " vstore3( iout, 0, out2 + 3*i );\n"
61 " }\n"
62 " else\n"
63 " {\n"
64 " size_t parity = i & 1; // Figure out how many elements are "
65 "left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two "
66 "buffer size \n"
67 " float3 iout = NAN;\n"
68 " float3 f0;\n"
69 " switch( parity )\n"
70 " {\n"
71 " case 1:\n"
72 " f0 = (float3)( in[3*i], NAN, NAN ); \n"
73 " break;\n"
74 " case 0:\n"
75 " f0 = (float3)( in[3*i], in[3*i+1], NAN ); \n"
76 " break;\n"
77 " }\n"
78 " f0 = ",
79 name,
80 "( f0, &iout );\n"
81 " switch( parity )\n"
82 " {\n"
83 " case 0:\n"
84 " out[3*i+1] = f0.y; \n"
85 " out2[3*i+1] = iout.y; \n"
86 " // fall through\n"
87 " case 1:\n"
88 " out[3*i] = f0.x; \n"
89 " out2[3*i] = iout.x; \n"
90 " break;\n"
91 " }\n"
92 " }\n"
93 "}\n"
94 };
95
96 const char **kern = c;
97 size_t kernSize = sizeof(c) / sizeof(c[0]);
98
99 if (sizeValues[vectorSize] == 3)
100 {
101 kern = c3;
102 kernSize = sizeof(c3) / sizeof(c3[0]);
103 }
104
105 char testName[32];
106 snprintf(testName, sizeof(testName) - 1, "math_kernel%s",
107 sizeNames[vectorSize]);
108
109 return MakeKernel(kern, (cl_uint)kernSize, testName, k, p, relaxedMode);
110 }
111
112 struct BuildKernelInfo2
113 {
114 cl_kernel *kernels;
115 Programs &programs;
116 const char *nameInCode;
117 bool relaxedMode; // Whether to build with -cl-fast-relaxed-math.
118 };
119
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)120 cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
121 {
122 BuildKernelInfo2 *info = (BuildKernelInfo2 *)p;
123 cl_uint vectorSize = gMinVectorSizeIndex + job_id;
124 return BuildKernel(info->nameInCode, vectorSize, info->kernels + vectorSize,
125 &(info->programs[vectorSize]), info->relaxedMode);
126 }
127
128 } // anonymous namespace
129
TestFunc_Float2_Float(const Func * f,MTdata d,bool relaxedMode)130 int TestFunc_Float2_Float(const Func *f, MTdata d, bool relaxedMode)
131 {
132 int error;
133 Programs programs;
134 cl_kernel kernels[VECTOR_SIZE_COUNT];
135 float maxError0 = 0.0f;
136 float maxError1 = 0.0f;
137 int ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
138 float maxErrorVal0 = 0.0f;
139 float maxErrorVal1 = 0.0f;
140 uint64_t step = getTestStep(sizeof(float), BUFFER_SIZE);
141 int scale = (int)((1ULL << 32) / (16 * BUFFER_SIZE / sizeof(float)) + 1);
142 cl_uchar overflow[BUFFER_SIZE / sizeof(float)];
143 int isFract = 0 == strcmp("fract", f->nameInCode);
144 int skipNanInf = isFract && !gInfNanSupport;
145
146 logFunctionInfo(f->name, sizeof(cl_float), relaxedMode);
147
148 float float_ulps = getAllowedUlpError(f, relaxedMode);
149 // Init the kernels
150 {
151 BuildKernelInfo2 build_info{ kernels, programs, f->nameInCode,
152 relaxedMode };
153 if ((error = ThreadPool_Do(BuildKernelFn,
154 gMaxVectorSizeIndex - gMinVectorSizeIndex,
155 &build_info)))
156 return error;
157 }
158
159 for (uint64_t i = 0; i < (1ULL << 32); i += step)
160 {
161 // Init input array
162 uint32_t *p = (uint32_t *)gIn;
163 if (gWimpyMode)
164 {
165 for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
166 {
167 p[j] = (uint32_t)i + j * scale;
168 if (relaxedMode && strcmp(f->name, "sincos") == 0)
169 {
170 float pj = *(float *)&p[j];
171 if (fabs(pj) > M_PI) ((float *)p)[j] = NAN;
172 }
173 }
174 }
175 else
176 {
177 for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
178 {
179 p[j] = (uint32_t)i + j;
180 if (relaxedMode && strcmp(f->name, "sincos") == 0)
181 {
182 float pj = *(float *)&p[j];
183 if (fabs(pj) > M_PI) ((float *)p)[j] = NAN;
184 }
185 }
186 }
187
188 if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0,
189 BUFFER_SIZE, gIn, 0, NULL, NULL)))
190 {
191 vlog_error("\n*** Error %d in clEnqueueWriteBuffer ***\n", error);
192 return error;
193 }
194
195 // write garbage into output arrays
196 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
197 {
198 uint32_t pattern = 0xffffdead;
199 memset_pattern4(gOut[j], &pattern, BUFFER_SIZE);
200 if ((error =
201 clEnqueueWriteBuffer(gQueue, gOutBuffer[j], CL_FALSE, 0,
202 BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
203 {
204 vlog_error("\n*** Error %d in clEnqueueWriteBuffer2(%d) ***\n",
205 error, j);
206 goto exit;
207 }
208
209 memset_pattern4(gOut2[j], &pattern, BUFFER_SIZE);
210 if ((error = clEnqueueWriteBuffer(gQueue, gOutBuffer2[j], CL_FALSE,
211 0, BUFFER_SIZE, gOut2[j], 0, NULL,
212 NULL)))
213 {
214 vlog_error("\n*** Error %d in clEnqueueWriteBuffer2b(%d) ***\n",
215 error, j);
216 goto exit;
217 }
218 }
219
220 // Run the kernels
221 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
222 {
223 size_t vectorSize = sizeValues[j] * sizeof(cl_float);
224 size_t localCount = (BUFFER_SIZE + vectorSize - 1) / vectorSize;
225 if ((error = clSetKernelArg(kernels[j], 0, sizeof(gOutBuffer[j]),
226 &gOutBuffer[j])))
227 {
228 LogBuildError(programs[j]);
229 goto exit;
230 }
231 if ((error = clSetKernelArg(kernels[j], 1, sizeof(gOutBuffer2[j]),
232 &gOutBuffer2[j])))
233 {
234 LogBuildError(programs[j]);
235 goto exit;
236 }
237 if ((error = clSetKernelArg(kernels[j], 2, sizeof(gInBuffer),
238 &gInBuffer)))
239 {
240 LogBuildError(programs[j]);
241 goto exit;
242 }
243
244 if ((error =
245 clEnqueueNDRangeKernel(gQueue, kernels[j], 1, NULL,
246 &localCount, NULL, 0, NULL, NULL)))
247 {
248 vlog_error("FAILED -- could not execute kernel\n");
249 goto exit;
250 }
251 }
252
253 // Get that moving
254 if ((error = clFlush(gQueue))) vlog("clFlush failed\n");
255
256 FPU_mode_type oldMode;
257 RoundingMode oldRoundMode = kRoundToNearestEven;
258 if (isFract)
259 {
260 // Calculate the correctly rounded reference result
261 memset(&oldMode, 0, sizeof(oldMode));
262 if (ftz || relaxedMode) ForceFTZ(&oldMode);
263
264 // Set the rounding mode to match the device
265 if (gIsInRTZMode)
266 oldRoundMode = set_round(kRoundTowardZero, kfloat);
267 }
268
269 // Calculate the correctly rounded reference result
270 float *r = (float *)gOut_Ref;
271 float *r2 = (float *)gOut_Ref2;
272 float *s = (float *)gIn;
273
274 if (skipNanInf)
275 {
276 for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
277 {
278 double dd;
279 feclearexcept(FE_OVERFLOW);
280
281 if (relaxedMode)
282 r[j] = (float)f->rfunc.f_fpf(s[j], &dd);
283 else
284 r[j] = (float)f->func.f_fpf(s[j], &dd);
285
286 r2[j] = (float)dd;
287 overflow[j] =
288 FE_OVERFLOW == (FE_OVERFLOW & fetestexcept(FE_OVERFLOW));
289 }
290 }
291 else
292 {
293 for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
294 {
295 double dd;
296 if (relaxedMode)
297 r[j] = (float)f->rfunc.f_fpf(s[j], &dd);
298 else
299 r[j] = (float)f->func.f_fpf(s[j], &dd);
300
301 r2[j] = (float)dd;
302 }
303 }
304
305 if (isFract && ftz) RestoreFPState(&oldMode);
306
307 // Read the data back
308 for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
309 {
310 if ((error =
311 clEnqueueReadBuffer(gQueue, gOutBuffer[j], CL_TRUE, 0,
312 BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
313 {
314 vlog_error("ReadArray failed %d\n", error);
315 goto exit;
316 }
317 if ((error =
318 clEnqueueReadBuffer(gQueue, gOutBuffer2[j], CL_TRUE, 0,
319 BUFFER_SIZE, gOut2[j], 0, NULL, NULL)))
320 {
321 vlog_error("ReadArray2 failed %d\n", error);
322 goto exit;
323 }
324 }
325
326 if (gSkipCorrectnessTesting)
327 {
328 if (isFract && gIsInRTZMode) (void)set_round(oldRoundMode, kfloat);
329 break;
330 }
331
332 // Verify data
333 uint32_t *t = (uint32_t *)gOut_Ref;
334 uint32_t *t2 = (uint32_t *)gOut_Ref2;
335 for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
336 {
337 for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
338 {
339 uint32_t *q = (uint32_t *)gOut[k];
340 uint32_t *q2 = (uint32_t *)gOut2[k];
341
342 // If we aren't getting the correctly rounded result
343 if (t[j] != q[j] || t2[j] != q2[j])
344 {
345 double correct, correct2;
346 float err, err2;
347 float test = ((float *)q)[j];
348 float test2 = ((float *)q2)[j];
349
350 if (relaxedMode)
351 correct = f->rfunc.f_fpf(s[j], &correct2);
352 else
353 correct = f->func.f_fpf(s[j], &correct2);
354
355 // Per section 10 paragraph 6, accept any result if an input
356 // or output is a infinity or NaN or overflow
357 if (relaxedMode || skipNanInf)
358 {
359 if (skipNanInf && overflow[j]) continue;
360 // Note: no double rounding here. Reference functions
361 // calculate in single precision.
362 if (IsFloatInfinity(correct) || IsFloatNaN(correct)
363 || IsFloatInfinity(correct2) || IsFloatNaN(correct2)
364 || IsFloatInfinity(s[j]) || IsFloatNaN(s[j]))
365 continue;
366 }
367
368 typedef int (*CheckForSubnormal)(
369 double, float); // If we are in fast relaxed math, we
370 // have a different calculation for the
371 // subnormal threshold.
372 CheckForSubnormal isFloatResultSubnormalPtr;
373 if (relaxedMode)
374 {
375 err = Abs_Error(test, correct);
376 err2 = Abs_Error(test2, correct2);
377 isFloatResultSubnormalPtr =
378 &IsFloatResultSubnormalAbsError;
379 }
380 else
381 {
382 err = Ulp_Error(test, correct);
383 err2 = Ulp_Error(test2, correct2);
384 isFloatResultSubnormalPtr = &IsFloatResultSubnormal;
385 }
386 int fail = !(fabsf(err) <= float_ulps
387 && fabsf(err2) <= float_ulps);
388
389 if (ftz || relaxedMode)
390 {
391 // retry per section 6.5.3.2
392 if ((*isFloatResultSubnormalPtr)(correct, float_ulps))
393 {
394 if ((*isFloatResultSubnormalPtr)(correct2,
395 float_ulps))
396 {
397 fail = fail && !(test == 0.0f && test2 == 0.0f);
398 if (!fail)
399 {
400 err = 0.0f;
401 err2 = 0.0f;
402 }
403 }
404 else
405 {
406 fail = fail
407 && !(test == 0.0f
408 && fabsf(err2) <= float_ulps);
409 if (!fail) err = 0.0f;
410 }
411 }
412 else if ((*isFloatResultSubnormalPtr)(correct2,
413 float_ulps))
414 {
415 fail = fail
416 && !(test2 == 0.0f && fabsf(err) <= float_ulps);
417 if (!fail) err2 = 0.0f;
418 }
419
420
421 // retry per section 6.5.3.3
422 if (IsFloatSubnormal(s[j]))
423 {
424 double correctp, correctn;
425 double correct2p, correct2n;
426 float errp, err2p, errn, err2n;
427
428 if (skipNanInf) feclearexcept(FE_OVERFLOW);
429 if (relaxedMode)
430 {
431 correctp = f->rfunc.f_fpf(0.0, &correct2p);
432 correctn = f->rfunc.f_fpf(-0.0, &correct2n);
433 }
434 else
435 {
436 correctp = f->func.f_fpf(0.0, &correct2p);
437 correctn = f->func.f_fpf(-0.0, &correct2n);
438 }
439
440 // Per section 10 paragraph 6, accept any result if
441 // an input or output is a infinity or NaN or
442 // overflow
443 if (skipNanInf)
444 {
445 if (fetestexcept(FE_OVERFLOW)) continue;
446
447 // Note: no double rounding here. Reference
448 // functions calculate in single precision.
449 if (IsFloatInfinity(correctp)
450 || IsFloatNaN(correctp)
451 || IsFloatInfinity(correctn)
452 || IsFloatNaN(correctn)
453 || IsFloatInfinity(correct2p)
454 || IsFloatNaN(correct2p)
455 || IsFloatInfinity(correct2n)
456 || IsFloatNaN(correct2n))
457 continue;
458 }
459
460 if (relaxedMode)
461 {
462 errp = Abs_Error(test, correctp);
463 err2p = Abs_Error(test, correct2p);
464 errn = Abs_Error(test, correctn);
465 err2n = Abs_Error(test, correct2n);
466 }
467 else
468 {
469 errp = Ulp_Error(test, correctp);
470 err2p = Ulp_Error(test, correct2p);
471 errn = Ulp_Error(test, correctn);
472 err2n = Ulp_Error(test, correct2n);
473 }
474
475 fail = fail
476 && ((!(fabsf(errp) <= float_ulps))
477 && (!(fabsf(err2p) <= float_ulps))
478 && ((!(fabsf(errn) <= float_ulps))
479 && (!(fabsf(err2n) <= float_ulps))));
480 if (fabsf(errp) < fabsf(err)) err = errp;
481 if (fabsf(errn) < fabsf(err)) err = errn;
482 if (fabsf(err2p) < fabsf(err2)) err2 = err2p;
483 if (fabsf(err2n) < fabsf(err2)) err2 = err2n;
484
485 // retry per section 6.5.3.4
486 if ((*isFloatResultSubnormalPtr)(correctp,
487 float_ulps)
488 || (*isFloatResultSubnormalPtr)(correctn,
489 float_ulps))
490 {
491 if ((*isFloatResultSubnormalPtr)(correct2p,
492 float_ulps)
493 || (*isFloatResultSubnormalPtr)(correct2n,
494 float_ulps))
495 {
496 fail = fail
497 && !(test == 0.0f && test2 == 0.0f);
498 if (!fail) err = err2 = 0.0f;
499 }
500 else
501 {
502 fail = fail
503 && !(test == 0.0f
504 && fabsf(err2) <= float_ulps);
505 if (!fail) err = 0.0f;
506 }
507 }
508 else if ((*isFloatResultSubnormalPtr)(correct2p,
509 float_ulps)
510 || (*isFloatResultSubnormalPtr)(
511 correct2n, float_ulps))
512 {
513 fail = fail
514 && !(test2 == 0.0f
515 && (fabsf(err) <= float_ulps));
516 if (!fail) err2 = 0.0f;
517 }
518 }
519 }
520 if (fabsf(err) > maxError0)
521 {
522 maxError0 = fabsf(err);
523 maxErrorVal0 = s[j];
524 }
525 if (fabsf(err2) > maxError1)
526 {
527 maxError1 = fabsf(err2);
528 maxErrorVal1 = s[j];
529 }
530 if (fail)
531 {
532 vlog_error("\nERROR: %s%s: {%f, %f} ulp error at %a: "
533 "*{%a, %a} vs. {%a, %a}\n",
534 f->name, sizeNames[k], err, err2,
535 ((float *)gIn)[j], ((float *)gOut_Ref)[j],
536 ((float *)gOut_Ref2)[j], test, test2);
537 error = -1;
538 goto exit;
539 }
540 }
541 }
542 }
543
544 if (isFract && gIsInRTZMode) (void)set_round(oldRoundMode, kfloat);
545
546 if (0 == (i & 0x0fffffff))
547 {
548 if (gVerboseBruteForce)
549 {
550 vlog("base:%14" PRIu64 " step:%10" PRIu64
551 " bufferSize:%10d \n",
552 i, step, BUFFER_SIZE);
553 }
554 else
555 {
556 vlog(".");
557 }
558 fflush(stdout);
559 }
560 }
561
562 if (!gSkipCorrectnessTesting)
563 {
564 if (gWimpyMode)
565 vlog("Wimp pass");
566 else
567 vlog("passed");
568
569 vlog("\t{%8.2f, %8.2f} @ {%a, %a}", maxError0, maxError1, maxErrorVal0,
570 maxErrorVal1);
571 }
572
573 vlog("\n");
574
575 exit:
576 // Release
577 for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
578 {
579 clReleaseKernel(kernels[k]);
580 }
581
582 return error;
583 }
584