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 #include "Utility.h"
17
18 #include <string.h>
19 #include "FunctionList.h"
20
21 int TestFunc_mad(const Func *f, MTdata);
22 int TestFunc_mad_Double(const Func *f, MTdata);
23
24 extern const vtbl _mad_tbl = { "ternary", TestFunc_mad, TestFunc_mad_Double };
25
26 static int BuildKernel( const char *name, int vectorSize, cl_kernel *k, cl_program *p );
27 static int BuildKernelDouble( const char *name, int vectorSize, cl_kernel *k, cl_program *p );
28
BuildKernel(const char * name,int vectorSize,cl_kernel * k,cl_program * p)29 static int BuildKernel( const char *name, int vectorSize, cl_kernel *k, cl_program *p )
30 {
31 const char *c[] = {
32 "__kernel void math_kernel", sizeNames[vectorSize], "( __global float", sizeNames[vectorSize], "* out, __global float", sizeNames[vectorSize], "* in1, __global float", sizeNames[vectorSize], "* in2, __global float", sizeNames[vectorSize], "* in3 )\n"
33 "{\n"
34 " int i = get_global_id(0);\n"
35 " out[i] = ", name, "( in1[i], in2[i], in3[i] );\n"
36 "}\n"
37 };
38 const char *c3[] = { "__kernel void math_kernel", sizeNames[vectorSize], "( __global float* out, __global float* in, __global float* in2, __global float* in3)\n"
39 "{\n"
40 " size_t i = get_global_id(0);\n"
41 " if( i + 1 < get_global_size(0) )\n"
42 " {\n"
43 " float3 f0 = vload3( 0, in + 3 * i );\n"
44 " float3 f1 = vload3( 0, in2 + 3 * i );\n"
45 " float3 f2 = vload3( 0, in3 + 3 * i );\n"
46 " f0 = ", name, "( f0, f1, f2 );\n"
47 " vstore3( f0, 0, out + 3*i );\n"
48 " }\n"
49 " else\n"
50 " {\n"
51 " size_t parity = i & 1; // Figure out how many elements are left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two buffer size \n"
52 " float3 f0, f1, f2;\n"
53 " switch( parity )\n"
54 " {\n"
55 " case 1:\n"
56 " f0 = (float3)( in[3*i], NAN, NAN ); \n"
57 " f1 = (float3)( in2[3*i], NAN, NAN ); \n"
58 " f2 = (float3)( in3[3*i], NAN, NAN ); \n"
59 " break;\n"
60 " case 0:\n"
61 " f0 = (float3)( in[3*i], in[3*i+1], NAN ); \n"
62 " f1 = (float3)( in2[3*i], in2[3*i+1], NAN ); \n"
63 " f2 = (float3)( in3[3*i], in3[3*i+1], NAN ); \n"
64 " break;\n"
65 " }\n"
66 " f0 = ", name, "( f0, f1, f2 );\n"
67 " switch( parity )\n"
68 " {\n"
69 " case 0:\n"
70 " out[3*i+1] = f0.y; \n"
71 " // fall through\n"
72 " case 1:\n"
73 " out[3*i] = f0.x; \n"
74 " break;\n"
75 " }\n"
76 " }\n"
77 "}\n"
78 };
79
80 const char **kern = c;
81 size_t kernSize = sizeof(c)/sizeof(c[0]);
82
83 if( sizeValues[vectorSize] == 3 )
84 {
85 kern = c3;
86 kernSize = sizeof(c3)/sizeof(c3[0]);
87 }
88
89 char testName[32];
90 snprintf( testName, sizeof( testName ) -1, "math_kernel%s", sizeNames[vectorSize] );
91
92 return MakeKernel(kern, (cl_uint) kernSize, testName, k, p);
93 }
94
BuildKernelDouble(const char * name,int vectorSize,cl_kernel * k,cl_program * p)95 static int BuildKernelDouble( const char *name, int vectorSize, cl_kernel *k, cl_program *p )
96 {
97 const char *c[] = {
98 "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n",
99 "__kernel void math_kernel", sizeNames[vectorSize], "( __global double", sizeNames[vectorSize], "* out, __global double", sizeNames[vectorSize], "* in1, __global double", sizeNames[vectorSize], "* in2, __global double", sizeNames[vectorSize], "* in3 )\n"
100 "{\n"
101 " int i = get_global_id(0);\n"
102 " out[i] = ", name, "( in1[i], in2[i], in3[i] );\n"
103 "}\n"
104 };
105 const char *c3[] = { "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n",
106 "__kernel void math_kernel", sizeNames[vectorSize], "( __global double* out, __global double* in, __global double* in2, __global double* in3)\n"
107 "{\n"
108 " size_t i = get_global_id(0);\n"
109 " if( i + 1 < get_global_size(0) )\n"
110 " {\n"
111 " double3 d0 = vload3( 0, in + 3 * i );\n"
112 " double3 d1 = vload3( 0, in2 + 3 * i );\n"
113 " double3 d2 = vload3( 0, in3 + 3 * i );\n"
114 " d0 = ", name, "( d0, d1, d2 );\n"
115 " vstore3( d0, 0, out + 3*i );\n"
116 " }\n"
117 " else\n"
118 " {\n"
119 " size_t parity = i & 1; // Figure out how many elements are left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two buffer size \n"
120 " double3 d0, d1, d2;\n"
121 " switch( parity )\n"
122 " {\n"
123 " case 1:\n"
124 " d0 = (double3)( in[3*i], NAN, NAN ); \n"
125 " d1 = (double3)( in2[3*i], NAN, NAN ); \n"
126 " d2 = (double3)( in3[3*i], NAN, NAN ); \n"
127 " break;\n"
128 " case 0:\n"
129 " d0 = (double3)( in[3*i], in[3*i+1], NAN ); \n"
130 " d1 = (double3)( in2[3*i], in2[3*i+1], NAN ); \n"
131 " d2 = (double3)( in3[3*i], in3[3*i+1], NAN ); \n"
132 " break;\n"
133 " }\n"
134 " d0 = ", name, "( d0, d1, d2 );\n"
135 " switch( parity )\n"
136 " {\n"
137 " case 0:\n"
138 " out[3*i+1] = d0.y; \n"
139 " // fall through\n"
140 " case 1:\n"
141 " out[3*i] = d0.x; \n"
142 " break;\n"
143 " }\n"
144 " }\n"
145 "}\n"
146 };
147
148 const char **kern = c;
149 size_t kernSize = sizeof(c)/sizeof(c[0]);
150
151 if( sizeValues[vectorSize] == 3 )
152 {
153 kern = c3;
154 kernSize = sizeof(c3)/sizeof(c3[0]);
155 }
156
157 char testName[32];
158 snprintf( testName, sizeof( testName ) -1, "math_kernel%s", sizeNames[vectorSize] );
159
160 return MakeKernel(kern, (cl_uint) kernSize, testName, k, p);
161 }
162
163 typedef struct BuildKernelInfo
164 {
165 cl_uint offset; // the first vector size to build
166 cl_kernel *kernels;
167 cl_program *programs;
168 const char *nameInCode;
169 }BuildKernelInfo;
170
171 static cl_int BuildKernel_FloatFn( cl_uint job_id, cl_uint thread_id UNUSED, void *p );
BuildKernel_FloatFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)172 static cl_int BuildKernel_FloatFn( cl_uint job_id, cl_uint thread_id UNUSED, void *p )
173 {
174 BuildKernelInfo *info = (BuildKernelInfo*) p;
175 cl_uint i = info->offset + job_id;
176 return BuildKernel( info->nameInCode, i, info->kernels + i, info->programs + i );
177 }
178
179 static cl_int BuildKernel_DoubleFn( cl_uint job_id, cl_uint thread_id UNUSED, void *p );
BuildKernel_DoubleFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)180 static cl_int BuildKernel_DoubleFn( cl_uint job_id, cl_uint thread_id UNUSED, void *p )
181 {
182 BuildKernelInfo *info = (BuildKernelInfo*) p;
183 cl_uint i = info->offset + job_id;
184 return BuildKernelDouble( info->nameInCode, i, info->kernels + i, info->programs + i );
185 }
186
TestFunc_mad(const Func * f,MTdata d)187 int TestFunc_mad(const Func *f, MTdata d)
188 {
189 uint64_t i;
190 uint32_t j, k;
191 int error;
192
193 logFunctionInfo(f->name,sizeof(cl_float),gTestFastRelaxed);
194
195 cl_program programs[ VECTOR_SIZE_COUNT ];
196 cl_kernel kernels[ VECTOR_SIZE_COUNT ];
197 float maxError = 0.0f;
198 // int ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
199 float maxErrorVal = 0.0f;
200 float maxErrorVal2 = 0.0f;
201 float maxErrorVal3 = 0.0f;
202 size_t bufferSize = (gWimpyMode)? gWimpyBufferSize: BUFFER_SIZE;
203 uint64_t step = bufferSize / sizeof( float );
204
205 if( gWimpyMode )
206 {
207 step = (1ULL<<32) * gWimpyReductionFactor / (512);
208 }
209 // Init the kernels
210 BuildKernelInfo build_info = { gMinVectorSizeIndex, kernels, programs, f->nameInCode };
211 if( (error = ThreadPool_Do( BuildKernel_FloatFn, gMaxVectorSizeIndex - gMinVectorSizeIndex, &build_info ) ))
212 return error;
213 /*
214 for( i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++ )
215 if( (error = BuildKernel( f->nameInCode, (int) i, kernels + i, programs + i) ) )
216 return error;
217 */
218
219 for( i = 0; i < (1ULL<<32); i += step )
220 {
221 //Init input array
222 uint32_t *p = (uint32_t *)gIn;
223 uint32_t *p2 = (uint32_t *)gIn2;
224 uint32_t *p3 = (uint32_t *)gIn3;
225 for( j = 0; j < bufferSize / sizeof( float ); j++ )
226 {
227 p[j] = genrand_int32(d);
228 p2[j] = genrand_int32(d);
229 p3[j] = genrand_int32(d);
230 }
231 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0, bufferSize, gIn, 0, NULL, NULL) ))
232 {
233 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer ***\n", error );
234 return error;
235 }
236 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer2, CL_FALSE, 0, bufferSize, gIn2, 0, NULL, NULL) ))
237 {
238 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer2 ***\n", error );
239 return error;
240 }
241 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer3, CL_FALSE, 0, bufferSize, gIn3, 0, NULL, NULL) ))
242 {
243 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer3 ***\n", error );
244 return error;
245 }
246
247 // write garbage into output arrays
248 for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ )
249 {
250 uint32_t pattern = 0xffffdead;
251 memset_pattern4(gOut[j], &pattern, bufferSize);
252 if( (error = clEnqueueWriteBuffer(gQueue, gOutBuffer[j], CL_FALSE, 0, bufferSize, gOut[j], 0, NULL, NULL) ))
253 {
254 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer2(%d) ***\n", error, j );
255 goto exit;
256 }
257 }
258
259 // Run the kernels
260 for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ )
261 {
262 size_t vectorSize = sizeof( cl_float ) * sizeValues[j];
263 size_t localCount = (bufferSize + vectorSize - 1) / vectorSize; // bufferSize / vectorSize rounded up
264 if( ( error = clSetKernelArg(kernels[j], 0, sizeof( gOutBuffer[j] ), &gOutBuffer[j] ) )) { LogBuildError(programs[j]); goto exit; }
265 if( ( error = clSetKernelArg( kernels[j], 1, sizeof( gInBuffer ), &gInBuffer ) )) { LogBuildError(programs[j]); goto exit; }
266 if( ( error = clSetKernelArg( kernels[j], 2, sizeof( gInBuffer2 ), &gInBuffer2 ) )) { LogBuildError(programs[j]); goto exit; }
267 if( ( error = clSetKernelArg( kernels[j], 3, sizeof( gInBuffer3 ), &gInBuffer3 ) )) { LogBuildError(programs[j]); goto exit; }
268
269 if( (error = clEnqueueNDRangeKernel(gQueue, kernels[j], 1, NULL, &localCount, NULL, 0, NULL, NULL)) )
270 {
271 vlog_error( "FAILED -- could not execute kernel\n" );
272 goto exit;
273 }
274 }
275
276 // Get that moving
277 if( (error = clFlush(gQueue) ))
278 vlog( "clFlush failed\n" );
279
280 //Calculate the correctly rounded reference result
281 float *r = (float *)gOut_Ref;
282 float *s = (float *)gIn;
283 float *s2 = (float *)gIn2;
284 float *s3 = (float *)gIn3;
285 for( j = 0; j < bufferSize / sizeof( float ); j++ )
286 r[j] = (float) f->func.f_fff( s[j], s2[j], s3[j] );
287
288 // Read the data back
289 for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ )
290 {
291 if( (error = clEnqueueReadBuffer(gQueue, gOutBuffer[j], CL_TRUE, 0, bufferSize, gOut[j], 0, NULL, NULL)) )
292 {
293 vlog_error( "ReadArray failed %d\n", error );
294 goto exit;
295 }
296 }
297
298 if( gSkipCorrectnessTesting )
299 break;
300
301 //Verify data -- Commented out on purpose. no verification possible. MAD is a random number generator.
302 /*
303 uint32_t *t = gOut_Ref;
304 for( j = 0; j < bufferSize / sizeof( float ); j++ )
305 {
306 for( k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++ )
307 {
308 uint32_t *q = gOut[k];
309
310 // If we aren't getting the correctly rounded result
311 if( t[j] != q[j] )
312 {
313 float test = ((float*) q)[j];
314 double correct = f->func.f_fff( s[j], s2[j], s3[j] );
315 float err = Ulp_Error( test, correct );
316 int fail = ! (fabsf(err) <= f->float_ulps);
317
318 if( fail && ftz )
319 {
320 // retry per section 6.5.3.2
321 if( IsFloatSubnormal(correct) )
322 { // look at me,
323 fail = fail && ( test != 0.0f );
324 if( ! fail )
325 err = 0.0f;
326 }
327
328 // retry per section 6.5.3.3
329 if( fail && IsFloatSubnormal( s[j] ) )
330 { // look at me,
331 double correct2 = f->func.f_fff( 0.0, s2[j], s3[j] );
332 double correct3 = f->func.f_fff( -0.0, s2[j], s3[j] );
333 float err2 = Ulp_Error( test, correct2 );
334 float err3 = Ulp_Error( test, correct3 );
335 fail = fail && ((!(fabsf(err2) <= f->float_ulps)) && (!(fabsf(err3) <= f->float_ulps)));
336 if( fabsf( err2 ) < fabsf(err ) )
337 err = err2;
338 if( fabsf( err3 ) < fabsf(err ) )
339 err = err3;
340
341 // retry per section 6.5.3.4
342 if( IsFloatResultSubnormal(correct2, f->float_ulps ) || IsFloatResultSubnormal(correct3, f->float_ulps ) )
343 { // look at me now,
344 fail = fail && ( test != 0.0f);
345 if( ! fail )
346 err = 0.0f;
347 }
348
349 //try with first two args as zero
350 if( IsFloatSubnormal( s2[j] ) )
351 { // its fun to have fun,
352 correct2 = f->func.f_fff( 0.0, 0.0, s3[j] );
353 correct3 = f->func.f_fff( -0.0, 0.0, s3[j] );
354 double correct4 = f->func.f_fff( 0.0, -0.0, s3[j] );
355 double correct5 = f->func.f_fff( -0.0, -0.0, s3[j] );
356 err2 = Ulp_Error( test, correct2 );
357 err3 = Ulp_Error( test, correct3 );
358 float err4 = Ulp_Error( test, correct4 );
359 float err5 = Ulp_Error( test, correct5 );
360 fail = fail && ((!(fabsf(err2) <= f->float_ulps)) && (!(fabsf(err3) <= f->float_ulps)) &&
361 (!(fabsf(err4) <= f->float_ulps)) && (!(fabsf(err5) <= f->float_ulps)));
362 if( fabsf( err2 ) < fabsf(err ) )
363 err = err2;
364 if( fabsf( err3 ) < fabsf(err ) )
365 err = err3;
366 if( fabsf( err4 ) < fabsf(err ) )
367 err = err4;
368 if( fabsf( err5 ) < fabsf(err ) )
369 err = err5;
370
371 // retry per section 6.5.3.4
372 if( IsFloatResultSubnormal(correct2, f->float_ulps ) || IsFloatResultSubnormal(correct3, f->float_ulps ) ||
373 IsFloatResultSubnormal(correct4, f->float_ulps ) || IsFloatResultSubnormal(correct5, f->float_ulps ) )
374 {
375 fail = fail && ( test != 0.0f);
376 if( ! fail )
377 err = 0.0f;
378 }
379
380 if( IsFloatSubnormal( s3[j] ) )
381 { // but you have to know how!
382 correct2 = f->func.f_fff( 0.0, 0.0, 0.0f );
383 correct3 = f->func.f_fff( -0.0, 0.0, 0.0f );
384 correct4 = f->func.f_fff( 0.0, -0.0, 0.0f );
385 correct5 = f->func.f_fff( -0.0, -0.0, 0.0f );
386 double correct6 = f->func.f_fff( 0.0, 0.0, -0.0f );
387 double correct7 = f->func.f_fff( -0.0, 0.0, -0.0f );
388 double correct8 = f->func.f_fff( 0.0, -0.0, -0.0f );
389 double correct9 = f->func.f_fff( -0.0, -0.0, -0.0f );
390 err2 = Ulp_Error( test, correct2 );
391 err3 = Ulp_Error( test, correct3 );
392 err4 = Ulp_Error( test, correct4 );
393 err5 = Ulp_Error( test, correct5 );
394 float err6 = Ulp_Error( test, correct6 );
395 float err7 = Ulp_Error( test, correct7 );
396 float err8 = Ulp_Error( test, correct8 );
397 float err9 = Ulp_Error( test, correct9 );
398 fail = fail && ((!(fabsf(err2) <= f->float_ulps)) && (!(fabsf(err3) <= f->float_ulps)) &&
399 (!(fabsf(err4) <= f->float_ulps)) && (!(fabsf(err5) <= f->float_ulps)) &&
400 (!(fabsf(err5) <= f->float_ulps)) && (!(fabsf(err6) <= f->float_ulps)) &&
401 (!(fabsf(err7) <= f->float_ulps)) && (!(fabsf(err8) <= f->float_ulps)));
402 if( fabsf( err2 ) < fabsf(err ) )
403 err = err2;
404 if( fabsf( err3 ) < fabsf(err ) )
405 err = err3;
406 if( fabsf( err4 ) < fabsf(err ) )
407 err = err4;
408 if( fabsf( err5 ) < fabsf(err ) )
409 err = err5;
410 if( fabsf( err6 ) < fabsf(err ) )
411 err = err6;
412 if( fabsf( err7 ) < fabsf(err ) )
413 err = err7;
414 if( fabsf( err8 ) < fabsf(err ) )
415 err = err8;
416 if( fabsf( err9 ) < fabsf(err ) )
417 err = err9;
418
419 // retry per section 6.5.3.4
420 if( IsFloatResultSubnormal(correct2, f->float_ulps ) || IsFloatResultSubnormal(correct3, f->float_ulps ) ||
421 IsFloatResultSubnormal(correct4, f->float_ulps ) || IsFloatResultSubnormal(correct5, f->float_ulps ) ||
422 IsFloatResultSubnormal( correct6, f->float_ulps ) || IsFloatResultSubnormal(correct7, f->float_ulps ) ||
423 IsFloatResultSubnormal(correct8, f->float_ulps ) || IsFloatResultSubnormal( correct9, f->float_ulps ) )
424 {
425 fail = fail && ( test != 0.0f);
426 if( ! fail )
427 err = 0.0f;
428 }
429 }
430 }
431 else if( IsFloatSubnormal( s3[j] ) )
432 {
433 correct2 = f->func.f_fff( 0.0, s2[j], 0.0 );
434 correct3 = f->func.f_fff( -0.0, s2[j], 0.0 );
435 double correct4 = f->func.f_fff( 0.0, s2[j], -0.0 );
436 double correct5 = f->func.f_fff( -0.0, s2[j], -0.0 );
437 err2 = Ulp_Error( test, correct2 );
438 err3 = Ulp_Error( test, correct3 );
439 float err4 = Ulp_Error( test, correct4 );
440 float err5 = Ulp_Error( test, correct5 );
441 fail = fail && ((!(fabsf(err2) <= f->float_ulps)) && (!(fabsf(err3) <= f->float_ulps)) &&
442 (!(fabsf(err4) <= f->float_ulps)) && (!(fabsf(err5) <= f->float_ulps)));
443 if( fabsf( err2 ) < fabsf(err ) )
444 err = err2;
445 if( fabsf( err3 ) < fabsf(err ) )
446 err = err3;
447 if( fabsf( err4 ) < fabsf(err ) )
448 err = err4;
449 if( fabsf( err5 ) < fabsf(err ) )
450 err = err5;
451
452 // retry per section 6.5.3.4
453 if( IsFloatResultSubnormal(correct2, f->float_ulps ) || IsFloatResultSubnormal(correct3, f->float_ulps ) ||
454 IsFloatResultSubnormal(correct4, f->float_ulps ) || IsFloatResultSubnormal(correct5, f->float_ulps ) )
455 {
456 fail = fail && ( test != 0.0f);
457 if( ! fail )
458 err = 0.0f;
459 }
460 }
461 }
462 else if( fail && IsFloatSubnormal( s2[j] ) )
463 {
464 double correct2 = f->func.f_fff( s[j], 0.0, s3[j] );
465 double correct3 = f->func.f_fff( s[j], -0.0, s3[j] );
466 float err2 = Ulp_Error( test, correct2 );
467 float err3 = Ulp_Error( test, correct3 );
468 fail = fail && ((!(fabsf(err2) <= f->float_ulps)) && (!(fabsf(err3) <= f->float_ulps)));
469 if( fabsf( err2 ) < fabsf(err ) )
470 err = err2;
471 if( fabsf( err3 ) < fabsf(err ) )
472 err = err3;
473
474 // retry per section 6.5.3.4
475 if( IsFloatResultSubnormal(correct2, f->float_ulps ) || IsFloatResultSubnormal(correct3, f->float_ulps ) )
476 {
477 fail = fail && ( test != 0.0f);
478 if( ! fail )
479 err = 0.0f;
480 }
481
482 //try with second two args as zero
483 if( IsFloatSubnormal( s3[j] ) )
484 {
485 correct2 = f->func.f_fff( s[j], 0.0, 0.0 );
486 correct3 = f->func.f_fff( s[j], -0.0, 0.0 );
487 double correct4 = f->func.f_fff( s[j], 0.0, -0.0 );
488 double correct5 = f->func.f_fff( s[j], -0.0, -0.0 );
489 err2 = Ulp_Error( test, correct2 );
490 err3 = Ulp_Error( test, correct3 );
491 float err4 = Ulp_Error( test, correct4 );
492 float err5 = Ulp_Error( test, correct5 );
493 fail = fail && ((!(fabsf(err2) <= f->float_ulps)) && (!(fabsf(err3) <= f->float_ulps)) &&
494 (!(fabsf(err4) <= f->float_ulps)) && (!(fabsf(err5) <= f->float_ulps)));
495 if( fabsf( err2 ) < fabsf(err ) )
496 err = err2;
497 if( fabsf( err3 ) < fabsf(err ) )
498 err = err3;
499 if( fabsf( err4 ) < fabsf(err ) )
500 err = err4;
501 if( fabsf( err5 ) < fabsf(err ) )
502 err = err5;
503
504 // retry per section 6.5.3.4
505 if( IsFloatResultSubnormal(correct2, f->float_ulps ) || IsFloatResultSubnormal(correct3, f->float_ulps ) ||
506 IsFloatResultSubnormal(correct4, f->float_ulps ) || IsFloatResultSubnormal(correct5, f->float_ulps ) )
507 {
508 fail = fail && ( test != 0.0f);
509 if( ! fail )
510 err = 0.0f;
511 }
512 }
513 }
514 else if( fail && IsFloatSubnormal(s3[j]) )
515 {
516 double correct2 = f->func.f_fff( s[j], s2[j], 0.0 );
517 double correct3 = f->func.f_fff( s[j], s2[j], -0.0 );
518 float err2 = Ulp_Error( test, correct2 );
519 float err3 = Ulp_Error( test, correct3 );
520 fail = fail && ((!(fabsf(err2) <= f->float_ulps)) && (!(fabsf(err3) <= f->float_ulps)));
521 if( fabsf( err2 ) < fabsf(err ) )
522 err = err2;
523 if( fabsf( err3 ) < fabsf(err ) )
524 err = err3;
525
526 // retry per section 6.5.3.4
527 if( IsFloatResultSubnormal(correct2, f->float_ulps ) || IsFloatResultSubnormal(correct3, f->float_ulps ) )
528 {
529 fail = fail && ( test != 0.0f);
530 if( ! fail )
531 err = 0.0f;
532 }
533 }
534 }
535
536 if( fabsf(err ) > maxError )
537 {
538 maxError = fabsf(err);
539 maxErrorVal = s[j];
540 maxErrorVal2 = s2[j];
541 maxErrorVal3 = s3[j];
542 }
543
544 if( fail )
545 {
546 vlog_error( "\nERROR: %s%s: %f ulp error at {%a, %a, %a}: *%a vs. %a\n", f->name, sizeNames[k], err, s[j], s2[j], s3[j], ((float*) gOut_Ref)[j], test );
547 error = -1;
548 goto exit;
549 }
550 }
551 }
552 }
553 */
554 if( 0 == (i & 0x0fffffff) )
555 {
556 vlog("." );
557 fflush(stdout);
558 }
559 }
560
561 if( ! gSkipCorrectnessTesting )
562 {
563 if( gWimpyMode )
564 vlog( "Wimp pass" );
565 else
566 vlog( "pass" );
567 }
568
569 if( gMeasureTimes )
570 {
571 //Init input array
572 uint32_t *p = (uint32_t *)gIn;
573 uint32_t *p2 = (uint32_t *)gIn2;
574 uint32_t *p3 = (uint32_t *)gIn3;
575 for( j = 0; j < bufferSize / sizeof( float ); j++ )
576 {
577 p[j] = genrand_int32(d);
578 p2[j] = genrand_int32(d);
579 p3[j] = genrand_int32(d);
580 }
581 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0, bufferSize, gIn, 0, NULL, NULL) ))
582 {
583 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer ***\n", error );
584 return error;
585 }
586 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer2, CL_FALSE, 0, bufferSize, gIn2, 0, NULL, NULL) ))
587 {
588 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer2 ***\n", error );
589 return error;
590 }
591 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer3, CL_FALSE, 0, bufferSize, gIn3, 0, NULL, NULL) ))
592 {
593 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer3 ***\n", error );
594 return error;
595 }
596
597
598 // Run the kernels
599 for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ )
600 {
601 size_t vectorSize = sizeof( cl_float ) * sizeValues[j];
602 size_t localCount = (bufferSize + vectorSize - 1) / vectorSize; // bufferSize / vectorSize rounded up
603 if( ( error = clSetKernelArg(kernels[j], 0, sizeof( gOutBuffer[j] ), &gOutBuffer[j] ) )) { LogBuildError(programs[j]); goto exit; }
604 if( ( error = clSetKernelArg( kernels[j], 1, sizeof( gInBuffer ), &gInBuffer ) )) { LogBuildError(programs[j]); goto exit; }
605 if( ( error = clSetKernelArg( kernels[j], 2, sizeof( gInBuffer2 ), &gInBuffer2 ) )) { LogBuildError(programs[j]); goto exit; }
606 if( ( error = clSetKernelArg( kernels[j], 3, sizeof( gInBuffer3 ), &gInBuffer3 ) )) { LogBuildError(programs[j]); goto exit; }
607
608 double sum = 0.0;
609 double bestTime = INFINITY;
610 for( k = 0; k < PERF_LOOP_COUNT; k++ )
611 {
612 uint64_t startTime = GetTime();
613 if( (error = clEnqueueNDRangeKernel(gQueue, kernels[j], 1, NULL, &localCount, NULL, 0, NULL, NULL)) )
614 {
615 vlog_error( "FAILED -- could not execute kernel\n" );
616 goto exit;
617 }
618
619 // Make sure OpenCL is done
620 if( (error = clFinish(gQueue) ) )
621 {
622 vlog_error( "Error %d at clFinish\n", error );
623 goto exit;
624 }
625
626 uint64_t endTime = GetTime();
627 double time = SubtractTime( endTime, startTime );
628 sum += time;
629 if( time < bestTime )
630 bestTime = time;
631 }
632
633 if( gReportAverageTimes )
634 bestTime = sum / PERF_LOOP_COUNT;
635 double clocksPerOp = bestTime * (double) gDeviceFrequency * gComputeDevices * gSimdSize * 1e6 / (bufferSize / sizeof( float ) );
636 vlog_perf( clocksPerOp, LOWER_IS_BETTER, "clocks / element", "%sf%s", f->name, sizeNames[j] );
637 }
638 }
639
640 if( ! gSkipCorrectnessTesting )
641 vlog( "\t%8.2f @ {%a, %a, %a}", maxError, maxErrorVal, maxErrorVal2, maxErrorVal3 );
642 vlog( "\n" );
643
644 exit:
645 // Release
646 for( k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++ )
647 {
648 clReleaseKernel(kernels[k]);
649 clReleaseProgram(programs[k]);
650 }
651
652 return error;
653 }
654
TestFunc_mad_Double(const Func * f,MTdata d)655 int TestFunc_mad_Double(const Func *f, MTdata d)
656 {
657 uint64_t i;
658 uint32_t j, k;
659 int error;
660 cl_program programs[ VECTOR_SIZE_COUNT ];
661 cl_kernel kernels[ VECTOR_SIZE_COUNT ];
662 float maxError = 0.0f;
663 // int ftz = f->ftz || gForceFTZ;
664 double maxErrorVal = 0.0f;
665 double maxErrorVal2 = 0.0f;
666 double maxErrorVal3 = 0.0f;
667 size_t bufferSize = (gWimpyMode)? gWimpyBufferSize: BUFFER_SIZE;
668
669 logFunctionInfo(f->name,sizeof(cl_double),gTestFastRelaxed);
670 uint64_t step = bufferSize / sizeof( double );
671 if( gWimpyMode )
672 {
673 step = (1ULL<<32) * gWimpyReductionFactor / (512);
674 }
675 // Init the kernels
676 BuildKernelInfo build_info = { gMinVectorSizeIndex, kernels, programs, f->nameInCode };
677 if( (error = ThreadPool_Do( BuildKernel_DoubleFn,
678 gMaxVectorSizeIndex - gMinVectorSizeIndex,
679 &build_info ) ))
680 {
681 return error;
682 }
683 /*
684 for( i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++ )
685 if( (error = BuildKernelDouble( f->nameInCode, (int) i, kernels + i, programs + i) ) )
686 return error;
687 */
688
689 for( i = 0; i < (1ULL<<32); i += step )
690 {
691 //Init input array
692 double *p = (double *)gIn;
693 double *p2 = (double *)gIn2;
694 double *p3 = (double *)gIn3;
695 for( j = 0; j < bufferSize / sizeof( double ); j++ )
696 {
697 p[j] = DoubleFromUInt32(genrand_int32(d));
698 p2[j] = DoubleFromUInt32(genrand_int32(d));
699 p3[j] = DoubleFromUInt32(genrand_int32(d));
700 }
701 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0, bufferSize, gIn, 0, NULL, NULL) ))
702 {
703 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer ***\n", error );
704 return error;
705 }
706 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer2, CL_FALSE, 0, bufferSize, gIn2, 0, NULL, NULL) ))
707 {
708 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer2 ***\n", error );
709 return error;
710 }
711 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer3, CL_FALSE, 0, bufferSize, gIn3, 0, NULL, NULL) ))
712 {
713 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer3 ***\n", error );
714 return error;
715 }
716
717 // write garbage into output arrays
718 for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ )
719 {
720 uint32_t pattern = 0xffffdead;
721 memset_pattern4(gOut[j], &pattern, bufferSize);
722 if( (error = clEnqueueWriteBuffer(gQueue, gOutBuffer[j], CL_FALSE, 0, bufferSize, gOut[j], 0, NULL, NULL) ))
723 {
724 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer2(%d) ***\n", error, j );
725 goto exit;
726 }
727 }
728
729 // Run the kernels
730 for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ )
731 {
732 size_t vectorSize = sizeof( cl_double ) * sizeValues[j];
733 size_t localCount = (bufferSize + vectorSize - 1) / vectorSize; // bufferSize / vectorSize rounded up
734 if( ( error = clSetKernelArg(kernels[j], 0, sizeof( gOutBuffer[j] ), &gOutBuffer[j] ) )) { LogBuildError(programs[j]); goto exit; }
735 if( ( error = clSetKernelArg( kernels[j], 1, sizeof( gInBuffer ), &gInBuffer ) )) { LogBuildError(programs[j]); goto exit; }
736 if( ( error = clSetKernelArg( kernels[j], 2, sizeof( gInBuffer2 ), &gInBuffer2 ) )) { LogBuildError(programs[j]); goto exit; }
737 if( ( error = clSetKernelArg( kernels[j], 3, sizeof( gInBuffer3 ), &gInBuffer3 ) )) { LogBuildError(programs[j]); goto exit; }
738
739 if( (error = clEnqueueNDRangeKernel(gQueue, kernels[j], 1, NULL, &localCount, NULL, 0, NULL, NULL)) )
740 {
741 vlog_error( "FAILED -- could not execute kernel\n" );
742 goto exit;
743 }
744 }
745
746 // Get that moving
747 if( (error = clFlush(gQueue) ))
748 vlog( "clFlush failed\n" );
749
750 //Calculate the correctly rounded reference result
751 double *r = (double *)gOut_Ref;
752 double *s = (double *)gIn;
753 double *s2 = (double *)gIn2;
754 double *s3 = (double *)gIn3;
755 for( j = 0; j < bufferSize / sizeof( double ); j++ )
756 r[j] = (double) f->dfunc.f_fff( s[j], s2[j], s3[j] );
757
758 // Read the data back
759 for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ )
760 {
761 if( (error = clEnqueueReadBuffer(gQueue, gOutBuffer[j], CL_TRUE, 0, bufferSize, gOut[j], 0, NULL, NULL)) )
762 {
763 vlog_error( "ReadArray failed %d\n", error );
764 goto exit;
765 }
766 }
767
768 if( gSkipCorrectnessTesting )
769 break;
770
771 //Verify data -- Commented out on purpose. no verification possible. MAD is a random number generator.
772 /*
773 uint64_t *t = gOut_Ref;
774 for( j = 0; j < bufferSize / sizeof( double ); j++ )
775 {
776 for( k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++ )
777 {
778 uint64_t *q = gOut[k];
779
780 // If we aren't getting the correctly rounded result
781 if( t[j] != q[j] )
782 {
783 double test = ((double*) q)[j];
784 long double correct = f->dfunc.f_fff( s[j], s2[j], s3[j] );
785 float err = Bruteforce_Ulp_Error_Double( test, correct );
786 int fail = ! (fabsf(err) <= f->double_ulps);
787
788 if( fail && ftz )
789 {
790 // retry per section 6.5.3.2
791 if( IsDoubleResultSubnormal(correct, f->double_ulps) )
792 { // look at me,
793 fail = fail && ( test != 0.0f );
794 if( ! fail )
795 err = 0.0f;
796 }
797
798 // retry per section 6.5.3.3
799 if( fail && IsDoubleSubnormal( s[j] ) )
800 { // look at me,
801 long double correct2 = f->dfunc.f_fff( 0.0, s2[j], s3[j] );
802 long double correct3 = f->dfunc.f_fff( -0.0, s2[j], s3[j] );
803 float err2 = Bruteforce_Ulp_Error_Double( test, correct2 );
804 float err3 = Bruteforce_Ulp_Error_Double( test, correct3 );
805 fail = fail && ((!(fabsf(err2) <= f->double_ulps)) && (!(fabsf(err3) <= f->double_ulps)));
806 if( fabsf( err2 ) < fabsf(err ) )
807 err = err2;
808 if( fabsf( err3 ) < fabsf(err ) )
809 err = err3;
810
811 // retry per section 6.5.3.4
812 if( IsDoubleResultSubnormal( correct2, f->double_ulps ) || IsDoubleResultSubnormal( correct3, f->double_ulps ) )
813 { // look at me now,
814 fail = fail && ( test != 0.0f);
815 if( ! fail )
816 err = 0.0f;
817 }
818
819 //try with first two args as zero
820 if( IsDoubleSubnormal( s2[j] ) )
821 { // its fun to have fun,
822 correct2 = f->dfunc.f_fff( 0.0, 0.0, s3[j] );
823 correct3 = f->dfunc.f_fff( -0.0, 0.0, s3[j] );
824 long double correct4 = f->dfunc.f_fff( 0.0, -0.0, s3[j] );
825 long double correct5 = f->dfunc.f_fff( -0.0, -0.0, s3[j] );
826 err2 = Bruteforce_Ulp_Error_Double( test, correct2 );
827 err3 = Bruteforce_Ulp_Error_Double( test, correct3 );
828 float err4 = Bruteforce_Ulp_Error_Double( test, correct4 );
829 float err5 = Bruteforce_Ulp_Error_Double( test, correct5 );
830 fail = fail && ((!(fabsf(err2) <= f->double_ulps)) && (!(fabsf(err3) <= f->double_ulps)) &&
831 (!(fabsf(err4) <= f->double_ulps)) && (!(fabsf(err5) <= f->double_ulps)));
832 if( fabsf( err2 ) < fabsf(err ) )
833 err = err2;
834 if( fabsf( err3 ) < fabsf(err ) )
835 err = err3;
836 if( fabsf( err4 ) < fabsf(err ) )
837 err = err4;
838 if( fabsf( err5 ) < fabsf(err ) )
839 err = err5;
840
841 // retry per section 6.5.3.4
842 if( IsDoubleResultSubnormal( correct2, f->double_ulps ) || IsDoubleResultSubnormal( correct3, f->double_ulps ) ||
843 IsDoubleResultSubnormal( correct4, f->double_ulps ) || IsDoubleResultSubnormal( correct5, f->double_ulps ) )
844 {
845 fail = fail && ( test != 0.0f);
846 if( ! fail )
847 err = 0.0f;
848 }
849
850 if( IsDoubleSubnormal( s3[j] ) )
851 { // but you have to know how!
852 correct2 = f->dfunc.f_fff( 0.0, 0.0, 0.0f );
853 correct3 = f->dfunc.f_fff( -0.0, 0.0, 0.0f );
854 correct4 = f->dfunc.f_fff( 0.0, -0.0, 0.0f );
855 correct5 = f->dfunc.f_fff( -0.0, -0.0, 0.0f );
856 long double correct6 = f->dfunc.f_fff( 0.0, 0.0, -0.0f );
857 long double correct7 = f->dfunc.f_fff( -0.0, 0.0, -0.0f );
858 long double correct8 = f->dfunc.f_fff( 0.0, -0.0, -0.0f );
859 long double correct9 = f->dfunc.f_fff( -0.0, -0.0, -0.0f );
860 err2 = Bruteforce_Ulp_Error_Double( test, correct2 );
861 err3 = Bruteforce_Ulp_Error_Double( test, correct3 );
862 err4 = Bruteforce_Ulp_Error_Double( test, correct4 );
863 err5 = Bruteforce_Ulp_Error_Double( test, correct5 );
864 float err6 = Bruteforce_Ulp_Error_Double( test, correct6 );
865 float err7 = Bruteforce_Ulp_Error_Double( test, correct7 );
866 float err8 = Bruteforce_Ulp_Error_Double( test, correct8 );
867 float err9 = Bruteforce_Ulp_Error_Double( test, correct9 );
868 fail = fail && ((!(fabsf(err2) <= f->double_ulps)) && (!(fabsf(err3) <= f->double_ulps)) &&
869 (!(fabsf(err4) <= f->double_ulps)) && (!(fabsf(err5) <= f->double_ulps)) &&
870 (!(fabsf(err5) <= f->double_ulps)) && (!(fabsf(err6) <= f->double_ulps)) &&
871 (!(fabsf(err7) <= f->double_ulps)) && (!(fabsf(err8) <= f->double_ulps)));
872 if( fabsf( err2 ) < fabsf(err ) )
873 err = err2;
874 if( fabsf( err3 ) < fabsf(err ) )
875 err = err3;
876 if( fabsf( err4 ) < fabsf(err ) )
877 err = err4;
878 if( fabsf( err5 ) < fabsf(err ) )
879 err = err5;
880 if( fabsf( err6 ) < fabsf(err ) )
881 err = err6;
882 if( fabsf( err7 ) < fabsf(err ) )
883 err = err7;
884 if( fabsf( err8 ) < fabsf(err ) )
885 err = err8;
886 if( fabsf( err9 ) < fabsf(err ) )
887 err = err9;
888
889 // retry per section 6.5.3.4
890 if( IsDoubleResultSubnormal( correct2, f->double_ulps ) || IsDoubleResultSubnormal( correct3, f->double_ulps ) ||
891 IsDoubleResultSubnormal( correct4, f->double_ulps ) || IsDoubleResultSubnormal( correct5, f->double_ulps ) ||
892 IsDoubleResultSubnormal( correct6, f->double_ulps ) || IsDoubleResultSubnormal( correct7, f->double_ulps ) ||
893 IsDoubleResultSubnormal( correct8, f->double_ulps ) || IsDoubleResultSubnormal( correct9, f->double_ulps ) )
894 {
895 fail = fail && ( test != 0.0f);
896 if( ! fail )
897 err = 0.0f;
898 }
899 }
900 }
901 else if( IsDoubleSubnormal( s3[j] ) )
902 {
903 correct2 = f->dfunc.f_fff( 0.0, s2[j], 0.0 );
904 correct3 = f->dfunc.f_fff( -0.0, s2[j], 0.0 );
905 long double correct4 = f->dfunc.f_fff( 0.0, s2[j], -0.0 );
906 long double correct5 = f->dfunc.f_fff( -0.0, s2[j], -0.0 );
907 err2 = Bruteforce_Ulp_Error_Double( test, correct2 );
908 err3 = Bruteforce_Ulp_Error_Double( test, correct3 );
909 float err4 = Bruteforce_Ulp_Error_Double( test, correct4 );
910 float err5 = Bruteforce_Ulp_Error_Double( test, correct5 );
911 fail = fail && ((!(fabsf(err2) <= f->double_ulps)) && (!(fabsf(err3) <= f->double_ulps)) &&
912 (!(fabsf(err4) <= f->double_ulps)) && (!(fabsf(err5) <= f->double_ulps)));
913 if( fabsf( err2 ) < fabsf(err ) )
914 err = err2;
915 if( fabsf( err3 ) < fabsf(err ) )
916 err = err3;
917 if( fabsf( err4 ) < fabsf(err ) )
918 err = err4;
919 if( fabsf( err5 ) < fabsf(err ) )
920 err = err5;
921
922 // retry per section 6.5.3.4
923 if( IsDoubleResultSubnormal( correct2, f->double_ulps ) || IsDoubleResultSubnormal( correct3, f->double_ulps ) ||
924 IsDoubleResultSubnormal( correct4, f->double_ulps ) || IsDoubleResultSubnormal( correct5, f->double_ulps ) )
925 {
926 fail = fail && ( test != 0.0f);
927 if( ! fail )
928 err = 0.0f;
929 }
930 }
931 }
932 else if( fail && IsDoubleSubnormal( s2[j] ) )
933 {
934 long double correct2 = f->dfunc.f_fff( s[j], 0.0, s3[j] );
935 long double correct3 = f->dfunc.f_fff( s[j], -0.0, s3[j] );
936 float err2 = Bruteforce_Ulp_Error_Double( test, correct2 );
937 float err3 = Bruteforce_Ulp_Error_Double( test, correct3 );
938 fail = fail && ((!(fabsf(err2) <= f->double_ulps)) && (!(fabsf(err3) <= f->double_ulps)));
939 if( fabsf( err2 ) < fabsf(err ) )
940 err = err2;
941 if( fabsf( err3 ) < fabsf(err ) )
942 err = err3;
943
944 // retry per section 6.5.3.4
945 if( IsDoubleResultSubnormal( correct2, f->double_ulps ) || IsDoubleResultSubnormal( correct3, f->double_ulps ) )
946 {
947 fail = fail && ( test != 0.0f);
948 if( ! fail )
949 err = 0.0f;
950 }
951
952 //try with second two args as zero
953 if( IsDoubleSubnormal( s3[j] ) )
954 {
955 correct2 = f->dfunc.f_fff( s[j], 0.0, 0.0 );
956 correct3 = f->dfunc.f_fff( s[j], -0.0, 0.0 );
957 long double correct4 = f->dfunc.f_fff( s[j], 0.0, -0.0 );
958 long double correct5 = f->dfunc.f_fff( s[j], -0.0, -0.0 );
959 err2 = Bruteforce_Ulp_Error_Double( test, correct2 );
960 err3 = Bruteforce_Ulp_Error_Double( test, correct3 );
961 float err4 = Bruteforce_Ulp_Error_Double( test, correct4 );
962 float err5 = Bruteforce_Ulp_Error_Double( test, correct5 );
963 fail = fail && ((!(fabsf(err2) <= f->double_ulps)) && (!(fabsf(err3) <= f->double_ulps)) &&
964 (!(fabsf(err4) <= f->double_ulps)) && (!(fabsf(err5) <= f->double_ulps)));
965 if( fabsf( err2 ) < fabsf(err ) )
966 err = err2;
967 if( fabsf( err3 ) < fabsf(err ) )
968 err = err3;
969 if( fabsf( err4 ) < fabsf(err ) )
970 err = err4;
971 if( fabsf( err5 ) < fabsf(err ) )
972 err = err5;
973
974 // retry per section 6.5.3.4
975 if( IsDoubleResultSubnormal( correct2, f->double_ulps ) || IsDoubleResultSubnormal( correct3, f->double_ulps ) ||
976 IsDoubleResultSubnormal( correct4, f->double_ulps ) || IsDoubleResultSubnormal( correct5, f->double_ulps ) )
977 {
978 fail = fail && ( test != 0.0f);
979 if( ! fail )
980 err = 0.0f;
981 }
982 }
983 }
984 else if( fail && IsDoubleSubnormal(s3[j]) )
985 {
986 long double correct2 = f->dfunc.f_fff( s[j], s2[j], 0.0 );
987 long double correct3 = f->dfunc.f_fff( s[j], s2[j], -0.0 );
988 float err2 = Bruteforce_Ulp_Error_Double( test, correct2 );
989 float err3 = Bruteforce_Ulp_Error_Double( test, correct3 );
990 fail = fail && ((!(fabsf(err2) <= f->double_ulps)) && (!(fabsf(err3) <= f->double_ulps)));
991 if( fabsf( err2 ) < fabsf(err ) )
992 err = err2;
993 if( fabsf( err3 ) < fabsf(err ) )
994 err = err3;
995
996 // retry per section 6.5.3.4
997 if( IsDoubleResultSubnormal( correct2, f->double_ulps ) || IsDoubleResultSubnormal( correct3, f->double_ulps ) )
998 {
999 fail = fail && ( test != 0.0f);
1000 if( ! fail )
1001 err = 0.0f;
1002 }
1003 }
1004 }
1005
1006 if( fabsf(err ) > maxError )
1007 {
1008 maxError = fabsf(err);
1009 maxErrorVal = s[j];
1010 maxErrorVal2 = s2[j];
1011 maxErrorVal3 = s3[j];
1012 }
1013
1014 if( fail )
1015 {
1016 vlog_error( "\nERROR: %sD%s: %f ulp error at {%a, %a, %a}: *%a vs. %a\n", f->name, sizeNames[k], err, s[j], s2[j], s3[j], ((double*) gOut_Ref)[j], test );
1017 error = -1;
1018 goto exit;
1019 }
1020 }
1021 }
1022 }
1023 */
1024 if( 0 == (i & 0x0fffffff) )
1025 {
1026 vlog("." );
1027 fflush(stdout);
1028 }
1029 }
1030
1031 if( ! gSkipCorrectnessTesting )
1032 {
1033 if( gWimpyMode )
1034 vlog( "Wimp pass" );
1035 else
1036 vlog( "pass" );
1037 }
1038
1039 if( gMeasureTimes )
1040 {
1041 //Init input array
1042 double *p = (double *)gIn;
1043 double *p2 = (double *)gIn2;
1044 double *p3 = (double *)gIn3;
1045 for( j = 0; j < bufferSize / sizeof( double ); j++ )
1046 {
1047 p[j] = DoubleFromUInt32(genrand_int32(d));
1048 p2[j] = DoubleFromUInt32(genrand_int32(d));
1049 p3[j] = DoubleFromUInt32(genrand_int32(d));
1050 }
1051 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0, bufferSize, gIn, 0, NULL, NULL) ))
1052 {
1053 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer ***\n", error );
1054 return error;
1055 }
1056 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer2, CL_FALSE, 0, bufferSize, gIn2, 0, NULL, NULL) ))
1057 {
1058 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer2 ***\n", error );
1059 return error;
1060 }
1061 if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer3, CL_FALSE, 0, bufferSize, gIn3, 0, NULL, NULL) ))
1062 {
1063 vlog_error( "\n*** Error %d in clEnqueueWriteBuffer3 ***\n", error );
1064 return error;
1065 }
1066
1067
1068 // Run the kernels
1069 for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ )
1070 {
1071 size_t vectorSize = sizeof( cl_double ) * sizeValues[j];
1072 size_t localCount = (bufferSize + vectorSize - 1) / vectorSize; // bufferSize / vectorSize rounded up
1073 if( ( error = clSetKernelArg(kernels[j], 0, sizeof( gOutBuffer[j] ), &gOutBuffer[j] ) )) { LogBuildError(programs[j]); goto exit; }
1074 if( ( error = clSetKernelArg( kernels[j], 1, sizeof( gInBuffer ), &gInBuffer ) )) { LogBuildError(programs[j]); goto exit; }
1075 if( ( error = clSetKernelArg( kernels[j], 2, sizeof( gInBuffer2 ), &gInBuffer2 ) )) { LogBuildError(programs[j]); goto exit; }
1076 if( ( error = clSetKernelArg( kernels[j], 3, sizeof( gInBuffer3 ), &gInBuffer3 ) )) { LogBuildError(programs[j]); goto exit; }
1077
1078 double sum = 0.0;
1079 double bestTime = INFINITY;
1080 for( k = 0; k < PERF_LOOP_COUNT; k++ )
1081 {
1082 uint64_t startTime = GetTime();
1083 if( (error = clEnqueueNDRangeKernel(gQueue, kernels[j], 1, NULL, &localCount, NULL, 0, NULL, NULL)) )
1084 {
1085 vlog_error( "FAILED -- could not execute kernel\n" );
1086 goto exit;
1087 }
1088
1089 // Make sure OpenCL is done
1090 if( (error = clFinish(gQueue) ) )
1091 {
1092 vlog_error( "Error %d at clFinish\n", error );
1093 goto exit;
1094 }
1095
1096 uint64_t endTime = GetTime();
1097 double time = SubtractTime( endTime, startTime );
1098 sum += time;
1099 if( time < bestTime )
1100 bestTime = time;
1101 }
1102
1103 if( gReportAverageTimes )
1104 bestTime = sum / PERF_LOOP_COUNT;
1105 double clocksPerOp = bestTime * (double) gDeviceFrequency * gComputeDevices * gSimdSize * 1e6 / (bufferSize / sizeof( double ) );
1106 vlog_perf( clocksPerOp, LOWER_IS_BETTER, "clocks / element", "%sD%s", f->name, sizeNames[j] );
1107 }
1108 for( ; j < gMaxVectorSizeIndex; j++ )
1109 vlog( "\t -- " );
1110 }
1111
1112 if( ! gSkipCorrectnessTesting )
1113 vlog( "\t%8.2f @ {%a, %a, %a}", maxError, maxErrorVal, maxErrorVal2, maxErrorVal3 );
1114 vlog( "\n" );
1115
1116 exit:
1117 // Release
1118 for( k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++ )
1119 {
1120 clReleaseKernel(kernels[k]);
1121 clReleaseProgram(programs[k]);
1122 }
1123
1124 return error;
1125 }
1126
1127
1128
1129