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 <climits>
22 #include <cstring>
23 
BuildKernel(const char * name,int vectorSize,cl_kernel * k,cl_program * p,bool relaxedMode)24 static int BuildKernel(const char *name, int vectorSize, cl_kernel *k,
25                        cl_program *p, bool relaxedMode)
26 {
27     const char *c[] = { "__kernel void math_kernel",
28                         sizeNames[vectorSize],
29                         "( __global float",
30                         sizeNames[vectorSize],
31                         "* out, __global int",
32                         sizeNames[vectorSize],
33                         "* out2, __global float",
34                         sizeNames[vectorSize],
35                         "* in )\n"
36                         "{\n"
37                         "   size_t i = get_global_id(0);\n"
38                         "   out[i] = ",
39                         name,
40                         "( in[i], out2 + i );\n"
41                         "}\n" };
42 
43     const char *c3[] = {
44         "__kernel void math_kernel",
45         sizeNames[vectorSize],
46         "( __global float* out, __global int* out2, __global float* in)\n"
47         "{\n"
48         "   size_t i = get_global_id(0);\n"
49         "   if( i + 1 < get_global_size(0) )\n"
50         "   {\n"
51         "       float3 f0 = vload3( 0, in + 3 * i );\n"
52         "       int3 iout = INT_MIN;\n"
53         "       f0 = ",
54         name,
55         "( f0, &iout );\n"
56         "       vstore3( f0, 0, out + 3*i );\n"
57         "       vstore3( iout, 0, out2 + 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         "       int3 iout = INT_MIN;\n"
65         "       float3 f0;\n"
66         "       switch( parity )\n"
67         "       {\n"
68         "           case 1:\n"
69         "               f0 = (float3)( in[3*i], NAN, NAN ); \n"
70         "               break;\n"
71         "           case 0:\n"
72         "               f0 = (float3)( in[3*i], in[3*i+1], NAN ); \n"
73         "               break;\n"
74         "       }\n"
75         "       f0 = ",
76         name,
77         "( f0, &iout );\n"
78         "       switch( parity )\n"
79         "       {\n"
80         "           case 0:\n"
81         "               out[3*i+1] = f0.y; \n"
82         "               out2[3*i+1] = iout.y; \n"
83         "               // fall through\n"
84         "           case 1:\n"
85         "               out[3*i] = f0.x; \n"
86         "               out2[3*i] = iout.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 MakeKernel(kern, (cl_uint)kernSize, testName, k, p, relaxedMode);
107 }
108 
109 typedef struct BuildKernelInfo
110 {
111     cl_uint offset; // the first vector size to build
112     cl_kernel *kernels;
113     cl_program *programs;
114     const char *nameInCode;
115     bool relaxedMode; // Whether to build with -cl-fast-relaxed-math.
116 } BuildKernelInfo;
117 
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)118 static cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
119 {
120     BuildKernelInfo *info = (BuildKernelInfo *)p;
121     cl_uint i = info->offset + job_id;
122     return BuildKernel(info->nameInCode, i, info->kernels + i,
123                        info->programs + i, info->relaxedMode);
124 }
125 
abs_cl_long(cl_long i)126 static cl_ulong abs_cl_long(cl_long i)
127 {
128     cl_long mask = i >> 63;
129     return (i ^ mask) - mask;
130 }
131 
TestFunc_FloatI_Float(const Func * f,MTdata d,bool relaxedMode)132 int TestFunc_FloatI_Float(const Func *f, MTdata d, bool relaxedMode)
133 {
134     int error;
135     cl_program programs[VECTOR_SIZE_COUNT];
136     cl_kernel kernels[VECTOR_SIZE_COUNT];
137     float maxError = 0.0f;
138     int64_t maxError2 = 0;
139     int ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
140     float maxErrorVal = 0.0f;
141     float maxErrorVal2 = 0.0f;
142     uint64_t step = getTestStep(sizeof(float), BUFFER_SIZE);
143     int scale = (int)((1ULL << 32) / (16 * BUFFER_SIZE / sizeof(float)) + 1);
144     cl_ulong maxiError;
145 
146     logFunctionInfo(f->name, sizeof(cl_float), relaxedMode);
147 
148     float float_ulps;
149     if (gIsEmbedded)
150         float_ulps = f->float_embedded_ulps;
151     else
152         float_ulps = f->float_ulps;
153 
154     maxiError = float_ulps == INFINITY ? CL_ULONG_MAX : 0;
155 
156     // Init the kernels
157     {
158         BuildKernelInfo build_info = { gMinVectorSizeIndex, kernels, programs,
159                                        f->nameInCode, relaxedMode };
160         if ((error = ThreadPool_Do(BuildKernelFn,
161                                    gMaxVectorSizeIndex - gMinVectorSizeIndex,
162                                    &build_info)))
163             return error;
164     }
165 
166     for (uint64_t i = 0; i < (1ULL << 32); i += step)
167     {
168         // Init input array
169         uint32_t *p = (uint32_t *)gIn;
170         if (gWimpyMode)
171         {
172             for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
173                 p[j] = (uint32_t)i + j * scale;
174         }
175         else
176         {
177             for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
178                 p[j] = (uint32_t)i + j;
179         }
180         if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0,
181                                           BUFFER_SIZE, gIn, 0, NULL, NULL)))
182         {
183             vlog_error("\n*** Error %d in clEnqueueWriteBuffer ***\n", error);
184             return error;
185         }
186 
187         // write garbage into output arrays
188         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
189         {
190             uint32_t pattern = 0xffffdead;
191             memset_pattern4(gOut[j], &pattern, BUFFER_SIZE);
192             if ((error =
193                      clEnqueueWriteBuffer(gQueue, gOutBuffer[j], CL_FALSE, 0,
194                                           BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
195             {
196                 vlog_error("\n*** Error %d in clEnqueueWriteBuffer2(%d) ***\n",
197                            error, j);
198                 goto exit;
199             }
200 
201             memset_pattern4(gOut2[j], &pattern, BUFFER_SIZE);
202             if ((error = clEnqueueWriteBuffer(gQueue, gOutBuffer2[j], CL_FALSE,
203                                               0, BUFFER_SIZE, gOut2[j], 0, NULL,
204                                               NULL)))
205             {
206                 vlog_error("\n*** Error %d in clEnqueueWriteBuffer2b(%d) ***\n",
207                            error, j);
208                 goto exit;
209             }
210         }
211 
212         // Run the kernels
213         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
214         {
215             size_t vectorSize = sizeValues[j] * sizeof(cl_float);
216             size_t localCount = (BUFFER_SIZE + vectorSize - 1) / vectorSize;
217             if ((error = clSetKernelArg(kernels[j], 0, sizeof(gOutBuffer[j]),
218                                         &gOutBuffer[j])))
219             {
220                 LogBuildError(programs[j]);
221                 goto exit;
222             }
223             if ((error = clSetKernelArg(kernels[j], 1, sizeof(gOutBuffer2[j]),
224                                         &gOutBuffer2[j])))
225             {
226                 LogBuildError(programs[j]);
227                 goto exit;
228             }
229             if ((error = clSetKernelArg(kernels[j], 2, sizeof(gInBuffer),
230                                         &gInBuffer)))
231             {
232                 LogBuildError(programs[j]);
233                 goto exit;
234             }
235 
236             if ((error =
237                      clEnqueueNDRangeKernel(gQueue, kernels[j], 1, NULL,
238                                             &localCount, NULL, 0, NULL, NULL)))
239             {
240                 vlog_error("FAILED -- could not execute kernel\n");
241                 goto exit;
242             }
243         }
244 
245         // Get that moving
246         if ((error = clFlush(gQueue))) vlog("clFlush failed\n");
247 
248         // Calculate the correctly rounded reference result
249         float *r = (float *)gOut_Ref;
250         int *r2 = (int *)gOut_Ref2;
251         float *s = (float *)gIn;
252         for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
253             r[j] = (float)f->func.f_fpI(s[j], r2 + j);
254 
255         // Read the data back
256         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
257         {
258             if ((error =
259                      clEnqueueReadBuffer(gQueue, gOutBuffer[j], CL_TRUE, 0,
260                                          BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
261             {
262                 vlog_error("ReadArray failed %d\n", error);
263                 goto exit;
264             }
265             if ((error =
266                      clEnqueueReadBuffer(gQueue, gOutBuffer2[j], CL_TRUE, 0,
267                                          BUFFER_SIZE, gOut2[j], 0, NULL, NULL)))
268             {
269                 vlog_error("ReadArray2 failed %d\n", error);
270                 goto exit;
271             }
272         }
273 
274         if (gSkipCorrectnessTesting) break;
275 
276         // Verify data
277         uint32_t *t = (uint32_t *)gOut_Ref;
278         int32_t *t2 = (int32_t *)gOut_Ref2;
279         for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
280         {
281             for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
282             {
283                 uint32_t *q = (uint32_t *)(gOut[k]);
284                 int32_t *q2 = (int32_t *)(gOut2[k]);
285 
286                 // If we aren't getting the correctly rounded result
287                 if (t[j] != q[j] || t2[j] != q2[j])
288                 {
289                     float test = ((float *)q)[j];
290                     int correct2 = INT_MIN;
291                     double correct = f->func.f_fpI(s[j], &correct2);
292                     float err = Ulp_Error(test, correct);
293                     cl_long iErr = (int64_t)q2[j] - (int64_t)correct2;
294                     int fail = !(fabsf(err) <= float_ulps
295                                  && abs_cl_long(iErr) <= maxiError);
296                     if (ftz)
297                     {
298                         // retry per section 6.5.3.2
299                         if (IsFloatResultSubnormal(correct, float_ulps))
300                         {
301                             fail = fail && !(test == 0.0f && iErr == 0);
302                             if (!fail) err = 0.0f;
303                         }
304 
305                         // retry per section 6.5.3.3
306                         if (IsFloatSubnormal(s[j]))
307                         {
308                             int correct5, correct6;
309                             double correct3 = f->func.f_fpI(0.0, &correct5);
310                             double correct4 = f->func.f_fpI(-0.0, &correct6);
311                             float err2 = Ulp_Error(test, correct3);
312                             float err3 = Ulp_Error(test, correct4);
313                             cl_long iErr2 =
314                                 (long long)q2[j] - (long long)correct5;
315                             cl_long iErr3 =
316                                 (long long)q2[j] - (long long)correct6;
317 
318                             // Did +0 work?
319                             if (fabsf(err2) <= float_ulps
320                                 && abs_cl_long(iErr2) <= maxiError)
321                             {
322                                 err = err2;
323                                 iErr = iErr2;
324                                 fail = 0;
325                             }
326                             // Did -0 work?
327                             else if (fabsf(err3) <= float_ulps
328                                      && abs_cl_long(iErr3) <= maxiError)
329                             {
330                                 err = err3;
331                                 iErr = iErr3;
332                                 fail = 0;
333                             }
334 
335                             // retry per section 6.5.3.4
336                             if (fail
337                                 && (IsFloatResultSubnormal(correct2, float_ulps)
338                                     || IsFloatResultSubnormal(correct3,
339                                                               float_ulps)))
340                             {
341                                 fail = fail
342                                     && !(test == 0.0f
343                                          && (abs_cl_long(iErr2) <= maxiError
344                                              || abs_cl_long(iErr3)
345                                                  <= maxiError));
346                                 if (!fail)
347                                 {
348                                     err = 0.0f;
349                                     iErr = 0;
350                                 }
351                             }
352                         }
353                     }
354                     if (fabsf(err) > maxError)
355                     {
356                         maxError = fabsf(err);
357                         maxErrorVal = s[j];
358                     }
359                     if (llabs(iErr) > maxError2)
360                     {
361                         maxError2 = llabs(iErr);
362                         maxErrorVal2 = s[j];
363                     }
364 
365                     if (fail)
366                     {
367                         vlog_error("\nERROR: %s%s: {%f, %d} ulp error at %a: "
368                                    "*{%a, %d} vs. {%a, %d}\n",
369                                    f->name, sizeNames[k], err, (int)iErr,
370                                    ((float *)gIn)[j], ((float *)gOut_Ref)[j],
371                                    ((int *)gOut_Ref2)[j], test, q2[j]);
372                         error = -1;
373                         goto exit;
374                     }
375                 }
376             }
377         }
378 
379         if (0 == (i & 0x0fffffff))
380         {
381             if (gVerboseBruteForce)
382             {
383                 vlog("base:%14u step:%10zu  bufferSize:%10zd \n", i, step,
384                      BUFFER_SIZE);
385             }
386             else
387             {
388                 vlog(".");
389             }
390             fflush(stdout);
391         }
392     }
393 
394     if (!gSkipCorrectnessTesting)
395     {
396         if (gWimpyMode)
397             vlog("Wimp pass");
398         else
399             vlog("passed");
400 
401         vlog("\t{%8.2f, %lld} @ {%a, %a}", maxError, maxError2, maxErrorVal,
402              maxErrorVal2);
403     }
404 
405     vlog("\n");
406 
407 exit:
408     // Release
409     for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
410     {
411         clReleaseKernel(kernels[k]);
412         clReleaseProgram(programs[k]);
413     }
414 
415     return error;
416 }
417