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 
BuildKernel(const char * name,int vectorSize,cl_kernel * k,cl_program * p,bool relaxedMode)23 static int BuildKernel(const char *name, int vectorSize, cl_kernel *k,
24                        cl_program *p, bool relaxedMode)
25 {
26     const char *c[] = { "__kernel void math_kernel",
27                         sizeNames[vectorSize],
28                         "( __global float",
29                         sizeNames[vectorSize],
30                         "* out, __global float",
31                         sizeNames[vectorSize],
32                         "* in1, __global float",
33                         sizeNames[vectorSize],
34                         "* in2,  __global float",
35                         sizeNames[vectorSize],
36                         "* in3 )\n"
37                         "{\n"
38                         "   size_t i = get_global_id(0);\n"
39                         "   out[i] = ",
40                         name,
41                         "( in1[i], in2[i], in3[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, "
48         "__global float* in3)\n"
49         "{\n"
50         "   size_t i = get_global_id(0);\n"
51         "   if( i + 1 < get_global_size(0) )\n"
52         "   {\n"
53         "       float3 f0 = vload3( 0, in + 3 * i );\n"
54         "       float3 f1 = vload3( 0, in2 + 3 * i );\n"
55         "       float3 f2 = vload3( 0, in3 + 3 * i );\n"
56         "       f0 = ",
57         name,
58         "( f0, f1, f2 );\n"
59         "       vstore3( f0, 0, out + 3*i );\n"
60         "   }\n"
61         "   else\n"
62         "   {\n"
63         "       size_t parity = i & 1;   // Figure out how many elements are "
64         "left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two "
65         "buffer size \n"
66         "       float3 f0;\n"
67         "       float3 f1;\n"
68         "       float3 f2;\n"
69         "       switch( parity )\n"
70         "       {\n"
71         "           case 1:\n"
72         "               f0 = (float3)( in[3*i], NAN, NAN ); \n"
73         "               f1 = (float3)( in2[3*i], NAN, NAN ); \n"
74         "               f2 = (float3)( in3[3*i], NAN, NAN ); \n"
75         "               break;\n"
76         "           case 0:\n"
77         "               f0 = (float3)( in[3*i], in[3*i+1], NAN ); \n"
78         "               f1 = (float3)( in2[3*i], in2[3*i+1], NAN ); \n"
79         "               f2 = (float3)( in3[3*i], in3[3*i+1], NAN ); \n"
80         "               break;\n"
81         "       }\n"
82         "       f0 = ",
83         name,
84         "( f0, f1, f2 );\n"
85         "       switch( parity )\n"
86         "       {\n"
87         "           case 0:\n"
88         "               out[3*i+1] = f0.y; \n"
89         "               // fall through\n"
90         "           case 1:\n"
91         "               out[3*i] = f0.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 typedef struct BuildKernelInfo
115 {
116     cl_uint offset; // the first vector size to build
117     cl_kernel *kernels;
118     cl_program *programs;
119     const char *nameInCode;
120     bool relaxedMode; // Whether to build with -cl-fast-relaxed-math.
121 } BuildKernelInfo;
122 
BuildKernelFn(cl_uint job_id,cl_uint thread_id UNUSED,void * p)123 static cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
124 {
125     BuildKernelInfo *info = (BuildKernelInfo *)p;
126     cl_uint i = info->offset + job_id;
127     return BuildKernel(info->nameInCode, i, info->kernels + i,
128                        info->programs + i, info->relaxedMode);
129 }
130 
TestFunc_mad_Float(const Func * f,MTdata d,bool relaxedMode)131 int TestFunc_mad_Float(const Func *f, MTdata d, bool relaxedMode)
132 {
133     int error;
134 
135     logFunctionInfo(f->name, sizeof(cl_float), relaxedMode);
136 
137     cl_program programs[VECTOR_SIZE_COUNT];
138     cl_kernel kernels[VECTOR_SIZE_COUNT];
139     float maxError = 0.0f;
140     float maxErrorVal = 0.0f;
141     float maxErrorVal2 = 0.0f;
142     float maxErrorVal3 = 0.0f;
143     uint64_t step = getTestStep(sizeof(float), BUFFER_SIZE);
144 
145     // Init the kernels
146     {
147         BuildKernelInfo build_info = { gMinVectorSizeIndex, kernels, programs,
148                                        f->nameInCode, relaxedMode };
149         if ((error = ThreadPool_Do(BuildKernelFn,
150                                    gMaxVectorSizeIndex - gMinVectorSizeIndex,
151                                    &build_info)))
152             return error;
153     }
154 
155     for (uint64_t i = 0; i < (1ULL << 32); i += step)
156     {
157         // Init input array
158         cl_uint *p = (cl_uint *)gIn;
159         cl_uint *p2 = (cl_uint *)gIn2;
160         cl_uint *p3 = (cl_uint *)gIn3;
161         for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
162         {
163             p[j] = genrand_int32(d);
164             p2[j] = genrand_int32(d);
165             p3[j] = genrand_int32(d);
166         }
167 
168         if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0,
169                                           BUFFER_SIZE, gIn, 0, NULL, NULL)))
170         {
171             vlog_error("\n*** Error %d in clEnqueueWriteBuffer ***\n", error);
172             return error;
173         }
174 
175         if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer2, CL_FALSE, 0,
176                                           BUFFER_SIZE, gIn2, 0, NULL, NULL)))
177         {
178             vlog_error("\n*** Error %d in clEnqueueWriteBuffer2 ***\n", error);
179             return error;
180         }
181 
182         if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer3, CL_FALSE, 0,
183                                           BUFFER_SIZE, gIn3, 0, NULL, NULL)))
184         {
185             vlog_error("\n*** Error %d in clEnqueueWriteBuffer3 ***\n", error);
186             return error;
187         }
188 
189         // write garbage into output arrays
190         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
191         {
192             uint32_t pattern = 0xffffdead;
193             memset_pattern4(gOut[j], &pattern, BUFFER_SIZE);
194             if ((error =
195                      clEnqueueWriteBuffer(gQueue, gOutBuffer[j], CL_FALSE, 0,
196                                           BUFFER_SIZE, gOut[j], 0, NULL, NULL)))
197             {
198                 vlog_error("\n*** Error %d in clEnqueueWriteBuffer2(%d) ***\n",
199                            error, j);
200                 goto exit;
201             }
202         }
203 
204         // Run the kernels
205         for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
206         {
207             size_t vectorSize = sizeof(cl_float) * sizeValues[j];
208             size_t localCount = (BUFFER_SIZE + vectorSize - 1)
209                 / vectorSize; // BUFFER_SIZE / vectorSize  rounded up
210             if ((error = clSetKernelArg(kernels[j], 0, sizeof(gOutBuffer[j]),
211                                         &gOutBuffer[j])))
212             {
213                 LogBuildError(programs[j]);
214                 goto exit;
215             }
216             if ((error = clSetKernelArg(kernels[j], 1, sizeof(gInBuffer),
217                                         &gInBuffer)))
218             {
219                 LogBuildError(programs[j]);
220                 goto exit;
221             }
222             if ((error = clSetKernelArg(kernels[j], 2, sizeof(gInBuffer2),
223                                         &gInBuffer2)))
224             {
225                 LogBuildError(programs[j]);
226                 goto exit;
227             }
228             if ((error = clSetKernelArg(kernels[j], 3, sizeof(gInBuffer3),
229                                         &gInBuffer3)))
230             {
231                 LogBuildError(programs[j]);
232                 goto exit;
233             }
234 
235             if ((error =
236                      clEnqueueNDRangeKernel(gQueue, kernels[j], 1, NULL,
237                                             &localCount, NULL, 0, NULL, NULL)))
238             {
239                 vlog_error("FAILED -- could not execute kernel\n");
240                 goto exit;
241             }
242         }
243 
244         // Get that moving
245         if ((error = clFlush(gQueue))) vlog("clFlush failed\n");
246 
247         // Calculate the correctly rounded reference result
248         float *r = (float *)gOut_Ref;
249         float *s = (float *)gIn;
250         float *s2 = (float *)gIn2;
251         float *s3 = (float *)gIn3;
252         for (size_t j = 0; j < BUFFER_SIZE / sizeof(float); j++)
253             r[j] = (float)f->func.f_fff(s[j], s2[j], s3[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         }
266 
267         if (gSkipCorrectnessTesting) break;
268 
269         // Verify data -- No verification possible.
270         // MAD is a random number generator.
271         if (0 == (i & 0x0fffffff))
272         {
273             vlog(".");
274             fflush(stdout);
275         }
276     }
277 
278     if (!gSkipCorrectnessTesting)
279     {
280         if (gWimpyMode)
281             vlog("Wimp pass");
282         else
283             vlog("passed");
284 
285         vlog("\t%8.2f @ {%a, %a, %a}", maxError, maxErrorVal, maxErrorVal2,
286              maxErrorVal3);
287     }
288 
289     vlog("\n");
290 
291 exit:
292     // Release
293     for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
294     {
295         clReleaseKernel(kernels[k]);
296         clReleaseProgram(programs[k]);
297     }
298 
299     return error;
300 }
301