xref: /aosp_15_r20/external/OpenCL-CTS/test_conformance/basic/test_async_copy.cpp (revision 6467f958c7de8070b317fc65bcb0f6472e388d82)
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 "harness/compat.h"
17 
18 #include <stdio.h>
19 #include <stdlib.h>
20 #include <string.h>
21 #include <sys/types.h>
22 #include <sys/stat.h>
23 #include <vector>
24 
25 #include "procs.h"
26 #include "harness/conversions.h"
27 
28 static const char *async_global_to_local_kernel =
29 "%s\n" // optional pragma string
30 "__kernel void test_fn( const __global %s *src, __global %s *dst, __local %s *localBuffer, int copiesPerWorkgroup, int copiesPerWorkItem )\n"
31 "{\n"
32 " int i;\n"
33 // Zero the local storage first
34 " for(i=0; i<copiesPerWorkItem; i++)\n"
35 "     localBuffer[ get_local_id( 0 )*copiesPerWorkItem+i ] = (%s)(%s)0;\n"
36 // Do this to verify all kernels are done zeroing the local buffer before we try the copy
37 "    barrier( CLK_LOCAL_MEM_FENCE );\n"
38 "    event_t event;\n"
39 "    event = async_work_group_copy( (__local %s*)localBuffer, (__global const %s*)(src+copiesPerWorkgroup*get_group_id(0)), (size_t)copiesPerWorkgroup, 0 );\n"
40 // Wait for the copy to complete, then verify by manually copying to the dest
41 "    wait_group_events( 1, &event );\n"
42 " for(i=0; i<copiesPerWorkItem; i++)\n"
43 "  dst[ get_global_id( 0 )*copiesPerWorkItem+i ] = localBuffer[ get_local_id( 0 )*copiesPerWorkItem+i ];\n"
44 "}\n" ;
45 
46 static const char *async_local_to_global_kernel =
47 "%s\n" // optional pragma string
48 "__kernel void test_fn( const __global %s *src, __global %s *dst, __local %s *localBuffer, int copiesPerWorkgroup, int copiesPerWorkItem )\n"
49 "{\n"
50 " int i;\n"
51 // Zero the local storage first
52 " for(i=0; i<copiesPerWorkItem; i++)\n"
53 "  localBuffer[ get_local_id( 0 )*copiesPerWorkItem+i ] = (%s)(%s)0;\n"
54 // Do this to verify all kernels are done zeroing the local buffer before we try the copy
55 "    barrier( CLK_LOCAL_MEM_FENCE );\n"
56 " for(i=0; i<copiesPerWorkItem; i++)\n"
57 "  localBuffer[ get_local_id( 0 )*copiesPerWorkItem+i ] = src[ get_global_id( 0 )*copiesPerWorkItem+i ];\n"
58 // Do this to verify all kernels are done copying to the local buffer before we try the copy
59 "    barrier( CLK_LOCAL_MEM_FENCE );\n"
60 "    event_t event;\n"
61 "    event = async_work_group_copy((__global %s*)(dst+copiesPerWorkgroup*get_group_id(0)), (__local const %s*)localBuffer, (size_t)copiesPerWorkgroup, 0 );\n"
62 "    wait_group_events( 1, &event );\n"
63 "}\n" ;
64 
65 
66 static const char *prefetch_kernel =
67 "%s\n" // optional pragma string
68 "__kernel void test_fn( const __global %s *src, __global %s *dst, __local %s *localBuffer, int copiesPerWorkgroup, int copiesPerWorkItem )\n"
69 "{\n"
70 " // Ignore this: %s%s%s\n"
71 " int i;\n"
72 " prefetch( (const __global %s*)(src+copiesPerWorkItem*get_global_id(0)), copiesPerWorkItem);\n"
73 " for(i=0; i<copiesPerWorkItem; i++)\n"
74 "  dst[ get_global_id( 0 )*copiesPerWorkItem+i ] = src[ get_global_id( 0 )*copiesPerWorkItem+i ];\n"
75 "}\n" ;
76 
77 
78 
test_copy(cl_device_id deviceID,cl_context context,cl_command_queue queue,const char * kernelCode,ExplicitType vecType,int vecSize)79 int test_copy(cl_device_id deviceID, cl_context context, cl_command_queue queue, const char *kernelCode,
80               ExplicitType vecType, int vecSize
81               )
82 {
83     int error;
84     clProgramWrapper program;
85     clKernelWrapper kernel;
86     clMemWrapper streams[ 2 ];
87     size_t threads[ 1 ], localThreads[ 1 ];
88     MTdataHolder d(gRandomSeed);
89     char vecNameString[64]; vecNameString[0] = 0;
90     if (vecSize == 1)
91         sprintf(vecNameString, "%s", get_explicit_type_name(vecType));
92     else
93         sprintf(vecNameString, "%s%d", get_explicit_type_name(vecType), vecSize);
94 
95 
96     size_t elementSize = get_explicit_type_size(vecType)*vecSize;
97     log_info("Testing %s\n", vecNameString);
98 
99     cl_long max_local_mem_size;
100     error = clGetDeviceInfo(deviceID, CL_DEVICE_LOCAL_MEM_SIZE, sizeof(max_local_mem_size), &max_local_mem_size, NULL);
101     test_error( error, "clGetDeviceInfo for CL_DEVICE_LOCAL_MEM_SIZE failed.");
102 
103     unsigned int num_of_compute_devices;
104     error = clGetDeviceInfo(deviceID, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(num_of_compute_devices), &num_of_compute_devices, NULL);
105     test_error( error, "clGetDeviceInfo for CL_DEVICE_MAX_COMPUTE_UNITS failed.");
106 
107     char programSource[4096]; programSource[0]=0;
108     char *programPtr;
109 
110     std::string extStr = "";
111     if (vecType == kDouble)
112         extStr = "#pragma OPENCL EXTENSION cl_khr_fp64 : enable";
113     else if (vecType == kHalf)
114         extStr = "#pragma OPENCL EXTENSION cl_khr_fp16 : enable";
115 
116     sprintf(programSource, kernelCode, extStr.c_str(), vecNameString,
117             vecNameString, vecNameString, vecNameString,
118             get_explicit_type_name(vecType), vecNameString, vecNameString);
119     //log_info("program: %s\n", programSource);
120     programPtr = programSource;
121 
122     error = create_single_kernel_helper( context, &program, &kernel, 1, (const char **)&programPtr, "test_fn" );
123     test_error( error, "Unable to create testing kernel" );
124 
125     size_t max_workgroup_size;
126     error = clGetKernelWorkGroupInfo(kernel, deviceID, CL_KERNEL_WORK_GROUP_SIZE, sizeof(max_workgroup_size), &max_workgroup_size, NULL);
127     test_error (error, "clGetKernelWorkGroupInfo failed for CL_KERNEL_WORK_GROUP_SIZE.");
128 
129     size_t max_local_workgroup_size[3];
130     error = clGetDeviceInfo(deviceID, CL_DEVICE_MAX_WORK_ITEM_SIZES, sizeof(max_local_workgroup_size), max_local_workgroup_size, NULL);
131     test_error (error, "clGetDeviceInfo failed for CL_DEVICE_MAX_WORK_ITEM_SIZES");
132 
133     // Pick the minimum of the device and the kernel
134     if (max_workgroup_size > max_local_workgroup_size[0])
135         max_workgroup_size = max_local_workgroup_size[0];
136 
137     size_t numberOfCopiesPerWorkitem = 13;
138     elementSize = get_explicit_type_size(vecType)* ((vecSize == 3) ? 4 : vecSize);
139     size_t localStorageSpacePerWorkitem = numberOfCopiesPerWorkitem*elementSize;
140     size_t maxLocalWorkgroupSize = (((int)max_local_mem_size/2)/localStorageSpacePerWorkitem);
141 
142     // Calculation can return 0 on embedded devices due to 1KB local mem limit
143     if(maxLocalWorkgroupSize == 0)
144     {
145         maxLocalWorkgroupSize = 1;
146     }
147 
148     size_t localWorkgroupSize = maxLocalWorkgroupSize;
149     if (maxLocalWorkgroupSize > max_workgroup_size)
150         localWorkgroupSize = max_workgroup_size;
151 
152     size_t localBufferSize = localWorkgroupSize*elementSize*numberOfCopiesPerWorkitem;
153     size_t numberOfLocalWorkgroups = 1111;
154     size_t globalBufferSize = numberOfLocalWorkgroups*localBufferSize;
155     size_t globalWorkgroupSize = numberOfLocalWorkgroups*localWorkgroupSize;
156 
157     std::vector<unsigned char> inBuffer(globalBufferSize);
158     std::vector<unsigned char> outBuffer(globalBufferSize);
159 
160     outBuffer.assign(globalBufferSize, 0);
161 
162     cl_int copiesPerWorkItemInt, copiesPerWorkgroup;
163     copiesPerWorkItemInt = (int)numberOfCopiesPerWorkitem;
164     copiesPerWorkgroup = (int)(numberOfCopiesPerWorkitem*localWorkgroupSize);
165 
166     log_info("Global: %d, local %d, local buffer %db, global buffer %db, each work group will copy %d elements and each work item item will copy %d elements.\n",
167              (int) globalWorkgroupSize, (int)localWorkgroupSize, (int)localBufferSize, (int)globalBufferSize, copiesPerWorkgroup, copiesPerWorkItemInt);
168 
169     threads[0] = globalWorkgroupSize;
170     localThreads[0] = localWorkgroupSize;
171 
172     generate_random_data(vecType,
173                          globalBufferSize / get_explicit_type_size(vecType), d,
174                          &inBuffer.front());
175 
176     streams[0] = clCreateBuffer(context, CL_MEM_COPY_HOST_PTR, globalBufferSize,
177                                 &inBuffer.front(), &error);
178     test_error( error, "Unable to create input buffer" );
179     streams[1] = clCreateBuffer(context, CL_MEM_COPY_HOST_PTR, globalBufferSize,
180                                 &outBuffer.front(), &error);
181     test_error( error, "Unable to create output buffer" );
182 
183     error = clSetKernelArg( kernel, 0, sizeof( streams[ 0 ] ), &streams[ 0 ] );
184     test_error( error, "Unable to set kernel argument" );
185     error = clSetKernelArg( kernel, 1, sizeof( streams[ 1 ] ), &streams[ 1 ] );
186     test_error( error, "Unable to set kernel argument" );
187     error = clSetKernelArg( kernel, 2, localBufferSize, NULL );
188     test_error( error, "Unable to set kernel argument" );
189     error = clSetKernelArg( kernel, 3, sizeof(copiesPerWorkgroup), &copiesPerWorkgroup );
190     test_error( error, "Unable to set kernel argument" );
191     error = clSetKernelArg( kernel, 4, sizeof(copiesPerWorkItemInt), &copiesPerWorkItemInt );
192     test_error( error, "Unable to set kernel argument" );
193 
194     // Enqueue
195     error = clEnqueueNDRangeKernel( queue, kernel, 1, NULL, threads, localThreads, 0, NULL, NULL );
196     test_error( error, "Unable to queue kernel" );
197 
198     // Read
199     error = clEnqueueReadBuffer(queue, streams[1], CL_TRUE, 0, globalBufferSize,
200                                 &outBuffer.front(), 0, NULL, NULL);
201     test_error( error, "Unable to read results" );
202 
203     // Verify
204     int failuresPrinted = 0;
205     if (memcmp(&inBuffer.front(), &outBuffer.front(), globalBufferSize) != 0)
206     {
207         size_t typeSize = get_explicit_type_size(vecType)* vecSize;
208         unsigned char *inchar = static_cast<unsigned char *>(&inBuffer.front());
209         unsigned char *outchar =
210             static_cast<unsigned char *>(&outBuffer.front());
211         for (int i=0; i< (int)globalBufferSize; i+=(int)elementSize) {
212             if (memcmp( ((char *)inchar)+i, ((char *)outchar)+i, typeSize) != 0 )
213             {
214                 char values[4096];
215                 values[0] = 0;
216                 if ( failuresPrinted == 0 ) {
217                     // Print first failure message
218                     log_error( "ERROR: Results of copy did not validate!\n" );
219                 }
220                 sprintf(values + strlen( values), "%d -> [", i);
221                 for (int j=0; j<(int)elementSize; j++)
222                     sprintf(values + strlen( values), "%2x ", inchar[i+j]);
223                 sprintf(values + strlen(values), "] != [");
224                 for (int j=0; j<(int)elementSize; j++)
225                     sprintf(values + strlen( values), "%2x ", outchar[i+j]);
226                 sprintf(values + strlen(values), "]");
227                 log_error("%s\n", values);
228                 failuresPrinted++;
229             }
230 
231             if (failuresPrinted > 5) {
232                 log_error("Not printing further failures...\n");
233                 break;
234             }
235         }
236     }
237 
238     return failuresPrinted ? -1 : 0;
239 }
240 
test_copy_all_types(cl_device_id deviceID,cl_context context,cl_command_queue queue,const char * kernelCode)241 int test_copy_all_types(cl_device_id deviceID, cl_context context, cl_command_queue queue, const char *kernelCode) {
242     const std::vector<ExplicitType> vecType = { kChar,  kUChar, kShort, kUShort,
243                                                 kInt,   kUInt,  kLong,  kULong,
244                                                 kFloat, kHalf,  kDouble };
245     unsigned int vecSizes[] = { 1, 2, 3, 4, 8, 16, 0 };
246     unsigned int size, typeIndex;
247 
248     int errors = 0;
249 
250     bool fp16Support = is_extension_available(deviceID, "cl_khr_fp16");
251     bool fp64Support = is_extension_available(deviceID, "cl_khr_fp64");
252 
253     for (typeIndex = 0; typeIndex < vecType.size(); typeIndex++)
254     {
255         if (( vecType[ typeIndex ] == kLong || vecType[ typeIndex ] == kULong ) && !gHasLong )
256             continue;
257         else if (vecType[typeIndex] == kDouble && !fp64Support)
258             continue;
259         else if (vecType[typeIndex] == kHalf && !fp16Support)
260             continue;
261 
262         for( size = 0; vecSizes[ size ] != 0; size++ )
263         {
264             if (test_copy( deviceID, context, queue, kernelCode, vecType[typeIndex],vecSizes[size] )) {
265                 errors++;
266             }
267         }
268     }
269     if (errors)
270         return -1;
271     return 0;
272 }
273 
test_async_copy_global_to_local(cl_device_id deviceID,cl_context context,cl_command_queue queue,int num_elements)274 int test_async_copy_global_to_local(cl_device_id deviceID, cl_context context, cl_command_queue queue, int num_elements)
275 {
276     return test_copy_all_types( deviceID, context, queue, async_global_to_local_kernel );
277 }
278 
test_async_copy_local_to_global(cl_device_id deviceID,cl_context context,cl_command_queue queue,int num_elements)279 int test_async_copy_local_to_global(cl_device_id deviceID, cl_context context, cl_command_queue queue, int num_elements)
280 {
281     return test_copy_all_types( deviceID, context, queue, async_local_to_global_kernel );
282 }
283 
test_prefetch(cl_device_id deviceID,cl_context context,cl_command_queue queue,int num_elements)284 int test_prefetch(cl_device_id deviceID, cl_context context, cl_command_queue queue, int num_elements)
285 {
286     return test_copy_all_types( deviceID, context, queue, prefetch_kernel );
287 }
288 
289