xref: /aosp_15_r20/external/OpenCL-CTS/test_conformance/basic/test_loop.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 
24 #include <vector>
25 
26 #include "procs.h"
27 
28 namespace {
29 const char *loop_kernel_code = R"(
30 __kernel void test_loop(__global int *src, __global int *loopindx, __global int *loopcnt, __global int *dst)
31 {
32     int  tid = get_global_id(0);
33     int  n = get_global_size(0);
34     int  i, j;
35 
36     dst[tid] = 0;
37     for (i=0, j=loopindx[tid]; i<loopcnt[tid]; i++, j++)
38     {
39         if (j >= n)
40             j = 0;
41         dst[tid] += src[j];
42     }
43 }
44 )";
45 
46 
verify_loop(std::vector<cl_int> inptr,std::vector<cl_int> loopindx,std::vector<cl_int> loopcnt,std::vector<cl_int> outptr,int n)47 int verify_loop(std::vector<cl_int> inptr, std::vector<cl_int> loopindx,
48                 std::vector<cl_int> loopcnt, std::vector<cl_int> outptr, int n)
49 {
50     for (int i = 0; i < n; i++)
51     {
52         int r = 0;
53         for (int j = 0, k = loopindx[i]; j < loopcnt[i]; j++, k++)
54         {
55             if (k >= n) k = 0;
56             r += inptr[k];
57         }
58 
59         if (r != outptr[i])
60         {
61             log_error("LOOP test failed: %d found, expected %d\n", outptr[i],
62                       r);
63             return -1;
64         }
65     }
66 
67     log_info("LOOP test passed\n");
68     return 0;
69 }
70 }
test_loop(cl_device_id device,cl_context context,cl_command_queue queue,int num_elements)71 int test_loop(cl_device_id device, cl_context context, cl_command_queue queue,
72               int num_elements)
73 {
74     clMemWrapper streams[4];
75     clProgramWrapper program;
76     clKernelWrapper kernel;
77     int err;
78 
79     size_t length = sizeof(cl_int) * num_elements;
80     std::vector<cl_int> input(length);
81     std::vector<cl_int> loop_indx(length);
82     std::vector<cl_int> loop_cnt(length);
83     std::vector<cl_int> output(length);
84 
85     for (auto &stream : streams)
86     {
87         stream =
88             clCreateBuffer(context, CL_MEM_READ_WRITE, length, nullptr, &err);
89         test_error(err, "clCreateBuffer failed.");
90     }
91 
92     RandomSeed seed(gRandomSeed);
93     for (int i = 0; i < num_elements; i++)
94     {
95         input[i] = static_cast<int>(genrand_int32(seed));
96         loop_indx[i] =
97             static_cast<int>(get_random_float(0, num_elements - 1, seed));
98         loop_cnt[i] =
99             static_cast<int>(get_random_float(0, num_elements / 32, seed));
100     };
101 
102     err = clEnqueueWriteBuffer(queue, streams[0], CL_TRUE, 0, length,
103                                input.data(), 0, nullptr, nullptr);
104     test_error(err, "clEnqueueWriteBuffer failed.");
105     err = clEnqueueWriteBuffer(queue, streams[1], CL_TRUE, 0, length,
106                                loop_indx.data(), 0, nullptr, nullptr);
107     test_error(err, "clEnqueueWriteBuffer failed.");
108     err = clEnqueueWriteBuffer(queue, streams[2], CL_TRUE, 0, length,
109                                loop_cnt.data(), 0, nullptr, nullptr);
110     test_error(err, "clEnqueueWriteBuffer failed.");
111 
112     err = create_single_kernel_helper(context, &program, &kernel, 1,
113                                       &loop_kernel_code, "test_loop");
114     test_error(err, "create_single_kernel_helper failed.");
115 
116     for (int i = 0; i < ARRAY_SIZE(streams); i++)
117     {
118         err = clSetKernelArg(kernel, i, sizeof streams[i], &streams[i]);
119         test_error(err, "clSetKernelArgs failed\n");
120     }
121 
122     size_t threads[] = { (size_t)num_elements };
123     err = clEnqueueNDRangeKernel(queue, kernel, 1, nullptr, threads, nullptr, 0,
124                                  nullptr, nullptr);
125     test_error(err, "clEnqueueNDRangeKernel failed\n");
126 
127     err = clEnqueueReadBuffer(queue, streams[3], CL_TRUE, 0, length,
128                               output.data(), 0, nullptr, nullptr);
129     test_error(err, "clEnqueueReadBuffer failed\n");
130 
131     err = verify_loop(input, loop_indx, loop_cnt, output, num_elements);
132 
133 
134     return err;
135 }
136