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 "procs.h"
17 #include "subhelpers.h"
18 #include "subgroup_common_kernels.h"
19 #include "subgroup_common_templates.h"
20 #include "harness/conversions.h"
21 #include "harness/typeWrappers.h"
22
23 namespace {
24 // Any/All test functions
25 template <NonUniformVoteOp operation> struct AA
26 {
log_test__anon4a8586e70111::AA27 static void log_test(const WorkGroupParams &test_params,
28 const char *extra_text)
29 {
30 log_info(" sub_group_%s...%s\n", operation_names(operation),
31 extra_text);
32 }
33
gen__anon4a8586e70111::AA34 static void gen(cl_int *x, cl_int *t, cl_int *m,
35 const WorkGroupParams &test_params)
36 {
37 int i, ii, j, k, n;
38 int ng = test_params.global_workgroup_size;
39 int nw = test_params.local_workgroup_size;
40 int ns = test_params.subgroup_size;
41 int nj = (nw + ns - 1) / ns;
42 int e;
43 ng = ng / nw;
44 ii = 0;
45 for (k = 0; k < ng; ++k)
46 {
47 for (j = 0; j < nj; ++j)
48 {
49 ii = j * ns;
50 n = ii + ns > nw ? nw - ii : ns;
51 e = (int)(genrand_int32(gMTdata) % 3);
52
53 // Initialize data matrix indexed by local id and sub group id
54 switch (e)
55 {
56 case 0: memset(&t[ii], 0, n * sizeof(cl_int)); break;
57 case 1:
58 memset(&t[ii], 0, n * sizeof(cl_int));
59 i = (int)(genrand_int32(gMTdata) % (cl_uint)n);
60 t[ii + i] = 41;
61 break;
62 case 2: memset(&t[ii], 0xff, n * sizeof(cl_int)); break;
63 }
64 }
65
66 // Now map into work group using map from device
67 for (j = 0; j < nw; ++j)
68 {
69 x[j] = t[j];
70 }
71
72 x += nw;
73 m += 4 * nw;
74 }
75 }
76
chk__anon4a8586e70111::AA77 static test_status chk(cl_int *x, cl_int *y, cl_int *mx, cl_int *my,
78 cl_int *m, const WorkGroupParams &test_params)
79 {
80 int ii, i, j, k, n;
81 int ng = test_params.global_workgroup_size;
82 int nw = test_params.local_workgroup_size;
83 int ns = test_params.subgroup_size;
84 int nj = (nw + ns - 1) / ns;
85 cl_int taa, raa;
86 ng = ng / nw;
87
88 for (k = 0; k < ng; ++k)
89 {
90 // Map to array indexed to array indexed by local ID and sub group
91 for (j = 0; j < nw; ++j)
92 {
93 mx[j] = x[j];
94 my[j] = y[j];
95 }
96
97 for (j = 0; j < nj; ++j)
98 {
99 ii = j * ns;
100 n = ii + ns > nw ? nw - ii : ns;
101
102 // Compute target
103 if (operation == NonUniformVoteOp::any)
104 {
105 taa = 0;
106 for (i = 0; i < n; ++i) taa |= mx[ii + i] != 0;
107 }
108
109 if (operation == NonUniformVoteOp::all)
110 {
111 taa = 1;
112 for (i = 0; i < n; ++i) taa &= mx[ii + i] != 0;
113 }
114
115 // Check result
116 for (i = 0; i < n; ++i)
117 {
118 raa = my[ii + i] != 0;
119 if (raa != taa)
120 {
121 log_error("ERROR: sub_group_%s mismatch for local id "
122 "%d in sub group %d in group %d\n",
123 operation_names(operation), i, j, k);
124 return TEST_FAIL;
125 }
126 }
127 }
128
129 x += nw;
130 y += nw;
131 m += 4 * nw;
132 }
133 return TEST_PASS;
134 }
135 };
136
137 template <typename T>
run_broadcast_scan_reduction_for_type(RunTestForType rft)138 int run_broadcast_scan_reduction_for_type(RunTestForType rft)
139 {
140 int error = rft.run_impl<T, BC<T, SubgroupsBroadcastOp::broadcast>>(
141 "sub_group_broadcast");
142 error |=
143 rft.run_impl<T, RED_NU<T, ArithmeticOp::add_>>("sub_group_reduce_add");
144 error |=
145 rft.run_impl<T, RED_NU<T, ArithmeticOp::max_>>("sub_group_reduce_max");
146 error |=
147 rft.run_impl<T, RED_NU<T, ArithmeticOp::min_>>("sub_group_reduce_min");
148 error |= rft.run_impl<T, SCIN_NU<T, ArithmeticOp::add_>>(
149 "sub_group_scan_inclusive_add");
150 error |= rft.run_impl<T, SCIN_NU<T, ArithmeticOp::max_>>(
151 "sub_group_scan_inclusive_max");
152 error |= rft.run_impl<T, SCIN_NU<T, ArithmeticOp::min_>>(
153 "sub_group_scan_inclusive_min");
154 error |= rft.run_impl<T, SCEX_NU<T, ArithmeticOp::add_>>(
155 "sub_group_scan_exclusive_add");
156 error |= rft.run_impl<T, SCEX_NU<T, ArithmeticOp::max_>>(
157 "sub_group_scan_exclusive_max");
158 error |= rft.run_impl<T, SCEX_NU<T, ArithmeticOp::min_>>(
159 "sub_group_scan_exclusive_min");
160 return error;
161 }
162
163 }
164 // Entry point from main
test_subgroup_functions(cl_device_id device,cl_context context,cl_command_queue queue,int num_elements,bool useCoreSubgroups)165 int test_subgroup_functions(cl_device_id device, cl_context context,
166 cl_command_queue queue, int num_elements,
167 bool useCoreSubgroups)
168 {
169 constexpr size_t global_work_size = 2000;
170 constexpr size_t local_work_size = 200;
171 WorkGroupParams test_params(global_work_size, local_work_size);
172 test_params.save_kernel_source(sub_group_reduction_scan_source);
173 test_params.save_kernel_source(sub_group_generic_source,
174 "sub_group_broadcast");
175
176 RunTestForType rft(device, context, queue, num_elements, test_params);
177 int error =
178 rft.run_impl<cl_int, AA<NonUniformVoteOp::any>>("sub_group_any");
179 error |= rft.run_impl<cl_int, AA<NonUniformVoteOp::all>>("sub_group_all");
180 error |= run_broadcast_scan_reduction_for_type<cl_int>(rft);
181 error |= run_broadcast_scan_reduction_for_type<cl_uint>(rft);
182 error |= run_broadcast_scan_reduction_for_type<cl_long>(rft);
183 error |= run_broadcast_scan_reduction_for_type<cl_ulong>(rft);
184 error |= run_broadcast_scan_reduction_for_type<cl_float>(rft);
185 error |= run_broadcast_scan_reduction_for_type<cl_double>(rft);
186 error |= run_broadcast_scan_reduction_for_type<subgroups::cl_half>(rft);
187 return error;
188 }
189
test_subgroup_functions_core(cl_device_id device,cl_context context,cl_command_queue queue,int num_elements)190 int test_subgroup_functions_core(cl_device_id device, cl_context context,
191 cl_command_queue queue, int num_elements)
192 {
193 return test_subgroup_functions(device, context, queue, num_elements, true);
194 }
195
test_subgroup_functions_ext(cl_device_id device,cl_context context,cl_command_queue queue,int num_elements)196 int test_subgroup_functions_ext(cl_device_id device, cl_context context,
197 cl_command_queue queue, int num_elements)
198 {
199 bool hasExtension = is_extension_available(device, "cl_khr_subgroups");
200
201 if (!hasExtension)
202 {
203 log_info(
204 "Device does not support 'cl_khr_subgroups'. Skipping the test.\n");
205 return TEST_SKIPPED_ITSELF;
206 }
207 return test_subgroup_functions(device, context, queue, num_elements, false);
208 }
209