1 //
2 // Copyright (c) 2021 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 "harness/typeWrappers.h"
19 #include <set>
20
21 namespace {
22
23 template <typename T, NonUniformVoteOp operation> struct VOTE
24 {
gen__anon20349c9e0111::VOTE25 static void gen(T *x, T *t, cl_int *m, const WorkGroupParams &test_params)
26 {
27 int i, ii, j, k, n;
28 int nw = test_params.local_workgroup_size;
29 int ns = test_params.subgroup_size;
30 int ng = test_params.global_workgroup_size;
31 uint32_t work_items_mask = test_params.work_items_mask;
32 int nj = (nw + ns - 1) / ns;
33 int non_uniform_size = ng % nw;
34 ng = ng / nw;
35 int last_subgroup_size = 0;
36 ii = 0;
37
38 log_info(" sub_group_%s%s... \n",
39 (operation == NonUniformVoteOp::elect) ? "" : "non_uniform_",
40 operation_names(operation));
41
42 log_info(" test params: global size = %d local size = %d subgroups "
43 "size = %d work item mask = 0x%x data type (%s)\n",
44 test_params.global_workgroup_size, nw, ns, work_items_mask,
45 TypeManager<T>::name());
46 if (non_uniform_size)
47 {
48 log_info(" non uniform work group size mode ON\n");
49 }
50 if (operation == NonUniformVoteOp::elect) return;
51
52 for (k = 0; k < ng; ++k)
53 { // for each work_group
54 if (non_uniform_size && k == ng - 1)
55 {
56 set_last_workgroup_params(non_uniform_size, nj, ns, nw,
57 last_subgroup_size);
58 }
59 for (j = 0; j < nj; ++j)
60 { // for each subgroup
61 ii = j * ns;
62 if (last_subgroup_size && j == nj - 1)
63 {
64 n = last_subgroup_size;
65 }
66 else
67 {
68 n = ii + ns > nw ? nw - ii : ns;
69 }
70 int e = genrand_int32(gMTdata) % 3;
71
72 for (i = 0; i < n; i++)
73 {
74 if (e == 2)
75 { // set once 0 and once 1 alternately
76 int value = i % 2;
77 set_value(t[ii + i], value);
78 }
79 else
80 { // set 0/1 for all work items in subgroup
81 set_value(t[ii + i], e);
82 }
83 }
84 }
85 // Now map into work group using map from device
86 for (j = 0; j < nw; ++j)
87 {
88 x[j] = t[j];
89 }
90 x += nw;
91 m += 4 * nw;
92 }
93 }
94
chk__anon20349c9e0111::VOTE95 static int chk(T *x, T *y, T *mx, T *my, cl_int *m,
96 const WorkGroupParams &test_params)
97 {
98 int ii, i, j, k, n;
99 int nw = test_params.local_workgroup_size;
100 int ns = test_params.subgroup_size;
101 int ng = test_params.global_workgroup_size;
102 uint32_t work_items_mask = test_params.work_items_mask;
103 int nj = (nw + ns - 1) / ns;
104 cl_int tr, rr;
105 int non_uniform_size = ng % nw;
106 ng = ng / nw;
107 if (non_uniform_size) ng++;
108 int last_subgroup_size = 0;
109
110 for (k = 0; k < ng; ++k)
111 { // for each work_group
112 if (non_uniform_size && k == ng - 1)
113 {
114 set_last_workgroup_params(non_uniform_size, nj, ns, nw,
115 last_subgroup_size);
116 }
117 for (j = 0; j < nw; ++j)
118 { // inside the work_group
119 mx[j] = x[j]; // read host inputs for work_group
120 my[j] = y[j]; // read device outputs for work_group
121 }
122
123 for (j = 0; j < nj; ++j)
124 { // for each subgroup
125 ii = j * ns;
126 if (last_subgroup_size && j == nj - 1)
127 {
128 n = last_subgroup_size;
129 }
130 else
131 {
132 n = ii + ns > nw ? nw - ii : ns;
133 }
134
135 rr = 0;
136 if (operation == NonUniformVoteOp::all
137 || operation == NonUniformVoteOp::all_equal)
138 tr = 1;
139 if (operation == NonUniformVoteOp::any) tr = 0;
140
141 std::set<int> active_work_items;
142 for (i = 0; i < n; ++i)
143 {
144 uint32_t check_work_item = 1 << (i % 32);
145 if (work_items_mask & check_work_item)
146 {
147 active_work_items.insert(i);
148 switch (operation)
149 {
150 case NonUniformVoteOp::elect: break;
151
152 case NonUniformVoteOp::all:
153 tr &=
154 !compare_ordered<T>(mx[ii + i], 0) ? 1 : 0;
155 break;
156 case NonUniformVoteOp::any:
157 tr |=
158 !compare_ordered<T>(mx[ii + i], 0) ? 1 : 0;
159 break;
160 case NonUniformVoteOp::all_equal:
161 tr &= compare_ordered<T>(
162 mx[ii + i],
163 mx[ii + *active_work_items.begin()])
164 ? 1
165 : 0;
166 break;
167 default:
168 log_error("Unknown operation\n");
169 return TEST_FAIL;
170 }
171 }
172 }
173 if (active_work_items.empty())
174 {
175 log_info(" no one workitem acitve... in workgroup id = %d "
176 "subgroup id = %d\n",
177 k, j);
178 }
179 else
180 {
181 auto lowest_active = active_work_items.begin();
182 for (const int &active_work_item : active_work_items)
183 {
184 i = active_work_item;
185 if (operation == NonUniformVoteOp::elect)
186 {
187 i == *lowest_active ? tr = 1 : tr = 0;
188 }
189
190 // normalize device values on host, non zero set 1.
191 rr = compare_ordered<T>(my[ii + i], 0) ? 0 : 1;
192
193 if (rr != tr)
194 {
195 log_error("ERROR: sub_group_%s() \n",
196 operation_names(operation));
197 log_error(
198 "mismatch for work item %d sub group %d in "
199 "work group %d. Expected: %d Obtained: %d\n",
200 i, j, k, tr, rr);
201 return TEST_FAIL;
202 }
203 }
204 }
205 }
206
207 x += nw;
208 y += nw;
209 m += 4 * nw;
210 }
211
212 log_info(" sub_group_%s%s... passed\n",
213 (operation == NonUniformVoteOp::elect) ? "" : "non_uniform_",
214 operation_names(operation));
215 return TEST_PASS;
216 }
217 };
218 static const char *elect_source = R"(
219 __kernel void test_elect(const __global Type *in, __global int4 *xy, __global Type *out) {
220 int gid = get_global_id(0);
221 XY(xy,gid);
222 uint elect_work_item = 1 << (get_sub_group_local_id() % 32);
223 if (elect_work_item & WORK_ITEMS_MASK){
224 out[gid] = sub_group_elect();
225 }
226 }
227 )";
228
229 static const char *non_uniform_any_source = R"(
230 __kernel void test_non_uniform_any(const __global Type *in, __global int4 *xy, __global Type *out) {
231 int gid = get_global_id(0);
232 XY(xy,gid);
233 uint elect_work_item = 1 << (get_sub_group_local_id() % 32);
234 if (elect_work_item & WORK_ITEMS_MASK){
235 out[gid] = sub_group_non_uniform_any(in[gid]);
236 }
237 }
238 )";
239
240 static const char *non_uniform_all_source = R"(
241 __kernel void test_non_uniform_all(const __global Type *in, __global int4 *xy, __global Type *out) {
242 int gid = get_global_id(0);
243 XY(xy,gid);
244 uint elect_work_item = 1 << (get_sub_group_local_id() % 32);
245 if (elect_work_item & WORK_ITEMS_MASK){
246 out[gid] = sub_group_non_uniform_all(in[gid]);
247 }
248 }
249 )";
250
251 static const char *non_uniform_all_equal_source = R"(
252 __kernel void test_non_uniform_all_equal(const __global Type *in, __global int4 *xy, __global Type *out) {
253 int gid = get_global_id(0);
254 XY(xy,gid);
255 uint elect_work_item = 1 << (get_sub_group_local_id() % 32);
256 if (elect_work_item & WORK_ITEMS_MASK){
257 out[gid] = sub_group_non_uniform_all_equal(in[gid]);
258 }
259 }
260 )";
261
run_vote_all_equal_for_type(RunTestForType rft)262 template <typename T> int run_vote_all_equal_for_type(RunTestForType rft)
263 {
264 int error = rft.run_impl<T, VOTE<T, NonUniformVoteOp::all_equal>>(
265 "test_non_uniform_all_equal", non_uniform_all_equal_source);
266 return error;
267 }
268 }
269
test_subgroup_functions_non_uniform_vote(cl_device_id device,cl_context context,cl_command_queue queue,int num_elements)270 int test_subgroup_functions_non_uniform_vote(cl_device_id device,
271 cl_context context,
272 cl_command_queue queue,
273 int num_elements)
274 {
275 std::vector<std::string> required_extensions = {
276 "cl_khr_subgroup_non_uniform_vote"
277 };
278
279 std::vector<uint32_t> masks{ 0xffffffff, 0x55aaaa55, 0x5555aaaa, 0xaaaa5555,
280 0x0f0ff0f0, 0x0f0f0f0f, 0xff0000ff, 0xff00ff00,
281 0x00ffff00, 0x80000000 };
282 constexpr size_t global_work_size = 170;
283 constexpr size_t local_work_size = 64;
284 WorkGroupParams test_params(global_work_size, local_work_size,
285 required_extensions, masks);
286 RunTestForType rft(device, context, queue, num_elements, test_params);
287
288 int error = run_vote_all_equal_for_type<cl_int>(rft);
289 error |= run_vote_all_equal_for_type<cl_uint>(rft);
290 error |= run_vote_all_equal_for_type<cl_long>(rft);
291 error |= run_vote_all_equal_for_type<cl_ulong>(rft);
292 error |= run_vote_all_equal_for_type<cl_float>(rft);
293 error |= run_vote_all_equal_for_type<cl_double>(rft);
294 error |= run_vote_all_equal_for_type<subgroups::cl_half>(rft);
295
296 error |= rft.run_impl<cl_int, VOTE<cl_int, NonUniformVoteOp::all>>(
297 "test_non_uniform_all", non_uniform_all_source);
298 error |= rft.run_impl<cl_int, VOTE<cl_int, NonUniformVoteOp::elect>>(
299 "test_elect", elect_source);
300 error |= rft.run_impl<cl_int, VOTE<cl_int, NonUniformVoteOp::any>>(
301 "test_non_uniform_any", non_uniform_any_source);
302 return error;
303 }
304