1 /*
2 * Copyright (c) 2017-2022 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "arm_compute/runtime/OMP/OMPScheduler.h"
25
26 #include "arm_compute/core/CPP/ICPPKernel.h"
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/Helpers.h"
29 #include "arm_compute/core/Utils.h"
30 #include <omp.h>
31
32 namespace arm_compute
33 {
OMPScheduler()34 OMPScheduler::OMPScheduler() // NOLINT
35 : _num_threads(omp_get_max_threads())
36 {
37 }
38
num_threads() const39 unsigned int OMPScheduler::num_threads() const
40 {
41 return _num_threads;
42 }
43
set_num_threads(unsigned int num_threads)44 void OMPScheduler::set_num_threads(unsigned int num_threads)
45 {
46 const unsigned int num_cores = omp_get_max_threads();
47 _num_threads = (num_threads == 0) ? num_cores : num_threads;
48 }
49
schedule(ICPPKernel * kernel,const Hints & hints)50 void OMPScheduler::schedule(ICPPKernel *kernel, const Hints &hints)
51 {
52 ITensorPack tensors;
53 schedule_common(kernel, hints, kernel->window(), tensors);
54 }
55
schedule_op(ICPPKernel * kernel,const Hints & hints,const Window & window,ITensorPack & tensors)56 void OMPScheduler::schedule_op(ICPPKernel *kernel, const Hints &hints, const Window &window, ITensorPack &tensors)
57 {
58 ARM_COMPUTE_ERROR_ON_MSG(!kernel, "The child class didn't set the kernel");
59 ARM_COMPUTE_ERROR_ON_MSG(hints.strategy() == StrategyHint::DYNAMIC,
60 "Dynamic scheduling is not supported in OMPScheduler");
61
62 const Window & max_window = window;
63 const unsigned int num_iterations = max_window.num_iterations(hints.split_dimension());
64 const unsigned int num_threads = std::min(num_iterations, _num_threads);
65
66 if(!kernel->is_parallelisable() || num_threads == 1)
67 {
68 ThreadInfo info;
69 info.cpu_info = &cpu_info();
70 kernel->run_op(tensors, max_window, info);
71 }
72 else
73 {
74 const unsigned int num_windows = num_threads;
75 std::vector<IScheduler::Workload> workloads(num_windows);
76 for(unsigned int t = 0; t < num_windows; t++)
77 {
78 //Capture 't' by copy, all the other variables by reference:
79 workloads[t] = [t, &hints, &max_window, &num_windows, &kernel, &tensors](const ThreadInfo &info) {
80 Window win = max_window.split_window(hints.split_dimension(), t, num_windows);
81 win.validate();
82 kernel->run_op(tensors, win, info);
83 };
84 }
85 run_workloads(workloads);
86 }
87 }
88 #ifndef DOXYGEN_SKIP_THIS
run_workloads(std::vector<arm_compute::IScheduler::Workload> & workloads)89 void OMPScheduler::run_workloads(std::vector<arm_compute::IScheduler::Workload> &workloads)
90 {
91 const unsigned int amount_of_work = static_cast<unsigned int>(workloads.size());
92 const unsigned int num_threads_to_use = std::min(_num_threads, amount_of_work);
93
94 if(amount_of_work < 1 || num_threads_to_use == 1)
95 {
96 return;
97 }
98
99 ThreadInfo info;
100 info.cpu_info = &cpu_info();
101 info.num_threads = num_threads_to_use;
102 #pragma omp parallel for firstprivate(info) num_threads(num_threads_to_use) default(shared) proc_bind(close) schedule(static, 1)
103 for(unsigned int wid = 0; wid < amount_of_work; ++wid)
104 {
105 const int tid = omp_get_thread_num();
106
107 info.thread_id = tid;
108 workloads[wid](info);
109 }
110 }
111 #endif /* DOXYGEN_SKIP_THIS */
112 } // namespace arm_compute
113