1 /*
2 * Copyright (c) 2016-2022 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/IScheduler.h"
25
26 #include "arm_compute/core/CPP/ICPPKernel.h"
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/Log.h"
29 #include "arm_compute/core/Window.h"
30 #include "src/common/cpuinfo/CpuInfo.h"
31 #include "src/runtime/SchedulerUtils.h"
32
33 namespace arm_compute
34 {
IScheduler()35 IScheduler::IScheduler()
36 {
37 // Work out the best possible number of execution threads
38 _num_threads_hint = cpuinfo::num_threads_hint();
39 }
40
cpu_info()41 CPUInfo &IScheduler::cpu_info()
42 {
43 return CPUInfo::get();
44 }
45
set_num_threads_with_affinity(unsigned int num_threads,BindFunc func)46 void IScheduler::set_num_threads_with_affinity(unsigned int num_threads, BindFunc func)
47 {
48 ARM_COMPUTE_UNUSED(num_threads, func);
49 ARM_COMPUTE_ERROR("Feature for affinity setting is not implemented");
50 }
51
num_threads_hint() const52 unsigned int IScheduler::num_threads_hint() const
53 {
54 return _num_threads_hint;
55 }
56
schedule_common(ICPPKernel * kernel,const Hints & hints,const Window & window,ITensorPack & tensors)57 void IScheduler::schedule_common(ICPPKernel *kernel, const Hints &hints, const Window &window, ITensorPack &tensors)
58 {
59 ARM_COMPUTE_ERROR_ON_MSG(!kernel, "The child class didn't set the kernel");
60 #ifndef BARE_METAL
61 const Window &max_window = window;
62 if(hints.split_dimension() == IScheduler::split_dimensions_all)
63 {
64 /*
65 * if the split dim is size_t max then this signals we should parallelise over
66 * all dimensions
67 */
68 const std::size_t m = max_window.num_iterations(Window::DimX);
69 const std::size_t n = max_window.num_iterations(Window::DimY);
70
71 //in c++17 this can be swapped for auto [ m_threads, n_threads ] = split_2d(...
72 unsigned m_threads, n_threads;
73 std::tie(m_threads, n_threads) = scheduler_utils::split_2d(this->num_threads(), m, n);
74
75 std::vector<IScheduler::Workload> workloads;
76 for(unsigned int ni = 0; ni != n_threads; ++ni)
77 {
78 for(unsigned int mi = 0; mi != m_threads; ++mi)
79 {
80 workloads.push_back(
81 [ni, mi, m_threads, n_threads, &max_window, &kernel](const ThreadInfo & info)
82 {
83 //narrow the window to our mi-ni workload
84 Window win = max_window.split_window(Window::DimX, mi, m_threads)
85 .split_window(Window::DimY, ni, n_threads);
86
87 win.validate();
88
89 Window thread_locator;
90 thread_locator.set(Window::DimX, Window::Dimension(mi, m_threads));
91 thread_locator.set(Window::DimY, Window::Dimension(ni, n_threads));
92
93 thread_locator.validate();
94
95 kernel->run_nd(win, info, thread_locator);
96 });
97 }
98 }
99 run_workloads(workloads);
100 }
101 else
102 {
103 const unsigned int num_iterations = max_window.num_iterations(hints.split_dimension());
104 const unsigned int num_threads = std::min(num_iterations, this->num_threads());
105
106 if(num_iterations == 0)
107 {
108 return;
109 }
110
111 if(!kernel->is_parallelisable() || num_threads == 1)
112 {
113 ThreadInfo info;
114 info.cpu_info = &cpu_info();
115 if(tensors.empty())
116 {
117 kernel->run(max_window, info);
118 }
119 else
120 {
121 kernel->run_op(tensors, max_window, info);
122 }
123 }
124 else
125 {
126 unsigned int num_windows = 0;
127 switch(hints.strategy())
128 {
129 case StrategyHint::STATIC:
130 num_windows = num_threads;
131 break;
132 case StrategyHint::DYNAMIC:
133 {
134 const unsigned int granule_threshold = (hints.threshold() <= 0) ? num_threads : static_cast<unsigned int>(hints.threshold());
135 // Make sure we don't use some windows which are too small as this might create some contention on the ThreadFeeder
136 num_windows = num_iterations > granule_threshold ? granule_threshold : num_iterations;
137 break;
138 }
139 default:
140 ARM_COMPUTE_ERROR("Unknown strategy");
141 }
142 // Make sure the smallest window is larger than minimim workload size
143 num_windows = adjust_num_of_windows(max_window, hints.split_dimension(), num_windows, *kernel, cpu_info());
144
145 std::vector<IScheduler::Workload> workloads(num_windows);
146 for(unsigned int t = 0; t < num_windows; ++t)
147 {
148 //Capture 't' by copy, all the other variables by reference:
149 workloads[t] = [t, &hints, &max_window, &num_windows, &kernel, &tensors](const ThreadInfo & info)
150 {
151 Window win = max_window.split_window(hints.split_dimension(), t, num_windows);
152 win.validate();
153
154 if(tensors.empty())
155 {
156 kernel->run(win, info);
157 }
158 else
159 {
160 kernel->run_op(tensors, win, info);
161 }
162 };
163 }
164 run_workloads(workloads);
165 }
166 }
167 #else /* !BARE_METAL */
168 ARM_COMPUTE_UNUSED(kernel, hints, window, tensors);
169 #endif /* !BARE_METAL */
170 }
171
run_tagged_workloads(std::vector<Workload> & workloads,const char * tag)172 void IScheduler::run_tagged_workloads(std::vector<Workload> &workloads, const char *tag)
173 {
174 ARM_COMPUTE_UNUSED(tag);
175 run_workloads(workloads);
176 }
177
adjust_num_of_windows(const Window & window,std::size_t split_dimension,std::size_t init_num_windows,const ICPPKernel & kernel,const CPUInfo & cpu_info)178 std::size_t IScheduler::adjust_num_of_windows(const Window &window, std::size_t split_dimension, std::size_t init_num_windows, const ICPPKernel &kernel, const CPUInfo &cpu_info)
179 {
180 // Mitigation of the narrow split issue, which occurs when the split dimension is too small to split (hence "narrow").
181 if(window.num_iterations(split_dimension) < init_num_windows )
182 {
183 auto recommended_split_dim = Window::DimX;
184 for(std::size_t dims = Window::DimY; dims <= Window::DimW; ++dims)
185 {
186 if(window.num_iterations(recommended_split_dim) < window.num_iterations(dims))
187 {
188 recommended_split_dim = dims;
189 }
190 }
191 ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("%zu dimension is not a suitable dimension to split the workload. Recommended: %zu recommended_split_dim", split_dimension,
192 recommended_split_dim);
193 }
194
195 for(auto t = init_num_windows; t > 0; --t) // Trying the highest number of windows ,init_num_windows, first
196 {
197 // Try splitting the workload into t, subject to each subworkload size <= mws.
198 if((window.num_iterations(split_dimension) / kernel.get_mws(cpu_info, t)) >= t)
199 {
200 if(t != init_num_windows)
201 {
202 ARM_COMPUTE_LOG_INFO_MSG_CORE("The scheduler is using a different thread count than the one assigned by the user.");
203 }
204 return t;
205 }
206 }
207 ARM_COMPUTE_LOG_INFO_MSG_CORE("The scheduler is using single thread instead of the thread count assigned by the user.");
208 return 1; // If the workload is so small that it can't be split, we should run a single thread
209 }
210
211 } // namespace arm_compute
212