Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
SiCong Li | dba672c | 2023-04-06 16:30:18 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2021, 2023 Arm Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 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 | */ |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 24 | #ifndef ARM_COMPUTE_ISCHEDULER_H |
| 25 | #define ARM_COMPUTE_ISCHEDULER_H |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 26 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/CPP/CPPTypes.h" |
Michalis Spyrou | bcd2352 | 2020-05-21 15:02:36 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/experimental/Types.h" |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 29 | #include "arm_compute/core/Types.h" |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 30 | |
Anthony Barbier | 52ecb06 | 2018-05-25 13:32:10 +0100 | [diff] [blame] | 31 | #include <functional> |
Joseph Dobson | 6f8b17d | 2020-02-11 19:32:11 +0000 | [diff] [blame] | 32 | #include <limits> |
Anthony Barbier | 52ecb06 | 2018-05-25 13:32:10 +0100 | [diff] [blame] | 33 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 34 | namespace arm_compute |
| 35 | { |
| 36 | class ICPPKernel; |
Michalis Spyrou | bcd2352 | 2020-05-21 15:02:36 +0100 | [diff] [blame] | 37 | class ITensor; |
Sang-Hoon Park | 0094c02 | 2021-01-20 18:16:47 +0000 | [diff] [blame] | 38 | class Window; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 39 | |
| 40 | /** Scheduler interface to run kernels */ |
| 41 | class IScheduler |
| 42 | { |
| 43 | public: |
Anthony Barbier | 376c85f | 2018-05-25 14:17:21 +0100 | [diff] [blame] | 44 | /** Strategies available to split a workload */ |
| 45 | enum class StrategyHint |
| 46 | { |
| 47 | STATIC, /**< Split the workload evenly among the threads */ |
| 48 | DYNAMIC, /**< Split the workload dynamically using a bucket system */ |
| 49 | }; |
Joseph Dobson | 6f8b17d | 2020-02-11 19:32:11 +0000 | [diff] [blame] | 50 | |
Georgios Pinitas | 06e890b | 2020-07-09 18:38:34 +0100 | [diff] [blame] | 51 | /** Function to be used and map a given thread id to a logical core id |
| 52 | * |
| 53 | * Mapping function expects the thread index and total number of cores as input, |
| 54 | * and returns the logical core index to bind against |
| 55 | */ |
| 56 | using BindFunc = std::function<int(int, int)>; |
| 57 | |
Joseph Dobson | 6f8b17d | 2020-02-11 19:32:11 +0000 | [diff] [blame] | 58 | /** When arm_compute::ISchedular::Hints::_split_dimension is initialized with this value |
| 59 | * then the schedular is free to break down the problem space over as many dimensions |
| 60 | * as it wishes |
| 61 | */ |
| 62 | static constexpr unsigned int split_dimensions_all = std::numeric_limits<unsigned>::max(); |
| 63 | |
Anthony Barbier | 376c85f | 2018-05-25 14:17:21 +0100 | [diff] [blame] | 64 | /** Scheduler hints |
| 65 | * |
| 66 | * Collection of preferences set by the function regarding how to split a given workload |
| 67 | */ |
| 68 | class Hints |
| 69 | { |
| 70 | public: |
| 71 | /** Constructor |
| 72 | * |
| 73 | * @param[in] split_dimension Dimension along which to split the kernel's execution window. |
| 74 | * @param[in] strategy (Optional) Split strategy. |
Georgios Pinitas | 77d4252 | 2019-11-05 13:35:47 +0000 | [diff] [blame] | 75 | * @param[in] threshold (Optional) Dynamic scheduling capping threshold. |
Anthony Barbier | 376c85f | 2018-05-25 14:17:21 +0100 | [diff] [blame] | 76 | */ |
Georgios Pinitas | 77d4252 | 2019-11-05 13:35:47 +0000 | [diff] [blame] | 77 | Hints(unsigned int split_dimension, StrategyHint strategy = StrategyHint::STATIC, int threshold = 0) |
| 78 | : _split_dimension(split_dimension), _strategy(strategy), _threshold(threshold) |
Anthony Barbier | 376c85f | 2018-05-25 14:17:21 +0100 | [diff] [blame] | 79 | { |
| 80 | } |
| 81 | /** Set the split_dimension hint |
| 82 | * |
| 83 | * @param[in] split_dimension Dimension along which to split the kernel's execution window. |
| 84 | * |
| 85 | * @return the Hints object |
| 86 | */ |
| 87 | Hints &set_split_dimension(unsigned int split_dimension) |
| 88 | { |
| 89 | _split_dimension = split_dimension; |
| 90 | return *this; |
| 91 | } |
| 92 | /** Return the prefered split dimension |
| 93 | * |
| 94 | * @return The split dimension |
| 95 | */ |
| 96 | unsigned int split_dimension() const |
| 97 | { |
| 98 | return _split_dimension; |
| 99 | } |
| 100 | |
| 101 | /** Set the strategy hint |
| 102 | * |
| 103 | * @param[in] strategy Prefered strategy to use to split the workload |
| 104 | * |
| 105 | * @return the Hints object |
| 106 | */ |
| 107 | Hints &set_strategy(StrategyHint strategy) |
| 108 | { |
| 109 | _strategy = strategy; |
| 110 | return *this; |
| 111 | } |
| 112 | /** Return the prefered strategy to use to split workload. |
| 113 | * |
| 114 | * @return The strategy |
| 115 | */ |
| 116 | StrategyHint strategy() const |
| 117 | { |
| 118 | return _strategy; |
| 119 | } |
Georgios Pinitas | 77d4252 | 2019-11-05 13:35:47 +0000 | [diff] [blame] | 120 | /** Return the granule capping threshold to be used by dynamic scheduling. |
| 121 | * |
| 122 | * @return The capping threshold |
| 123 | */ |
| 124 | int threshold() const |
| 125 | { |
| 126 | return _threshold; |
| 127 | } |
Anthony Barbier | 376c85f | 2018-05-25 14:17:21 +0100 | [diff] [blame] | 128 | |
| 129 | private: |
Georgios Pinitas | 08302c1 | 2021-06-09 10:08:27 +0100 | [diff] [blame] | 130 | unsigned int _split_dimension{}; |
| 131 | StrategyHint _strategy{}; |
| 132 | int _threshold{}; |
Anthony Barbier | 376c85f | 2018-05-25 14:17:21 +0100 | [diff] [blame] | 133 | }; |
Anthony Barbier | 52ecb06 | 2018-05-25 13:32:10 +0100 | [diff] [blame] | 134 | /** Signature for the workloads to execute */ |
| 135 | using Workload = std::function<void(const ThreadInfo &)>; |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 136 | /** Default constructor. */ |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 137 | IScheduler(); |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 138 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 139 | /** Destructor. */ |
| 140 | virtual ~IScheduler() = default; |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 141 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 142 | /** Sets the number of threads the scheduler will use to run the kernels. |
| 143 | * |
| 144 | * @param[in] num_threads If set to 0, then one thread per CPU core available on the system will be used, otherwise the number of threads specified. |
| 145 | */ |
| 146 | virtual void set_num_threads(unsigned int num_threads) = 0; |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 147 | |
Georgios Pinitas | 06e890b | 2020-07-09 18:38:34 +0100 | [diff] [blame] | 148 | /** Sets the number of threads the scheduler will use to run the kernels but also using a binding function to pin the threads to given logical cores |
| 149 | * |
| 150 | * @param[in] num_threads If set to 0, then one thread per CPU core available on the system will be used, otherwise the number of threads specified. |
| 151 | * @param[in] func Binding function to use. |
| 152 | */ |
| 153 | virtual void set_num_threads_with_affinity(unsigned int num_threads, BindFunc func); |
| 154 | |
ramelg01 | b2eba7f | 2021-12-23 08:32:08 +0000 | [diff] [blame] | 155 | /** Returns the number of threads that the SingleThreadScheduler has in its pool. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 156 | * |
| 157 | * @return Number of threads available in SingleThreadScheduler. |
| 158 | */ |
| 159 | virtual unsigned int num_threads() const = 0; |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 160 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 161 | /** Runs the kernel in the same thread as the caller synchronously. |
| 162 | * |
Anthony Barbier | 376c85f | 2018-05-25 14:17:21 +0100 | [diff] [blame] | 163 | * @param[in] kernel Kernel to execute. |
| 164 | * @param[in] hints Hints for the scheduler. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 165 | */ |
Anthony Barbier | 376c85f | 2018-05-25 14:17:21 +0100 | [diff] [blame] | 166 | virtual void schedule(ICPPKernel *kernel, const Hints &hints) = 0; |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 167 | |
Michalis Spyrou | bcd2352 | 2020-05-21 15:02:36 +0100 | [diff] [blame] | 168 | /** Runs the kernel in the same thread as the caller synchronously. |
| 169 | * |
| 170 | * @param[in] kernel Kernel to execute. |
| 171 | * @param[in] hints Hints for the scheduler. |
Sang-Hoon Park | 0094c02 | 2021-01-20 18:16:47 +0000 | [diff] [blame] | 172 | * @param[in] window Window to use for kernel execution. |
Georgios Pinitas | 0499dff | 2020-07-31 22:21:38 +0100 | [diff] [blame] | 173 | * @param[in] tensors Vector containing the tensors to operate on. |
Michalis Spyrou | bcd2352 | 2020-05-21 15:02:36 +0100 | [diff] [blame] | 174 | */ |
Sang-Hoon Park | 0094c02 | 2021-01-20 18:16:47 +0000 | [diff] [blame] | 175 | virtual void schedule_op(ICPPKernel *kernel, const Hints &hints, const Window &window, ITensorPack &tensors) = 0; |
Michalis Spyrou | bcd2352 | 2020-05-21 15:02:36 +0100 | [diff] [blame] | 176 | |
Anthony Barbier | 52ecb06 | 2018-05-25 13:32:10 +0100 | [diff] [blame] | 177 | /** Execute all the passed workloads |
| 178 | * |
SiCong Li | dba672c | 2023-04-06 16:30:18 +0100 | [diff] [blame] | 179 | * @note There is no guarantee regarding the order in which the workloads will be executed or whether or not they will be executed in parallel. |
Anthony Barbier | 52ecb06 | 2018-05-25 13:32:10 +0100 | [diff] [blame] | 180 | * |
SiCong Li | dba672c | 2023-04-06 16:30:18 +0100 | [diff] [blame] | 181 | * @param[in] workloads List of workloads to run |
Anthony Barbier | 148b075 | 2018-09-11 14:19:39 +0100 | [diff] [blame] | 182 | * @param[in] tag String that can be used by profiling tools to identify the workloads run by the scheduler (Can be null). |
Anthony Barbier | 52ecb06 | 2018-05-25 13:32:10 +0100 | [diff] [blame] | 183 | */ |
Anthony Barbier | 148b075 | 2018-09-11 14:19:39 +0100 | [diff] [blame] | 184 | virtual void run_tagged_workloads(std::vector<Workload> &workloads, const char *tag); |
Anthony Barbier | 52ecb06 | 2018-05-25 13:32:10 +0100 | [diff] [blame] | 185 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 186 | /** Get CPU info. |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 187 | * |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 188 | * @return CPU info. |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 189 | */ |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 190 | CPUInfo &cpu_info(); |
Georgios Pinitas | 53d1227 | 2018-02-01 20:23:25 +0000 | [diff] [blame] | 191 | /** Get a hint for the best possible number of execution threads |
| 192 | * |
| 193 | * @warning In case we can't work out the best number of threads, |
| 194 | * std::thread::hardware_concurrency() is returned else 1 in case of bare metal builds |
| 195 | * |
| 196 | * @return Best possible number of execution threads to use |
| 197 | */ |
| 198 | unsigned int num_threads_hint() const; |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 199 | |
| 200 | protected: |
Anthony Barbier | 148b075 | 2018-09-11 14:19:39 +0100 | [diff] [blame] | 201 | /** Execute all the passed workloads |
| 202 | * |
| 203 | * @note there is no guarantee regarding the order in which the workloads will be executed or whether or not they will be executed in parallel. |
| 204 | * |
| 205 | * @param[in] workloads Array of workloads to run |
| 206 | */ |
| 207 | virtual void run_workloads(std::vector<Workload> &workloads) = 0; |
Georgios Pinitas | 53d1227 | 2018-02-01 20:23:25 +0000 | [diff] [blame] | 208 | |
Sang-Hoon Park | 0094c02 | 2021-01-20 18:16:47 +0000 | [diff] [blame] | 209 | /** Common scheduler logic to execute the given kernel |
| 210 | * |
| 211 | * @param[in] kernel Kernel to execute. |
| 212 | * @param[in] hints Hints for the scheduler. |
| 213 | * @param[in] window Window to use for kernel execution. |
| 214 | * @param[in] tensors Vector containing the tensors to operate on. |
| 215 | */ |
| 216 | void schedule_common(ICPPKernel *kernel, const Hints &hints, const Window &window, ITensorPack &tensors); |
morgolock | 5111264 | 2020-08-20 14:51:39 +0100 | [diff] [blame] | 217 | |
Dana Zlotnik | d7154db | 2021-11-10 11:50:58 +0200 | [diff] [blame] | 218 | /** Adjust the number of windows to the optimize performance |
| 219 | * (used for small workloads where smaller number of threads might improve the performance) |
| 220 | * |
| 221 | * @param[in] window Window to use for kernel execution |
| 222 | * @param[in] split_dimension Axis of dimension to split |
| 223 | * @param[in] init_num_windows Initial number of sub-windows to split |
| 224 | * @param[in] kernel Kernel to execute |
| 225 | * @param[in] cpu_info The CPU platform used to create the context. |
| 226 | * |
| 227 | * @return Adjusted number of windows |
| 228 | */ |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 229 | std::size_t adjust_num_of_windows(const Window &window, |
| 230 | std::size_t split_dimension, |
| 231 | std::size_t init_num_windows, |
| 232 | const ICPPKernel &kernel, |
| 233 | const CPUInfo &cpu_info); |
Dana Zlotnik | d7154db | 2021-11-10 11:50:58 +0200 | [diff] [blame] | 234 | |
Georgios Pinitas | 53d1227 | 2018-02-01 20:23:25 +0000 | [diff] [blame] | 235 | private: |
| 236 | unsigned int _num_threads_hint = {}; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 237 | }; |
Georgios Pinitas | 77d4252 | 2019-11-05 13:35:47 +0000 | [diff] [blame] | 238 | } // namespace arm_compute |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 239 | #endif /* ARM_COMPUTE_ISCHEDULER_H */ |