1/* 2 * Copyright (C) 2019 The Android Open Source Project 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 17package android.hardware.neuralnetworks@1.3; 18 19import @1.2::IPreparedModel; 20import @1.2::MeasureTiming; 21import @1.2::OutputShape; 22import @1.2::Timing; 23import ErrorStatus; 24import OptionalTimeoutDuration; 25import OptionalTimePoint; 26import Request; 27import IExecutionCallback; 28import IFencedExecutionCallback; 29 30/** 31 * IPreparedModel describes a model that has been prepared for execution and 32 * is used to launch executions. 33 */ 34interface IPreparedModel extends @1.2::IPreparedModel { 35 /** 36 * Launches an asynchronous execution on a prepared model. 37 * 38 * The execution is performed asynchronously with respect to the caller. 39 * execute_1_3 must verify the inputs to the function are correct, and the usages 40 * of memory pools allocated by IDevice::allocate are valid. If there is 41 * an error, execute_1_3 must immediately invoke the callback with the 42 * appropriate ErrorStatus value, then return with the same ErrorStatus. If 43 * the inputs to the function are valid and there is no error, execute_1_3 must 44 * launch an asynchronous task to perform the execution in the background, 45 * and immediately return with ErrorStatus::NONE. If the asynchronous task 46 * fails to launch, execute_1_3 must immediately invoke the callback with 47 * ErrorStatus::GENERAL_FAILURE, then return with 48 * ErrorStatus::GENERAL_FAILURE. 49 * 50 * When the asynchronous task has finished its execution, it must 51 * immediately invoke the callback object provided as an input to the 52 * execute_1_3 function. This callback must be provided with the ErrorStatus of 53 * the execution. 54 * 55 * If the launch is successful, the caller must not change the content of 56 * any data object referenced by 'request' (described by the 57 * {@link @1.0::DataLocation} of a {@link @1.0::RequestArgument}) until the 58 * asynchronous task has invoked the callback object. The asynchronous task 59 * must not change the content of any of the data objects corresponding to 60 * 'request' inputs. 61 * 62 * If the prepared model was prepared from a model wherein all tensor 63 * operands have fully specified dimensions, and the inputs to the function 64 * are valid, then: 65 * - the execution should launch successfully (ErrorStatus::NONE): There 66 * must be no failure unless the device itself is in a bad state. 67 * - if at execution time every operation's input operands have legal 68 * values, the execution should complete successfully (ErrorStatus::NONE): 69 * There must be no failure unless the device itself is in a bad state. 70 * 71 * execute_1_3 can be called with an optional deadline. If the execution 72 * is not able to be completed before the provided deadline, the execution 73 * may be aborted, and either {@link 74 * ErrorStatus::MISSED_DEADLINE_TRANSIENT} or {@link 75 * ErrorStatus::MISSED_DEADLINE_PERSISTENT} may be returned. The error due 76 * to an abort must be sent the same way as other errors, described above. 77 * The deadline is represented as nanoseconds since the epoch of the steady 78 * clock (as if from std::chrono::steady_clock::time_point), but the service 79 * may convert it to the nanoseconds since boot time (as if from 80 * clock_gettime(CLOCK_BOOTTIME, &ts) or 81 * android::base::boot_clock::time_point) to account for time when the 82 * system is suspended. This conversion can by done by finding the timeout 83 * duration remaining compared to the steady_clock and adding it to the 84 * current boot_clock time. 85 * 86 * Any number of calls to the execute* and executeSynchronously* functions, 87 * in any combination, may be made concurrently, even on the same 88 * IPreparedModel object. 89 * 90 * @param request The input and output information on which the prepared 91 * model is to be executed. 92 * @param measure Specifies whether or not to measure duration of the execution. 93 * The duration runs from the time the driver sees the call 94 * to the execute_1_3 function to the time the driver invokes 95 * the callback. 96 * @param deadline The time by which the execution is expected to complete. 97 * If the execution cannot be completed by the deadline, the 98 * execution may be aborted. 99 * @param loopTimeoutDuration The maximum amount of time that should be spent 100 * executing a {@link OperationType::WHILE} 101 * operation. If a loop condition model does not 102 * output false within this duration, the 103 * execution must be aborted. If no loop timeout 104 * duration is provided, the maximum amount of 105 * time is {@link LoopTimeoutDurationNs::DEFAULT}. 106 * When provided, the duration must not exceed 107 * {@link LoopTimeoutDurationNs::MAXIMUM}. 108 * @param callback A callback object used to return the error status of 109 * the execution, shape information of model output operands, and 110 * duration of execution. The callback object's notify function must 111 * be called exactly once, even if the execution was 112 * unsuccessful. 113 * @return status Error status of the call, must be: 114 * - NONE if task is successfully launched 115 * - DEVICE_UNAVAILABLE if driver is offline or busy 116 * - GENERAL_FAILURE if there is an unspecified error 117 * - OUTPUT_INSUFFICIENT_SIZE if provided output buffer is 118 * not large enough to store the resultant values 119 * - INVALID_ARGUMENT if one of the input arguments is 120 * invalid 121 * - MISSED_DEADLINE_* if the execution is aborted because it 122 * cannot be completed by the deadline 123 * - RESOURCE_EXHAUSTED_* if the task was aborted by the 124 * driver 125 */ 126 execute_1_3(Request request, MeasureTiming measure, OptionalTimePoint deadline, 127 OptionalTimeoutDuration loopTimeoutDuration, IExecutionCallback callback) 128 generates (ErrorStatus status); 129 130 /** 131 * Performs a synchronous execution on a prepared model. 132 * 133 * The execution is performed synchronously with respect to the caller. 134 * executeSynchronously_1_3 must verify the inputs to the function are 135 * correct, and the usages of memory pools allocated by IDevice::allocate 136 * are valid. If there is an error, executeSynchronously_1_3 must immediately 137 * return with the appropriate ErrorStatus value. If the inputs to the 138 * function are valid and there is no error, executeSynchronously_1_3 must 139 * perform the execution, and must not return until the execution is 140 * complete. 141 * 142 * The caller must not change the content of any data object referenced by 143 * 'request' (described by the {@link @1.0::DataLocation} of a 144 * {@link @1.0::RequestArgument}) until executeSynchronously_1_3 145 * returns. executeSynchronously_1_3 must not change the content of any of the 146 * data objects corresponding to 'request' inputs. 147 * 148 * If the prepared model was prepared from a model wherein all tensor 149 * operands have fully specified dimensions, and the inputs to the function 150 * are valid, and at execution time every operation's input operands have 151 * legal values, then the execution should complete successfully 152 * (ErrorStatus::NONE): There must be no failure unless the device itself is 153 * in a bad state. 154 * 155 * executeSynchronously_1_3 may be called with an optional deadline. If the 156 * execution is not able to be completed before the provided deadline, the 157 * execution may be aborted, and either {@link 158 * ErrorStatus::MISSED_DEADLINE_TRANSIENT} or {@link 159 * ErrorStatus::MISSED_DEADLINE_PERSISTENT} may be returned. The error due 160 * to an abort must be sent the same way as other errors, described above. 161 * The deadline is represented as nanoseconds since the epoch of the steady 162 * clock (as if from std::chrono::steady_clock::time_point), but the service 163 * may convert it to the nanoseconds since boot time (as if from 164 * clock_gettime(CLOCK_BOOTTIME, &ts) or 165 * android::base::boot_clock::time_point) to account for time when the 166 * system is suspended. This conversion can by done by finding the timeout 167 * duration remaining compared to the steady_clock and adding it to the 168 * current boot_clock time. 169 * 170 * Any number of calls to the execute* and executeSynchronously* functions, 171 * in any combination, may be made concurrently, even on the same 172 * IPreparedModel object. 173 * 174 * @param request The input and output information on which the prepared 175 * model is to be executed. 176 * @param measure Specifies whether or not to measure duration of the execution. 177 * The duration runs from the time the driver sees the call 178 * to the executeSynchronously_1_3 function to the time the driver 179 * returns from the function. 180 * @param deadline The time by which the execution is expected to complete. 181 * If the execution cannot be finished by the deadline, the 182 * execution may be aborted. 183 * @param loopTimeoutDuration The maximum amount of time that should be spent 184 * executing a {@link OperationType::WHILE} 185 * operation. If a loop condition model does not 186 * output false within this duration, the 187 * execution must be aborted. If no loop timeout 188 * duration is provided, the maximum amount of 189 * time is {@link LoopTimeoutDurationNs::DEFAULT}. 190 * When provided, the duration must not exceed 191 * {@link LoopTimeoutDurationNs::MAXIMUM}. 192 * @return status Error status of the execution, must be: 193 * - NONE if execution is performed successfully 194 * - DEVICE_UNAVAILABLE if driver is offline or busy 195 * - GENERAL_FAILURE if there is an unspecified error 196 * - OUTPUT_INSUFFICIENT_SIZE if at least one output 197 * operand buffer is not large enough to store the 198 * corresponding output 199 * - INVALID_ARGUMENT if one of the input arguments is 200 * invalid 201 * - MISSED_DEADLINE_* if the execution is aborted because it 202 * cannot be completed by the deadline 203 * - RESOURCE_EXHAUSTED_* if the task was aborted by the 204 * driver 205 * @return outputShapes A list of shape information of model output operands. 206 * The index into "outputShapes" corresponds to the index 207 * of the output operand in the Request outputs vector. 208 * outputShapes must be empty unless the status is either 209 * NONE or OUTPUT_INSUFFICIENT_SIZE. 210 * @return timing Duration of execution. Unless measure is YES and status is 211 * NONE, all times must be reported as UINT64_MAX. A driver may 212 * choose to report any time as UINT64_MAX, indicating that 213 * measurement is not available. 214 */ 215 executeSynchronously_1_3(Request request, MeasureTiming measure, 216 OptionalTimePoint deadline, 217 OptionalTimeoutDuration loopTimeoutDuration) 218 generates (ErrorStatus status, vec<OutputShape> outputShapes, 219 Timing timing); 220 221 /** 222 * Launch a fenced asynchronous execution on a prepared model. 223 * 224 * The execution is performed asynchronously with respect to the caller. 225 * executeFenced must verify the inputs to the function are correct, and the usages 226 * of memory pools allocated by IDevice::allocate are valid. If there is an error, 227 * executeFenced must immediately return with the corresponding ErrorStatus, an empty 228 * handle for syncFence, and nullptr for callback. If the inputs to the function 229 * are valid and there is no error, executeFenced must dispatch an asynchronous task 230 * to perform the execution in the background, and immediately return with 231 * ErrorStatus::NONE, a sync fence that will be signaled once the execution is completed, 232 * and a callback that can be used by the client to query the duration and runtime error 233 * status. If the task has finished before the call returns, an empty handle may be returned 234 * for syncFence. The execution must wait for all the sync fences (if any) in waitFor 235 * to be signaled before starting the actual execution. 236 * 237 * When the asynchronous task has finished its execution, it must 238 * immediately signal the syncFence returned from the executeFenced call. After 239 * the syncFence is signaled, the task must not modify the content of 240 * any data object referenced by 'request' (described by the 241 * {@link @1.0::DataLocation} of a {@link @1.0::RequestArgument}). 242 * 243 * executeFenced may be called with an optional deadline and an optional duration. 244 * If the execution is not able to be completed before the provided deadline or 245 * within the timeout duration (measured from when all sync fences in waitFor are 246 * signaled), whichever comes earlier, the execution may be aborted, and either 247 * {@link ErrorStatus::MISSED_DEADLINE_TRANSIENT} or {@link 248 * ErrorStatus::MISSED_DEADLINE_PERSISTENT} may be returned. The error due 249 * to an abort must be sent the same way as other errors, described above. 250 * The deadline is represented as nanoseconds since the epoch of the steady 251 * clock (as if from std::chrono::steady_clock::time_point), but the service 252 * may convert it to the nanoseconds since boot time (as if from 253 * clock_gettime(CLOCK_BOOTTIME, &ts) or 254 * android::base::boot_clock::time_point) to account for time when the 255 * system is suspended. This conversion can by done by finding the timeout 256 * duration remaining compared to the steady_clock and adding it to the 257 * current boot_clock time. 258 * 259 * If any of the sync fences in waitFor changes to error status after the executeFenced 260 * call succeeds, or the execution is aborted because it cannot finish before the deadline 261 * has been reached or the duration has elapsed, the driver must immediately set the returned 262 * syncFence to error status. 263 * 264 * Any number of calls to the executeFenced, execute* and executeSynchronously* 265 * functions, in any combination, may be made concurrently, even on the same 266 * IPreparedModel object. 267 * 268 * @param request The input and output information on which the prepared 269 * model is to be executed. The outputs in the request must have 270 * fully specified dimensions. 271 * @param waitFor A vector of sync fence file descriptors. 272 * Execution must not start until all sync fences have been signaled. 273 * @param measure Specifies whether or not to measure duration of the execution. 274 * @param deadline The time by which the execution is expected to complete. 275 * If the execution cannot be finished by the deadline, the 276 * execution may be aborted. 277 * @param loopTimeoutDuration The maximum amount of time that should be spent 278 * executing a {@link OperationType::WHILE} 279 * operation. If a loop condition model does not 280 * output false within this duration, the 281 * execution must be aborted. If no loop timeout 282 * duration is provided, the maximum amount of 283 * time is {@link LoopTimeoutDurationNs::DEFAULT}. 284 * When provided, the duration must not exceed 285 * {@link LoopTimeoutDurationNs::MAXIMUM}. 286 * @param duration The length of time within which the execution is expected 287 * to complete after all sync fences in waitFor are signaled. 288 * If the execution cannot be finished within the duration, 289 * the execution may be aborted. 290 * @return status Error status of the call, must be: 291 * - NONE if task is successfully launched 292 * - DEVICE_UNAVAILABLE if driver is offline or busy 293 * - GENERAL_FAILURE if there is an unspecified error 294 * - INVALID_ARGUMENT if one of the input arguments is invalid, including 295 * fences in error states. 296 * - MISSED_DEADLINE_* if the execution is aborted because it 297 * cannot be completed by the deadline 298 * - RESOURCE_EXHAUSTED_* if the task was aborted by the 299 * driver 300 * @return syncFence The sync fence that will be signaled when the task is completed. 301 * The sync fence will be set to error if a critical error, 302 * e.g. hardware failure or kernel panic, occurs when doing execution. 303 * @return callback The IFencedExecutionCallback can be used to query information like duration 304 * and error status when the execution is completed. 305 */ 306 executeFenced(Request request, vec<handle> waitFor, MeasureTiming measure, 307 OptionalTimePoint deadline, OptionalTimeoutDuration loopTimeoutDuration, 308 OptionalTimeoutDuration duration) 309 generates (ErrorStatus status, handle syncFence, IFencedExecutionCallback callback); 310}; 311