2 * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
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
8 * http://www.apache.org/licenses/LICENSE-2.0
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.
19 * @brief This file describes runtime API
32 * @brief Session to query with runtime
34 * <p>nnfw_session is started and passed by calling {@link nnfw_create_session}.
35 * Each session has its own inference environment, such as model to inference, backend usage, etc.
37 * <p>Load model by calling {@link nnfw_load_model_from_file}
39 * <p>After loading, prepare inference by calling {@link nnfw_prepare}.
40 * Application can set runtime environment before prepare by calling
41 * {@link nnfw_set_available_backends} and {@link nnfw_set_op_backend}, and it is optional.
43 * <p>Application can inference by calling {@link nnfw_run}.
44 * Before inference, application has responsibility to set input tensor to set input data by calling
45 * {@link nnfw_set_output}, and output tensor to get output by calling {@link nnfw_set_input}
47 * <p>To support input and output setting, application can get
48 * input and output tensor information by calling<ul>
49 * <li>{@link nnfw_input_size}</li>
50 * <li>{@link nnfw_output_size}</li>
51 * <li>{@link nnfw_input_tensorinfo}</li>
52 * <li>{@link nnfw_output_tensorinfo}</li>
55 * <p>Application can inference many times using one session,
56 * but next inference can do after prior inference end
58 * <p>Application cannot use muitiple model using one session
60 typedef struct nnfw_session nnfw_session;
65 * The type of tensor represented in {@link nnfw_tensorinfo}
68 /** A tensor of 32 bit floating point */
69 NNFW_TYPE_TENSOR_FLOAT32 = 0,
70 /** A tensor of 32 bit signed integer */
71 NNFW_TYPE_TENSOR_INT32 = 1,
73 * A tensor of 8 bit integers that represent real numbers.
75 * real_value = (integer_value - zeroPoint) * scale.
77 NNFW_TYPE_TENSOR_QUANT8_ASYMM = 2,
78 /** A tensor of boolean */
79 NNFW_TYPE_TENSOR_BOOL = 3,
81 /** A tensor of 8 bit unsigned integer */
82 NNFW_TYPE_TENSOR_UINT8 = 4,
84 /** A tensor of 64 bit signed integer */
85 NNFW_TYPE_TENSOR_INT64 = 5,
90 * @brief Result values returned from a call to an API function
94 NNFW_STATUS_NO_ERROR = 0,
96 * An error code for general use.
97 * Mostly used when there is no specific value for that certain situation.
99 NNFW_STATUS_ERROR = 1,
100 /** Unexpected null argument is given. */
101 NNFW_STATUS_UNEXPECTED_NULL = 2,
102 /** When a function was called but it is not valid for the current session state. */
103 NNFW_STATUS_INVALID_STATE = 3,
104 /** When it is out of memory */
105 NNFW_STATUS_OUT_OF_MEMORY = 4,
109 * @brief Data format of a tensor
112 /** Don't care layout */
113 NNFW_LAYOUT_NONE = 0,
115 * Channel last layout
116 * If rank is 4, layout is NHWC
118 NNFW_LAYOUT_CHANNELS_LAST = 1,
120 * Channel first layout
121 * If rank is 4, layout is NCHW
123 NNFW_LAYOUT_CHANNELS_FIRST = 2,
127 * @brief Information ID for retrieving information on nnfw (e.g. version)
130 /** nnfw runtime version
131 * Its value is uint32 in 0xMMmmmmPP, where MM = major, mmmm = minor, PP = patch.
133 NNFW_INFO_ID_VERSION = 0,
137 * @brief Maximum rank expressible with nnfw
139 #define NNFW_MAX_RANK (6)
142 * @brief tensor info describes the type and shape of tensors
144 * <p>This structure is used to describe input and output tensors.
145 * Application can get input and output tensor type and shape described in model by using
146 * {@link nnfw_input_tensorinfo} and {@link nnfw_output_tensorinfo}
148 * <p>Maximum rank is 6 (NNFW_MAX_RANK). And tensor's dimension value is filled in 'dims' field from
150 * For example, if tensor's rank is 4,
151 * application can get dimension value from dims[0], dims[1], dims[2], and dims[3]
153 typedef struct nnfw_tensorinfo
157 /** The number of dimensions (rank) */
160 * The dimension of tensor.
161 * Maximum rank is 6 (NNFW_MAX_RANK).
163 int32_t dims[NNFW_MAX_RANK];
167 * @brief Create a new session instance.
169 * <p>This only creates a session.
170 * Model is loaded after {@link nnfw_load_model_from_file} is invoked.
171 * And inference is performed after {@link nnfw_run} is invoked.
173 * <p>{@link nnfw_close_session} should be called once
174 * if session is no longer need
176 * @param[out] session The session to be created
177 * @return NNFW_STATUS_NO_ERROR if successful
179 NNFW_STATUS nnfw_create_session(nnfw_session **session);
182 * @brief Close a session instance
184 * After called, access to closed session by application will be invalid
186 * @param[in] session The session to be closed
187 * @return @c NNFW_STATUS_NO_ERROR if successful
189 NNFW_STATUS nnfw_close_session(nnfw_session *session);
192 * @brief Load model from nnpackage file or directory
194 * The length of \p package_file_path must not execeed 1024 bytes including zero at the end.
196 * @param[in] session nnfw_session loading the given nnpackage file/dir
197 * @param[in] package_file_path Path to the nnpackage file or unzipped directory to be loaded
199 * @return @c NNFW_STATUS_NO_ERROR if successful
201 NNFW_STATUS nnfw_load_model_from_file(nnfw_session *session, const char *package_file_path);
204 * @brief Apply i-th input's tensor info to resize input tensor
206 * This function should be called before {@link nnfw_prepare} is invoked, and
207 * should be called after {@link nnfw_load_model_from_file} is invoked
208 * See {@link nnfw_prepare} for information applying updated tensor info
209 * If this function is called many times for same index, tensor info is overwritten
211 * @deprecated Deprecated since 1.7.0. Use {@link nnfw_set_input_tensorinfo} instead.
213 * @param[in] session Session to the input tensor info is to be set
214 * @param[in] index Index of input to be applied (0-indexed)
215 * @param[in] tensor_info Tensor info to be applied
216 * @return @c NNFW_STATUS_NO_ERROR if successful, otherwise return @c NNFW_STATUS_ERROR
218 NNFW_STATUS nnfw_apply_tensorinfo(nnfw_session *session, uint32_t index,
219 nnfw_tensorinfo tensor_info);
222 * @brief Set input model's tensor info for resizing
224 * This function can be called at any time after calling {@link nnfw_model_load_from_file}. Changing
225 * input tensor's shape will cause shape inference for the model. There are two different types of
226 * shape inference - static and dynamic. Which one to use is depend on the current state of the
228 * When it is called after calling {@link nnfw_model_load_from_file} and before calling {@link
229 * nnfw_prepare}, this info will be used when {@link nnfw_prepare}. And it will perform static shape
230 * inference for all tensors.
231 * When it is called after calling {@link nnfw_prepare} or even after {@link nnfw_run}, this info
232 * will be used when {@link nnfw_run}. And the shapes of the tensors are determined on the fly.
233 * If this function is called many times for the same index, it is overwritten.
235 * @param[in] session Session to the input tensor info is to be set
236 * @param[in] index Index of input to be set (0-indexed)
237 * @param[in] tensor_info Tensor info to be set
238 * @return @c NNFW_STATUS_NO_ERROR if successful, otherwise return @c NNFW_STATUS_ERROR
240 NNFW_STATUS nnfw_set_input_tensorinfo(nnfw_session *session, uint32_t index,
241 const nnfw_tensorinfo *tensor_info);
244 * @brief Prepare session to be ready for inference
246 * This phase may finalize model compilation, scheduling, and additional settings.
247 * If {@link nnfw_apply_tensor} is called to apply input tensor info different with model
248 * before this function, tries to resize all tensors.
250 * @param[in] session the session to be prepared
251 * @return @c NNFW_STATUS_NO_ERROR if successful, otherwise return @c NNFW_STATUS_ERROR
253 NNFW_STATUS nnfw_prepare(nnfw_session *session);
256 * @brief Run inference
258 * <p>This function should be called after model is loaded by {@link nnfw_load_model_from_file},
259 * session is prepared for inference by {@link nnfw_prepare}, set input and output buffers
260 * by {@link nnfw_set_input} and {@link nnfw_set_output}.</p>
262 * <p>This function return after inference is finished.</p>
264 * @param[in] session The session to run inference
265 * @return @c NNFW_STATUS_NO_ERROR if successful
267 NNFW_STATUS nnfw_run(nnfw_session *session);
270 * @brief Run inference asynchronously
272 * <p>This function must be called after model is loaded by {@link nnfw_load_model_from_file},
273 * session is prepared for inference by {@link nnfw_prepare}, set input and output buffers
274 * by {@link nnfw_set_input} and {@link nnfw_set_output}.</p>
276 * <p>This function returns immediately after starting a thread to run the inference.
277 * To get the result of it or to do the next inference with {@link nnfw_run} or
278 * {@link nnfw_run_async}, {@link nnfw_await} must be called to ensure the current asynchronous
279 * inference has finished. Only one asynchronous inference is allowed at a time for a session.
280 * If this function is called while the previous one is still running, it returns an error.</p>
282 * @param[in] session The session to run inference
283 * @return @c NNFW_STATUS_NO_ERROR if successful
285 NNFW_STATUS nnfw_run_async(nnfw_session *session);
288 * @brief Wait for asynchronous run to finish
290 * <p>This function must be called after calling {@link nnfw_run_asnyc}, and can be called only once
291 * for a {@link nnfw_run_async} call.
293 * <p>When this function returns, it means that this session has finished the asynchronous run. Then
294 * the user can safely use the output data.</p>
296 * <p>This function returns after the asynchronous inference is finished.</p>
298 * @param[in] session The session to run inference
299 * @return @c NNFW_STATUS_NO_ERROR if successful
301 NNFW_STATUS nnfw_await(nnfw_session *session);
304 * @brief Set input buffer
306 * This function must be called after {@link nnfw_prepare}, \p buffer given to this function can be
307 * reused for many inferences. \p length must be greater or equal than the operand requires. To
308 * specify an optional input, you can either not call this for that input or call this with \p
309 * buffer of NULL and \p length of 0.
311 * @param[in] session Session to the input is to be set
312 * @param[in] index Index of input to be set (0-indexed)
313 * @param[in] type Type of the input
314 * @param[in] buffer Raw buffer for input
315 * @param[in] length Size of bytes of input buffer
317 * @return @c NNFW_STATUS_NO_ERROR if successful
319 NNFW_STATUS nnfw_set_input(nnfw_session *session, uint32_t index, NNFW_TYPE type,
320 const void *buffer, size_t length);
323 * @brief Set output buffer
325 * This function must be called after {@link nnfw_prepare}, \p buffer given to this function can be
326 * reused for many inferences. \p length must be greater or equal than the operand requires. An
327 * output operand can have unspecified shape and deduced dynamically during the execution. You must
328 * provide \p buffer large enough.
330 * @param[in] session Session from inference output is to be extracted
331 * @param[in] index Index of output to be set (0-indexed)
332 * @param[in] type Type of the output
333 * @param[out] buffer Raw buffer for output
334 * @param[in] length Size of bytes of output buffer
336 * @return @c NNFW_STATUS_NO_ERROR if successful
338 NNFW_STATUS nnfw_set_output(nnfw_session *session, uint32_t index, NNFW_TYPE type, void *buffer,
342 * @brief Get the number of inputs
344 * Application can call this function to get number of inputs defined in loaded model.
345 * This function should be called after {@link nnfw_load_model_from_file} is invoked to load model
347 * @param[in] session Session from input information is to be extracted
348 * @param[out] number Variable which the number of inputs is put into
350 * @return @c NNFW_STATUS_NO_ERROR if successful
352 NNFW_STATUS nnfw_input_size(nnfw_session *session, uint32_t *number);
355 * @brief Get the number of outputs
357 * Application can call this function to get number of outputs defined in loaded model.
358 * This function should be called after {@link nnfw_load_model_from_file} is invoked to load model
360 * @param[in] session Session from output information is to be extracted
361 * @param[out] number Variable which the number of outputs is put into
363 * @return @c NNFW_STATUS_NO_ERROR if successful
365 NNFW_STATUS nnfw_output_size(nnfw_session *session, uint32_t *number);
368 * @brief Set the layout of an input
370 * The input that does not call this has NNFW_LAYOUT_NHWC layout
372 * @param[in] session session from inference input is to be extracted
373 * @param[in] index index of input to be set (0-indexed)
374 * @param[in] layout layout to set to target input
376 * @return NNFW_STATUS_NO_ERROR if successful
378 NNFW_STATUS nnfw_set_input_layout(nnfw_session *session, uint32_t index, NNFW_LAYOUT layout);
381 * @brief Set the layout of an output
383 * The output that does not call this has NNFW_LAYOUT_NHWC layout
385 * @param[in] session session from inference output is to be extracted
386 * @param[in] index index of output to be set (0-indexed)
387 * @param[in] layout layout to set to target output
389 * @return NNFW_STATUS_NO_ERROR if successful
391 NNFW_STATUS nnfw_set_output_layout(nnfw_session *session, uint32_t index, NNFW_LAYOUT layout);
394 * @brief Get i-th input tensor info
396 * <p>Before {@link nnfw_prepare} is invoked, this function return tensor info in model,
397 * so updated tensor info by {@link nnfw_apply_tensorinfo} is not returned.</p>
399 * <p>After {@link nnfw_prepare} is invoked, this function return updated tensor info
400 * if tensor info is updated by {@link nnfw_apply_tensorinfo}.</p>
402 * @param[in] session Session from input information is to be extracted
403 * @param[in] index Index of input
404 * @param[out] tensor_info Tensor info (shape, type, etc)
406 * @return @c NNFW_STATUS_NO_ERROR if successful
408 NNFW_STATUS nnfw_input_tensorinfo(nnfw_session *session, uint32_t index,
409 nnfw_tensorinfo *tensor_info);
412 * @brief Get i-th output tensor info
414 * <p>After {@link nnfw_load_model_from_file} and before {@link nnfw_prepare} is invoked, it returns
415 * tensor info in the model.</p>
417 * <p>After {@link nnfw_prepare} and before {@link nnfw_run} is invoked, this function returns
418 * updated tensor info if tensor info is updated by {@link nnfw_set_input_tensorinfo}.</p>
420 * <p>After {@link nnfw_run} is invoked(at least once), it returns the updated tensor info during
421 * the latest execution.</p>
423 * @param[in] session Session from output information is to be extracted
424 * @param[in] index Index of output
425 * @param[out] tensor_info Tensor info (shape, type, etc)
427 * @return @c NNFW_STATUS_NO_ERROR if successful
429 NNFW_STATUS nnfw_output_tensorinfo(nnfw_session *session, uint32_t index,
430 nnfw_tensorinfo *tensor_info);
433 * @brief Set available backends
435 * This function should be called before {@link nnfw_prepare} is invoked.
437 * <p>Supported backends differs on each platforms.
438 * For example, `x86_64` supports "cpu" only.
439 * Multiple backends can be set and they must be separated by a semicolon (ex: "acl_cl;cpu").
440 * For each backend string, `libbackend_{backend}.so` will be dynamically loaded during
441 * {@link nnfw_prepare}.
442 * Among the multiple backends, the 1st element is used as the default backend.</p>
444 * @param[in] session session to which avilable backends are set
445 * @param[in] backends available backends on which nnfw uses
447 * @return @c NNFW_STATUS_NO_ERROR if successful
449 NNFW_STATUS nnfw_set_available_backends(nnfw_session *session, const char *backends);
452 * @brief Set the operation's backend
454 * This function should be called before {@link nnfw_prepare} is invoked.
456 * <p>The backend for op has higher priority than available backends specified by
457 * {@link nnfw_set_available_backends}.</p>
459 * @deprecated Deprecated since 1.8.0.
461 * @param[in] session session to be modified
462 * @param[in] op operation to be set
463 * @param[in] backend bakcend on which operation run
465 * @return @c NNFW_STATUS_NO_ERROR if successful
467 NNFW_STATUS nnfw_set_op_backend(nnfw_session *session, const char *op, const char *backend);
470 * @brief Retrieve uint32 type of nnfw information for given information ID.
472 * <p>Retrieves the information of property given by information id </p>
474 * @note: The input session could be null for global information (e.g. runtime version).*
476 * @param[in] session session to be queried on.
477 * @param[in] information ID to be queried
478 * @param[out] val uint32 value to be returned.
480 * @return @c NNFW_STATUS_NO_ERROR if successful
482 NNFW_STATUS nnfw_query_info_u32(nnfw_session *session, NNFW_INFO_ID id, uint32_t *val);