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,
106 /** When it was given an insufficient output buffer */
107 NNFW_STATUS_INSUFFICIENT_OUTPUT_SIZE = 5,
111 * @brief Data format of a tensor
114 /** Don't care layout */
115 NNFW_LAYOUT_NONE = 0,
117 * Channel last layout
118 * If rank is 4, layout is NHWC
120 NNFW_LAYOUT_CHANNELS_LAST = 1,
122 * Channel first layout
123 * If rank is 4, layout is NCHW
125 NNFW_LAYOUT_CHANNELS_FIRST = 2,
129 * @brief Information ID for retrieving information on nnfw (e.g. version)
132 /** nnfw runtime version
133 * Its value is uint32 in 0xMMmmmmPP, where MM = major, mmmm = minor, PP = patch.
135 NNFW_INFO_ID_VERSION = 0,
139 * @brief Maximum rank expressible with nnfw
141 #define NNFW_MAX_RANK (6)
144 * @brief tensor info describes the type and shape of tensors
146 * <p>This structure is used to describe input and output tensors.
147 * Application can get input and output tensor type and shape described in model by using
148 * {@link nnfw_input_tensorinfo} and {@link nnfw_output_tensorinfo}
150 * <p>Maximum rank is 6 (NNFW_MAX_RANK). And tensor's dimension value is filled in 'dims' field from
152 * For example, if tensor's rank is 4,
153 * application can get dimension value from dims[0], dims[1], dims[2], and dims[3]
155 typedef struct nnfw_tensorinfo
159 /** The number of dimensions (rank) */
162 * The dimension of tensor.
163 * Maximum rank is 6 (NNFW_MAX_RANK).
165 int32_t dims[NNFW_MAX_RANK];
169 * @brief Create a new session instance.
171 * <p>This only creates a session.
172 * Model is loaded after {@link nnfw_load_model_from_file} is invoked.
173 * And inference is performed after {@link nnfw_run} is invoked.
175 * <p>{@link nnfw_close_session} should be called once
176 * if session is no longer need
178 * @param[out] session The session to be created
179 * @return NNFW_STATUS_NO_ERROR if successful
181 NNFW_STATUS nnfw_create_session(nnfw_session **session);
184 * @brief Close a session instance
186 * After called, access to closed session by application will be invalid
188 * @param[in] session The session to be closed
189 * @return @c NNFW_STATUS_NO_ERROR if successful
191 NNFW_STATUS nnfw_close_session(nnfw_session *session);
194 * @brief Load model from nnpackage file or directory
196 * The length of \p package_file_path must not execeed 1024 bytes including zero at the end.
198 * @param[in] session nnfw_session loading the given nnpackage file/dir
199 * @param[in] package_file_path Path to the nnpackage file or unzipped directory to be loaded
201 * @return @c NNFW_STATUS_NO_ERROR if successful
203 NNFW_STATUS nnfw_load_model_from_file(nnfw_session *session, const char *package_file_path);
206 * @brief Apply i-th input's tensor info to resize input tensor
208 * This function should be called before {@link nnfw_prepare} is invoked, and
209 * should be called after {@link nnfw_load_model_from_file} is invoked
210 * See {@link nnfw_prepare} for information applying updated tensor info
211 * If this function is called many times for same index, tensor info is overwritten
213 * @deprecated Deprecated since 1.7.0. Use {@link nnfw_set_input_tensorinfo} instead.
215 * @param[in] session Session to the input tensor info is to be set
216 * @param[in] index Index of input to be applied (0-indexed)
217 * @param[in] tensor_info Tensor info to be applied
218 * @return @c NNFW_STATUS_NO_ERROR if successful, otherwise return @c NNFW_STATUS_ERROR
220 NNFW_STATUS nnfw_apply_tensorinfo(nnfw_session *session, uint32_t index,
221 nnfw_tensorinfo tensor_info);
224 * @brief Set input model's tensor info for resizing
226 * This function can be called at any time after calling {@link nnfw_model_load_from_file}. Changing
227 * input tensor's shape will cause shape inference for the model. There are two different types of
228 * shape inference - static and dynamic. Which one to use is depend on the current state of the
230 * When it is called after calling {@link nnfw_model_load_from_file} and before calling {@link
231 * nnfw_prepare}, this info will be used when {@link nnfw_prepare}. And it will perform static shape
232 * inference for all tensors.
233 * When it is called after calling {@link nnfw_prepare} or even after {@link nnfw_run}, this info
234 * will be used when {@link nnfw_run}. And the shapes of the tensors are determined on the fly.
235 * If this function is called many times for the same index, it is overwritten.
237 * @param[in] session Session to the input tensor info is to be set
238 * @param[in] index Index of input to be set (0-indexed)
239 * @param[in] tensor_info Tensor info to be set
240 * @return @c NNFW_STATUS_NO_ERROR if successful, otherwise return @c NNFW_STATUS_ERROR
242 NNFW_STATUS nnfw_set_input_tensorinfo(nnfw_session *session, uint32_t index,
243 const nnfw_tensorinfo *tensor_info);
246 * @brief Prepare session to be ready for inference
248 * This phase may finalize model compilation, scheduling, and additional settings.
249 * If {@link nnfw_apply_tensor} is called to apply input tensor info different with model
250 * before this function, tries to resize all tensors.
252 * @param[in] session the session to be prepared
253 * @return @c NNFW_STATUS_NO_ERROR if successful, otherwise return @c NNFW_STATUS_ERROR
255 NNFW_STATUS nnfw_prepare(nnfw_session *session);
258 * @brief Run inference
260 * <p>This function should be called after model is loaded by {@link nnfw_load_model_from_file},
261 * session is prepared for inference by {@link nnfw_prepare}, set input and output buffers
262 * by {@link nnfw_set_input} and {@link nnfw_set_output}.</p>
264 * <p>This function return after inference is finished.</p>
266 * @param[in] session The session to run inference
267 * @return @c NNFW_STATUS_NO_ERROR if successful
269 NNFW_STATUS nnfw_run(nnfw_session *session);
272 * @brief Run inference asynchronously
274 * <p>This function must be called after model is loaded by {@link nnfw_load_model_from_file},
275 * session is prepared for inference by {@link nnfw_prepare}, set input and output buffers
276 * by {@link nnfw_set_input} and {@link nnfw_set_output}.</p>
278 * <p>This function returns immediately after starting a thread to run the inference.
279 * To get the result of it or to do the next inference with {@link nnfw_run} or
280 * {@link nnfw_run_async}, {@link nnfw_await} must be called to ensure the current asynchronous
281 * inference has finished. Only one asynchronous inference is allowed at a time for a session.
282 * If this function is called while the previous one is still running, it returns an error.</p>
284 * @param[in] session The session to run inference
285 * @return @c NNFW_STATUS_NO_ERROR if successful
287 NNFW_STATUS nnfw_run_async(nnfw_session *session);
290 * @brief Wait for asynchronous run to finish
292 * <p>This function must be called after calling {@link nnfw_run_asnyc}, and can be called only once
293 * for a {@link nnfw_run_async} call.
295 * <p>When this function returns, it means that this session has finished the asynchronous run. Then
296 * the user can safely use the output data.</p>
298 * <p>This function returns after the asynchronous inference is finished.</p>
300 * @param[in] session The session to run inference
301 * @return @c NNFW_STATUS_NO_ERROR if successful
303 NNFW_STATUS nnfw_await(nnfw_session *session);
306 * @brief Set input buffer
308 * This function must be called after {@link nnfw_prepare}, \p buffer given to this function can be
309 * reused for many inferences. \p length must be greater or equal than the operand requires. To
310 * specify an optional input, you can either not call this for that input or call this with \p
311 * buffer of NULL and \p length of 0.
313 * @param[in] session Session to the input is to be set
314 * @param[in] index Index of input to be set (0-indexed)
315 * @param[in] type Type of the input
316 * @param[in] buffer Raw buffer for input
317 * @param[in] length Size of bytes of input buffer
319 * @return @c NNFW_STATUS_NO_ERROR if successful
321 NNFW_STATUS nnfw_set_input(nnfw_session *session, uint32_t index, NNFW_TYPE type,
322 const void *buffer, size_t length);
325 * @brief Set output buffer
327 * This function must be called after {@link nnfw_prepare}, \p buffer given to this function can be
328 * reused for many inferences. \p length must be greater or equal than the operand requires. An
329 * output operand can have unspecified shape and deduced dynamically during the execution. You must
330 * provide \p buffer large enough.
332 * @param[in] session Session from inference output is to be extracted
333 * @param[in] index Index of output to be set (0-indexed)
334 * @param[in] type Type of the output
335 * @param[out] buffer Raw buffer for output
336 * @param[in] length Size of bytes of output buffer
338 * @return @c NNFW_STATUS_NO_ERROR if successful
340 NNFW_STATUS nnfw_set_output(nnfw_session *session, uint32_t index, NNFW_TYPE type, void *buffer,
344 * @brief Get the number of inputs
346 * Application can call this function to get number of inputs defined in loaded model.
347 * This function should be called after {@link nnfw_load_model_from_file} is invoked to load model
349 * @param[in] session Session from input information is to be extracted
350 * @param[out] number Variable which the number of inputs is put into
352 * @return @c NNFW_STATUS_NO_ERROR if successful
354 NNFW_STATUS nnfw_input_size(nnfw_session *session, uint32_t *number);
357 * @brief Get the number of outputs
359 * Application can call this function to get number of outputs defined in loaded model.
360 * This function should be called after {@link nnfw_load_model_from_file} is invoked to load model
362 * @param[in] session Session from output information is to be extracted
363 * @param[out] number Variable which the number of outputs is put into
365 * @return @c NNFW_STATUS_NO_ERROR if successful
367 NNFW_STATUS nnfw_output_size(nnfw_session *session, uint32_t *number);
370 * @brief Set the layout of an input
372 * The input that does not call this has NNFW_LAYOUT_NHWC layout
374 * @param[in] session session from inference input is to be extracted
375 * @param[in] index index of input to be set (0-indexed)
376 * @param[in] layout layout to set to target input
378 * @return NNFW_STATUS_NO_ERROR if successful
380 NNFW_STATUS nnfw_set_input_layout(nnfw_session *session, uint32_t index, NNFW_LAYOUT layout);
383 * @brief Set the layout of an output
385 * The output that does not call this has NNFW_LAYOUT_NHWC layout
387 * @param[in] session session from inference output is to be extracted
388 * @param[in] index index of output to be set (0-indexed)
389 * @param[in] layout layout to set to target output
391 * @return NNFW_STATUS_NO_ERROR if successful
393 NNFW_STATUS nnfw_set_output_layout(nnfw_session *session, uint32_t index, NNFW_LAYOUT layout);
396 * @brief Get i-th input tensor info
398 * <p>Before {@link nnfw_prepare} is invoked, this function return tensor info in model,
399 * so updated tensor info by {@link nnfw_apply_tensorinfo} is not returned.</p>
401 * <p>After {@link nnfw_prepare} is invoked, this function return updated tensor info
402 * if tensor info is updated by {@link nnfw_apply_tensorinfo}.</p>
404 * @param[in] session Session from input information is to be extracted
405 * @param[in] index Index of input
406 * @param[out] tensor_info Tensor info (shape, type, etc)
408 * @return @c NNFW_STATUS_NO_ERROR if successful
410 NNFW_STATUS nnfw_input_tensorinfo(nnfw_session *session, uint32_t index,
411 nnfw_tensorinfo *tensor_info);
414 * @brief Get i-th output tensor info
416 * <p>After {@link nnfw_load_model_from_file} and before {@link nnfw_prepare} is invoked, it returns
417 * tensor info in the model.</p>
419 * <p>After {@link nnfw_prepare} and before {@link nnfw_run} is invoked, this function returns
420 * updated tensor info if tensor info is updated by {@link nnfw_set_input_tensorinfo}.</p>
422 * <p>After {@link nnfw_run} is invoked(at least once), it returns the updated tensor info during
423 * the latest execution.</p>
425 * @param[in] session Session from output information is to be extracted
426 * @param[in] index Index of output
427 * @param[out] tensor_info Tensor info (shape, type, etc)
429 * @return @c NNFW_STATUS_NO_ERROR if successful
431 NNFW_STATUS nnfw_output_tensorinfo(nnfw_session *session, uint32_t index,
432 nnfw_tensorinfo *tensor_info);
435 * @brief Set available backends
437 * This function should be called before {@link nnfw_prepare} is invoked.
439 * <p>Supported backends differs on each platforms.
440 * For example, `x86_64` supports "cpu" only.
441 * Multiple backends can be set and they must be separated by a semicolon (ex: "acl_cl;cpu").
442 * For each backend string, `libbackend_{backend}.so` will be dynamically loaded during
443 * {@link nnfw_prepare}.
444 * Among the multiple backends, the 1st element is used as the default backend.</p>
446 * @param[in] session session to which avilable backends are set
447 * @param[in] backends available backends on which nnfw uses
449 * @return @c NNFW_STATUS_NO_ERROR if successful
451 NNFW_STATUS nnfw_set_available_backends(nnfw_session *session, const char *backends);
454 * @brief Set the operation's backend
456 * This function should be called before {@link nnfw_prepare} is invoked.
458 * <p>The backend for op has higher priority than available backends specified by
459 * {@link nnfw_set_available_backends}.</p>
461 * @deprecated Deprecated since 1.8.0.
463 * @param[in] session session to be modified
464 * @param[in] op operation to be set
465 * @param[in] backend bakcend on which operation run
467 * @return @c NNFW_STATUS_NO_ERROR if successful
469 NNFW_STATUS nnfw_set_op_backend(nnfw_session *session, const char *op, const char *backend);
472 * @brief Retrieve uint32 type of nnfw information for given information ID.
474 * <p>Retrieves the information of property given by information id </p>
476 * @note: The input session could be null for global information (e.g. runtime version).*
478 * @param[in] session session to be queried on.
479 * @param[in] information ID to be queried
480 * @param[out] val uint32 value to be returned.
482 * @return @c NNFW_STATUS_NO_ERROR if successful
484 NNFW_STATUS nnfw_query_info_u32(nnfw_session *session, NNFW_INFO_ID id, uint32_t *val);