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,
105 * @brief Data format of a tensor
108 /** Don't care layout */
109 NNFW_LAYOUT_NONE = 0,
111 * Channel last layout
112 * If rank is 4, layout is NHWC
114 NNFW_LAYOUT_CHANNELS_LAST = 1,
116 * Channel first layout
117 * If rank is 4, layout is NCHW
119 NNFW_LAYOUT_CHANNELS_FIRST = 2,
123 * @brief Information ID for retrieving information on nnfw (e.g. version)
126 /** nnfw runtime version
127 * Its value is uint32 in 0xMMmmmmPP, where MM = major, mmmm = minor, PP = patch.
129 NNFW_INFO_ID_VERSION = 0,
133 * @brief Maximum rank expressible with nnfw
135 #define NNFW_MAX_RANK (6)
138 * @brief tensor info describes the type and shape of tensors
140 * <p>This structure is used to describe input and output tensors.
141 * Application can get input and output tensor type and shape described in model by using
142 * {@link nnfw_input_tensorinfo} and {@link nnfw_output_tensorinfo}
144 * <p>Maximum rank is 6 (NNFW_MAX_RANK). And tensor's dimension value is filled in 'dims' field from
146 * For example, if tensor's rank is 4,
147 * application can get dimension value from dims[0], dims[1], dims[2], and dims[3]
149 typedef struct nnfw_tensorinfo
153 /** The number of dimensions (rank) */
156 * The dimension of tensor.
157 * Maximum rank is 6 (NNFW_MAX_RANK).
159 int32_t dims[NNFW_MAX_RANK];
163 * @brief Create a new session instance.
165 * <p>This only creates a session.
166 * Model is loaded after {@link nnfw_load_model_from_file} is invoked.
167 * And inference is performed after {@link nnfw_run} is invoked.
169 * <p>{@link nnfw_close_session} should be called once
170 * if session is no longer need
172 * @param[out] session The session to be created
173 * @return NNFW_STATUS_NO_ERROR if successful
175 NNFW_STATUS nnfw_create_session(nnfw_session **session);
178 * @brief Close a session instance
180 * After called, access to closed session by application will be invalid
182 * @param[in] session The session to be closed
183 * @return @c NNFW_STATUS_NO_ERROR if successful
185 NNFW_STATUS nnfw_close_session(nnfw_session *session);
188 * @brief Load model from nnpackage file or directory
190 * The length of \p package_file_path must not execeed 1024 bytes including zero at the end.
192 * @param[in] session nnfw_session loading the given nnpackage file/dir
193 * @param[in] package_file_path Path to the nnpackage file or unzipped directory to be loaded
195 * @return @c NNFW_STATUS_NO_ERROR if successful
197 NNFW_STATUS nnfw_load_model_from_file(nnfw_session *session, const char *package_file_path);
200 * @brief Apply i-th input's tensor info to resize input tensor
202 * This function should be called before {@link nnfw_prepare} is invoked, and
203 * should be called after {@link nnfw_load_model_from_file} is invoked
204 * See {@link nnfw_prepare} for information applying updated tensor info
205 * If this function is called many times for same index, tensor info is overwritten
207 * @deprecated Deprecated since 1.7.0. Use {@link nnfw_set_input_tensorinfo} instead.
209 * @param[in] session Session to the input tensor info is to be set
210 * @param[in] index Index of input to be applied (0-indexed)
211 * @param[in] tensor_info Tensor info to be applied
212 * @return @c NNFW_STATUS_NO_ERROR if successful, otherwise return @c NNFW_STATUS_ERROR
214 NNFW_STATUS nnfw_apply_tensorinfo(nnfw_session *session, uint32_t index,
215 nnfw_tensorinfo tensor_info);
218 * @brief Set input model's tensor info for resizing
220 * This function can be called at any time after calling {@link nnfw_model_load_from_file}. Changing
221 * input tensor's shape will cause shape inference for the model. There are two different types of
222 * shape inference - static and dynamic. Which one to use is depend on the current state of the
224 * When it is called after calling {@link nnfw_model_load_from_file} and before calling {@link
225 * nnfw_prepare}, this info will be used when {@link nnfw_prepare}. And it will perform static shape
226 * inference for all tensors.
227 * When it is called after calling {@link nnfw_prepare} or even after {@link nnfw_run}, this info
228 * will be used when {@link nnfw_run}. And the shapes of the tensors are determined on the fly.
229 * If this function is called many times for the same index, it is overwritten.
231 * @param[in] session Session to the input tensor info is to be set
232 * @param[in] index Index of input to be set (0-indexed)
233 * @param[in] tensor_info Tensor info to be set
234 * @return @c NNFW_STATUS_NO_ERROR if successful, otherwise return @c NNFW_STATUS_ERROR
236 NNFW_STATUS nnfw_set_input_tensorinfo(nnfw_session *session, uint32_t index,
237 const nnfw_tensorinfo *tensor_info);
240 * @brief Prepare session to be ready for inference
242 * This phase may finalize model compilation, scheduling, and additional settings.
243 * If {@link nnfw_apply_tensor} is called to apply input tensor info different with model
244 * before this function, tries to resize all tensors.
246 * @param[in] session the session to be prepared
247 * @return @c NNFW_STATUS_NO_ERROR if successful, otherwise return @c NNFW_STATUS_ERROR
249 NNFW_STATUS nnfw_prepare(nnfw_session *session);
252 * @brief Run inference
254 * <p>This function should be called after model is loaded by {@link nnfw_load_model_from_file},
255 * session is prepared for inference by {@link nnfw_prepare}, set input and output buffers
256 * by {@link nnfw_set_input} and {@link nnfw_set_output}.</p>
258 * <p>This function return after inference is finished.</p>
260 * @param[in] session The session to run inference
261 * @return @c NNFW_STATUS_NO_ERROR if successful
263 NNFW_STATUS nnfw_run(nnfw_session *session);
266 * @brief Run inference asynchronously
268 * <p>This function must be called after model is loaded by {@link nnfw_load_model_from_file},
269 * session is prepared for inference by {@link nnfw_prepare}, set input and output buffers
270 * by {@link nnfw_set_input} and {@link nnfw_set_output}.</p>
272 * <p>This function returns immediately after starting a thread to run the inference.
273 * To get the result of it or to do the next inference with {@link nnfw_run} or
274 * {@link nnfw_run_async}, {@link nnfw_await} must be called to ensure the current asynchronous
275 * inference has finished. Only one asynchronous inference is allowed at a time for a session.
276 * If this function is called while the previous one is still running, it returns an error.</p>
278 * @param[in] session The session to run inference
279 * @return @c NNFW_STATUS_NO_ERROR if successful
281 NNFW_STATUS nnfw_run_async(nnfw_session *session);
284 * @brief Wait for asynchronous run to finish
286 * <p>This function must be called after calling {@link nnfw_run_asnyc}, and can be called only once
287 * for a {@link nnfw_run_async} call.
289 * <p>When this function returns, it means that this session has finished the asynchronous run. Then
290 * the user can safely use the output data.</p>
292 * <p>This function returns after the asynchronous inference is finished.</p>
294 * @param[in] session The session to run inference
295 * @return @c NNFW_STATUS_NO_ERROR if successful
297 NNFW_STATUS nnfw_await(nnfw_session *session);
300 * @brief Set input buffer
302 * This function must be called after {@link nnfw_prepare}, \p buffer given to this function can be
303 * reused for many inferences. \p length must be greater or equal than the operand requires. To
304 * specify an optional input, you can either not call this for that input or call this with \p
305 * buffer of NULL and \p length of 0.
307 * @param[in] session Session to the input is to be set
308 * @param[in] index Index of input to be set (0-indexed)
309 * @param[in] type Type of the input
310 * @param[in] buffer Raw buffer for input
311 * @param[in] length Size of bytes of input buffer
313 * @return @c NNFW_STATUS_NO_ERROR if successful
315 NNFW_STATUS nnfw_set_input(nnfw_session *session, uint32_t index, NNFW_TYPE type,
316 const void *buffer, size_t length);
319 * @brief Set output buffer
321 * This function must be called after {@link nnfw_prepare}, \p buffer given to this function can be
322 * reused for many inferences. \p length must be greater or equal than the operand requires. An
323 * output operand can have unspecified shape and deduced dynamically during the execution. You must
324 * provide \p buffer large enough.
326 * @param[in] session Session from inference output is to be extracted
327 * @param[in] index Index of output to be set (0-indexed)
328 * @param[in] type Type of the output
329 * @param[out] buffer Raw buffer for output
330 * @param[in] length Size of bytes of output buffer
332 * @return @c NNFW_STATUS_NO_ERROR if successful
334 NNFW_STATUS nnfw_set_output(nnfw_session *session, uint32_t index, NNFW_TYPE type, void *buffer,
338 * @brief Get the number of inputs
340 * Application can call this function to get number of inputs defined in loaded model.
341 * This function should be called after {@link nnfw_load_model_from_file} is invoked to load model
343 * @param[in] session Session from input information is to be extracted
344 * @param[out] number Variable which the number of inputs is put into
346 * @return @c NNFW_STATUS_NO_ERROR if successful
348 NNFW_STATUS nnfw_input_size(nnfw_session *session, uint32_t *number);
351 * @brief Get the number of outputs
353 * Application can call this function to get number of outputs defined in loaded model.
354 * This function should be called after {@link nnfw_load_model_from_file} is invoked to load model
356 * @param[in] session Session from output information is to be extracted
357 * @param[out] number Variable which the number of outputs is put into
359 * @return @c NNFW_STATUS_NO_ERROR if successful
361 NNFW_STATUS nnfw_output_size(nnfw_session *session, uint32_t *number);
364 * @brief Set the layout of an input
366 * The input that does not call this has NNFW_LAYOUT_NHWC layout
368 * @param[in] session session from inference input is to be extracted
369 * @param[in] index index of input to be set (0-indexed)
370 * @param[in] layout layout to set to target input
372 * @return NNFW_STATUS_NO_ERROR if successful
374 NNFW_STATUS nnfw_set_input_layout(nnfw_session *session, uint32_t index, NNFW_LAYOUT layout);
377 * @brief Set the layout of an output
379 * The output that does not call this has NNFW_LAYOUT_NHWC layout
381 * @param[in] session session from inference output is to be extracted
382 * @param[in] index index of output to be set (0-indexed)
383 * @param[in] layout layout to set to target output
385 * @return NNFW_STATUS_NO_ERROR if successful
387 NNFW_STATUS nnfw_set_output_layout(nnfw_session *session, uint32_t index, NNFW_LAYOUT layout);
390 * @brief Get i-th input tensor info
392 * <p>Before {@link nnfw_prepare} is invoked, this function return tensor info in model,
393 * so updated tensor info by {@link nnfw_apply_tensorinfo} is not returned.</p>
395 * <p>After {@link nnfw_prepare} is invoked, this function return updated tensor info
396 * if tensor info is updated by {@link nnfw_apply_tensorinfo}.</p>
398 * @param[in] session Session from input information is to be extracted
399 * @param[in] index Index of input
400 * @param[out] tensor_info Tensor info (shape, type, etc)
402 * @return @c NNFW_STATUS_NO_ERROR if successful
404 NNFW_STATUS nnfw_input_tensorinfo(nnfw_session *session, uint32_t index,
405 nnfw_tensorinfo *tensor_info);
408 * @brief Get i-th output tensor info
410 * <p>After {@link nnfw_load_model_from_file} and before {@link nnfw_prepare} is invoked, it returns
411 * tensor info in the model.</p>
413 * <p>After {@link nnfw_prepare} and before {@link nnfw_run} is invoked, this function returns
414 * updated tensor info if tensor info is updated by {@link nnfw_set_input_tensorinfo}.</p>
416 * <p>After {@link nnfw_run} is invoked(at least once), it returns the updated tensor info during
417 * the latest execution.</p>
419 * @param[in] session Session from output information is to be extracted
420 * @param[in] index Index of output
421 * @param[out] tensor_info Tensor info (shape, type, etc)
423 * @return @c NNFW_STATUS_NO_ERROR if successful
425 NNFW_STATUS nnfw_output_tensorinfo(nnfw_session *session, uint32_t index,
426 nnfw_tensorinfo *tensor_info);
429 * @brief Set available backends
431 * This function should be called before {@link nnfw_prepare} is invoked.
433 * <p>Supported backends differs on each platforms.
434 * For example, `x86_64` supports "cpu" only.
435 * Can set multiple backends by semicolon (ex: "acl_cl;cpu").
436 * Among the multiple backends, the 1st element is used as default backend.</p>
438 * @note Possible backend strings are: "cpu", "acl_cl", "acl_neon", "srcn"
440 * @param[in] session session to which avilable backends are set
441 * @param[in] backends available backends on which nnfw uses
443 * @return @c NNFW_STATUS_NO_ERROR if successful
445 NNFW_STATUS nnfw_set_available_backends(nnfw_session *session, const char *backends);
448 * @brief Set the operation's backend
450 * This function should be called before {@link nnfw_prepare} is invoked.
452 * <p>Supported backends differs on each platforms.
453 * For example, `x86_64` supports "cpu" only.
454 * The backend for op has higher priority than available backends specified by
455 * nnfw_set_available_backends.</p>
457 * @note Possible backend strings are: "cpu", "acl_cl", "acl_neon"
459 * @param[in] session session to be modified
460 * @param[in] op operation to be set
461 * @param[in] backend bakcend on which operation run
463 * @return @c NNFW_STATUS_NO_ERROR if successful
465 NNFW_STATUS nnfw_set_op_backend(nnfw_session *session, const char *op, const char *backend);
468 * @brief Retrieve uint32 type of nnfw information for given information ID.
470 * <p>Retrieves the information of property given by information id </p>
472 * @note: The input session could be null for global information (e.g. runtime version).*
474 * @param[in] session session to be queried on.
475 * @param[in] information ID to be queried
476 * @param[out] val uint32 value to be returned.
478 * @return @c NNFW_STATUS_NO_ERROR if successful
480 NNFW_STATUS nnfw_query_info_u32(nnfw_session *session, NNFW_INFO_ID id, uint32_t *val);