FFmpeg: libavfilter/dnn/dnn_backend_torch.cpp Source File
Go to the documentation of this file. 1 /*
2 * Copyright (c) 2024
3 *
4 * This file is part of FFmpeg.
5 *
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
10 *
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 */
20
21 /**
22 * @file
23 * DNN Torch backend implementation.
24 */
25
26 #include <torch/torch.h>
27 #include <torch/script.h>
28
29 extern "C" {
36 }
37
46
51
57
58
59 #define OFFSET(x) offsetof(THOptions, x)
60 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM
64 };
65
67 {
71 if (!lltask) {
74 }
77 lltask->task = task;
82 }
83 return 0;
84 }
85
87 {
88 if (!request)
89 return;
93 }
97 }
98 return;
99 }
100
102 {
105 return;
106 }
113 }
114
116 {
118 if (!model || !*model)
119 return;
120
121 th_model = (
THModel *) (*model);
125 }
127
131 }
133
139 }
144 }
145
147 {
155 return 0;
156 }
157
159 {
161 }
162
164 {
170 int ret, width_idx, height_idx, channel_idx;
171
173 if (!lltask) {
175 goto err;
176 }
180
183 goto err;
184 }
191 input.dims[channel_idx] *
sizeof(
float));
195 infer_request->
output =
new torch::Tensor();
196
203 } else {
205 }
206 }
207 break;
208 default:
210 break;
211 }
213 {1, input.dims[channel_idx], input.dims[height_idx], input.dims[width_idx]},
215 return 0;
216
217 err:
220 }
221
223 {
230 std::vector<torch::jit::IValue>
inputs;
231 torch::NoGradGuard no_grad;
232
233 if (!request) {
236 }
242
243 if (
ctx->torch_option.optimize)
244 torch::jit::setGraphExecutorOptimize(true);
245 else
246 torch::jit::setGraphExecutorOptimize(false);
247
251 }
252 // Transfer tensor to the same device as model
253 c10::Device device = (*th_model->
jit_model->parameters().begin()).device();
257
259
260 return 0;
261 }
262
271
276 if (
sizes.size() == 4) {
277 // 4 dimensions: [batch_size, channel, height, width]
278 // this format of data is normally used for video frame SR
283 } else {
285 goto err;
286 }
287
291 // Post process can only deal with CPU memory.
292 if (
output->device() != torch::kCPU)
298 } else {
300 }
301 } else {
304 }
305 break;
306 default:
308 goto err;
309 }
312 err:
314
317 av_log(th_model->
ctx,
AV_LOG_ERROR,
"Unable to push back request_queue when failed to start inference.\n");
318 }
319 }
320
322 {
327
330 return 0;
331 }
332
334 if (lltask ==
NULL) {
337 goto err;
338 }
341
344 goto err;
345 }
348 } else {
351 goto err;
352 }
355 }
356
357 err:
361 }
363 }
364
366 const char *output_name, int *output_width, int *output_height)
367 {
375 .output_names = &output_name,
376 .nb_output = 1,
379 };
382 goto err;
383 }
384
388 goto err;
389 }
390
392 if (!request) {
395 goto err;
396 }
397
401
402 err:
406 }
407
409 {
411 if (!request) {
413 }
416 return request;
417 }
418
420 {
424 const char *device_name =
ctx->device ?
ctx->device :
"cpu";
425
427 if (!th_model)
429 model = &th_model->
model;
431
432 c10::Device device = c10::Device(device_name);
433 if (device.is_xpu()) {
434 if (!at::hasXPU()) {
437 }
438 at::detail::getXPUHooks().initXPU();
439 } else if (!device.is_cpu()) {
442 }
443
444 try {
445 th_model->
jit_model =
new torch::jit::Module;
446 (*th_model->
jit_model) = torch::jit::load(
ctx->model_filename);
448 } catch (const c10::Error& e) {
451 }
452
456 }
457
459 if (!item) {
461 }
467 }
471
474 }
476
480 }
481
485 }
486
491 return model;
492
494 if (item) {
497 }
500 }
501
503 {
509
514 }
515
517 if (!task) {
520 }
521
527 }
528
534 }
535
540 }
541
543 if (!request) {
546 }
547
549 }
550
552 {
555 }
556
558 {
561
563 // no pending task need to flush
564 return 0;
565
567 if (!request) {
570 }
571
573 }
574
583 };
LastLevelTaskItem * lltask
THInferRequest * infer_request
Filter the word "frame" indicates either a video frame or a group of audio as stored in an AVFrame structure Format for each input and each output the list of supported formats For video that means pixel format For audio that means channel sample they are references to shared objects When the negotiation mechanism computes the intersection of the formats supported at each end of a all references to both lists are replaced with a reference to the intersection And when a single format is eventually chosen for a link amongst the remaining all references to the list are updated That means that if a filter requires that its input and output have the same format amongst a supported all it has to do is use a reference to the same list of formats query_formats can leave some formats unset and return AVERROR(EAGAIN) to cause the negotiation mechanism toagain later. That can be used by filters with complex requirements to use the format negotiated on one link to set the formats supported on another. Frame references ownership and permissions
void * ff_safe_queue_pop_front(SafeQueue *sq)
Remove and free first element from the queue in SafeQueue.
static void deleter(void *arg)
Common Async Execution Mechanism for the DNN Backends.
filter_frame For filters that do not use the this method is called when a frame is pushed to the filter s input It can be called at any time except in a reentrant way If the input frame is enough to produce output
void * ff_queue_pop_front(Queue *q)
Remove and free first element from the Queue.
int ff_check_exec_params(void *ctx, DNNBackendType backend, DNNFunctionType func_type, DNNExecBaseParams *exec_params)
size_t ff_queue_size(Queue *q)
Return the length of the Queue.
#define DNN_GENERIC_ERROR
void av_frame_free(AVFrame **frame)
Free the frame and any dynamically allocated objects in it, e.g.
const DNNModule ff_dnn_backend_torch
This structure describes decoded (raw) audio or video data.
Double-ended queue with mutex locks ensuring data consistency while multithreading.
static int dnn_execute_model_th(const DNNModel *model, DNNExecBaseParams *exec_params)
FramePrePostProc frame_pre_proc
void(* callback)(void *args)
Completion Callback for the backend.
AVFilterContext * filter_ctx
Queue * ff_queue_create(void)
Create a Queue instance.
static int dnn_get_width_idx_by_layout(DNNLayout layout)
static FilteringContext * filter_ctx
Linear double-ended data structure.
int ff_queue_push_back(Queue *q, void *v)
Add data to the tail of the queue.
torch::jit::Module * jit_model
#define AV_LOG_ERROR
Something went wrong and cannot losslessly be recovered.
static void destroy_request_item(THRequestItem **arg)
static THInferRequest * th_create_inference_request(void)
void ff_queue_destroy(Queue *q)
Destroy the Queue instance.
int ff_dnn_fill_gettingoutput_task(TaskItem *task, DNNExecBaseParams *exec_params, void *backend_model, int input_height, int input_width, void *ctx)
Allocate input and output frames and fill the Task with execution parameters.
int(* get_output)(struct DNNModel *model, const char *input_name, int input_width, int input_height, const char *output_name, int *output_width, int *output_height)
static DNNModel * dnn_load_model_th(DnnContext *ctx, DNNFunctionType func_type, AVFilterContext *filter_ctx)
size_t ff_safe_queue_size(SafeQueue *sq)
Return the length of the SafeQueue.
int ff_proc_from_frame_to_dnn(AVFrame *frame, DNNData *input, void *log_ctx)
DNNAsyncExecModule exec_module
static const int sizes[][2]
static int get_input_th(DNNModel *model, DNNData *input, const char *input_name)
SafeQueue * ff_safe_queue_create(void)
Create and initialize a SafeQueue instance.
FramePrePostProc frame_post_proc
static int get_output_th(DNNModel *model, const char *input_name, int input_width, int input_height, const char *output_name, int *output_width, int *output_height)
int ff_dnn_async_module_cleanup(DNNAsyncExecModule *async_module)
Join the Async Execution thread and set module pointers to NULL.
static void infer_completion_callback(void *args)
static int extract_lltask_from_task(TaskItem *task, Queue *lltask_queue)
these buffered frames must be flushed immediately if a new input produces new the filter must not call request_frame to get more It must just process the frame or queue it The task of requesting more frames is left to the filter s request_frame method or the application If a filter has several inputs
DNNFunctionType func_type
void avpriv_report_missing_feature(void *avc, const char *msg,...) av_printf_format(2
Log a generic warning message about a missing feature.
void ff_safe_queue_destroy(SafeQueue *sq)
Destroy the SafeQueue instance.
static DNNAsyncStatusType dnn_get_result_th(const DNNModel *model, AVFrame **in, AVFrame **out)
int ff_dnn_fill_task(TaskItem *task, DNNExecBaseParams *exec_params, void *backend_model, int async, int do_ioproc)
Fill the Task for Backend Execution.
and forward the test the status of outputs and forward it to the corresponding return FFERROR_NOT_READY If the filters stores internally one or a few frame for some input
#define DNN_DEFINE_CLASS(fname)
int ff_safe_queue_push_back(SafeQueue *sq, void *v)
Add data to the tail of queue in the SafeQueue after locking mutex.
static int th_start_inference(void *args)
torch::Tensor * input_tensor
int(* start_inference)(void *request)
Synchronous inference function for the backend with corresponding request item as the argument.
void * args
Argument for the execution functions.
void * av_mallocz(size_t size)
Allocate a memory block with alignment suitable for all memory accesses (including vectors if availab...
static const AVFilterPad outputs[]
static const AVOption dnn_th_options[]
static int execute_model_th(THRequestItem *request, Queue *lltask_queue)
@ AV_OPT_TYPE_INT
Underlying C type is int.
DNNAsyncStatusType ff_dnn_get_result_common(Queue *task_queue, AVFrame **in, AVFrame **out)
Extract input and output frame from the Task Queue after asynchronous inference.
void * ff_queue_peek_front(Queue *q)
Return a pointer to the data at the head of the queue.
static int dnn_get_height_idx_by_layout(DNNLayout layout)
static int dnn_flush_th(const DNNModel *model)
static int dnn_get_channel_idx_by_layout(DNNLayout layout)
int(* get_input)(struct DNNModel *model, DNNData *input, const char *input_name)
static void dnn_free_model_th(DNNModel **model)
static int fill_model_input_th(THModel *th_model, THRequestItem *request)
SafeQueue * request_queue
int ff_proc_from_dnn_to_frame(AVFrame *frame, DNNData *output, void *log_ctx)
static void th_free_request(THInferRequest *request)
Generated on Fri Aug 22 2025 13:59:09 for FFmpeg by
doxygen
1.8.17