4 * This file is part of FFmpeg.
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.
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.
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
23 * DNN OpenVINO backend implementation.
26 #include "dnn_backend_openvino.h"
27 #include "dnn_io_proc.h"
28 #include "libavformat/avio.h"
29 #include "libavutil/avassert.h"
30 #include "libavutil/opt.h"
31 #include "libavutil/avstring.h"
32 #include "../internal.h"
34 #include "safe_queue.h"
35 #include <c_api/ie_c_api.h>
37 typedef struct OVOptions{
44 typedef struct OVContext {
49 typedef struct OVModel{
53 ie_network_t *network;
54 ie_executable_network_t *exe_network;
55 ie_infer_request_t *infer_request;
57 /* for async execution */
58 SafeQueue *request_queue; // holds RequestItem
59 Queue *task_queue; // holds TaskItem
62 typedef struct TaskItem {
64 const char *input_name;
66 const char *output_name;
73 typedef struct RequestItem {
74 ie_infer_request_t *infer_request;
77 ie_complete_call_back_t callback;
80 #define APPEND_STRING(generated_string, iterate_string) \
81 generated_string = generated_string ? av_asprintf("%s %s", generated_string, iterate_string) : \
82 av_asprintf("%s", iterate_string);
84 #define OFFSET(x) offsetof(OVContext, x)
85 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM
86 static const AVOption dnn_openvino_options[] = {
87 { "device", "device to run model", OFFSET(options.device_type), AV_OPT_TYPE_STRING, { .str = "CPU" }, 0, 0, FLAGS },
88 { "nireq", "number of request", OFFSET(options.nireq), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, FLAGS },
89 { "batch_size", "batch size per request", OFFSET(options.batch_size), AV_OPT_TYPE_INT, { .i64 = 1 }, 1, 1000, FLAGS},
90 { "input_resizable", "can input be resizable or not", OFFSET(options.input_resizable), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS },
94 AVFILTER_DEFINE_CLASS(dnn_openvino);
96 static DNNDataType precision_to_datatype(precision_e precision)
105 av_assert0(!"not supported yet.");
110 static int get_datatype_size(DNNDataType dt)
115 return sizeof(float);
117 return sizeof(uint8_t);
119 av_assert0(!"not supported yet.");
124 static DNNReturnType fill_model_input_ov(OVModel *ov_model, RequestItem *request)
127 precision_e precision;
128 ie_blob_buffer_t blob_buffer;
129 OVContext *ctx = &ov_model->ctx;
132 ie_blob_t *input_blob = NULL;
133 TaskItem *task = request->tasks[0];
135 status = ie_infer_request_get_blob(request->infer_request, task->input_name, &input_blob);
137 av_log(ctx, AV_LOG_ERROR, "Failed to get input blob with name %s\n", task->input_name);
141 status |= ie_blob_get_dims(input_blob, &dims);
142 status |= ie_blob_get_precision(input_blob, &precision);
144 av_log(ctx, AV_LOG_ERROR, "Failed to get input blob dims/precision\n");
148 status = ie_blob_get_buffer(input_blob, &blob_buffer);
150 av_log(ctx, AV_LOG_ERROR, "Failed to get input blob buffer\n");
154 input.height = dims.dims[2];
155 input.width = dims.dims[3];
156 input.channels = dims.dims[1];
157 input.data = blob_buffer.buffer;
158 input.dt = precision_to_datatype(precision);
159 // all models in openvino open model zoo use BGR as input,
160 // change to be an option when necessary.
161 input.order = DCO_BGR;
163 av_assert0(request->task_count <= dims.dims[0]);
164 for (int i = 0; i < request->task_count; ++i) {
165 task = request->tasks[i];
166 if (task->do_ioproc) {
167 if (ov_model->model->pre_proc != NULL) {
168 ov_model->model->pre_proc(task->in_frame, &input, ov_model->model->filter_ctx);
170 ff_proc_from_frame_to_dnn(task->in_frame, &input, ov_model->model->func_type, ctx);
173 input.data = (uint8_t *)input.data
174 + input.width * input.height * input.channels * get_datatype_size(input.dt);
176 ie_blob_free(&input_blob);
181 static void infer_completion_callback(void *args)
184 precision_e precision;
186 RequestItem *request = args;
187 TaskItem *task = request->tasks[0];
188 SafeQueue *requestq = task->ov_model->request_queue;
189 ie_blob_t *output_blob = NULL;
190 ie_blob_buffer_t blob_buffer;
192 OVContext *ctx = &task->ov_model->ctx;
194 status = ie_infer_request_get_blob(request->infer_request, task->output_name, &output_blob);
196 //incorrect output name
197 char *model_output_name = NULL;
198 char *all_output_names = NULL;
199 size_t model_output_count = 0;
200 av_log(ctx, AV_LOG_ERROR, "Failed to get model output data\n");
201 status = ie_network_get_outputs_number(task->ov_model->network, &model_output_count);
202 for (size_t i = 0; i < model_output_count; i++) {
203 status = ie_network_get_output_name(task->ov_model->network, i, &model_output_name);
204 APPEND_STRING(all_output_names, model_output_name)
206 av_log(ctx, AV_LOG_ERROR,
207 "output \"%s\" may not correct, all output(s) are: \"%s\"\n",
208 task->output_name, all_output_names);
212 status = ie_blob_get_buffer(output_blob, &blob_buffer);
214 av_log(ctx, AV_LOG_ERROR, "Failed to access output memory\n");
218 status |= ie_blob_get_dims(output_blob, &dims);
219 status |= ie_blob_get_precision(output_blob, &precision);
221 av_log(ctx, AV_LOG_ERROR, "Failed to get dims or precision of output\n");
225 output.channels = dims.dims[1];
226 output.height = dims.dims[2];
227 output.width = dims.dims[3];
228 output.dt = precision_to_datatype(precision);
229 output.data = blob_buffer.buffer;
231 av_assert0(request->task_count <= dims.dims[0]);
232 av_assert0(request->task_count >= 1);
233 for (int i = 0; i < request->task_count; ++i) {
234 task = request->tasks[i];
235 if (task->do_ioproc) {
236 if (task->ov_model->model->post_proc != NULL) {
237 task->ov_model->model->post_proc(task->out_frame, &output, task->ov_model->model->filter_ctx);
239 ff_proc_from_dnn_to_frame(task->out_frame, &output, ctx);
242 task->out_frame->width = output.width;
243 task->out_frame->height = output.height;
246 output.data = (uint8_t *)output.data
247 + output.width * output.height * output.channels * get_datatype_size(output.dt);
249 ie_blob_free(&output_blob);
251 request->task_count = 0;
254 if (ff_safe_queue_push_back(requestq, request) < 0) {
255 av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
261 static DNNReturnType init_model_ov(OVModel *ov_model, const char *input_name, const char *output_name)
263 OVContext *ctx = &ov_model->ctx;
265 ie_available_devices_t a_dev;
266 ie_config_t config = {NULL, NULL, NULL};
267 char *all_dev_names = NULL;
270 if (ctx->options.batch_size <= 0) {
271 ctx->options.batch_size = 1;
274 if (ctx->options.batch_size > 1) {
275 input_shapes_t input_shapes;
276 status = ie_network_get_input_shapes(ov_model->network, &input_shapes);
279 for (int i = 0; i < input_shapes.shape_num; i++)
280 input_shapes.shapes[i].shape.dims[0] = ctx->options.batch_size;
281 status = ie_network_reshape(ov_model->network, input_shapes);
282 ie_network_input_shapes_free(&input_shapes);
287 // The order of dims in the openvino is fixed and it is always NCHW for 4-D data.
288 // while we pass NHWC data from FFmpeg to openvino
289 status = ie_network_set_input_layout(ov_model->network, input_name, NHWC);
291 av_log(ctx, AV_LOG_ERROR, "Failed to set layout as NHWC for input %s\n", input_name);
294 status = ie_network_set_output_layout(ov_model->network, output_name, NHWC);
296 av_log(ctx, AV_LOG_ERROR, "Failed to set layout as NHWC for output %s\n", output_name);
300 // all models in openvino open model zoo use BGR with range [0.0f, 255.0f] as input,
301 // we don't have a AVPixelFormat to descibe it, so we'll use AV_PIX_FMT_BGR24 and
302 // ask openvino to do the conversion internally.
303 // the current supported SR model (frame processing) is generated from tensorflow model,
304 // and its input is Y channel as float with range [0.0f, 1.0f], so do not set for this case.
305 // TODO: we need to get a final clear&general solution with all backends/formats considered.
306 if (ov_model->model->func_type != DFT_PROCESS_FRAME) {
307 status = ie_network_set_input_precision(ov_model->network, input_name, U8);
309 av_log(ctx, AV_LOG_ERROR, "Failed to set input precision as U8 for %s\n", input_name);
314 status = ie_core_load_network(ov_model->core, ov_model->network, ctx->options.device_type, &config, &ov_model->exe_network);
316 av_log(ctx, AV_LOG_ERROR, "Failed to load OpenVINO model network\n");
317 status = ie_core_get_available_devices(ov_model->core, &a_dev);
319 av_log(ctx, AV_LOG_ERROR, "Failed to get available devices\n");
322 for (int i = 0; i < a_dev.num_devices; i++) {
323 APPEND_STRING(all_dev_names, a_dev.devices[i])
325 av_log(ctx, AV_LOG_ERROR,"device %s may not be supported, all available devices are: \"%s\"\n",
326 ctx->options.device_type, all_dev_names);
330 // create infer_request for sync execution
331 status = ie_exec_network_create_infer_request(ov_model->exe_network, &ov_model->infer_request);
335 // create infer_requests for async execution
336 if (ctx->options.nireq <= 0) {
337 // the default value is a rough estimation
338 ctx->options.nireq = av_cpu_count() / 2 + 1;
341 ov_model->request_queue = ff_safe_queue_create();
342 if (!ov_model->request_queue) {
346 for (int i = 0; i < ctx->options.nireq; i++) {
347 RequestItem *item = av_mallocz(sizeof(*item));
352 status = ie_exec_network_create_infer_request(ov_model->exe_network, &item->infer_request);
358 item->tasks = av_malloc_array(ctx->options.batch_size, sizeof(*item->tasks));
363 item->task_count = 0;
365 item->callback.completeCallBackFunc = infer_completion_callback;
366 item->callback.args = item;
367 if (ff_safe_queue_push_back(ov_model->request_queue, item) < 0) {
373 ov_model->task_queue = ff_queue_create();
374 if (!ov_model->task_queue) {
381 ff_dnn_free_model_ov(&ov_model->model);
385 static DNNReturnType execute_model_ov(RequestItem *request)
389 TaskItem *task = request->tasks[0];
390 OVContext *ctx = &task->ov_model->ctx;
393 if (request->task_count < ctx->options.batch_size) {
394 if (ff_safe_queue_push_front(task->ov_model->request_queue, request) < 0) {
395 av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
400 ret = fill_model_input_ov(task->ov_model, request);
401 if (ret != DNN_SUCCESS) {
404 status = ie_infer_set_completion_callback(request->infer_request, &request->callback);
406 av_log(ctx, AV_LOG_ERROR, "Failed to set completion callback for inference\n");
409 status = ie_infer_request_infer_async(request->infer_request);
411 av_log(ctx, AV_LOG_ERROR, "Failed to start async inference\n");
416 ret = fill_model_input_ov(task->ov_model, request);
417 if (ret != DNN_SUCCESS) {
420 status = ie_infer_request_infer(request->infer_request);
422 av_log(ctx, AV_LOG_ERROR, "Failed to start synchronous model inference\n");
425 infer_completion_callback(request);
426 return task->done ? DNN_SUCCESS : DNN_ERROR;
430 static DNNReturnType get_input_ov(void *model, DNNData *input, const char *input_name)
432 OVModel *ov_model = model;
433 OVContext *ctx = &ov_model->ctx;
434 char *model_input_name = NULL;
435 char *all_input_names = NULL;
437 size_t model_input_count = 0;
439 precision_e precision;
440 int input_resizable = ctx->options.input_resizable;
442 status = ie_network_get_inputs_number(ov_model->network, &model_input_count);
444 av_log(ctx, AV_LOG_ERROR, "Failed to get input count\n");
448 for (size_t i = 0; i < model_input_count; i++) {
449 status = ie_network_get_input_name(ov_model->network, i, &model_input_name);
451 av_log(ctx, AV_LOG_ERROR, "Failed to get No.%d input's name\n", (int)i);
454 if (strcmp(model_input_name, input_name) == 0) {
455 ie_network_name_free(&model_input_name);
456 status |= ie_network_get_input_dims(ov_model->network, input_name, &dims);
457 status |= ie_network_get_input_precision(ov_model->network, input_name, &precision);
459 av_log(ctx, AV_LOG_ERROR, "Failed to get No.%d input's dims or precision\n", (int)i);
463 input->channels = dims.dims[1];
464 input->height = input_resizable ? -1 : dims.dims[2];
465 input->width = input_resizable ? -1 : dims.dims[3];
466 input->dt = precision_to_datatype(precision);
469 //incorrect input name
470 APPEND_STRING(all_input_names, model_input_name)
473 ie_network_name_free(&model_input_name);
476 av_log(ctx, AV_LOG_ERROR, "Could not find \"%s\" in model, all input(s) are: \"%s\"\n", input_name, all_input_names);
480 static DNNReturnType get_output_ov(void *model, const char *input_name, int input_width, int input_height,
481 const char *output_name, int *output_width, int *output_height)
484 OVModel *ov_model = model;
485 OVContext *ctx = &ov_model->ctx;
488 AVFrame *in_frame = av_frame_alloc();
489 AVFrame *out_frame = NULL;
490 TaskItem *ptask = &task;
492 input_shapes_t input_shapes;
495 av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for input frame\n");
498 out_frame = av_frame_alloc();
500 av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for output frame\n");
501 av_frame_free(&in_frame);
504 in_frame->width = input_width;
505 in_frame->height = input_height;
507 if (ctx->options.input_resizable) {
508 status = ie_network_get_input_shapes(ov_model->network, &input_shapes);
509 input_shapes.shapes->shape.dims[2] = input_height;
510 input_shapes.shapes->shape.dims[3] = input_width;
511 status |= ie_network_reshape(ov_model->network, input_shapes);
512 ie_network_input_shapes_free(&input_shapes);
514 av_log(ctx, AV_LOG_ERROR, "Failed to reshape input size for %s\n", input_name);
519 if (!ov_model->exe_network) {
520 if (init_model_ov(ov_model, input_name, output_name) != DNN_SUCCESS) {
521 av_log(ctx, AV_LOG_ERROR, "Failed init OpenVINO exectuable network or inference request\n");
529 task.input_name = input_name;
530 task.in_frame = in_frame;
531 task.output_name = output_name;
532 task.out_frame = out_frame;
533 task.ov_model = ov_model;
535 request.infer_request = ov_model->infer_request;
536 request.task_count = 1;
537 request.tasks = &ptask;
539 ret = execute_model_ov(&request);
540 *output_width = out_frame->width;
541 *output_height = out_frame->height;
543 av_frame_free(&out_frame);
544 av_frame_free(&in_frame);
548 DNNModel *ff_dnn_load_model_ov(const char *model_filename, DNNFunctionType func_type, const char *options, AVFilterContext *filter_ctx)
550 DNNModel *model = NULL;
551 OVModel *ov_model = NULL;
552 OVContext *ctx = NULL;
555 model = av_mallocz(sizeof(DNNModel));
560 ov_model = av_mallocz(sizeof(OVModel));
565 model->model = ov_model;
566 ov_model->model = model;
567 ov_model->ctx.class = &dnn_openvino_class;
568 ctx = &ov_model->ctx;
571 av_opt_set_defaults(ctx);
572 if (av_opt_set_from_string(ctx, options, NULL, "=", "&") < 0) {
573 av_log(ctx, AV_LOG_ERROR, "Failed to parse options \"%s\"\n", options);
577 status = ie_core_create("", &ov_model->core);
581 status = ie_core_read_network(ov_model->core, model_filename, NULL, &ov_model->network);
584 ver = ie_c_api_version();
585 av_log(ctx, AV_LOG_ERROR, "Failed to read the network from model file %s,\n"
586 "Please check if the model version matches the runtime OpenVINO %s\n",
587 model_filename, ver.api_version);
588 ie_version_free(&ver);
592 model->get_input = &get_input_ov;
593 model->get_output = &get_output_ov;
594 model->options = options;
595 model->filter_ctx = filter_ctx;
596 model->func_type = func_type;
601 ff_dnn_free_model_ov(&model);
605 DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
606 const char **output_names, uint32_t nb_output, AVFrame *out_frame)
608 OVModel *ov_model = model->model;
609 OVContext *ctx = &ov_model->ctx;
612 TaskItem *ptask = &task;
615 av_log(ctx, AV_LOG_ERROR, "in frame is NULL when execute model.\n");
620 av_log(ctx, AV_LOG_ERROR, "out frame is NULL when execute model.\n");
624 if (nb_output != 1) {
625 // currently, the filter does not need multiple outputs,
626 // so we just pending the support until we really need it.
627 avpriv_report_missing_feature(ctx, "multiple outputs");
631 if (ctx->options.batch_size > 1) {
632 avpriv_report_missing_feature(ctx, "batch mode for sync execution");
636 if (!ov_model->exe_network) {
637 if (init_model_ov(ov_model, input_name, output_names[0]) != DNN_SUCCESS) {
638 av_log(ctx, AV_LOG_ERROR, "Failed init OpenVINO exectuable network or inference request\n");
646 task.input_name = input_name;
647 task.in_frame = in_frame;
648 task.output_name = output_names[0];
649 task.out_frame = out_frame;
650 task.ov_model = ov_model;
652 request.infer_request = ov_model->infer_request;
653 request.task_count = 1;
654 request.tasks = &ptask;
656 return execute_model_ov(&request);
659 DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
660 const char **output_names, uint32_t nb_output, AVFrame *out_frame)
662 OVModel *ov_model = model->model;
663 OVContext *ctx = &ov_model->ctx;
664 RequestItem *request;
668 av_log(ctx, AV_LOG_ERROR, "in frame is NULL when async execute model.\n");
673 av_log(ctx, AV_LOG_ERROR, "out frame is NULL when async execute model.\n");
677 task = av_malloc(sizeof(*task));
679 av_log(ctx, AV_LOG_ERROR, "unable to alloc memory for task item.\n");
683 if (!ov_model->exe_network) {
684 if (init_model_ov(ov_model, input_name, output_names[0]) != DNN_SUCCESS) {
685 av_log(ctx, AV_LOG_ERROR, "Failed init OpenVINO exectuable network or inference request\n");
693 task->input_name = input_name;
694 task->in_frame = in_frame;
695 task->output_name = output_names[0];
696 task->out_frame = out_frame;
697 task->ov_model = ov_model;
698 if (ff_queue_push_back(ov_model->task_queue, task) < 0) {
700 av_log(ctx, AV_LOG_ERROR, "unable to push back task_queue.\n");
704 request = ff_safe_queue_pop_front(ov_model->request_queue);
706 av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
710 request->tasks[request->task_count++] = task;
711 return execute_model_ov(request);
714 DNNAsyncStatusType ff_dnn_get_async_result_ov(const DNNModel *model, AVFrame **in, AVFrame **out)
716 OVModel *ov_model = model->model;
717 TaskItem *task = ff_queue_peek_front(ov_model->task_queue);
720 return DAST_EMPTY_QUEUE;
724 return DAST_NOT_READY;
727 *in = task->in_frame;
728 *out = task->out_frame;
729 ff_queue_pop_front(ov_model->task_queue);
735 DNNReturnType ff_dnn_flush_ov(const DNNModel *model)
737 OVModel *ov_model = model->model;
738 OVContext *ctx = &ov_model->ctx;
739 RequestItem *request;
743 request = ff_safe_queue_pop_front(ov_model->request_queue);
745 av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
749 if (request->task_count == 0) {
750 // no pending task need to flush
751 if (ff_safe_queue_push_back(ov_model->request_queue, request) < 0) {
752 av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
758 ret = fill_model_input_ov(ov_model, request);
759 if (ret != DNN_SUCCESS) {
760 av_log(ctx, AV_LOG_ERROR, "Failed to fill model input.\n");
763 status = ie_infer_set_completion_callback(request->infer_request, &request->callback);
765 av_log(ctx, AV_LOG_ERROR, "Failed to set completion callback for inference\n");
768 status = ie_infer_request_infer_async(request->infer_request);
770 av_log(ctx, AV_LOG_ERROR, "Failed to start async inference\n");
777 void ff_dnn_free_model_ov(DNNModel **model)
780 OVModel *ov_model = (*model)->model;
781 while (ff_safe_queue_size(ov_model->request_queue) != 0) {
782 RequestItem *item = ff_safe_queue_pop_front(ov_model->request_queue);
783 if (item && item->infer_request) {
784 ie_infer_request_free(&item->infer_request);
786 av_freep(&item->tasks);
789 ff_safe_queue_destroy(ov_model->request_queue);
791 while (ff_queue_size(ov_model->task_queue) != 0) {
792 TaskItem *item = ff_queue_pop_front(ov_model->task_queue);
793 av_frame_free(&item->in_frame);
794 av_frame_free(&item->out_frame);
797 ff_queue_destroy(ov_model->task_queue);
799 if (ov_model->infer_request)
800 ie_infer_request_free(&ov_model->infer_request);
801 if (ov_model->exe_network)
802 ie_exec_network_free(&ov_model->exe_network);
803 if (ov_model->network)
804 ie_network_free(&ov_model->network);
806 ie_core_free(&ov_model->core);