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 ie_blob_free(&input_blob);
145 av_log(ctx, AV_LOG_ERROR, "Failed to get input blob dims/precision\n");
149 status = ie_blob_get_buffer(input_blob, &blob_buffer);
151 ie_blob_free(&input_blob);
152 av_log(ctx, AV_LOG_ERROR, "Failed to get input blob buffer\n");
156 input.height = dims.dims[2];
157 input.width = dims.dims[3];
158 input.channels = dims.dims[1];
159 input.data = blob_buffer.buffer;
160 input.dt = precision_to_datatype(precision);
161 // all models in openvino open model zoo use BGR as input,
162 // change to be an option when necessary.
163 input.order = DCO_BGR;
165 av_assert0(request->task_count <= dims.dims[0]);
166 for (int i = 0; i < request->task_count; ++i) {
167 task = request->tasks[i];
168 if (task->do_ioproc) {
169 if (ov_model->model->pre_proc != NULL) {
170 ov_model->model->pre_proc(task->in_frame, &input, ov_model->model->filter_ctx);
172 ff_proc_from_frame_to_dnn(task->in_frame, &input, ov_model->model->func_type, ctx);
175 input.data = (uint8_t *)input.data
176 + input.width * input.height * input.channels * get_datatype_size(input.dt);
178 ie_blob_free(&input_blob);
183 static void infer_completion_callback(void *args)
186 precision_e precision;
188 RequestItem *request = args;
189 TaskItem *task = request->tasks[0];
190 SafeQueue *requestq = task->ov_model->request_queue;
191 ie_blob_t *output_blob = NULL;
192 ie_blob_buffer_t blob_buffer;
194 OVContext *ctx = &task->ov_model->ctx;
196 status = ie_infer_request_get_blob(request->infer_request, task->output_name, &output_blob);
198 //incorrect output name
199 char *model_output_name = NULL;
200 char *all_output_names = NULL;
201 size_t model_output_count = 0;
202 av_log(ctx, AV_LOG_ERROR, "Failed to get model output data\n");
203 status = ie_network_get_outputs_number(task->ov_model->network, &model_output_count);
204 for (size_t i = 0; i < model_output_count; i++) {
205 status = ie_network_get_output_name(task->ov_model->network, i, &model_output_name);
206 APPEND_STRING(all_output_names, model_output_name)
208 av_log(ctx, AV_LOG_ERROR,
209 "output \"%s\" may not correct, all output(s) are: \"%s\"\n",
210 task->output_name, all_output_names);
214 status = ie_blob_get_buffer(output_blob, &blob_buffer);
216 ie_blob_free(&output_blob);
217 av_log(ctx, AV_LOG_ERROR, "Failed to access output memory\n");
221 status |= ie_blob_get_dims(output_blob, &dims);
222 status |= ie_blob_get_precision(output_blob, &precision);
224 ie_blob_free(&output_blob);
225 av_log(ctx, AV_LOG_ERROR, "Failed to get dims or precision of output\n");
229 output.channels = dims.dims[1];
230 output.height = dims.dims[2];
231 output.width = dims.dims[3];
232 output.dt = precision_to_datatype(precision);
233 output.data = blob_buffer.buffer;
235 av_assert0(request->task_count <= dims.dims[0]);
236 av_assert0(request->task_count >= 1);
237 for (int i = 0; i < request->task_count; ++i) {
238 task = request->tasks[i];
239 if (task->do_ioproc) {
240 if (task->ov_model->model->post_proc != NULL) {
241 task->ov_model->model->post_proc(task->out_frame, &output, task->ov_model->model->filter_ctx);
243 ff_proc_from_dnn_to_frame(task->out_frame, &output, ctx);
246 task->out_frame->width = output.width;
247 task->out_frame->height = output.height;
250 output.data = (uint8_t *)output.data
251 + output.width * output.height * output.channels * get_datatype_size(output.dt);
253 ie_blob_free(&output_blob);
255 request->task_count = 0;
258 if (ff_safe_queue_push_back(requestq, request) < 0) {
259 av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
265 static DNNReturnType init_model_ov(OVModel *ov_model, const char *input_name, const char *output_name)
267 OVContext *ctx = &ov_model->ctx;
269 ie_available_devices_t a_dev;
270 ie_config_t config = {NULL, NULL, NULL};
271 char *all_dev_names = NULL;
274 if (ctx->options.batch_size <= 0) {
275 ctx->options.batch_size = 1;
278 if (ctx->options.batch_size > 1) {
279 input_shapes_t input_shapes;
280 status = ie_network_get_input_shapes(ov_model->network, &input_shapes);
283 for (int i = 0; i < input_shapes.shape_num; i++)
284 input_shapes.shapes[i].shape.dims[0] = ctx->options.batch_size;
285 status = ie_network_reshape(ov_model->network, input_shapes);
286 ie_network_input_shapes_free(&input_shapes);
291 // The order of dims in the openvino is fixed and it is always NCHW for 4-D data.
292 // while we pass NHWC data from FFmpeg to openvino
293 status = ie_network_set_input_layout(ov_model->network, input_name, NHWC);
295 av_log(ctx, AV_LOG_ERROR, "Failed to set layout as NHWC for input %s\n", input_name);
298 status = ie_network_set_output_layout(ov_model->network, output_name, NHWC);
300 av_log(ctx, AV_LOG_ERROR, "Failed to set layout as NHWC for output %s\n", output_name);
304 // all models in openvino open model zoo use BGR with range [0.0f, 255.0f] as input,
305 // we don't have a AVPixelFormat to descibe it, so we'll use AV_PIX_FMT_BGR24 and
306 // ask openvino to do the conversion internally.
307 // the current supported SR model (frame processing) is generated from tensorflow model,
308 // and its input is Y channel as float with range [0.0f, 1.0f], so do not set for this case.
309 // TODO: we need to get a final clear&general solution with all backends/formats considered.
310 if (ov_model->model->func_type != DFT_PROCESS_FRAME) {
311 status = ie_network_set_input_precision(ov_model->network, input_name, U8);
313 av_log(ctx, AV_LOG_ERROR, "Failed to set input precision as U8 for %s\n", input_name);
318 status = ie_core_load_network(ov_model->core, ov_model->network, ctx->options.device_type, &config, &ov_model->exe_network);
320 av_log(ctx, AV_LOG_ERROR, "Failed to load OpenVINO model network\n");
321 status = ie_core_get_available_devices(ov_model->core, &a_dev);
323 av_log(ctx, AV_LOG_ERROR, "Failed to get available devices\n");
326 for (int i = 0; i < a_dev.num_devices; i++) {
327 APPEND_STRING(all_dev_names, a_dev.devices[i])
329 av_log(ctx, AV_LOG_ERROR,"device %s may not be supported, all available devices are: \"%s\"\n",
330 ctx->options.device_type, all_dev_names);
334 // create infer_request for sync execution
335 status = ie_exec_network_create_infer_request(ov_model->exe_network, &ov_model->infer_request);
339 // create infer_requests for async execution
340 if (ctx->options.nireq <= 0) {
341 // the default value is a rough estimation
342 ctx->options.nireq = av_cpu_count() / 2 + 1;
345 ov_model->request_queue = ff_safe_queue_create();
346 if (!ov_model->request_queue) {
350 for (int i = 0; i < ctx->options.nireq; i++) {
351 RequestItem *item = av_mallocz(sizeof(*item));
356 status = ie_exec_network_create_infer_request(ov_model->exe_network, &item->infer_request);
362 item->tasks = av_malloc_array(ctx->options.batch_size, sizeof(*item->tasks));
367 item->task_count = 0;
369 item->callback.completeCallBackFunc = infer_completion_callback;
370 item->callback.args = item;
371 if (ff_safe_queue_push_back(ov_model->request_queue, item) < 0) {
377 ov_model->task_queue = ff_queue_create();
378 if (!ov_model->task_queue) {
385 ff_dnn_free_model_ov(&ov_model->model);
389 static DNNReturnType execute_model_ov(RequestItem *request)
393 TaskItem *task = request->tasks[0];
394 OVContext *ctx = &task->ov_model->ctx;
397 if (request->task_count < ctx->options.batch_size) {
398 if (ff_safe_queue_push_front(task->ov_model->request_queue, request) < 0) {
399 av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
404 ret = fill_model_input_ov(task->ov_model, request);
405 if (ret != DNN_SUCCESS) {
408 status = ie_infer_set_completion_callback(request->infer_request, &request->callback);
410 av_log(ctx, AV_LOG_ERROR, "Failed to set completion callback for inference\n");
413 status = ie_infer_request_infer_async(request->infer_request);
415 av_log(ctx, AV_LOG_ERROR, "Failed to start async inference\n");
420 ret = fill_model_input_ov(task->ov_model, request);
421 if (ret != DNN_SUCCESS) {
424 status = ie_infer_request_infer(request->infer_request);
426 av_log(ctx, AV_LOG_ERROR, "Failed to start synchronous model inference\n");
429 infer_completion_callback(request);
430 return task->done ? DNN_SUCCESS : DNN_ERROR;
434 static DNNReturnType get_input_ov(void *model, DNNData *input, const char *input_name)
436 OVModel *ov_model = model;
437 OVContext *ctx = &ov_model->ctx;
438 char *model_input_name = NULL;
439 char *all_input_names = NULL;
441 size_t model_input_count = 0;
443 precision_e precision;
444 int input_resizable = ctx->options.input_resizable;
446 status = ie_network_get_inputs_number(ov_model->network, &model_input_count);
448 av_log(ctx, AV_LOG_ERROR, "Failed to get input count\n");
452 for (size_t i = 0; i < model_input_count; i++) {
453 status = ie_network_get_input_name(ov_model->network, i, &model_input_name);
455 av_log(ctx, AV_LOG_ERROR, "Failed to get No.%d input's name\n", (int)i);
458 if (strcmp(model_input_name, input_name) == 0) {
459 ie_network_name_free(&model_input_name);
460 status |= ie_network_get_input_dims(ov_model->network, input_name, &dims);
461 status |= ie_network_get_input_precision(ov_model->network, input_name, &precision);
463 av_log(ctx, AV_LOG_ERROR, "Failed to get No.%d input's dims or precision\n", (int)i);
467 input->channels = dims.dims[1];
468 input->height = input_resizable ? -1 : dims.dims[2];
469 input->width = input_resizable ? -1 : dims.dims[3];
470 input->dt = precision_to_datatype(precision);
473 //incorrect input name
474 APPEND_STRING(all_input_names, model_input_name)
477 ie_network_name_free(&model_input_name);
480 av_log(ctx, AV_LOG_ERROR, "Could not find \"%s\" in model, all input(s) are: \"%s\"\n", input_name, all_input_names);
484 static DNNReturnType get_output_ov(void *model, const char *input_name, int input_width, int input_height,
485 const char *output_name, int *output_width, int *output_height)
488 OVModel *ov_model = model;
489 OVContext *ctx = &ov_model->ctx;
492 AVFrame *in_frame = NULL;
493 AVFrame *out_frame = NULL;
494 TaskItem *ptask = &task;
496 input_shapes_t input_shapes;
498 if (ctx->options.input_resizable) {
499 status = ie_network_get_input_shapes(ov_model->network, &input_shapes);
500 input_shapes.shapes->shape.dims[2] = input_height;
501 input_shapes.shapes->shape.dims[3] = input_width;
502 status |= ie_network_reshape(ov_model->network, input_shapes);
503 ie_network_input_shapes_free(&input_shapes);
505 av_log(ctx, AV_LOG_ERROR, "Failed to reshape input size for %s\n", input_name);
510 if (!ov_model->exe_network) {
511 if (init_model_ov(ov_model, input_name, output_name) != DNN_SUCCESS) {
512 av_log(ctx, AV_LOG_ERROR, "Failed init OpenVINO exectuable network or inference request\n");
517 in_frame = av_frame_alloc();
519 av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for input frame\n");
522 in_frame->width = input_width;
523 in_frame->height = input_height;
525 out_frame = av_frame_alloc();
527 av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for output frame\n");
528 av_frame_free(&in_frame);
535 task.input_name = input_name;
536 task.in_frame = in_frame;
537 task.output_name = output_name;
538 task.out_frame = out_frame;
539 task.ov_model = ov_model;
541 request.infer_request = ov_model->infer_request;
542 request.task_count = 1;
543 request.tasks = &ptask;
545 ret = execute_model_ov(&request);
546 *output_width = out_frame->width;
547 *output_height = out_frame->height;
549 av_frame_free(&out_frame);
550 av_frame_free(&in_frame);
554 DNNModel *ff_dnn_load_model_ov(const char *model_filename, DNNFunctionType func_type, const char *options, AVFilterContext *filter_ctx)
556 DNNModel *model = NULL;
557 OVModel *ov_model = NULL;
558 OVContext *ctx = NULL;
561 model = av_mallocz(sizeof(DNNModel));
566 ov_model = av_mallocz(sizeof(OVModel));
571 model->model = ov_model;
572 ov_model->model = model;
573 ov_model->ctx.class = &dnn_openvino_class;
574 ctx = &ov_model->ctx;
577 av_opt_set_defaults(ctx);
578 if (av_opt_set_from_string(ctx, options, NULL, "=", "&") < 0) {
579 av_log(ctx, AV_LOG_ERROR, "Failed to parse options \"%s\"\n", options);
583 status = ie_core_create("", &ov_model->core);
587 status = ie_core_read_network(ov_model->core, model_filename, NULL, &ov_model->network);
590 ver = ie_c_api_version();
591 av_log(ctx, AV_LOG_ERROR, "Failed to read the network from model file %s,\n"
592 "Please check if the model version matches the runtime OpenVINO %s\n",
593 model_filename, ver.api_version);
594 ie_version_free(&ver);
598 model->get_input = &get_input_ov;
599 model->get_output = &get_output_ov;
600 model->options = options;
601 model->filter_ctx = filter_ctx;
602 model->func_type = func_type;
607 ff_dnn_free_model_ov(&model);
611 DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
612 const char **output_names, uint32_t nb_output, AVFrame *out_frame)
614 OVModel *ov_model = model->model;
615 OVContext *ctx = &ov_model->ctx;
618 TaskItem *ptask = &task;
621 av_log(ctx, AV_LOG_ERROR, "in frame is NULL when execute model.\n");
625 if (!out_frame && model->func_type == DFT_PROCESS_FRAME) {
626 av_log(ctx, AV_LOG_ERROR, "out frame is NULL when execute model.\n");
630 if (nb_output != 1) {
631 // currently, the filter does not need multiple outputs,
632 // so we just pending the support until we really need it.
633 avpriv_report_missing_feature(ctx, "multiple outputs");
637 if (ctx->options.batch_size > 1) {
638 avpriv_report_missing_feature(ctx, "batch mode for sync execution");
642 if (!ov_model->exe_network) {
643 if (init_model_ov(ov_model, input_name, output_names[0]) != DNN_SUCCESS) {
644 av_log(ctx, AV_LOG_ERROR, "Failed init OpenVINO exectuable network or inference request\n");
652 task.input_name = input_name;
653 task.in_frame = in_frame;
654 task.output_name = output_names[0];
655 task.out_frame = out_frame;
656 task.ov_model = ov_model;
658 request.infer_request = ov_model->infer_request;
659 request.task_count = 1;
660 request.tasks = &ptask;
662 return execute_model_ov(&request);
665 DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
666 const char **output_names, uint32_t nb_output, AVFrame *out_frame)
668 OVModel *ov_model = model->model;
669 OVContext *ctx = &ov_model->ctx;
670 RequestItem *request;
674 av_log(ctx, AV_LOG_ERROR, "in frame is NULL when async execute model.\n");
678 if (!out_frame && model->func_type == DFT_PROCESS_FRAME) {
679 av_log(ctx, AV_LOG_ERROR, "out frame is NULL when async execute model.\n");
683 task = av_malloc(sizeof(*task));
685 av_log(ctx, AV_LOG_ERROR, "unable to alloc memory for task item.\n");
689 if (!ov_model->exe_network) {
690 if (init_model_ov(ov_model, input_name, output_names[0]) != DNN_SUCCESS) {
691 av_log(ctx, AV_LOG_ERROR, "Failed init OpenVINO exectuable network or inference request\n");
699 task->input_name = input_name;
700 task->in_frame = in_frame;
701 task->output_name = output_names[0];
702 task->out_frame = out_frame;
703 task->ov_model = ov_model;
704 if (ff_queue_push_back(ov_model->task_queue, task) < 0) {
706 av_log(ctx, AV_LOG_ERROR, "unable to push back task_queue.\n");
710 request = ff_safe_queue_pop_front(ov_model->request_queue);
712 av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
716 request->tasks[request->task_count++] = task;
717 return execute_model_ov(request);
720 DNNAsyncStatusType ff_dnn_get_async_result_ov(const DNNModel *model, AVFrame **in, AVFrame **out)
722 OVModel *ov_model = model->model;
723 TaskItem *task = ff_queue_peek_front(ov_model->task_queue);
726 return DAST_EMPTY_QUEUE;
730 return DAST_NOT_READY;
733 *in = task->in_frame;
734 *out = task->out_frame;
735 ff_queue_pop_front(ov_model->task_queue);
741 DNNReturnType ff_dnn_flush_ov(const DNNModel *model)
743 OVModel *ov_model = model->model;
744 OVContext *ctx = &ov_model->ctx;
745 RequestItem *request;
749 request = ff_safe_queue_pop_front(ov_model->request_queue);
751 av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
755 if (request->task_count == 0) {
756 // no pending task need to flush
757 if (ff_safe_queue_push_back(ov_model->request_queue, request) < 0) {
758 av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
764 ret = fill_model_input_ov(ov_model, request);
765 if (ret != DNN_SUCCESS) {
766 av_log(ctx, AV_LOG_ERROR, "Failed to fill model input.\n");
769 status = ie_infer_set_completion_callback(request->infer_request, &request->callback);
771 av_log(ctx, AV_LOG_ERROR, "Failed to set completion callback for inference\n");
774 status = ie_infer_request_infer_async(request->infer_request);
776 av_log(ctx, AV_LOG_ERROR, "Failed to start async inference\n");
783 void ff_dnn_free_model_ov(DNNModel **model)
786 OVModel *ov_model = (*model)->model;
787 while (ff_safe_queue_size(ov_model->request_queue) != 0) {
788 RequestItem *item = ff_safe_queue_pop_front(ov_model->request_queue);
789 if (item && item->infer_request) {
790 ie_infer_request_free(&item->infer_request);
792 av_freep(&item->tasks);
795 ff_safe_queue_destroy(ov_model->request_queue);
797 while (ff_queue_size(ov_model->task_queue) != 0) {
798 TaskItem *item = ff_queue_pop_front(ov_model->task_queue);
799 av_frame_free(&item->in_frame);
800 av_frame_free(&item->out_frame);
803 ff_queue_destroy(ov_model->task_queue);
805 if (ov_model->infer_request)
806 ie_infer_request_free(&ov_model->infer_request);
807 if (ov_model->exe_network)
808 ie_exec_network_free(&ov_model->exe_network);
809 if (ov_model->network)
810 ie_network_free(&ov_model->network);
812 ie_core_free(&ov_model->core);