diff mbox series

[FFmpeg-devel,V2] libavfilter/dnn: add batch mode for async execution

Message ID 20210110131601.22685-1-yejun.guo@intel.com
State New
Headers show
Series [FFmpeg-devel,V2] libavfilter/dnn: add batch mode for async execution
Related show

Checks

Context Check Description
andriy/x86_make success Make finished
andriy/x86_make_fate success Make fate finished
andriy/PPC64_make success Make finished
andriy/PPC64_make_fate success Make fate finished

Commit Message

Guo, Yejun Jan. 10, 2021, 1:16 p.m. UTC
the default number of batch_size is 1

Signed-off-by: Xie, Lin <lin.xie@intel.com>
Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
---
 libavfilter/dnn/dnn_backend_openvino.c | 187 ++++++++++++++++++++-----
 libavfilter/dnn/dnn_backend_openvino.h |   1 +
 libavfilter/dnn/dnn_interface.c        |   1 +
 libavfilter/dnn_interface.h            |   2 +
 libavfilter/vf_dnn_processing.c        |  36 ++++-
 5 files changed, 194 insertions(+), 33 deletions(-)

Comments

Fu, Ting Jan. 14, 2021, 2:45 p.m. UTC | #1
> -----Original Message-----
> From: ffmpeg-devel <ffmpeg-devel-bounces@ffmpeg.org> On Behalf Of Guo,
> Yejun
> Sent: Sunday, January 10, 2021 09:16 PM
> To: ffmpeg-devel@ffmpeg.org
> Cc: Guo, Yejun <yejun.guo@intel.com>
> Subject: [FFmpeg-devel] [PATCH V2] libavfilter/dnn: add batch mode for async
> execution
> 
> the default number of batch_size is 1
> 
> Signed-off-by: Xie, Lin <lin.xie@intel.com>
> Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>
> Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
> ---
>  libavfilter/dnn/dnn_backend_openvino.c | 187 ++++++++++++++++++++-----
>  libavfilter/dnn/dnn_backend_openvino.h |   1 +
>  libavfilter/dnn/dnn_interface.c        |   1 +
>  libavfilter/dnn_interface.h            |   2 +
>  libavfilter/vf_dnn_processing.c        |  36 ++++-
>  5 files changed, 194 insertions(+), 33 deletions(-)
> 
[...]
>      if (ff_inlink_acknowledge_status(inlink, &status, &pts)) {
>          if (status == AVERROR_EOF) {
> -            ff_outlink_set_status(outlink, status, pts);
> +            int64_t out_pts = pts;
> +            ret = flush_frame(outlink, pts, &out_pts);
> +            ff_outlink_set_status(outlink, status, out_pts);
>              return ret;
>          }
>      }
> --
> 2.17.1

Hi Yejun,
This patch works well for me.
Testing was carried on my machine, which CPU is i7-8700K 3.7Ghz and iGPU is UHD630.
The patch was tested by using espcn super resolution model (950*540 video as input), with async on and off. The fps increased from 11fps to 13fps (~18% up) on CPU, from 8fps to 11fps (~37% up) on iGPU.

On CPU with async off:
./ffmpeg -i input_video.mp4 -vf dnn_processing=dnn_backend=openvino:model=espcn1080p.xml:input=x:output=espcn/prediction:async=0:options=device=CPU\&batch_size=1 -y output_video.mp4
On CPU with async on:
./ffmpeg -i input_video.mp4 -vf dnn_processing=dnn_backend=openvino:model=espcn1080p.xml:input=x:output=espcn/prediction:async=1:options=device=CPU\&batch_size=2 -y output_video.mp4

On GPU with async off:
./ffmpeg -i input_video.mp4 -vf dnn_processing=dnn_backend=openvino:model=espcn1080p.xml:input=x:output=espcn/prediction:async=0:options=device=GPU\&batch_size=1 -y output_video.mp4
On GPU with async on:
./ffmpeg -i input_video.mp4 -vf dnn_processing=dnn_backend=openvino:model=espcn1080p.xml:input=x:output=espcn/prediction:async=1:options=device=GPU\&batch_size=2 -y output_video.mp4

> 
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email ffmpeg-devel-request@ffmpeg.org
> with subject "unsubscribe".
Guo, Yejun Jan. 15, 2021, 1:04 a.m. UTC | #2
> -----Original Message-----
> From: ffmpeg-devel <ffmpeg-devel-bounces@ffmpeg.org> On Behalf Of Fu,
> Ting
> Sent: 2021年1月14日 22:45
> To: FFmpeg development discussions and patches <ffmpeg-devel@ffmpeg.org>
> Subject: Re: [FFmpeg-devel] [PATCH V2] libavfilter/dnn: add batch mode for
> async execution
> 
> 
> 
> > -----Original Message-----
> > From: ffmpeg-devel <ffmpeg-devel-bounces@ffmpeg.org> On Behalf Of Guo,
> > Yejun
> > Sent: Sunday, January 10, 2021 09:16 PM
> > To: ffmpeg-devel@ffmpeg.org
> > Cc: Guo, Yejun <yejun.guo@intel.com>
> > Subject: [FFmpeg-devel] [PATCH V2] libavfilter/dnn: add batch mode for
> > async execution
> >
> > the default number of batch_size is 1
> >
> > Signed-off-by: Xie, Lin <lin.xie@intel.com>
> > Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>
> > Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
> > ---
> >  libavfilter/dnn/dnn_backend_openvino.c | 187
> ++++++++++++++++++++-----
> >  libavfilter/dnn/dnn_backend_openvino.h |   1 +
> >  libavfilter/dnn/dnn_interface.c        |   1 +
> >  libavfilter/dnn_interface.h            |   2 +
> >  libavfilter/vf_dnn_processing.c        |  36 ++++-
> >  5 files changed, 194 insertions(+), 33 deletions(-)
> >
> > --
> > 2.17.1
> 
> Hi Yejun,
> This patch works well for me.
> Testing was carried on my machine, which CPU is i7-8700K 3.7Ghz and iGPU is
> UHD630.
> The patch was tested by using espcn super resolution model (950*540 video as
> input), with async on and off. The fps increased from 11fps to 13fps (~18% up)
> on CPU, from 8fps to 11fps (~37% up) on iGPU.
> 
> On CPU with async off:
> ./ffmpeg -i input_video.mp4 -vf
> dnn_processing=dnn_backend=openvino:model=espcn1080p.xml:input=x:outp
> ut=espcn/prediction:async=0:options=device=CPU\&batch_size=1 -y
> output_video.mp4 On CPU with async on:
> ./ffmpeg -i input_video.mp4 -vf
> dnn_processing=dnn_backend=openvino:model=espcn1080p.xml:input=x:outp
> ut=espcn/prediction:async=1:options=device=CPU\&batch_size=2 -y
> output_video.mp4
> 
> On GPU with async off:
> ./ffmpeg -i input_video.mp4 -vf
> dnn_processing=dnn_backend=openvino:model=espcn1080p.xml:input=x:outp
> ut=espcn/prediction:async=0:options=device=GPU\&batch_size=1 -y
> output_video.mp4 On GPU with async on:
> ./ffmpeg -i input_video.mp4 -vf
> dnn_processing=dnn_backend=openvino:model=espcn1080p.xml:input=x:outp
> ut=espcn/prediction:async=1:options=device=GPU\&batch_size=2 -y
> output_video.mp4
> 
thanks Ting for the test, will push soon.
Steven Liu Jan. 15, 2021, 3:33 a.m. UTC | #3
> 2021年1月10日 下午9:16,Guo, Yejun <yejun.guo@intel.com> 写道:
> 
> the default number of batch_size is 1
> 
> Signed-off-by: Xie, Lin <lin.xie@intel.com>
> Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>
> Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
> ---
> libavfilter/dnn/dnn_backend_openvino.c | 187 ++++++++++++++++++++-----
> libavfilter/dnn/dnn_backend_openvino.h |   1 +
> libavfilter/dnn/dnn_interface.c        |   1 +
> libavfilter/dnn_interface.h            |   2 +
> libavfilter/vf_dnn_processing.c        |  36 ++++-
> 5 files changed, 194 insertions(+), 33 deletions(-)
> 
> diff --git a/libavfilter/dnn/dnn_backend_openvino.c b/libavfilter/dnn/dnn_backend_openvino.c
> index d27e451eea..5271d1caa5 100644
> --- a/libavfilter/dnn/dnn_backend_openvino.c
> +++ b/libavfilter/dnn/dnn_backend_openvino.c
> @@ -37,6 +37,7 @@
> typedef struct OVOptions{
>     char *device_type;
>     int nireq;
> +    int batch_size;
> } OVOptions;
> 
> typedef struct OVContext {
> @@ -70,7 +71,8 @@ typedef struct TaskItem {
> 
> typedef struct RequestItem {
>     ie_infer_request_t *infer_request;
> -    TaskItem *task;
> +    TaskItem **tasks;
> +    int task_count;
>     ie_complete_call_back_t callback;
> } RequestItem;
> 
> @@ -83,6 +85,7 @@ typedef struct RequestItem {
> static const AVOption dnn_openvino_options[] = {
>     { "device", "device to run model", OFFSET(options.device_type), AV_OPT_TYPE_STRING, { .str = "CPU" }, 0, 0, FLAGS },
>     { "nireq",  "number of request",   OFFSET(options.nireq),       AV_OPT_TYPE_INT,    { .i64 = 0 },     0, INT_MAX, FLAGS },
> +    { "batch_size",  "batch size per request", OFFSET(options.batch_size),  AV_OPT_TYPE_INT,    { .i64 = 1 },     1, 1000, FLAGS},
>     { NULL }
> };
> 
> @@ -100,7 +103,19 @@ static DNNDataType precision_to_datatype(precision_e precision)
>     }
> }
> 
> -static DNNReturnType fill_model_input_ov(OVModel *ov_model, TaskItem *task, RequestItem *request)
> +static int get_datatype_size(DNNDataType dt)
> +{
> +    switch (dt)
> +    {
> +    case DNN_FLOAT:
> +        return sizeof(float);
> +    default:
> +        av_assert0(!"not supported yet.");
> +        return 1;
Why don’t try about this way ? :D
avpriv_request_sample()
AVERROR_PATCHWELCOME;

> +    }
> +}
> +
> +static DNNReturnType fill_model_input_ov(OVModel *ov_model, RequestItem *request)
> {
>     dimensions_t dims;
>     precision_e precision;
> @@ -109,6 +124,7 @@ static DNNReturnType fill_model_input_ov(OVModel *ov_model, TaskItem *task, Requ
>     IEStatusCode status;
>     DNNData input;
>     ie_blob_t *input_blob = NULL;
> +    TaskItem *task = request->tasks[0];
> 
>     status = ie_infer_request_get_blob(request->infer_request, task->input_name, &input_blob);
>     if (status != OK) {
> @@ -134,12 +150,19 @@ static DNNReturnType fill_model_input_ov(OVModel *ov_model, TaskItem *task, Requ
>     input.channels = dims.dims[1];
>     input.data = blob_buffer.buffer;
>     input.dt = precision_to_datatype(precision);
> -    if (task->do_ioproc) {
> -        if (ov_model->model->pre_proc != NULL) {
> -            ov_model->model->pre_proc(task->in_frame, &input, ov_model->model->filter_ctx);
> -        } else {
> -            proc_from_frame_to_dnn(task->in_frame, &input, ctx);
> +
> +    av_assert0(request->task_count <= dims.dims[0]);
> +    for (int i = 0; i < request->task_count; ++i) {
> +        task = request->tasks[i];
> +        if (task->do_ioproc) {
> +            if (ov_model->model->pre_proc != NULL) {
> +                ov_model->model->pre_proc(task->in_frame, &input, ov_model->model->filter_ctx);
> +            } else {
> +                proc_from_frame_to_dnn(task->in_frame, &input, ctx);
> +            }
>         }
> +        input.data = (uint8_t *)input.data
> +                     + input.width * input.height * input.channels * get_datatype_size(input.dt);
>     }
>     ie_blob_free(&input_blob);
> 
> @@ -152,7 +175,7 @@ static void infer_completion_callback(void *args)
>     precision_e precision;
>     IEStatusCode status;
>     RequestItem *request = args;
> -    TaskItem *task = request->task;
> +    TaskItem *task = request->tasks[0];
>     ie_blob_t *output_blob = NULL;
>     ie_blob_buffer_t blob_buffer;
>     DNNData output;
> @@ -194,41 +217,56 @@ static void infer_completion_callback(void *args)
>     output.width    = dims.dims[3];
>     output.dt       = precision_to_datatype(precision);
>     output.data     = blob_buffer.buffer;
> -    if (task->do_ioproc) {
> -        if (task->ov_model->model->post_proc != NULL) {
> -            task->ov_model->model->post_proc(task->out_frame, &output, task->ov_model->model->filter_ctx);
> +
> +    av_assert0(request->task_count <= dims.dims[0]);
> +    av_assert0(request->task_count >= 1);
> +    for (int i = 0; i < request->task_count; ++i) {
> +        task = request->tasks[i];
> +        if (task->do_ioproc) {
> +            if (task->ov_model->model->post_proc != NULL) {
> +                task->ov_model->model->post_proc(task->out_frame, &output, task->ov_model->model->filter_ctx);
> +            } else {
> +                proc_from_dnn_to_frame(task->out_frame, &output, ctx);
> +            }
>         } else {
> -            proc_from_dnn_to_frame(task->out_frame, &output, ctx);
> +            task->out_frame->width = output.width;
> +            task->out_frame->height = output.height;
>         }
> -    } else {
> -        task->out_frame->width = output.width;
> -        task->out_frame->height = output.height;
> +        task->done = 1;
> +        output.data = (uint8_t *)output.data
> +                      + output.width * output.height * output.channels * get_datatype_size(output.dt);
>     }
>     ie_blob_free(&output_blob);
> 
> +    request->task_count = 0;
> +
>     if (task->async) {
> -        request->task = NULL;
>         if (ff_safe_queue_push_back(task->ov_model->request_queue, request) < 0) {
>             av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
>             return;
>         }
>     }
> -
> -    task->done = 1;
> }
> 
> -static DNNReturnType execute_model_ov(TaskItem *task, RequestItem *request)
> +static DNNReturnType execute_model_ov(RequestItem *request)
> {
>     IEStatusCode status;
> +    DNNReturnType ret;
> +    TaskItem *task = request->tasks[0];
>     OVContext *ctx = &task->ov_model->ctx;
> 
> -    DNNReturnType ret = fill_model_input_ov(task->ov_model, task, request);
> -    if (ret != DNN_SUCCESS) {
> -        return ret;
> -    }
> -
>     if (task->async) {
> -        request->task = task;
> +        if (request->task_count < ctx->options.batch_size) {
> +            if (ff_safe_queue_push_front(task->ov_model->request_queue, request) < 0) {
> +                av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
> +                return DNN_ERROR;
> +            }
> +            return DNN_SUCCESS;
> +        }
> +        ret = fill_model_input_ov(task->ov_model, request);
> +        if (ret != DNN_SUCCESS) {
> +            return ret;
> +        }
>         status = ie_infer_set_completion_callback(request->infer_request, &request->callback);
>         if (status != OK) {
>             av_log(ctx, AV_LOG_ERROR, "Failed to set completion callback for inference\n");
> @@ -241,12 +279,15 @@ static DNNReturnType execute_model_ov(TaskItem *task, RequestItem *request)
>         }
>         return DNN_SUCCESS;
>     } else {
> +        ret = fill_model_input_ov(task->ov_model, request);
> +        if (ret != DNN_SUCCESS) {
> +            return ret;
> +        }
>         status = ie_infer_request_infer(request->infer_request);
>         if (status != OK) {
>             av_log(ctx, AV_LOG_ERROR, "Failed to start synchronous model inference\n");
>             return DNN_ERROR;
>         }
> -        request->task = task;
>         infer_completion_callback(request);
>         return task->done ? DNN_SUCCESS : DNN_ERROR;
>     }
> @@ -319,6 +360,7 @@ static DNNReturnType get_output_ov(void *model, const char *input_name, int inpu
>     RequestItem request;
>     AVFrame *in_frame = av_frame_alloc();
>     AVFrame *out_frame = NULL;
> +    TaskItem *ptask = &task;
> 
>     if (!in_frame) {
>         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for input frame\n");
> @@ -343,8 +385,10 @@ static DNNReturnType get_output_ov(void *model, const char *input_name, int inpu
>     task.ov_model = ov_model;
> 
>     request.infer_request = ov_model->infer_request;
> +    request.task_count = 1;
> +    request.tasks = &ptask;
> 
> -    ret = execute_model_ov(&task, &request);
> +    ret = execute_model_ov(&request);
>     *output_width = out_frame->width;
>     *output_height = out_frame->height;
> 
> @@ -393,6 +437,24 @@ DNNModel *ff_dnn_load_model_ov(const char *model_filename, const char *options,
>     if (status != OK)
>         goto err;
> 
> +    // batch size
> +    if (ctx->options.batch_size <= 0) {
> +        ctx->options.batch_size = 1;
> +    }
> +
> +    if (ctx->options.batch_size > 1) {
> +        input_shapes_t input_shapes;
> +        status = ie_network_get_input_shapes(ov_model->network, &input_shapes);
> +        if (status != OK)
> +            goto err;
> +        for (int i = 0; i < input_shapes.shape_num; i++)
> +            input_shapes.shapes[i].shape.dims[0] = ctx->options.batch_size;
> +        status = ie_network_reshape(ov_model->network, input_shapes);
> +        ie_network_input_shapes_free(&input_shapes);
> +        if (status != OK)
> +            goto err;
> +    }
> +
>     status = ie_core_load_network(ov_model->core, ov_model->network, ctx->options.device_type, &config, &ov_model->exe_network);
>     if (status != OK) {
>         av_log(ctx, AV_LOG_ERROR, "Failed to init OpenVINO model\n");
> @@ -426,17 +488,24 @@ DNNModel *ff_dnn_load_model_ov(const char *model_filename, const char *options,
>     }
> 
>     for (int i = 0; i < ctx->options.nireq; i++) {
> -        ie_infer_request_t *request;
>         RequestItem *item = av_mallocz(sizeof(*item));
>         if (!item) {
>             goto err;
>         }
> -        status = ie_exec_network_create_infer_request(ov_model->exe_network, &request);
> +
> +        status = ie_exec_network_create_infer_request(ov_model->exe_network, &item->infer_request);
>         if (status != OK) {
>             av_freep(&item);
>             goto err;
>         }
> -        item->infer_request = request;
> +
> +        item->tasks = av_malloc_array(ctx->options.batch_size, sizeof(*item->tasks));
> +        if (!item->tasks) {
> +            av_freep(&item);
> +            goto err;
> +        }
> +        item->task_count = 0;
> +
>         item->callback.completeCallBackFunc = infer_completion_callback;
>         item->callback.args = item;
>         if (ff_safe_queue_push_back(ov_model->request_queue, item) < 0) {
> @@ -469,6 +538,7 @@ DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
>     OVContext *ctx = &ov_model->ctx;
>     TaskItem task;
>     RequestItem request;
> +    TaskItem *ptask = &task;
> 
>     if (!in_frame) {
>         av_log(ctx, AV_LOG_ERROR, "in frame is NULL when execute model.\n");
> @@ -487,6 +557,11 @@ DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
>         return DNN_ERROR;
>     }
> 
> +    if (ctx->options.batch_size > 1) {
> +        av_log(ctx, AV_LOG_ERROR, "do not support batch mode for sync execution.\n");
> +        return DNN_ERROR;
> +    }
> +
>     task.done = 0;
>     task.do_ioproc = 1;
>     task.async = 0;
> @@ -497,8 +572,10 @@ DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
>     task.ov_model = ov_model;
> 
>     request.infer_request = ov_model->infer_request;
> +    request.task_count = 1;
> +    request.tasks = &ptask;
> 
> -    return execute_model_ov(&task, &request);
> +    return execute_model_ov(&request);
> }
> 
> DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
> @@ -545,7 +622,8 @@ DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, const char *i
>         return DNN_ERROR;
>     }
> 
> -    return execute_model_ov(task, request);
> +    request->tasks[request->task_count++] = task;
> +    return execute_model_ov(request);
> }
> 
> DNNAsyncStatusType ff_dnn_get_async_result_ov(const DNNModel *model, AVFrame **in, AVFrame **out)
> @@ -569,6 +647,48 @@ DNNAsyncStatusType ff_dnn_get_async_result_ov(const DNNModel *model, AVFrame **i
>     return DAST_SUCCESS;
> }
> 
> +DNNReturnType ff_dnn_flush_ov(const DNNModel *model)
> +{
> +    OVModel *ov_model = (OVModel *)model->model;
> +    OVContext *ctx = &ov_model->ctx;
> +    RequestItem *request;
> +    IEStatusCode status;
> +    DNNReturnType ret;
> +
> +    request = ff_safe_queue_pop_front(ov_model->request_queue);
> +    if (!request) {
> +        av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
> +        return DNN_ERROR;
> +    }
> +
> +    if (request->task_count == 0) {
> +        // no pending task need to flush
> +        if (ff_safe_queue_push_back(ov_model->request_queue, request) < 0) {
> +            av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
> +            return DNN_ERROR;
> +        }
> +        return DNN_SUCCESS;
> +    }
> +
> +    ret = fill_model_input_ov(ov_model, request);
> +    if (ret != DNN_SUCCESS) {
> +        av_log(ctx, AV_LOG_ERROR, "Failed to fill model input.\n");
> +        return ret;
> +    }
> +    status = ie_infer_set_completion_callback(request->infer_request, &request->callback);
> +    if (status != OK) {
> +        av_log(ctx, AV_LOG_ERROR, "Failed to set completion callback for inference\n");
> +        return DNN_ERROR;
> +    }
> +    status = ie_infer_request_infer_async(request->infer_request);
> +    if (status != OK) {
> +        av_log(ctx, AV_LOG_ERROR, "Failed to start async inference\n");
> +        return DNN_ERROR;
> +    }
> +
> +    return DNN_SUCCESS;
> +}
> +
> void ff_dnn_free_model_ov(DNNModel **model)
> {
>     if (*model){
> @@ -578,12 +698,15 @@ void ff_dnn_free_model_ov(DNNModel **model)
>             if (item && item->infer_request) {
>                 ie_infer_request_free(&item->infer_request);
>             }
> +            av_freep(&item->tasks);
>             av_freep(&item);
>         }
>         ff_safe_queue_destroy(ov_model->request_queue);
> 
>         while (ff_queue_size(ov_model->task_queue) != 0) {
>             TaskItem *item = ff_queue_pop_front(ov_model->task_queue);
> +            av_frame_free(&item->in_frame);
> +            av_frame_free(&item->out_frame);
>             av_freep(&item);
>         }
>         ff_queue_destroy(ov_model->task_queue);
> diff --git a/libavfilter/dnn/dnn_backend_openvino.h b/libavfilter/dnn/dnn_backend_openvino.h
> index 1b70150040..23b819440e 100644
> --- a/libavfilter/dnn/dnn_backend_openvino.h
> +++ b/libavfilter/dnn/dnn_backend_openvino.h
> @@ -36,6 +36,7 @@ DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
> DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
>                                             const char **output_names, uint32_t nb_output, AVFrame *out_frame);
> DNNAsyncStatusType ff_dnn_get_async_result_ov(const DNNModel *model, AVFrame **in, AVFrame **out);
> +DNNReturnType ff_dnn_flush_ov(const DNNModel *model);
> 
> void ff_dnn_free_model_ov(DNNModel **model);
> 
> diff --git a/libavfilter/dnn/dnn_interface.c b/libavfilter/dnn/dnn_interface.c
> index e1b41a21e1..02e532fc1b 100644
> --- a/libavfilter/dnn/dnn_interface.c
> +++ b/libavfilter/dnn/dnn_interface.c
> @@ -60,6 +60,7 @@ DNNModule *ff_get_dnn_module(DNNBackendType backend_type)
>         dnn_module->execute_model = &ff_dnn_execute_model_ov;
>         dnn_module->execute_model_async = &ff_dnn_execute_model_async_ov;
>         dnn_module->get_async_result = &ff_dnn_get_async_result_ov;
> +        dnn_module->flush = &ff_dnn_flush_ov;
>         dnn_module->free_model = &ff_dnn_free_model_ov;
>     #else
>         av_freep(&dnn_module);
> diff --git a/libavfilter/dnn_interface.h b/libavfilter/dnn_interface.h
> index 9533c88829..ff338ea084 100644
> --- a/libavfilter/dnn_interface.h
> +++ b/libavfilter/dnn_interface.h
> @@ -82,6 +82,8 @@ typedef struct DNNModule{
>                                          const char **output_names, uint32_t nb_output, AVFrame *out_frame);
>     // Retrieve inference result.
>     DNNAsyncStatusType (*get_async_result)(const DNNModel *model, AVFrame **in, AVFrame **out);
> +    // Flush all the pending tasks.
> +    DNNReturnType (*flush)(const DNNModel *model);
>     // Frees memory allocated for model.
>     void (*free_model)(DNNModel **model);
> } DNNModule;
> diff --git a/libavfilter/vf_dnn_processing.c b/libavfilter/vf_dnn_processing.c
> index fff5696a31..be48631782 100644
> --- a/libavfilter/vf_dnn_processing.c
> +++ b/libavfilter/vf_dnn_processing.c
> @@ -33,6 +33,7 @@
> #include "formats.h"
> #include "internal.h"
> #include "libswscale/swscale.h"
> +#include "libavutil/time.h"
> 
> typedef struct DnnProcessingContext {
>     const AVClass *class;
> @@ -369,6 +370,37 @@ static int activate_sync(AVFilterContext *filter_ctx)
>     return FFERROR_NOT_READY;
> }
> 
> +static int flush_frame(AVFilterLink *outlink, int64_t pts, int64_t *out_pts)
> +{
> +    DnnProcessingContext *ctx = outlink->src->priv;
> +    int ret;
> +    DNNAsyncStatusType async_state;
> +
> +    ret = (ctx->dnn_module->flush)(ctx->model);
> +    if (ret != DNN_SUCCESS) {
> +        return -1;
> +    }
> +
> +    do {
> +        AVFrame *in_frame = NULL;
> +        AVFrame *out_frame = NULL;
> +        async_state = (ctx->dnn_module->get_async_result)(ctx->model, &in_frame, &out_frame);
> +        if (out_frame) {
> +            if (isPlanarYUV(in_frame->format))
> +                copy_uv_planes(ctx, out_frame, in_frame);
> +            av_frame_free(&in_frame);
> +            ret = ff_filter_frame(outlink, out_frame);
> +            if (ret < 0)
> +                return ret;
> +            if (out_pts)
> +                *out_pts = out_frame->pts + pts;
> +        }
> +        av_usleep(5000);
> +    } while (async_state >= DAST_NOT_READY);
> +
> +    return 0;
> +}
> +
> static int activate_async(AVFilterContext *filter_ctx)
> {
>     AVFilterLink *inlink = filter_ctx->inputs[0];
> @@ -423,7 +455,9 @@ static int activate_async(AVFilterContext *filter_ctx)
> 
>     if (ff_inlink_acknowledge_status(inlink, &status, &pts)) {
>         if (status == AVERROR_EOF) {
> -            ff_outlink_set_status(outlink, status, pts);
> +            int64_t out_pts = pts;
> +            ret = flush_frame(outlink, pts, &out_pts);
> +            ff_outlink_set_status(outlink, status, out_pts);
>             return ret;
>         }
>     }
> -- 
> 2.17.1
> 
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

Thanks

Steven Liu
Guo, Yejun Jan. 15, 2021, 11:16 a.m. UTC | #4
> -----Original Message-----
> From: Steven Liu <lq@chinaffmpeg.org>
> Sent: 2021年1月15日 11:34
> To: FFmpeg development discussions and patches <ffmpeg-devel@ffmpeg.org>
> Cc: Steven Liu <lq@chinaffmpeg.org>; Guo, Yejun <yejun.guo@intel.com>
> Subject: Re: [FFmpeg-devel] [PATCH V2] libavfilter/dnn: add batch mode for
> async execution
> 
> 
> 
> > 2021年1月10日 下午9:16,Guo, Yejun <yejun.guo@intel.com> 写道:
> >
> > the default number of batch_size is 1
> >
> > Signed-off-by: Xie, Lin <lin.xie@intel.com>
> > Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>
> > Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
> > ---
> > libavfilter/dnn/dnn_backend_openvino.c | 187 ++++++++++++++++++++-----
> > libavfilter/dnn/dnn_backend_openvino.h |   1 +
> > libavfilter/dnn/dnn_interface.c        |   1 +
> > libavfilter/dnn_interface.h            |   2 +
> > libavfilter/vf_dnn_processing.c        |  36 ++++-
> > 5 files changed, 194 insertions(+), 33 deletions(-)
> >
> > diff --git a/libavfilter/dnn/dnn_backend_openvino.c
> > b/libavfilter/dnn/dnn_backend_openvino.c
> > index d27e451eea..5271d1caa5 100644
> > --- a/libavfilter/dnn/dnn_backend_openvino.c
> > +++ b/libavfilter/dnn/dnn_backend_openvino.c
> > @@ -37,6 +37,7 @@
> > typedef struct OVOptions{
> >     char *device_type;
> >     int nireq;
> > +    int batch_size;
> > } OVOptions;
> >
> > typedef struct OVContext {
> > @@ -70,7 +71,8 @@ typedef struct TaskItem {
> >
> > typedef struct RequestItem {
> >     ie_infer_request_t *infer_request;
> > -    TaskItem *task;
> > +    TaskItem **tasks;
> > +    int task_count;
> >     ie_complete_call_back_t callback;
> > } RequestItem;
> >
> > @@ -83,6 +85,7 @@ typedef struct RequestItem { static const AVOption
> > dnn_openvino_options[] = {
> >     { "device", "device to run model", OFFSET(options.device_type),
> AV_OPT_TYPE_STRING, { .str = "CPU" }, 0, 0, FLAGS },
> >     { "nireq",  "number of request",   OFFSET(options.nireq),
> AV_OPT_TYPE_INT,    { .i64 = 0 },     0, INT_MAX, FLAGS },
> > +    { "batch_size",  "batch size per request", OFFSET(options.batch_size),
> AV_OPT_TYPE_INT,    { .i64 = 1 },     1, 1000, FLAGS},
> >     { NULL }
> > };
> >
> > @@ -100,7 +103,19 @@ static DNNDataType
> precision_to_datatype(precision_e precision)
> >     }
> > }
> >
> > -static DNNReturnType fill_model_input_ov(OVModel *ov_model, TaskItem
> > *task, RequestItem *request)
> > +static int get_datatype_size(DNNDataType dt) {
> > +    switch (dt)
> > +    {
> > +    case DNN_FLOAT:
> > +        return sizeof(float);
> > +    default:
> > +        av_assert0(!"not supported yet.");
> > +        return 1;
> Why don’t try about this way ? :D
> avpriv_request_sample()
> AVERROR_PATCHWELCOME;
thanks, good point, will do it this way.
diff mbox series

Patch

diff --git a/libavfilter/dnn/dnn_backend_openvino.c b/libavfilter/dnn/dnn_backend_openvino.c
index d27e451eea..5271d1caa5 100644
--- a/libavfilter/dnn/dnn_backend_openvino.c
+++ b/libavfilter/dnn/dnn_backend_openvino.c
@@ -37,6 +37,7 @@ 
 typedef struct OVOptions{
     char *device_type;
     int nireq;
+    int batch_size;
 } OVOptions;
 
 typedef struct OVContext {
@@ -70,7 +71,8 @@  typedef struct TaskItem {
 
 typedef struct RequestItem {
     ie_infer_request_t *infer_request;
-    TaskItem *task;
+    TaskItem **tasks;
+    int task_count;
     ie_complete_call_back_t callback;
 } RequestItem;
 
@@ -83,6 +85,7 @@  typedef struct RequestItem {
 static const AVOption dnn_openvino_options[] = {
     { "device", "device to run model", OFFSET(options.device_type), AV_OPT_TYPE_STRING, { .str = "CPU" }, 0, 0, FLAGS },
     { "nireq",  "number of request",   OFFSET(options.nireq),       AV_OPT_TYPE_INT,    { .i64 = 0 },     0, INT_MAX, FLAGS },
+    { "batch_size",  "batch size per request", OFFSET(options.batch_size),  AV_OPT_TYPE_INT,    { .i64 = 1 },     1, 1000, FLAGS},
     { NULL }
 };
 
@@ -100,7 +103,19 @@  static DNNDataType precision_to_datatype(precision_e precision)
     }
 }
 
-static DNNReturnType fill_model_input_ov(OVModel *ov_model, TaskItem *task, RequestItem *request)
+static int get_datatype_size(DNNDataType dt)
+{
+    switch (dt)
+    {
+    case DNN_FLOAT:
+        return sizeof(float);
+    default:
+        av_assert0(!"not supported yet.");
+        return 1;
+    }
+}
+
+static DNNReturnType fill_model_input_ov(OVModel *ov_model, RequestItem *request)
 {
     dimensions_t dims;
     precision_e precision;
@@ -109,6 +124,7 @@  static DNNReturnType fill_model_input_ov(OVModel *ov_model, TaskItem *task, Requ
     IEStatusCode status;
     DNNData input;
     ie_blob_t *input_blob = NULL;
+    TaskItem *task = request->tasks[0];
 
     status = ie_infer_request_get_blob(request->infer_request, task->input_name, &input_blob);
     if (status != OK) {
@@ -134,12 +150,19 @@  static DNNReturnType fill_model_input_ov(OVModel *ov_model, TaskItem *task, Requ
     input.channels = dims.dims[1];
     input.data = blob_buffer.buffer;
     input.dt = precision_to_datatype(precision);
-    if (task->do_ioproc) {
-        if (ov_model->model->pre_proc != NULL) {
-            ov_model->model->pre_proc(task->in_frame, &input, ov_model->model->filter_ctx);
-        } else {
-            proc_from_frame_to_dnn(task->in_frame, &input, ctx);
+
+    av_assert0(request->task_count <= dims.dims[0]);
+    for (int i = 0; i < request->task_count; ++i) {
+        task = request->tasks[i];
+        if (task->do_ioproc) {
+            if (ov_model->model->pre_proc != NULL) {
+                ov_model->model->pre_proc(task->in_frame, &input, ov_model->model->filter_ctx);
+            } else {
+                proc_from_frame_to_dnn(task->in_frame, &input, ctx);
+            }
         }
+        input.data = (uint8_t *)input.data
+                     + input.width * input.height * input.channels * get_datatype_size(input.dt);
     }
     ie_blob_free(&input_blob);
 
@@ -152,7 +175,7 @@  static void infer_completion_callback(void *args)
     precision_e precision;
     IEStatusCode status;
     RequestItem *request = args;
-    TaskItem *task = request->task;
+    TaskItem *task = request->tasks[0];
     ie_blob_t *output_blob = NULL;
     ie_blob_buffer_t blob_buffer;
     DNNData output;
@@ -194,41 +217,56 @@  static void infer_completion_callback(void *args)
     output.width    = dims.dims[3];
     output.dt       = precision_to_datatype(precision);
     output.data     = blob_buffer.buffer;
-    if (task->do_ioproc) {
-        if (task->ov_model->model->post_proc != NULL) {
-            task->ov_model->model->post_proc(task->out_frame, &output, task->ov_model->model->filter_ctx);
+
+    av_assert0(request->task_count <= dims.dims[0]);
+    av_assert0(request->task_count >= 1);
+    for (int i = 0; i < request->task_count; ++i) {
+        task = request->tasks[i];
+        if (task->do_ioproc) {
+            if (task->ov_model->model->post_proc != NULL) {
+                task->ov_model->model->post_proc(task->out_frame, &output, task->ov_model->model->filter_ctx);
+            } else {
+                proc_from_dnn_to_frame(task->out_frame, &output, ctx);
+            }
         } else {
-            proc_from_dnn_to_frame(task->out_frame, &output, ctx);
+            task->out_frame->width = output.width;
+            task->out_frame->height = output.height;
         }
-    } else {
-        task->out_frame->width = output.width;
-        task->out_frame->height = output.height;
+        task->done = 1;
+        output.data = (uint8_t *)output.data
+                      + output.width * output.height * output.channels * get_datatype_size(output.dt);
     }
     ie_blob_free(&output_blob);
 
+    request->task_count = 0;
+
     if (task->async) {
-        request->task = NULL;
         if (ff_safe_queue_push_back(task->ov_model->request_queue, request) < 0) {
             av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
             return;
         }
     }
-
-    task->done = 1;
 }
 
-static DNNReturnType execute_model_ov(TaskItem *task, RequestItem *request)
+static DNNReturnType execute_model_ov(RequestItem *request)
 {
     IEStatusCode status;
+    DNNReturnType ret;
+    TaskItem *task = request->tasks[0];
     OVContext *ctx = &task->ov_model->ctx;
 
-    DNNReturnType ret = fill_model_input_ov(task->ov_model, task, request);
-    if (ret != DNN_SUCCESS) {
-        return ret;
-    }
-
     if (task->async) {
-        request->task = task;
+        if (request->task_count < ctx->options.batch_size) {
+            if (ff_safe_queue_push_front(task->ov_model->request_queue, request) < 0) {
+                av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
+                return DNN_ERROR;
+            }
+            return DNN_SUCCESS;
+        }
+        ret = fill_model_input_ov(task->ov_model, request);
+        if (ret != DNN_SUCCESS) {
+            return ret;
+        }
         status = ie_infer_set_completion_callback(request->infer_request, &request->callback);
         if (status != OK) {
             av_log(ctx, AV_LOG_ERROR, "Failed to set completion callback for inference\n");
@@ -241,12 +279,15 @@  static DNNReturnType execute_model_ov(TaskItem *task, RequestItem *request)
         }
         return DNN_SUCCESS;
     } else {
+        ret = fill_model_input_ov(task->ov_model, request);
+        if (ret != DNN_SUCCESS) {
+            return ret;
+        }
         status = ie_infer_request_infer(request->infer_request);
         if (status != OK) {
             av_log(ctx, AV_LOG_ERROR, "Failed to start synchronous model inference\n");
             return DNN_ERROR;
         }
-        request->task = task;
         infer_completion_callback(request);
         return task->done ? DNN_SUCCESS : DNN_ERROR;
     }
@@ -319,6 +360,7 @@  static DNNReturnType get_output_ov(void *model, const char *input_name, int inpu
     RequestItem request;
     AVFrame *in_frame = av_frame_alloc();
     AVFrame *out_frame = NULL;
+    TaskItem *ptask = &task;
 
     if (!in_frame) {
         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory for input frame\n");
@@ -343,8 +385,10 @@  static DNNReturnType get_output_ov(void *model, const char *input_name, int inpu
     task.ov_model = ov_model;
 
     request.infer_request = ov_model->infer_request;
+    request.task_count = 1;
+    request.tasks = &ptask;
 
-    ret = execute_model_ov(&task, &request);
+    ret = execute_model_ov(&request);
     *output_width = out_frame->width;
     *output_height = out_frame->height;
 
@@ -393,6 +437,24 @@  DNNModel *ff_dnn_load_model_ov(const char *model_filename, const char *options,
     if (status != OK)
         goto err;
 
+    // batch size
+    if (ctx->options.batch_size <= 0) {
+        ctx->options.batch_size = 1;
+    }
+
+    if (ctx->options.batch_size > 1) {
+        input_shapes_t input_shapes;
+        status = ie_network_get_input_shapes(ov_model->network, &input_shapes);
+        if (status != OK)
+            goto err;
+        for (int i = 0; i < input_shapes.shape_num; i++)
+            input_shapes.shapes[i].shape.dims[0] = ctx->options.batch_size;
+        status = ie_network_reshape(ov_model->network, input_shapes);
+        ie_network_input_shapes_free(&input_shapes);
+        if (status != OK)
+            goto err;
+    }
+
     status = ie_core_load_network(ov_model->core, ov_model->network, ctx->options.device_type, &config, &ov_model->exe_network);
     if (status != OK) {
         av_log(ctx, AV_LOG_ERROR, "Failed to init OpenVINO model\n");
@@ -426,17 +488,24 @@  DNNModel *ff_dnn_load_model_ov(const char *model_filename, const char *options,
     }
 
     for (int i = 0; i < ctx->options.nireq; i++) {
-        ie_infer_request_t *request;
         RequestItem *item = av_mallocz(sizeof(*item));
         if (!item) {
             goto err;
         }
-        status = ie_exec_network_create_infer_request(ov_model->exe_network, &request);
+
+        status = ie_exec_network_create_infer_request(ov_model->exe_network, &item->infer_request);
         if (status != OK) {
             av_freep(&item);
             goto err;
         }
-        item->infer_request = request;
+
+        item->tasks = av_malloc_array(ctx->options.batch_size, sizeof(*item->tasks));
+        if (!item->tasks) {
+            av_freep(&item);
+            goto err;
+        }
+        item->task_count = 0;
+
         item->callback.completeCallBackFunc = infer_completion_callback;
         item->callback.args = item;
         if (ff_safe_queue_push_back(ov_model->request_queue, item) < 0) {
@@ -469,6 +538,7 @@  DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
     OVContext *ctx = &ov_model->ctx;
     TaskItem task;
     RequestItem request;
+    TaskItem *ptask = &task;
 
     if (!in_frame) {
         av_log(ctx, AV_LOG_ERROR, "in frame is NULL when execute model.\n");
@@ -487,6 +557,11 @@  DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
         return DNN_ERROR;
     }
 
+    if (ctx->options.batch_size > 1) {
+        av_log(ctx, AV_LOG_ERROR, "do not support batch mode for sync execution.\n");
+        return DNN_ERROR;
+    }
+
     task.done = 0;
     task.do_ioproc = 1;
     task.async = 0;
@@ -497,8 +572,10 @@  DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
     task.ov_model = ov_model;
 
     request.infer_request = ov_model->infer_request;
+    request.task_count = 1;
+    request.tasks = &ptask;
 
-    return execute_model_ov(&task, &request);
+    return execute_model_ov(&request);
 }
 
 DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
@@ -545,7 +622,8 @@  DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, const char *i
         return DNN_ERROR;
     }
 
-    return execute_model_ov(task, request);
+    request->tasks[request->task_count++] = task;
+    return execute_model_ov(request);
 }
 
 DNNAsyncStatusType ff_dnn_get_async_result_ov(const DNNModel *model, AVFrame **in, AVFrame **out)
@@ -569,6 +647,48 @@  DNNAsyncStatusType ff_dnn_get_async_result_ov(const DNNModel *model, AVFrame **i
     return DAST_SUCCESS;
 }
 
+DNNReturnType ff_dnn_flush_ov(const DNNModel *model)
+{
+    OVModel *ov_model = (OVModel *)model->model;
+    OVContext *ctx = &ov_model->ctx;
+    RequestItem *request;
+    IEStatusCode status;
+    DNNReturnType ret;
+
+    request = ff_safe_queue_pop_front(ov_model->request_queue);
+    if (!request) {
+        av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
+        return DNN_ERROR;
+    }
+
+    if (request->task_count == 0) {
+        // no pending task need to flush
+        if (ff_safe_queue_push_back(ov_model->request_queue, request) < 0) {
+            av_log(ctx, AV_LOG_ERROR, "Failed to push back request_queue.\n");
+            return DNN_ERROR;
+        }
+        return DNN_SUCCESS;
+    }
+
+    ret = fill_model_input_ov(ov_model, request);
+    if (ret != DNN_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to fill model input.\n");
+        return ret;
+    }
+    status = ie_infer_set_completion_callback(request->infer_request, &request->callback);
+    if (status != OK) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to set completion callback for inference\n");
+        return DNN_ERROR;
+    }
+    status = ie_infer_request_infer_async(request->infer_request);
+    if (status != OK) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to start async inference\n");
+        return DNN_ERROR;
+    }
+
+    return DNN_SUCCESS;
+}
+
 void ff_dnn_free_model_ov(DNNModel **model)
 {
     if (*model){
@@ -578,12 +698,15 @@  void ff_dnn_free_model_ov(DNNModel **model)
             if (item && item->infer_request) {
                 ie_infer_request_free(&item->infer_request);
             }
+            av_freep(&item->tasks);
             av_freep(&item);
         }
         ff_safe_queue_destroy(ov_model->request_queue);
 
         while (ff_queue_size(ov_model->task_queue) != 0) {
             TaskItem *item = ff_queue_pop_front(ov_model->task_queue);
+            av_frame_free(&item->in_frame);
+            av_frame_free(&item->out_frame);
             av_freep(&item);
         }
         ff_queue_destroy(ov_model->task_queue);
diff --git a/libavfilter/dnn/dnn_backend_openvino.h b/libavfilter/dnn/dnn_backend_openvino.h
index 1b70150040..23b819440e 100644
--- a/libavfilter/dnn/dnn_backend_openvino.h
+++ b/libavfilter/dnn/dnn_backend_openvino.h
@@ -36,6 +36,7 @@  DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
 DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
                                             const char **output_names, uint32_t nb_output, AVFrame *out_frame);
 DNNAsyncStatusType ff_dnn_get_async_result_ov(const DNNModel *model, AVFrame **in, AVFrame **out);
+DNNReturnType ff_dnn_flush_ov(const DNNModel *model);
 
 void ff_dnn_free_model_ov(DNNModel **model);
 
diff --git a/libavfilter/dnn/dnn_interface.c b/libavfilter/dnn/dnn_interface.c
index e1b41a21e1..02e532fc1b 100644
--- a/libavfilter/dnn/dnn_interface.c
+++ b/libavfilter/dnn/dnn_interface.c
@@ -60,6 +60,7 @@  DNNModule *ff_get_dnn_module(DNNBackendType backend_type)
         dnn_module->execute_model = &ff_dnn_execute_model_ov;
         dnn_module->execute_model_async = &ff_dnn_execute_model_async_ov;
         dnn_module->get_async_result = &ff_dnn_get_async_result_ov;
+        dnn_module->flush = &ff_dnn_flush_ov;
         dnn_module->free_model = &ff_dnn_free_model_ov;
     #else
         av_freep(&dnn_module);
diff --git a/libavfilter/dnn_interface.h b/libavfilter/dnn_interface.h
index 9533c88829..ff338ea084 100644
--- a/libavfilter/dnn_interface.h
+++ b/libavfilter/dnn_interface.h
@@ -82,6 +82,8 @@  typedef struct DNNModule{
                                          const char **output_names, uint32_t nb_output, AVFrame *out_frame);
     // Retrieve inference result.
     DNNAsyncStatusType (*get_async_result)(const DNNModel *model, AVFrame **in, AVFrame **out);
+    // Flush all the pending tasks.
+    DNNReturnType (*flush)(const DNNModel *model);
     // Frees memory allocated for model.
     void (*free_model)(DNNModel **model);
 } DNNModule;
diff --git a/libavfilter/vf_dnn_processing.c b/libavfilter/vf_dnn_processing.c
index fff5696a31..be48631782 100644
--- a/libavfilter/vf_dnn_processing.c
+++ b/libavfilter/vf_dnn_processing.c
@@ -33,6 +33,7 @@ 
 #include "formats.h"
 #include "internal.h"
 #include "libswscale/swscale.h"
+#include "libavutil/time.h"
 
 typedef struct DnnProcessingContext {
     const AVClass *class;
@@ -369,6 +370,37 @@  static int activate_sync(AVFilterContext *filter_ctx)
     return FFERROR_NOT_READY;
 }
 
+static int flush_frame(AVFilterLink *outlink, int64_t pts, int64_t *out_pts)
+{
+    DnnProcessingContext *ctx = outlink->src->priv;
+    int ret;
+    DNNAsyncStatusType async_state;
+
+    ret = (ctx->dnn_module->flush)(ctx->model);
+    if (ret != DNN_SUCCESS) {
+        return -1;
+    }
+
+    do {
+        AVFrame *in_frame = NULL;
+        AVFrame *out_frame = NULL;
+        async_state = (ctx->dnn_module->get_async_result)(ctx->model, &in_frame, &out_frame);
+        if (out_frame) {
+            if (isPlanarYUV(in_frame->format))
+                copy_uv_planes(ctx, out_frame, in_frame);
+            av_frame_free(&in_frame);
+            ret = ff_filter_frame(outlink, out_frame);
+            if (ret < 0)
+                return ret;
+            if (out_pts)
+                *out_pts = out_frame->pts + pts;
+        }
+        av_usleep(5000);
+    } while (async_state >= DAST_NOT_READY);
+
+    return 0;
+}
+
 static int activate_async(AVFilterContext *filter_ctx)
 {
     AVFilterLink *inlink = filter_ctx->inputs[0];
@@ -423,7 +455,9 @@  static int activate_async(AVFilterContext *filter_ctx)
 
     if (ff_inlink_acknowledge_status(inlink, &status, &pts)) {
         if (status == AVERROR_EOF) {
-            ff_outlink_set_status(outlink, status, pts);
+            int64_t out_pts = pts;
+            ret = flush_frame(outlink, pts, &out_pts);
+            ff_outlink_set_status(outlink, status, out_pts);
             return ret;
         }
     }