Message ID | 20210110131601.22685-1-yejun.guo@intel.com |
---|---|
State | Accepted |
Headers | show |
Series | [FFmpeg-devel,V2] libavfilter/dnn: add batch mode for async execution | expand |
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 |
> -----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".
> -----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.
> 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
> -----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 --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; } }