History

lavfi/dnn: Async Support for TensorFlow Backend

This commit enables async execution in the TensorFlow backend and adds function to flush extra frames. The async execution mechanism executes the TFInferRequests on a separate thread which is joined before the next execution of same TFRequestItem/while freeing the model. The following is the comparison of this mechanism with the existing sync mechanism on TensorFlow C API 2.5 CPU variant. Async Mode: 4m32.846s Sync Mode: 5m17.582s The above was performed on super resolution filter using SRCNN model.

Shubhanshu Saxena committed 5 years ago

Commit 0985e921 parent e6ae8fc3 files · +109 −18Browse files

Unified · Split

3 files
  1. libavfilter/dnn/dnn_backend_tf.c +103 −18
  2. libavfilter/dnn/dnn_backend_tf.h +3 −0
  3. libavfilter/dnn/dnn_interface.c +3 −0
libavfilter/dnn/dnn_backend_tf.c+103−18
@@ -38,7 +38,6 @@ · 37 unmodified lines
3838 #include "dnn_io_proc.h"
3939 #include "dnn_backend_common.h"
4040 #include "safe_queue.h"
41−#include "queue.h"
4241 #include <tensorflow/c/c_api.h>
4342
4443 typedef struct TFOptions{
@@ -59,6 +58,7 @@ typedef struct TFModel{ · 14 unmodified lines
5958 TF_Status *status;
6059 SafeQueue *request_queue;
6160 Queue *inference_queue;
61+ Queue *task_queue;
6262 } TFModel;
6363
6464 /**
@@ -75,7 +75,7 @@ typedef struct TFInferRequest { · 10 unmodified lines
7575 typedef struct TFRequestItem {
7676 TFInferRequest *infer_request;
7777 InferenceItem *inference;
78− // further properties will be added later for async
78+ DNNAsyncExecModule exec_module;
7979 } TFRequestItem;
8080
8181 #define OFFSET(x) offsetof(TFContext, x)
@@ -89,6 +89,7 @@ static const AVOption dnn_tensorflow_options[] = { · 7 unmodified lines
8989 AVFILTER_DEFINE_CLASS(dnn_tensorflow);
9090
9191 static DNNReturnType execute_model_tf(TFRequestItem *request, Queue *inference_queue);
92+static void infer_completion_callback(void *args);
9293
9394 static void free_buffer(void *data, size_t length)
9495 {
@@ -886,6 +887,9 @@ DNNModel *ff_dnn_load_model_tf(const char *model_filename, DNNFunctionType func_ · 791 unmodified lines
886887 av_freep(&item);
887888 goto err;
888889 }
890+ item->exec_module.start_inference = &tf_start_inference;
891+ item->exec_module.callback = &infer_completion_callback;
892+ item->exec_module.args = item;
889893
890894 if (ff_safe_queue_push_back(tf_model->request_queue, item) < 0) {
891895 av_freep(&item->infer_request);
@@ -899,6 +903,11 @@ DNNModel *ff_dnn_load_model_tf(const char *model_filename, DNNFunctionType func_ · 7 unmodified lines
899903 goto err;
900904 }
901905
906+ tf_model->task_queue = ff_queue_create();
907+ if (!tf_model->task_queue) {
908+ goto err;
909+ }
910+
902911 model->model = tf_model;
903912 model->get_input = &get_input_tf;
904913 model->get_output = &get_output_tf;
@@ -1061,7 +1070,6 @@ static DNNReturnType execute_model_tf(TFRequestItem *request, Queue *inference_q · 156 unmodified lines
10611070 {
10621071 TFModel *tf_model;
10631072 TFContext *ctx;
1064− TFInferRequest *infer_request;
10651073 InferenceItem *inference;
10661074 TaskItem *task;
10671075
@@ -1074,23 +1082,14 @@ static DNNReturnType execute_model_tf(TFRequestItem *request, Queue *inference_q · 6 unmodified lines
10741082 tf_model = task->model;
10751083 ctx = &tf_model->ctx;
10761084
1077− if (task->async) {
1078− avpriv_report_missing_feature(ctx, "Async execution not supported");
1085+ if (fill_model_input_tf(tf_model, request) != DNN_SUCCESS) {
10791086 return DNN_ERROR;
1080− } else {
1081− if (fill_model_input_tf(tf_model, request) != DNN_SUCCESS) {
1082− return DNN_ERROR;
1083− }
1087+ }
10841088
1085− infer_request = request->infer_request;
1086− TF_SessionRun(tf_model->session, NULL,
1087− infer_request->tf_input, &infer_request->input_tensor, 1,
1088− infer_request->tf_outputs, infer_request->output_tensors,
1089− task->nb_output, NULL, 0, NULL,
1090− tf_model->status);
1091− if (TF_GetCode(tf_model->status) != TF_OK) {
1092− tf_free_request(infer_request);
1093− av_log(ctx, AV_LOG_ERROR, "Failed to run session when executing model\n");
1089+ if (task->async) {
1090+ return ff_dnn_start_inference_async(ctx, &request->exec_module);
1091+ } else {
1092+ if (tf_start_inference(request) != DNN_SUCCESS) {
10941093 return DNN_ERROR;
10951094 }
10961095 infer_completion_callback(request);
@@ -1127,6 +1126,83 @@ DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNExecBaseParams * · 30 unmodified lines
11271126 return execute_model_tf(request, tf_model->inference_queue);
11281127 }
11291128
1129+DNNReturnType ff_dnn_execute_model_async_tf(const DNNModel *model, DNNExecBaseParams *exec_params) {
1130+ TFModel *tf_model = model->model;
1131+ TFContext *ctx = &tf_model->ctx;
1132+ TaskItem *task;
1133+ TFRequestItem *request;
1134+
1135+ if (ff_check_exec_params(ctx, DNN_TF, model->func_type, exec_params) != 0) {
1136+ return DNN_ERROR;
1137+ }
1138+
1139+ task = av_malloc(sizeof(*task));
1140+ if (!task) {
1141+ av_log(ctx, AV_LOG_ERROR, "unable to alloc memory for task item.\n");
1142+ return DNN_ERROR;
1143+ }
1144+
1145+ if (ff_dnn_fill_task(task, exec_params, tf_model, 1, 1) != DNN_SUCCESS) {
1146+ av_freep(&task);
1147+ return DNN_ERROR;
1148+ }
1149+
1150+ if (ff_queue_push_back(tf_model->task_queue, task) < 0) {
1151+ av_freep(&task);
1152+ av_log(ctx, AV_LOG_ERROR, "unable to push back task_queue.\n");
1153+ return DNN_ERROR;
1154+ }
1155+
1156+ if (extract_inference_from_task(task, tf_model->inference_queue) != DNN_SUCCESS) {
1157+ av_log(ctx, AV_LOG_ERROR, "unable to extract inference from task.\n");
1158+ return DNN_ERROR;
1159+ }
1160+
1161+ request = ff_safe_queue_pop_front(tf_model->request_queue);
1162+ if (!request) {
1163+ av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
1164+ return DNN_ERROR;
1165+ }
1166+ return execute_model_tf(request, tf_model->inference_queue);
1167+}
1168+
1169+DNNAsyncStatusType ff_dnn_get_async_result_tf(const DNNModel *model, AVFrame **in, AVFrame **out)
1170+{
1171+ TFModel *tf_model = model->model;
1172+ return ff_dnn_get_async_result_common(tf_model->task_queue, in, out);
1173+}
1174+
1175+DNNReturnType ff_dnn_flush_tf(const DNNModel *model)
1176+{
1177+ TFModel *tf_model = model->model;
1178+ TFContext *ctx = &tf_model->ctx;
1179+ TFRequestItem *request;
1180+ DNNReturnType ret;
1181+
1182+ if (ff_queue_size(tf_model->inference_queue) == 0) {
1183+ // no pending task need to flush
1184+ return DNN_SUCCESS;
1185+ }
1186+
1187+ request = ff_safe_queue_pop_front(tf_model->request_queue);
1188+ if (!request) {
1189+ av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
1190+ return DNN_ERROR;
1191+ }
1192+
1193+ ret = fill_model_input_tf(tf_model, request);
1194+ if (ret != DNN_SUCCESS) {
1195+ av_log(ctx, AV_LOG_ERROR, "Failed to fill model input.\n");
1196+ if (ff_safe_queue_push_back(tf_model->request_queue, request) < 0) {
1197+ av_freep(&request->infer_request);
1198+ av_freep(&request);
1199+ }
1200+ return ret;
1201+ }
1202+
1203+ return ff_dnn_start_inference_async(ctx, &request->exec_module);
1204+}
1205+
11301206 void ff_dnn_free_model_tf(DNNModel **model)
11311207 {
11321208 TFModel *tf_model;
@@ -1135,6 +1211,7 @@ void ff_dnn_free_model_tf(DNNModel **model) · 2 unmodified lines
11351211 tf_model = (*model)->model;
11361212 while (ff_safe_queue_size(tf_model->request_queue) != 0) {
11371213 TFRequestItem *item = ff_safe_queue_pop_front(tf_model->request_queue);
1214+ ff_dnn_async_module_cleanup(&item->exec_module);
11381215 tf_free_request(item->infer_request);
11391216 av_freep(&item->infer_request);
11401217 av_freep(&item);
@@ -1147,6 +1224,14 @@ void ff_dnn_free_model_tf(DNNModel **model) · 6 unmodified lines
11471224 }
11481225 ff_queue_destroy(tf_model->inference_queue);
11491226
1227+ while (ff_queue_size(tf_model->task_queue) != 0) {
1228+ TaskItem *item = ff_queue_pop_front(tf_model->task_queue);
1229+ av_frame_free(&item->in_frame);
1230+ av_frame_free(&item->out_frame);
1231+ av_freep(&item);
1232+ }
1233+ ff_queue_destroy(tf_model->task_queue);
1234+
11501235 if (tf_model->graph){
11511236 TF_DeleteGraph(tf_model->graph);
11521237 }
libavfilter/dnn/dnn_backend_tf.h+3−0
@@ -32,6 +32,9 @@ · 31 unmodified lines
3232 DNNModel *ff_dnn_load_model_tf(const char *model_filename, DNNFunctionType func_type, const char *options, AVFilterContext *filter_ctx);
3333
3434 DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNExecBaseParams *exec_params);
35+DNNReturnType ff_dnn_execute_model_async_tf(const DNNModel *model, DNNExecBaseParams *exec_params);
36+DNNAsyncStatusType ff_dnn_get_async_result_tf(const DNNModel *model, AVFrame **in, AVFrame **out);
37+DNNReturnType ff_dnn_flush_tf(const DNNModel *model);
3538
3639 void ff_dnn_free_model_tf(DNNModel **model);
3740
libavfilter/dnn/dnn_interface.c+3−0
@@ -48,6 +48,9 @@ DNNModule *ff_get_dnn_module(DNNBackendType backend_type) · 47 unmodified lines
4848 #if (CONFIG_LIBTENSORFLOW == 1)
4949 dnn_module->load_model = &ff_dnn_load_model_tf;
5050 dnn_module->execute_model = &ff_dnn_execute_model_tf;
51+ dnn_module->execute_model_async = &ff_dnn_execute_model_async_tf;
52+ dnn_module->get_async_result = &ff_dnn_get_async_result_tf;
53+ dnn_module->flush = &ff_dnn_flush_tf;
5154 dnn_module->free_model = &ff_dnn_free_model_tf;
5255 #else
5356 av_freep(&dnn_module);

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t