libavfilter: Unify Execution Modes in DNN Filters

This commit unifies the async and sync mode from the DNN filters'
perspective. As of this commit, the Native backend only supports
synchronous execution mode.

Now the user can switch between async and sync mode by using the
'async' option in the backend_configs. The values can be 1 for
async and 0 for sync mode of execution.

This commit affects the following filters:
1. vf_dnn_classify
2. vf_dnn_detect
3. vf_dnn_processing
4. vf_sr
5. vf_derain

This commit also updates the filters vf_dnn_detect and vf_dnn_classify
to send only the input frame and send NULL as output frame instead of
input frame to the DNN backends.

Signed-off-by: Shubhanshu Saxena <shubhanshu.e01@gmail.com>

9 files changed, 115 insertions, 109 deletions

diff --git a/libavfilter/dnn/dnn_backend_common.c b/libavfilter/dnn/dnn_backend_common.c
index 426683b73d..6a9c4cc87f 100644
--- a/libavfilter/dnn/dnn_backend_common.c
+++ b/libavfilter/dnn/dnn_backend_common.c
@@ -38,7 +38,7 @@ int ff_check_exec_params(void *ctx, DNNBackendType backend, DNNFunctionType func
         return AVERROR(EINVAL);
     }
 
-    if (!exec_params->out_frame) {
+    if (!exec_params->out_frame && func_type == DFT_PROCESS_FRAME) {
         av_log(ctx, AV_LOG_ERROR, "out frame is NULL when execute model.\n");
         return AVERROR(EINVAL);
     }
@@ -138,7 +138,7 @@ DNNReturnType ff_dnn_start_inference_async(void *ctx, DNNAsyncExecModule *async_
     return DNN_SUCCESS;
 }
 
-DNNAsyncStatusType ff_dnn_get_async_result_common(Queue *task_queue, AVFrame **in, AVFrame **out)
+DNNAsyncStatusType ff_dnn_get_result_common(Queue *task_queue, AVFrame **in, AVFrame **out)
 {
     TaskItem *task = ff_queue_peek_front(task_queue);
 
diff --git a/libavfilter/dnn/dnn_backend_common.h b/libavfilter/dnn/dnn_backend_common.h
index 604c1d3bd7..78e62a94a2 100644
--- a/libavfilter/dnn/dnn_backend_common.h
+++ b/libavfilter/dnn/dnn_backend_common.h
@@ -29,7 +29,8 @@
 #include "libavutil/thread.h"
 
 #define DNN_BACKEND_COMMON_OPTIONS \
-    { "nireq",           "number of request",             OFFSET(options.nireq),           AV_OPT_TYPE_INT,    { .i64 = 0 },     0, INT_MAX, FLAGS },
+    { "nireq",           "number of request",             OFFSET(options.nireq),           AV_OPT_TYPE_INT,    { .i64 = 0 },     0, INT_MAX, FLAGS }, \
+    { "async",           "use DNN async inference",       OFFSET(options.async),           AV_OPT_TYPE_BOOL,   { .i64 = 1 },     0,       1, FLAGS },
 
 // one task for one function call from dnn interface
 typedef struct TaskItem {
@@ -135,7 +136,7 @@ DNNReturnType ff_dnn_start_inference_async(void *ctx, DNNAsyncExecModule *async_
  * @retval DAST_NOT_READY if inference not completed yet.
  * @retval DAST_SUCCESS if result successfully extracted
  */
-DNNAsyncStatusType ff_dnn_get_async_result_common(Queue *task_queue, AVFrame **in, AVFrame **out);
+DNNAsyncStatusType ff_dnn_get_result_common(Queue *task_queue, AVFrame **in, AVFrame **out);
 
 /**
  * Allocate input and output frames and fill the Task
diff --git a/libavfilter/dnn/dnn_backend_native.c b/libavfilter/dnn/dnn_backend_native.c
index 3b2a3aa55d..2d34b88f8a 100644
--- a/libavfilter/dnn/dnn_backend_native.c
+++ b/libavfilter/dnn/dnn_backend_native.c
@@ -34,6 +34,7 @@
 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM
 static const AVOption dnn_native_options[] = {
     { "conv2d_threads", "threads num for conv2d layer", OFFSET(options.conv2d_threads), AV_OPT_TYPE_INT,  { .i64 = 0 }, INT_MIN, INT_MAX, FLAGS },
+    { "async",          "use DNN async inference",      OFFSET(options.async),          AV_OPT_TYPE_BOOL, { .i64 = 0 },       0,       1, FLAGS },
     { NULL },
 };
 
@@ -189,6 +190,11 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename, DNNFunctionType f
         goto fail;
     native_model->model = model;
 
+    if (native_model->ctx.options.async) {
+        av_log(&native_model->ctx, AV_LOG_WARNING, "Async not supported. Rolling back to sync\n");
+        native_model->ctx.options.async = 0;
+    }
+
 #if !HAVE_PTHREAD_CANCEL
     if (native_model->ctx.options.conv2d_threads > 1){
         av_log(&native_model->ctx, AV_LOG_WARNING, "'conv2d_threads' option was set but it is not supported "
@@ -212,6 +218,11 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename, DNNFunctionType f
         goto fail;
     }
 
+    native_model->task_queue = ff_queue_create();
+    if (!native_model->task_queue) {
+        goto fail;
+    }
+
     native_model->inference_queue = ff_queue_create();
     if (!native_model->inference_queue) {
         goto fail;
@@ -425,17 +436,30 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNExecBasePara
 {
     NativeModel *native_model = model->model;
     NativeContext *ctx = &native_model->ctx;
-    TaskItem task;
+    TaskItem *task;
 
     if (ff_check_exec_params(ctx, DNN_NATIVE, model->func_type, exec_params) != 0) {
         return DNN_ERROR;
     }
 
-    if (ff_dnn_fill_task(&task, exec_params, native_model, 0, 1) != DNN_SUCCESS) {
+    task = av_malloc(sizeof(*task));
+    if (!task) {
+        av_log(ctx, AV_LOG_ERROR, "unable to alloc memory for task item.\n");
         return DNN_ERROR;
     }
 
-    if (extract_inference_from_task(&task, native_model->inference_queue) != DNN_SUCCESS) {
+    if (ff_dnn_fill_task(task, exec_params, native_model, ctx->options.async, 1) != DNN_SUCCESS) {
+        av_freep(&task);
+        return DNN_ERROR;
+    }
+
+    if (ff_queue_push_back(native_model->task_queue, task) < 0) {
+        av_freep(&task);
+        av_log(ctx, AV_LOG_ERROR, "unable to push back task_queue.\n");
+        return DNN_ERROR;
+    }
+
+    if (extract_inference_from_task(task, native_model->inference_queue) != DNN_SUCCESS) {
         av_log(ctx, AV_LOG_ERROR, "unable to extract inference from task.\n");
         return DNN_ERROR;
     }
@@ -443,6 +467,26 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNExecBasePara
     return execute_model_native(native_model->inference_queue);
 }
 
+DNNReturnType ff_dnn_flush_native(const DNNModel *model)
+{
+    NativeModel *native_model = model->model;
+
+    if (ff_queue_size(native_model->inference_queue) == 0) {
+        // no pending task need to flush
+        return DNN_SUCCESS;
+    }
+
+    // for now, use sync node with flush operation
+    // Switch to async when it is supported
+    return execute_model_native(native_model->inference_queue);
+}
+
+DNNAsyncStatusType ff_dnn_get_result_native(const DNNModel *model, AVFrame **in, AVFrame **out)
+{
+    NativeModel *native_model = model->model;
+    return ff_dnn_get_result_common(native_model->task_queue, in, out);
+}
+
 int32_t ff_calculate_operand_dims_count(const DnnOperand *oprd)
 {
     int32_t result = 1;
@@ -497,6 +541,15 @@ void ff_dnn_free_model_native(DNNModel **model)
                 av_freep(&item);
             }
             ff_queue_destroy(native_model->inference_queue);
+
+            while (ff_queue_size(native_model->task_queue) != 0) {
+                TaskItem *item = ff_queue_pop_front(native_model->task_queue);
+                av_frame_free(&item->in_frame);
+                av_frame_free(&item->out_frame);
+                av_freep(&item);
+            }
+            ff_queue_destroy(native_model->task_queue);
+
             av_freep(&native_model);
         }
         av_freep(model);
diff --git a/libavfilter/dnn/dnn_backend_native.h b/libavfilter/dnn/dnn_backend_native.h
index 1b9d5bdf2d..ca61bb353f 100644
--- a/libavfilter/dnn/dnn_backend_native.h
+++ b/libavfilter/dnn/dnn_backend_native.h
@@ -111,6 +111,7 @@ typedef struct InputParams{
 } InputParams;
 
 typedef struct NativeOptions{
+    uint8_t async;
     uint32_t conv2d_threads;
 } NativeOptions;
 
@@ -127,6 +128,7 @@ typedef struct NativeModel{
     int32_t layers_num;
     DnnOperand *operands;
     int32_t operands_num;
+    Queue *task_queue;
     Queue *inference_queue;
 } NativeModel;
 
@@ -134,6 +136,10 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename, DNNFunctionType f
 
 DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNExecBaseParams *exec_params);
 
+DNNAsyncStatusType ff_dnn_get_result_native(const DNNModel *model, AVFrame **in, AVFrame **out);
+
+DNNReturnType ff_dnn_flush_native(const DNNModel *model);
+
 void ff_dnn_free_model_native(DNNModel **model);
 
 // NOTE: User must check for error (return value <= 0) to handle
diff --git a/libavfilter/dnn/dnn_backend_openvino.c b/libavfilter/dnn/dnn_backend_openvino.c
index c825e70c82..bf13b017fb 100644
--- a/libavfilter/dnn/dnn_backend_openvino.c
+++ b/libavfilter/dnn/dnn_backend_openvino.c
@@ -39,6 +39,7 @@
 typedef struct OVOptions{
     char *device_type;
     int nireq;
+    uint8_t async;
     int batch_size;
     int input_resizable;
 } OVOptions;
@@ -271,14 +272,14 @@ static void infer_completion_callback(void *args)
                 av_log(ctx, AV_LOG_ERROR, "detect filter needs to provide post proc\n");
                 return;
             }
-            ov_model->model->detect_post_proc(task->out_frame, &output, 1, ov_model->model->filter_ctx);
+            ov_model->model->detect_post_proc(task->in_frame, &output, 1, ov_model->model->filter_ctx);
             break;
         case DFT_ANALYTICS_CLASSIFY:
             if (!ov_model->model->classify_post_proc) {
                 av_log(ctx, AV_LOG_ERROR, "classify filter needs to provide post proc\n");
                 return;
             }
-            ov_model->model->classify_post_proc(task->out_frame, &output, request->inferences[i]->bbox_index, ov_model->model->filter_ctx);
+            ov_model->model->classify_post_proc(task->in_frame, &output, request->inferences[i]->bbox_index, ov_model->model->filter_ctx);
             break;
         default:
             av_assert0(!"should not reach here");
@@ -761,55 +762,6 @@ DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, DNNExecBaseParams *
 {
     OVModel *ov_model = model->model;
     OVContext *ctx = &ov_model->ctx;
-    TaskItem task;
-    OVRequestItem *request;
-
-    if (ff_check_exec_params(ctx, DNN_OV, model->func_type, exec_params) != 0) {
-        return DNN_ERROR;
-    }
-
-    if (model->func_type == DFT_ANALYTICS_CLASSIFY) {
-        // Once we add async support for tensorflow backend and native backend,
-        // we'll combine the two sync/async functions in dnn_interface.h to
-        // simplify the code in filter, and async will be an option within backends.
-        // so, do not support now, and classify filter will not call this function.
-        return DNN_ERROR;
-    }
-
-    if (ctx->options.batch_size > 1) {
-        avpriv_report_missing_feature(ctx, "batch mode for sync execution");
-        return DNN_ERROR;
-    }
-
-    if (!ov_model->exe_network) {
-        if (init_model_ov(ov_model, exec_params->input_name, exec_params->output_names[0]) != DNN_SUCCESS) {
-            av_log(ctx, AV_LOG_ERROR, "Failed init OpenVINO exectuable network or inference request\n");
-            return DNN_ERROR;
-        }
-    }
-
-    if (ff_dnn_fill_task(&task, exec_params, ov_model, 0, 1) != DNN_SUCCESS) {
-        return DNN_ERROR;
-    }
-
-    if (extract_inference_from_task(ov_model->model->func_type, &task, ov_model->inference_queue, exec_params) != DNN_SUCCESS) {
-        av_log(ctx, AV_LOG_ERROR, "unable to extract inference from task.\n");
-        return DNN_ERROR;
-    }
-
-    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;
-    }
-
-    return execute_model_ov(request, ov_model->inference_queue);
-}
-
-DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, DNNExecBaseParams *exec_params)
-{
-    OVModel *ov_model = model->model;
-    OVContext *ctx = &ov_model->ctx;
     OVRequestItem *request;
     TaskItem *task;
     DNNReturnType ret;
@@ -831,7 +783,8 @@ DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, DNNExecBasePa
         return DNN_ERROR;
     }
 
-    if (ff_dnn_fill_task(task, exec_params, ov_model, 1, 1) != DNN_SUCCESS) {
+    if (ff_dnn_fill_task(task, exec_params, ov_model, ctx->options.async, 1) != DNN_SUCCESS) {
+        av_freep(&task);
         return DNN_ERROR;
     }
 
@@ -846,26 +799,48 @@ DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, DNNExecBasePa
         return DNN_ERROR;
     }
 
-    while (ff_queue_size(ov_model->inference_queue) >= ctx->options.batch_size) {
+    if (ctx->options.async) {
+        while (ff_queue_size(ov_model->inference_queue) >= ctx->options.batch_size) {
+            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;
+            }
+
+            ret = execute_model_ov(request, ov_model->inference_queue);
+            if (ret != DNN_SUCCESS) {
+                return ret;
+            }
+        }
+
+        return DNN_SUCCESS;
+    }
+    else {
+        if (model->func_type == DFT_ANALYTICS_CLASSIFY) {
+            // Classification filter has not been completely
+            // tested with the sync mode. So, do not support now.
+            avpriv_report_missing_feature(ctx, "classify for sync execution");
+            return DNN_ERROR;
+        }
+
+        if (ctx->options.batch_size > 1) {
+            avpriv_report_missing_feature(ctx, "batch mode for sync execution");
+            return DNN_ERROR;
+        }
+
         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;
         }
-
-        ret = execute_model_ov(request, ov_model->inference_queue);
-        if (ret != DNN_SUCCESS) {
-            return ret;
-        }
+        return execute_model_ov(request, ov_model->inference_queue);
     }
-
-    return DNN_SUCCESS;
 }
 
-DNNAsyncStatusType ff_dnn_get_async_result_ov(const DNNModel *model, AVFrame **in, AVFrame **out)
+DNNAsyncStatusType ff_dnn_get_result_ov(const DNNModel *model, AVFrame **in, AVFrame **out)
 {
     OVModel *ov_model = model->model;
-    return ff_dnn_get_async_result_common(ov_model->task_queue, in, out);
+    return ff_dnn_get_result_common(ov_model->task_queue, in, out);
 }
 
 DNNReturnType ff_dnn_flush_ov(const DNNModel *model)
diff --git a/libavfilter/dnn/dnn_backend_openvino.h b/libavfilter/dnn/dnn_backend_openvino.h
index 046d0c5b5a..0bbca0c057 100644
--- a/libavfilter/dnn/dnn_backend_openvino.h
+++ b/libavfilter/dnn/dnn_backend_openvino.h
@@ -32,8 +32,7 @@
 DNNModel *ff_dnn_load_model_ov(const char *model_filename, DNNFunctionType func_type, const char *options, AVFilterContext *filter_ctx);
 
 DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, DNNExecBaseParams *exec_params);
-DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, DNNExecBaseParams *exec_params);
-DNNAsyncStatusType ff_dnn_get_async_result_ov(const DNNModel *model, AVFrame **in, AVFrame **out);
+DNNAsyncStatusType ff_dnn_get_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_backend_tf.c b/libavfilter/dnn/dnn_backend_tf.c
index ffec1b1328..84952ece5a 100644
--- a/libavfilter/dnn/dnn_backend_tf.c
+++ b/libavfilter/dnn/dnn_backend_tf.c
@@ -42,6 +42,7 @@
 
 typedef struct TFOptions{
     char *sess_config;
+    uint8_t async;
     uint32_t nireq;
 } TFOptions;
 
@@ -1061,7 +1062,7 @@ static void infer_completion_callback(void *args) {
             av_log(ctx, AV_LOG_ERROR, "Detect filter needs provide post proc\n");
             return;
         }
-        tf_model->model->detect_post_proc(task->out_frame, outputs, task->nb_output, tf_model->model->filter_ctx);
+        tf_model->model->detect_post_proc(task->in_frame, outputs, task->nb_output, tf_model->model->filter_ctx);
         break;
     default:
         av_log(ctx, AV_LOG_ERROR, "Tensorflow backend does not support this kind of dnn filter now\n");
@@ -1123,34 +1124,6 @@ DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNExecBaseParams *
 {
     TFModel *tf_model = model->model;
     TFContext *ctx = &tf_model->ctx;
-    TaskItem task;
-    TFRequestItem *request;
-
-    if (ff_check_exec_params(ctx, DNN_TF, model->func_type, exec_params) != 0) {
-        return DNN_ERROR;
-    }
-
-    if (ff_dnn_fill_task(&task, exec_params, tf_model, 0, 1) != DNN_SUCCESS) {
-        return DNN_ERROR;
-    }
-
-    if (extract_inference_from_task(&task, tf_model->inference_queue) != DNN_SUCCESS) {
-        av_log(ctx, AV_LOG_ERROR, "unable to extract inference from task.\n");
-        return DNN_ERROR;
-    }
-
-    request = ff_safe_queue_pop_front(tf_model->request_queue);
-    if (!request) {
-        av_log(ctx, AV_LOG_ERROR, "unable to get infer request.\n");
-        return DNN_ERROR;
-    }
-
-    return execute_model_tf(request, tf_model->inference_queue);
-}
-
-DNNReturnType ff_dnn_execute_model_async_tf(const DNNModel *model, DNNExecBaseParams *exec_params) {
-    TFModel *tf_model = model->model;
-    TFContext *ctx = &tf_model->ctx;
     TaskItem *task;
     TFRequestItem *request;
 
@@ -1164,7 +1137,7 @@ DNNReturnType ff_dnn_execute_model_async_tf(const DNNModel *model, DNNExecBasePa
         return DNN_ERROR;
     }
 
-    if (ff_dnn_fill_task(task, exec_params, tf_model, 1, 1) != DNN_SUCCESS) {
+    if (ff_dnn_fill_task(task, exec_params, tf_model, ctx->options.async, 1) != DNN_SUCCESS) {
         av_freep(&task);
         return DNN_ERROR;
     }
@@ -1188,10 +1161,10 @@ DNNReturnType ff_dnn_execute_model_async_tf(const DNNModel *model, DNNExecBasePa
     return execute_model_tf(request, tf_model->inference_queue);
 }
 
-DNNAsyncStatusType ff_dnn_get_async_result_tf(const DNNModel *model, AVFrame **in, AVFrame **out)
+DNNAsyncStatusType ff_dnn_get_result_tf(const DNNModel *model, AVFrame **in, AVFrame **out)
 {
     TFModel *tf_model = model->model;
-    return ff_dnn_get_async_result_common(tf_model->task_queue, in, out);
+    return ff_dnn_get_result_common(tf_model->task_queue, in, out);
 }
 
 DNNReturnType ff_dnn_flush_tf(const DNNModel *model)
diff --git a/libavfilter/dnn/dnn_backend_tf.h b/libavfilter/dnn/dnn_backend_tf.h
index aec0fc2011..f14ea8c47a 100644
--- a/libavfilter/dnn/dnn_backend_tf.h
+++ b/libavfilter/dnn/dnn_backend_tf.h
@@ -32,8 +32,7 @@
 DNNModel *ff_dnn_load_model_tf(const char *model_filename, DNNFunctionType func_type, const char *options, AVFilterContext *filter_ctx);
 
 DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNExecBaseParams *exec_params);
-DNNReturnType ff_dnn_execute_model_async_tf(const DNNModel *model, DNNExecBaseParams *exec_params);
-DNNAsyncStatusType ff_dnn_get_async_result_tf(const DNNModel *model, AVFrame **in, AVFrame **out);
+DNNAsyncStatusType ff_dnn_get_result_tf(const DNNModel *model, AVFrame **in, AVFrame **out);
 DNNReturnType ff_dnn_flush_tf(const DNNModel *model);
 
 void ff_dnn_free_model_tf(DNNModel **model);
diff --git a/libavfilter/dnn/dnn_interface.c b/libavfilter/dnn/dnn_interface.c
index 81af934dd5..554a36b0dc 100644
--- a/libavfilter/dnn/dnn_interface.c
+++ b/libavfilter/dnn/dnn_interface.c
@@ -42,14 +42,15 @@ DNNModule *ff_get_dnn_module(DNNBackendType backend_type)
     case DNN_NATIVE:
         dnn_module->load_model = &ff_dnn_load_model_native;
         dnn_module->execute_model = &ff_dnn_execute_model_native;
+        dnn_module->get_result = &ff_dnn_get_result_native;
+        dnn_module->flush = &ff_dnn_flush_native;
         dnn_module->free_model = &ff_dnn_free_model_native;
         break;
     case DNN_TF:
     #if (CONFIG_LIBTENSORFLOW == 1)
         dnn_module->load_model = &ff_dnn_load_model_tf;
         dnn_module->execute_model = &ff_dnn_execute_model_tf;
-        dnn_module->execute_model_async = &ff_dnn_execute_model_async_tf;
-        dnn_module->get_async_result = &ff_dnn_get_async_result_tf;
+        dnn_module->get_result = &ff_dnn_get_result_tf;
         dnn_module->flush = &ff_dnn_flush_tf;
         dnn_module->free_model = &ff_dnn_free_model_tf;
     #else
@@ -61,8 +62,7 @@ DNNModule *ff_get_dnn_module(DNNBackendType backend_type)
     #if (CONFIG_LIBOPENVINO == 1)
         dnn_module->load_model = &ff_dnn_load_model_ov;
         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->get_result = &ff_dnn_get_result_ov;
         dnn_module->flush = &ff_dnn_flush_ov;
         dnn_module->free_model = &ff_dnn_free_model_ov;
     #else