From: Guo, Yejun <yejun.guo@intel.com>
Date: Fri, 11 Sep 2020 14:15:04 +0000 (+0800)
Subject: dnn: add a new interface DNNModel.get_output
X-Git-Url: https://git.sesse.net/?a=commitdiff_plain;h=e71d73b09652f4fc96e512a7d6d4c2ab41860f27;p=ffmpeg

dnn: add a new interface DNNModel.get_output

for some cases (for example, super resolution), the DNN model changes
the frame size which impacts the filter behavior, so the filter needs
to know the out frame size at very beginning.

Currently, the filter reuses DNNModule.execute_model to query the
out frame size, it is not clear from interface perspective, so add
a new explict interface DNNModel.get_output for such query.
---

diff --git a/libavfilter/dnn/dnn_backend_native.c b/libavfilter/dnn/dnn_backend_native.c
index dc47c9b5429..d45e211f0c8 100644
--- a/libavfilter/dnn/dnn_backend_native.c
+++ b/libavfilter/dnn/dnn_backend_native.c
@@ -44,6 +44,10 @@ const AVClass dnn_native_class = {
     .category   = AV_CLASS_CATEGORY_FILTER,
 };
 
+static DNNReturnType execute_model_native(const DNNModel *model, const char *input_name, AVFrame *in_frame,
+                                          const char **output_names, uint32_t nb_output, AVFrame *out_frame,
+                                          int do_ioproc);
+
 static DNNReturnType get_input_native(void *model, DNNData *input, const char *input_name)
 {
     NativeModel *native_model = (NativeModel *)model;
@@ -70,6 +74,25 @@ static DNNReturnType get_input_native(void *model, DNNData *input, const char *i
     return DNN_ERROR;
 }
 
+static DNNReturnType get_output_native(void *model, const char *input_name, int input_width, int input_height,
+                                       const char *output_name, int *output_width, int *output_height)
+{
+    DNNReturnType ret;
+    NativeModel *native_model = (NativeModel *)model;
+    AVFrame *in_frame = av_frame_alloc();
+    AVFrame *out_frame = av_frame_alloc();
+    in_frame->width = input_width;
+    in_frame->height = input_height;
+
+    ret = execute_model_native(native_model->model, input_name, in_frame, &output_name, 1, out_frame, 0);
+    *output_width = out_frame->width;
+    *output_height = out_frame->height;
+
+    av_frame_free(&out_frame);
+    av_frame_free(&in_frame);
+    return ret;
+}
+
 // Loads model and its parameters that are stored in a binary file with following structure:
 // layers_num,layer_type,layer_parameterss,layer_type,layer_parameters...
 // For CONV layer: activation_function, input_num, output_num, kernel_size, kernel, biases
@@ -216,6 +239,7 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename, const char *optio
     }
 
     model->get_input = &get_input_native;
+    model->get_output = &get_output_native;
     model->userdata = userdata;
 
     return model;
@@ -226,8 +250,9 @@ fail:
     return NULL;
 }
 
-DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, const char *input_name, AVFrame *in_frame,
-                                          const char **output_names, uint32_t nb_output, AVFrame *out_frame)
+static DNNReturnType execute_model_native(const DNNModel *model, const char *input_name, AVFrame *in_frame,
+                                          const char **output_names, uint32_t nb_output, AVFrame *out_frame,
+                                          int do_ioproc)
 {
     NativeModel *native_model = (NativeModel *)model->model;
     NativeContext *ctx = &native_model->ctx;
@@ -276,10 +301,12 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, const char *inp
     input.channels = oprd->dims[3];
     input.data = oprd->data;
     input.dt = oprd->data_type;
-    if (native_model->model->pre_proc != NULL) {
-        native_model->model->pre_proc(in_frame, &input, native_model->model->userdata);
-    } else {
-        proc_from_frame_to_dnn(in_frame, &input, ctx);
+    if (do_ioproc) {
+        if (native_model->model->pre_proc != NULL) {
+            native_model->model->pre_proc(in_frame, &input, native_model->model->userdata);
+        } else {
+            proc_from_frame_to_dnn(in_frame, &input, ctx);
+        }
     }
 
     if (nb_output != 1) {
@@ -322,21 +349,40 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, const char *inp
         output.channels = oprd->dims[3];
         output.dt = oprd->data_type;
 
-        if (out_frame->width != output.width || out_frame->height != output.height) {
-            out_frame->width = output.width;
-            out_frame->height = output.height;
-        } else {
+        if (do_ioproc) {
             if (native_model->model->post_proc != NULL) {
                 native_model->model->post_proc(out_frame, &output, native_model->model->userdata);
             } else {
                 proc_from_dnn_to_frame(out_frame, &output, ctx);
             }
+        } else {
+            out_frame->width = output.width;
+            out_frame->height = output.height;
         }
     }
 
     return DNN_SUCCESS;
 }
 
+DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, const char *input_name, AVFrame *in_frame,
+                                          const char **output_names, uint32_t nb_output, AVFrame *out_frame)
+{
+    NativeModel *native_model = (NativeModel *)model->model;
+    NativeContext *ctx = &native_model->ctx;
+
+    if (!in_frame) {
+        av_log(ctx, AV_LOG_ERROR, "in frame is NULL when execute model.\n");
+        return DNN_ERROR;
+    }
+
+    if (!out_frame) {
+        av_log(ctx, AV_LOG_ERROR, "out frame is NULL when execute model.\n");
+        return DNN_ERROR;
+    }
+
+    return execute_model_native(model, input_name, in_frame, output_names, nb_output, out_frame, 1);
+}
+
 int32_t calculate_operand_dims_count(const DnnOperand *oprd)
 {
     int32_t result = 1;
diff --git a/libavfilter/dnn/dnn_backend_openvino.c b/libavfilter/dnn/dnn_backend_openvino.c
index 0dba1c1adcf..495225d0b3d 100644
--- a/libavfilter/dnn/dnn_backend_openvino.c
+++ b/libavfilter/dnn/dnn_backend_openvino.c
@@ -63,6 +63,10 @@ static const AVOption dnn_openvino_options[] = {
 
 AVFILTER_DEFINE_CLASS(dnn_openvino);
 
+static DNNReturnType execute_model_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
+                                      const char **output_names, uint32_t nb_output, AVFrame *out_frame,
+                                      int do_ioproc);
+
 static DNNDataType precision_to_datatype(precision_e precision)
 {
     switch (precision)
@@ -132,6 +136,25 @@ static DNNReturnType get_input_ov(void *model, DNNData *input, const char *input
     return DNN_ERROR;
 }
 
+static DNNReturnType get_output_ov(void *model, const char *input_name, int input_width, int input_height,
+                                   const char *output_name, int *output_width, int *output_height)
+{
+    DNNReturnType ret;
+    OVModel *ov_model = (OVModel *)model;
+    AVFrame *in_frame = av_frame_alloc();
+    AVFrame *out_frame = av_frame_alloc();
+    in_frame->width = input_width;
+    in_frame->height = input_height;
+
+    ret = execute_model_ov(ov_model->model, input_name, in_frame, &output_name, 1, out_frame, 0);
+    *output_width = out_frame->width;
+    *output_height = out_frame->height;
+
+    av_frame_free(&out_frame);
+    av_frame_free(&in_frame);
+    return ret;
+}
+
 DNNModel *ff_dnn_load_model_ov(const char *model_filename, const char *options, void *userdata)
 {
     char *all_dev_names = NULL;
@@ -191,6 +214,7 @@ DNNModel *ff_dnn_load_model_ov(const char *model_filename, const char *options,
 
     model->model = (void *)ov_model;
     model->get_input = &get_input_ov;
+    model->get_output = &get_output_ov;
     model->options = options;
     model->userdata = userdata;
 
@@ -213,8 +237,9 @@ err:
     return NULL;
 }
 
-DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
-                                      const char **output_names, uint32_t nb_output, AVFrame *out_frame)
+static DNNReturnType execute_model_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
+                                      const char **output_names, uint32_t nb_output, AVFrame *out_frame,
+                                      int do_ioproc)
 {
     char *model_output_name = NULL;
     char *all_output_names = NULL;
@@ -252,10 +277,12 @@ DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
     input.channels = dims.dims[1];
     input.data = blob_buffer.buffer;
     input.dt = precision_to_datatype(precision);
-    if (ov_model->model->pre_proc != NULL) {
-        ov_model->model->pre_proc(in_frame, &input, ov_model->model->userdata);
-    } else {
-        proc_from_frame_to_dnn(in_frame, &input, ctx);
+    if (do_ioproc) {
+        if (ov_model->model->pre_proc != NULL) {
+            ov_model->model->pre_proc(in_frame, &input, ov_model->model->userdata);
+        } else {
+            proc_from_frame_to_dnn(in_frame, &input, ctx);
+        }
     }
     ie_blob_free(&input_blob);
 
@@ -308,15 +335,15 @@ DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
         output.width    = dims.dims[3];
         output.dt       = precision_to_datatype(precision);
         output.data     = blob_buffer.buffer;
-        if (out_frame->width != output.width || out_frame->height != output.height) {
-            out_frame->width = output.width;
-            out_frame->height = output.height;
-        } else {
+        if (do_ioproc) {
             if (ov_model->model->post_proc != NULL) {
                 ov_model->model->post_proc(out_frame, &output, ov_model->model->userdata);
             } else {
                 proc_from_dnn_to_frame(out_frame, &output, ctx);
             }
+        } else {
+            out_frame->width = output.width;
+            out_frame->height = output.height;
         }
         ie_blob_free(&output_blob);
     }
@@ -324,6 +351,25 @@ DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
     return DNN_SUCCESS;
 }
 
+DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_name, AVFrame *in_frame,
+                                      const char **output_names, uint32_t nb_output, AVFrame *out_frame)
+{
+    OVModel *ov_model = (OVModel *)model->model;
+    OVContext *ctx = &ov_model->ctx;
+
+    if (!in_frame) {
+        av_log(ctx, AV_LOG_ERROR, "in frame is NULL when execute model.\n");
+        return DNN_ERROR;
+    }
+
+    if (!out_frame) {
+        av_log(ctx, AV_LOG_ERROR, "out frame is NULL when execute model.\n");
+        return DNN_ERROR;
+    }
+
+    return execute_model_ov(model, input_name, in_frame, output_names, nb_output, out_frame, 1);
+}
+
 void ff_dnn_free_model_ov(DNNModel **model)
 {
     if (*model){
diff --git a/libavfilter/dnn/dnn_backend_tf.c b/libavfilter/dnn/dnn_backend_tf.c
index 8467f8a459b..be860b11b57 100644
--- a/libavfilter/dnn/dnn_backend_tf.c
+++ b/libavfilter/dnn/dnn_backend_tf.c
@@ -55,6 +55,10 @@ static const AVClass dnn_tensorflow_class = {
     .category   = AV_CLASS_CATEGORY_FILTER,
 };
 
+static DNNReturnType execute_model_tf(const DNNModel *model, const char *input_name, AVFrame *in_frame,
+                                      const char **output_names, uint32_t nb_output, AVFrame *out_frame,
+                                      int do_ioproc);
+
 static void free_buffer(void *data, size_t length)
 {
     av_freep(&data);
@@ -150,6 +154,25 @@ static DNNReturnType get_input_tf(void *model, DNNData *input, const char *input
     return DNN_SUCCESS;
 }
 
+static DNNReturnType get_output_tf(void *model, const char *input_name, int input_width, int input_height,
+                                   const char *output_name, int *output_width, int *output_height)
+{
+    DNNReturnType ret;
+    TFModel *tf_model = (TFModel *)model;
+    AVFrame *in_frame = av_frame_alloc();
+    AVFrame *out_frame = av_frame_alloc();
+    in_frame->width = input_width;
+    in_frame->height = input_height;
+
+    ret = execute_model_tf(tf_model->model, input_name, in_frame, &output_name, 1, out_frame, 0);
+    *output_width = out_frame->width;
+    *output_height = out_frame->height;
+
+    av_frame_free(&out_frame);
+    av_frame_free(&in_frame);
+    return ret;
+}
+
 static DNNReturnType load_tf_model(TFModel *tf_model, const char *model_filename)
 {
     TFContext *ctx = &tf_model->ctx;
@@ -583,14 +606,16 @@ DNNModel *ff_dnn_load_model_tf(const char *model_filename, const char *options,
 
     model->model = (void *)tf_model;
     model->get_input = &get_input_tf;
+    model->get_output = &get_output_tf;
     model->options = options;
     model->userdata = userdata;
 
     return model;
 }
 
-DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, const char *input_name, AVFrame *in_frame,
-                                      const char **output_names, uint32_t nb_output, AVFrame *out_frame)
+static DNNReturnType execute_model_tf(const DNNModel *model, const char *input_name, AVFrame *in_frame,
+                                      const char **output_names, uint32_t nb_output, AVFrame *out_frame,
+                                      int do_ioproc)
 {
     TF_Output *tf_outputs;
     TFModel *tf_model = (TFModel *)model->model;
@@ -618,10 +643,12 @@ DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, const char *input_n
     }
     input.data = (float *)TF_TensorData(input_tensor);
 
-    if (tf_model->model->pre_proc != NULL) {
-        tf_model->model->pre_proc(in_frame, &input, tf_model->model->userdata);
-    } else {
-        proc_from_frame_to_dnn(in_frame, &input, ctx);
+    if (do_ioproc) {
+        if (tf_model->model->pre_proc != NULL) {
+            tf_model->model->pre_proc(in_frame, &input, tf_model->model->userdata);
+        } else {
+            proc_from_frame_to_dnn(in_frame, &input, ctx);
+        }
     }
 
     if (nb_output != 1) {
@@ -673,15 +700,15 @@ DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, const char *input_n
         output.data = TF_TensorData(output_tensors[i]);
         output.dt = TF_TensorType(output_tensors[i]);
 
-        if (out_frame->width != output.width || out_frame->height != output.height) {
-            out_frame->width = output.width;
-            out_frame->height = output.height;
-        } else {
+        if (do_ioproc) {
             if (tf_model->model->post_proc != NULL) {
                 tf_model->model->post_proc(out_frame, &output, tf_model->model->userdata);
             } else {
                 proc_from_dnn_to_frame(out_frame, &output, ctx);
             }
+        } else {
+            out_frame->width = output.width;
+            out_frame->height = output.height;
         }
     }
 
@@ -696,6 +723,25 @@ DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, const char *input_n
     return DNN_SUCCESS;
 }
 
+DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, const char *input_name, AVFrame *in_frame,
+                                      const char **output_names, uint32_t nb_output, AVFrame *out_frame)
+{
+    TFModel *tf_model = (TFModel *)model->model;
+    TFContext *ctx = &tf_model->ctx;
+
+    if (!in_frame) {
+        av_log(ctx, AV_LOG_ERROR, "in frame is NULL when execute model.\n");
+        return DNN_ERROR;
+    }
+
+    if (!out_frame) {
+        av_log(ctx, AV_LOG_ERROR, "out frame is NULL when execute model.\n");
+        return DNN_ERROR;
+    }
+
+    return execute_model_tf(model, input_name, in_frame, output_names, nb_output, out_frame, 1);
+}
+
 void ff_dnn_free_model_tf(DNNModel **model)
 {
     TFModel *tf_model;
diff --git a/libavfilter/dnn_interface.h b/libavfilter/dnn_interface.h
index 0369ee4f716..2f129d535e5 100644
--- a/libavfilter/dnn_interface.h
+++ b/libavfilter/dnn_interface.h
@@ -51,6 +51,9 @@ typedef struct DNNModel{
     // Gets model input information
     // Just reuse struct DNNData here, actually the DNNData.data field is not needed.
     DNNReturnType (*get_input)(void *model, DNNData *input, const char *input_name);
+    // Gets model output width/height with given input w/h
+    DNNReturnType (*get_output)(void *model, const char *input_name, int input_width, int input_height,
+                                const char *output_name, int *output_width, int *output_height);
     // set the pre process to transfer data from AVFrame to DNNData
     // the default implementation within DNN is used if it is not provided by the filter
     int (*pre_proc)(AVFrame *frame_in, DNNData *model_input, void *user_data);
diff --git a/libavfilter/vf_dnn_processing.c b/libavfilter/vf_dnn_processing.c
index 2c8578c9b06..334243bd2b3 100644
--- a/libavfilter/vf_dnn_processing.c
+++ b/libavfilter/vf_dnn_processing.c
@@ -233,24 +233,15 @@ static int config_output(AVFilterLink *outlink)
     DnnProcessingContext *ctx = context->priv;
     DNNReturnType result;
     AVFilterLink *inlink = context->inputs[0];
-    AVFrame *out = NULL;
-
-    AVFrame *fake_in = ff_get_video_buffer(inlink, inlink->w, inlink->h);
 
     // have a try run in case that the dnn model resize the frame
-    out = ff_get_video_buffer(inlink, inlink->w, inlink->h);
-    result = (ctx->dnn_module->execute_model)(ctx->model, ctx->model_inputname, fake_in,
-                                              (const char **)&ctx->model_outputname, 1, out);
-    if (result != DNN_SUCCESS){
-        av_log(ctx, AV_LOG_ERROR, "failed to execute model\n");
+    result = ctx->model->get_output(ctx->model->model, ctx->model_inputname, inlink->w, inlink->h,
+                                    ctx->model_outputname, &outlink->w, &outlink->h);
+    if (result != DNN_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "could not get output from the model\n");
         return AVERROR(EIO);
     }
 
-    outlink->w = out->width;
-    outlink->h = out->height;
-
-    av_frame_free(&fake_in);
-    av_frame_free(&out);
     prepare_uv_scale(outlink);
 
     return 0;
diff --git a/libavfilter/vf_sr.c b/libavfilter/vf_sr.c
index 72a31372627..fe6c5d3c0da 100644
--- a/libavfilter/vf_sr.c
+++ b/libavfilter/vf_sr.c
@@ -111,23 +111,20 @@ static int config_output(AVFilterLink *outlink)
     SRContext *ctx = context->priv;
     DNNReturnType result;
     AVFilterLink *inlink = context->inputs[0];
-    AVFrame *out = NULL;
-    const char *model_output_name = "y";
+    int out_width, out_height;
 
     // have a try run in case that the dnn model resize the frame
-    AVFrame *fake_in = ff_get_video_buffer(inlink, inlink->w, inlink->h);
-    out = ff_get_video_buffer(inlink, inlink->w, inlink->h);
-    result = (ctx->dnn_module->execute_model)(ctx->model, "x", fake_in,
-                                              (const char **)&model_output_name, 1, out);
-    if (result != DNN_SUCCESS){
-        av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
+    result = ctx->model->get_output(ctx->model->model, "x", inlink->w, inlink->h,
+                                    "y", &out_width, &out_height);
+    if (result != DNN_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "could not get output from the model\n");
         return AVERROR(EIO);
     }
 
-    if (fake_in->width != out->width || fake_in->height != out->height) {
+    if (inlink->w != out_width || inlink->h != out_height) {
         //espcn
-        outlink->w = out->width;
-        outlink->h = out->height;
+        outlink->w = out_width;
+        outlink->h = out_height;
         if (inlink->format != AV_PIX_FMT_GRAY8){
             const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
             int sws_src_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
@@ -141,15 +138,13 @@ static int config_output(AVFilterLink *outlink)
         }
     } else {
         //srcnn
-        outlink->w = out->width * ctx->scale_factor;
-        outlink->h = out->height * ctx->scale_factor;
+        outlink->w = out_width * ctx->scale_factor;
+        outlink->h = out_height * ctx->scale_factor;
         ctx->sws_pre_scale = sws_getContext(inlink->w, inlink->h, inlink->format,
                                         outlink->w, outlink->h, outlink->format,
                                         SWS_BICUBIC, NULL, NULL, NULL);
     }
 
-    av_frame_free(&fake_in);
-    av_frame_free(&out);
     return 0;
 }