so, we can make a filter more general to accept different network models, by adding a data type convertion after getting data from network. After we add dt field into struct DNNData, it becomes the same as DNNInputData, so merge them with one struct: DNNData.
Showing the changes to libavfilter (8 of 8 files). Show all
Unified · Split
| @@ -28,7 +28,7 @@ · 27 unmodified lines | ||
| 28 | 28 | #include "dnn_backend_native_layer_conv2d.h" |
| 29 | 29 | #include "dnn_backend_native_layers.h" |
| 30 | 30 | |
| 31 | −static DNNReturnType set_input_output_native(void *model, DNNInputData *input, const char *input_name, const char **output_names, uint32_t nb_output) | |
| 31 | +static DNNReturnType set_input_output_native(void *model, DNNData *input, const char *input_name, const char **output_names, uint32_t nb_output) | |
| 32 | 32 | { |
| 33 | 33 | ConvolutionalNetwork *network = (ConvolutionalNetwork *)model; |
| 34 | 34 | DnnOperand *oprd = NULL; |
| @@ -263,6 +263,7 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *output · 228 unmodified lines | ||
| 263 | 263 | outputs[i].height = oprd->dims[1]; |
| 264 | 264 | outputs[i].width = oprd->dims[2]; |
| 265 | 265 | outputs[i].channels = oprd->dims[3]; |
| 266 | + outputs[i].dt = oprd->data_type; | |
| 266 | 267 | } |
| 267 | 268 | |
| 268 | 269 | return DNN_SUCCESS; |
| @@ -106,6 +106,7 @@ int dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_ · 105 unmodified lines | ||
| 106 | 106 | output_operand->dims[1] = height - pad_size * 2; |
| 107 | 107 | output_operand->dims[2] = width - pad_size * 2; |
| 108 | 108 | output_operand->dims[3] = conv_params->output_num; |
| 109 | + output_operand->data_type = operands[input_operand_index].data_type; | |
| 109 | 110 | output_operand->length = calculate_operand_data_length(output_operand); |
| 110 | 111 | output_operand->data = av_realloc(output_operand->data, output_operand->length); |
| 111 | 112 | if (!output_operand->data) |
| @@ -69,6 +69,7 @@ int dnn_execute_layer_depth2space(DnnOperand *operands, const int32_t *input_ope · 68 unmodified lines | ||
| 69 | 69 | output_operand->dims[1] = height * block_size; |
| 70 | 70 | output_operand->dims[2] = width * block_size; |
| 71 | 71 | output_operand->dims[3] = new_channels; |
| 72 | + output_operand->data_type = operands[input_operand_index].data_type; | |
| 72 | 73 | output_operand->length = calculate_operand_data_length(output_operand); |
| 73 | 74 | output_operand->data = av_realloc(output_operand->data, output_operand->length); |
| 74 | 75 | if (!output_operand->data) |
| @@ -105,6 +105,7 @@ int dnn_execute_layer_pad(DnnOperand *operands, const int32_t *input_operand_ind · 104 unmodified lines | ||
| 105 | 105 | output_operand->dims[1] = new_height; |
| 106 | 106 | output_operand->dims[2] = new_width; |
| 107 | 107 | output_operand->dims[3] = new_channel; |
| 108 | + output_operand->data_type = operands[input_operand_index].data_type; | |
| 108 | 109 | output_operand->length = calculate_operand_data_length(output_operand); |
| 109 | 110 | output_operand->data = av_realloc(output_operand->data, output_operand->length); |
| 110 | 111 | if (!output_operand->data) |
| @@ -83,7 +83,7 @@ static TF_Buffer *read_graph(const char *model_filename) · 82 unmodified lines | ||
| 83 | 83 | return graph_buf; |
| 84 | 84 | } |
| 85 | 85 | |
| 86 | −static TF_Tensor *allocate_input_tensor(const DNNInputData *input) | |
| 86 | +static TF_Tensor *allocate_input_tensor(const DNNData *input) | |
| 87 | 87 | { |
| 88 | 88 | TF_DataType dt; |
| 89 | 89 | size_t size; |
| @@ -105,7 +105,7 @@ static TF_Tensor *allocate_input_tensor(const DNNInputData *input) · 15 unmodified lines | ||
| 105 | 105 | input_dims[1] * input_dims[2] * input_dims[3] * size); |
| 106 | 106 | } |
| 107 | 107 | |
| 108 | −static DNNReturnType set_input_output_tf(void *model, DNNInputData *input, const char *input_name, const char **output_names, uint32_t nb_output) | |
| 108 | +static DNNReturnType set_input_output_tf(void *model, DNNData *input, const char *input_name, const char **output_names, uint32_t nb_output) | |
| 109 | 109 | { |
| 110 | 110 | TFModel *tf_model = (TFModel *)model; |
| 111 | 111 | TF_SessionOptions *sess_opts; |
| @@ -603,6 +603,7 @@ DNNReturnType ff_dnn_execute_model_tf(const DNNModel *model, DNNData *outputs, u · 491 unmodified lines | ||
| 603 | 603 | outputs[i].width = TF_Dim(tf_model->output_tensors[i], 2); |
| 604 | 604 | outputs[i].channels = TF_Dim(tf_model->output_tensors[i], 3); |
| 605 | 605 | outputs[i].data = TF_TensorData(tf_model->output_tensors[i]); |
| 606 | + outputs[i].dt = TF_TensorType(tf_model->output_tensors[i]); | |
| 606 | 607 | } |
| 607 | 608 | |
| 608 | 609 | return DNN_SUCCESS; |
| @@ -34,15 +34,10 @@ typedef enum {DNN_NATIVE, DNN_TF} DNNBackendType; · 33 unmodified lines | ||
| 34 | 34 | |
| 35 | 35 | typedef enum {DNN_FLOAT = 1, DNN_UINT8 = 4} DNNDataType; |
| 36 | 36 | |
| 37 | −typedef struct DNNInputData{ | |
| 37 | +typedef struct DNNData{ | |
| 38 | 38 | void *data; |
| 39 | 39 | DNNDataType dt; |
| 40 | 40 | int width, height, channels; |
| 41 | −} DNNInputData; | |
| 42 | − | |
| 43 | −typedef struct DNNData{ | |
| 44 | − float *data; | |
| 45 | − int width, height, channels; | |
| 46 | 41 | } DNNData; |
| 47 | 42 | |
| 48 | 43 | typedef struct DNNModel{ |
| @@ -50,7 +45,7 @@ typedef struct DNNModel{ · 1 unmodified line | ||
| 50 | 45 | void *model; |
| 51 | 46 | // Sets model input and output. |
| 52 | 47 | // Should be called at least once before model execution. |
| 53 | − DNNReturnType (*set_input_output)(void *model, DNNInputData *input, const char *input_name, const char **output_names, uint32_t nb_output); | |
| 48 | + DNNReturnType (*set_input_output)(void *model, DNNData *input, const char *input_name, const char **output_names, uint32_t nb_output); | |
| 54 | 49 | } DNNModel; |
| 55 | 50 | |
| 56 | 51 | // Stores pointers to functions for loading, executing, freeing DNN models for one of the backends. |
| @@ -39,7 +39,7 @@ typedef struct DRContext { · 38 unmodified lines | ||
| 39 | 39 | DNNBackendType backend_type; |
| 40 | 40 | DNNModule *dnn_module; |
| 41 | 41 | DNNModel *model; |
| 42 | − DNNInputData input; | |
| 42 | + DNNData input; | |
| 43 | 43 | DNNData output; |
| 44 | 44 | } DRContext; |
| 45 | 45 | |
| @@ -137,7 +137,7 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in) · 91 unmodified lines | ||
| 137 | 137 | int t = i * out->width * 3 + j; |
| 138 | 138 | |
| 139 | 139 | int t_in = (i + pad_size) * in->width * 3 + j + pad_size * 3; |
| 140 | − out->data[0][k] = CLIP((int)((((float *)dr_context->input.data)[t_in] - dr_context->output.data[t]) * 255), 0, 255); | |
| 140 | + out->data[0][k] = CLIP((int)((((float *)dr_context->input.data)[t_in] - ((float *)dr_context->output.data)[t]) * 255), 0, 255); | |
| 141 | 141 | } |
| 142 | 142 | } |
| 143 | 143 | |
| @@ -41,7 +41,7 @@ typedef struct SRContext { · 40 unmodified lines | ||
| 41 | 41 | DNNBackendType backend_type; |
| 42 | 42 | DNNModule *dnn_module; |
| 43 | 43 | DNNModel *model; |
| 44 | − DNNInputData input; | |
| 44 | + DNNData input; | |
| 45 | 45 | DNNData output; |
| 46 | 46 | int scale_factor; |
| 47 | 47 | struct SwsContext *sws_contexts[3]; |