1 /* 2 * Copyright (c) 2018 Sergey Lavrushkin 3 * 4 * This file is part of FFmpeg. 5 * 6 * FFmpeg is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU Lesser General Public 8 * License as published by the Free Software Foundation; either 9 * version 2.1 of the License, or (at your option) any later version. 10 * 11 * FFmpeg is distributed in the hope that it will be useful, 12 * but WITHOUT ANY WARRANTY; without even the implied warranty of 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 14 * Lesser General Public License for more details. 15 * 16 * You should have received a copy of the GNU Lesser General Public 17 * License along with FFmpeg; if not, write to the Free Software 18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 19 */ 20 21 /** 22 * @file 23 * DNN inference functions interface for native backend. 24 */ 25 26 27 #ifndef AVFILTER_DNN_DNN_BACKEND_NATIVE_H 28 #define AVFILTER_DNN_DNN_BACKEND_NATIVE_H 29 30 #include "../dnn_interface.h" 31 #include "libavformat/avio.h" 32 33 /** 34 * the enum value of DNNLayerType should not be changed, 35 * the same values are used in convert_from_tensorflow.py 36 * and, it is used to index the layer execution/load function pointer. 37 */ 38 typedef enum { 39 DLT_INPUT = 0, 40 DLT_CONV2D = 1, 41 DLT_DEPTH_TO_SPACE = 2, 42 DLT_MIRROR_PAD = 3, 43 DLT_MAXIMUM = 4, 44 DLT_MATH_BINARY = 5, 45 DLT_MATH_UNARY = 6, 46 DLT_COUNT 47 } DNNLayerType; 48 49 typedef enum {DOT_INPUT = 1, DOT_OUTPUT = 2, DOT_INTERMEDIATE = DOT_INPUT | DOT_OUTPUT} DNNOperandType; 50 51 typedef struct Layer{ 52 DNNLayerType type; 53 /** 54 * a layer can have multiple inputs and one output. 55 * 4 is just a big enough number for input operands (increase it if necessary), 56 * do not use 'int32_t *input_operand_indexes', so we don't worry about mem leaks. 57 */ 58 int32_t input_operand_indexes[4]; 59 int32_t output_operand_index; 60 void *params; 61 } Layer; 62 63 typedef struct DnnOperand{ 64 /** 65 * there are two memory layouts, NHWC or NCHW, so we use dims, 66 * dims[0] is Number. 67 */ 68 int32_t dims[4]; 69 70 /** 71 * input/output/intermediate operand of the network 72 */ 73 DNNOperandType type; 74 75 /** 76 * support different kinds of data type such as float, half float, int8 etc, 77 * first support float now. 78 */ 79 DNNDataType data_type; 80 81 /** 82 * NHWC if 1, otherwise NCHW. 83 * let's first support NHWC only, this flag is for extensive usage. 84 */ 85 int8_t isNHWC; 86 87 /** 88 * to avoid possible memory leak, do not use char *name 89 */ 90 char name[128]; 91 92 /** 93 * data pointer with data length in bytes. 94 * usedNumbersLeft is only valid for intermediate operand, 95 * it means how many layers still depend on this operand, 96 * todo: the memory can be reused when usedNumbersLeft is zero. 97 */ 98 void *data; 99 int32_t length; 100 int32_t usedNumbersLeft; 101 }DnnOperand; 102 103 typedef struct InputParams{ 104 int height, width, channels; 105 } InputParams; 106 107 // Represents simple feed-forward convolutional network. 108 typedef struct ConvolutionalNetwork{ 109 Layer *layers; 110 int32_t layers_num; 111 DnnOperand *operands; 112 int32_t operands_num; 113 int32_t *output_indexes; 114 uint32_t nb_output; 115 } ConvolutionalNetwork; 116 117 DNNModel *ff_dnn_load_model_native(const char *model_filename); 118 119 DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *outputs, uint32_t nb_output); 120 121 void ff_dnn_free_model_native(DNNModel **model); 122 123 // NOTE: User must check for error (return value <= 0) to handle 124 // case like integer overflow. 125 int32_t calculate_operand_data_length(const DnnOperand *oprd); 126 int32_t calculate_operand_dims_count(const DnnOperand *oprd); 127 #endif 128