1 /*
2 * Copyright (c) 2020
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 native backend implementation.
24 */
25
26 #include "dnn_backend_native.h"
27 #include "libavutil/avassert.h"
28 #include "dnn_backend_native_layer_mathbinary.h"
29
30 typedef float (*FunType)(float src0, float src1);
31
sub(float src0,float src1)32 static float sub(float src0, float src1)
33 {
34 return src0 - src1;
35 }
add(float src0,float src1)36 static float add(float src0, float src1)
37 {
38 return src0 + src1;
39 }
mul(float src0,float src1)40 static float mul(float src0, float src1)
41 {
42 return src0 * src1;
43 }
realdiv(float src0,float src1)44 static float realdiv(float src0, float src1)
45 {
46 return src0 / src1;
47 }
minimum(float src0,float src1)48 static float minimum(float src0, float src1)
49 {
50 return FFMIN(src0, src1);
51 }
floormod(float src0,float src1)52 static float floormod(float src0, float src1)
53 {
54 return (float)((int)(src0) % (int)(src1));
55 }
56
math_binary_commutative(FunType pfun,const DnnLayerMathBinaryParams * params,const DnnOperand * input,DnnOperand * output,DnnOperand * operands,const int32_t * input_operand_indexes)57 static void math_binary_commutative(FunType pfun, const DnnLayerMathBinaryParams *params, const DnnOperand *input, DnnOperand *output, DnnOperand *operands, const int32_t *input_operand_indexes)
58 {
59 int dims_count;
60 const float *src;
61 float *dst;
62 dims_count = ff_calculate_operand_dims_count(output);
63 src = input->data;
64 dst = output->data;
65 if (params->input0_broadcast || params->input1_broadcast) {
66 for (int i = 0; i < dims_count; ++i) {
67 dst[i] = pfun(params->v, src[i]);
68 }
69 } else {
70 const DnnOperand *input1 = &operands[input_operand_indexes[1]];
71 const float *src1 = input1->data;
72 for (int i = 0; i < dims_count; ++i) {
73 dst[i] = pfun(src[i], src1[i]);
74 }
75 }
76 }
math_binary_not_commutative(FunType pfun,const DnnLayerMathBinaryParams * params,const DnnOperand * input,DnnOperand * output,DnnOperand * operands,const int32_t * input_operand_indexes)77 static void math_binary_not_commutative(FunType pfun, const DnnLayerMathBinaryParams *params, const DnnOperand *input, DnnOperand *output, DnnOperand *operands, const int32_t *input_operand_indexes)
78 {
79 int dims_count;
80 const float *src;
81 float *dst;
82 dims_count = ff_calculate_operand_dims_count(output);
83 src = input->data;
84 dst = output->data;
85 if (params->input0_broadcast) {
86 for (int i = 0; i < dims_count; ++i) {
87 dst[i] = pfun(params->v, src[i]);
88 }
89 } else if (params->input1_broadcast) {
90 for (int i = 0; i < dims_count; ++i) {
91 dst[i] = pfun(src[i], params->v);
92 }
93 } else {
94 const DnnOperand *input1 = &operands[input_operand_indexes[1]];
95 const float *src1 = input1->data;
96 for (int i = 0; i < dims_count; ++i) {
97 dst[i] = pfun(src[i], src1[i]);
98 }
99 }
100 }
ff_dnn_load_layer_math_binary(Layer * layer,AVIOContext * model_file_context,int file_size,int operands_num)101 int ff_dnn_load_layer_math_binary(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
102 {
103 DnnLayerMathBinaryParams params = { 0 };
104 int dnn_size = 0;
105 int input_index = 0;
106
107 params.bin_op = (int32_t)avio_rl32(model_file_context);
108 dnn_size += 4;
109
110 params.input0_broadcast = (int32_t)avio_rl32(model_file_context);
111 dnn_size += 4;
112 if (params.input0_broadcast) {
113 params.v = av_int2float(avio_rl32(model_file_context));
114 } else {
115 layer->input_operand_indexes[input_index] = (int32_t)avio_rl32(model_file_context);
116 if (layer->input_operand_indexes[input_index] >= operands_num) {
117 return 0;
118 }
119 input_index++;
120 }
121 dnn_size += 4;
122
123 params.input1_broadcast = (int32_t)avio_rl32(model_file_context);
124 dnn_size += 4;
125 if (params.input1_broadcast) {
126 params.v = av_int2float(avio_rl32(model_file_context));
127 } else {
128 layer->input_operand_indexes[input_index] = (int32_t)avio_rl32(model_file_context);
129 if (layer->input_operand_indexes[input_index] >= operands_num) {
130 return 0;
131 }
132 input_index++;
133 }
134 dnn_size += 4;
135
136 layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
137 dnn_size += 4;
138
139 if (layer->output_operand_index >= operands_num) {
140 return 0;
141 }
142 layer->params = av_memdup(¶ms, sizeof(params));
143 if (!layer->params)
144 return 0;
145
146 return dnn_size;
147 }
148
ff_dnn_execute_layer_math_binary(DnnOperand * operands,const int32_t * input_operand_indexes,int32_t output_operand_index,const void * parameters,NativeContext * ctx)149 int ff_dnn_execute_layer_math_binary(DnnOperand *operands, const int32_t *input_operand_indexes,
150 int32_t output_operand_index, const void *parameters, NativeContext *ctx)
151 {
152 const DnnOperand *input = &operands[input_operand_indexes[0]];
153 DnnOperand *output = &operands[output_operand_index];
154 const DnnLayerMathBinaryParams *params = parameters;
155
156 for (int i = 0; i < 4; ++i)
157 output->dims[i] = input->dims[i];
158
159 output->data_type = input->data_type;
160 output->length = ff_calculate_operand_data_length(output);
161 if (output->length <= 0) {
162 av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n");
163 return DNN_ERROR;
164 }
165 output->data = av_realloc(output->data, output->length);
166 if (!output->data) {
167 av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n");
168 return DNN_ERROR;
169 }
170
171 switch (params->bin_op) {
172 case DMBO_SUB:
173 math_binary_not_commutative(sub, params, input, output, operands, input_operand_indexes);
174 return 0;
175 case DMBO_ADD:
176 math_binary_commutative(add, params, input, output, operands, input_operand_indexes);
177 return 0;
178 case DMBO_MUL:
179 math_binary_commutative(mul, params, input, output, operands, input_operand_indexes);
180 return 0;
181 case DMBO_REALDIV:
182 math_binary_not_commutative(realdiv, params, input, output, operands, input_operand_indexes);
183 return 0;
184 case DMBO_MINIMUM:
185 math_binary_commutative(minimum, params, input, output, operands, input_operand_indexes);
186 return 0;
187 case DMBO_FLOORMOD:
188 math_binary_not_commutative(floormod, params, input, output, operands, input_operand_indexes);
189 return 0;
190 default:
191 av_log(ctx, AV_LOG_ERROR, "Unmatch math binary operator\n");
192 return DNN_ERROR;
193 }
194 }
195