FFmpeg
dnn_backend_native_layer_mathbinary.c
Go to the documentation of this file.
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"
29 
30 typedef float (*FunType)(float src0, float src1);
31 
32 static float sub(float src0, float src1)
33 {
34  return src0 - src1;
35 }
36 static float add(float src0, float src1)
37 {
38  return src0 + src1;
39 }
40 static float mul(float src0, float src1)
41 {
42  return src0 * src1;
43 }
44 static float realdiv(float src0, float src1)
45 {
46  return src0 / src1;
47 }
48 static float minimum(float src0, float src1)
49 {
50  return FFMIN(src0, src1);
51 }
52 static float floormod(float src0, float src1)
53 {
54  return (float)((int)(src0) % (int)(src1));
55 }
56 
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 }
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 }
101 int ff_dnn_load_layer_math_binary(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
102 {
104  int dnn_size = 0;
105  int input_index = 0;
106  params = av_malloc(sizeof(*params));
107  if (!params)
108  return 0;
109 
110  params->bin_op = (int32_t)avio_rl32(model_file_context);
111  dnn_size += 4;
112 
113  params->input0_broadcast = (int32_t)avio_rl32(model_file_context);
114  dnn_size += 4;
115  if (params->input0_broadcast) {
116  params->v = av_int2float(avio_rl32(model_file_context));
117  } else {
118  layer->input_operand_indexes[input_index] = (int32_t)avio_rl32(model_file_context);
119  if (layer->input_operand_indexes[input_index] >= operands_num) {
120  return 0;
121  }
122  input_index++;
123  }
124  dnn_size += 4;
125 
126  params->input1_broadcast = (int32_t)avio_rl32(model_file_context);
127  dnn_size += 4;
128  if (params->input1_broadcast) {
129  params->v = av_int2float(avio_rl32(model_file_context));
130  } else {
131  layer->input_operand_indexes[input_index] = (int32_t)avio_rl32(model_file_context);
132  if (layer->input_operand_indexes[input_index] >= operands_num) {
133  return 0;
134  }
135  input_index++;
136  }
137  dnn_size += 4;
138 
139  layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
140  dnn_size += 4;
141  layer->params = params;
142 
143  if (layer->output_operand_index >= operands_num) {
144  return 0;
145  }
146 
147  return dnn_size;
148 }
149 
150 int ff_dnn_execute_layer_math_binary(DnnOperand *operands, const int32_t *input_operand_indexes,
151  int32_t output_operand_index, const void *parameters, NativeContext *ctx)
152 {
153  const DnnOperand *input = &operands[input_operand_indexes[0]];
154  DnnOperand *output = &operands[output_operand_index];
155  const DnnLayerMathBinaryParams *params = (const DnnLayerMathBinaryParams *)parameters;
156 
157  for (int i = 0; i < 4; ++i)
158  output->dims[i] = input->dims[i];
159 
160  output->data_type = input->data_type;
161  output->length = ff_calculate_operand_data_length(output);
162  if (output->length <= 0) {
163  av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n");
164  return DNN_ERROR;
165  }
166  output->data = av_realloc(output->data, output->length);
167  if (!output->data) {
168  av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n");
169  return DNN_ERROR;
170  }
171 
172  switch (params->bin_op) {
173  case DMBO_SUB:
174  math_binary_not_commutative(sub, params, input, output, operands, input_operand_indexes);
175  return 0;
176  case DMBO_ADD:
177  math_binary_commutative(add, params, input, output, operands, input_operand_indexes);
178  return 0;
179  case DMBO_MUL:
180  math_binary_commutative(mul, params, input, output, operands, input_operand_indexes);
181  return 0;
182  case DMBO_REALDIV:
183  math_binary_not_commutative(realdiv, params, input, output, operands, input_operand_indexes);
184  return 0;
185  case DMBO_MINIMUM:
186  math_binary_commutative(minimum, params, input, output, operands, input_operand_indexes);
187  return 0;
188  case DMBO_FLOORMOD:
189  math_binary_not_commutative(floormod, params, input, output, operands, input_operand_indexes);
190  return 0;
191  default:
192  av_log(ctx, AV_LOG_ERROR, "Unmatch math binary operator\n");
193  return DNN_ERROR;
194  }
195 }
Bytestream IO Context.
Definition: avio.h:161
void * av_realloc(void *ptr, size_t size)
Allocate, reallocate, or free a block of memory.
Definition: mem.c:134
static float floormod(float src0, float src1)
static av_always_inline float av_int2float(uint32_t i)
Reinterpret a 32-bit integer as a float.
Definition: intfloat.h:40
DNN inference functions interface for native backend.
int32_t input_operand_indexes[4]
a layer can have multiple inputs and one output.
DNN inference functions interface for native backend.
#define av_malloc(s)
static float realdiv(float src0, float src1)
filter_frame For filters that do not use the this method is called when a frame is pushed to the filter s input It can be called at any time except in a reentrant way If the input frame is enough to produce output
DNNDataType data_type
support different kinds of data type such as float, half float, int8 etc, first support float now...
int32_t ff_calculate_operand_dims_count(const DnnOperand *oprd)
#define av_log(a,...)
#define src
Definition: vp8dsp.c:255
#define AV_LOG_ERROR
Something went wrong and cannot losslessly be recovered.
Definition: log.h:194
unsigned int avio_rl32(AVIOContext *s)
Definition: aviobuf.c:754
void * data
data pointer with data length in bytes.
GLenum GLint * params
Definition: opengl_enc.c:113
simple assert() macros that are a bit more flexible than ISO C assert().
int32_t dims[4]
there are two memory layouts, NHWC or NCHW, so we use dims, dims[0] is Number.
static void math_binary_not_commutative(FunType pfun, const DnnLayerMathBinaryParams *params, const DnnOperand *input, DnnOperand *output, DnnOperand *operands, const int32_t *input_operand_indexes)
static float mul(float src0, float src1)
#define FFMIN(a, b)
Definition: common.h:96
static float minimum(float src0, float src1)
int32_t
AVFormatContext * ctx
Definition: movenc.c:48
#define src1
Definition: h264pred.c:140
int ff_dnn_load_layer_math_binary(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
#define src0
Definition: h264pred.c:139
and forward the test the status of outputs and forward it to the corresponding return FFERROR_NOT_READY If the filters stores internally one or a few frame for some input
int
int ff_dnn_execute_layer_math_binary(DnnOperand *operands, const int32_t *input_operand_indexes, int32_t output_operand_index, const void *parameters, NativeContext *ctx)
static void math_binary_commutative(FunType pfun, const DnnLayerMathBinaryParams *params, const DnnOperand *input, DnnOperand *output, DnnOperand *operands, const int32_t *input_operand_indexes)
void * params
static float add(float src0, float src1)
static float sub(float src0, float src1)
float(* FunType)(float src0, float src1)
int32_t ff_calculate_operand_data_length(const DnnOperand *oprd)
int i
Definition: input.c:407
int32_t output_operand_index