1 /*
2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include <assert.h>
13 #include <math.h>
14
15 #include "aom_dsp/aom_dsp_common.h"
16 #include "av1/encoder/ml.h"
17
av1_nn_output_prec_reduce(float * const output,int num_output)18 void av1_nn_output_prec_reduce(float *const output, int num_output) {
19 const int prec_bits = 9;
20 const int prec = 1 << prec_bits;
21 const float inv_prec = (float)(1.0 / prec);
22 for (int i = 0; i < num_output; i++) {
23 output[i] = ((int)(output[i] * prec + 0.5)) * inv_prec;
24 }
25 }
26
27 // Calculate prediction based on the given input features and neural net config.
28 // Assume there are no more than NN_MAX_NODES_PER_LAYER nodes in each hidden
29 // layer.
av1_nn_predict_c(const float * input_nodes,const NN_CONFIG * const nn_config,int reduce_prec,float * const output)30 void av1_nn_predict_c(const float *input_nodes,
31 const NN_CONFIG *const nn_config, int reduce_prec,
32 float *const output) {
33 int num_input_nodes = nn_config->num_inputs;
34 int buf_index = 0;
35 float buf[2][NN_MAX_NODES_PER_LAYER];
36
37 // Propagate hidden layers.
38 const int num_layers = nn_config->num_hidden_layers;
39 assert(num_layers <= NN_MAX_HIDDEN_LAYERS);
40 for (int layer = 0; layer < num_layers; ++layer) {
41 const float *layer_weights = nn_config->weights[layer];
42 const float *layer_bias = nn_config->bias[layer];
43 float *output_nodes = buf[buf_index];
44 const int num_output_nodes = nn_config->num_hidden_nodes[layer];
45 assert(num_output_nodes < NN_MAX_NODES_PER_LAYER);
46 for (int node = 0; node < num_output_nodes; ++node) {
47 float val = layer_bias[node];
48 for (int i = 0; i < num_input_nodes; ++i)
49 val += layer_weights[node * num_input_nodes + i] * input_nodes[i];
50 // ReLU as activation function.
51 val = val > 0.0f ? val : 0.0f; // Could use AOMMAX().
52 output_nodes[node] = val;
53 }
54 num_input_nodes = num_output_nodes;
55 input_nodes = output_nodes;
56 buf_index = 1 - buf_index;
57 }
58
59 // Final output layer.
60 const float *layer_weights = nn_config->weights[num_layers];
61 const float *layer_bias = nn_config->bias[num_layers];
62 for (int node = 0; node < nn_config->num_outputs; ++node) {
63 float val = layer_bias[node];
64 for (int i = 0; i < num_input_nodes; ++i)
65 val += layer_weights[node * num_input_nodes + i] * input_nodes[i];
66 output[node] = val;
67 }
68 if (reduce_prec) av1_nn_output_prec_reduce(output, nn_config->num_outputs);
69 }
70
71 #if CONFIG_NN_V2
72 // Applies the ReLu activation to one fc layer
73 // output[i] = Max(input[i],0.0f)
nn_relu(const float * input,FC_LAYER * layer)74 static float *nn_relu(const float *input, FC_LAYER *layer) {
75 for (int i = 0; i < layer->num_outputs; ++i) {
76 layer->output[i] = AOMMAX(input[i], 0.0f);
77 }
78
79 return layer->output;
80 }
81
82 // Applies the Sigmoid activation to one fc layer
83 // output[i] = 1/(1+exp(input[i]))
nn_sigmoid(const float * input,FC_LAYER * layer)84 static float *nn_sigmoid(const float *input, FC_LAYER *layer) {
85 for (int i = 0; i < layer->num_outputs; ++i) {
86 const float tmp = AOMMIN(AOMMAX(input[i], -10.0f), 10.0f);
87 layer->output[i] = 1.0f / (1.0f + expf(-tmp));
88 }
89
90 return layer->output;
91 }
92
93 // Forward prediction in one fc layer, used in function av1_nn_predict_V2
nn_fc_forward(const float * input,FC_LAYER * layer)94 static float *nn_fc_forward(const float *input, FC_LAYER *layer) {
95 const float *weights = layer->weights;
96 const float *bias = layer->bias;
97 assert(layer->num_outputs < NN_MAX_NODES_PER_LAYER);
98 // fc
99 for (int node = 0; node < layer->num_outputs; ++node) {
100 float val = bias[node];
101 for (int i = 0; i < layer->num_inputs; ++i) val += weights[i] * input[i];
102 layer->output[node] = val;
103 weights += layer->num_inputs;
104 }
105
106 // activation
107 switch (layer->activation) {
108 case NONE: return layer->output;
109 case RELU: return nn_relu(layer->output, layer);
110 case SIGMOID: return nn_sigmoid(layer->output, layer);
111 case SOFTSIGN:
112 assert(0 && "Softsign has not been supported in NN."); // TO DO
113 return NULL;
114 default:
115 assert(0 && "Unknown activation"); // Unknown activation
116 return NULL;
117 }
118 }
119
av1_nn_predict_v2(const float * feature,NN_CONFIG_V2 * nn_config,int reduce_prec,float * output)120 void av1_nn_predict_v2(const float *feature, NN_CONFIG_V2 *nn_config,
121 int reduce_prec, float *output) {
122 const float *input_nodes = feature;
123
124 // Propagate the layers.
125 const int num_layers = nn_config->num_hidden_layers;
126 assert(num_layers <= NN_MAX_HIDDEN_LAYERS);
127 for (int i = 0; i < num_layers; ++i) {
128 input_nodes = nn_fc_forward(input_nodes, nn_config->layer + i);
129 assert(nn_config->layer[i + 1].num_inputs ==
130 nn_config->layer[i].num_outputs);
131 }
132
133 // Final layer
134 input_nodes = nn_fc_forward(input_nodes, nn_config->layer + num_layers);
135 assert(nn_config->layer[num_layers].num_outputs == nn_config->num_logits);
136 // Copy the final layer output
137 memcpy(output, input_nodes, sizeof(*input_nodes) * nn_config->num_logits);
138 if (reduce_prec) av1_nn_output_prec_reduce(output, nn_config->num_logits);
139 }
140 #endif // CONFIG_NN_V2
141
av1_nn_softmax(const float * input,float * output,int n)142 void av1_nn_softmax(const float *input, float *output, int n) {
143 // Softmax function is invariant to adding the same constant
144 // to all input values, so we subtract the maximum input to avoid
145 // possible overflow.
146 float max_input = input[0];
147 for (int i = 1; i < n; i++) max_input = AOMMAX(max_input, input[i]);
148 float sum_out = 0.0f;
149 for (int i = 0; i < n; i++) {
150 // Clamp to range [-10.0, 0.0] to prevent FE_UNDERFLOW errors.
151 const float normalized_input = AOMMAX(input[i] - max_input, -10.0f);
152 output[i] = expf(normalized_input);
153 sum_out += output[i];
154 }
155 for (int i = 0; i < n; i++) output[i] /= sum_out;
156 }
157
approx_exp(float y)158 static AOM_INLINE float approx_exp(float y) {
159 #define A ((1 << 23) / 0.69314718056f) // (1 << 23) / ln(2)
160 #define B \
161 127 // Offset for the exponent according to IEEE floating point standard.
162 #define C 60801 // Magic number controls the accuracy of approximation
163 union {
164 float as_float;
165 int32_t as_int32;
166 } container;
167 container.as_int32 = ((int32_t)(y * A)) + ((B << 23) - C);
168 return container.as_float;
169 #undef A
170 #undef B
171 #undef C
172 }
173
av1_nn_fast_softmax_16_c(const float * input,float * output)174 void av1_nn_fast_softmax_16_c(const float *input, float *output) {
175 const int kNumClasses = 16;
176 float max_input = input[0];
177 for (int i = 1; i < kNumClasses; i++) max_input = AOMMAX(max_input, input[i]);
178 float sum_out = 0.0f;
179 for (int i = 0; i < kNumClasses; i++) {
180 // Clamp to range [-10.0, 0.0] to prevent FE_UNDERFLOW errors.
181 const float normalized_input = AOMMAX(input[i] - max_input, -10.0f);
182 output[i] = approx_exp(normalized_input);
183 sum_out += output[i];
184 }
185 for (int i = 0; i < kNumClasses; i++) output[i] /= sum_out;
186 }
187