39 const auto *w_data = w.
data();
40 const auto *b_data = b.
data();
42 auto *w_fused_data = w_fused.
data();
43 auto *b_fused_data = b_fused.
data();
45 const unsigned int width = w.
shape()[0];
46 const unsigned int height = w.
shape()[1];
47 const unsigned int dim2 = w.
shape()[2];
50 #pragma omp parallel for 52 for(
unsigned int b = 0; b < dim2; ++
b)
54 const auto mean_val = mean.
data()[
b];
55 const auto var_val = var.
data()[
b];
56 const auto beta_val = beta.
data()[
b];
57 const auto gamma_val = gamma.
data()[
b];
59 for(
unsigned int i = 0; i < width * height; ++i)
61 unsigned int index = i + b * width * height;
63 w_fused_data[index] = (gamma_val * (w_data[index])) / sqrt(var_val + epsilon);
66 b_fused_data[
b] = (b_data[
b] - mean_val) / sqrt(var_val + epsilon) * gamma_val + beta_val;
75 const auto *w_data = w.
data();
76 const auto *b_data = b.
data();
78 auto *w_fused_data = w_fused.
data();
79 auto *b_fused_data = b_fused.
data();
81 const unsigned int width = w.
shape()[0];
82 const unsigned int height = w.
shape()[1];
83 const unsigned int dim2 = w.
shape()[2];
84 const unsigned int dim3 = w.
shape()[3];
86 for(
unsigned int b = 0; b < dim3; ++
b)
88 const auto mean_val = mean.
data()[
b];
89 const auto var_val = var.
data()[
b];
90 const auto beta_val = beta.
data()[
b];
91 const auto gamma_val = gamma.
data()[
b];
93 for(
unsigned int i = 0; i < width * height * dim2; ++i)
95 unsigned int index = i + b * width * height * dim2;
97 w_fused_data[index] = (gamma_val * (w_data[index])) / sqrt(var_val + epsilon);
100 b_fused_data[
b] = (b_data[
b] - mean_val) / sqrt(var_val + epsilon) * gamma_val + beta_val;
void fuse_batch_normalization_dwc_layer(const SimpleTensor< T > &w, const SimpleTensor< T > &mean, const SimpleTensor< T > &var, SimpleTensor< T > &w_fused, SimpleTensor< T > &b_fused, const SimpleTensor< T > &b, const SimpleTensor< T > &beta, const SimpleTensor< T > &gamma, float epsilon)
TensorShape shape() const override
Shape of the tensor.
Copyright (c) 2017-2021 Arm Limited.
Simple tensor object that stores elements in a consecutive chunk of memory.
void fuse_batch_normalization_conv_layer(const SimpleTensor< T > &w, const SimpleTensor< T > &mean, const SimpleTensor< T > &var, SimpleTensor< T > &w_fused, SimpleTensor< T > &b_fused, const SimpleTensor< T > &b, const SimpleTensor< T > &beta, const SimpleTensor< T > &gamma, float epsilon)
const T * data() const
Constant pointer to the underlying buffer.