44 template <
typename T,
typename TW,
typename TB>
51 const int width_in =
src.shape().x();
52 const int height_in =
src.shape().y();
53 const int depth_in =
src.shape().z();
54 const int width_out =
dst.shape().x();
55 const int height_out =
dst.shape().y();
56 const int depth_out =
dst.shape().z();
57 const int width_weights = weights.
shape().x();
58 const int height_weights = weights.
shape().y();
59 const int depth_weights = weights.
shape().z();
60 const int pad_left =
info.pad_left();
61 const int pad_top =
info.pad_top();
62 const int stride_xi =
info.stride().first;
63 const int stride_yi =
info.stride().second;
67 const int start_xi = (dilation.
x() * (width_weights - 1) + 1) / 2 - pad_left;
68 const int start_yi = (dilation.
y() * (height_weights - 1) + 1) / 2 - pad_top;
69 const int end_xi =
output_wh.first * stride_xi;
70 const int end_yi =
output_wh.second * stride_yi;
71 const int num_batches =
src.shape().total_size() / (width_in * height_in * depth_in);
73 #if defined(_OPENMP) && !( defined(__arm__) && defined(__ANDROID__)) 74 #pragma omp parallel for collapse(5) 76 for(
int r = 0; r < num_batches; ++r)
78 for(
int yi = start_yi; yi < start_yi + end_yi; yi += stride_yi)
80 for(
int xi = start_xi; xi < start_xi + end_xi; xi += stride_xi)
82 for(
int group = 0; group < static_cast<int>(
num_groups); ++group)
84 for(
int ofm = 0; ofm < static_cast<int>(depth_out /
num_groups); ++ofm)
87 const int offset_in = r * width_in * height_in * depth_in + (group * (depth_in /
num_groups) * width_in * height_in);
88 const int xo = (xi - start_xi) / stride_xi;
89 const int yo = (yi - start_yi) / stride_yi;
90 const int offset_out = xo + yo * width_out + ((ofm + group * (depth_out /
num_groups)) * width_out * height_out) + (r * width_out * height_out * depth_out);
91 const int offset_w = (ofm + group * (depth_out /
num_groups)) * width_weights * height_weights * depth_weights;
92 const int offset_b = (ofm + group * (depth_out /
num_groups));
99 offset_in, offset_w, offset_b, offset_out,
102 width_weights, height_weights, dilation.
x(), dilation.
y(), ofm);
110 template <
typename T,
typename TW,
typename TB>
117 out_quant_info =
src.quantization_info();
#define ARM_COMPUTE_ASSERT(cond)
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
TensorShape shape() const override
Shape of the tensor.
size_t x() const
Semantic accessor for width as x.
SimpleTensor< float > src
Copyright (c) 2017-2021 Arm Limited.
std::pair< unsigned int, unsigned int > scaled_dimensions(int width, int height, int kernel_width, int kernel_height, const PadStrideInfo &pad_stride_info, const Size2D &dilation=Size2D(1U, 1U))
Returns expected width and height of output scaled tensor depending on dimensions rounding mode.
void convolution3d(const SimpleTensor< T > &in, const SimpleTensor< TW > &weights, const SimpleTensor< TB > &bias, SimpleTensor< T > &out, int i_offset, int w_offset, int b_offset, int o_offset, int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights, int dilation_x=1, int dilation_y=1, int filter_id=0)
Quantization information.
const unsigned int num_groups
Padding and stride information class.
size_t y() const
Semantic accessor for height as y.
Simple tensor object that stores elements in a consecutive chunk of memory.
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
SimpleTensor< T > convolution_layer_nchw(const SimpleTensor< T > &src, const SimpleTensor< TW > &weights, const SimpleTensor< TB > &bias, SimpleTensor< T > &dst, const PadStrideInfo &info, const Size2D &dilation, unsigned int num_groups)
SimpleTensor< T > convolution_layer(const SimpleTensor< T > &src, const SimpleTensor< TW > &weights, const SimpleTensor< TB > &bias, const TensorShape &output_shape, const PadStrideInfo &info, const Size2D &dilation, unsigned int num_groups, QuantizationInfo out_quant_info)
Class for specifying the size of an image or rectangle.