1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
|
/*
* Copyright (c) 2017-2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "ConvolutionLayer.h"
#include "tests/validation/Helpers.h"
namespace arm_compute
{
namespace test
{
namespace validation
{
namespace reference
{
template <typename T, typename TB>
SimpleTensor<T> deconvolution_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, const TensorShape &output_shape,
const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &a)
{
// Create reference
const int stride_x = info.stride().first;
const int stride_y = info.stride().second;
const int weights_width = weights.shape().x();
const int weights_height = weights.shape().y();
const int weights_upper_dims = weights.shape().total_size() / (weights_width * weights_height);
// Find the upsampled dimensions
unsigned int out_x = (src.shape().x() - 1) * stride_x + a.first + 1;
unsigned int out_y = (src.shape().y() - 1) * stride_y + a.second + 1;
// Find the padding needed for the convolution with stride 1 in order to match output shape
unsigned int padx = output_shape.x() - (out_x - weights_width + 1);
unsigned int pady = output_shape.y() - (out_y - weights_height + 1);
out_x += padx;
out_y += pady;
TensorShape scaled_shape = src.shape();
scaled_shape.set(0, out_x);
scaled_shape.set(1, out_y);
SimpleTensor<T> scaled{ scaled_shape, src.data_type(), 1, src.quantization_info() };
const int width_in = src.shape().x();
const int height_in = src.shape().y();
const int width_scaled = scaled.shape().x();
const int height_scaled = scaled.shape().y();
const int num_2d_slices = src.shape().total_size() / (width_in * height_in);
const int ax = a.first; // The number of zeros added to right edge of the input.
const int ay = a.second; // The number of zeros added to top edge of the input.
ARM_COMPUTE_ERROR_ON(info.pad().first > (weights.shape().x() - 1));
ARM_COMPUTE_ERROR_ON_MSG(ax > stride_x - 1, "ax must be smaller than stride_x");
ARM_COMPUTE_ERROR_ON_MSG(ay > stride_y - 1, "ay must be smaller than stride_y");
if(src.data_type() == DataType::QASYMM8)
{
const uint8_t quantized_zero = src.quantization_info().offset;
std::fill_n(scaled.data(), scaled.num_elements(), quantized_zero);
}
else
{
std::fill_n(scaled.data(), scaled.num_elements(), T(0));
}
// Flip weights by 180 degrees
SimpleTensor<T> weights_flipped{ weights.shape(), weights.data_type(), 1, weights.quantization_info() };
for(int ud = 0; ud < weights_upper_dims; ++ud)
{
const int offset = ud * weights_width * weights_height;
for(int y = 0; y < weights_height; ++y)
{
for(int x = 0; x < weights_width; ++x)
{
weights_flipped[offset + (weights_height - 1 - y) * weights_width + (weights_width - 1 - x)] = weights[offset + y * weights_width + x];
}
}
}
for(int slice = 0; slice < num_2d_slices; ++slice)
{
const int offset_slice_in = slice * width_in * height_in;
const int offset_slice_out = slice * width_scaled * height_scaled;
const int start_x = padx / 2;
const int start_y = ay + pady / 2;
const int end_y = height_scaled - pady / 2;
const int end_x = width_scaled - ax - padx / 2;
for(int yi = start_y, in_y = 0; yi < end_y; yi += stride_y, in_y++)
{
for(int xi = start_x, in_x = 0; xi < end_x; xi += stride_x, in_x++)
{
const T *in = src.data() + offset_slice_in + in_y * width_in + in_x;
T *out = scaled.data() + offset_slice_out + xi + yi * width_scaled;
*out = *in;
}
}
}
const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
return convolution_layer(scaled, weights_flipped, bias, output_shape, conv_info);
}
template SimpleTensor<uint8_t> deconvolution_layer(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &bias, const TensorShape &output_shape,
const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &a);
template SimpleTensor<float> deconvolution_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &bias, const TensorShape &output_shape,
const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &a);
template SimpleTensor<half> deconvolution_layer(const SimpleTensor<half> &src, const SimpleTensor<half> &weights, const SimpleTensor<half> &bias, const TensorShape &output_shape,
const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &a);
} // namespace reference
} // namespace validation
} // namespace test
} // namespace arm_compute
|
