Moritz Pflanzer | 3ce3ff4 | 2017-07-21 17:41:02 +0100 | [diff] [blame] | 1 | /* |
Pablo Tello | 54e98d9 | 2019-02-05 16:16:19 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2019 ARM Limited. |
Moritz Pflanzer | 3ce3ff4 | 2017-07-21 17:41:02 +0100 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "DepthConcatenateLayer.h" |
| 25 | |
Moritz Pflanzer | a09de0c | 2017-09-01 20:41:12 +0100 | [diff] [blame] | 26 | #include "tests/validation/Helpers.h" |
Moritz Pflanzer | 3ce3ff4 | 2017-07-21 17:41:02 +0100 | [diff] [blame] | 27 | |
| 28 | namespace arm_compute |
| 29 | { |
| 30 | namespace test |
| 31 | { |
| 32 | namespace validation |
| 33 | { |
| 34 | namespace reference |
| 35 | { |
| 36 | template <typename T> |
Pablo Tello | 54e98d9 | 2019-02-05 16:16:19 +0000 | [diff] [blame] | 37 | SimpleTensor<T> depthconcatenate_layer(const std::vector<SimpleTensor<T>> &srcs, SimpleTensor<T> &dst) |
Moritz Pflanzer | 3ce3ff4 | 2017-07-21 17:41:02 +0100 | [diff] [blame] | 38 | { |
| 39 | // Create reference |
| 40 | std::vector<TensorShape> shapes; |
Michalis Spyrou | bcfd09a | 2019-05-01 13:03:59 +0100 | [diff] [blame] | 41 | shapes.reserve(srcs.size()); |
Moritz Pflanzer | 3ce3ff4 | 2017-07-21 17:41:02 +0100 | [diff] [blame] | 42 | for(const auto &src : srcs) |
| 43 | { |
| 44 | shapes.emplace_back(src.shape()); |
| 45 | } |
| 46 | |
Moritz Pflanzer | 3ce3ff4 | 2017-07-21 17:41:02 +0100 | [diff] [blame] | 47 | // Compute reference |
Georgios Pinitas | 8c8b748 | 2019-02-20 11:16:57 +0000 | [diff] [blame] | 48 | int depth_offset = 0; |
| 49 | const int width_out = dst.shape().x(); |
| 50 | const int height_out = dst.shape().y(); |
| 51 | const int depth_out = dst.shape().z(); |
| 52 | const int out_stride_z = width_out * height_out; |
| 53 | const int batches = dst.shape().total_size_upper(3); |
| 54 | auto have_different_quantization_info = [&](const SimpleTensor<T> &tensor) |
| 55 | { |
| 56 | return tensor.quantization_info() != dst.quantization_info(); |
| 57 | }; |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 58 | |
Georgios Pinitas | 8c8b748 | 2019-02-20 11:16:57 +0000 | [diff] [blame] | 59 | if(srcs[0].data_type() == DataType::QASYMM8 && std::any_of(srcs.cbegin(), srcs.cend(), have_different_quantization_info)) |
Pablo Tello | 661c252 | 2019-02-11 17:03:15 +0000 | [diff] [blame] | 60 | { |
Pablo Tello | 0dde667 | 2019-02-12 11:09:48 +0000 | [diff] [blame] | 61 | for(int b = 0; b < batches; ++b) |
Pablo Tello | 661c252 | 2019-02-11 17:03:15 +0000 | [diff] [blame] | 62 | { |
Pablo Tello | 0dde667 | 2019-02-12 11:09:48 +0000 | [diff] [blame] | 63 | // input tensors can have smaller width and height than the output, so for each output's slice we need to requantize 0 (as this is the value |
| 64 | // used in NEFillBorderKernel by NEDepthConcatenateLayer) using the corresponding quantization info for that particular slice/input tensor. |
| 65 | int slice = 0; |
| 66 | for(const auto &src : srcs) |
Pablo Tello | 661c252 | 2019-02-11 17:03:15 +0000 | [diff] [blame] | 67 | { |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 68 | auto ptr_slice = static_cast<T *>(dst(Coordinates(0, 0, slice, b))); |
| 69 | const auto num_elems_in_slice((dst.num_elements() / depth_out) * src.shape().z()); |
| 70 | const UniformQuantizationInfo iq_info = src.quantization_info().uniform(); |
| 71 | const UniformQuantizationInfo oq_info = dst.quantization_info().uniform(); |
| 72 | |
| 73 | std::transform(ptr_slice, ptr_slice + num_elems_in_slice, ptr_slice, [&](T) |
Pablo Tello | 0dde667 | 2019-02-12 11:09:48 +0000 | [diff] [blame] | 74 | { |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 75 | return quantize_qasymm8(dequantize_qasymm8(0, iq_info), oq_info); |
Pablo Tello | 0dde667 | 2019-02-12 11:09:48 +0000 | [diff] [blame] | 76 | }); |
| 77 | slice += src.shape().z(); |
| 78 | } |
Pablo Tello | 661c252 | 2019-02-11 17:03:15 +0000 | [diff] [blame] | 79 | } |
| 80 | } |
| 81 | else |
| 82 | { |
| 83 | std::fill_n(dst.data(), dst.num_elements(), 0); |
| 84 | } |
Moritz Pflanzer | 3ce3ff4 | 2017-07-21 17:41:02 +0100 | [diff] [blame] | 85 | |
| 86 | for(const auto &src : srcs) |
| 87 | { |
| 88 | ARM_COMPUTE_ERROR_ON(depth_offset >= depth_out); |
| 89 | ARM_COMPUTE_ERROR_ON(batches != static_cast<int>(src.shape().total_size_upper(3))); |
| 90 | |
| 91 | const int width = src.shape().x(); |
| 92 | const int height = src.shape().y(); |
| 93 | const int depth = src.shape().z(); |
| 94 | const int x_diff = (width_out - width) / 2; |
| 95 | const int y_diff = (height_out - height) / 2; |
| 96 | |
| 97 | const T *src_ptr = src.data(); |
| 98 | |
| 99 | for(int b = 0; b < batches; ++b) |
| 100 | { |
| 101 | const size_t offset_to_first_element = b * out_stride_z * depth_out + depth_offset * out_stride_z + y_diff * width_out + x_diff; |
| 102 | |
| 103 | for(int d = 0; d < depth; ++d) |
| 104 | { |
| 105 | for(int r = 0; r < height; ++r) |
| 106 | { |
Pablo Tello | 54e98d9 | 2019-02-05 16:16:19 +0000 | [diff] [blame] | 107 | if(src.data_type() == DataType::QASYMM8 && src.quantization_info() != dst.quantization_info()) |
| 108 | { |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 109 | const UniformQuantizationInfo iq_info = src.quantization_info().uniform(); |
| 110 | const UniformQuantizationInfo oq_info = dst.quantization_info().uniform(); |
| 111 | std::transform(src_ptr, src_ptr + width, dst.data() + offset_to_first_element + d * out_stride_z + r * width_out, [&](T t) |
Pablo Tello | 54e98d9 | 2019-02-05 16:16:19 +0000 | [diff] [blame] | 112 | { |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 113 | const float dequantized_input = dequantize_qasymm8(t, iq_info); |
| 114 | return quantize_qasymm8(dequantized_input, oq_info); |
Pablo Tello | 54e98d9 | 2019-02-05 16:16:19 +0000 | [diff] [blame] | 115 | }); |
| 116 | src_ptr += width; |
| 117 | } |
| 118 | else |
| 119 | { |
| 120 | std::copy(src_ptr, src_ptr + width, dst.data() + offset_to_first_element + d * out_stride_z + r * width_out); |
| 121 | src_ptr += width; |
| 122 | } |
Moritz Pflanzer | 3ce3ff4 | 2017-07-21 17:41:02 +0100 | [diff] [blame] | 123 | } |
| 124 | } |
| 125 | } |
| 126 | |
| 127 | depth_offset += depth; |
| 128 | } |
| 129 | |
| 130 | return dst; |
| 131 | } |
| 132 | |
Pablo Tello | 54e98d9 | 2019-02-05 16:16:19 +0000 | [diff] [blame] | 133 | template SimpleTensor<uint8_t> depthconcatenate_layer(const std::vector<SimpleTensor<uint8_t>> &srcs, SimpleTensor<uint8_t> &dst); |
| 134 | template SimpleTensor<float> depthconcatenate_layer(const std::vector<SimpleTensor<float>> &srcs, SimpleTensor<float> &dst); |
| 135 | template SimpleTensor<half> depthconcatenate_layer(const std::vector<SimpleTensor<half>> &srcs, SimpleTensor<half> &dst); |
Moritz Pflanzer | 3ce3ff4 | 2017-07-21 17:41:02 +0100 | [diff] [blame] | 136 | } // namespace reference |
| 137 | } // namespace validation |
| 138 | } // namespace test |
| 139 | } // namespace arm_compute |