Suhail Munshi | ab84088 | 2021-02-09 16:31:00 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 Arm Limited. |
| 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 | |
| 25 | #include "ROIPoolingLayer.h" |
| 26 | #include "arm_compute/core/Types.h" |
| 27 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 28 | #include "tests/validation/Helpers.h" |
| 29 | #include <algorithm> |
| 30 | |
| 31 | namespace arm_compute |
| 32 | { |
| 33 | namespace test |
| 34 | { |
| 35 | namespace validation |
| 36 | { |
| 37 | namespace reference |
| 38 | { |
| 39 | template <> |
| 40 | SimpleTensor<float> roi_pool_layer(const SimpleTensor<float> &src, const SimpleTensor<uint16_t> &rois, const ROIPoolingLayerInfo &pool_info, const QuantizationInfo &output_qinfo) |
| 41 | { |
| 42 | ARM_COMPUTE_UNUSED(output_qinfo); |
| 43 | |
| 44 | const size_t num_rois = rois.shape()[1]; |
| 45 | const size_t values_per_roi = rois.shape()[0]; |
| 46 | DataType output_data_type = src.data_type(); |
| 47 | |
| 48 | TensorShape input_shape = src.shape(); |
| 49 | TensorShape output_shape(pool_info.pooled_width(), pool_info.pooled_height(), src.shape()[2], num_rois); |
| 50 | SimpleTensor<float> output(output_shape, output_data_type); |
| 51 | |
| 52 | const int pooled_w = pool_info.pooled_width(); |
| 53 | const int pooled_h = pool_info.pooled_height(); |
| 54 | const float spatial_scale = pool_info.spatial_scale(); |
| 55 | |
| 56 | // get sizes of x and y dimensions in src tensor |
| 57 | const int width = src.shape()[0]; |
| 58 | const int height = src.shape()[1]; |
| 59 | |
| 60 | // Move pointer across the fourth dimension |
| 61 | const size_t input_stride_w = input_shape[0] * input_shape[1] * input_shape[2]; |
| 62 | const size_t output_stride_w = output_shape[0] * output_shape[1] * output_shape[2]; |
| 63 | |
| 64 | const auto *rois_ptr = reinterpret_cast<const uint16_t *>(rois.data()); |
| 65 | |
| 66 | // Iterate through pixel width (X-Axis) |
| 67 | for(size_t pw = 0; pw < num_rois; ++pw) |
| 68 | { |
| 69 | const unsigned int roi_batch = rois_ptr[values_per_roi * pw]; |
| 70 | const auto x1 = rois_ptr[values_per_roi * pw + 1]; |
| 71 | const auto y1 = rois_ptr[values_per_roi * pw + 2]; |
| 72 | const auto x2 = rois_ptr[values_per_roi * pw + 3]; |
| 73 | const auto y2 = rois_ptr[values_per_roi * pw + 4]; |
| 74 | |
| 75 | //Iterate through pixel height (Y-Axis) |
| 76 | for(size_t fm = 0; fm < input_shape[2]; ++fm) |
| 77 | { |
| 78 | // Iterate through regions of interest index |
| 79 | for(size_t py = 0; py < pool_info.pooled_height(); ++py) |
| 80 | { |
| 81 | // Scale ROI |
| 82 | const int roi_anchor_x = support::cpp11::round(x1 * spatial_scale); |
| 83 | const int roi_anchor_y = support::cpp11::round(y1 * spatial_scale); |
| 84 | const int roi_width = std::max(support::cpp11::round((x2 - x1) * spatial_scale), 1.f); |
| 85 | const int roi_height = std::max(support::cpp11::round((y2 - y1) * spatial_scale), 1.f); |
| 86 | |
| 87 | // Iterate over feature map (Z axis) |
| 88 | for(size_t px = 0; px < pool_info.pooled_width(); ++px) |
| 89 | { |
| 90 | auto region_start_x = static_cast<int>(std::floor((static_cast<float>(px) / pooled_w) * roi_width)); |
| 91 | auto region_end_x = static_cast<int>(std::floor((static_cast<float>(px + 1) / pooled_w) * roi_width)); |
| 92 | auto region_start_y = static_cast<int>(std::floor((static_cast<float>(py) / pooled_h) * roi_height)); |
| 93 | auto region_end_y = static_cast<int>(std::floor((static_cast<float>(py + 1) / pooled_h) * roi_height)); |
| 94 | |
| 95 | region_start_x = std::min(std::max(region_start_x + roi_anchor_x, 0), width); |
| 96 | region_end_x = std::min(std::max(region_end_x + roi_anchor_x, 0), width); |
| 97 | region_start_y = std::min(std::max(region_start_y + roi_anchor_y, 0), height); |
| 98 | region_end_y = std::min(std::max(region_end_y + roi_anchor_y, 0), height); |
| 99 | |
| 100 | // Iterate through the pooling region |
| 101 | if((region_end_x <= region_start_x) || (region_end_y <= region_start_y)) |
| 102 | { |
| 103 | /* Assign element in tensor 'output' at coordinates px, py, fm, roi_indx, to 0 */ |
| 104 | auto out_ptr = output.data() + px + py * output_shape[0] + fm * output_shape[0] * output_shape[1] + pw * output_stride_w; |
| 105 | *out_ptr = 0; |
| 106 | } |
| 107 | else |
| 108 | { |
| 109 | float curr_max = -std::numeric_limits<float>::max(); |
| 110 | for(int j = region_start_y; j < region_end_y; ++j) |
| 111 | { |
| 112 | for(int i = region_start_x; i < region_end_x; ++i) |
| 113 | { |
| 114 | /* Retrieve element from input tensor at coordinates(i, j, fm, roi_batch) */ |
| 115 | float in_element = *(src.data() + i + j * input_shape[0] + fm * input_shape[0] * input_shape[1] + roi_batch * input_stride_w); |
| 116 | curr_max = std::max(in_element, curr_max); |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | /* Assign element in tensor 'output' at coordinates px, py, fm, roi_indx, to curr_max */ |
| 121 | auto out_ptr = output.data() + px + py * output_shape[0] + fm * output_shape[0] * output_shape[1] + pw * output_stride_w; |
| 122 | *out_ptr = curr_max; |
| 123 | } |
| 124 | } |
| 125 | } |
| 126 | } |
| 127 | } |
| 128 | |
| 129 | return output; |
| 130 | } |
| 131 | |
| 132 | /* |
| 133 | Template genericised method to allow calling of roi_pooling_layer with quantized 8 bit datatype |
| 134 | */ |
| 135 | template <> |
| 136 | SimpleTensor<uint8_t> roi_pool_layer(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint16_t> &rois, const ROIPoolingLayerInfo &pool_info, const QuantizationInfo &output_qinfo) |
| 137 | { |
| 138 | const SimpleTensor<float> src_tmp = convert_from_asymmetric(src); |
| 139 | SimpleTensor<float> dst_tmp = roi_pool_layer<float>(src_tmp, rois, pool_info, output_qinfo); |
| 140 | SimpleTensor<uint8_t> dst = convert_to_asymmetric<uint8_t>(dst_tmp, output_qinfo); |
| 141 | return dst; |
| 142 | } |
| 143 | |
| 144 | } // namespace reference |
| 145 | } // namespace validation |
| 146 | } // namespace test |
| 147 | } // namespace arm_compute |