| 84 | |
| 85 | template <typename TIndex> |
| 86 | void NEGatherKernel::gather_common(const Window &window, const ThreadInfo &info) |
| 87 | { |
| 88 | ARM_COMPUTE_UNUSED(info); |
| 89 | |
| 90 | auto dst_win = window; |
| 91 | |
| 92 | const auto src_info = _input->info(); |
| 93 | const auto idx_info = _indices->info(); |
| 94 | const auto dst_info = _output->info(); |
| 95 | |
| 96 | const auto num_dims = dst_info->num_dimensions(); |
| 97 | const auto chunk_stride = src_info->strides_in_bytes()[_axis]; |
| 98 | |
| 99 | const auto window_start_x = window.x().start(); |
| 100 | const auto window_end_x = window.x().end(); |
| 101 | auto window_size_x = src_info->element_size(); |
| 102 | |
| 103 | const auto idx_limit = static_cast<TIndex>(src_info->tensor_shape()[_axis]); |
| 104 | |
| 105 | if (_axis != 0) |
| 106 | { |
| 107 | dst_win.set(0, Window::Dimension(window_start_x, window_start_x + 1, 1)); |
| 108 | window_size_x *= window_end_x - window_start_x; |
| 109 | } |
| 110 | |
| 111 | // Compute source and index tensors window based on the output window. |
| 112 | auto src_win = dst_win; |
| 113 | Window idx_win; |
| 114 | |
| 115 | for (size_t i = 0; i < idx_info->num_dimensions(); ++i) |
| 116 | { |
| 117 | src_win.set(_axis + i, Window::Dimension(0, 1, 1)); |
| 118 | idx_win.set(_axis + i, window[_axis + i]); |
| 119 | } |
| 120 | |
| 121 | // Use the custom strides to access all three tensors using the same loop. |
| 122 | Iterator src_it(num_dims, _src_it_strides, _input->buffer(), src_info->offset_first_element_in_bytes(), src_win); |
| 123 | Iterator idx_it(num_dims, _idx_it_strides, _indices->buffer(), idx_info->offset_first_element_in_bytes(), idx_win); |
| 124 | Iterator dst_it(num_dims, dst_info->strides_in_bytes(), _output->buffer(), |
| 125 | dst_info->offset_first_element_in_bytes(), dst_win); |
| 126 | |
| 127 | execute_window_loop( |
| 128 | dst_win, |
| 129 | [&](const Coordinates &) |
| 130 | { |
| 131 | const auto idx = *reinterpret_cast<const TIndex *>(idx_it.ptr()); |
| 132 | |
| 133 | if (idx >= 0 && idx < idx_limit) |
| 134 | { |
| 135 | const auto src_ptr = src_it.ptr() + idx * chunk_stride; |
| 136 | |
| 137 | std::copy_n(src_ptr, window_size_x, dst_it.ptr()); |
| 138 | } |
| 139 | else |
| 140 | { |
| 141 | std::fill_n(dst_it.ptr(), window_size_x, 0); |
| 142 | } |
| 143 | }, |
nothing calls this directly
no test coverage detected