| 44 | namespace |
| 45 | { |
| 46 | void fp16_sve_scale_nearest(const ITensor *src, |
| 47 | ITensor *dst, |
| 48 | const ITensor *offsets, |
| 49 | float sampling_offset, |
| 50 | bool align_corners, |
| 51 | const Window &window) |
| 52 | { |
| 53 | const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right; |
| 54 | const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom; |
| 55 | const size_t in_stride_wc = in_stride_w * in_stride_c; |
| 56 | const size_t in_dim_h = src->info()->dimension(2); |
| 57 | |
| 58 | // Compute the ratio between source height and destination height |
| 59 | const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners); |
| 60 | const auto window_start_x = static_cast<int32_t>(window.x().start()); |
| 61 | const auto window_end_x = static_cast<int32_t>(window.x().end()); |
| 62 | |
| 63 | Window win(window); |
| 64 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 65 | Iterator out(dst, win); |
| 66 | |
| 67 | const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes(); |
| 68 | const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3]; |
| 69 | |
| 70 | execute_window_loop( |
| 71 | win, |
| 72 | [&](const Coordinates &id) |
| 73 | { |
| 74 | const int32_t offset = |
| 75 | *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c; |
| 76 | const auto in_hi = static_cast<int>( |
| 77 | align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) |
| 78 | : std::floor((id.z() + sampling_offset) * hr)); |
| 79 | const int offset_row = in_hi * in_stride_wc; |
| 80 | const auto in_ptr = reinterpret_cast<const float16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]); |
| 81 | const auto out_ptr = reinterpret_cast<float16_t *>(out.ptr()); |
| 82 | |
| 83 | // Compute S elements per iteration |
| 84 | int x = window_start_x; |
| 85 | svbool_t pg = svwhilelt_b16(x, window_end_x); |
| 86 | do |
| 87 | { |
| 88 | // Store results |
| 89 | svst1_f16(pg, out_ptr + x, svld1_f16(pg, in_ptr + offset + offset_row + x)); |
| 90 | |
| 91 | x += svcntw(); |
| 92 | pg = svwhilelt_b16(x, window_end_x); |
| 93 | } while (svptest_any(svptrue_b16(), pg)); |
| 94 | }, |
| 95 | out); |
| 96 | } |
| 97 | } // namespace |
| 98 | namespace cpu |
| 99 | { |
no test coverage detected