Pablo Marquez Tello | 68b6dce | 2023-10-05 11:28:15 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2023 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 | #ifndef ACL_SRC_CPU_KERNELS_POOL2D_NEON_IMPL_H |
| 26 | #define ACL_SRC_CPU_KERNELS_POOL2D_NEON_IMPL_H |
| 27 | |
| 28 | #include "arm_compute/core/Helpers.h" |
| 29 | #include "arm_compute/core/ITensor.h" |
| 30 | #include "arm_compute/core/Types.h" |
| 31 | #include "arm_compute/core/utils/misc/Traits.h" |
| 32 | |
| 33 | #include "src/core/helpers/WindowHelpers.h" |
| 34 | #include "src/core/NEON/wrapper/intrinsics/intrinsics.h" |
| 35 | #include "src/cpu/kernels/pool2d/neon/list.h" |
| 36 | |
| 37 | #include <limits> |
| 38 | |
| 39 | #ifdef ENABLE_NCHW_KERNELS |
| 40 | namespace arm_compute |
| 41 | { |
| 42 | namespace cpu |
| 43 | { |
| 44 | |
| 45 | namespace |
| 46 | { |
| 47 | template <typename T> |
| 48 | auto read_2_boundary_aware_as_f32(int srcw, int srch, int pad_l, int pad_t, int x, int y, const T *ptr, T fval) |
| 49 | { |
| 50 | T vec[2]; |
| 51 | const bool row_in_bounds((y >= pad_t) && (y < (srch + pad_t))); |
| 52 | for (int i = 0; i < 2; i++) |
| 53 | { |
| 54 | if (row_in_bounds && (x + i >= pad_l) && (x + i < (srcw + pad_l))) |
| 55 | { |
| 56 | vec[i] = *(ptr + i); |
| 57 | } |
| 58 | else |
| 59 | { |
| 60 | vec[i] = fval; |
| 61 | } |
| 62 | } |
| 63 | float32_t vec_f32[2] = {vec[0], vec[1]}; |
| 64 | return wrapper::vload(vec_f32); |
| 65 | } |
| 66 | } // namespace |
| 67 | |
| 68 | template <typename T> |
| 69 | void pooling2_nchw_maxpool_indices(const ITensor *src, |
| 70 | ITensor *dst0, |
| 71 | ITensor *dst1, |
| 72 | PoolingLayerInfo &pool_info, |
| 73 | const Window &window_src, |
| 74 | const Window &window) |
| 75 | { |
| 76 | Iterator in(src, window_src); |
| 77 | Iterator out(dst0, window); |
| 78 | Iterator indices(dst1, window); |
| 79 | const int pool_pad_top = pool_info.pad_stride_info.pad_top(); |
| 80 | const int pool_pad_left = pool_info.pad_stride_info.pad_left(); |
| 81 | int pool_stride_x = 0; |
| 82 | int pool_stride_y = 0; |
| 83 | std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride(); |
| 84 | const int src_w = src->info()->dimension(0); |
| 85 | const int src_h = src->info()->dimension(1); |
| 86 | const uint8_t *const src_top_ptr = |
| 87 | src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top))); |
| 88 | const uint8_t *const src_bottom_ptr = |
| 89 | src->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_left), -static_cast<int>(pool_pad_top) + 1)); |
| 90 | const int pad_left = src->info()->padding().left; |
| 91 | const int pad_right = src->info()->padding().right; |
| 92 | const int in_stride_y = static_cast<int>(src->info()->strides_in_bytes().y()); |
| 93 | const T float_min = get_initial_min<T>(pool_info.use_inf_as_limit); |
| 94 | const T fill_value = (pool_info.pool_type == PoolingType::MAX) ? float_min : 0.f; |
| 95 | |
| 96 | execute_window_loop( |
| 97 | window, |
| 98 | [&](const Coordinates &id) |
| 99 | { |
| 100 | const auto x_val = id.x() * pool_stride_x; |
| 101 | const auto y_val_0 = id.y() * pool_stride_y; |
| 102 | const auto y_val_1 = (id.y() * pool_stride_y) + 1; |
| 103 | auto top_data = |
| 104 | read_2_boundary_aware_as_f32(src_w, src_h, pool_pad_left, pool_pad_top, x_val, y_val_0, |
| 105 | reinterpret_cast<const T *>(src_top_ptr + in.offset()), fill_value); |
| 106 | auto bottom_data = |
| 107 | read_2_boundary_aware_as_f32(src_w, src_h, pool_pad_left, pool_pad_top, x_val, y_val_1, |
| 108 | reinterpret_cast<const T *>(src_bottom_ptr + in.offset()), fill_value); |
| 109 | |
| 110 | // Calculate max data, compare top first, then bottom, to make sue the first max is recorded. |
| 111 | const float32x2_t max_data_top = vpmax_f32(top_data, top_data); |
| 112 | const float32x2_t max_data_bottom = vpmax_f32(bottom_data, bottom_data); |
| 113 | const float32x2_t max_data = vmax_f32(max_data_top, max_data_bottom); |
| 114 | *(reinterpret_cast<T *>(out.ptr())) = static_cast<T>(vget_lane_f32(max_data, 0)); |
| 115 | |
| 116 | // Calculate max data indice, which will be used in max unpool. |
| 117 | const uint32_t offset_base = |
| 118 | offset_no_padding<T>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y, DataLayout::NCHW); |
| 119 | const uint32_t offset_top = (uint32_t)(offset_base / sizeof(T)); |
| 120 | const uint32_t offset_bottom = offset_top + in_stride_y / sizeof(T) - pad_right - pad_left; |
| 121 | const uint32x2_t voffset_top = {offset_top, offset_top + 1u}; |
| 122 | const uint32x2_t voffset_bottom = {offset_bottom, offset_bottom + 1u}; |
| 123 | const uint32x2_t tmp_indices_top = |
| 124 | vbsl_u32(vcge_f32(top_data, vrev64_f32(top_data)), voffset_top, vrev64_u32(voffset_top)); |
| 125 | const uint32x2_t tmp_indices_bottom = |
| 126 | vbsl_u32(vcge_f32(bottom_data, vrev64_f32(bottom_data)), voffset_bottom, vrev64_u32(voffset_bottom)); |
| 127 | *(reinterpret_cast<int *>(indices.ptr())) = vget_lane_u32( |
| 128 | vbsl_u32(vcge_f32(max_data_top, max_data_bottom), tmp_indices_top, tmp_indices_bottom), 0); |
| 129 | }, |
| 130 | in, out, indices); |
| 131 | } |
| 132 | |
| 133 | } // namespace cpu |
| 134 | } // namespace arm_compute |
| 135 | |
| 136 | #endif // ENABLE_NCHW_KERNELS |
| 137 | |
| 138 | #endif // ACL_SRC_CPU_KERNELS_POOL2D_NEON_IMPL_H |