| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // SPDX-FileCopyrightText: 2024 - 2025 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 2 | // | ||
| 3 | // SPDX-License-Identifier: Apache-2.0 | ||
| 4 | |||
| 5 | #include <cassert> | ||
| 6 | #include <type_traits> | ||
| 7 | |||
| 8 | #include "kleidicv/kleidicv.h" | ||
| 9 | #include "kleidicv/neon.h" | ||
| 10 | #include "kleidicv/transform/remap.h" | ||
| 11 | |||
| 12 | namespace kleidicv::neon { | ||
| 13 | |||
| 14 | template <typename ScalarType> | ||
| 15 | class RemapS16Replicate { | ||
| 16 | public: | ||
| 17 | using MapVecTraits = neon::VecTraits<int16_t>; | ||
| 18 | using MapVectorType = typename MapVecTraits::VectorType; | ||
| 19 | using MapVector2Type = typename MapVecTraits::Vector2Type; | ||
| 20 | |||
| 21 | 268 | RemapS16Replicate(Rows<const ScalarType> src_rows, size_t src_width, | |
| 22 | size_t src_height) | ||
| 23 | 268 | : src_rows_{src_rows}, | |
| 24 | 536 | v_src_element_stride_{vdupq_n_u16( | |
| 25 | 268 | static_cast<uint16_t>(src_rows_.stride() / sizeof(ScalarType)))}, | |
| 26 | 268 | v_xmax_{vdupq_n_s16(static_cast<int16_t>(src_width - 1))}, | |
| 27 | 268 | v_ymax_{vdupq_n_s16(static_cast<int16_t>(src_height - 1))} {} | |
| 28 | |||
| 29 | 10788 | void transform_pixels(uint32x4_t indices_low, uint32x4_t indices_high, | |
| 30 | Columns<ScalarType> dst) { | ||
| 31 | // Copy pixels from source | ||
| 32 | 10788 | dst[0] = src_rows_[vgetq_lane_u32(indices_low, 0)]; | |
| 33 | 10788 | dst[1] = src_rows_[vgetq_lane_u32(indices_low, 1)]; | |
| 34 | 10788 | dst[2] = src_rows_[vgetq_lane_u32(indices_low, 2)]; | |
| 35 | 10788 | dst[3] = src_rows_[vgetq_lane_u32(indices_low, 3)]; | |
| 36 | |||
| 37 | 10788 | dst[4] = src_rows_[vgetq_lane_u32(indices_high, 0)]; | |
| 38 | 10788 | dst[5] = src_rows_[vgetq_lane_u32(indices_high, 1)]; | |
| 39 | 10788 | dst[6] = src_rows_[vgetq_lane_u32(indices_high, 2)]; | |
| 40 | 10788 | dst[7] = src_rows_[vgetq_lane_u32(indices_high, 3)]; | |
| 41 | 10788 | } | |
| 42 | |||
| 43 | 364 | void process_row(size_t width, Columns<const int16_t> mapxy, | |
| 44 | Columns<ScalarType> dst) { | ||
| 45 | 11152 | auto vector_path = [&](size_t step) { | |
| 46 | 10788 | MapVector2Type xy = vld2q_s16(&mapxy[0]); | |
| 47 | // Clamp coordinates to within the dimensions of the source image | ||
| 48 | 21576 | uint16x8_t x = vreinterpretq_u16_s16( | |
| 49 | 10788 | vmaxq_s16(vdupq_n_s16(0), vminq_s16(xy.val[0], v_xmax_))); | |
| 50 | 21576 | uint16x8_t y = vreinterpretq_u16_s16( | |
| 51 | 10788 | vmaxq_s16(vdupq_n_s16(0), vminq_s16(xy.val[1], v_ymax_))); | |
| 52 | // Calculate offsets from coordinates (y * stride/sizeof(ScalarType) + x) | ||
| 53 | 21576 | uint32x4_t indices_low = | |
| 54 | 21576 | vmlal_u16(vmovl_u16(vget_low_u16(x)), vget_low_u16(y), | |
| 55 | 10788 | vget_low_u16(v_src_element_stride_)); | |
| 56 | 21576 | uint32x4_t indices_high = | |
| 57 | 10788 | vmlal_high_u16(vmovl_high_u16(x), y, v_src_element_stride_); | |
| 58 | |||
| 59 | 10788 | transform_pixels(indices_low, indices_high, dst); | |
| 60 | |||
| 61 | 10788 | mapxy += ptrdiff_t(step); | |
| 62 | 10788 | dst += ptrdiff_t(step); | |
| 63 | 10788 | }; | |
| 64 | |||
| 65 | 364 | LoopUnroll loop{width, MapVecTraits::num_lanes()}; | |
| 66 | 364 | loop.unroll_once(vector_path); | |
| 67 | 728 | ptrdiff_t back_step = static_cast<ptrdiff_t>(loop.step()) - | |
| 68 | 364 | static_cast<ptrdiff_t>(loop.remaining_length()); | |
| 69 | 364 | mapxy -= back_step; | |
| 70 | 364 | dst -= back_step; | |
| 71 | 560 | loop.remaining([&](size_t, size_t step) { vector_path(step); }); | |
| 72 | 364 | } | |
| 73 | |||
| 74 | private: | ||
| 75 | Rows<const ScalarType> src_rows_; | ||
| 76 | uint16x8_t v_src_element_stride_; | ||
| 77 | int16x8_t v_xmax_; | ||
| 78 | int16x8_t v_ymax_; | ||
| 79 | }; // end of class RemapS16Replicate<ScalarType> | ||
| 80 | |||
| 81 | template <typename ScalarType> | ||
| 82 | class RemapS16ConstantBorder { | ||
| 83 | public: | ||
| 84 | using SrcVecTraits = neon::VecTraits<ScalarType>; | ||
| 85 | using SrcVecType = typename SrcVecTraits::VectorType; | ||
| 86 | |||
| 87 | using MapVecTraits = neon::VecTraits<int16_t>; | ||
| 88 | using MapVectorType = typename MapVecTraits::VectorType; | ||
| 89 | using MapVector2Type = typename MapVecTraits::Vector2Type; | ||
| 90 | |||
| 91 | 264 | RemapS16ConstantBorder(Rows<const ScalarType> src_rows, size_t src_width, | |
| 92 | size_t src_height, const ScalarType *border_value) | ||
| 93 | 264 | : src_rows_{src_rows}, | |
| 94 | 528 | v_src_element_stride_{vdupq_n_u16( | |
| 95 | 264 | static_cast<uint16_t>(src_rows_.stride() / sizeof(ScalarType)))}, | |
| 96 | 264 | v_width_{vdupq_n_u16(static_cast<uint16_t>(src_width))}, | |
| 97 | 264 | v_height_{vdupq_n_u16(static_cast<uint16_t>(src_height))}, | |
| 98 | 264 | v_border_{vdupq_n_u16(*border_value)} {} | |
| 99 | |||
| 100 | void transform_pixels(uint32x4_t indices_low, uint32x4_t indices_high, | ||
| 101 | uint16x8_t in_range, Columns<ScalarType> dst); | ||
| 102 | |||
| 103 | 360 | void process_row(size_t width, Columns<const int16_t> mapxy, | |
| 104 | Columns<ScalarType> dst) { | ||
| 105 | 11144 | auto vector_path = [&](size_t step) { | |
| 106 | 10784 | MapVector2Type xy = vld2q_s16(&mapxy[0]); | |
| 107 | |||
| 108 | 10784 | uint16x8_t x = vreinterpretq_u16_s16(xy.val[0]); | |
| 109 | 10784 | uint16x8_t y = vreinterpretq_u16_s16(xy.val[1]); | |
| 110 | |||
| 111 | // Find whether coordinates are within the image dimensions. | ||
| 112 | // Negative coordinates are interpreted as large values due to the | ||
| 113 | // s16->u16 reinterpretation. | ||
| 114 | 21568 | uint16x8_t in_range = | |
| 115 | 10784 | vandq_u16(vcltq_u16(x, v_width_), vcltq_u16(y, v_height_)); | |
| 116 | |||
| 117 | // Zero out-of-range coordinates. | ||
| 118 | 10784 | x = vandq_u16(in_range, x); | |
| 119 | 10784 | y = vandq_u16(in_range, y); | |
| 120 | |||
| 121 | // Calculate offsets from coordinates (y * stride/sizeof(ScalarType) + x) | ||
| 122 | 21568 | uint32x4_t indices_low = | |
| 123 | 21568 | vmlal_u16(vmovl_u16(vget_low_u16(x)), vget_low_u16(y), | |
| 124 | 10784 | vget_low_u16(v_src_element_stride_)); | |
| 125 | 21568 | uint32x4_t indices_high = | |
| 126 | 10784 | vmlal_high_u16(vmovl_high_u16(x), y, v_src_element_stride_); | |
| 127 | |||
| 128 | 10784 | transform_pixels(indices_low, indices_high, in_range, dst); | |
| 129 | |||
| 130 | 10784 | mapxy += ptrdiff_t(step); | |
| 131 | 10784 | dst += ptrdiff_t(step); | |
| 132 | 10784 | }; | |
| 133 | |||
| 134 | 360 | LoopUnroll loop{width, MapVecTraits::num_lanes()}; | |
| 135 | 360 | loop.unroll_once(vector_path); | |
| 136 | 720 | ptrdiff_t back_step = static_cast<ptrdiff_t>(loop.step()) - | |
| 137 | 360 | static_cast<ptrdiff_t>(loop.remaining_length()); | |
| 138 | 360 | mapxy -= back_step; | |
| 139 | 360 | dst -= back_step; | |
| 140 | 556 | loop.remaining([&](size_t, size_t step) { vector_path(step); }); | |
| 141 | 360 | } | |
| 142 | |||
| 143 | private: | ||
| 144 | Rows<const ScalarType> src_rows_; | ||
| 145 | uint16x8_t v_src_element_stride_; | ||
| 146 | uint16x8_t v_width_; | ||
| 147 | uint16x8_t v_height_; | ||
| 148 | uint16x8_t v_border_; | ||
| 149 | }; // end of class RemapS16ConstantBorder<ScalarType> | ||
| 150 | |||
| 151 | template <> | ||
| 152 | 5632 | void RemapS16ConstantBorder<uint8_t>::transform_pixels(uint32x4_t indices_low, | |
| 153 | uint32x4_t indices_high, | ||
| 154 | uint16x8_t in_range, | ||
| 155 | Columns<uint8_t> dst) { | ||
| 156 | 50688 | uint8x8_t pixels = { | |
| 157 | 5632 | src_rows_[vgetq_lane_u32(indices_low, 0)], | |
| 158 | 5632 | src_rows_[vgetq_lane_u32(indices_low, 1)], | |
| 159 | 5632 | src_rows_[vgetq_lane_u32(indices_low, 2)], | |
| 160 | 5632 | src_rows_[vgetq_lane_u32(indices_low, 3)], | |
| 161 | 5632 | src_rows_[vgetq_lane_u32(indices_high, 0)], | |
| 162 | 5632 | src_rows_[vgetq_lane_u32(indices_high, 1)], | |
| 163 | 5632 | src_rows_[vgetq_lane_u32(indices_high, 2)], | |
| 164 | 5632 | src_rows_[vgetq_lane_u32(indices_high, 3)], | |
| 165 | }; | ||
| 166 | // Select between source pixels and border colour | ||
| 167 | 11264 | uint8x8_t pixels_or_border = | |
| 168 | 5632 | vbsl_u8(vmovn_u16(in_range), pixels, vmovn_u16(v_border_)); | |
| 169 | |||
| 170 | 5632 | vst1_u8(&dst[0], pixels_or_border); | |
| 171 | 5632 | } | |
| 172 | |||
| 173 | template <> | ||
| 174 | 5152 | void RemapS16ConstantBorder<uint16_t>::transform_pixels(uint32x4_t indices_low, | |
| 175 | uint32x4_t indices_high, | ||
| 176 | uint16x8_t in_range, | ||
| 177 | Columns<uint16_t> dst) { | ||
| 178 | 46368 | uint16x8_t pixels = { | |
| 179 | 5152 | src_rows_[vgetq_lane_u32(indices_low, 0)], | |
| 180 | 5152 | src_rows_[vgetq_lane_u32(indices_low, 1)], | |
| 181 | 5152 | src_rows_[vgetq_lane_u32(indices_low, 2)], | |
| 182 | 5152 | src_rows_[vgetq_lane_u32(indices_low, 3)], | |
| 183 | 5152 | src_rows_[vgetq_lane_u32(indices_high, 0)], | |
| 184 | 5152 | src_rows_[vgetq_lane_u32(indices_high, 1)], | |
| 185 | 5152 | src_rows_[vgetq_lane_u32(indices_high, 2)], | |
| 186 | 5152 | src_rows_[vgetq_lane_u32(indices_high, 3)], | |
| 187 | }; | ||
| 188 | |||
| 189 | // Select between source pixels and border colour | ||
| 190 | 5152 | uint16x8_t pixels_or_border = vbslq_u16(in_range, pixels, v_border_); | |
| 191 | |||
| 192 | 5152 | vst1q_u16(&dst[0], pixels_or_border); | |
| 193 | 5152 | } | |
| 194 | |||
| 195 | // Most of the complexity comes from parameter checking. | ||
| 196 | // NOLINTBEGIN(readability-function-cognitive-complexity) | ||
| 197 | template <typename T> | ||
| 198 | 600 | kleidicv_error_t remap_s16(const T *src, size_t src_stride, size_t src_width, | |
| 199 | size_t src_height, T *dst, size_t dst_stride, | ||
| 200 | size_t dst_width, size_t dst_height, size_t channels, | ||
| 201 | const int16_t *mapxy, size_t mapxy_stride, | ||
| 202 | kleidicv_border_type_t border_type, | ||
| 203 | [[maybe_unused]] const T *border_value) { | ||
| 204 |
8/8✓ Branch 0 taken 2 times.
✓ Branch 1 taken 298 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 298 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 298 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 298 times.
|
600 | CHECK_POINTER_AND_STRIDE(src, src_stride, src_height); |
| 205 |
8/8✓ Branch 0 taken 2 times.
✓ Branch 1 taken 296 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 296 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 296 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 296 times.
|
596 | CHECK_POINTER_AND_STRIDE(dst, dst_stride, dst_height); |
| 206 |
8/8✓ Branch 0 taken 2 times.
✓ Branch 1 taken 294 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 294 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 294 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 294 times.
|
592 | CHECK_POINTER_AND_STRIDE(mapxy, mapxy_stride, dst_height); |
| 207 |
12/12✓ Branch 0 taken 2 times.
✓ Branch 1 taken 292 times.
✓ Branch 2 taken 4 times.
✓ Branch 3 taken 288 times.
✓ Branch 4 taken 6 times.
✓ Branch 5 taken 288 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 292 times.
✓ Branch 8 taken 4 times.
✓ Branch 9 taken 288 times.
✓ Branch 10 taken 6 times.
✓ Branch 11 taken 288 times.
|
588 | CHECK_IMAGE_SIZE(src_width, src_height); |
| 208 |
12/12✓ Branch 0 taken 2 times.
✓ Branch 1 taken 286 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 284 times.
✓ Branch 4 taken 4 times.
✓ Branch 5 taken 284 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 286 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 284 times.
✓ Branch 10 taken 4 times.
✓ Branch 11 taken 284 times.
|
576 | CHECK_IMAGE_SIZE(dst_width, dst_height); |
| 209 |
8/8✓ Branch 0 taken 134 times.
✓ Branch 1 taken 150 times.
✓ Branch 2 taken 132 times.
✓ Branch 3 taken 2 times.
✓ Branch 4 taken 134 times.
✓ Branch 5 taken 150 times.
✓ Branch 6 taken 132 times.
✓ Branch 7 taken 2 times.
|
568 | if (border_type == KLEIDICV_BORDER_TYPE_CONSTANT && nullptr == border_value) { |
| 210 | 4 | return KLEIDICV_ERROR_NULL_POINTER; | |
| 211 | } | ||
| 212 | |||
| 213 |
8/8✓ Branch 0 taken 266 times.
✓ Branch 1 taken 16 times.
✓ Branch 2 taken 266 times.
✓ Branch 3 taken 16 times.
✓ Branch 4 taken 266 times.
✓ Branch 5 taken 16 times.
✓ Branch 6 taken 266 times.
✓ Branch 7 taken 16 times.
|
1128 | if (!remap_s16_is_implemented<T>(src_stride, src_width, src_height, dst_width, |
| 214 | 564 | border_type, channels)) { | |
| 215 | 32 | return KLEIDICV_ERROR_NOT_IMPLEMENTED; | |
| 216 | } | ||
| 217 | |||
| 218 | 532 | Rows<const T> src_rows{src, src_stride, channels}; | |
| 219 | 532 | Rows<const int16_t> mapxy_rows{mapxy, mapxy_stride, 2}; | |
| 220 | 532 | Rows<T> dst_rows{dst, dst_stride, channels}; | |
| 221 | 532 | Rectangle rect{dst_width, dst_height}; | |
| 222 |
4/4✓ Branch 0 taken 134 times.
✓ Branch 1 taken 132 times.
✓ Branch 2 taken 134 times.
✓ Branch 3 taken 132 times.
|
532 | if (border_type == KLEIDICV_BORDER_TYPE_CONSTANT) { |
| 223 | 528 | RemapS16ConstantBorder<T> operation{src_rows, src_width, src_height, | |
| 224 | 264 | border_value}; | |
| 225 | 264 | zip_rows(operation, rect, mapxy_rows, dst_rows); | |
| 226 | 264 | } else { | |
| 227 | assert(border_type == KLEIDICV_BORDER_TYPE_REPLICATE); | ||
| 228 | 268 | RemapS16Replicate<T> operation{src_rows, src_width, src_height}; | |
| 229 | 268 | zip_rows(operation, rect, mapxy_rows, dst_rows); | |
| 230 | 268 | } | |
| 231 | 532 | return KLEIDICV_OK; | |
| 232 | 600 | } | |
| 233 | // NOLINTEND(readability-function-cognitive-complexity) | ||
| 234 | |||
| 235 | #define KLEIDICV_INSTANTIATE_TEMPLATE_REMAP_S16(type) \ | ||
| 236 | template KLEIDICV_TARGET_FN_ATTRS kleidicv_error_t remap_s16<type>( \ | ||
| 237 | const type *src, size_t src_stride, size_t src_width, size_t src_height, \ | ||
| 238 | type *dst, size_t dst_stride, size_t dst_width, size_t dst_height, \ | ||
| 239 | size_t channels, const int16_t *mapxy, size_t mapxy_stride, \ | ||
| 240 | kleidicv_border_type_t border_type, const type *border_value) | ||
| 241 | |||
| 242 | KLEIDICV_INSTANTIATE_TEMPLATE_REMAP_S16(uint8_t); | ||
| 243 | KLEIDICV_INSTANTIATE_TEMPLATE_REMAP_S16(uint16_t); | ||
| 244 | |||
| 245 | } // namespace kleidicv::neon | ||
| 246 |