| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // SPDX-FileCopyrightText: 2023 - 2024 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 2 | // | ||
| 3 | // SPDX-License-Identifier: Apache-2.0 | ||
| 4 | |||
| 5 | #include "kleidicv/filters/separable_filter_3x3_neon.h" | ||
| 6 | #include "kleidicv/filters/sobel.h" | ||
| 7 | #include "kleidicv/kleidicv.h" | ||
| 8 | #include "kleidicv/neon.h" | ||
| 9 | #include "kleidicv/workspace/separable.h" | ||
| 10 | |||
| 11 | namespace kleidicv::neon { | ||
| 12 | |||
| 13 | // Template for 3x3 Sobel filters which calculate horizontal derivative | ||
| 14 | // approximations, often denoted as Gx. | ||
| 15 | // | ||
| 16 | // The applied weights, as the kernel is mirrored both vertically and | ||
| 17 | // horizontally during the convolution: | ||
| 18 | // [ -1, 0, 1 ] [ 1 ] | ||
| 19 | // F = [ -2, 0, 2 ] = [ 2 ] * [ -1, 0, 1 ] | ||
| 20 | // [ -1, 0, 1 ] [ 1 ] | ||
| 21 | template <typename T> | ||
| 22 | class HorizontalSobel3x3; | ||
| 23 | |||
| 24 | // 3x3 Sobel filter for uint8_t types which calculates horizontal derivative | ||
| 25 | // approximations, often denoted as Gx. | ||
| 26 | template <> | ||
| 27 | class HorizontalSobel3x3<uint8_t> { | ||
| 28 | public: | ||
| 29 | using SourceType = uint8_t; | ||
| 30 | using BufferType = int16_t; | ||
| 31 | using DestinationType = int16_t; | ||
| 32 | |||
| 33 | // Applies vertical filtering vector using SIMD operations. | ||
| 34 | // | ||
| 35 | // DST = [ SRC0, SRC1, SRC2 ] * [ 1, 2, 1 ]T | ||
| 36 | 1494 | void vertical_vector_path(uint8x16_t src[3], BufferType *dst) const { | |
| 37 | 1494 | int16x8_t acc_l = vaddl_u8(vget_low_u8(src[0]), vget_low_u8(src[2])); | |
| 38 | 1494 | int16x8_t acc_h = vaddl_u8(vget_high_u8(src[0]), vget_high_u8(src[2])); | |
| 39 | 1494 | uint8x16_t shift_l = vshll_n_u8(vget_low_u8(src[1]), 1); | |
| 40 | 1494 | uint8x16_t shift_h = vshll_n_u8(vget_high_u8(src[1]), 1); | |
| 41 | 1494 | acc_l = vaddq_u16(acc_l, shift_l); | |
| 42 | 1494 | acc_h = vaddq_u16(acc_h, shift_h); | |
| 43 | 1494 | vst1q(&dst[0], acc_l); | |
| 44 | 1494 | vst1q(&dst[VecTraits<BufferType>::num_lanes()], acc_h); | |
| 45 | 1494 | } | |
| 46 | |||
| 47 | // Applies vertical filtering vector using scalar operations. | ||
| 48 | // | ||
| 49 | // DST = [ SRC0, SRC1, SRC2 ] * [ 1, 2, 1 ]T | ||
| 50 | 8556 | void vertical_scalar_path(const SourceType src[3], BufferType *dst) const { | |
| 51 | // Explicitly narrow. Overflow is permitted. | ||
| 52 | 8556 | dst[0] = static_cast<DestinationType>(src[0] + 2 * src[1] + src[2]); | |
| 53 | 8556 | } | |
| 54 | |||
| 55 | // Applies horizontal filtering vector using SIMD operations. | ||
| 56 | // | ||
| 57 | // DST = [ SRC0, SRC1, SRC2 ] * [ -1, 0, 1 ]T | ||
| 58 | 2560 | void horizontal_vector_path(int16x8_t src[3], DestinationType *dst) const { | |
| 59 | 2560 | vst1q(&dst[0], vsubq_s16(src[2], src[0])); | |
| 60 | 2560 | } | |
| 61 | |||
| 62 | // Applies horizontal filtering vector using scalar operations. | ||
| 63 | // | ||
| 64 | // DST = [ SRC0, SRC1, SRC2 ] * [ -1, 0, 1 ]T | ||
| 65 | 9456 | void horizontal_scalar_path(const BufferType src[3], | |
| 66 | DestinationType *dst) const { | ||
| 67 | // Explicitly narrow. Overflow is permitted. | ||
| 68 | 9456 | dst[0] = static_cast<DestinationType>(src[2] - src[0]); | |
| 69 | 9456 | } | |
| 70 | }; // end of class HorizontalSobel3x3<uint8_t> | ||
| 71 | |||
| 72 | // Template for 3x3 Sobel filters which calculate vertical derivative | ||
| 73 | // approximations, often denoted as Gy. | ||
| 74 | // | ||
| 75 | // The applied weights, as the kernel is mirrored both vertically and | ||
| 76 | // horizontally during the convolution: | ||
| 77 | // [ -1, -2, -1 ] [ -1 ] | ||
| 78 | // F = [ 0, 0, 0 ] = [ 0 ] * [ 1, 2, 1 ] | ||
| 79 | // [ 1, 2, 1 ] [ 1 ] | ||
| 80 | template <typename T> | ||
| 81 | class VerticalSobel3x3; | ||
| 82 | |||
| 83 | // 3x3 Sobel filter for uint8_t types which calculates vertical derivative | ||
| 84 | // approximations, often denoted as Gy. | ||
| 85 | template <> | ||
| 86 | class VerticalSobel3x3<uint8_t> { | ||
| 87 | public: | ||
| 88 | using SourceType = uint8_t; | ||
| 89 | using BufferType = int16_t; | ||
| 90 | using DestinationType = int16_t; | ||
| 91 | |||
| 92 | // Applies vertical filtering vector using SIMD operations. | ||
| 93 | // | ||
| 94 | // DST = [ SRC0, SRC1, SRC2 ] * [ -1, 0, 1 ]T | ||
| 95 | 1494 | void vertical_vector_path(uint8x16_t src[3], BufferType *dst) const { | |
| 96 | 1494 | uint16x8_t acc_l = vsubl_u8(vget_low_u8(src[2]), vget_low_u8(src[0])); | |
| 97 | 1494 | uint16x8_t acc_h = vsubl_u8(vget_high_u8(src[2]), vget_high_u8(src[0])); | |
| 98 | 1494 | vst1q(&dst[0], vreinterpretq_s16_u16(acc_l)); | |
| 99 | 2988 | vst1q(&dst[VecTraits<BufferType>::num_lanes()], | |
| 100 | 1494 | vreinterpretq_s16_u16(acc_h)); | |
| 101 | 1494 | } | |
| 102 | |||
| 103 | // Applies vertical filtering vector using scalar operations. | ||
| 104 | // | ||
| 105 | // DST = [ SRC0, SRC1, SRC2 ] * [ -1, 0, 1 ]T | ||
| 106 | 8556 | void vertical_scalar_path(const SourceType src[3], BufferType *dst) const { | |
| 107 | // Explicitly narrow. Overflow is permitted. | ||
| 108 | 8556 | dst[0] = static_cast<DestinationType>(src[2] - src[0]); | |
| 109 | 8556 | } | |
| 110 | |||
| 111 | // Applies horizontal filtering vector using SIMD operations. | ||
| 112 | // | ||
| 113 | // DST = [ SRC0, SRC1, SRC2 ] * [ 1, 2, 1 ]T | ||
| 114 | 2560 | void horizontal_vector_path(int16x8_t src[3], DestinationType *dst) const { | |
| 115 | 2560 | int16x8_t acc = vaddq_s16(src[0], src[2]); | |
| 116 | 2560 | acc = vaddq_s16(acc, vshlq_n_s16(src[1], 1)); | |
| 117 | 2560 | vst1q(&dst[0], acc); | |
| 118 | 2560 | } | |
| 119 | |||
| 120 | // Applies horizontal filtering vector using scalar operations. | ||
| 121 | // | ||
| 122 | // DST = [ SRC0, SRC1, SRC2 ] * [ 1, 2, 1 ]T | ||
| 123 | 9456 | void horizontal_scalar_path(const BufferType src[3], | |
| 124 | DestinationType *dst) const { | ||
| 125 | // Explicitly narrow. Overflow is permitted. | ||
| 126 | 9456 | dst[0] = static_cast<DestinationType>(src[0] + 2 * src[1] + src[2]); | |
| 127 | 9456 | } | |
| 128 | }; // end of class VerticalSobel3x3<uint8_t> | ||
| 129 | |||
| 130 | KLEIDICV_TARGET_FN_ATTRS | ||
| 131 | 90 | kleidicv_error_t sobel_3x3_horizontal_stripe_s16_u8( | |
| 132 | const uint8_t *src, size_t src_stride, int16_t *dst, size_t dst_stride, | ||
| 133 | size_t width, size_t height, size_t y_begin, size_t y_end, | ||
| 134 | size_t channels) { | ||
| 135 |
4/4✓ Branch 0 taken 1 times.
✓ Branch 1 taken 89 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 89 times.
|
90 | CHECK_POINTER_AND_STRIDE(src, src_stride, height); |
| 136 |
4/4✓ Branch 0 taken 2 times.
✓ Branch 1 taken 87 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 87 times.
|
89 | CHECK_POINTER_AND_STRIDE(dst, dst_stride, height); |
| 137 |
6/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 86 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 85 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 85 times.
|
87 | CHECK_IMAGE_SIZE(width, height); |
| 138 | |||
| 139 |
2/2✓ Branch 0 taken 1 times.
✓ Branch 1 taken 84 times.
|
85 | if (channels > KLEIDICV_MAXIMUM_CHANNEL_COUNT) { |
| 140 | 1 | return KLEIDICV_ERROR_NOT_IMPLEMENTED; | |
| 141 | } | ||
| 142 | |||
| 143 | 84 | Rectangle rect{width, height}; | |
| 144 | 84 | Rows<const uint8_t> src_rows{src, src_stride, channels}; | |
| 145 | 84 | Rows<int16_t> dst_rows{dst, dst_stride, channels}; | |
| 146 | |||
| 147 | 84 | auto workspace = | |
| 148 | 84 | SeparableFilterWorkspace::create(rect, channels, sizeof(int16_t)); | |
| 149 |
2/2✓ Branch 0 taken 83 times.
✓ Branch 1 taken 1 times.
|
84 | if (!workspace) { |
| 150 | 1 | return KLEIDICV_ERROR_ALLOCATION; | |
| 151 | } | ||
| 152 | |||
| 153 | 83 | HorizontalSobel3x3<uint8_t> horizontal_sobel; | |
| 154 | 83 | SeparableFilter3x3<HorizontalSobel3x3<uint8_t>> filter{horizontal_sobel}; | |
| 155 | 166 | workspace->process(rect, y_begin, y_end, src_rows, dst_rows, channels, | |
| 156 | 83 | FixedBorderType::REPLICATE, filter); | |
| 157 | 83 | return KLEIDICV_OK; | |
| 158 | 90 | } | |
| 159 | |||
| 160 | KLEIDICV_TARGET_FN_ATTRS | ||
| 161 | 90 | kleidicv_error_t sobel_3x3_vertical_stripe_s16_u8( | |
| 162 | const uint8_t *src, size_t src_stride, int16_t *dst, size_t dst_stride, | ||
| 163 | size_t width, size_t height, size_t y_begin, size_t y_end, | ||
| 164 | size_t channels) { | ||
| 165 |
4/4✓ Branch 0 taken 1 times.
✓ Branch 1 taken 89 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 89 times.
|
90 | CHECK_POINTER_AND_STRIDE(src, src_stride, height); |
| 166 |
4/4✓ Branch 0 taken 2 times.
✓ Branch 1 taken 87 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 87 times.
|
89 | CHECK_POINTER_AND_STRIDE(dst, dst_stride, height); |
| 167 |
6/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 86 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 85 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 85 times.
|
87 | CHECK_IMAGE_SIZE(width, height); |
| 168 | |||
| 169 |
2/2✓ Branch 0 taken 1 times.
✓ Branch 1 taken 84 times.
|
85 | if (channels > KLEIDICV_MAXIMUM_CHANNEL_COUNT) { |
| 170 | 1 | return KLEIDICV_ERROR_NOT_IMPLEMENTED; | |
| 171 | } | ||
| 172 | |||
| 173 | 84 | Rectangle rect{width, height}; | |
| 174 | 84 | Rows<const uint8_t> src_rows{src, src_stride, channels}; | |
| 175 | 84 | Rows<int16_t> dst_rows{dst, dst_stride, channels}; | |
| 176 | |||
| 177 | 84 | auto workspace = | |
| 178 | 84 | SeparableFilterWorkspace::create(rect, channels, sizeof(int16_t)); | |
| 179 |
2/2✓ Branch 0 taken 83 times.
✓ Branch 1 taken 1 times.
|
84 | if (!workspace) { |
| 180 | 1 | return KLEIDICV_ERROR_ALLOCATION; | |
| 181 | } | ||
| 182 | |||
| 183 | 83 | VerticalSobel3x3<uint8_t> vertical_sobel; | |
| 184 | 83 | SeparableFilter3x3<VerticalSobel3x3<uint8_t>> filter{vertical_sobel}; | |
| 185 | 166 | workspace->process(rect, y_begin, y_end, src_rows, dst_rows, channels, | |
| 186 | 83 | FixedBorderType::REPLICATE, filter); | |
| 187 | 83 | return KLEIDICV_OK; | |
| 188 | 90 | } | |
| 189 | |||
| 190 | } // namespace kleidicv::neon | ||
| 191 |