| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // SPDX-FileCopyrightText: 2024 - 2026 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 2 | // | ||
| 3 | // SPDX-License-Identifier: Apache-2.0 | ||
| 4 | |||
| 5 | #include <cassert> | ||
| 6 | |||
| 7 | #include "kleidicv/dispatch.h" | ||
| 8 | #include "kleidicv/kleidicv.h" | ||
| 9 | #include "kleidicv/resize/resize_linear.h" | ||
| 10 | #include "kleidicv/utils.h" | ||
| 11 | |||
| 12 | #define KLEIDICV_DEFINE_C_API_ALL(name, called_name) \ | ||
| 13 | KLEIDICV_MULTIVERSION_C_API( \ | ||
| 14 | name, &kleidicv::neon::called_name, \ | ||
| 15 | KLEIDICV_SVE2_IMPL_IF(kleidicv::sve2::called_name), \ | ||
| 16 | &kleidicv::sme::called_name, &kleidicv::sme2::called_name) | ||
| 17 | |||
| 18 | #define KLEIDICV_DEFINE_C_API_NEON(name, called_name) \ | ||
| 19 | KLEIDICV_MULTIVERSION_C_API(name, &kleidicv::neon::called_name, nullptr, \ | ||
| 20 | nullptr, nullptr) | ||
| 21 | |||
| 22 |
6/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 3 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 1 times.
|
10 | KLEIDICV_DEFINE_C_API_ALL(kleidicv_resize_2x2_stripe_u8, |
| 23 | kleidicv_resize_2x2_stripe_u8); | ||
| 24 |
6/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 3 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 1 times.
|
10 | KLEIDICV_DEFINE_C_API_ALL(kleidicv_resize_4x4_stripe_u8, |
| 25 | kleidicv_resize_4x4_stripe_u8); | ||
| 26 | |||
| 27 |
10/12✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 1 times.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 3 times.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 1 times.
✓ Branch 10 taken 1 times.
✓ Branch 11 taken 1 times.
|
10 | KLEIDICV_MULTIVERSION_C_API_VECLEN( |
| 28 | kleidicv_resize_1ch_r2_stripe_u8, | ||
| 29 | (&kleidicv::neon::kleidicv_resize_generic_stripe_u8<2, 1>), | ||
| 30 | (&kleidicv::sve2::kleidicv_resize_generic_stripe_u8<2, 1>), | ||
| 31 | (&kleidicv::sme::kleidicv_resize_generic_stripe_u8<2, 1>), | ||
| 32 | (&kleidicv::sme2::kleidicv_resize_generic_stripe_u8<2, 1>), 16, 64); | ||
| 33 |
10/12✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 1 times.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 3 times.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 1 times.
✓ Branch 10 taken 1 times.
✓ Branch 11 taken 1 times.
|
10 | KLEIDICV_MULTIVERSION_C_API_VECLEN( |
| 34 | kleidicv_resize_1ch_r3_stripe_u8, | ||
| 35 | (&kleidicv::neon::kleidicv_resize_generic_stripe_u8<3, 1>), | ||
| 36 | (&kleidicv::sve2::kleidicv_resize_generic_stripe_u8<3, 1>), | ||
| 37 | (&kleidicv::sme::kleidicv_resize_generic_stripe_u8<3, 1>), | ||
| 38 | (&kleidicv::sme2::kleidicv_resize_generic_stripe_u8<3, 1>), 16, 64); | ||
| 39 |
10/12✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 1 times.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 3 times.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 1 times.
✓ Branch 10 taken 1 times.
✓ Branch 11 taken 1 times.
|
10 | KLEIDICV_MULTIVERSION_C_API_VECLEN( |
| 40 | kleidicv_resize_2ch_r2_stripe_u8, | ||
| 41 | (&kleidicv::neon::kleidicv_resize_generic_stripe_u8<2, 2>), | ||
| 42 | (&kleidicv::sve2::kleidicv_resize_generic_stripe_u8<2, 2>), | ||
| 43 | (&kleidicv::sme::kleidicv_resize_generic_stripe_u8<2, 2>), | ||
| 44 | (&kleidicv::sme2::kleidicv_resize_generic_stripe_u8<2, 2>), 16, 64); | ||
| 45 |
10/12✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 1 times.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 3 times.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 1 times.
✓ Branch 10 taken 1 times.
✓ Branch 11 taken 1 times.
|
10 | KLEIDICV_MULTIVERSION_C_API_VECLEN( |
| 46 | kleidicv_resize_2ch_r3_stripe_u8, | ||
| 47 | (&kleidicv::neon::kleidicv_resize_generic_stripe_u8<3, 2>), | ||
| 48 | (&kleidicv::sve2::kleidicv_resize_generic_stripe_u8<3, 2>), | ||
| 49 | (&kleidicv::sme::kleidicv_resize_generic_stripe_u8<3, 2>), | ||
| 50 | (&kleidicv::sme2::kleidicv_resize_generic_stripe_u8<3, 2>), 16, 64); | ||
| 51 |
10/12✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 1 times.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 3 times.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 1 times.
✓ Branch 10 taken 1 times.
✓ Branch 11 taken 1 times.
|
10 | KLEIDICV_MULTIVERSION_C_API_VECLEN( |
| 52 | kleidicv_resize_3ch_r2_stripe_u8, | ||
| 53 | (&kleidicv::neon::kleidicv_resize_generic_stripe_u8<2, 3>), | ||
| 54 | (&kleidicv::sve2::kleidicv_resize_generic_stripe_u8<2, 3>), | ||
| 55 | (&kleidicv::sme::kleidicv_resize_generic_stripe_u8<2, 3>), | ||
| 56 | (&kleidicv::sme2::kleidicv_resize_generic_stripe_u8<2, 3>), 16, 64); | ||
| 57 |
10/12✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 1 times.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 3 times.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 1 times.
✓ Branch 10 taken 1 times.
✓ Branch 11 taken 1 times.
|
10 | KLEIDICV_MULTIVERSION_C_API_VECLEN( |
| 58 | kleidicv_resize_3ch_r3_stripe_u8, | ||
| 59 | (&kleidicv::neon::kleidicv_resize_generic_stripe_u8<3, 3>), | ||
| 60 | (&kleidicv::sve2::kleidicv_resize_generic_stripe_u8<3, 3>), | ||
| 61 | (&kleidicv::sme::kleidicv_resize_generic_stripe_u8<3, 3>), | ||
| 62 | (&kleidicv::sme2::kleidicv_resize_generic_stripe_u8<3, 3>), 16, 64); | ||
| 63 | |||
| 64 |
6/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 3 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 1 times.
|
10 | KLEIDICV_DEFINE_C_API_ALL(kleidicv_resize_linear_stripe_f32, |
| 65 | kleidicv_resize_linear_stripe_f32); | ||
| 66 | |||
| 67 |
6/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 3 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 1 times.
|
10 | KLEIDICV_MULTIVERSION_C_API( |
| 68 | kleidicv_resize_to_quarter_u8, &kleidicv::neon::resize_to_quarter_u8, | ||
| 69 | KLEIDICV_SVE2_IMPL_IF(&kleidicv::sve2::resize_to_quarter_u8), | ||
| 70 | &kleidicv::sme::resize_to_quarter_u8, | ||
| 71 | KLEIDICV_SME2_IMPL_IF(&kleidicv::sme2::resize_to_quarter_u8)); | ||
| 72 | |||
| 73 | extern "C" { | ||
| 74 | |||
| 75 | 1275 | kleidicv_error_t kleidicv_resize_linear_u8(const uint8_t *src, | |
| 76 | size_t src_stride, size_t src_width, | ||
| 77 | size_t src_height, uint8_t *dst, | ||
| 78 | size_t dst_stride, size_t dst_width, | ||
| 79 | size_t dst_height, size_t channels) { | ||
| 80 |
2/2✓ Branch 0 taken 260 times.
✓ Branch 1 taken 1015 times.
|
1275 | if (!kleidicv::resize_linear_u8_is_implemented( |
| 81 | 1275 | src_width, src_height, dst_width, dst_height, channels)) { | |
| 82 | 260 | return KLEIDICV_ERROR_NOT_IMPLEMENTED; | |
| 83 | } | ||
| 84 | |||
| 85 | // For upsampling, process by src rows | ||
| 86 | // For resize_generic, process by dst rows | ||
| 87 |
2/2✓ Branch 0 taken 350 times.
✓ Branch 1 taken 665 times.
|
1015 | size_t y_end = (src_width < dst_width) ? src_height : dst_height; |
| 88 | |||
| 89 | 2030 | return kleidicv_resize_linear_stripe_u8(src, src_stride, src_width, | |
| 90 | 1015 | src_height, 0, y_end, dst, dst_stride, | |
| 91 | 1015 | dst_width, dst_height, channels); | |
| 92 | 1275 | } | |
| 93 | |||
| 94 | // This function is too complex, but disable the warning for now. | ||
| 95 | // NOLINTBEGIN(readability-function-cognitive-complexity) | ||
| 96 | 2025 | kleidicv_error_t kleidicv_resize_linear_stripe_u8( | |
| 97 | const uint8_t *src, size_t src_stride, size_t src_width, size_t src_height, | ||
| 98 | size_t y_begin, size_t y_end, uint8_t *dst, size_t dst_stride, | ||
| 99 | size_t dst_width, size_t dst_height, size_t channels) { | ||
| 100 |
4/4✓ Branch 0 taken 20 times.
✓ Branch 1 taken 2005 times.
✓ Branch 2 taken 20 times.
✓ Branch 3 taken 2005 times.
|
2025 | CHECK_POINTER_AND_STRIDE(src, src_stride, src_height); |
| 101 |
4/4✓ Branch 0 taken 20 times.
✓ Branch 1 taken 1985 times.
✓ Branch 2 taken 20 times.
✓ Branch 3 taken 1985 times.
|
2005 | CHECK_POINTER_AND_STRIDE(dst, dst_stride, dst_height); |
| 102 | |||
| 103 |
4/4✓ Branch 0 taken 1980 times.
✓ Branch 1 taken 5 times.
✓ Branch 2 taken 5 times.
✓ Branch 3 taken 1975 times.
|
1985 | if (src_width == 0 || src_height == 0) { |
| 104 | 10 | return KLEIDICV_OK; | |
| 105 | } | ||
| 106 | |||
| 107 |
2/2✓ Branch 0 taken 1455 times.
✓ Branch 1 taken 520 times.
|
1975 | if (channels == 1) { |
| 108 |
4/4✓ Branch 0 taken 1010 times.
✓ Branch 1 taken 445 times.
✓ Branch 2 taken 80 times.
✓ Branch 3 taken 930 times.
|
1580 | if ((src_width / 2 == dst_width || (src_width + 1) / 2 == dst_width) && |
| 109 |
2/2✓ Branch 0 taken 125 times.
✓ Branch 1 taken 400 times.
|
525 | (src_height / 2 == dst_height || (src_height + 1) / 2 == dst_height)) { |
| 110 | 525 | size_t src_begin = y_begin * 2; | |
| 111 | 525 | size_t src_end = std::min<size_t>(src_height, y_end * 2); | |
| 112 | 525 | size_t dst_begin = y_begin; | |
| 113 | 525 | size_t dst_end = std::min<size_t>(dst_height, y_end); | |
| 114 | |||
| 115 |
2/2✓ Branch 0 taken 30 times.
✓ Branch 1 taken 495 times.
|
525 | if (dst_begin == dst_end) { |
| 116 | 30 | return KLEIDICV_OK; | |
| 117 | } | ||
| 118 | |||
| 119 | 990 | return kleidicv_resize_to_quarter_u8( | |
| 120 | 495 | src + src_begin * src_stride, src_stride, src_width, | |
| 121 | 495 | src_end - src_begin, dst + dst_begin * dst_stride, dst_stride, | |
| 122 | 495 | dst_width, dst_end - dst_begin); | |
| 123 | 525 | } | |
| 124 |
3/4✓ Branch 0 taken 390 times.
✓ Branch 1 taken 540 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 390 times.
|
930 | if (src_width * 2 == dst_width && src_height * 2 == dst_height) { |
| 125 | 780 | return kleidicv_resize_2x2_stripe_u8(src, src_stride, src_width, | |
| 126 | 390 | src_height, y_begin, y_end, dst, | |
| 127 | 390 | dst_stride); | |
| 128 | } | ||
| 129 |
3/4✓ Branch 0 taken 290 times.
✓ Branch 1 taken 250 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 290 times.
|
540 | if (src_width * 4 == dst_width && src_height * 4 == dst_height) { |
| 130 | 580 | return kleidicv_resize_4x4_stripe_u8(src, src_stride, src_width, | |
| 131 | 290 | src_height, y_begin, y_end, dst, | |
| 132 | 290 | dst_stride); | |
| 133 | } | ||
| 134 | |||
| 135 |
2/2✓ Branch 0 taken 125 times.
✓ Branch 1 taken 125 times.
|
250 | if (dst_width * 2 >= src_width) { |
| 136 | 250 | return kleidicv_resize_1ch_r2_stripe_u8( | |
| 137 | 125 | src, src_stride, src_width, src_height, y_begin, y_end, dst, | |
| 138 | 125 | dst_stride, dst_width, dst_height); | |
| 139 | } | ||
| 140 | 250 | return kleidicv_resize_1ch_r3_stripe_u8(src, src_stride, src_width, | |
| 141 | 125 | src_height, y_begin, y_end, dst, | |
| 142 | 125 | dst_stride, dst_width, dst_height); | |
| 143 | } | ||
| 144 | |||
| 145 |
2/2✓ Branch 0 taken 280 times.
✓ Branch 1 taken 240 times.
|
520 | if (channels == 2) { |
| 146 |
2/2✓ Branch 0 taken 140 times.
✓ Branch 1 taken 140 times.
|
280 | if (dst_width * 2 >= src_width) { |
| 147 | 280 | return kleidicv_resize_2ch_r2_stripe_u8( | |
| 148 | 140 | src, src_stride, src_width, src_height, y_begin, y_end, dst, | |
| 149 | 140 | dst_stride, dst_width, dst_height); | |
| 150 | } | ||
| 151 | 280 | return kleidicv_resize_2ch_r3_stripe_u8(src, src_stride, src_width, | |
| 152 | 140 | src_height, y_begin, y_end, dst, | |
| 153 | 140 | dst_stride, dst_width, dst_height); | |
| 154 | } | ||
| 155 | |||
| 156 | assert(channels == 3); | ||
| 157 | 480 | double inverz_scale = | |
| 158 | 240 | static_cast<double>(src_width) / static_cast<double>(dst_width); | |
| 159 | // SVE variant does not handle the rightmost lanes of b and d vectors for 3 | ||
| 160 | // channel images, so use the r3 and use the Neon variant only over 2.8 | ||
| 161 |
2/2✓ Branch 0 taken 112 times.
✓ Branch 1 taken 128 times.
|
240 | if (inverz_scale < 1.8) { |
| 162 | 224 | return kleidicv_resize_3ch_r2_stripe_u8(src, src_stride, src_width, | |
| 163 | 112 | src_height, y_begin, y_end, dst, | |
| 164 | 112 | dst_stride, dst_width, dst_height); | |
| 165 | } | ||
| 166 |
2/2✓ Branch 0 taken 120 times.
✓ Branch 1 taken 8 times.
|
128 | if (inverz_scale < 2.8) { |
| 167 | 240 | return kleidicv_resize_3ch_r3_stripe_u8(src, src_stride, src_width, | |
| 168 | 120 | src_height, y_begin, y_end, dst, | |
| 169 | 120 | dst_stride, dst_width, dst_height); | |
| 170 | } | ||
| 171 | 8 | return kleidicv::neon::kleidicv_resize_generic_stripe_u8<3, 3>( | |
| 172 | 8 | src, src_stride, src_width, src_height, y_begin, y_end, dst, dst_stride, | |
| 173 | 8 | dst_width, dst_height); | |
| 174 | 2025 | } | |
| 175 | // NOLINTEND(readability-function-cognitive-complexity) | ||
| 176 | |||
| 177 | 565 | kleidicv_error_t kleidicv_resize_linear_f32(const float *src, size_t src_stride, | |
| 178 | size_t src_width, size_t src_height, | ||
| 179 | float *dst, size_t dst_stride, | ||
| 180 | size_t dst_width, size_t dst_height, | ||
| 181 | size_t channels) { | ||
| 182 |
2/2✓ Branch 0 taken 80 times.
✓ Branch 1 taken 485 times.
|
565 | if (!kleidicv::resize_linear_f32_is_implemented( |
| 183 | 565 | src_width, src_height, dst_width, dst_height, channels)) { | |
| 184 | 80 | return KLEIDICV_ERROR_NOT_IMPLEMENTED; | |
| 185 | } | ||
| 186 | 970 | return kleidicv_resize_linear_stripe_f32(src, src_stride, src_width, | |
| 187 | 485 | src_height, 0, src_height, dst, | |
| 188 | 485 | dst_stride, dst_width, dst_height); | |
| 189 | 565 | } | |
| 190 | |||
| 191 | } // extern "C" | ||
| 192 |