| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // SPDX-FileCopyrightText: 2026 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 2 | // | ||
| 3 | // SPDX-License-Identifier: Apache-2.0 | ||
| 4 | |||
| 5 | #include <cstddef> | ||
| 6 | #include <cstdint> | ||
| 7 | #include <cstdlib> | ||
| 8 | #include <variant> | ||
| 9 | |||
| 10 | #include "kleidicv/ctypes.h" | ||
| 11 | #include "kleidicv/resize/resize_linear.h" | ||
| 12 | #include "resize_linear_generic_u8_neon.h" | ||
| 13 | |||
| 14 | namespace kleidicv::neon { | ||
| 15 | |||
| 16 | // ratio: number of vectors to load and resize to 1 vector | ||
| 17 | // supported ratio values: 2 and 3 | ||
| 18 | // supported channel counts: 1, 2 and 3 | ||
| 19 | template <ptrdiff_t kRatio, ptrdiff_t kChannels> | ||
| 20 | 313 | kleidicv_error_t kleidicv_resize_generic_stripe_u8( | |
| 21 | const uint8_t *src, size_t src_stride, size_t src_width, size_t src_height, | ||
| 22 | size_t y_begin, size_t y_end, uint8_t *dst, // NOLINT | ||
| 23 | size_t dst_stride, size_t dst_width, size_t dst_height) { | ||
| 24 | 313 | resize_linear_generic_u8::RowInterpolationConstantsGenerator<kRatio, | |
| 25 | kChannels> | ||
| 26 | 313 | generator{src_width, dst_width}; | |
| 27 | |||
| 28 | 313 | auto row_interpolation_constants_variant = generator(); | |
| 29 | |||
| 30 |
36/36✓ Branch 0 taken 2 times.
✓ Branch 1 taken 48 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 48 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 48 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 54 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 54 times.
✓ Branch 10 taken 2 times.
✓ Branch 11 taken 54 times.
✓ Branch 12 taken 2 times.
✓ Branch 13 taken 43 times.
✓ Branch 14 taken 2 times.
✓ Branch 15 taken 43 times.
✓ Branch 16 taken 2 times.
✓ Branch 17 taken 43 times.
✓ Branch 18 taken 2 times.
✓ Branch 19 taken 48 times.
✓ Branch 20 taken 2 times.
✓ Branch 21 taken 48 times.
✓ Branch 22 taken 2 times.
✓ Branch 23 taken 48 times.
✓ Branch 24 taken 2 times.
✓ Branch 25 taken 54 times.
✓ Branch 26 taken 2 times.
✓ Branch 27 taken 54 times.
✓ Branch 28 taken 2 times.
✓ Branch 29 taken 54 times.
✓ Branch 30 taken 2 times.
✓ Branch 31 taken 54 times.
✓ Branch 32 taken 2 times.
✓ Branch 33 taken 54 times.
✓ Branch 34 taken 2 times.
✓ Branch 35 taken 54 times.
|
638 | if (auto *err = |
| 31 | 313 | std::get_if<kleidicv_error_t>(&row_interpolation_constants_variant)) { | |
| 32 | 12 | return *err; | |
| 33 | } | ||
| 34 | 602 | auto &row_interpolation_constants = | |
| 35 | 301 | *std::get_if<resize_linear_generic_u8::RowInterpolationConstants>( | |
| 36 | &row_interpolation_constants_variant); | ||
| 37 | |||
| 38 | if constexpr (kChannels == 3) { | ||
| 39 | 194 | double inverz_scale = | |
| 40 | 97 | static_cast<double>(src_width) / static_cast<double>(dst_width); | |
| 41 |
3/4✗ Branch 0 not taken.
✓ Branch 1 taken 43 times.
✓ Branch 2 taken 8 times.
✓ Branch 3 taken 46 times.
|
97 | if ((kRatio == 2 && inverz_scale >= 1.8) || |
| 42 | 54 | (kRatio == 3 && inverz_scale >= 2.8)) { | |
| 43 | // Rightmost lanes of b and d vectors need to be set by scalar code, as | ||
| 44 | // the table lookup overindexes the src registers | ||
| 45 | 8 | resize_linear_generic_u8::ResizeGenericU8Operation<kRatio, kChannels, | |
| 46 | true> | ||
| 47 | 16 | operation(src, src_stride, src_height, y_begin, y_end, dst, | |
| 48 | 8 | dst_stride, dst_height); | |
| 49 | 8 | operation.process_rows(row_interpolation_constants); | |
| 50 | 8 | return KLEIDICV_OK; | |
| 51 | 8 | } | |
| 52 | 97 | } | |
| 53 | |||
| 54 | 293 | resize_linear_generic_u8::ResizeGenericU8Operation<kRatio, kChannels> | |
| 55 | 586 | operation(src, src_stride, src_height, y_begin, y_end, dst, dst_stride, | |
| 56 | 293 | dst_height); | |
| 57 | 293 | operation.process_rows(row_interpolation_constants); | |
| 58 | |||
| 59 | 293 | return KLEIDICV_OK; | |
| 60 | 313 | } | |
| 61 | |||
| 62 | #define KLEIDICV_INSTANTIATE_TEMPLATE(ratio, channels) \ | ||
| 63 | template kleidicv_error_t \ | ||
| 64 | kleidicv_resize_generic_stripe_u8<ratio, channels>( \ | ||
| 65 | const uint8_t *src, size_t src_stride, size_t src_width, \ | ||
| 66 | size_t src_height, size_t y_begin, size_t y_end, uint8_t *dst, \ | ||
| 67 | size_t dst_stride, size_t dst_width, size_t dst_height) | ||
| 68 | |||
| 69 | KLEIDICV_INSTANTIATE_TEMPLATE(2L, 1L); | ||
| 70 | KLEIDICV_INSTANTIATE_TEMPLATE(2L, 2L); | ||
| 71 | KLEIDICV_INSTANTIATE_TEMPLATE(2L, 3L); | ||
| 72 | KLEIDICV_INSTANTIATE_TEMPLATE(3L, 1L); | ||
| 73 | KLEIDICV_INSTANTIATE_TEMPLATE(3L, 2L); | ||
| 74 | KLEIDICV_INSTANTIATE_TEMPLATE(3L, 3L); | ||
| 75 | |||
| 76 | } // namespace kleidicv::neon | ||
| 77 |