| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // SPDX-FileCopyrightText: 2025 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 2 | // | ||
| 3 | // SPDX-License-Identifier: Apache-2.0 | ||
| 4 | |||
| 5 | #ifndef MULTIPLY_SC_H | ||
| 6 | #define MULTIPLY_SC_H | ||
| 7 | |||
| 8 | #include "kleidicv/sve2.h" | ||
| 9 | |||
| 10 | namespace KLEIDICV_TARGET_NAMESPACE { | ||
| 11 | |||
| 12 | template <typename ScalarType> | ||
| 13 | class SaturatingMultiply final : public UnrollTwice { | ||
| 14 | public: | ||
| 15 | using ContextType = Context; | ||
| 16 | using VecTraits = KLEIDICV_TARGET_NAMESPACE::VecTraits<ScalarType>; | ||
| 17 | using VectorType = typename VecTraits::VectorType; | ||
| 18 | |||
| 19 | 1245 | explicit SaturatingMultiply(double scale = 1.0) : scale_{scale} {}; | |
| 20 | |||
| 21 | 8334 | VectorType vector_path(ContextType ctx, VectorType src_a, | |
| 22 | VectorType src_b) KLEIDICV_STREAMING { | ||
| 23 | 8334 | VectorType result; | |
| 24 | (void)ctx; | ||
| 25 | |||
| 26 | // multiply-widen even-indexed elements | ||
| 27 | 8334 | auto bottom_part = svmullb(src_a, src_b); | |
| 28 | // multiply-widen odd-indexed elements | ||
| 29 | 8334 | auto top_part = svmullt(src_a, src_b); | |
| 30 | // saturating-narrow even-indexed | ||
| 31 | 8334 | auto narrow_bottom = svqxtnb(bottom_part); | |
| 32 | // saturaning-narrow odd-indexed and merge with even-indexed | ||
| 33 | 8334 | result = svqxtnt(narrow_bottom, top_part); | |
| 34 | |||
| 35 | /* TODO: figure out the way to multiply by double or some | ||
| 36 | fixed supported scale. | ||
| 37 | */ | ||
| 38 | |||
| 39 | 16668 | return result; | |
| 40 | 8334 | } | |
| 41 | |||
| 42 | private: | ||
| 43 | double scale_; | ||
| 44 | }; | ||
| 45 | |||
| 46 | template <typename T> | ||
| 47 | 1347 | kleidicv_error_t saturating_multiply_sc(const T *src_a, size_t src_a_stride, | |
| 48 | const T *src_b, size_t src_b_stride, | ||
| 49 | T *dst, size_t dst_stride, size_t width, | ||
| 50 | size_t height, | ||
| 51 | double scale) KLEIDICV_STREAMING { | ||
| 52 |
20/20✓ Branch 0 taken 3 times.
✓ Branch 1 taken 261 times.
✓ Branch 2 taken 3 times.
✓ Branch 3 taken 261 times.
✓ Branch 4 taken 3 times.
✓ Branch 5 taken 261 times.
✓ Branch 6 taken 3 times.
✓ Branch 7 taken 261 times.
✓ Branch 8 taken 6 times.
✓ Branch 9 taken 267 times.
✓ Branch 10 taken 6 times.
✓ Branch 11 taken 267 times.
✓ Branch 12 taken 6 times.
✓ Branch 13 taken 267 times.
✓ Branch 14 taken 6 times.
✓ Branch 15 taken 267 times.
✓ Branch 16 taken 6 times.
✓ Branch 17 taken 267 times.
✓ Branch 18 taken 6 times.
✓ Branch 19 taken 267 times.
|
1347 | CHECK_POINTER_AND_STRIDE(src_a, src_a_stride, height); |
| 53 |
20/20✓ Branch 0 taken 3 times.
✓ Branch 1 taken 258 times.
✓ Branch 2 taken 3 times.
✓ Branch 3 taken 258 times.
✓ Branch 4 taken 3 times.
✓ Branch 5 taken 258 times.
✓ Branch 6 taken 3 times.
✓ Branch 7 taken 258 times.
✓ Branch 8 taken 6 times.
✓ Branch 9 taken 261 times.
✓ Branch 10 taken 6 times.
✓ Branch 11 taken 261 times.
✓ Branch 12 taken 6 times.
✓ Branch 13 taken 261 times.
✓ Branch 14 taken 6 times.
✓ Branch 15 taken 261 times.
✓ Branch 16 taken 6 times.
✓ Branch 17 taken 261 times.
✓ Branch 18 taken 6 times.
✓ Branch 19 taken 261 times.
|
1323 | CHECK_POINTER_AND_STRIDE(src_b, src_b_stride, height); |
| 54 |
20/20✓ Branch 0 taken 3 times.
✓ Branch 1 taken 255 times.
✓ Branch 2 taken 3 times.
✓ Branch 3 taken 255 times.
✓ Branch 4 taken 3 times.
✓ Branch 5 taken 255 times.
✓ Branch 6 taken 3 times.
✓ Branch 7 taken 255 times.
✓ Branch 8 taken 6 times.
✓ Branch 9 taken 255 times.
✓ Branch 10 taken 6 times.
✓ Branch 11 taken 255 times.
✓ Branch 12 taken 6 times.
✓ Branch 13 taken 255 times.
✓ Branch 14 taken 6 times.
✓ Branch 15 taken 255 times.
✓ Branch 16 taken 6 times.
✓ Branch 17 taken 255 times.
✓ Branch 18 taken 6 times.
✓ Branch 19 taken 255 times.
|
1299 | CHECK_POINTER_AND_STRIDE(dst, dst_stride, height); |
| 55 |
30/30✓ Branch 0 taken 3 times.
✓ Branch 1 taken 252 times.
✓ Branch 2 taken 3 times.
✓ Branch 3 taken 249 times.
✓ Branch 4 taken 6 times.
✓ Branch 5 taken 249 times.
✓ Branch 6 taken 3 times.
✓ Branch 7 taken 252 times.
✓ Branch 8 taken 3 times.
✓ Branch 9 taken 249 times.
✓ Branch 10 taken 6 times.
✓ Branch 11 taken 249 times.
✓ Branch 12 taken 3 times.
✓ Branch 13 taken 252 times.
✓ Branch 14 taken 3 times.
✓ Branch 15 taken 249 times.
✓ Branch 16 taken 6 times.
✓ Branch 17 taken 249 times.
✓ Branch 18 taken 3 times.
✓ Branch 19 taken 252 times.
✓ Branch 20 taken 3 times.
✓ Branch 21 taken 249 times.
✓ Branch 22 taken 6 times.
✓ Branch 23 taken 249 times.
✓ Branch 24 taken 3 times.
✓ Branch 25 taken 252 times.
✓ Branch 26 taken 3 times.
✓ Branch 27 taken 249 times.
✓ Branch 28 taken 6 times.
✓ Branch 29 taken 249 times.
|
1275 | CHECK_IMAGE_SIZE(width, height); |
| 56 | |||
| 57 | (void)scale; // TODO: figure out the way to process the scale. | ||
| 58 | 1245 | SaturatingMultiply<T> operation; | |
| 59 | 1245 | Rectangle rect{width, height}; | |
| 60 | 1245 | Rows<const T> src_a_rows{src_a, src_a_stride}; | |
| 61 | 1245 | Rows<const T> src_b_rows{src_b, src_b_stride}; | |
| 62 | 1245 | Rows<T> dst_rows{dst, dst_stride}; | |
| 63 | 1245 | apply_operation_by_rows(operation, rect, src_a_rows, src_b_rows, dst_rows); | |
| 64 | 1245 | return KLEIDICV_OK; | |
| 65 | 1347 | } | |
| 66 | |||
| 67 | } // namespace KLEIDICV_TARGET_NAMESPACE | ||
| 68 | |||
| 69 | #endif // MULTIPLY_SC_H | ||
| 70 |