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 <limits> | ||
6 | |||
7 | #include "kleidicv/kleidicv.h" | ||
8 | #include "kleidicv/neon.h" | ||
9 | #include "kleidicv/types.h" | ||
10 | |||
11 | namespace kleidicv::neon { | ||
12 | |||
13 | template <typename ScalarType> | ||
14 | class SaturatingMultiply final : public UnrollTwice { | ||
15 | public: | ||
16 | using VecTraits = neon::VecTraits<ScalarType>; | ||
17 | using VectorType = typename VecTraits::VectorType; | ||
18 | |||
19 | 325 | explicit SaturatingMultiply(double scale = 1.0) : scale_{scale} {}; | |
20 | |||
21 | 4664 | VectorType vector_path(VectorType src_a, VectorType src_b) { | |
22 | 4664 | VectorType result; | |
23 | |||
24 | // multiply-widen high part of the vectors. | ||
25 | // results in e.g. int16x8 -> int32x4 | ||
26 | 4664 | auto high_part = vmull_high(src_a, src_b); | |
27 | |||
28 | // get low part of the vectors and multiply-widen these. | ||
29 | 4664 | auto low_part = vmull(vget_low(src_a), vget_low(src_b)); | |
30 | |||
31 | // narrow-saturate low_part back to int16x4 | ||
32 | // narrow-saturate high_part back to int16x4 | ||
33 | // and stitch them back together into int16x8 | ||
34 | 4664 | result = vqmovn_high(vqmovn(low_part), high_part); | |
35 | |||
36 | /* TODO: figure out the way to multiply by double or some | ||
37 | fixed supported scale. Note that vmulq_n does not support 8x8 vectors. | ||
38 | */ | ||
39 | // result = vmulq_n(result, this->scale_); | ||
40 | |||
41 | 9328 | return result; | |
42 | 4664 | } | |
43 | |||
44 | 1736 | ScalarType scalar_path(ScalarType src_a, ScalarType src_b) { | |
45 | 1736 | ScalarType result; | |
46 | if (std::numeric_limits<ScalarType>::is_signed) { | ||
47 |
6/6✓ Branch 0 taken 416 times.
✓ Branch 1 taken 232 times.
✓ Branch 2 taken 140 times.
✓ Branch 3 taken 108 times.
✓ Branch 4 taken 76 times.
✓ Branch 5 taken 44 times.
|
1016 | if (__builtin_mul_overflow(src_a, src_b, &result)) { |
48 |
18/18✓ Branch 0 taken 180 times.
✓ Branch 1 taken 236 times.
✓ Branch 2 taken 77 times.
✓ Branch 3 taken 103 times.
✓ Branch 4 taken 103 times.
✓ Branch 5 taken 236 times.
✓ Branch 6 taken 61 times.
✓ Branch 7 taken 79 times.
✓ Branch 8 taken 19 times.
✓ Branch 9 taken 42 times.
✓ Branch 10 taken 42 times.
✓ Branch 11 taken 79 times.
✓ Branch 12 taken 29 times.
✓ Branch 13 taken 47 times.
✓ Branch 14 taken 9 times.
✓ Branch 15 taken 20 times.
✓ Branch 16 taken 20 times.
✓ Branch 17 taken 47 times.
|
632 | return (src_a < 0 && src_b > 0) || (src_a > 0 && src_b < 0) |
49 | 632 | ? std::numeric_limits<ScalarType>::lowest() | |
50 | 632 | : std::numeric_limits<ScalarType>::max(); | |
51 | } | ||
52 | 384 | return result; | |
53 | } | ||
54 | |||
55 |
4/4✓ Branch 0 taken 418 times.
✓ Branch 1 taken 118 times.
✓ Branch 2 taken 136 times.
✓ Branch 3 taken 48 times.
|
720 | if (__builtin_mul_overflow(src_a, src_b, &result)) { |
56 | 554 | return std::numeric_limits<ScalarType>::max(); | |
57 | } | ||
58 | 166 | return result; | |
59 | 1736 | } | |
60 | |||
61 | private: | ||
62 | double scale_; | ||
63 | }; | ||
64 | |||
65 | template <typename T> | ||
66 | 359 | kleidicv_error_t saturating_multiply(const T *src_a, size_t src_a_stride, | |
67 | const T *src_b, size_t src_b_stride, | ||
68 | T *dst, size_t dst_stride, size_t width, | ||
69 | size_t height, double scale) { | ||
70 |
20/20✓ Branch 0 taken 1 times.
✓ Branch 1 taken 69 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 69 times.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 69 times.
✓ Branch 6 taken 1 times.
✓ Branch 7 taken 69 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 71 times.
✓ Branch 10 taken 2 times.
✓ Branch 11 taken 71 times.
✓ Branch 12 taken 2 times.
✓ Branch 13 taken 71 times.
✓ Branch 14 taken 2 times.
✓ Branch 15 taken 71 times.
✓ Branch 16 taken 2 times.
✓ Branch 17 taken 71 times.
✓ Branch 18 taken 2 times.
✓ Branch 19 taken 71 times.
|
359 | CHECK_POINTER_AND_STRIDE(src_a, src_a_stride, height); |
71 |
20/20✓ Branch 0 taken 1 times.
✓ Branch 1 taken 68 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 68 times.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 68 times.
✓ Branch 6 taken 1 times.
✓ Branch 7 taken 68 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 69 times.
✓ Branch 10 taken 2 times.
✓ Branch 11 taken 69 times.
✓ Branch 12 taken 2 times.
✓ Branch 13 taken 69 times.
✓ Branch 14 taken 2 times.
✓ Branch 15 taken 69 times.
✓ Branch 16 taken 2 times.
✓ Branch 17 taken 69 times.
✓ Branch 18 taken 2 times.
✓ Branch 19 taken 69 times.
|
351 | CHECK_POINTER_AND_STRIDE(src_b, src_b_stride, height); |
72 |
20/20✓ Branch 0 taken 1 times.
✓ Branch 1 taken 67 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 67 times.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 67 times.
✓ Branch 6 taken 1 times.
✓ Branch 7 taken 67 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 67 times.
✓ Branch 10 taken 2 times.
✓ Branch 11 taken 67 times.
✓ Branch 12 taken 2 times.
✓ Branch 13 taken 67 times.
✓ Branch 14 taken 2 times.
✓ Branch 15 taken 67 times.
✓ Branch 16 taken 2 times.
✓ Branch 17 taken 67 times.
✓ Branch 18 taken 2 times.
✓ Branch 19 taken 67 times.
|
343 | CHECK_POINTER_AND_STRIDE(dst, dst_stride, height); |
73 |
30/30✓ Branch 0 taken 1 times.
✓ Branch 1 taken 66 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 65 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 65 times.
✓ Branch 6 taken 1 times.
✓ Branch 7 taken 66 times.
✓ Branch 8 taken 1 times.
✓ Branch 9 taken 65 times.
✓ Branch 10 taken 2 times.
✓ Branch 11 taken 65 times.
✓ Branch 12 taken 1 times.
✓ Branch 13 taken 66 times.
✓ Branch 14 taken 1 times.
✓ Branch 15 taken 65 times.
✓ Branch 16 taken 2 times.
✓ Branch 17 taken 65 times.
✓ Branch 18 taken 1 times.
✓ Branch 19 taken 66 times.
✓ Branch 20 taken 1 times.
✓ Branch 21 taken 65 times.
✓ Branch 22 taken 2 times.
✓ Branch 23 taken 65 times.
✓ Branch 24 taken 1 times.
✓ Branch 25 taken 66 times.
✓ Branch 26 taken 1 times.
✓ Branch 27 taken 65 times.
✓ Branch 28 taken 2 times.
✓ Branch 29 taken 65 times.
|
335 | CHECK_IMAGE_SIZE(width, height); |
74 | |||
75 | (void)scale; // TODO: figure out the way to process the scale. | ||
76 | 325 | SaturatingMultiply<T> operation; | |
77 | 325 | Rectangle rect{width, height}; | |
78 | 325 | Rows<const T> src_a_rows{src_a, src_a_stride}; | |
79 | 325 | Rows<const T> src_b_rows{src_b, src_b_stride}; | |
80 | 325 | Rows<T> dst_rows{dst, dst_stride}; | |
81 | 325 | neon::apply_operation_by_rows(operation, rect, src_a_rows, src_b_rows, | |
82 | dst_rows); | ||
83 | 325 | return KLEIDICV_OK; | |
84 | 359 | } | |
85 | |||
86 | #define KLEIDICV_INSTANTIATE_TEMPLATE(type) \ | ||
87 | template KLEIDICV_TARGET_FN_ATTRS kleidicv_error_t \ | ||
88 | saturating_multiply<type>(const type *src_a, size_t src_a_stride, \ | ||
89 | const type *src_b, size_t src_b_stride, type *dst, \ | ||
90 | size_t dst_stride, size_t width, size_t height, \ | ||
91 | double scale) | ||
92 | |||
93 | KLEIDICV_INSTANTIATE_TEMPLATE(uint8_t); | ||
94 | KLEIDICV_INSTANTIATE_TEMPLATE(int8_t); | ||
95 | KLEIDICV_INSTANTIATE_TEMPLATE(uint16_t); | ||
96 | KLEIDICV_INSTANTIATE_TEMPLATE(int16_t); | ||
97 | KLEIDICV_INSTANTIATE_TEMPLATE(int32_t); | ||
98 | |||
99 | } // namespace kleidicv::neon | ||
100 |