| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // | ||
| 2 | // SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 3 | // | ||
| 4 | // SPDX-License-Identifier: Apache-2.0 | ||
| 5 | // | ||
| 6 | |||
| 7 | #include "test/reference/transpose.hpp" | ||
| 8 | |||
| 9 | #include <cstddef> | ||
| 10 | #include <cstdint> | ||
| 11 | #include <cstring> | ||
| 12 | |||
| 13 | #include "kai/kai_common.h" | ||
| 14 | #include "test/common/buffer.hpp" | ||
| 15 | #include "test/common/data_type.hpp" | ||
| 16 | #include "test/common/memory.hpp" | ||
| 17 | #include "test/common/round.hpp" | ||
| 18 | |||
| 19 | namespace kai::test { | ||
| 20 | |||
| 21 | 250 | Buffer transpose(const void* data, DataType data_type, size_t height, size_t width) { | |
| 22 | − | KAI_ASSUME(data_type_size_in_bits(data_type) % 8 == 0); | |
| 23 | 250 | const auto element_size = data_type_size_in_bits(data_type) / 8; | |
| 24 | |||
| 25 | 250 | Buffer output(height * width * element_size); | |
| 26 | |||
| 27 | 250 | const auto* src_ptr = reinterpret_cast<const uint8_t*>(data); | |
| 28 | |||
| 29 |
2/2✓ Branch 0 taken 250 times.
✓ Branch 1 taken 34410 times.
|
34660 | for (size_t y = 0; y < width; ++y) { |
| 30 |
2/2✓ Branch 0 taken 34410 times.
✓ Branch 1 taken 1722400 times.
|
1756810 | for (size_t x = 0; x < height; ++x) { |
| 31 | 1722400 | memcpy( | |
| 32 |
1/2✓ Branch 0 taken 1722400 times.
✗ Branch 1 not taken.
|
1722400 | output.data() + (y * height + x) * element_size, src_ptr + (x * width + y) * element_size, |
| 33 | 1722400 | element_size); | |
| 34 | 1722400 | } | |
| 35 | 34410 | } | |
| 36 | |||
| 37 | 250 | return output; | |
| 38 | 250 | } | |
| 39 | |||
| 40 | template <typename T> | ||
| 41 | 2430 | Buffer transpose_with_padding( | |
| 42 | const void* data, const size_t height, const size_t width, const size_t src_stride, const size_t dst_stride, | ||
| 43 | const size_t dst_size) { | ||
| 44 | 2430 | Buffer output(dst_size); | |
| 45 | |||
| 46 |
4/4✓ Branch 0 taken 132268 times.
✓ Branch 1 taken 1814 times.
✓ Branch 2 taken 23576 times.
✓ Branch 3 taken 616 times.
|
158274 | for (size_t y = 0; y < width; ++y) { |
| 47 |
4/4✓ Branch 0 taken 11946068 times.
✓ Branch 1 taken 132268 times.
✓ Branch 2 taken 1339128 times.
✓ Branch 3 taken 23576 times.
|
13441040 | for (size_t x = 0; x < height; ++x) { |
| 48 |
2/4✓ Branch 0 taken 11946068 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1339128 times.
✗ Branch 3 not taken.
|
13285196 | auto element = read_array<T>(data, (x * src_stride) + y); |
| 49 |
4/8✓ Branch 0 taken 11946068 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 11946068 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 1339128 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 1339128 times.
✗ Branch 7 not taken.
|
13285196 | write_array<T>(output.data(), (y * dst_stride) + x, element); |
| 50 | 13285196 | } | |
| 51 | 155844 | } | |
| 52 | |||
| 53 | 2430 | return output; | |
| 54 | 2430 | } | |
| 55 | |||
| 56 | template Buffer transpose_with_padding<Int4>( | ||
| 57 | const void* data, const size_t height, const size_t width, const size_t src_stride, const size_t dst_stride, | ||
| 58 | const size_t dst_size); | ||
| 59 | |||
| 60 | template Buffer transpose_with_padding<int8_t>( | ||
| 61 | const void* data, const size_t height, const size_t width, const size_t src_stride, const size_t dst_stride, | ||
| 62 | const size_t dst_size); | ||
| 63 | |||
| 64 | template <typename T> | ||
| 65 | 1854 | Buffer transpose(const void* src, size_t height, size_t width) { | |
| 66 | 1854 | Buffer dst(round_up_division(height * width * size_in_bits<T>, 8)); | |
| 67 | |||
| 68 |
4/4✓ Branch 0 taken 105378 times.
✓ Branch 1 taken 927 times.
✓ Branch 2 taken 170643 times.
✓ Branch 3 taken 927 times.
|
277875 | for (size_t y = 0; y < width; ++y) { |
| 69 |
4/4✓ Branch 0 taken 16220061 times.
✓ Branch 1 taken 105378 times.
✓ Branch 2 taken 16220061 times.
✓ Branch 3 taken 170643 times.
|
32716143 | for (size_t x = 0; x < height; ++x) { |
| 70 |
6/12✓ Branch 0 taken 16220061 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 16220061 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 16220061 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 16220061 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 16220061 times.
✗ Branch 9 not taken.
✓ Branch 10 taken 16220061 times.
✗ Branch 11 not taken.
|
32440122 | write_array<T>(dst.data(), y * height + x, read_array<T>(src, x * width + y)); |
| 71 | 32440122 | } | |
| 72 | 276021 | } | |
| 73 | |||
| 74 | 1854 | return dst; | |
| 75 | 1854 | } | |
| 76 | |||
| 77 | template Buffer transpose<float>(const void* src, size_t height, size_t width); | ||
| 78 | template Buffer transpose<int8_t>(const void* src, size_t height, size_t width); | ||
| 79 | |||
| 80 | } // namespace kai::test | ||
| 81 |