| 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/cast.hpp" | ||
| 8 | |||
| 9 | #include <cstddef> | ||
| 10 | #include <cstdint> | ||
| 11 | |||
| 12 | #include "kai/kai_common.h" | ||
| 13 | #include "test/common/bfloat16.hpp" | ||
| 14 | #include "test/common/buffer.hpp" | ||
| 15 | #include "test/common/data_type.hpp" | ||
| 16 | #include "test/common/float16.hpp" | ||
| 17 | #include "test/common/memory.hpp" | ||
| 18 | #include "test/common/round.hpp" | ||
| 19 | |||
| 20 | namespace kai::test { | ||
| 21 | |||
| 22 | template <typename DstType, typename SrcType> | ||
| 23 | 11192 | Buffer cast(const void* src, size_t length) { | |
| 24 | 11192 | Buffer dst(round_up_division(length * size_in_bits<DstType>, 8)); | |
| 25 | |||
| 26 |
10/12✓ Branch 0 taken 3130780 times.
✓ Branch 1 taken 3984 times.
✓ Branch 2 taken 2721210 times.
✓ Branch 3 taken 2384 times.
✓ Branch 4 taken 3475372 times.
✓ Branch 5 taken 108 times.
✓ Branch 6 taken 2361560 times.
✓ Branch 7 taken 3736 times.
✓ Branch 8 taken 1942616 times.
✓ Branch 9 taken 980 times.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
|
13642730 | for (size_t i = 0; i < length; ++i) { |
| 27 |
20/48✓ Branch 0 taken 3130780 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 3130780 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 3130780 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 3130780 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2721210 times.
✗ Branch 9 not taken.
✓ Branch 10 taken 2721210 times.
✗ Branch 11 not taken.
✓ Branch 12 taken 2721210 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2721210 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 3475372 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 3475372 times.
✗ Branch 19 not taken.
✓ Branch 20 taken 3475372 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 3475372 times.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2361560 times.
✓ Branch 26 taken 2361560 times.
✗ Branch 27 not taken.
✓ Branch 28 taken 2361560 times.
✗ Branch 29 not taken.
✓ Branch 30 taken 2361560 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✓ Branch 33 taken 1942616 times.
✓ Branch 34 taken 1942616 times.
✗ Branch 35 not taken.
✓ Branch 36 taken 1942616 times.
✗ Branch 37 not taken.
✓ Branch 38 taken 1942616 times.
✗ Branch 39 not taken.
✗ Branch 40 not taken.
✗ Branch 41 not taken.
✗ Branch 42 not taken.
✗ Branch 43 not taken.
✗ Branch 44 not taken.
✗ Branch 45 not taken.
✗ Branch 46 not taken.
✗ Branch 47 not taken.
|
13631538 | write_array(dst.data(), i, static_cast<DstType>(read_array<SrcType>(src, i))); |
| 28 | 13631538 | } | |
| 29 | |||
| 30 | 11192 | return dst; | |
| 31 | 11192 | } | |
| 32 | |||
| 33 | template <> | ||
| 34 | ✗ | Buffer cast<BFloat16<false>, Float16>(const void* src, size_t length) { | |
| 35 | ✗ | Buffer dst(round_up_division(length * size_in_bits<BFloat16<>>, 8)); | |
| 36 | |||
| 37 | ✗ | for (size_t i = 0; i < length; ++i) { | |
| 38 | ✗ | float interim = static_cast<float>(read_array<Float16>(src, i)); | |
| 39 | ✗ | write_array(dst.data(), i, BFloat16<false>(interim)); | |
| 40 | ✗ | } | |
| 41 | |||
| 42 | ✗ | return dst; | |
| 43 | ✗ | } | |
| 44 | |||
| 45 | template <> | ||
| 46 | 40 | Buffer cast<BFloat16<true>, Float16>(const void* src, size_t length) { | |
| 47 | 40 | Buffer dst(round_up_division(length * size_in_bits<BFloat16<>>, 8)); | |
| 48 | |||
| 49 |
2/2✓ Branch 0 taken 593616 times.
✓ Branch 1 taken 40 times.
|
593656 | for (size_t i = 0; i < length; ++i) { |
| 50 |
2/4✓ Branch 0 taken 593616 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 593616 times.
✗ Branch 3 not taken.
|
593616 | float interim = static_cast<float>(read_array<Float16>(src, i)); |
| 51 |
3/6✓ Branch 0 taken 593616 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 593616 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 593616 times.
✗ Branch 5 not taken.
|
593616 | write_array(dst.data(), i, BFloat16<true>(interim)); |
| 52 | 593616 | } | |
| 53 | |||
| 54 | 40 | return dst; | |
| 55 | 40 | } | |
| 56 | |||
| 57 | template Buffer cast<Float16, float>(const void* src, size_t length); | ||
| 58 | template Buffer cast<BFloat16<false>, float>(const void* src, size_t length); | ||
| 59 | template Buffer cast<BFloat16<true>, float>(const void* src, size_t length); | ||
| 60 | template Buffer cast<float, Float16>(const void* src, size_t length); | ||
| 61 | template Buffer cast<float, BFloat16<false>>(const void* src, size_t length); | ||
| 62 | template Buffer cast<float, BFloat16<true>>(const void* src, size_t length); | ||
| 63 | |||
| 64 | 148 | Buffer cast(const void* src, kai::test::DataType src_dt, DataType dst_dt, size_t height, size_t width) { | |
| 65 | 148 | const auto length = height * width; | |
| 66 | |||
| 67 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 148 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
148 | if (src_dt == DataType::BF16 && dst_dt == DataType::FP32) { |
| 68 | ✗ | return cast<float, BFloat16<>>(src, length); | |
| 69 |
3/4✓ Branch 0 taken 40 times.
✓ Branch 1 taken 108 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 40 times.
|
148 | } else if (src_dt == DataType::FP16 && dst_dt == DataType::BF16) { |
| 70 | 40 | return cast<BFloat16<>, Float16>(src, length); | |
| 71 |
1/2✓ Branch 0 taken 108 times.
✗ Branch 1 not taken.
|
108 | } else if (src_dt == DataType::FP32 && dst_dt == DataType::BF16) { |
| 72 | 108 | return cast<BFloat16<>, float>(src, length); | |
| 73 | } | ||
| 74 | |||
| 75 | − | KAI_ERROR("Unsupported cast data type!"); | |
| 76 | 148 | } | |
| 77 | |||
| 78 | 8262 | Buffer cast_qsu4_qsi4(const void* src, size_t length) { | |
| 79 | 8262 | Buffer dst(round_up_division(length, 2)); | |
| 80 | |||
| 81 |
2/2✓ Branch 0 taken 49710168 times.
✓ Branch 1 taken 8262 times.
|
49718430 | for (size_t i = 0; i < length; ++i) { |
| 82 |
5/10✓ Branch 0 taken 49710168 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 49710168 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 49710168 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 49710168 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 49710168 times.
✗ Branch 9 not taken.
|
49710168 | write_array(dst.data(), i, static_cast<UInt4>(static_cast<int32_t>(read_array<Int4>(src, i)) + 8)); |
| 83 | 49710168 | } | |
| 84 | |||
| 85 | 8262 | return dst; | |
| 86 | 8262 | } | |
| 87 | |||
| 88 | } // namespace kai::test | ||
| 89 |