test/nextgen/reference/matmul.cpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // | ||
| 2 | // SPDX-FileCopyrightText: Copyright 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/nextgen/reference/matmul.hpp" | ||
| 8 | |||
| 9 | #include <cstddef> | ||
| 10 | |||
| 11 | #include "test/common/assert.hpp" | ||
| 12 | #include "test/common/buffer.hpp" | ||
| 13 | #include "test/common/data_type.hpp" | ||
| 14 | #include "test/common/memory.hpp" | ||
| 15 | #include "test/common/round.hpp" | ||
| 16 | #include "test/common/span.hpp" | ||
| 17 | #include "test/nextgen/functions/fused_mul_add.hpp" | ||
| 18 | |||
| 19 | namespace kai::test { | ||
| 20 | |||
| 21 | namespace { | ||
| 22 | |||
| 23 | template <typename T> | ||
| 24 | 200 | [[nodiscard]] Buffer matmul_nt_t( | |
| 25 | size_t shape_m, size_t shape_n, size_t shape_k, Span<const std::byte> lhs, Span<const std::byte> rhs) { | ||
| 26 | 200 | Buffer dst(shape_m * round_up_division(shape_n * size_in_bits<T>, 8), 0); | |
| 27 | |||
| 28 |
2/2✓ Branch 0 taken 200 times.
✓ Branch 1 taken 15452 times.
|
15652 | for (size_t row = 0; row < shape_m; ++row) { |
| 29 |
2/2✓ Branch 0 taken 15452 times.
✓ Branch 1 taken 1246953 times.
|
1262405 | for (size_t col = 0; col < shape_n; ++col) { |
| 30 | 1246953 | T acc = static_cast<T>(0); | |
| 31 | |||
| 32 |
2/2✓ Branch 0 taken 100601313 times.
✓ Branch 1 taken 1246953 times.
|
101848266 | for (size_t depth = 0; depth < shape_k; ++depth) { |
| 33 |
1/2✓ Branch 0 taken 100601313 times.
✗ Branch 1 not taken.
|
100601313 | const T lhs_value = read_2d<T>(lhs, shape_k, row, depth); |
| 34 |
1/2✓ Branch 0 taken 100601313 times.
✗ Branch 1 not taken.
|
100601313 | const T rhs_value = read_2d<T>(rhs, shape_k, col, depth); |
| 35 | |||
| 36 |
1/2✓ Branch 0 taken 100601313 times.
✗ Branch 1 not taken.
|
100601313 | acc = fused_mul_add<T>(lhs_value, rhs_value, acc); |
| 37 | 100601313 | } | |
| 38 | |||
| 39 |
2/4✓ Branch 0 taken 1246953 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1246953 times.
✗ Branch 3 not taken.
|
1246953 | write_2d<T>(dst.view(), shape_n, row, col, acc); |
| 40 | 1246953 | } | |
| 41 | 15452 | } | |
| 42 | |||
| 43 | 200 | return dst; | |
| 44 | 200 | } | |
| 45 | |||
| 46 | } // namespace | ||
| 47 | |||
| 48 | 200 | MatMulFn make_matmul_nt_t(DataType dtype) { | |
| 49 |
1/2✓ Branch 0 taken 200 times.
✗ Branch 1 not taken.
|
200 | switch (dtype) { |
| 50 | case DataType::FP32: | ||
| 51 | 200 | return matmul_nt_t<float>; | |
| 52 | |||
| 53 | default: | ||
| 54 | ✗ | KAI_TEST_ERROR("Not implemented."); | |
| 55 | } | ||
| 56 | ✗ | } | |
| 57 | |||
| 58 | } // namespace kai::test | ||
| 59 |