|
| 1 | +// Copyright 2019 Google LLC |
| 2 | +// SPDX-License-Identifier: Apache-2.0 |
| 3 | +// |
| 4 | +// Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | +// you may not use this file except in compliance with the License. |
| 6 | +// You may obtain a copy of the License at |
| 7 | +// |
| 8 | +// http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | +// |
| 10 | +// Unless required by applicable law or agreed to in writing, software |
| 11 | +// distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | +// See the License for the specific language governing permissions and |
| 14 | +// limitations under the License. |
| 15 | + |
| 16 | +#include <stddef.h> |
| 17 | +#include <stdint.h> |
| 18 | + |
| 19 | +#undef HWY_TARGET_INCLUDE |
| 20 | +#define HWY_TARGET_INCLUDE "tests/minmax128_test.cc" |
| 21 | +#include "hwy/foreach_target.h" // IWYU pragma: keep |
| 22 | +#include "hwy/highway.h" |
| 23 | +#include "hwy/tests/test_util-inl.h" |
| 24 | + |
| 25 | +HWY_BEFORE_NAMESPACE(); |
| 26 | +namespace hwy { |
| 27 | +namespace HWY_NAMESPACE { |
| 28 | +namespace { |
| 29 | + |
| 30 | +template <class D> |
| 31 | +static HWY_NOINLINE Vec<D> Make128(D d, uint64_t hi, uint64_t lo) { |
| 32 | + alignas(16) uint64_t in[2]; |
| 33 | + in[0] = lo; |
| 34 | + in[1] = hi; |
| 35 | + return LoadDup128(d, in); |
| 36 | +} |
| 37 | + |
| 38 | +struct TestMinMax128 { |
| 39 | + template <typename T, class D> |
| 40 | + HWY_NOINLINE void operator()(T /*unused*/, D d) { |
| 41 | + using V = Vec<D>; |
| 42 | + const size_t N = Lanes(d); |
| 43 | + auto a_lanes = AllocateAligned<T>(N); |
| 44 | + auto b_lanes = AllocateAligned<T>(N); |
| 45 | + auto min_lanes = AllocateAligned<T>(N); |
| 46 | + auto max_lanes = AllocateAligned<T>(N); |
| 47 | + HWY_ASSERT(a_lanes && b_lanes && min_lanes && max_lanes); |
| 48 | + RandomState rng; |
| 49 | + |
| 50 | + const V v00 = Zero(d); |
| 51 | + const V v01 = Make128(d, 0, 1); |
| 52 | + const V v10 = Make128(d, 1, 0); |
| 53 | + const V v11 = Add(v01, v10); |
| 54 | + |
| 55 | + // Same arg |
| 56 | + HWY_ASSERT_VEC_EQ(d, v00, Min128(d, v00, v00)); |
| 57 | + HWY_ASSERT_VEC_EQ(d, v01, Min128(d, v01, v01)); |
| 58 | + HWY_ASSERT_VEC_EQ(d, v10, Min128(d, v10, v10)); |
| 59 | + HWY_ASSERT_VEC_EQ(d, v11, Min128(d, v11, v11)); |
| 60 | + HWY_ASSERT_VEC_EQ(d, v00, Max128(d, v00, v00)); |
| 61 | + HWY_ASSERT_VEC_EQ(d, v01, Max128(d, v01, v01)); |
| 62 | + HWY_ASSERT_VEC_EQ(d, v10, Max128(d, v10, v10)); |
| 63 | + HWY_ASSERT_VEC_EQ(d, v11, Max128(d, v11, v11)); |
| 64 | + |
| 65 | + // First arg less |
| 66 | + HWY_ASSERT_VEC_EQ(d, v00, Min128(d, v00, v01)); |
| 67 | + HWY_ASSERT_VEC_EQ(d, v01, Min128(d, v01, v10)); |
| 68 | + HWY_ASSERT_VEC_EQ(d, v10, Min128(d, v10, v11)); |
| 69 | + HWY_ASSERT_VEC_EQ(d, v01, Max128(d, v00, v01)); |
| 70 | + HWY_ASSERT_VEC_EQ(d, v10, Max128(d, v01, v10)); |
| 71 | + HWY_ASSERT_VEC_EQ(d, v11, Max128(d, v10, v11)); |
| 72 | + |
| 73 | + // Second arg less |
| 74 | + HWY_ASSERT_VEC_EQ(d, v00, Min128(d, v01, v00)); |
| 75 | + HWY_ASSERT_VEC_EQ(d, v01, Min128(d, v10, v01)); |
| 76 | + HWY_ASSERT_VEC_EQ(d, v10, Min128(d, v11, v10)); |
| 77 | + HWY_ASSERT_VEC_EQ(d, v01, Max128(d, v01, v00)); |
| 78 | + HWY_ASSERT_VEC_EQ(d, v10, Max128(d, v10, v01)); |
| 79 | + HWY_ASSERT_VEC_EQ(d, v11, Max128(d, v11, v10)); |
| 80 | + |
| 81 | + // Also check 128-bit blocks are independent |
| 82 | + for (size_t rep = 0; rep < AdjustedReps(1000); ++rep) { |
| 83 | + for (size_t i = 0; i < N; ++i) { |
| 84 | + a_lanes[i] = Random64(&rng); |
| 85 | + b_lanes[i] = Random64(&rng); |
| 86 | + } |
| 87 | + const V a = Load(d, a_lanes.get()); |
| 88 | + const V b = Load(d, b_lanes.get()); |
| 89 | + for (size_t i = 0; i < N; i += 2) { |
| 90 | + const bool lt = a_lanes[i + 1] == b_lanes[i + 1] |
| 91 | + ? (a_lanes[i] < b_lanes[i]) |
| 92 | + : (a_lanes[i + 1] < b_lanes[i + 1]); |
| 93 | + min_lanes[i + 0] = lt ? a_lanes[i + 0] : b_lanes[i + 0]; |
| 94 | + min_lanes[i + 1] = lt ? a_lanes[i + 1] : b_lanes[i + 1]; |
| 95 | + max_lanes[i + 0] = lt ? b_lanes[i + 0] : a_lanes[i + 0]; |
| 96 | + max_lanes[i + 1] = lt ? b_lanes[i + 1] : a_lanes[i + 1]; |
| 97 | + } |
| 98 | + HWY_ASSERT_VEC_EQ(d, min_lanes.get(), Min128(d, a, b)); |
| 99 | + HWY_ASSERT_VEC_EQ(d, max_lanes.get(), Max128(d, a, b)); |
| 100 | + } |
| 101 | + } |
| 102 | +}; |
| 103 | + |
| 104 | +HWY_NOINLINE void TestAllMinMax128() { |
| 105 | + ForGEVectors<128, TestMinMax128>()(uint64_t()); |
| 106 | +} |
| 107 | + |
| 108 | +struct TestMinMax128Upper { |
| 109 | + template <typename T, class D> |
| 110 | + HWY_NOINLINE void operator()(T /*unused*/, D d) { |
| 111 | + using V = Vec<D>; |
| 112 | + const size_t N = Lanes(d); |
| 113 | + auto a_lanes = AllocateAligned<T>(N); |
| 114 | + auto b_lanes = AllocateAligned<T>(N); |
| 115 | + auto min_lanes = AllocateAligned<T>(N); |
| 116 | + auto max_lanes = AllocateAligned<T>(N); |
| 117 | + RandomState rng; |
| 118 | + |
| 119 | + const V v00 = Zero(d); |
| 120 | + const V v01 = Make128(d, 0, 1); |
| 121 | + const V v10 = Make128(d, 1, 0); |
| 122 | + const V v11 = Add(v01, v10); |
| 123 | + |
| 124 | + // Same arg |
| 125 | + HWY_ASSERT_VEC_EQ(d, v00, Min128Upper(d, v00, v00)); |
| 126 | + HWY_ASSERT_VEC_EQ(d, v01, Min128Upper(d, v01, v01)); |
| 127 | + HWY_ASSERT_VEC_EQ(d, v10, Min128Upper(d, v10, v10)); |
| 128 | + HWY_ASSERT_VEC_EQ(d, v11, Min128Upper(d, v11, v11)); |
| 129 | + HWY_ASSERT_VEC_EQ(d, v00, Max128Upper(d, v00, v00)); |
| 130 | + HWY_ASSERT_VEC_EQ(d, v01, Max128Upper(d, v01, v01)); |
| 131 | + HWY_ASSERT_VEC_EQ(d, v10, Max128Upper(d, v10, v10)); |
| 132 | + HWY_ASSERT_VEC_EQ(d, v11, Max128Upper(d, v11, v11)); |
| 133 | + |
| 134 | + // Equivalent but not equal (chooses second arg) |
| 135 | + HWY_ASSERT_VEC_EQ(d, v01, Min128Upper(d, v00, v01)); |
| 136 | + HWY_ASSERT_VEC_EQ(d, v11, Min128Upper(d, v10, v11)); |
| 137 | + HWY_ASSERT_VEC_EQ(d, v00, Min128Upper(d, v01, v00)); |
| 138 | + HWY_ASSERT_VEC_EQ(d, v10, Min128Upper(d, v11, v10)); |
| 139 | + HWY_ASSERT_VEC_EQ(d, v00, Max128Upper(d, v01, v00)); |
| 140 | + HWY_ASSERT_VEC_EQ(d, v10, Max128Upper(d, v11, v10)); |
| 141 | + HWY_ASSERT_VEC_EQ(d, v01, Max128Upper(d, v00, v01)); |
| 142 | + HWY_ASSERT_VEC_EQ(d, v11, Max128Upper(d, v10, v11)); |
| 143 | + |
| 144 | + // First arg less |
| 145 | + HWY_ASSERT_VEC_EQ(d, v01, Min128Upper(d, v01, v10)); |
| 146 | + HWY_ASSERT_VEC_EQ(d, v10, Max128Upper(d, v01, v10)); |
| 147 | + |
| 148 | + // Second arg less |
| 149 | + HWY_ASSERT_VEC_EQ(d, v01, Min128Upper(d, v10, v01)); |
| 150 | + HWY_ASSERT_VEC_EQ(d, v10, Max128Upper(d, v10, v01)); |
| 151 | + |
| 152 | + // Also check 128-bit blocks are independent |
| 153 | + for (size_t rep = 0; rep < AdjustedReps(1000); ++rep) { |
| 154 | + for (size_t i = 0; i < N; ++i) { |
| 155 | + a_lanes[i] = Random64(&rng); |
| 156 | + b_lanes[i] = Random64(&rng); |
| 157 | + } |
| 158 | + const V a = Load(d, a_lanes.get()); |
| 159 | + const V b = Load(d, b_lanes.get()); |
| 160 | + for (size_t i = 0; i < N; i += 2) { |
| 161 | + const bool lt = a_lanes[i + 1] < b_lanes[i + 1]; |
| 162 | + min_lanes[i + 0] = lt ? a_lanes[i + 0] : b_lanes[i + 0]; |
| 163 | + min_lanes[i + 1] = lt ? a_lanes[i + 1] : b_lanes[i + 1]; |
| 164 | + max_lanes[i + 0] = lt ? b_lanes[i + 0] : a_lanes[i + 0]; |
| 165 | + max_lanes[i + 1] = lt ? b_lanes[i + 1] : a_lanes[i + 1]; |
| 166 | + } |
| 167 | + HWY_ASSERT_VEC_EQ(d, min_lanes.get(), Min128Upper(d, a, b)); |
| 168 | + HWY_ASSERT_VEC_EQ(d, max_lanes.get(), Max128Upper(d, a, b)); |
| 169 | + } |
| 170 | + } |
| 171 | +}; |
| 172 | + |
| 173 | +HWY_NOINLINE void TestAllMinMax128Upper() { |
| 174 | + ForGEVectors<128, TestMinMax128Upper>()(uint64_t()); |
| 175 | +} |
| 176 | + |
| 177 | +} // namespace |
| 178 | +// NOLINTNEXTLINE(google-readability-namespace-comments) |
| 179 | +} // namespace HWY_NAMESPACE |
| 180 | +} // namespace hwy |
| 181 | +HWY_AFTER_NAMESPACE(); |
| 182 | + |
| 183 | +#if HWY_ONCE |
| 184 | +namespace hwy { |
| 185 | +namespace { |
| 186 | +HWY_BEFORE_TEST(HwyMinMax128Test); |
| 187 | +HWY_EXPORT_AND_TEST_P(HwyMinMax128Test, TestAllMinMax128); |
| 188 | +HWY_EXPORT_AND_TEST_P(HwyMinMax128Test, TestAllMinMax128Upper); |
| 189 | +HWY_AFTER_TEST(); |
| 190 | +} // namespace |
| 191 | +} // namespace hwy |
| 192 | +HWY_TEST_MAIN(); |
| 193 | +#endif // HWY_ONCE |
0 commit comments