1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
|
// ----------------------------------------------------------------------------
// - Open3D: www.open3d.org -
// ----------------------------------------------------------------------------
// Copyright (c) 2018-2024 www.open3d.org
// SPDX-License-Identifier: MIT
// ----------------------------------------------------------------------------
#include <benchmark/benchmark.h>
#include <type_traits>
#include "benchmarks/benchmark_utilities/Rand.h"
#include "open3d/core/CUDAUtils.h"
#include "open3d/core/Indexer.h"
#include "open3d/core/ParallelFor.h"
#include "open3d/core/Tensor.h"
#include "open3d/utility/Logging.h"
namespace open3d {
namespace core {
enum class BinaryOpCode {
Add,
Sub,
Mul,
Div,
LogicalAnd,
LogicalOr,
LogicalXor,
Gt,
Ge,
Lt,
Le,
Eq,
Neq,
};
static std::function<Tensor(const Tensor&, const Tensor&)> MakeOperation(
BinaryOpCode op) {
switch (op) {
case BinaryOpCode::Add:
return std::plus<Tensor>();
case BinaryOpCode::Sub:
return std::minus<Tensor>();
case BinaryOpCode::Mul:
return std::multiplies<Tensor>();
case BinaryOpCode::Div:
return std::divides<Tensor>();
case BinaryOpCode::LogicalAnd:
return [](const Tensor& lhs, const Tensor& rhs) -> Tensor {
return lhs && rhs;
};
case BinaryOpCode::LogicalOr:
return [](const Tensor& lhs, const Tensor& rhs) -> Tensor {
return lhs || rhs;
};
case BinaryOpCode::LogicalXor:
return [](const Tensor& lhs, const Tensor& rhs) -> Tensor {
return lhs.LogicalXor(rhs);
};
case BinaryOpCode::Gt:
return [](const Tensor& lhs, const Tensor& rhs) -> Tensor {
return lhs > rhs;
};
case BinaryOpCode::Ge:
return [](const Tensor& lhs, const Tensor& rhs) -> Tensor {
return lhs >= rhs;
};
case BinaryOpCode::Lt:
return [](const Tensor& lhs, const Tensor& rhs) -> Tensor {
return lhs < rhs;
};
case BinaryOpCode::Le:
return [](const Tensor& lhs, const Tensor& rhs) -> Tensor {
return lhs <= rhs;
};
case BinaryOpCode::Eq:
return [](const Tensor& lhs, const Tensor& rhs) -> Tensor {
return lhs == rhs;
};
case BinaryOpCode::Neq:
return [](const Tensor& lhs, const Tensor& rhs) -> Tensor {
return lhs != rhs;
};
default:
utility::LogError("Unknown operation {}", static_cast<int>(op));
}
}
void BinaryEW(benchmark::State& state,
int size,
BinaryOpCode op_code,
const Dtype& dtype,
const Device& device) {
Tensor lhs = benchmarks::Rand({1, size}, 1, {1, 127}, dtype, device);
Tensor rhs = benchmarks::Rand({1, size}, 2, {1, 127}, dtype, device);
auto op = MakeOperation(op_code);
Tensor result = op(lhs, rhs);
benchmark::DoNotOptimize(result);
for (auto _ : state) {
Tensor result = op(lhs, rhs);
benchmark::DoNotOptimize(result);
cuda::Synchronize(device);
}
}
#define ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, DTYPE) \
BENCHMARK_CAPTURE(FN, OP##__##DEVICE_NAME##_##DTYPE##__100, 100, \
BinaryOpCode::OP, DTYPE, DEVICE) \
->Unit(benchmark::kMillisecond); \
BENCHMARK_CAPTURE(FN, OP##__##DEVICE_NAME##_##DTYPE##__100000, 100000, \
BinaryOpCode::OP, DTYPE, DEVICE) \
->Unit(benchmark::kMillisecond); \
BENCHMARK_CAPTURE(FN, OP##__##DEVICE_NAME##_##DTYPE##__100000000, \
100000000, BinaryOpCode::OP, DTYPE, DEVICE) \
->Unit(benchmark::kMillisecond);
#define ENUM_BM_DTYPE(FN, OP, DEVICE, DEVICE_NAME) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Int8) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, UInt8) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Int16) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, UInt16) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Int32) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, UInt32) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Int64) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, UInt64) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Float32) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Float64)
#define ENUM_BM_DTYPE_WITH_BOOL(FN, OP, DEVICE, DEVICE_NAME) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Bool) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Int8) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, UInt8) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Int16) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, UInt16) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Int32) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, UInt32) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Int64) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, UInt64) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Float32) \
ENUM_BM_SIZE(FN, OP, DEVICE, DEVICE_NAME, Float64)
// #ifdef BUILD_CUDA_MODULE
// #define ENUM_BM_TENSOR(FN, OP)
// ENUM_BM_DTYPE(FN, OP, Device("CPU:0"), CPU)
// ENUM_BM_DTYPE(FN, OP, Device("CUDA:0"), CUDA)
// #else
#define ENUM_BM_TENSOR(FN, OP) ENUM_BM_DTYPE(FN, OP, Device("CPU:0"), CPU)
// #endif
// #ifdef BUILD_CUDA_MODULE
// #define ENUM_BM_TENSOR_WTIH_BOOL(FN, OP)
// ENUM_BM_DTYPE_WITH_BOOL(FN, OP, Device("CPU:0"), CPU)
// ENUM_BM_DTYPE_WITH_BOOL(FN, OP, Device("CUDA:0"), CUDA)
// #else
#define ENUM_BM_TENSOR_WTIH_BOOL(FN, OP) \
ENUM_BM_DTYPE_WITH_BOOL(FN, OP, Device("CPU:0"), CPU)
// #endif
ENUM_BM_TENSOR(BinaryEW, Add)
ENUM_BM_TENSOR(BinaryEW, Sub)
ENUM_BM_TENSOR(BinaryEW, Mul)
ENUM_BM_TENSOR(BinaryEW, Div)
ENUM_BM_TENSOR_WTIH_BOOL(BinaryEW, LogicalAnd)
ENUM_BM_TENSOR_WTIH_BOOL(BinaryEW, LogicalOr)
ENUM_BM_TENSOR_WTIH_BOOL(BinaryEW, LogicalXor)
ENUM_BM_TENSOR_WTIH_BOOL(BinaryEW, Gt)
ENUM_BM_TENSOR_WTIH_BOOL(BinaryEW, Ge)
ENUM_BM_TENSOR_WTIH_BOOL(BinaryEW, Lt)
ENUM_BM_TENSOR_WTIH_BOOL(BinaryEW, Le)
ENUM_BM_TENSOR_WTIH_BOOL(BinaryEW, Eq)
ENUM_BM_TENSOR_WTIH_BOOL(BinaryEW, Neq)
} // namespace core
} // namespace open3d
|
