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
|
CPU Architecture-Specific Implementations
=========================================
* [Overview](#overview)
* [Runtime CPU Detection](#runtime-cpu-detection)
* [Inspecting the Detected Implementation](#inspecting-the-detected-implementation)
* [Querying Available Implementations](#querying-available-implementations)
* [Manually Selecting the Implementation](#manually-selecting-the-implementation)
* [Checking that an Implementation can Run on your System](#checking-that-an-implementation-can-run-on-your-system)
Overview
--------
The simdjson library takes advantage of SIMD instruction sets such as NEON, SSE and AVX to achieve
much of its speed. Because these instruction sets work differently, simdjson has to compile a
different version of the JSON parser for different CPU architectures, often with different
algorithms to take better advantage of a given CPU!
The current implementations are:
* icelake: AVX-512F, AVX-512_VBMI, AVX-512_VBMI2, AVX-512_DQ, AVX-512512_CD, AVX-512_BW, AVX-512_VL (2019 Intel Ice Lake, Intel Rocket Lake, Intel Sapphire Rapids, AMD Zen 4)
* haswell: AVX2 (2013 Intel Haswell or later, all AMD Zen processors)
* westmere: SSE4.2 (2010 Westmere or later).
* arm64: 64-bit ARMv8-A NEON
* ppc64: 64-bit POWER8 and POWER9 with VSX and ALTIVEC extensions. Both big endian and little endian are implemented, depending on the compiler you are using. The library is tested on recent, little-endian, POWER systems.
* lasx: Loongson Advanced SIMD EXtension (LASX), a 256-bit vector expansion for the LoongArch architecture.
* lsx: Loongson SIMD EXtension (LSX), a 128-bit vector expansion for the LoongArch architecture.
* fallback: A generic implementation that runs on any 64-bit processor.
In many cases, you don't know where your compiled binary is going to run, so simdjson automatically
compiles *all* the implementations into the executable. On Intel, it will include 4 implementations
(icelake, haswell, westmere and fallback), on 64-bit ARM it will include just one since running dispatching is unnecessary, and on PPC
it will include 2 (ppc64 and fallback).
If you know more about where you're going to run and want to save the space, you can disable any of
these implementations at compile time with `-DSIMDJSON_IMPLEMENTATION_X=0` (where X is ICELAKE, HASWELL,
WESTMERE, ARM64, PPC64, LSX, LASX and FALLBACK).
The simdjson library automatically sets header flags for each implementation as it compiles; there
is no need to set architecture-specific flags yourself (e.g., `-mavx2`, `/AVX2` or
`-march=haswell`), and it may even break runtime dispatch and your binaries will fail to run on
older processors. _Note:_ for POWER9 processors make sure you compile it with `-mcpu=power9` and `-mtune=power9` to
get maximum performance.
Runtime CPU Detection
---------------------
When you first use simdjson, it will detect the CPU you're running on, and swap over to the fastest
implementation for it. This is a small, one-time cost and for many people will be paid the first
time they call `parse()` or `load()`.
Inspecting the Detected Implementation
--------------------------------------
You can check what implementation is running with `active_implementation`:
```cpp
cout << "simdjson v" << SIMDJSON_VERSION << endl;
cout << "Detected the best implementation for your machine: " << simdjson::get_active_implementation()->name();
cout << "(" << simdjson::get_active_implementation()->description() << ")" << endl;
```
Implementation detection will happen in this case when you first call `name()`.
Querying Available Implementations
----------------------------------
You can list all available implementations, regardless of which one was selected:
```cpp
for (auto implementation : simdjson::get_available_implementations()) {
cout << implementation->name() << ": " << implementation->description() << endl;
}
```
And look them up by name:
```cpp
cout << simdjson::get_available_implementations()["fallback"]->description() << endl;
```
When an implementation is not available, the bracket call `simdjson::get_available_implementations()[name]`
will return the null pointer.
The available implementations have been compiled but may not necessarily be run safely on your system
see [Checking that an Implementation can Run on your System](#checking-that-an-implementation-can-run-on-your-system).
Manually Selecting the Implementation
-------------------------------------
If you're trying to do performance tests or see how different implementations of simdjson run, you
can select the CPU architecture yourself:
```cpp
// Use the fallback implementation, even though my machine is fast enough for anything
simdjson::get_active_implementation() = simdjson::get_available_implementations()["fallback"];
```
You are responsible for ensuring that the requirements of the selected implementation match your current system.
Furthermore, you should check that the implementation is available before setting it to `simdjson::get_active_implementation()`
by comparing it with the null pointer.
```cpp
auto my_implementation = simdjson::get_available_implementations()["haswell"];
if (! my_implementation) { exit(1); }
if (! my_implementation->supported_by_runtime_system()) { exit(1); }
simdjson::get_active_implementation() = my_implementation;
```
Checking that an Implementation can Run on your System
-------------------------------------
You should call `supported_by_runtime_system()` to compare the processor's features with the need of the implementation.
```cpp
for (auto implementation : simdjson::get_available_implementations()) {
if (implementation->supported_by_runtime_system()) {
cout << implementation->name() << ": " << implementation->description() << endl;
}
}
```
The call to `supported_by_runtime_system()` may be relatively expensive. Do not call `supported_by_runtime_system()` each
time you parse a JSON input (for example). It is meant to be called a handful of times at most in the life of a program.
|
