在C#中学习如何使用SIMD
前言
在前面说 在C#中如何提高Linq的性能 ,其中就提到在.Net 8之前的版本(.Net Core 3.0之后的版本,)可以通过使用SimdLinq库来提高性能.SimdLinq源码还是比较轻量级的,因为轻量级只提供了以下方法的支持:- Sum(支持的类型: int, uint, long, ulong, float, double)
- LongSum(支持的类型: int, uint)
- Average(支持的类型: int, uint, long, ulong, float, double)
- Min(支持的类型: byte, sbyte, short, ushort, int, uint, long, ulong, float, double)
- Max(支持的类型: byte, sbyte, short, ushort, int, uint, long, ulong, float, double)
- MinMax(支持的类型: byte, sbyte, short, ushort, int, uint, long, ulong, float, double)
- Contains(支持的类型: byte, sbyte, short, ushort, int, uint, long, ulong, float, double)
- SequenceEqual(支持的类型:byte, sbyte, short, ushort, int, uint, long, ulong, float, double)
支持的集合有: T[], List<T>, Span<T>, Memory<T>, ReadOnlyMemory<T>, Span<T>, ReadOnlySpan<T>.
理解SIMD(单指令多数据)
这里从汇编代码来说SIMD的.extern printf
section .data
dummy db 13
align 16
;pdivector1相当于数组
pdivector1 dd 1
dd 2
dd 3
dd 4
;pdivector2相当于数组
pdivector2 dd 5
dd 6
dd 7
dd 8
fmt db "Sum Vector:%d %d %d %d",10,0
section .bss
alignb 16
pdivector_res resd 4
section .text
global main
main:
;序言
push rbp
mov rbp,rsp
;*****将pdivector1,加载到xmm0寄存器中*****
movdqa xmm0,[pdivector1]
;*****将pdivector2和xmm0寄存器内的值,进行加法运算***
paddd xmm0,[pdivector2]
;将结果保存在内存中
movdqa [pdivector_res],xmm0
;打印内存中的向量
mov rsi,pdivector_res
mov rdi,fmt
call printpdi
;尾言
mov rsp,rbp
pop rbp
ret
;打印-----------------------------------
printpdi:
push rbp
mov rbp,rsp
movdqa xmm0,[rsi]
;从xmmo0中提取打包的值
pextrd esi,xmm0,0
pextrd edx,xmm0,1
pextrd ecx,xmm0,2
pextrd r8d,xmm0,3
;没有浮点数
mov rax,0
call printf
mov rsp,rbp
pop rbp
ret
重点就这两行代码:
;*****将pdivector1,加载到xmm0寄存器中*****
movdqa xmm0,[pdivector1]
;*****将pdivector2和xmm0寄存器内的值,进行加法运算***
paddd xmm0,[pdivector2]
接着看SimdLinq源码
先看一下SimdLinq源码目录:
在学习源码的时候,只需要关心文件名带有Core的源码,包含Core源码就是具体实现.
接着看一下Sum的源码:
static T SumCore<T>(ReadOnlySpan<T> source)
where T : struct, INumber<T>
{
T sum = T.Zero;
if (!Vector128.IsHardwareAccelerated || source.Length < Vector128<T>.Count)
{
// Not SIMD supported or small source.
//1. 当硬件不支持,会退变为for循环
//2. 集合内的数量小于Vector128的支持的数量,如int->4 long->2,会退变for循环
unchecked // SIMD operation is unchecked so keep same behaviour
{
for (int i = 0; i < source.Length; i++)
{
sum += source[i];
}
}
}
else if (!Vector256.IsHardwareAccelerated || source.Length < Vector256<T>.Count)
{
// Only 128bit SIMD supported or small source.
//满足128bit,不足256bit的,数量少的时候
//获取开始元素
ref var begin = ref MemoryMarshal.GetReference(source);
//获取结尾的元素
ref var last = ref Unsafe.Add(ref begin, source.Length);
ref var current = ref begin;
//获取一个初始值为0的Vector128
var vectorSum = Vector128<T>.Zero;
//集合的长度减去Vector128<T>的数量,让开始元素进行偏移
ref var to = ref Unsafe.Add(ref begin, source.Length - Vector128<T>.Count);
//开始元素的地址是否和小于一次Vector的地址
while (Unsafe.IsAddressLessThan(ref current, ref to))
{
//如果是int类型, 就一次加载前4个元素
vectorSum += Vector128.LoadUnsafe(ref current);
current = ref Unsafe.Add(ref current, Vector128<T>.Count); //如果是int类型,偏移4个元素
}
//判断current的地址是否小于结尾元素的地址
//处理不够一次Vector的时候,退变循环处理
while (Unsafe.IsAddressLessThan(ref current, ref last))
{
unchecked // SIMD operation is unchecked so keep same behaviour
{
sum += current;
}
current = ref Unsafe.Add(ref current, 1); //每次偏移1个元素
}
sum += Vector128.Sum(vectorSum); //进行求和计算
}
else
{
// 256bit SIMD supported
//Vector256就不注释了
ref var begin = ref MemoryMarshal.GetReference(source);
ref var last = ref Unsafe.Add(ref begin, source.Length);
ref var current = ref begin;
var vectorSum = Vector256<T>.Zero;
ref var to = ref Unsafe.Add(ref begin, source.Length - Vector256<T>.Count);
while (Unsafe.IsAddressLessThan(ref current, ref to))
{
vectorSum += Vector256.LoadUnsafe(ref current);
current = ref Unsafe.Add(ref current, Vector256<T>.Count);
}
while (Unsafe.IsAddressLessThan(ref current, ref last))
{
unchecked // SIMD operation is unchecked so keep same behaviour
{
sum += current;
}
current = ref Unsafe.Add(ref current, 1);
}
sum += Vector256.Sum(vectorSum);
}
return sum;
}
秋风
2024-03-05