更新：之前提到的跨度问题已在.net核心2.1版本(目前正在预览中)中修复.这些实际上使得Span Vector *比数组Vector更快* …
注意：在“Intel Xeon E5-1660 v4”上进行测试,CPU-Z告诉我有“MMX,SSE,SSE2,SSE3,SSSE3,SSE4.1,SSE4.2,EM64T,VT-x,AES, AVX,AVX2,FMA3,RSX“所以它应该没问题……
在回答Vector based question之后,我想我会尝试实现一些BLAS功能.我发现那些正在阅读/求和的产品如dot产品非常好,但是我回写一个阵列是坏的 – 比非SIMD更好,但几乎没有.
我做错了什么,或者是否需要在JIT中做更多的工作？
示例(假设x.Length = y.Length,not null等等等等等等)：
public static void daxpy(double alpha, double[] x, double[] y)
{
for (var i = 0; i < x.Length; ++i)
y[i] = y[i] + x[i] * alpha;
}
在矢量形式变为：
public static void daxpy(double alpha, double[] x, double[] y)
{
var i = 0;
if (Vector.IsHardwareAccelerated)
{
var length = x.Length + 1 - Vector<double>.Count;
for (; i < length; i += Vector<double>.Count)
{
var valpha = new Vector<double>(alpha);
var vx = new Vector<double>(x, i);
var vy = new Vector<double>(y, i);
(vy + vx * valpha).CopyTo(y, i);
}
}
for (; i < x.Length; ++i)
y[i] = y[i] + x[i] * alpha;
}
而且,在.NET Core 2.0中玩游戏,虽然我会尝试Span,无论是天真还是矢量形式：
public static void daxpy(double alpha, Span<double> x, Span<double> y)
{
for (var i = 0; i < x.Length; ++i)
y[i] += x[i] * alpha;
}
和矢量
public static void daxpy(double alpha, Span<double> x, Span<double> y)
{
if (Vector.IsHardwareAccelerated)
{
var vx = x.NonPortableCast<double, Vector<double>>();
var vy = y.NonPortableCast<double, Vector<double>>();
var valpha = new Vector<double>(alpha);
for (var i = 0; i < vx.Length; ++i)
vy[i] += vx[i] * valpha;
x = x.Slice(Vector<double>.Count * vx.Length);
y = y.Slice(Vector<double>.Count * vy.Length);
}
for (var i = 0; i < x.Length; ++i)
y[i] += x[i] * alpha;
}
所以这些的相对时间是：
Naive 1.0
Vector 0.8
Span Naive 2.5 ==> Update: Span Naive 1.1
Span Vector 0.9 ==> Update: Span Vector 0.6
我做错了什么？我几乎无法想到一个更简单的例子,所以我不这么认为？
最佳答案 你可能想用2.1以上的测试;
在我的笔记本电脑上(SIMD与我的桌面相比较差),我得到：
daxpy_naive x10000: 144ms
daxpy_arr_vector x10000: 77ms
daxpy_span x10000: 173ms
daxpy_vector x10000: 67ms
daxpy_vector_no_slice x10000: 67ms
使用代码：
using System;
using System.Diagnostics;
using System.Numerics;
class Program
{
static void Main(string[] args)
{
double alpha = 0.5;
double[] x = new double[16 * 1024], y = new double[x.Length];
var rand = new Random(12345);
for (int i = 0; i < x.Length; i++)
x[i] = rand.NextDouble();
RunAll(alpha, x, y, 1, false);
RunAll(alpha, x, y, 10000, true);
}
private static void RunAll(double alpha, double[] x, double[] y, int loop, bool log)
{
GC.Collect(GC.MaxGeneration);
GC.WaitForPendingFinalizers();
var watch = Stopwatch.StartNew();
for(int i = 0; i < loop; i++)
{
daxpy_naive(alpha, x, y);
}
watch.Stop();
if (log) Console.WriteLine($"{nameof(daxpy_naive)} x{loop}: {watch.ElapsedMilliseconds}ms");
watch = Stopwatch.StartNew();
for (int i = 0; i < loop; i++)
{
daxpy_arr_vector(alpha, x, y);
}
watch.Stop();
if (log) Console.WriteLine($"{nameof(daxpy_arr_vector)} x{loop}: {watch.ElapsedMilliseconds}ms");
watch = Stopwatch.StartNew();
for (int i = 0; i < loop; i++)
{
daxpy_span(alpha, x, y);
}
watch.Stop();
if (log) Console.WriteLine($"{nameof(daxpy_span)} x{loop}: {watch.ElapsedMilliseconds}ms");
watch = Stopwatch.StartNew();
for (int i = 0; i < loop; i++)
{
daxpy_vector(alpha, x, y);
}
watch.Stop();
if (log) Console.WriteLine($"{nameof(daxpy_vector)} x{loop}: {watch.ElapsedMilliseconds}ms");
watch = Stopwatch.StartNew();
for (int i = 0; i < loop; i++)
{
daxpy_vector_no_slice(alpha, x, y);
}
watch.Stop();
if (log) Console.WriteLine($"{nameof(daxpy_vector_no_slice)} x{loop}: {watch.ElapsedMilliseconds}ms");
}
public static void daxpy_naive(double alpha, double[] x, double[] y)
{
for (var i = 0; i < x.Length; ++i)
y[i] = y[i] + x[i] * alpha;
}
public static void daxpy_arr_vector(double alpha, double[] x, double[] y)
{
var i = 0;
if (Vector.IsHardwareAccelerated)
{
var length = x.Length + 1 - Vector<double>.Count;
for (; i < length; i += Vector<double>.Count)
{
var valpha = new Vector<double>(alpha);
var vx = new Vector<double>(x, i);
var vy = new Vector<double>(y, i);
(vy + vx * valpha).CopyTo(y, i);
}
}
for (; i < x.Length; ++i)
y[i] = y[i] + x[i] * alpha;
}
public static void daxpy_span(double alpha, Span<double> x, Span<double> y)
{
for (var i = 0; i < x.Length; ++i)
y[i] += x[i] * alpha;
}
public static void daxpy_vector(double alpha, Span<double> x, Span<double> y)
{
if (Vector.IsHardwareAccelerated)
{
var vx = x.NonPortableCast<double, Vector<double>>();
var vy = y.NonPortableCast<double, Vector<double>>();
var valpha = new Vector<double>(alpha);
for (var i = 0; i < vx.Length; ++i)
vy[i] += vx[i] * valpha;
x = x.Slice(Vector<double>.Count * vx.Length);
y = y.Slice(Vector<double>.Count * vy.Length);
}
for (var i = 0; i < x.Length; ++i)
y[i] += x[i] * alpha;
}
public static void daxpy_vector_no_slice(double alpha, Span<double> x, Span<double> y)
{
int i = 0;
if (Vector.IsHardwareAccelerated)
{
var vx = x.NonPortableCast<double, Vector<double>>();
var vy = y.NonPortableCast<double, Vector<double>>();
var valpha = new Vector<double>(alpha);
for (i = 0; i < vx.Length; ++i)
vy[i] += vx[i] * valpha;
i = Vector<double>.Count * vx.Length;
}
for (; i < x.Length; ++i)
y[i] += x[i] * alpha;
}
}
这是使用dotnet build -c Release和dotnet run -c Release,dotnet –version报告“2.2.0-preview1-008000”(不久之前的“每日”).
在我的桌面上,我希望差异会更好.
