Notifications
Article
Daily Pathtracer!安利下不错的Pathtracer学习资料
Published 3 months ago
402
0
Daily Pathtracer!安利下不错的Pathtracer学习资料

0x00 前言

最近看到了我司大网红aras-p的博客开了一个很有趣的新系列《Daily Pathtracer》,来实现一个简单的ToyPathTracer。
除了使用C++,aras还使用了两种不同的C#运行时来实现,即.NET Standard和Xamarin/Mono。效率上自然C++要强于.NET Standard,而.NET Standard的表现要强于Xamarin/Mono。
除此之外,aras当然还会用到Unity来实现,比较有意思的是Unity的实现中使用了2018中的Burst来提升性能,而结果甚至有点小惊人呢,之后可以再说。
而aras的这个小巧的ToyPathTracer事实上和最近在知乎上小火的小书《Ray Tracing in One Weekend》还有些关系。其实这是一套3本中的第一本书,另外还有两本,分别叫《Ray Tracing: the Next Week》和《Ray Tracing: The Rest Of Your Life》。
一个周末不够,还要一周,最后甚至还要赔上余生。嗯,真是一入码门深似海。而图形学作为程序员的三大浪漫之一,真是诚不我欺也。
当然,这本书的作者Peter Shirley已经将此书作为PDF免费放出了, 但是各位如果有兴趣的话建议还是买一下吧(seriously, just buy that minibook)。
所以前言部分先安利的,是这套小书。

0x01 Unity中实现PathTracer

aras的这一系列博客很长,目录摘录在下面。
  • Part 0: Intro
  • Part 1: Initial C++ and walkthrough
  • Part 2: Fix stupid performance issue
  • Part 3: C#, Unity and Burst
  • Part 4: Correctness fixes and Mitsuba
  • Part 5: simple GPU version via Metal
  • Part 6: simple GPU version via D3D11
  • Part 7: initial C++ SIMD & SoA
  • Part 8: SSE SIMD for HitSpheres
  • Part 9: ryg optimizes my code
  • Part 10: Update all implementations to match
  • Part 11: Buffer-oriented approach on CPU
  • Part 12: Buffer-oriented approach on GPU D3D11
  • Part 13: GPU thread group data optimization
我的这篇小文作为安利文,也就不详细介绍了每一篇的内容了,相信有兴趣的小伙伴都会去看的。
但是其中有一篇文章还是蛮有趣的,因为对比了不同语言、不同运行时的性能差异,而且还演示了一下Unity2018中的Burst compiler对C#代码性能的提升。所以稍微介绍下这部分吧。
当然,首先可以看到,场景是作为硬编码写死的:
static Sphere[] s_SpheresData = { new Sphere(new float3(0,-100.5f,-1), 100), new Sphere(new float3(2,0,-1), 0.5f), new Sphere(new float3(0,0,-1), 0.5f), new Sphere(new float3(-2,0,-1), 0.5f), new Sphere(new float3(2,0,1), 0.5f), new Sphere(new float3(0,0,1), 0.5f), new Sphere(new float3(-2,0,1), 0.5f), new Sphere(new float3(0.5f,1,0.5f), 0.5f), new Sphere(new float3(-1.5f,1.5f,0f), 0.3f), }; static Material[] s_SphereMatsData = { new Material(Material.Type.Lambert, new float3(0.8f, 0.8f, 0.8f), new float3(0,0,0), 0, 0), new Material(Material.Type.Lambert, new float3(0.8f, 0.4f, 0.4f), new float3(0,0,0), 0, 0), new Material(Material.Type.Lambert, new float3(0.4f, 0.8f, 0.4f), new float3(0,0,0), 0, 0), new Material(Material.Type.Metal, new float3(0.4f, 0.4f, 0.8f), new float3(0,0,0), 0, 0), new Material(Material.Type.Metal, new float3(0.4f, 0.8f, 0.4f), new float3(0,0,0), 0, 0), new Material(Material.Type.Metal, new float3(0.4f, 0.8f, 0.4f), new float3(0,0,0), 0.2f, 0), new Material(Material.Type.Metal, new float3(0.4f, 0.8f, 0.4f), new float3(0,0,0), 0.6f, 0), new Material(Material.Type.Dielectric, new float3(0.4f, 0.4f, 0.4f), new float3(0,0,0), 0, 1.5f), new Material(Material.Type.Lambert, new float3(0.8f, 0.6f, 0.2f), new float3(30,25,15), 0, 0),
当然需要来判断是否相交的:
static bool HitWorld(Ray r, float tMin, float tMax, ref Hit outHit, ref int outID, ref SpheresSoA spheres) { outID = spheres.HitSpheres(ref r, tMin, tMax, ref outHit); return outID != -1; }
光线的追踪:
static float3 Trace(Ray r, int depth, ref int inoutRayCount, ref SpheresSoA spheres, NativeArray<Material> materials, ref uint randState, bool doMaterialE = true) { Hit rec = default(Hit); int id = 0; ++inoutRayCount; if (HitWorld(r, kMinT, kMaxT, ref rec, ref id, ref spheres)) { Ray scattered; float3 attenuation; float3 lightE; var mat = materials[id]; var matE = mat.emissive; ...
物体表面如何处理接收到的光线呢?可以看这里:
static bool Scatter(Material mat, Ray r_in, Hit rec, out float3 attenuation, out Ray scattered, out float3 outLightE, ref int inoutRayCount, ref SpheresSoA spheres, NativeArray<Material> materials, ref uint randState) { outLightE = new float3(0, 0, 0); if (mat.type == Material.Type.Lambert) { // random point inside unit sphere that is tangent to the hit point float3 target = rec.pos + rec.normal + MathUtil.RandomUnitVector(ref randState); scattered = new Ray(rec.pos, normalize(target - rec.pos)); attenuation = mat.albedo; // sample lights ...
最后,在Unity中使用了burst,并分发了任务:
[ComputeJobOptimization] struct TraceRowJob : IJobParallelFor { public int screenWidth, screenHeight, frameCount; [NativeDisableParallelForRestriction] public NativeArray<UnityEngine.Color> backbuffer; public Camera cam; [NativeDisableParallelForRestriction] public NativeArray<int> rayCounter; [NativeDisableParallelForRestriction] public SpheresSoA spheres; [NativeDisableParallelForRestriction] public NativeArray<Material> materials; public void Execute(int y) { int backbufferIdx = y * screenWidth; float invWidth = 1.0f / screenWidth; float invHeight = 1.0f / screenHeight; float lerpFac = (float)frameCount / (float)(frameCount + 1); ...
通过UnityProfiler,我们可以看到经过Burst优化后的运行状态:
不过让我感到比较吃惊的是,在Unity2018中使用C# + Burst时的效率竟然超过了C++的实现,看来Burst的效率还是很惊人的。
下面是aras那里的结论:
  • .NET Core is about 2x slower than vanilla C++.
  • Mono (with default settings) is about 3x slower than .NET Core.
  • IL2CPP is 2x-3x faster than Mono, which is roughly .NET Core performance level.
  • Unity’s Burst compiler can get our C# code faster than vanilla C++. Note that right now Burst is very early tech, I expect it will get even better performance later on.
当然,aras的博客中有更详细的对比数据。
所以这部分安利的,是aras的博客系列。

0x02 小结

当然,aras除了写了博客之外,也开源了这个ToyPathTracer(也许先有了ToyPathTracer才有了这系列的博客也未可知)。
所以各位可以去github上clone一份代码,除了可以分别在windows和mac上编译c++、c#的实现,也可以直接在unity中使用。
所以最后安利一下这个工程。
怎么样,是不是还挺有趣的?

Ref

《Ray Tracing in One Weekend》
https://drive.google.com/drive/folders/14yayBb9XiL16lmuhbYhhvea8mKUUK77W
《Daily Pathtracer》
http://aras-p.info/blog/2018/03/28/Daily-Pathtracer-Part-0-Intro/
《ToyPathTracer》
https://github.com/aras-p/ToyPathTracer

Tags:
Jiadong Chen 陈嘉栋
Field Engineer - Programmer
24
Comments