Port PacketTracer benchmarks

[adamsitnik]

Fixes #113

Please see the commits history for what I have copied and what I have modified.

This is the only and the last missing CoreCLR benchmark

[jorive]

@adamsitnik Would it make more sense to place this benchmark under a SoA.RayTracer namespace (what is attempting to measure and what it is doing)?

[jorive]

This benchmark only runs on AVX2 and not in netcoreapp3.0? Is there a way to not bother building on non-x86* architecture?

[adamsitnik]

@jorive it looks like we have what we want - this benchmark won't be built/run on ARM. Do you want me to rename the namespace/type/method name as suggested?

I have also sent a PR to BDN that will make it much easier for us in the future: dotnet/BenchmarkDotNet#1097

[jorive]

@adamsitnik Thanks Adam. I would prefer to rename the namespace/class/method (from PacketTracer/Render/Run to something like StructureOfArray/RayTracer/Render), but maybe @billwert have a different opinion.

[adamsitnik]

I think that this explains why the author of the benchmark did not add a namespace to it in the first place, the naming is just hard ;)

My proposal: HardwareIntrinsics/RayTracer/Render

@jorive what do you say?

[adamsitnik]

@jorive @billwert which name should we use? ;)

[billwert]

I think if we believe these benchmarks are solely for the purpose of testing hardware intrinsicics, we should go with HardwareIntrinsics/RayTracer/Render. If we think there is value in keeping PacketTracer for name recognition perhaps PacketTracer/HardwareIntrinsics/Render?

[jorive]

The benchmark is more about data storage/layout than "HardwareIntrinsics". The benchmarks is more about: Data Layout > StructureOfArray > RayTracer > Render

[adamsitnik]

The benchmark is more about data storage/layout than "HardwareIntrinsics".

I think that the original concept behind it was to test the performance of the features exposed by the new Hardware Intrinsics API but using a SoA algorithm.

@fiigii as the author of this benchmark, what name do you recommend for it? I know it sounds silly ;)

[fiigii]

what name do you recommend for it?

Packet tracing is a convention name in the graphics area that indicates the SoA vectorized ray tracing algorithm (but this benchmark is a simplified version without BVH). I am okay to rename it in this dotnet repo. Maybe HardwareIntrinsics/RayTracer/SoA? That gives us room to add other algorithms under /RayTracer/ in the future.

The benchmark is more about data storage/layout than "HardwareIntrinsics".

Originally, I created this benchmark for hardware intrinsic to leverage the 256-bit SIMD vectors (comparing to the AoS ray-tracer we had for benchmarking S.N.Vector<T>). In .NET Core, it is impossible to implement this program efficiently in S.N.Vector<T>, please see Transpose, public unsafe VectorPacket256(float* vectors), and methods in VectorMath.cs. So, I think it is definitely about hardware intrinsic.

[jorive]

Namespace: HardwareIntrinsics/RayTracer/
class/method: SoA/Render
?

[adamsitnik]

@fiigii thanks for help!

@jorive thanks for the suggestion! HardwareIntrinsics.RayTracer.SoA.Render is the new full benchmark name