F# without the allocs

[cartermp]

• Moves to .NET Core
• Marks various functions inline
• Value types instead of reference types, since most of the underlying data is just numeric primitives
• Adds a lot more to the .gitignore

These are the comparative results on my machine, which is different but I'd expect a similar improvement in timings:

Creating a scene:

BenchmarkDotNet=v0.12.1, OS=Windows 10.0.18363.720 (1909/November2018Update/19H2)
Intel Core i7-6700K CPU 4.00GHz (Skylake), 1 CPU, 8 logical and 4 physical cores
.NET Core SDK=3.1.300-preview-015048
  [Host]     : .NET Core 3.1.3 (CoreCLR 4.700.20.11803, CoreFX 4.700.20.12001), X64 RyuJIT DEBUG
  DefaultJob : .NET Core 3.1.3 (CoreCLR 4.700.20.11803, CoreFX 4.700.20.12001), X64 RyuJIT

Method	Mean	Error	StdDev	Ratio	Gen 0	Gen 1	Gen 2	Allocated
Original	438.6 μs	7.10 μs	6.29 μs	1.00	329.1016	23.4375	-	1333.62 KB
Modified	178.0 μs	3.53 μs	5.06 μs	0.41	146.7285	0.4883	-	558.04 KB

Rendering:

BenchmarkDotNet=v0.12.1, OS=Windows 10.0.18363.720 (1909/November2018Update/19H2)
Intel Core i7-6700K CPU 4.00GHz (Skylake), 1 CPU, 8 logical and 4 physical cores
.NET Core SDK=3.1.300-preview-015048
  [Host]     : .NET Core 3.1.3 (CoreCLR 4.700.20.11803, CoreFX 4.700.20.12001), X64 RyuJIT DEBUG
  DefaultJob : .NET Core 3.1.3 (CoreCLR 4.700.20.11803, CoreFX 4.700.20.12001), X64 RyuJIT

Method	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
Original	178.68 ms	3.550 ms	9.718 ms	1.00	0.00	275000.0000	46000.0000	1000.0000	1095.79 MB
Modified	63.01 ms	0.768 ms	0.681 ms	0.35	0.02	2555.5556	111.1111	111.1111	10.8 MB

Both seem to represent a nice improvement. However, there are likely more things to be explored:

• Using IsByRefLike structs and the Span<T> type to elide bounds checking
• Fiddling with more advanced .NET perf stuff

[athas]

Neat! Thanks!

[athas]

I don't have .NET Core 3.1 on this machine, but 2.1 seems to also run this pretty fast.

[cartermp]

If you upgrade to 3.1 there should be some improvements as well. Lots of little fixes in the standard library for .NET that make things run transparently faster. But I wouldn't expect a dramatic improvement 🙂

[battlebottle]

@cartermp you bet me to it ;). I believe changing the Pixel tuple to a struct also improves performance, though I did not test this thoroughly.

[cartermp]

@battlebottle The pixel tuple should be a struct here as well, unless I'm missing something: https://github.com/athas/raytracers/blob/master/fsharp/ray.fs#L281

[battlebottle]

@cartermp Ah you're right, my mistake. I misread the commit somehow 😩