I actually have a couple of very specific questions.

1) In general, nowadays are games compiled with fp:/fast or with fp:/precise flags turned on?

2) And what about games with large worlds, that usually have to deal with floating accuracy issues due to long distances from the camera: in those cases, the usual tricks like floating-origin and using doubles for calculating changes in transforms before reconverting back to float are used together with fp:/fast or even with them fp:/is needed?

Thanks for your insights.

Yeah if I was doing a large world, I'd use a floating origin for local coordinates, perhaps combined with 64bit integer / fixed point global coordinates. Using double precision anywhere is often a code smell honestly :wink:

I used to think like this as well, but most recently I changed my opinion on this. There are a places where doubles really make sense. In particular for large worlds and numerical sensitive operations double is the right choice in my opinion. Why 64bit integer / fixed point and not just double. What does this buy you? Because of memory? On modern machines double performance is nearly as good as float from my experience. I would argue if doubles are an option for your problem on your target machine use them.

I also spent quite some time implementing robust versions of Quickhull and Mesh Simplification (QEM) which needed to compile our entire asset database (15 years of models!) and I did it in 32 bit floating point. What a waste of time! With doubles it all just works and with float32 I spend lots of time debugging and tweaking epsilons. Sure I gained some great insight into the algorithms and numerical analysis from this, but in retrospect I could have spend this time with much more valuable things.

so sure, for anything smaller than that, "just use double" will work just fine

This is for precision. But what about quickness ? I highly suppose that CPUs can still compute 32-bits floating point calculations faster than with 64-bits floating points. Even on 64-bits architectures. Am I wrong ?

Also, as far as I remember, floats ruled on the GPU some years ago. Do you think that nowadays GPUs can operate on 64-bits float as fast as 32-bits floats ?

so sure, for anything smaller than that, "just use double" will work just fine

This is for precision. But what about quickness ? I highly suppose that CPUs can still compute 32-bits floating point calculations faster than with 64-bits floating points. Even on 64-bits architectures. Am I wrong ?

Also, as far as I remember, floats ruled on the GPU some years ago. Do you think that nowadays GPUs can operate on 64-bits float as fast as 32-bits floats ?

On GPU, double is possible these days but is a big no-no for performance reasons. IIRC on some GPU's it's 4x slower than float.

On x86, double and float are the same speed with fp:/fast (as the FPU internally uses 80-bit floating point) and float is actually slower with fp:/strict (not accounting for memory bandwidth -- in bandwidth-bottlenecked situations, float is twice as fast as double)!

On x64, double and float are the same speed... unless your code is vectorizable (SIMD-able, SSE/AVX...), in which case float can be twice as fast.

Thank you a lot. So these things did not improve that much on the GPU side since the years 2000...