Float32 function return and Cuda

General Usage
, , ,
1

I have GPU(GTX 1050) that performance for Float32 compared to Float64 is about 32 times better, so I should use Float32 values. I have function that return values in format of Float64(or Tuple of Float64). Function looks like that.

function FunctionName(x,y,z,...) 
   firstValue=...
   secondValue=...
   ...
   return firstValue, secondValue
end

and it returns values in format of Float64. All input data of this function are Float32 datatypes.
How to get Float32 return values?

2

Do you have a literal vaue like 0.0 in your code? It sounds like that is causing a promotion. Try using @code_warntype or Traceur.jl or the debugger to find where the promotion to Float64 occurs.

2 Likes
3

you can always convert back to Float32 but as dpsanders said, it’s better to debug and see where the promotion happens – usually when you do operation with some literal Float64 value (either ± or */)

4

The problem was, that in this function I created some helpful variable: buffer=[sin(x), cos(x),0.0], which was later multiplied by some input data. I corrected with: buffer=Float32[sin(x), cos(x),0.0] and everything works.

1 Like