From DoubleTensor to cuda.FloatTensor

Kiuhnm_Mnhuik · April 12, 2018, 9:08pm

I’m using float64 values on the cpu but I want to use float32 values on the gpu. What’s the fastest way to convert a DoubleTensor to a cuda.FloatTensor?

ptrblck · April 12, 2018, 9:14pm

torch.DoubleTensor(10).float().cuda() would be the way to go.
It would be faster, if you could load your data as floats, if that’s possible.

Kiuhnm_Mnhuik · April 12, 2018, 9:20pm

I’m doing RL so the data is generated as float64 for stability reasons. I don’t think this conversion will be a bottleneck but was wondering if there was a single instruction for this. It appears there isn’t.

ptrblck · April 12, 2018, 9:25pm

Ah ok, I understand.
Well you could use

a = torch.DoubleTensor(10)
a = a.type(torch.cuda.FloatTensor)

, if you would like to have a single command.

I tried to time it with torch.cuda.synchronize() and got mixed results.
Currently my GPU is busy, so the timing is most likely useless in this state.

Kiuhnm_Mnhuik · April 12, 2018, 10:40pm

Nice, I’ll try that!