Open drahnr opened 7 years ago
This would allow the individual backend to decide what accuracy would fit best.
On CPU i.e. there is no gain in using fp16 over fp32, on the GPU - if supported - there is. So this decision is backend specific.
This would allow the individual backend to decide what accuracy would fit best.
On CPU i.e. there is no gain in using fp16 over fp32, on the GPU - if supported - there is. So this decision is backend specific.