Default floating point compiler options

Does anyone know what the default floating point compiler flags are for CUDA 6.5 compiler? Fast or IEEE 754? I’m having some floating point inconsistencies during convolutions using a model generated on MacOS.

Hi,

Default flag is IEEE 754 but fast mode is also available for nvcc.

  • 754 mode (default):
    • -ftz=false
    • -prec-div=true
    • -prec-sqrt=true
  • fast mode:
    • -ftz=true
    • -prec-div=false
    • -prec-sqrt=false

Please find more information in our document here:
http://docs.nvidia.com/cuda/floating-point/index.html

Thanks.