-arch sm_13 vs -arch sm_20 (sm_20 slower on C2050)

[njuffa]

Correct, -use_fast_math implies -ftz=true -prec-div=false -prec-sqrt=false. I will followup with the documentation team to check where this is documented, or if it’s not documented yet, get it properly documented for the next CUDA release.

[njuffa]

The only documentation I could find is via nvcc --help (see below). I will file a bug to get this included in the appropriate document for the next version of CUDA. Thanks for alerting us to this documentation issue.

–ftz [true,false] (-ftz)

    When performing single-precision floating-point operations, flush 

    denormal values to zero or preserve denormal values. -use_fast_math 

    implies --ftz=true.

    Default value:  0.

–prec-div [true,false] (-prec-div)

    For single-precision floating-point division and reciprocals, use IEEE 

    round-to-nearest mode or use a faster approximation. -use_fast_math 

    implies --prec-div=false.

    Default value:  1.

–prec-sqrt [true,false] (-prec-sqrt)

    For single-precision floating-point square root, use IEEE 

    round-to-nearest mode or use a faster approximation. -use_fast_math 

    implies --prec-sqrt=false.

    Default value:  1.