Float constants are stored as double

[chak]

Due to the design of Language.C (used to represent C AST by the CUDA backend), Float constants are emitted without any indication that they are Floats (and not Doubles). Hence, they are interpreted as Double by nvcc. On devices with compute capability 1.3 and up, those Doubles are converted to Float at runtime (which causes some overhead). On devices below 1.3, nvcc emits an error message of the form

ptxas /tmp/tmpxft_000013d3_00000000-2_dragon3.ptx, line 87; warning : Double is not supported. Demoting to float

[seanl]

The patch for this ticket has been submitted.