.half? PTX code (i.e. mov.half.b32 or mul.half.b32)

Accelerated Computing CUDA CUDA Programming and Performance
1

just reading through some ptx code generated from decuda and i see .half a lot

examples:
rcp.half.f32 $r3, $r3
mul.half.rn.f32 $r10, $r5, $r3
mov.half.b32 $r3, $r4

does it mean it just operates on the lower 2 bytes? or does it do the full 32 bit operation but only returns 2 bytes
and w/ moving is it just transferring 2 bytes? am i correct to assume it defaults to the lower 2 bytes?

2

It just means that the instruction word is encoded in 32-bit instead of the usual 64-bit (which you can easily check by looking at the hex dump). It has no influence on behavior, nor performance (except a lower instruction cache footprint).

3

ah i see, thanks Sylvain

4

I do not think that this is actually ptx. I believe that decude actually produces something form of G80 assembly code