software implementation of double prec math?

Accelerated Computing CUDA CUDA Programming and Performance
1

Has anyone thought about just making (or knows of?) a software implementation of double precision arithmetic for
the non-DP chips? I know it would be slow, but I think a lot of regular CPUs did DP in microcode “back in the day”. It would
be good for testing, seeing that there are not DP Nvidia chips for laptops yet. You wouldn’t necessarily need the full complement of fp operations, maybe just the simplest IEEE ones to start:
add mult div

2

I am pretty sure that someone had ported at least some of the routines from DSFUN90 to CUDA in the pre GT200 era, although I am struggle to remember the who, where and when.

3

Quite a few of the functions are in the mandelbrot example program in the development kit, they are a lot slower than single precision, but a reasonable graphics card will still run faster than a CPU (well my GT220 does…

4

I’m pretty sure DSFUN90 doesn’t actually completely emulate full double precision. It is (approximately) the equivalent of a floating point number with a doubled mantissa, but with the same exponent as normal floats (8 bits).

Single = 23e8

Double = 52e11

emulated double = ~46e8

Full emulation of a double would be waaay too expensive (especially with things like denormals).

5

This is the thread you want.

6

Thanks, this is great.