float/int question

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,
I have a kernel that calculates a float value and another int value and writes it to gmem (actually an array of such vallues)
In another kernel I then read these values and use it.
I figured that the float value, as per the algorithm, is alway between zero and one. I thought to merge the two values into the float value like this:

First kernel:
float fValue = 1.f * IMUL(…);
pGMEMCalc[ threadId.x ] = fValue + fCalculatedFloatValue;

Second kernel:
float fValue = pGMEMCalc[ threadIdx.x ];
int iIntValue = ( int )floor( fValue );
float fFloatValue = fValue - floor( fValue );

However the results are not the same as when I use two seperate arrays. Any suggestions? Is it related to floating point arithmetic order or such a thing?

thanks
eyal

2

From your example it is not clear how large is the range of the integer part.

In any case you should bear in mind that the floats have about 24 bits of precision, so,
if you are using some (maybe all) of those bits for storing the integer value, it is clear
that you are losing bits to store the float itself, which cannot thus recovered.

I make an example with decimal digits instead of bits:
Assume you have just 5 (decimal) digits of precision.
you can store a number like 0.54331 which has significant 5 digits.
Now, if you try to pack also the integer number 788 you obtain
788.54 because, with just 5 digits of precision the 0.00331 gets lost forever.

If what I’m saying has no meaning for you, then it is a good time to read a tutorial
on the representation of floating point numbers in computers.

giovanni.

3

Hi,

Thanks for the quick response :)

The int part should be between 0 and 10,000. And the float accuracy I need is something like 5-7 digits after the dot.

Your explaination probably explains what I see.

How do I know how many digits I have to spare in the GPU?

thanks

eyal

4

I’m a totally newbie on CUDA, but the standard of floating point for 32 bits says that 24 bit are for the mantissa,

so you have about 24 bits of precision which means a little more than 7 decimal digits of precision, roughly

speaking.

This also means you cannot pack a number up to 10000 together with a float with 5 significant digits…

If you plan to program a computer and use float or double numbers, then you should really

read a very interesting article called

What every scientist should know about floating-point arithmetic:

http://www.physics.ohio-state.edu/~dws/gro…_point_math.pdf

giovanni