Initialising constant memory

Accelerated Computing CUDA CUDA Programming and Performance
1

When I try the following:

__constant__ float smoothingLength = h_epsilon*powf(particleMass/referenceDensity,(1.0f/3.0f));

It gives me the following compilation error:

/usr/include/c++/4.1.3/cstdlib(178): internal error: can't generate code for non empty constructors or destructors on device

1 catastrophic error detected in the compilation of "/tmp/tmpxft_0000235b_00000000-4_sphuniformgrid.cpp1.ii".

Compilation aborted.

Aborted (core dumped)

Is there any particular reason I can’t do this even though particleMass and referenceDensity are also constant values (written before the smoothingLength constant).

Obviously this would work in regular C, so I wanna know if there’s a workaround. (Other than to not put it into constant memory :P)

thx

2

constant variables have to be filled with cudaMemCpyToSymbol as far as I know (they are in constant memory)

if h_epsilon is not a constant value, I can imagine that this will not work btw…

3

h_epsilon is in fact also constant :)

I’ll check out the cudaMemCpyToSymbol, but really I was hoping the compiler would take care of somthing like this for me :) Especially for constant memory.

4

What happens if you do:

const float temp = h_epsilon*powf(particleMass/referenceDensity,(1.0f/3.0f));

__constant__ float smoothingLength = temp;

This should force evaluation at compile time, or tell you why it can’t be done.

5

That actually sounded like a great idea, but unfortunately I got the exact same generic error :)

Oh well, I guess I’ll just do it manually for now.

6

It sounds like the compiler ignored the construct and optimized it out. What about ‘constant const float smoothingLength’ or ‘volatile const float temp’?