using too much shared memory on fermi?

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi

While trying to compile a kernel on a Tesla C2050 that uses

[codebox]

shared double a[32][33], b[33][32];

[/codebox]

The compiler complains

[codebox]

Entry function ‘foo’ uses too much shared data (0x4238 bytes + 0x10 bytes system, 0x4000 max)

[/codebox]

And devicequery shows this:

[codebox]

Total amount of shared memory per block: 49152 bytes

[/codebox]

0x4000 is only 16kb. So any clue what is going on?

Thanks!

#2

oh, sm_20 does the trick.

#3

oh, sm_20 does the trick.