What is the default Shared Memory size per block in GTX 480 ?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi.

I want to know what is the default configured shared memory size per block in GeForce GTX 480…

Also, how can i verify this by code (e.g. by making a function call)?

Thanks :)

2

Hi,
By default, GTX480 along with all compute capability 2.x devices have 48KB of shared memory per SM and 16KB of L1 cache.
You can check this with cudaGetDeviceProperties() and checking cudaDeviceProp::sharedMemPerBlock.

3

Are you sure that the devices have 48KB per SM by default? Or just 48KB is the maximum size? I have read somewhere in that forum that the devices uses only 16KB of Shared Memory and if i want to switch to 48KB i need to call cudaFuncSetCacheConfig() with cudaFuncCachePreferShared argument… Is this true?

Is this somewhere stated formally by Nvidia?

How can i use cudaDeviceProp::sharedMemPerBlock to read the current values / sizes?

4

If you compile for sm_2x, the default is 48Kb of shared memory, 16Kb of L1 cache. You can change the setting with cudaFuncSetCacheConfig.
See page 145 of the CUDA C Programming Guide.

5

Great! Thank you mfatica.

One more thing. I compile my programs with nvcc without using any sm_xx architecture statement. What is the default architecture for my case? Does the defaults above apply to my case?

Also, Gilles_C, please give me a code example of how to use cudaDeviceProp::sharedMemPerBlock to read the current values / sizes if you can (because i programming in C)…

6

Default compute capability is 1.0 regardless of the machine you are compiling on.

Here you are:

int dev = 0; // to be adjusted or queried according to your needs

    cudaDeviceProp props;

    cudaGetDeviceProperties(&props, dev);

    printf("Amount of shared memory per block is %d bytes\n", props.sharedMemPerBlock);
7

That’s ok! Thank you very much, all!