Shared memory

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,
I am work with GTS450 and according spec, I have 1024thread / block and 48KB shared memory / block.

My question is, If some kernel works with 512thread / block, what is maximum shared memory per block? Its 24KB. My problem is that my kernel use 48KB(512thread/block) and it doesnt work well, but there is no error or compiler message.

thx for all answer.

2

What do you mean by ‘doesnt work well’ exactly?

You can use all the shared memory per SM for one block. Check out the CUDA GPU Occupancy Calculator.

3

Sorry, I aplologize, I found bug and thx for tip with “CUDA GPU Occupancy Calculator”, its very usefull :-)