Cost of accessing built-in variables

[weichertv]

Hello everyone!

I am sorry if this is documented somewhere, but I did not find it:

Accessing global memory costs between 400 and 600 cycles, accessing shared memory costs 4 cycles, registers add no overhead to the computation cost. Where do the built-in variables like threadIdx, blockDim, etc. reside? What is the access cost?

Thanks!

Volker

[jph4599]

It looks like this is discussed in a few other threads.

http://forums.nvidia.com/index.php?showtop…mp;#entry314099

threadIdx is initially in register R0 of each thread (the x and y components are in the 16 LSBs and MSBs, respectively), not in shared memory. It is placed there by the hardware on invocation of each block. If the kernel doesn’t use threadIdx the compiler may choose to use the register for something else.

blockIdx, blockDim, and gridDim are passed as parameters in shared memory, because all threads in a block will read the same location when these are read.

Mark

http://forums.nvidia.com/index.php?showtopic=79049

http://forums.nvidia.com/index.php?showtopic=58249

[avidday]

They come from shared memory

[weichertv]

Thank you!