Arguments of __global__ function is on __local__?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, all,

sorry, for the stupid question… Let I have a kernel:

void global kernel(int N, int M, float *A)
{ … }

are all N, M, A variables local, and very slow, or they definitelly stay on registers?

Thank you!

Sincerely

Ilghiz

2

Arguments are stored in shared memory. Since all threads typically access the same argument at the same time, it should be broadcast and thus have no bank conflicts.

3

MisterAnderson42 is right. Check the .ptx file. Kernel arguments should be accessed using ld.param which is shared mem space. The .cubin file lists the reserved space in smem = … If you see lmem != 0 here, some other kernel variables have been moved to local memory because you are out of registers.

Peter

4

Dear MisterAnderson42 and Peter,

thank you for your kind and helpful answer. Later I will look for .cubin and .ptx more often, thay have important informations.

Sincerely

Ilghiz