CUDA profiler not matching ptxas info

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

I am running same kernels using sm_1.x and sm_2.x, and profiling them using the command line profiler. I have set a configuration file to record the following attributes:

gridsize3d

threadblocksize

dynsmemperblock

stasmemperblock

regperthread

While compiling, the sm_1.x version shows:

ptxas info    : Used 52 registers, 144+16 bytes smem, 48 bytes cmem[1]

But upon executing the code and inspecting the cuda_profile log-file, I see:

dynsmemperblock    stasmemperblock          regperthread      occupancy

45552	                0	                63	         0.17

I expect the regperthread to be 52 and the static shared memory should be 160 bytes, why is it not showing?

Thank you,

Sayan