PTX assembler output is incorrect ?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi

I was tweaking my Ray Tracer and so I wanted to know the register and local memory usage for different configurations…I compiled it with

--ptxas-options=-v

option.

It says 50 registers in one case and it still runs perfectly.

Why is this the case ?

This is the exact ptxas output:

1>ptxas info    : Compiling entry function '_Z13cudaRayTracerPiii6float3PKf'

1>ptxas info    : Used 50 registers, 8448+0 bytes lmem, 48+48 bytes smem, 212 bytes cmem[1], 8 bytes cmem[14]

Program runs correctly.

Thanks

Akshay

2

I think I’m missing something. What’s the problem with the compiler using 50 registers? That’s more registers than I’ve seen a kernel use before, and it means you can only have 160 threads active per multiprocessor. Otherwise, seems fine to me.

3

One of my raytracing kernels is using more than 60 registers. I use 64 threads per block and it is working ok also.