Need help on elf format cubin file via decuda

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, all,

I have met one problem for the register and other resource occupancy analysis when applying DECUDA tool to analyze the elf-format cubin file.

The elf-format cubin file is not text format. So, I use the elfToCubin.py in DECUDA to change elf into readable cubin file. However, problem here is that the values of lmem, smem, reg and bar are always 0, i.e.

lmem=0
smem=0
reg=0
bar=0

But, the bincode is correctly demonstrated. Definitely, I use registers in my kernel function. (I paste my kernel function below):

global void matrixAdd_kernel( int **d_A, int **d_B, int **d_Result, int width)
{

int ix = blockDim.y * blockIdx.y + threadIdx.y;
int iy = blockDim.x * blockIdx.x + threadIdx.x;
	
    for(int i=0;i<width;i++)
	d_Result[ix][iy]+=d_A[ix][i] * d_B[i][iy];

} //end of kernel

Can anyone tell me how to get the right resource occupancy from elf-cubin file?

Some parameters setting: sm_12, cubda build rule 3.0.0.

2

As you still have the source code for the kernel it’s probably easier to compile again with [font=“Courier New”]–ptxas-options=-v[/font] to see resource usage, than to figure out from the object file.

3

Hi, tera, thanks for your advice. It works:)