Hi everybody around here
I have got a strange problem with warp serial.
In my program, I didn’t use shared memory.
I used local memory, const memory, global memory and texture.
when I analyst my program by used cuda visual profiler, I saw that warp serial counter is not zero , I don’t know why warp serial counter is not zero and what is the main problem.
I hope someone here understand and please give some ideas.
thank you.
From the profiler documentation:
My guess is that your memory access pattern of the constant memory is not 100% broadcast.
Thank you Mr MisterAnderson42. :)
I tried not to use const memory, but warp-serial still occurs.
So my program uses local memory, global memory, and texture.
Not used const memory and shared memory.
Does anything else makes warp-serial occurs?
I solve my problem.
before, when passing parameters to kernel function. I thought that parameters will stored in register.
by checked in *cubin. I realize that parameter stored in shared memory (smem = static shared memory).
by accessed these parameters. warp-serial may occurs.