About concurrent execution of kernel into SM. I had doubts when doing the experiment.

[weijizhen]

Experimental environment：
-gpgpu_concurrent_kernel_sm 1
-gpgpu_max_concurrent_kernel 8
kernel1<<<1,1>>>
kernel2<<<1,1>>>

question：
I printed this variable during gdb debugging.
My understanding of this variable is that since it is concurrent, it should be 0x1e00f40,0x1e5654d, 0,0,0,0,0, and each value is the address of the kernel.

But until the end of the first kernel, the print is 0x1e00f40,0, 0,0,0,0,0

So how should we understand concurrent kernel execution?

---

Sorry, I didn't understand the code. After many inquiries, this variable m_running_kernels just gets the kernel address in execution, so there is only one address.
I wonder if I add such modification (-gpgpu_concurrent_kernel_sm 1
-gpgpu_max_concurrent_kernel 8), is there any difference from the case without modification? I found no difference in my experiments.

[JRPan]

All kernels launched should show up here. Can you show the gpgpu-sim output?

[weijizhen]

log-0.txt

As you can see, I didn't increase the stream, the two kernels are in one stream.

Therefore, do I need to modify the CUDA program to achieve the condition of running multiple kernels?

[JRPan]

Yes, only kernels not on the same stream can run concurrently.

[weijizhen]

thank you for your help.