How does <<<>>> code translated to cudaLaunchKernel?

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi,

I wondering how does CUDA kernel code something like <<<>>> translates to the cudaLaunchKernel.
Is there any specific cuda runtime API? or translated during compiling?

Thanks

#2

It is translated during compiling by nvcc

#3

Thanks, @Robert_Crovella
I have one more follow-up question.

When I pass the device pointer to the kernel with shape <<<>>> as a function argument, the pointer address seems to change when it comes to cudaLaunchKernel.

For instance, I run the kernel with 3 arguments and the third one is the device pointer.
Before the kernel execution, the address where the pointer points is something like 0x3~~~
but inside of cudaLauncKernel, when i print the address of void**args (ex ags[2]) it shows 0x7~~~
It shows different addresses.

++ follow up information
When I access to the non-pointer arguments, I got the right values.

Can you tell me what happened here?

#4

No I’m not able to understand your description. When launching the kernel using cudaLaunchKernel, your arguments are passed using pointer-to-pointer. Its unlikely you also witnessed the pointer-to-pointer version inside device code. Probably a short, complete example that printed out the two values would clear things up.