Make "cooperativity" be part of the launch configuration

[eyalroz]

At the moment, the user must specify, when launching a kernel, whether thread blocks should be allowed to cooperate or not. This is a bit cumbersome, and also increases the number of kernel launch functions and methods, which is already not so small.

On the other hand, our wrapped kernel class does indicate whether thread blocks should be allowed to cooperate, and putting that flag into a launch config may remove the ability to hard-code that requirement.

Yet again on the first hand - why should we not have the ability to hard-code requirements for other aspects of the launch, into a kernel_t? Dynamic shared memory size requirements, and even block and grid size requirements? A kernel can certainly have those.

I would lean in favor of having less functionality in kernel_t and making the move, since this library is just for wrapping, not for beefier abstractions.

Why do you think @codecircuit , @raplonu, @quxflux et al. ?

[codecircuit]

As far as I understand your point you want to discuss how much 'kernel launch configuration' should be saved in class kernel_t. Is that correct?

I agree with your point of having less functionality in class kernel_t and e.g. move the thread block cooperation flag into the kernel launch config, instead of having it as a member variable in kernel_t. Due to issue #236, I think it is not completely safe to use the kernel_t class when launching a CUDA kernel and I personally avoid that completely. Nevertheless, I think the kernel_t class is handy, to e.g. check kernel attributes and GPU occupancy. As far as I see it, the member variable thread_block_cooperation of class kernel_t is only used in the ctor and getter function. Thus, the member functions of class kernel_t should not be affected if this member variable is removed.

[eyalroz]

As far as I understand your point you want to discuss how much 'kernel launch configuration' should be saved in class kernel_t. Is that correct?

No... I'm wondering whether to add bool block_cooperation to launch_configuration_t, and removing it from kernel_t...

[codecircuit]

To me this sounds reasonable.

[eyalroz]

@codecircuit : Reasonable, and better than what we have currently?

[eyalroz]

@codecircuit : Done.

[codecircuit]

@codecircuit : Reasonable, and better than what we have currently?

Yes, the commit looks good 👍.

[eyalroz]

@codecircuit : I've deleted 63 more lines than I've added, so it has to be good :-P

Reminding you to possibly have a look at the driver-wrappers branch.