[gpu]: ISTFT basic impl.

[pkowalc1]

Basic implementation of ISTFT for GPU.

Details:

• supports everything which supports CPU ref impl.

Tickets:

• CVS-162460

[praasz]

Can be removed.
Is not required right, shape inference not depends on this attribute

[pkowalc1]

This method is needed to properly construct op on gpu side(ctor params are not supported in gpu plugin... )..

Otherwise you will construct partially undefined op which will work at this point, but only because you and me know that shape inference currently does not depend on that state.

[praasz]

OK, for core part

[mlukasze]

@wilson-seok please review, or assign someone to review

[isanghao]

We have a new(and better) infra for adding kernel: ocl_v2. Could you implement kernel on that, instead of ocl?
Here's an example of migration: #29901

[pkowalc1]

Well I can do it in separate PR if really needed - this one has already passed all the tests(which took a week due to unrelated CI failures...)

[mlukasze]

I'm voting for this approach

[sshlyapn]

For proper synchronization this call should either be blocking, or result event should be added to kernel dependencies

[pkowalc1]

So the kernel may not run on the same stream?

[sshlyapn]

No, they are guaranteed to be executed on the same stream, but depending on the queue type there might be a problem:

1. in case of in_order queue all submitted tasks (memory fill and kernel enqueue) are executed in order they are enqueued
2. however, in case of out_of_order queue driver considers dependencies between the tasks to execute them. If there are no proper dependencies (using barriers or events), the driver doesn't guarantee that memory will be reset before the kernel launches in this case - as kernel execution might start in parallel or even before memory reset in theory

Additionally, considering that memory can be reused between primitives, there might be a situation where we reset the memory of some predecessor primitive right during its execution. Therefore, for out_of_order queue some additional barrier is required before this fill() call as well

[sshlyapn]

I think it's not actually needed to enumerate all the inputs, we can just slightly update ISTFT like this:

    ISTFT(const primitive_id& id,
          const std::vector<input_info>& inputs,
          const bool center,
          const bool normalized)
        : primitive_base(id, inputs),
          center(center),
          normalized(normalized) {}

And simplify primitive creation:

Suggested change

	if (inputs.size() == 4) {
	auto prim = cldnn::ISTFT(layer_type_name_ID(op), inputs[0], inputs[1], inputs[2], inputs[3], op->get_center(), op->get_normalized());
	p.add_primitive(*op, prim);
	} else {
	auto prim = cldnn::ISTFT(layer_type_name_ID(op), inputs[0], inputs[1], inputs[2], inputs[3], inputs[4], op->get_center(), op->get_normalized());
	p.add_primitive(*op, prim);
	}
	auto prim = cldnn::ISTFT(layer_type_name_ID(op), inputs, op->get_center(), op->get_normalized());
	p.add_primitive(*op, prim);

[sshlyapn]

nit:

Suggested change

	// Copyright (C) 2018-2025 Intel Corporation
	// Copyright (C) 2025 Intel Corporation