[STFT][CPU] Improve performance of STFT for CPU by reusage RDFT jit Executor

[mitruska]

Details:

• Improve performance of STFT for CPU plugin by reusage RDFT jit Executor
• Use parallel loops in stft
• No changes in the logic of the existing RDFT executor, RDFTExecutor::build function has been added to keep the RDFT details hidden in cpp as is.
• Perf numbers collected within the ticket

Tickets:

• 156115

[a-sidorova]

LGTM 👍🏼

[dmitry-gorokhov]

minor: the simpliest way to perform such things like transpose in CPU plugin is to use the following interface: https://github.com/openvinotoolkit/openvino/blob/master/src/plugins/intel_cpu/src/cpu_memory.h#L190 with correctly initialized memory descriptors. In theory it should provides better performance.
Anyway I don't insist it should be refactored in bounds of this PR.

[mitruska]

Thank you for this hint! Worth to consider as a follow up, we have also briefly discussed potential further optimizations for the transpose here: #26967 (comment)

[dmitry-gorokhov]

Overall the changes looks good to me. The only remaining question is about accuracy. @mitruska Do you have any updates?

[mitruska]

Overall the changes looks good to me. The only remaining question is about accuracy. @mitruska Do you have any updates?

@dmitry-gorokhov

The input test data range for RDFT is defined as:

openvino/src/tests/functional/shared_test_classes/include/shared_test_classes/base/utils/ranges.hpp

Lines 79 to 80 in 8f0094d

	{ov::op::v7::DFT::get_type_info_static(), Range({{0, 1}}, {{0, 1, 1000000}})},
	{ov::op::v9::RDFT::get_type_info_static(), Range({{0, 1}}, {{0, 1, 1000000}})},

For STFT it is:

openvino/src/tests/functional/shared_test_classes/include/shared_test_classes/base/utils/ranges.hpp

Line 243 in 8f0094d

{ov::op::v15::STFT::get_type_info_static(), Range({{16, 24}, {1, 16}}, {{-100, 100}, {-100, 100}})},

The abs_threshold for existing STFT tests can be smaller (1e-5) if the same test data range as for RDFT is set, while the abs_threshold for RDFT test is currently set to (1e-4).

We can align the threshold and input data range (also modified in STFT GPU PR)

[mitruska]

The input data range for common layer tests has been updated on master by other PR:

• [GPU]: Added basic STFT implementation #27794

Currently custom threshold is not needed, I suggest to apply and review any additional ideas for tests adjustments separately as it will affect both CPU and GPU.