Add support for fully configurable partial-pass 3D kernels

[eng-flavio-teixeira]

Add the ability to fully configure partial-pass 3D kernels in kernel-generator.py. This should remove the need for the remaining hard-coded 64x64x64 partial-pass parameters across the library.

The main changes include:

• kernel-generator.py now has a list of CS_3D_PP kernels, where partial-pass parameters are fully configurable.
• Stockham_gen and kernel-generator.py interaction mechanism via JSON needed to understand the concept of linked partial-pass kernels.
• Function pool and function map key needed a new concept for the linked partial-pass kernels, which are now queryable in the pool.
• New compute schemes specifically added for partial-pass (CS_3D_PP, CS_STOCKHAM_PP, and CS_STOCKHAM_PP_BLOCK_CC). These should simplify handling of nodes in the plan tree structure, and make them clearly distinct from CS_3D_RC nodes.

[ROCmMathLibrariesBot]

Performance reports:
Commit hashes:
64fd8ba
a454522
gfx942-perf single report
gfx942-perf double report
gfx90a-perf single report
gfx90a-perf double report

[malcolmroberts]

It's not clear from context what these new structs are for; would it be possible to add some code comments, please?

[eng-flavio-teixeira]

Sure, new commit is in with the comments.

[Kevonosdiaz]

Haven't looked much into this PR yet, but it might be a good time to update the default value of the variable at

rocFFT/CMakeLists.txt

Line 121 in dcfff3f

set(ROCFFT_FUNCTION_POOL_N 8 CACHE STRING "Number of files to split function_pool into for compilation")

to 16 or higher if the function pool is growing, since it might improve build times a bit.

[eng-flavio-teixeira]

Haven't looked much into this PR yet, but it might be a good time to update the default value of the variable at

rocFFT/CMakeLists.txt

Line 121 in dcfff3f

set(ROCFFT_FUNCTION_POOL_N 8 CACHE STRING "Number of files to split function_pool into for compilation")

to 16 or higher if the function pool is growing, since it might improve build times a bit.

Did you see any noticeable improvement going from 8 to 16? This PR does not add more kernels to the function pool, but the next one will (but not an unreasonable amount).

[Kevonosdiaz]

Did you see any noticeable improvement going from 8 to 16? This PR does not add more kernels to the function pool, but the next one will (but not an unreasonable amount).

It slowed down my build time by a couple seconds when going from 8 to 16, and building just the rocfft-test target with ninja, and without ccache. When there are more kernels added then maybe it'll be worth checking again.

[eng-flavio-teixeira]

Did you see any noticeable improvement going from 8 to 16? This PR does not add more kernels to the function pool, but the next one will (but not an unreasonable amount).

It slowed down my build time by a couple seconds when going from 8 to 16, and building just the rocfft-test target with ninja, and without ccache. When there are more kernels added then maybe it'll be worth checking again.

Sounds good, I will keep that in mind for the next PR.