Tuning interpolation_kernel()

[bartvstratum]

I'm also starting the tuning of interpolation_kernel().

[bartvstratum]

I added the kernel tuner script (interpolation_kernel.py). This one again requires binary files as input, generated from the cuda_dump_bins branch.

Results from tuning:

So the optimal block size is again very small...

Optimized, the kernel is about 3x faster. Old and new profile:

Time(%)  Total Time (ns)  Instances   Average   Minimum  Maximum                                                  Name                                                
 -------  ---------------  ---------  ---------  -------  -------  ----------------------------------------------------------
 9.3         30895235         61   506479.3   137631   620120  void (anonymous namespace)::interpolation_kernel<double>(int, int, int, int, int, int, int, double,¿
 3.6         11185705         61   183372.2    79967   212253  void (anonymous namespace)::interpolation_kernel<double>(int, int, int, int, int, int, int, double,¿

[bartvstratum]

Changing the order of the dimensions (x=col, y=lay) is a bit faster: adffcd1

Time(%)  Total Time (ns)  Instances   Average   Minimum  Maximum                                                  Name
-------  ---------------  ---------  ---------  -------  -------  -------------------------------------------------------------------------------------------
2.8          8696371         61   142563.5    74783   171934  void (anonymous namespace)::interpolation_kernel<double>(

[isazi]

I have been testing changing the order of dimensions in both Tau kernels, no major changes. I assume it is because most data accesses are not directly dependent on thread ID, but on the content of some other array.

[bartvstratum]

It's again a bit faster if I remove the ncol loop from the kernel: 1e57819

Time(%)  Total Time (ns)  Instances   Average   Minimum  Maximum                                                  Name                                        -------  ---------------  ---------  ---------  -------  -------  -------------------------------------------------------------------------------------------
    1.6          5043399         61    82678.7    43136   100063  void (anonymous namespace)::interpolation_kernel<double>(int, int, int, int, int, in

This also results in larger optimal block sizes: