Improvements in the code of the clients

I would be happy to receive this patch. As long as it's OpenCL, we should be able to incorporate it in the App.

Thanks, got it. Regarding the GPU code, this only changes the type of twiddle_dee and adds some options to the OpenCL compiler, in particular to use OpenCL 2.0 (*). I built app version 1.26 with that, please give it a try. It's the same plan class as 1.25, so the same restrictions apply.

(*) The other stuff in there just puts the CPU to sleep while the GPU is running, there is some other method implemented in our app.

First difference: a kernel that begins like this:

__attribute__(( reqd_work_group_size( 16, 16, 1 ) ))
kernel void
transpose_gcn_tw_fwd( global float2* restrict pmComplexIn, global float2* restrict pmComplexOut )
{
   const Tile localIndex = { get_local_id( 0 ), get_local_id( 1 ) };
   const Tile localExtent = { get_local_size( 0 ), get_local_size( 1 ) };

...

   local float2 lds[ 64 ][ 64 ];

should be changed to

   local float2 lds[ 64 ][ 65 ];

It uses a bit more memory, but the access alignment becomes better and the number of cache line conflicts is reduced.

The same thing applies to kernel called transpose gcn_tw_back

and both of those gcn_tw kernels can have the line:

const Tile localExtent = { get_local_size( 0 ), get_local_size( 1 ) };

changed to

const Tile localExtent = { 16,  16 };

to help compiler to generate better code at compile time. The (16, 16) is the required work group size and can be hard coded,

 

That should make the final speedup.