WIP: Cuda autotune

This PR introduces two one changes:

• rotate (32,1,1) depending on field strides to fastest dimension. So (1,1,32) for c-layout and (32,1,1) for fortran layout. So pystencils will be fast also for c-layout (this will always be performed)
• auto-tune the block dimensions to whatevers is fastest for a specific kernel on localhost. On first kernel call different layouts are tried and the kernel will be called henceforth with the fastest configuration (disk_cached). This could be intersting for OpenCL where we don't know which launch config is the fastest (on OpenCL the runtime can alternatively give a hint on that).

One drawback: the test calls are only correct if input and output fields do not overlap (so no in-place kernels).

pystencils/gpucuda/indexing.py

@@ -131,7 +131,7 @@ class AbstractIndexing(abc.ABC):
                     current_best = block_and_thread_numbers
 
             print(f'{current_best} is the best out of {self._autotune_block_sizes or self.AUTOTUNE_BLOCK_SIZES}')
-            self._block_size = current_best
+            self._block_size = current_best['block']
             return current_best
         return _autotune_call_parameters(self,
                                          call_shape,
@@ -158,7 +158,7 @@ class BlockIndexing(AbstractIndexing):
     AUTOTUNE_NUM_CALLS = 10
 
     def __init__(self, field, iteration_slice,
-                 block_size=get_cuda_config()['preferred_block_size'],
+                 block_size=tuple(get_cuda_config()['preferred_block_size']),
                  permute_block_size_dependent_on_layout=True,
                  compile_time_block_size=False,
 
@@ -308,6 +308,7 @@ class LineIndexing(AbstractIndexing):
         self._coordinates = coordinates
         self._iterationSlice = normalize_slice(iteration_slice, field.spatial_shape)
         self._symbolicShape = [e if isinstance(e, sp.Basic) else None for e in field.spatial_shape]
+        self._autotune_block_sizes = None
 
     @property
     def coordinates(self):