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/cpu/cpujit.py

@@ -175,9 +175,15 @@ def read_config():
         ('object_cache', os.path.join(user_cache_dir('pystencils'), 'objectcache')),
         ('clear_cache_on_start', False),
     ])
+    default_cuda_config = OrderedDict([
+        ('always_autotune', False),
+        ('preferred_block_size', (16, 16, 1)),
+    ])
 
     default_config = OrderedDict([('compiler', default_compiler_config),
-                                  ('cache', default_cache_config)])
+                                  ('cache', default_cache_config),
+                                  ('cuda', default_cuda_config)
+                                  ])
 
     config_path, config_exists = get_configuration_file_path()
     config = default_config.copy()
@@ -219,6 +225,10 @@ def get_cache_config():
     return _config['cache']
 
 
+def get_cuda_config():
+    return _config['cuda']
+
+
 def add_or_change_compiler_flags(flags):
     if not isinstance(flags, list) and not isinstance(flags, tuple):
         flags = [flags]