Add experimental half precison support

With this MR experimental half-precision support is added

pystencils/gpu/cudajit.py

@@ -39,17 +39,22 @@ def make_python_function(kernel_function_node, argument_dict=None, custom_backen
     if argument_dict is None:
         argument_dict = {}
 
+    half = True in (field.dtype.is_half() for field in kernel_function_node.fields_accessed)
+
     if cp.cuda.runtime.is_hip:
         header_list = ['"gpu_defines.h"'] + list(get_headers(kernel_function_node))
+        if half:
+            header_list += ['<hip_fp16.h>', ]
     else:
         header_list = ['"gpu_defines.h"', '<cstdint>'] + list(get_headers(kernel_function_node))
+        if half:
+            header_list += ['<cuda_fp16.h>', ]
     includes = "\n".join([f"#include {include_file}" for include_file in header_list])
 
     code = includes + "\n"
     code += "#define FUNC_PREFIX __global__\n"
     code += "#define RESTRICT __restrict__\n\n"
-    code += str(generate_cuda(kernel_function_node, custom_backend=custom_backend))
-    code = 'extern "C" {\n%s\n}\n' % code
+    code += 'extern "C" {\n%s\n}\n' % str(generate_cuda(kernel_function_node, custom_backend=custom_backend))
 
     options = ["-w", "-std=c++11"]
     if USE_FAST_MATH: