Replace PyCuda with CuPy

Replaces PyCuda with CuPy

Advantages of CuPy:

• AMD support
• probably higher maintained due to NVIDIA support
• SciPy compatible.

Fixes #70 (closed) Fixes #69 (closed)

pystencils/gpu/cudajit.py

@@ -39,14 +39,10 @@ def make_python_function(kernel_function_node, argument_dict=None, custom_backen
     if argument_dict is None:
         argument_dict = {}
 
-    # if cp.cuda.runtime.is_hip:
-    header_list = ['<cstdint>'] + list(get_headers(kernel_function_node))
-    backend = "nvrtc"
-    options = ["-w", "-std=c++11"]
-    # else:
-    #     header_list = ['<cstdint>'] + list(get_headers(kernel_function_node))
-    #     backend = "nvcc"
-    #     options = ["-w", "-std=c++11", "-Wno-deprecated-gpu-targets", "-m64"]
+    if cp.cuda.runtime.is_hip:
+        header_list = ['"gpu_defines.h"'] + list(get_headers(kernel_function_node))
+    else:
+        header_list = ['"gpu_defines.h"', '<cstdint>'] + list(get_headers(kernel_function_node))
     includes = "\n".join([f"#include {include_file}" for include_file in header_list])
 
     code = includes + "\n"
@@ -55,11 +51,12 @@ def make_python_function(kernel_function_node, argument_dict=None, custom_backen
     code += str(generate_cuda(kernel_function_node, custom_backend=custom_backend))
     code = 'extern "C" {\n%s\n}\n' % code
 
+    options = ["-w", "-std=c++11"]
     if USE_FAST_MATH:
         options.append("-use_fast_math")
     options.append("-I" + get_pystencils_include_path())
 
-    mod = cp.RawModule(code=code, options=tuple(options), backend=backend, jitify=True)
+    mod = cp.RawModule(code=code, options=tuple(options), backend="nvrtc", jitify=True)
     func = mod.get_function(kernel_function_node.function_name)
 
     parameters = kernel_function_node.get_parameters()