changed once again how gpu arguments get passed around --> ctypes structure

Author: MiloLurati

kernel_tuner/backends/hip.py

@@ -128,8 +128,8 @@ def ready_argument_list(self, arguments):
             The order should match the argument list on the HIP function.
             Allowed values are np.ndarray, and/or np.int32, np.float32, and so on.
         :type arguments: list(numpy objects)
-        :returns: List of ctypes arguments to be passed to the HIP function.
-        :rtype: list of ctypes
+        :returns: Ctypes structure of arguments to be passed to the HIP function.
+        :rtype: ctypes structure
         """
         logging.debug("HipFunction ready_argument_list called")
 
@@ -150,9 +150,17 @@ def ready_argument_list(self, arguments):
             elif isinstance(arg, np.generic):
                 data_ctypes = dtype_map[dtype_str](arg)
                 ctype_args.append(data_ctypes)  
+
+        # Determine the types of the fields in the structure
+        field_types = [type(x) for x in ctype_args]
+        # Define a new ctypes structure with the inferred layout
+        class ArgListStructure(ctypes.Structure):
+            _fields_ = [(f'field{i}', t) for i, t in enumerate(field_types)]
+            def __getitem__(self, key):
+                return self._fields_[key]
 
-        return ctype_args
-    
+        return ArgListStructure(*ctype_args)
+            
 
     def compile(self, kernel_instance):
         """call the HIP compiler to compile the kernel, return the function
@@ -219,10 +227,10 @@ def run_kernel(self, func, gpu_args, threads, grid, stream=None):
         :param func: A PyHIP kernel compiled for this specific kernel configuration
         :type func: ctypes pionter
 
-        :param gpu_args: A list of arguments to the kernel, order should match the
+        :param gpu_args: A ctypes structure of arguments to the kernel, order should match the
             order in the code. Allowed values are either variables in global memory
             or single values passed by value.
-        :type gpu_args: list of ctypes
+        :type gpu_args: ctypes structure
 
         :param threads: A tuple listing the number of threads in each dimension of
             the thread block
@@ -233,17 +241,10 @@ def run_kernel(self, func, gpu_args, threads, grid, stream=None):
         :type grid: tuple(int, int, int)
         """
         logging.debug("HipFunction run_kernel called")
+
         if stream is None:
             stream = self.stream
 
-        # Determine the types of the fields in the structure
-        field_types = [type(x) for x in gpu_args]
-        # Define a new ctypes structure with the inferred layout
-        class ArgListStructure(ctypes.Structure):
-            _fields_ = [(f'field{i}', t) for i, t in enumerate(field_types)]
-        
-        gpu_args = ArgListStructure(*gpu_args)
-
         hip.hipModuleLaunchKernel(func, 
                                   grid[0], grid[1], grid[2], 
                                   threads[0], threads[1], threads[2],

test/test_hip_functions.py

@@ -25,11 +25,8 @@ def test_ready_argument_list():
 
     dev = kt_hip.HipFunctions(0)
     gpu_args = dev.ready_argument_list(arguments)
-    
-    assert isinstance(gpu_args[0], ctypes.c_void_p)
-    assert isinstance(gpu_args[1], ctypes.c_int32)
-    assert isinstance(gpu_args[2], ctypes.c_void_p)
-    assert isinstance(gpu_args[3], ctypes.c_bool)
+
+    assert(gpu_args, ctypes.Structure)
 
 @skip_if_no_pyhip
 def test_compile():