def __call__(self, *args):
"""__call__
:param *args:
"""
if isinstance(args[0], hmarray):
output = empty_like(args[0])
else:
output = np.zeros_like(args[0])
# self.kernel.argtypes = tuple(
# cl_mem for _ in args + (output, )
# ) + (localmem, )
buffers = []
events = []
for index, arg in enumerate(args + (output, )):
if isinstance(arg, hmarray):
buffers.append(arg.ocl_buf)
else:
buf, evt = buffer_from_ndarray(self.queue, arg, blocking=True)
# evt.wait()
events.append(evt)
buffers.append(buf)
# self.kernel.setarg(index, buf, sizeof(cl_mem))
cl.clWaitForEvents(*events)
cl_error = 0
if isinstance(self.kernel, list):
kernels = len(self.kernel)
if kernels == 2:
cl_error = self._c_function(self.queue, self.kernel[0],
self.kernel[1], *buffers)
elif kernels == 3:
cl_error = self._c_function(self.queue, self.kernel[0],
self.kernel[1], self.kernel[2],
*buffers)
elif kernels == 4:
cl_error = self._c_function(
self.queue, self.kernel[0], self.kernel[1], self.kernel[2],
self.kernel[3], *buffers
)
else:
cl_error = self._c_function(self.queue, self.kernel, *buffers)
if cl.cl_errnum(cl_error) != cl.cl_errnum.CL_SUCCESS:
raise StencilException(
"Error executing stencil kernel: opencl {} {}".format(
cl_error,
cl.cl_errnum(cl_error)
)
)
if isinstance(output, hmarray):
return output
buf, evt = buffer_to_ndarray(
self.queue, buffers[-1], output
)
evt.wait()
return buf
def get_placeholder_output(self, args):
return empty_like(args[0])
