Compare commits

...
Author SHA1 Message Date
geohot e0bc99b6c9 check clSetKernelArg 2025-09-13 16:05:29 +08:00
+2 -2
View File
@@ -48,8 +48,8 @@ class CLProgram:
def __call__(self, *bufs:tuple[ctypes._CData, BufferSpec], global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]|None=None,
vals:tuple[int, ...]=(), wait=False) -> float|None:
for i,(b,_) in enumerate(bufs): cl.clSetKernelArg(self.kernel, i, ctypes.sizeof(b), ctypes.byref(b))
for i,v in enumerate(vals,start=len(bufs)): cl.clSetKernelArg(self.kernel, i, 4, ctypes.byref(ctypes.c_int32(v)))
for i,(b,_) in enumerate(bufs): check(cl.clSetKernelArg(self.kernel, i, ctypes.sizeof(b), ctypes.byref(b)))
for i,v in enumerate(vals,start=len(bufs)): check(cl.clSetKernelArg(self.kernel, i, 4, ctypes.byref(ctypes.c_int32(v))))
if local_size is not None: global_size = cast(tuple[int,int,int], tuple(int(g*l) for g,l in zip(global_size, local_size)))
event = cl.cl_event() if wait else None
check(cl.clEnqueueNDRangeKernel(self.dev.queue, self.kernel, len(global_size), None, (ctypes.c_size_t * len(global_size))(*global_size),