def __new__(subtype, shape, dtype=np.float32, buffer=None, offset=0, strides=None, order=None, info=None):
obj = np.ndarray.__new__(subtype, shape, dtype, buffer, offset, strides, order)
if backend in {"ocl", "opencl", "OCL"}:
obj.ocl_buf = cl.clCreateBuffer(context, np.prod(shape) * obj.itemsize)
obj.host_dirty = False
obj.ocl_dirty = False
obj.register = None
return obj
def __call__(self, im):
output = zeros_like(im.data)
in_buf, evt = buffer_from_ndarray(self.queue, im.data, blocking=False)
evt.wait()
self.kernel.setarg(0, in_buf, sizeof(cl_mem))
out_buf = clCreateBuffer(self.context, output.nbytes)
self.kernel.setarg(1, out_buf, sizeof(cl_mem))
evt = clEnqueueNDRangeKernel(self.queue, self.kernel, self.global_size)
evt.wait()
_, evt = buffer_to_ndarray(self.queue, out_buf, output)
evt.wait()
del in_buf
del out_buf
return Array(unique_name(), output)
def __new__(subtype,
shape,
dtype=np.float32,
buffer=None,
offset=0,
strides=None,
order=None,
info=None):
obj = np.ndarray.__new__(subtype, shape, dtype, buffer, offset,
strides, order)
if backend in {"ocl", "opencl", "OCL"}:
obj.ocl_buf = cl.clCreateBuffer(context,
np.prod(shape) * obj.itemsize)
obj.host_dirty = False
obj.ocl_dirty = False
obj.register = None
return obj
def __call__(self, im, num_powers, border):
out_shape = [num_powers] + list(im.shape)
output = np.empty(out_shape, dtype=np.float32)
in_buf, evt = buffer_from_ndarray(self.queue, im.data, blocking=False)
evt.wait()
self.kernel.setarg(0, in_buf, sizeof(cl_mem))
out_buf = clCreateBuffer(self.queue.context, output.nbytes)
self.kernel.setarg(1, out_buf, sizeof(cl_mem))
evt = clEnqueueNDRangeKernel(self.queue, self.kernel, self.global_size)
evt.wait()
self.kernel2.setarg(0, out_buf, sizeof(cl_mem))
for power in range(num_powers):
self.kernel2.setarg(1, power, sizeof(cl_int))
evt = clEnqueueNDRangeKernel(self.queue, self.kernel2, self.global_size)
evt.wait()
_, evt = buffer_to_ndarray(self.queue, out_buf, output)
evt.wait()
return Array(unique_name(), output)
def allocate(self, length, dtype, name):
mem = cl.clCreateBuffer(self.queue.context, length * dtype.itemsize)
mem.ptr_type = np.ctypeslib.ndpointer(dtype)()
mem.ptr_type._global = True
return mem, SymbolRef(name, mem)