def asarray(ctx, other, queue=None, copy=True):
if not isinstance(other, cl.DeviceMemoryView):
other = cl.from_host(ctx, other, copy=copy)
array = CLArray._view_as_this(other)
array.__array_init__(ctx, queue)
return array
def __call__(self, x, out=None, queue=None):
if queue is None:
queue = x.queue
if not isinstance(x, cl.DeviceMemoryView):
x = cl.from_host(queue.context, x)
if out is None:
out = cl.empty(queue.context, x.shape, x.format)
unary_ufunc_kernel(queue, self.device_func, x, out)
array = CLArray._view_as_this(out)
array.__array_init__(queue)
return array
def __call__(self, x, out=None, queue=None):
if queue is None:
queue = x.queue
if not isinstance(x, cl.DeviceMemoryView):
x = cl.from_host(queue.context, x)
if out is None:
out = cl.empty(queue.context, x.shape, x.format)
unary_ufunc_kernel(queue, self.device_func, x, out)
array = CLArray._view_as_this(out)
array.__array_init__(queue)
return array
def __call__(self, context, x, y, out=None, queue=None):
if queue is None:
if hasattr(x, 'queue'):
queue = x.queue
elif hasattr(y, 'queue'):
queue = y.queue
else:
queue = context.queue
if not isinstance(x, cl.DeviceMemoryView):
x = context.asarray(x)
if not isinstance(y, cl.DeviceMemoryView):
y = context.asarray(y)
if y.queue != queue:
queue.enqueue_wait_for_events(y.queue.marker())
if x.queue != queue:
queue.enqueue_wait_for_events(x.queue.marker())
new_shape = broadcast_shape(x.shape, y.shape)
a = cl.broadcast(x, new_shape)
b = cl.broadcast(y, new_shape)
if out is None:
out = context.empty(shape=new_shape, ctype=x.format, queue=queue)
# kernel_source = ufunc_kernel._compile(queue.context, function=self.device_func,
# a=cl.global_memory(a.format, flat=True),
# b=cl.global_memory(b.format, flat=True),
# out=cl.global_memory(out.format, flat=True), source_only=True)
kernel = ufunc_kernel.compile(context,
function=self.device_func,
a=cl.global_memory(a.format, flat=True),
b=cl.global_memory(b.format, flat=True),
out=cl.global_memory(out.format,
flat=True),
cly_meta=self.device_func.func_name)
kernel(queue, a, a.array_info, b, b.array_info, out, out.array_info)
array = CLArray._view_as_this(out)
array.__array_init__(context, queue)
return array
def __call__(self, context, x, y, out=None, queue=None):
if queue is None:
if hasattr(x,'queue'):
queue = x.queue
elif hasattr(y,'queue'):
queue = y.queue
else:
queue = context.queue
if not isinstance(x, cl.DeviceMemoryView):
x = context.asarray(x)
if not isinstance(y, cl.DeviceMemoryView):
y = context.asarray(y)
if y.queue != queue:
queue.enqueue_wait_for_events(y.queue.marker())
if x.queue != queue:
queue.enqueue_wait_for_events(x.queue.marker())
new_shape = broadcast_shape(x.shape, y.shape)
a = cl.broadcast(x, new_shape)
b = cl.broadcast(y, new_shape)
if out is None:
out = context.empty(shape=new_shape, ctype=x.format, queue=queue)
# kernel_source = ufunc_kernel._compile(queue.context, function=self.device_func,
# a=cl.global_memory(a.format, flat=True),
# b=cl.global_memory(b.format, flat=True),
# out=cl.global_memory(out.format, flat=True), source_only=True)
kernel = ufunc_kernel.compile(context, function=self.device_func,
a=cl.global_memory(a.format, flat=True),
b=cl.global_memory(b.format, flat=True),
out=cl.global_memory(out.format, flat=True),
cly_meta=self.device_func.func_name)
kernel(queue, a, a.array_info, b, b.array_info, out, out.array_info)
array = CLArray._view_as_this(out)
array.__array_init__(context, queue)
return array
def reduce(self, context, x, out=None, initial=0.0, queue=None):
if queue is None:
queue = x.queue
if not isinstance(x, cl.DeviceMemoryView):
x = cl.from_host(queue.context, x)
#output, input, shared, group_size, initial=0.0
size = x.size
shared = cl.local_memory(x.ctype, ndim=1, shape=[size])
group_size = size // 2
for item in [2, 4, 8, 16, 32, 64, 128, 256, 512]:
if group_size < item:
group_size = item // 2
break
else:
group_size = 512
if out is None:
out = cl.empty(queue.context, [1], x.format)
kernel = reduce_kernel.compile(queue.context,
function=self.device_func,
output=cl.global_memory(out.ctype,
flat=True),
array=cl.global_memory(x.ctype,
flat=True),
shared=shared,
group_size=cl.cl_uint,
cly_meta=self.device_func.func_name)
max_wgsize = kernel.work_group_size(queue.device)
group_size = min(max_wgsize, group_size)
kernel(queue, out, out.array_info, x, x.array_info, shared,
shared.local_info, group_size)
# reduce_kernel(queue, self.device_func, out, x, shared, group_size)
# reduce_kernel(queue, self.device_func, out, x, shared, group_size)
array = CLArray._view_as_this(out)
array.__array_init__(context, queue)
return array
def reduce(self, context, x, out=None, initial=0.0, queue=None):
if queue is None:
queue = x.queue
if not isinstance(x, cl.DeviceMemoryView):
x = cl.from_host(queue.context, x)
#output, input, shared, group_size, initial=0.0
size = x.size
shared = cl.local_memory(x.ctype, ndim=1, shape=[size])
group_size = size // 2
for item in [2, 4, 8, 16, 32, 64, 128, 256, 512]:
if group_size < item:
group_size = item // 2
break
else:
group_size = 512
if out is None:
out = cl.empty(queue.context, [1], x.format)
kernel = reduce_kernel.compile(queue.context,
function=self.device_func,
output=cl.global_memory(out.ctype, flat=True),
array=cl.global_memory(x.ctype, flat=True),
shared=shared,
group_size=cl.cl_uint,
cly_meta=self.device_func.func_name)
max_wgsize = kernel.work_group_size(queue.device)
group_size = min(max_wgsize, group_size)
kernel(queue, out, out.array_info, x, x.array_info, shared, shared.local_info, group_size)
# reduce_kernel(queue, self.device_func, out, x, shared, group_size)
# reduce_kernel(queue, self.device_func, out, x, shared, group_size)
array = CLArray._view_as_this(out)
array.__array_init__(context, queue)
return array