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 test_from_host_no_copy(self):
a = np.array([[1, 2], [3, 4]])
refcount = sys.getrefcount(a)
clmem = cl.from_host(ctx, a, copy=False)
# event = PyEvent()
# def set_event(mem):
# event.set()
# clmem.add_destructor_callback(set_event)
self.assertEqual(refcount + 1, sys.getrefcount(a))
del clmem
gc.collect()
# self.assertTrue(event.wait(1), 'event timed out. destructor_callback not called')
self.assertEqual(refcount, sys.getrefcount(a))
clmem = cl.from_host(ctx, a, copy=False)
view_a = memoryview(a)
self.assertEqual(clmem.format, view_a.format)
self.assertEqual(clmem.shape, view_a.shape)
self.assertEqual(clmem.strides, view_a.strides)
queue = cl.Queue(ctx)
if queue.device.host_unified_memory:
a[0, 0] = 100
with clmem.map(queue) as view:
b = np.asarray(view)
self.assertEqual(b[0, 0], 100)
else:
#TODO: should there be a test here?
pass
def test_from_host_no_copy(self):
a = np.array([[1, 2], [3, 4]])
refcount = sys.getrefcount(a)
clmem = cl.from_host(ctx, a, copy=False)
# event = PyEvent()
# def set_event(mem):
# event.set()
# clmem.add_destructor_callback(set_event)
self.assertEqual(refcount + 1, sys.getrefcount(a))
del clmem
gc.collect()
# self.assertTrue(event.wait(1), 'event timed out. destructor_callback not called')
self.assertEqual(refcount, sys.getrefcount(a))
clmem = cl.from_host(ctx, a, copy=False)
view_a = memoryview(a)
self.assertEqual(clmem.format, view_a.format)
self.assertEqual(clmem.shape, view_a.shape)
self.assertEqual(clmem.strides, view_a.strides)
queue = cl.Queue(ctx)
if queue.device.host_unified_memory:
a[0, 0] = 100
with clmem.map(queue) as view:
b = np.asarray(view)
self.assertEqual(b[0, 0], 100)
else:
#TODO: should there be a test here?
pass
def main():
size = 10
a = np.random.rand(size).astype('f')
b = np.random.rand(size).astype('f')
ctx = cl.Context()
queue = cl.Queue(ctx)
cla = cl.from_host(ctx, a, copy=True)
clb = cl.from_host(ctx, b, copy=True)
clc = cl.empty(ctx, [size], ctype='f')
prg = cl.Program(
ctx, """
__kernel void add(__global const float *a,
__global const float *b, __global float *c)
{
int gid = get_global_id(0);
c[gid] = a[gid] + b[gid];
}
""").build()
add = prg.add
add.argtypes = cl.global_memory('f'), cl.global_memory(
'f'), cl.global_memory('f')
add.argnames = 'a', 'b', 'c'
add.global_work_size = lambda a: a.shape
add(queue, a=cla, b=clb, c=clc)
with clc.map(queue) as view:
print "view is a python memoryview object", view
arr = np.asarray(view)
print "Answer should be zero:"
print(arr - (a + b)).sum()
def main():
size = 10
a = np.random.rand(size).astype('f')
b = np.random.rand(size).astype('f')
ctx = cl.Context()
queue = cl.Queue(ctx)
cla = cl.from_host(ctx, a, copy=True)
clb = cl.from_host(ctx, b, copy=True)
clc = cl.empty(ctx, [size], ctype='f')
prg = cl.Program(ctx, """
__kernel void add(__global const float *a,
__global const float *b, __global float *c)
{
int gid = get_global_id(0);
c[gid] = a[gid] + b[gid];
}
""").build()
add = prg.add
add.argtypes = cl.global_memory('f'), cl.global_memory('f'), cl.global_memory('f')
add.argnames = 'a', 'b', 'c'
add.global_work_size = lambda a: a.shape
add(queue, a=cla, b=clb, c=clc)
with clc.map(queue) as view:
print "view is a python memoryview object", view
arr = np.asarray(view)
print "Answer should be zero:"
print (arr - (a + b)).sum()
def test_broadcast_0D(self):
with self.assertRaises(TypeError):
cl.broadcast(None, [1])
one = cl.from_host(ctx, c_int(1))
a = cl.broadcast(one, [10, 10])
self.assertEqual(a.shape, (10, 10))
self.assertEqual(a.strides, (0, 0))
queue = cl.Queue(ctx)
with a.map(queue) as view:
b = np.asarray(view)
self.assertEqual(b.shape, (10, 10))
self.assertEqual(b.strides, (0, 0))
def test_broadcast_0D(self):
with self.assertRaises(TypeError):
cl.broadcast(None, [1])
one = cl.from_host(ctx, c_int(1))
a = cl.broadcast(one, [10, 10])
self.assertEqual(a.shape, (10, 10))
self.assertEqual(a.strides, (0, 0))
queue = cl.Queue(ctx)
with a.map(queue) as view:
b = np.asarray(view)
self.assertEqual(b.shape, (10, 10))
self.assertEqual(b.strides, (0, 0))
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 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
def test_broadcast_2D(self):
with self.assertRaises(TypeError):
cl.broadcast(None, [1])
npa = np.arange(10, dtype=c_float)
z = np.zeros([10, 1])
ten = cl.from_host(ctx, npa)
a = cl.broadcast(ten, [10, 10])
self.assertEqual(a.shape, (10, 10))
self.assertEqual(a.strides, (0, sizeof(c_float)))
queue = cl.Queue(ctx)
with a.map(queue) as view:
b = np.asarray(view)
self.assertEqual(b.shape, (10, 10))
self.assertEqual(b.strides, (0, sizeof(c_float)))
self.assertTrue(np.all(b == z + npa))
def test_broadcast_2D(self):
with self.assertRaises(TypeError):
cl.broadcast(None, [1])
npa = np.arange(10, dtype=c_float)
z = np.zeros([10, 1])
ten = cl.from_host(ctx, npa)
a = cl.broadcast(ten, [10, 10])
self.assertEqual(a.shape, (10, 10))
self.assertEqual(a.strides, (0, sizeof(c_float)))
queue = cl.Queue(ctx)
with a.map(queue) as view:
b = np.asarray(view)
self.assertEqual(b.shape, (10, 10))
self.assertEqual(b.strides, (0, sizeof(c_float)))
self.assertTrue(np.all(b == z + npa))
def blitz(queue, func, out=None):
'''
lets get blitzed!
'''
func_ast = decompile_func(func)
func_globals = func.func_globals.copy()
if func.func_closure:
func_globals.update({
name: cell.cell_contents
for name, cell in zip(func.func_code.co_freevars,
func.func_closure)
})
blitzer = BlitzVisitor(func.func_code.co_filename, func_globals)
blitzed = ast.Expression(blitzer.visit(func_ast))
blitzed_code = compile(blitzed, func.func_code.co_filename, 'eval')
blitzed_func = eval(blitzed_code)
blitz_kernel = create_n_arg_kernel(sorted(blitzer.locls.keys()))
args = {}
for key, var in blitzer.locls.items():
if not isinstance(var, cl.DeviceMemoryView):
var = cl.from_host(queue.context, var)
args[key] = var
shape = broadcast_shapes([var.shape for var in args.values()])
print "shape", shape
for key, var in args.items():
args[key] = cl.broadcast(var, shape)
print "out, **args", out, args
blitz_kernel(queue, blitzed_func, out, **args)
def main():
ctx = cl.Context(device_type=cl.Device.GPU)
queue = cl.Queue(ctx)
host_init = np.arange(8, dtype=c_float) + 1
device_input = cl.from_host(ctx, host_init)
output = ca.reduce(queue, lambda a, b: a + b, device_input)
print "-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- "
print "data:", host_init
print "-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- "
print "host sum:", host_init.sum()
with output.map(queue) as view:
print "device sum:", np.asarray(view).item()
output = ca.reduce(queue, lambda a, b: a * b, device_input, initial=1.0)
print "host product:", host_init.prod()
with output.map(queue) as view:
print "device product:", np.asarray(view).item()
def blitz(queue, func, out=None):
'''
lets get blitzed!
'''
func_ast = decompile_func(func)
func_globals = func.func_globals.copy()
if func.func_closure:
func_globals.update({name:cell.cell_contents for name, cell in zip(func.func_code.co_freevars, func.func_closure)})
blitzer = BlitzVisitor(func.func_code.co_filename, func_globals)
blitzed = ast.Expression(blitzer.visit(func_ast))
blitzed_code = compile(blitzed, func.func_code.co_filename, 'eval')
blitzed_func = eval(blitzed_code)
blitz_kernel = create_n_arg_kernel(sorted(blitzer.locls.keys()))
args = {}
for key, var in blitzer.locls.items():
if not isinstance(var, cl.DeviceMemoryView):
var = cl.from_host(queue.context, var)
args[key] = var
shape = broadcast_shapes([var.shape for var in args.values()])
print "shape", shape
for key, var in args.items():
args[key] = cl.broadcast(var, shape)
print "out, **args", out, args
blitz_kernel(queue, blitzed_func, out, **args)