Skip to content

P
pytensor
提交 e67bed64 authored 作者: Frederic's avatar Frederic
浏览文件
操作

add theano.scalar.get_scalar_type(dtype) to cache Scalar(dtype) object.

This speed up optimization.
上级 dc91b8f5
隐藏空白字符变更
内嵌 并排
正在显示 包含 64 行增加 和 51 行删除
theano/sandbox/gpuarray/elemwise.py
浏览文件 @ e67bed64
......@@ -72,11 +72,11 @@ class GpuElemwise(HideC, Elemwise):
try:
inps = [make_argument(i, 'i%d' % (n,)) for n, i in
enumerate(node.inputs)]
scal_ins = [scalar.Scalar(i.dtype) for i in node.inputs]
scal_ins = [scalar.get_scalar_type(i.dtype) for i in node.inputs]
outs = [make_argument(o, 'o%d' % (n,)) for n, o in
enumerate(node.outputs) if not n in self.inplace_pattern]
scal_out = [scalar.Scalar(o.dtype) for o in node.outputs]
scal_out = [scalar.get_scalar_type(o.dtype) for o in node.outputs]
fake_node = Apply(self.scalar_op, [i() for i in scal_ins],
[o() for o in scal_out])
......@@ -99,11 +99,11 @@ class GpuElemwise(HideC, Elemwise):
def generate_kernel(self, node, nodename):
inps = [make_argument(i, 'i%d' % (n,)) for n, i in
enumerate(node.inputs)]
scal_ins = [scalar.Scalar(i.dtype) for i in node.inputs]
scal_ins = [scalar.get_scalar_type(i.dtype) for i in node.inputs]
outs = [make_argument(o, 'o%d' % (n,)) for n, o in
enumerate(node.outputs) if not n in self.inplace_pattern]
scal_out = [scalar.Scalar(o.dtype) for o in node.outputs]
scal_out = [scalar.get_scalar_type(o.dtype) for o in node.outputs]
fake_node = Apply(self.scalar_op, [i() for i in scal_ins],
[o() for o in scal_out])
......
theano/scalar/basic.py
浏览文件 @ e67bed64
......@@ -69,6 +69,18 @@ def upcast(dtype, *dtypes):
return rval
def get_scalar_type(dtype):
"""
Return an Scalar(dtype) object.
This cache objects to save allocation and run time.
"""
if dtype not in get_scalar_type.cache:
get_scalar_type.cache[dtype] = Scalar(dtype=dtype)
return get_scalar_type.cache[dtype]
get_scalar_type.cache = {}
def as_scalar(x, name=None):
if isinstance(x, gof.Apply):
if len(x.outputs) != 1:
......@@ -91,7 +103,7 @@ def constant(x):
# purpose typically.
if hasattr(x, 'dtype'):
assert x.ndim == 0
return ScalarConstant(Scalar(str(x.dtype)), x)
return ScalarConstant(get_scalar_type(str(x.dtype)), x)
if isinstance(x, builtin_float):
for dtype in ['float32', 'float64']:
x_ = theano._asarray(x, dtype=dtype)
......@@ -99,7 +111,7 @@ def constant(x):
break
x_ = None
assert x_ is not None
return ScalarConstant(Scalar(str(x_.dtype)), x)
return ScalarConstant(get_scalar_type(str(x_.dtype)), x)
if isinstance(x, builtin_int):
for dtype in ['int8', 'int16', 'int32', 'int64']:
x_ = theano._asarray(x, dtype=dtype)
......@@ -107,7 +119,7 @@ def constant(x):
break
x_ = None
assert x_ is not None
return ScalarConstant(Scalar(str(x_.dtype)), x)
return ScalarConstant(get_scalar_type(str(x_.dtype)), x)
if isinstance(x, builtin_complex):
#TODO: We have added the complex type, so this should be tested
raise NotImplementedError()
......@@ -457,18 +469,18 @@ theano.compile.register_view_op_c_code(
1)
int8 = Scalar('int8')
int16 = Scalar('int16')
int32 = Scalar('int32')
int64 = Scalar('int64')
uint8 = Scalar('uint8')
uint16 = Scalar('uint16')
uint32 = Scalar('uint32')
uint64 = Scalar('uint64')
float32 = Scalar('float32')
float64 = Scalar('float64')
complex64 = Scalar('complex64')
complex128 = Scalar('complex128')
int8 = get_scalar_type('int8')
int16 = get_scalar_type('int16')
int32 = get_scalar_type('int32')
int64 = get_scalar_type('int64')
uint8 = get_scalar_type('uint8')
uint16 = get_scalar_type('uint16')
uint32 = get_scalar_type('uint32')
uint64 = get_scalar_type('uint64')
float32 = get_scalar_type('float32')
float64 = get_scalar_type('float64')
complex64 = get_scalar_type('complex64')
complex128 = get_scalar_type('complex128')
int_types = int8, int16, int32, int64
uint_types = uint8, uint16, uint32, uint64
......@@ -584,7 +596,7 @@ class _scalar_py_operators:
# The second is needed for Elemwise ops to work right
if dtype is None:
dtype = str(self.type.dtype)
return second(self, ScalarConstant(Scalar(dtype), 0))
return second(self, ScalarConstant(get_scalar_type(dtype), 0))
def astype(self, dtype):
return cast(self, dtype)
......@@ -628,7 +640,8 @@ complexs128 = _multi(complex128)
# necessary to use this same mechanism in other places as well in the future.
class upcast_out(object):
def __new__(self, *types):
return Scalar(dtype=Scalar.upcast(*types)),
dtype = Scalar.upcast(*types)
return get_scalar_type(dtype),
class upgrade_to_float(object):
......@@ -644,7 +657,7 @@ class upgrade_to_float(object):
uint16: float32,
uint32: float64,
uint64: float64}
return Scalar(Scalar.upcast(*[conv.get(type, type)
return get_scalar_type(Scalar.upcast(*[conv.get(type, type)
for type in types])),
......@@ -656,7 +669,7 @@ class same_out(object):
def upcast_out_no_complex(*types):
if any([type in complex_types for type in types]):
raise TypeError('complex type are not supported')
return Scalar(dtype=Scalar.upcast(*types)),
return get_scalar_type(dtype=Scalar.upcast(*types)),
def same_out_float_only(type):
......@@ -1455,7 +1468,7 @@ def div_proxy(x, y):
class TrueDiv(BinaryScalarOp):
def output_types(self, types):
if all(t in discrete_types for t in types):
return [Scalar(config.floatX)]
return [get_scalar_type(config.floatX)]
else:
return super(TrueDiv, self).output_types(types)
......
theano/scan_module/scan_utils.py
浏览文件 @ e67bed64
......@@ -59,7 +59,7 @@ def safe_new(x, tag='', dtype=None):
# making the pushout optimization fail
elif isinstance(x, scalar.ScalarVariable):
if dtype:
nw_x = scalar.Scalar(dtype=dtype)()
nw_x = scalar.get_scalar_type(dtype=dtype)()
else:
nw_x = x.type()
nw_x.name = nw_name
......
theano/tensor/basic.py
浏览文件 @ e67bed64
......@@ -991,7 +991,7 @@ class ScalarFromTensor(Op):
assert t.type.broadcastable == ()
return Apply(self,
[t],
[scal.Scalar(dtype=t.type.dtype).make_variable()])
[scal.get_scalar_type(dtype=t.type.dtype).make_variable()])
def perform(self, node, inp, out_):
s, = inp
......
theano/tensor/elemwise.py
浏览文件 @ e67bed64
......@@ -8,7 +8,7 @@ import theano
from theano import gof
from theano.gof import Apply, Op
from theano import scalar
from theano.scalar import Scalar
from theano.scalar import Scalar, get_scalar_type
from theano.printing import pprint
from theano.gof.python25 import all, any
from theano.tensor.utils import hash_from_dict
......@@ -515,7 +515,7 @@ class Elemwise(Op):
"""
inputs = map(as_tensor_variable, inputs)
shadow = self.scalar_op.make_node(
*[Scalar(dtype=i.type.dtype)() for i in inputs])
*[get_scalar_type(dtype=i.type.dtype)() for i in inputs])
target_length = max([input.type.ndim for input in inputs])
......@@ -718,7 +718,7 @@ class Elemwise(Op):
def as_scalar(t):
if isinstance(t.type, (NullType, DisconnectedType)):
return t
return Scalar(t.type.dtype)()
return get_scalar_type(t.type.dtype)()
scalar_inputs = map(as_scalar, inputs)
scalar_ograds = map(as_scalar, ograds)
......@@ -1039,9 +1039,9 @@ class Elemwise(Op):
# We generate the C code of the inner loop using the scalar op
task_code = self.scalar_op.c_code(
Apply(self.scalar_op,
[Scalar(dtype=input.type.dtype)()
[get_scalar_type(dtype=input.type.dtype)()
for input in node.inputs],
[Scalar(dtype=output.type.dtype)()
[get_scalar_type(dtype=output.type.dtype)()
for output in node.outputs]),
nodename + '_scalar_',
["%s_i" % s for s in _inames],
......@@ -1161,11 +1161,11 @@ class Elemwise(Op):
# now we insert versions for the ops on which we depend...
scalar_node = Apply(self.scalar_op,
[Scalar(dtype=input.type.dtype)() for input in node.inputs],
[Scalar(dtype=output.type.dtype)() for output in node.outputs])
[get_scalar_type(dtype=input.type.dtype)() for input in node.inputs],
[get_scalar_type(dtype=output.type.dtype)() for output in node.outputs])
version.append(self.scalar_op.c_code_cache_version_apply(scalar_node))
for i in node.inputs + node.outputs:
version.append(Scalar(dtype=i.type.dtype).c_code_cache_version())
version.append(get_scalar_type(dtype=i.type.dtype).c_code_cache_version())
if all(version):
return tuple(version)
else:
......@@ -1531,9 +1531,9 @@ for(int i=0;i<PyArray_NDIM(%(iname)s);i++){
task1_code = self.scalar_op.c_code(
Apply(
self.scalar_op,
[Scalar(dtype=input.type.dtype)()
[get_scalar_type(dtype=input.type.dtype)()
for input in (node.inputs * 2)],
[Scalar(dtype=output.type.dtype)()
[get_scalar_type(dtype=output.type.dtype)()
for input in node.outputs]),
None,
["%s_i" % aname, "%s_i" % inames[0]],
......@@ -1583,11 +1583,11 @@ for(int i=0;i<PyArray_NDIM(%(iname)s);i++){
# now we insert versions for the ops on which we depend...
scalar_node = Apply(self.scalar_op,
[Scalar(dtype=input.type.dtype)() for input in node.inputs],
[Scalar(dtype=output.type.dtype)() for output in node.outputs])
[get_scalar_type(dtype=input.type.dtype)() for input in node.inputs],
[get_scalar_type(dtype=output.type.dtype)() for output in node.outputs])
version.append(self.scalar_op.c_code_cache_version_apply(scalar_node))
for i in node.inputs + node.outputs:
version.append(Scalar(dtype=i.type.dtype).c_code_cache_version())
version.append(get_scalar_type(dtype=i.type.dtype).c_code_cache_version())
if all(version):
return tuple(version)
else:
......
theano/tensor/opt.py
浏览文件 @ e67bed64
......@@ -3583,7 +3583,7 @@ def local_pow_specialize_device(node):
# 512 is too small for the cpu and too big for some gpu!
if abs(y) == int(abs(y)) and abs(y) <= 512:
pow2 = [xsym]
pow2_scal = [theano.scalar.Scalar(xsym.dtype)()]
pow2_scal = [theano.scalar.get_scalar_type(xsym.dtype)()]
y_to_do = abs(y)
for i in xrange(int(numpy.log2(y_to_do))):
pow2.append(T.sqr(pow2[i]))
......@@ -4638,7 +4638,7 @@ def local_elemwise_fusion_op(OP, max_input_fct=lambda node: 1024):
elif ii in tmp_input:
tmp_s_input.append(tmp_scalar[tmp_input.index(ii)])
else:
tmp = scalar.Scalar(ii.dtype).make_variable()
tmp = scalar.get_scalar_type(ii.dtype).make_variable()
try:
tmp.tag.test_value = gof.op.get_test_value(ii).flatten()[0]
except AttributeError:
......@@ -4692,7 +4692,7 @@ def local_elemwise_fusion_op(OP, max_input_fct=lambda node: 1024):
if inputs.count(i) == node.inputs.count(i):
s = s_inputs[inputs.index(i)]
else:
s = scalar.Scalar(i.dtype).make_variable()
s = scalar.get_scalar_type(i.dtype).make_variable()
try:
if theano.config.compute_test_value != 'off':
v = gof.op.get_test_value(i)
......
theano/tensor/subtensor.py
浏览文件 @ e67bed64
......@@ -318,11 +318,11 @@ class Subtensor(Op):
if (isinstance(entry, gof.Variable)
and entry.type in tensor_types
and numpy.all(entry.type.broadcastable)):
return scal.Scalar(entry.type.dtype)
return scal.get_scalar_type(entry.type.dtype)
elif (isinstance(entry, gof.Type)
and entry in tensor_types
and numpy.all(entry.broadcastable)):
return scal.Scalar(entry.dtype)
return scal.get_scalar_type(entry.dtype)
elif slice_ok and isinstance(entry, slice):
a = entry.start
b = entry.stop
......
theano/tensor/type.py
浏览文件 @ e67bed64
......@@ -240,7 +240,7 @@ class TensorType(Type):
% (self.__class__.__name__, self.dtype))
def to_scalar_type(self):
return scal.Scalar(dtype=self.dtype)
return scal.get_scalar_type(dtype=self.dtype)
def __eq__(self, other):
"""Compare True iff other is the same kind of TensorType"""
......@@ -538,23 +538,23 @@ class TensorType(Type):
def c_headers(self):
"""Override `CLinkerObject.c_headers` """
return scal.Scalar(self.dtype).c_headers()
return scal.get_scalar_type(self.dtype).c_headers()
def c_libraries(self):
return scal.Scalar(self.dtype).c_libraries()
return scal.get_scalar_type(self.dtype).c_libraries()
def c_compile_args(self):
return scal.Scalar(self.dtype).c_compile_args()
return scal.get_scalar_type(self.dtype).c_compile_args()
def c_support_code(self):
"""Override `CLinkerObject.c_support_code` """
return scal.Scalar(self.dtype).c_support_code()
return scal.get_scalar_type(self.dtype).c_support_code()
def c_init_code(self):
return scal.Scalar(self.dtype).c_init_code()
return scal.get_scalar_type(self.dtype).c_init_code()
def c_code_cache_version(self):
scalar_version = scal.Scalar(self.dtype).c_code_cache_version()
scalar_version = scal.get_scalar_type(self.dtype).c_code_cache_version()
if scalar_version:
return (11,) + scalar_version
else:
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论