Skip to content

P
pytensor
提交 daf196e6 authored 作者: Frederic's avatar Frederic
浏览文件
操作

Make GpuCAReduceCuda support pre_scalar_op

上级 961c15e0
隐藏空白字符变更
内嵌 并排
正在显示 包含 107 行增加 和 34 行删除
theano/sandbox/gpuarray/elemwise.py
浏览文件 @ daf196e6
...@@ -570,10 +570,14 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -570,10 +570,14 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
GPUs are not especially well-suited to reduction operations so it is GPUs are not especially well-suited to reduction operations so it is
quite possible that the GPU might be slower for some cases. quite possible that the GPU might be slower for some cases.
pre_scalar_op: if present, must be a scalar op with only 1
input. We will execute it on the input value before reduction.
""" """
def __init__(self, scalar_op, axis=None, def __init__(self, scalar_op, axis=None,
reduce_mask=None, dtype=None, acc_dtype=None): reduce_mask=None, dtype=None, acc_dtype=None,
pre_scalar_op=None):
if reduce_mask is not None: if reduce_mask is not None:
reduce_mask = tuple(reduce_mask) reduce_mask = tuple(reduce_mask)
self.reduce_mask = reduce_mask self.reduce_mask = reduce_mask
...@@ -583,6 +587,9 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -583,6 +587,9 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
self._n_scalar_op_calls = 0 self._n_scalar_op_calls = 0
CAReduceDtype.__init__(self, scalar_op, axis=axis, CAReduceDtype.__init__(self, scalar_op, axis=axis,
dtype=dtype, acc_dtype=acc_dtype) dtype=dtype, acc_dtype=acc_dtype)
self.pre_scalar_op = pre_scalar_op
if pre_scalar_op:
assert pre_scalar_op.nin == 1
def __eq__(self, other): def __eq__(self, other):
return (type(self) == type(other) and return (type(self) == type(other) and
...@@ -590,7 +597,8 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -590,7 +597,8 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
self.reduce_mask == other.reduce_mask and self.reduce_mask == other.reduce_mask and
self.dtype == other.dtype and self.dtype == other.dtype and
self.acc_dtype == other.acc_dtype and self.acc_dtype == other.acc_dtype and
self.scalar_op == other.scalar_op) self.scalar_op == other.scalar_op and
self.pre_scalar_op == other.pre_scalar_op)
def __hash__(self): def __hash__(self):
return (hash(type(self)) ^ return (hash(type(self)) ^
...@@ -598,19 +606,35 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -598,19 +606,35 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
hash(self.reduce_mask) ^ hash(self.reduce_mask) ^
hash(self.dtype) ^ hash(self.dtype) ^
hash(self.acc_dtype) ^ hash(self.acc_dtype) ^
hash(type(self.scalar_op))) hash(type(self.scalar_op)) ^
hash(type(self.pre_scalar_op)))
def __str__(self): def __str__(self):
pre = ""
if self.pre_scalar_op:
pre = "pre=%s,red=" % str(self.pre_scalar_op)
ax = '' ax = ''
if self.axis is not None: if self.axis is not None:
ax = '{%s}' % (', '.join(str(x) for x in self.axis),) ax = '{%s}' % (', '.join(str(x) for x in self.axis),)
return "GpuCAReduceCuda{%s}%s" % (str(self.scalar_op), ax) return "GpuCAReduceCuda{%s%s}%s" % (pre,str(self.scalar_op), ax)
def __setstate__(self, d):
self.__dict__.update(d)
# For unpickling of old ops.
if not hasattr(self, "pre_scalar_op"):
self.pre_scalar_op = None
def make_node(self, x): def make_node(self, x):
x = as_gpuarray_variable(x) x = as_gpuarray_variable(x)
ret = super(GpuCAReduceCuda, self).make_node(x) ret = super(GpuCAReduceCuda, self).make_node(x)
self = copy.copy(self) self = copy.copy(self)
self.axis = ret.op.axis self.axis = ret.op.axis
if self.pre_scalar_op:
# Currently we only tested pre_scalar_op that don't cause
# upcast.
d1 = self.__class__(scalar_op=self.scalar_op)(Elemwise(self.pre_scalar_op)(x))
assert d1.dtype == ret.outputs[0].dtype
assert Elemwise(self.pre_scalar_op)(x).dtype == x.dtype
if self.reduce_mask is None: if self.reduce_mask is None:
if self.axis is None: if self.axis is None:
reduce_mask = [1] * x.type.ndim reduce_mask = [1] * x.type.ndim
...@@ -1010,15 +1034,33 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1010,15 +1034,33 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
else: else:
assert isinstance(self.scalar_op, (scal.Maximum, assert isinstance(self.scalar_op, (scal.Maximum,
scal.Minimum)) scal.Minimum))
if self.pre_scalar_op: # TODO, multi_dtype!
#dtype = node.inputs[0].dtype
dtype = 'float32'
dummy_var = scal.Scalar(dtype=dtype)()
dummy_node = self.pre_scalar_op.make_node(dummy_var)
dummy_name = 'assign_init_pre_scalar_op' + str(self._n_scalar_op_calls)
self._n_scalar_op_calls += 1
t = self.pre_scalar_op.c_code(dummy_node, dummy_name,
(first_item,), ("",), {})
assert t.startswith(' = ')
first_item = t[3:]
if first_item[-1] == ';':
first_item = first_item[:-1]
return first_item return first_item
def _assign_reduce(self, node, name, left, right, sub): def _assign_reduce(self, node, name, left, right, sub, pre):
""" """
node: the node argument to this op's c_code node: the node argument to this op's c_code
name: the name argument to this op's c_code name: the name argument to this op's c_code
left: a C code string identifying an lvalue left: a C code string identifying an lvalue
right: a C code string identifying an expression right: a C code string identifying an expression
sub: the sub argument to this op's c_code sub: the sub argument to this op's c_code
pre: If True, we will add the pre_scalar_op.c_code
returns C code to reduce left and right, assigning the returns C code to reduce left and right, assigning the
result to left.""" result to left."""
...@@ -1035,6 +1077,18 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1035,6 +1077,18 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
dummy_name = name + '_scalar_op' + str(self._n_scalar_op_calls) dummy_name = name + '_scalar_op' + str(self._n_scalar_op_calls)
self._n_scalar_op_calls += 1 self._n_scalar_op_calls += 1
if pre and self.pre_scalar_op:
assert left == "myresult"
dummy_node = self.pre_scalar_op.make_node(dummy_left)
dummy_name = name + '_scalar_op' + str(self._n_scalar_op_calls)
self._n_scalar_op_calls += 1
t = self.pre_scalar_op.c_code(dummy_node, dummy_name,
(right,), ("",), sub)
assert t.startswith(' = ')
right = t[3:]
if right[-1] == ';':
right = right[:-1]
return self.scalar_op.c_code(dummy_node, dummy_name, (left, right), return self.scalar_op.c_code(dummy_node, dummy_name, (left, right),
(left,), sub) (left,), sub)
...@@ -1064,7 +1118,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1064,7 +1118,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
int idx = threadNum - (threadCount >> 1) * 2;""" int idx = threadNum - (threadCount >> 1) * 2;"""
new_version += self._assign_reduce(node, name, 'buf[idx]', new_version += self._assign_reduce(node, name, 'buf[idx]',
'buf[threadNum]', sub) 'buf[threadNum]', sub, False)
new_version += """ new_version += """
} }
...@@ -1084,7 +1138,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1084,7 +1138,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
""" """
new_version += self._assign_reduce(node, name, new_version += self._assign_reduce(node, name,
'buf[threadNum]', 'temp', sub) 'buf[threadNum]', 'temp', sub, False)
new_version += """ new_version += """
} }
...@@ -1115,7 +1169,8 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1115,7 +1169,8 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
{ {
""" """
current_version += self._assign_reduce(node, name, current_version += self._assign_reduce(node, name,
'myresult', 'buf[i]', sub) + """ 'myresult', 'buf[i]',
sub, False) + """
} }
buf[threadNum] = myresult; buf[threadNum] = myresult;
/*Comment this optimization as it don't work on Fermi GPU. /*Comment this optimization as it don't work on Fermi GPU.
...@@ -1127,7 +1182,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1127,7 +1182,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
current_version += self._assign_reduce(node, name, current_version += self._assign_reduce(node, name,
'buf[threadNum]', 'buf[threadNum]',
'buf[threadNum+%d]' % num, 'buf[threadNum+%d]' % num,
sub) sub, False)
current_version += """ current_version += """
""" """
current_version += """ current_version += """
...@@ -1146,7 +1201,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1146,7 +1201,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
this_if = "if (threadNum + %d < threadCount) " % num + \ this_if = "if (threadNum + %d < threadCount) " % num + \
self._assign_reduce(node, name, self._assign_reduce(node, name,
'buf[threadNum]','buf[threadNum+%d]' % num, 'buf[threadNum]','buf[threadNum+%d]' % num,
sub) sub, False)
current_version += this_if current_version += this_if
current_version += """ current_version += """
""" """
...@@ -1166,7 +1221,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1166,7 +1221,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
#Threads must be organized as: threadNum%nb_reduce correspond to the same sum #Threads must be organized as: threadNum%nb_reduce correspond to the same sum
#nb_reduce<=warpSize #nb_reduce<=warpSize
def _k_reduce_buf_multiple(self, z_pos, node, name, nb_reduce): def _k_reduce_buf_multiple(self, z_pos, node, name, nb_reduce):
reduce_fct = self._assign_reduce(node, name, 'myresult', 'buf[i]', {}) reduce_fct = self._assign_reduce(node, name, 'myresult', 'buf[i]', {}, False)
return """ return """
__syncthreads(); // some kernel do multiple reduction. __syncthreads(); // some kernel do multiple reduction.
buf[threadNum] = myresult; buf[threadNum] = myresult;
...@@ -1767,7 +1822,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1767,7 +1822,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
reducebuf = self._k_reduce_buf('Z[0]', node, nodename, sub={}) reducebuf = self._k_reduce_buf('Z[0]', node, nodename, sub={})
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0]", "A[i0]",
{}) {}, True)
reduce_init = self._assign_init("A[0]") reduce_init = self._assign_init("A[0]")
print >> sio, """ print >> sio, """
static __global__ void kernel_reduce_ccontig_%(nodename)s( static __global__ void kernel_reduce_ccontig_%(nodename)s(
...@@ -1798,7 +1853,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1798,7 +1853,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
reducebuf = self._k_reduce_buf('Z[0]', node, nodename, sub={}) reducebuf = self._k_reduce_buf('Z[0]', node, nodename, sub={})
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0]", "A[i0 * sA0]",
{}) {}, True)
reduce_init = self._assign_init("A[0]") reduce_init = self._assign_init("A[0]")
print >> sio, """ print >> sio, """
static __global__ void kernel_reduce_1_%(nodename)s( static __global__ void kernel_reduce_1_%(nodename)s(
...@@ -1829,7 +1884,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1829,7 +1884,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
reducebuf = self._k_reduce_buf('Z[0]', node, nodename, sub={}) reducebuf = self._k_reduce_buf('Z[0]', node, nodename, sub={})
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + i1 * sA1]", "A[i0 * sA0 + i1 * sA1]",
{}) {}, True)
reduce_init = self._assign_init("A[0]") reduce_init = self._assign_init("A[0]")
print >> sio, """ print >> sio, """
...@@ -1913,7 +1968,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1913,7 +1968,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
reduce_fct = self._assign_reduce( reduce_fct = self._assign_reduce(
node, nodename, "myresult", node, nodename, "myresult",
"A[i3 * sA3 + i2 * sA2 + i1 * sA1 + i0 * sA0]", "A[i3 * sA3 + i2 * sA2 + i1 * sA1 + i0 * sA0]",
{}) {}, True)
print >> sio, """ print >> sio, """
%(decl)s{ %(decl)s{
%(init)s %(init)s
...@@ -1941,7 +1996,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1941,7 +1996,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
node, nodename, sub={}) node, nodename, sub={})
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + i1 * sA1 + i2 * sA2]", "A[i0 * sA0 + i1 * sA1 + i2 * sA2]",
{}) {}, True)
reduce_init = self._assign_init("A[i0 * sA0 + threadIdx.x * sA1 + i2 * sA2]") reduce_init = self._assign_init("A[i0 * sA0 + threadIdx.x * sA1 + i2 * sA2]")
print >> sio, """ print >> sio, """
static __global__ void kernel_reduce_010_%(nodename)s( static __global__ void kernel_reduce_010_%(nodename)s(
...@@ -1980,7 +2035,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -1980,7 +2035,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
if self.reduce_mask == (0, 1, 0): if self.reduce_mask == (0, 1, 0):
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"X[a * sX0 + b * sX1 + c * sX2]", "X[a * sX0 + b * sX1 + c * sX2]",
{}) {}, True)
reduce_init = self._assign_init("X[a * sX0 + 0 * sX1 + c * sX2]") reduce_init = self._assign_init("X[a * sX0 + 0 * sX1 + c * sX2]")
print >> sio, """ print >> sio, """
static __global__ void kernel_reduce_010_AD_%(nodename)s( static __global__ void kernel_reduce_010_AD_%(nodename)s(
...@@ -2040,7 +2095,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -2040,7 +2095,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
'blockDim.x') 'blockDim.x')
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + i1 * sA1 + i2 * sA2]", "A[i0 * sA0 + i1 * sA1 + i2 * sA2]",
{}) {}, True)
reduce_init = self._assign_init("A[i0 * sA0 + 0 * sA1 + i2 * sA2]") reduce_init = self._assign_init("A[i0 * sA0 + 0 * sA1 + i2 * sA2]")
print >> sio, """ print >> sio, """
%(decl)s %(decl)s
...@@ -2076,7 +2131,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -2076,7 +2131,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
reducebuf = self._k_reduce_buf('Z[blockIdx.x * sZ0]', node, nodename, sub = {}) reducebuf = self._k_reduce_buf('Z[blockIdx.x * sZ0]', node, nodename, sub = {})
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + i1 * sA1 + blockIdx.x * sA2]", "A[i0 * sA0 + i1 * sA1 + blockIdx.x * sA2]",
{}) {}, True)
reduce_init = self._assign_init("A[blockIdx.x * sA2]") reduce_init = self._assign_init("A[blockIdx.x * sA2]")
print >> sio, """ print >> sio, """
static __global__ void kernel_reduce_110_%(nodename)s( static __global__ void kernel_reduce_110_%(nodename)s(
...@@ -2117,7 +2172,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -2117,7 +2172,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
init = self._k_init(node, nodename) init = self._k_init(node, nodename)
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + i1 * sA1 + i2 * sA2]", "A[i0 * sA0 + i1 * sA1 + i2 * sA2]",
{}) {}, True)
reduce_init = self._assign_init("A[i1 * sA1 + i2 * sA2]") reduce_init = self._assign_init("A[i1 * sA1 + i2 * sA2]")
print >> sio, """ print >> sio, """
%(decl)s %(decl)s
...@@ -2144,7 +2199,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -2144,7 +2199,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
init = self._k_init(node, nodename) init = self._k_init(node, nodename)
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + i1 * sA1 + i2 * sA2]", "A[i0 * sA0 + i1 * sA1 + i2 * sA2]",
{}) {}, True)
reduce_init = self._assign_init("A[0]") reduce_init = self._assign_init("A[0]")
print >> sio, """ print >> sio, """
%(decl)s %(decl)s
...@@ -2171,7 +2226,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -2171,7 +2226,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
node, nodename, sub={}) node, nodename, sub={})
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + i1 * sA1 + i2 * sA2]", "A[i0 * sA0 + i1 * sA1 + i2 * sA2]",
{}) {}, True)
reduce_init = self._assign_init("A[i0 * sA0 + i1 * sA1]") reduce_init = self._assign_init("A[i0 * sA0 + i1 * sA1]")
print >> sio, """ print >> sio, """
static __global__ void kernel_reduce_001_%(nodename)s( static __global__ void kernel_reduce_001_%(nodename)s(
...@@ -2214,7 +2269,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -2214,7 +2269,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
init = self._k_init(node, nodename) init = self._k_init(node, nodename)
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + i1 * sA1 + i2 * sA2 + i3 * sA3]", "A[i0 * sA0 + i1 * sA1 + i2 * sA2 + i3 * sA3]",
{}) {}, True)
reduce_init = self._assign_init("A[i0 * sA0 + i1 * sA1]") reduce_init = self._assign_init("A[i0 * sA0 + i1 * sA1]")
print >> sio, """ print >> sio, """
%(decl)s %(decl)s
...@@ -2247,7 +2302,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -2247,7 +2302,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
init = self._k_init(node, nodename) init = self._k_init(node, nodename)
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + i1 * sA1 + i2 * sA2 + i3 * sA3]", "A[i0 * sA0 + i1 * sA1 + i2 * sA2 + i3 * sA3]",
{}) {}, True)
reduce_init = self._assign_init("A[i0 * sA0 + i2 * sA2]") reduce_init = self._assign_init("A[i0 * sA0 + i2 * sA2]")
print >> sio, """ print >> sio, """
%(decl)s %(decl)s
...@@ -2278,7 +2333,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -2278,7 +2333,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
init = self._k_init(node, nodename) init = self._k_init(node, nodename)
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + i1 * sA1 + i2 * sA2 + i3 * sA3]", "A[i0 * sA0 + i1 * sA1 + i2 * sA2 + i3 * sA3]",
{}) {}, True)
reduce_init = self._assign_init("A[0]") reduce_init = self._assign_init("A[0]")
print >> sio, """ print >> sio, """
%(decl)s %(decl)s
...@@ -2304,7 +2359,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype): ...@@ -2304,7 +2359,7 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
node, nodename, sub={}) node, nodename, sub={})
reduce_fct = self._assign_reduce(node, nodename, "myresult", reduce_fct = self._assign_reduce(node, nodename, "myresult",
"A[i0 * sA0 + blockIdx.x * sA1 + i2 * sA2 + i3 * sA3]", "A[i0 * sA0 + blockIdx.x * sA1 + i2 * sA2 + i3 * sA3]",
{}) {}, True)
reduce_init = self._assign_init("A[blockIdx.x * sA1]") reduce_init = self._assign_init("A[blockIdx.x * sA1]")
print >> sio, """ print >> sio, """
static __global__ void kernel_reduce_1011_%(nodename)s( static __global__ void kernel_reduce_1011_%(nodename)s(
......
theano/sandbox/gpuarray/tests/test_elemwise.py
浏览文件 @ daf196e6
...@@ -40,11 +40,13 @@ class test_GpuCAReduceCPY(test_CAReduce): ...@@ -40,11 +40,13 @@ class test_GpuCAReduceCPY(test_CAReduce):
bin_dtypes = ["uint8", "int8"] bin_dtypes = ["uint8", "int8"]
op = GpuCAReduceCPY op = GpuCAReduceCPY
reds = [scalar.add, scalar.mul] reds = [scalar.add, scalar.mul]
pre_scalar_op = None
def test_perform(self): def test_perform(self):
for dtype in self.dtypes + self.bin_dtypes: for dtype in self.dtypes + self.bin_dtypes:
for op in self.reds: for op in self.reds:
self.with_linker(gof.PerformLinker(), op, dtype=dtype) self.with_linker(gof.PerformLinker(), op, dtype=dtype,
pre_scalar_op=self.pre_scalar_op)
def test_perform_nan(self): def test_perform_nan(self):
for dtype in self.dtypes: for dtype in self.dtypes:
...@@ -52,12 +54,14 @@ class test_GpuCAReduceCPY(test_CAReduce): ...@@ -52,12 +54,14 @@ class test_GpuCAReduceCPY(test_CAReduce):
continue continue
for op in self.reds: for op in self.reds:
self.with_linker(gof.PerformLinker(), op, dtype=dtype, self.with_linker(gof.PerformLinker(), op, dtype=dtype,
test_nan=True) test_nan=True,
pre_scalar_op=self.pre_scalar_op)
def test_c(self): def test_c(self):
for dtype in self.dtypes + self.bin_dtypes: for dtype in self.dtypes + self.bin_dtypes:
for op in self.reds: for op in self.reds:
self.with_linker(gof.CLinker(), op, dtype=dtype) self.with_linker(gof.CLinker(), op, dtype=dtype,
pre_scalar_op=self.pre_scalar_op)
def test_c_nan(self): def test_c_nan(self):
for dtype in self.dtypes: for dtype in self.dtypes:
...@@ -65,7 +69,8 @@ class test_GpuCAReduceCPY(test_CAReduce): ...@@ -65,7 +69,8 @@ class test_GpuCAReduceCPY(test_CAReduce):
continue continue
for op in self.reds: for op in self.reds:
self.with_linker(gof.CLinker(), op, dtype=dtype, self.with_linker(gof.CLinker(), op, dtype=dtype,
test_nan=True) test_nan=True,
pre_scalar_op=self.pre_scalar_op)
def test_infer_shape(self): def test_infer_shape(self):
for dtype in self.dtypes: for dtype in self.dtypes:
...@@ -148,6 +153,7 @@ class test_GpuCAReduceCuda(test_GpuCAReduceCPY): ...@@ -148,6 +153,7 @@ class test_GpuCAReduceCuda(test_GpuCAReduceCPY):
op = GpuCAReduceCuda op = GpuCAReduceCuda
reds = [scalar.add, scalar.mul, reds = [scalar.add, scalar.mul,
scalar.maximum, scalar.minimum] scalar.maximum, scalar.minimum]
pre_scalar_op = scalar.sqr
def test_perform(self): def test_perform(self):
return return
......
theano/tensor/tests/test_elemwise.py
浏览文件 @ daf196e6
...@@ -308,15 +308,19 @@ class test_CAReduce(unittest_tools.InferShapeTester): ...@@ -308,15 +308,19 @@ class test_CAReduce(unittest_tools.InferShapeTester):
] ]
def with_linker(self, linker, scalar_op=scalar.add, dtype="floatX", def with_linker(self, linker, scalar_op=scalar.add, dtype="floatX",
pre_scalar_op=None,
test_nan=False, tensor_op=None): test_nan=False, tensor_op=None):
for xsh, tosum in self.cases: for xsh, tosum in self.cases:
if dtype == "floatX": if dtype == "floatX":
dtype = theano.config.floatX dtype = theano.config.floatX
x = TensorType(dtype, [(entry == 1) for entry in xsh])('x') x = TensorType(dtype, [(entry == 1) for entry in xsh])('x')
d = {}
if pre_scalar_op is not None:
d = {"pre_scalar_op": pre_scalar_op}
if tensor_op is None: if tensor_op is None:
e = as_tensor_variable(self.op(scalar_op, axis=tosum)(x)) e = as_tensor_variable(self.op(scalar_op, axis=tosum, **d)(x))
else: else:
e = as_tensor_variable(tensor_op(x, axis=tosum)) e = as_tensor_variable(tensor_op(x, axis=tosum, **d))
if tosum is None: if tosum is None:
tosum = range(len(xsh)) tosum = range(len(xsh))
...@@ -337,6 +341,8 @@ class test_CAReduce(unittest_tools.InferShapeTester): ...@@ -337,6 +341,8 @@ class test_CAReduce(unittest_tools.InferShapeTester):
else: else:
xv = numpy.asarray(numpy.nan, dtype=dtype) xv = numpy.asarray(numpy.nan, dtype=dtype)
zv = xv zv = xv
if pre_scalar_op is not None:
zv = Elemwise(scalar_op=pre_scalar_op)(x).eval({x: xv})
numpy_raised = False numpy_raised = False
if len(tosum) > 1 and any([a < 0 for a in tosum]): if len(tosum) > 1 and any([a < 0 for a in tosum]):
#In that case, we need to use the good order of axis #In that case, we need to use the good order of axis
...@@ -505,16 +511,22 @@ class test_CAReduce(unittest_tools.InferShapeTester): ...@@ -505,16 +511,22 @@ class test_CAReduce(unittest_tools.InferShapeTester):
self.with_linker(gof.CLinker(), scalar.maximum, dtype=dtype, self.with_linker(gof.CLinker(), scalar.maximum, dtype=dtype,
test_nan=True) test_nan=True)
def test_infer_shape(self, dtype=None): def test_infer_shape(self, dtype=None, pre_scalar_op=None):
if dtype is None: if dtype is None:
dtype = theano.config.floatX dtype = theano.config.floatX
for xsh, tosum in self.cases: for xsh, tosum in self.cases:
x = TensorType(dtype, [(entry == 1) for entry in xsh])('x') x = TensorType(dtype, [(entry == 1) for entry in xsh])('x')
if pre_scalar_op is not None:
x = pre_scalar_op(x)
if tosum is None: if tosum is None:
tosum = range(len(xsh)) tosum = range(len(xsh))
xv = numpy.asarray(numpy.random.rand(*xsh), dtype=dtype) xv = numpy.asarray(numpy.random.rand(*xsh), dtype=dtype)
d = {}
if pre_scalar_op is not None:
xv = x.eval({x.owner.inputs[0]: xv})
d = {pre_scalar_op: pre_scalar_op}
self._compile_and_check([x], self._compile_and_check([x],
[self.op(scalar.add, axis=tosum)(x)], [self.op(scalar.add, axis=tosum, *d)(x)],
[xv], self.op, [xv], self.op,
["local_cut_useless_reduce"], ["local_cut_useless_reduce"],
warn=0 not in xsh) warn=0 not in xsh)
......
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