Add a check at C code generation time for ops that are not explicitely

theano/compile/ops.py

@@ -215,6 +215,8 @@ class Shape(gof.Op):
 
     @note: Non-differentiable.
     """
+    _fp16_ok = True
+
     # Mapping from Type to C code (and version) to use.
     # In the C code, the name of the input variable is %(iname)s,
     # the output variable is %(oname)s.
@@ -308,6 +310,8 @@ class Shape_i(gof.Op):
 
     @note: Non-differentiable.
     """
+    _fp16_ok = True
+
     # Mapping from Type to C code (and version) to use.
     # In the C code, the name of the input variable is %(iname)s,
     # the output variable is %(oname)s.

theano/gof/op.py

@@ -726,9 +726,20 @@ class Op(utils.object2, PureOp, CLinkerOp):
         node_output_storage = [storage_map[r] for r in node.outputs]
         node_input_compute = [compute_map[r] for r in node.inputs]
         node_output_compute = [compute_map[r] for r in node.outputs]
-        #logger.debug('Compiling node %i of graph' % node_idx)
+
         if self._op_use_c_code:
             try:
+                # float16 get special treatment since running
+                # unprepared C code will get bad results.
+                if not getattr(self, '_f16_ok', False):
+                    def is_f16(t):
+                        return getattr(t, 'dtype', '') == 'float16'
+
+                    if (any(is_f16(i.type) for i in node.inputs) or
+                            any(is_f16(o.type) for o in node.outputs)):
+                        print ("Disabling C code for %s due to unsupported "
+                               "float16" % (self,))
+                        raise NotImplementedError("float16")
                 e = FunctionGraph(node.inputs, node.outputs)
 
                 e_no_recycling = [new_o

theano/sandbox/gpuarray/basic_ops.py

@@ -187,11 +187,8 @@ class GpuKernelBase(object):
 
 
 class HostFromGpu(Op):
-    def __eq__(self, other):
-        return type(self) == type(other)
-
-    def __hash__(self):
-        return hash(type(self))
+    __props__ = ()
+    _f16_ok = True
 
     def __str__(self):
         return 'HostFromGpu(gpuarray)'
@@ -270,11 +267,8 @@ host_from_gpu = HostFromGpu()
 
 
 class GpuFromHost(Op):
-    def __eq__(self, other):
-        return type(self) == type(other)
-
-    def __hash__(self):
-        return hash(type(self))
+    __props__ = ()
+    _f16_ok = True
 
     def __str__(self):
         return 'GpuFromHost(gpuarray)'
@@ -574,18 +568,15 @@ cuda_from_gpu = CudaFromGpu()
 
 
 class GpuAlloc(HideC, Alloc):
+    __props__ = ('memset_0',)
+    _f16_ok = True
+
     def __init__(self, memset_0=False):
         """memset_0 is only an optimized version. True, it mean the
         value is always 0, so the c code call memset as it is faster.
         """
         self.memset_0 = memset_0
 
-    def __eq__(self, other):
-        return type(self) == type(other) and self.memset_0 == other.memset_0
-
-    def __hash__(self):
-        return hash(type(self)) ^ hash(self.memset_0)
-
     def __str__(self):
         # Hide the memset parameter when not used to prevent confusion.
         if self.memset_0:
@@ -729,25 +720,17 @@ class GpuContiguous(Op):
     Always return a c contiguous output. Copy the input only if it is
     not already c contiguous.
     """
+    __props__ = ()
     view_map = {0: [0]}
-
-    def __eq__(self, other):
-        return type(self) == type(other)
-
-    def __hash__(self):
-        return hash(type(self))
+    _f16_ok = True
 
     def grad(self, inputs, dout):
-
         x, = inputs
         dout, = dout
         dout = as_gpuarray_variable(dout)
 
         return [dout]
 
-    def __str__(self):
-        return self.__class__.__name__
-
     def make_node(self, input):
         input = as_gpuarray_variable(input)
         return Apply(self, [input], [input.type()])
@@ -795,6 +778,8 @@ class GpuReshape(HideC, tensor.Reshape):
     """
     Implement Reshape on the gpu.
     """
+    _f16_ok = True
+
     # __hash__, __eq__, __str__ come from tensor.Reshape
     def make_node(self, x, shp):
         x = as_gpuarray_variable(x)
@@ -832,6 +817,8 @@ class GpuReshape(HideC, tensor.Reshape):
 
 
 class GpuJoin(HideC, Join):
+    _f16_ok = True
+
     def make_node(self, axis, *tensors):
         node = Join.make_node(self, axis, *tensors)
 
@@ -890,6 +877,7 @@ class GpuSplit(HideC, Split):
 
 class GpuEye(GpuKernelBase, Op):
     __props__ = ('dtype',)
+    _f16_ok = True
 
     def __init__(self, dtype=None):
         if dtype is None:

theano/sandbox/gpuarray/elemwise.py

@@ -61,6 +61,7 @@ def as_C_string_const(s):
 class GpuElemwise(HideC, Elemwise):
     nin = property(lambda self: self.scalar_op.nin)
     nout = property(lambda self: self.scalar_op.nout)
+    _f16_ok = True
 
     def __str__(self):
         if self.name is not None:
@@ -475,6 +476,8 @@ class SupportCodeError(Exception):
 
 
 class GpuDimShuffle(HideC, DimShuffle):
+    _f16_ok = True
+
     def make_node(self, input):
         res = DimShuffle.make_node(self, input)
         otype = GpuArrayType(dtype=res.outputs[0].type.dtype,
@@ -602,8 +605,8 @@ class GpuCAReduceCuda(HideC, CAReduceDtype):
 
     pre_scalar_op: if present, must be a scalar op with only 1
     input. We will execute it on the input value before reduction.
-
     """
+    _f16_ok = True
 
     def __init__(self, scalar_op, axis=None,
                  reduce_mask=None, dtype=None, acc_dtype=None,

theano/sandbox/gpuarray/nnet.py

@@ -25,15 +25,8 @@ class GpuCrossentropySoftmaxArgmax1HotWithBias(Op):
     """
     nin = 3
     nout = 3
-
-    def __eq__(self, other):
-        return type(self) == type(other)
-
-    def __hash__(self):
-        return hash(type(self))
-
-    def __str__(self):
-        return self.__class__.__name__
+    __props__ = ()
+    _f16_ok = True
 
     def make_node(self, x, b, y_idx):
         # N.B. won't work when we don't cast y_idx to float anymore
@@ -282,21 +275,12 @@ gpu_crossentropy_softmax_argmax_1hot_with_bias = GpuCrossentropySoftmaxArgmax1Ho
 class GpuCrossentropySoftmax1HotWithBiasDx(Op):
     """
     Implement CrossentropySoftmax1HotWithBiasDx on the gpu.
+
+    Gradient wrt x of the CrossentropySoftmax1Hot Op
     """
     nin = 3
     nout = 1
-    """Gradient wrt x of the CrossentropySoftmax1Hot Op"""
-    def __init__(self, **kwargs):
-        Op.__init__(self, **kwargs)
-
-    def __eq__(self, other):
-        return type(self) == type(other)
-
-    def __hash__(self):
-        return hash(type(self))
-
-    def __str__(self):
-        return self.__class__.__name__
+    __props__ = ()
 
     def make_node(self, dnll, sm, y_idx):
         dnll = as_gpuarray_variable(dnll)

theano/sandbox/gpuarray/subtensor.py

@@ -21,6 +21,8 @@ from .comp import NVCC_compiler
 
 
 class GpuSubtensor(HideC, Subtensor):
+    _fp16_ok = True
+
     def make_node(self, x, *inputs):
         rval = tensor.Subtensor.make_node(self, x, *inputs)
         otype = GpuArrayType(dtype=rval.outputs[0].type.dtype,

theano/tensor/elemwise.py

@@ -95,6 +95,7 @@ class DimShuffle(Op):
     transpose function.
     Adding, subtracting dimensions can be done with reshape.
     """
+    _f16_ok = True
 
     check_input = False