Added comments to Erfinv=>ErfinvGPU optimization

theano/sandbox/cuda/elemwise.py

@@ -1025,7 +1025,15 @@ nd_collapse_[i]=0;
 
 
 class ErfinvGPU(Erfinv):
-
+    """
+    Provides a c-code implementation of the inverse error function for GPU.
+    
+    Note: We do not add this c_code to theano.scalar.basic_scipy.Erfinv, as we
+    currently rely on Nvidia's cublas library to provide the erfinv
+    c-implementation (which requires different c_headers). As it stands,
+    theano.scalar.basic_scipy.Erfinv does not have c_code as scipy does not
+    export the required C function
+    """
     def c_libraries(self):
         return ['math.h']
 

theano/scalar/basic_scipy.py

@@ -78,6 +78,15 @@ erfc = Erfc(upgrade_to_float_no_complex, name='erfc')
 
 
 class Erfinv(UnaryScalarOp):
+    """
+    Implements the inverse error function.
+
+    Note: This op can still be executed on GPU, despite not having c_code.  When
+    running on GPU, sandbox.cuda.opt.local_gpu_elemwise_[0,1] replaces this op
+    with sandbox.cuda.elemwise.ErfinvGPU.
+
+    (TODO) Find a C implementation of erfinv for CPU.
+    """
     def impl(self, x):
         if imported_scipy_special:
             return scipy.special.erfinv(x)