added _dot22 and some gemm optimizations

theano/tensor/basic.py

@@ -1089,45 +1089,11 @@ pprint.assign(pow, printing.OperatorPrinter('**', 1, 'right'))
 # View Operations
 ##########################
 
-if 0:
-    class _TransposeInplace(Op):
-        view_map = {0: [0]}
-        
-        def make_node(self, input):
-            return Apply(self, [input], [tensor(dtype = input.type.dtype,
-                                                broadcastable = reversed(input.type.broadcastable))])
-        
-        def perform(self, node, (x, ), (z, )):
-            z[0] = x.T
-        
-        def grad(self, (x,), (gz,)):
-            return transpose(gz),
-        
-        def c_code(self, node, name, (x, ), (z, ), sub):
-            return """
-            PyArrayObject* transposed = (PyArrayObject*)PyArray_Transpose(%(x)s, NULL);
-            if (%(z)s) {
-                Py_XDECREF(%(z)s);
-            }
-            %(z)s = transposed;
-            """ % locals()
-
-        def __str__(self):
-            return "TransposeView"
-
-    _transpose_inplace = _TransposeInplace()
-
-    def _old_transpose(x, **kwargs):
-        """WRITEME"""
-        return _transpose_inplace(tensor_copy(x), **kwargs)
-
 def transpose(x, **kwargs):
     dims = range(x.ndim-1, -1, -1)
     return DimShuffle(x.broadcastable, dims, inplace=True)(tensor_copy(x))
 
 
-
-
 class Subtensor(Op):
     """Return a subtensor view
 
@@ -1786,6 +1752,7 @@ class Dot(Op):
             # The error raised by numpy has no shape information, we mean to add that
             e.args = e.args + (x.shape, y.shape)
             raise
+
     def grad(self, (x, y), (gz,)):
         if gz.type.ndim == 0:
             return gz * y, gz * x

theano/tensor/opt.py

@@ -625,6 +625,37 @@ def local_pow_specialize(node):
         return False
 register_specialize(local_pow_specialize)
 
+if 0: #TODO: replace this with a c version of any InplaceDimShuffle
+    class _TransposeInplace(T.Op):
+        view_map = {0: [0]}
+        
+        def make_node(self, input):
+            return T.Apply(self, [input], 
+                    [T.tensor(dtype = input.type.dtype,
+                        broadcastable = reversed(input.type.broadcastable))])
+        
+        def perform(self, node, (x, ), (z, )):
+            z[0] = x.T
+        
+        def c_code(self, node, name, (x, ), (z, ), sub):
+            return """
+            PyArrayObject* transposed = (PyArrayObject*)PyArray_Transpose(%(x)s, NULL);
+            if (%(z)s) {
+                Py_XDECREF(%(z)s);
+            }
+            %(z)s = transposed;
+            """ % locals()
+
+        def __str__(self):
+            return "_TransposeInplace"
+    _transpose_inplace = _TransposeInplace()
+
+    @gof.local_optimizer([T.DimShuffle([False,False],[1,0],inplace=True)])
+    def local_dimshuffle_transposeinplace(node):
+        if node.op == T.DimShuffle([False,False],[1,0],inplace=True):
+            return [_transpose_inplace(node.inputs[0])]
+        return False
+    register_specialize(local_dimshuffle_transposeinplace)
 
 register_canonicalize(local_mul_canonizer, name = 'local_mul_canonizer')