added an optimization phase: specialize_device and put the new POW optimization into that phase.

theano/compile/mode.py

@@ -137,6 +137,8 @@ optdb.register('stabilize', gof.EquilibriumDB(),            # replace unstable s
         1.5, 'fast_run')          
 optdb.register('specialize', gof.EquilibriumDB(),           # misc special cases for speed
         2, 'fast_run')
+optdb.register('specialize_device', gof.EquilibriumDB(),           # misc special cases for speed that are dependent on the device.
+        48.6, 'fast_run')#must be after gpu stuff at 48.5
 optdb.register('merge2', gof.MergeOptimizer(),              # especially constant merge
         49, 'fast_run')
 optdb.register('add_destroy_handler', AddDestroyHandler(), 

theano/tensor/opt.py

@@ -144,6 +144,11 @@ def register_specialize(lopt, *tags, **kwargs):
     compile.optdb['specialize'].register(name, lopt, 'fast_run', *tags)
     return lopt
 
+def register_specialize_device(lopt, *tags, **kwargs):
+    name = (kwargs and kwargs.pop('name')) or lopt.__name__
+    compile.optdb['specialize_device'].register(name, lopt, 'fast_run', *tags)
+    return lopt
+
 def register_stabilize(lopt, *tags, **kwargs):
     name = (kwargs and kwargs.pop('name')) or lopt.__name__
     compile.optdb['stabilize'].register(name, lopt, 'fast_run', *tags)
@@ -1829,9 +1834,31 @@ def local_pow_specialize(node):
                 rval = [T.inv(xsym)]
             if N.all(y == -2):
                 rval = [T.inv(T.sqr(xsym))]
+            if rval:
+                rval[0] = T.cast(rval[0], odtype)
+                assert rval[0].type == node.outputs[0].type, (rval, node.outputs)
+                return rval
+    else:
+        return False
+register_specialize(local_pow_specialize)
 
-            # Optimize all integral powers in [-RANGE, RANGE]
-            if config.experimental.pow and rval is None and abs(y)==int(abs(y)) and abs(y) <= 512:# 512 is too small for the cpu and too big for some gpu!
+@register_specialize_device
+@gof.local_optimizer([T.pow])
+def local_pow_specialize_device(node):
+    """
+    This optimization is not the same on all device. We do it only on cpu here.
+    """
+    if node.op == T.pow:
+        #the idea here is that we have pow(x, y)
+        odtype = node.outputs[0].dtype
+        xsym = node.inputs[0]
+        ysym = node.inputs[1]
+        y = local_mul_canonizer.get_constant(ysym)
+        if (y is not None) \
+                and encompasses_broadcastable(xsym.type.broadcastable, ysym.type.broadcastable):
+            rval = None
+            # 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)()]
                 y_to_do = abs(y)
@@ -1861,14 +1888,7 @@ def local_pow_specialize(node):
                 rval[0] = T.cast(rval[0], odtype)
                 assert rval[0].type == node.outputs[0].type, (rval, node.outputs)
                 return rval
-    else:
-        return False
-register_specialize(local_pow_specialize)
-theano.configparser.AddConfigVar('experimental.pow',
-        "Transform a pow to a constant integer to a graph of mul. Fast on cpu, but more work needed for gpu.",
-        theano.configparser.BoolParam(False),
-        )
-
+   
 @gof.local_optimizer([T.mul])
 def local_mul_specialize(node):
     """Remove special-case constants from mul arguments

theano/tensor/tests/test_opt.py

@@ -1259,6 +1259,7 @@ def test_local_pow_specialize():
     v = T.vector()
     val = numpy.arange(10,dtype=theano.config.floatX)
     val_no0 = numpy.arange(1,10,dtype=theano.config.floatX)
+
     f = function([v], v**0, mode=mode)
     nodes = [node.op for node in f.maker.env.toposort()]
     assert nodes == [Shape_i(0), T.alloc]
@@ -1300,33 +1301,44 @@ def test_local_pow_specialize():
 #    assert nodes == [T.sqrt,T.inv]#Why this don't work?
     assert numpy.allclose(f(val_no0),val_no0**(-.5))
 
-    if config.experimental.pow:
-        print "Test experimental.pow=True"
-        f = function([v], v**(15), mode=mode)
-        nodes = [node.op for node in f.maker.env.toposort()]
-        assert len(nodes)==1
-        assert isinstance(nodes[0].scalar_op,theano.scalar.Composite)
-        assert numpy.allclose(f(val),val**15)
-        
-        f = function([v], v**(-15), mode=mode)
-        nodes = [node.op for node in f.maker.env.toposort()]
-        assert len(nodes)==2
-        assert isinstance(nodes[0].scalar_op,theano.scalar.Composite)
-        assert isinstance(nodes[-1].scalar_op,theano.scalar.basic.Inv)
-        assert numpy.allclose(f(val_no0),val_no0**(-15))
-        
-        f = function([v], v**(16), mode=mode)
-        nodes = [node.op for node in f.maker.env.toposort()]
-        assert len(nodes) == 1
-        assert isinstance(nodes[0].scalar_op,theano.scalar.Composite)
-        assert numpy.allclose(f(val),val**16)
-        
-        f = function([v], v**(-16), mode=mode)
-        nodes = [node.op for node in f.maker.env.toposort()]
-        assert len(nodes) == 2
-        assert isinstance(nodes[0].scalar_op,theano.scalar.Composite)
-        assert isinstance(nodes[-1].scalar_op,theano.scalar.basic.Inv)
-        assert numpy.allclose(f(val_no0),val_no0**(-16))
+def test_local_pow_specialize_device():
+
+    # test that on cpu we use more agressive optimization
+
+    mode = theano.config.mode
+    if mode == 'FAST_COMPILE':
+       mode = 'FAST_RUN'
+    mode = compile.mode.get_mode(mode)
+    mode = mode.excluding('fusion').excluding('gpu')
+
+    v = T.vector()
+    val = numpy.arange(10,dtype=theano.config.floatX)
+    val_no0 = numpy.arange(1,10,dtype=theano.config.floatX)
+    f = function([v], v**(15), mode=mode)
+    nodes = [node.op for node in f.maker.env.toposort()]
+    assert len(nodes)==1
+    assert isinstance(nodes[0].scalar_op,theano.scalar.Composite)
+    assert numpy.allclose(f(val),val**15)
+    
+    f = function([v], v**(-15), mode=mode)
+    nodes = [node.op for node in f.maker.env.toposort()]
+    assert len(nodes)==2
+    assert isinstance(nodes[0].scalar_op,theano.scalar.Composite)
+    assert isinstance(nodes[-1].scalar_op,theano.scalar.basic.Inv)
+    assert numpy.allclose(f(val_no0),val_no0**(-15))
+    
+    f = function([v], v**(16), mode=mode)
+    nodes = [node.op for node in f.maker.env.toposort()]
+    assert len(nodes) == 1
+    assert isinstance(nodes[0].scalar_op,theano.scalar.Composite)
+    assert numpy.allclose(f(val),val**16)
+    
+    f = function([v], v**(-16), mode=mode)
+    nodes = [node.op for node in f.maker.env.toposort()]
+    assert len(nodes) == 2
+    assert isinstance(nodes[0].scalar_op,theano.scalar.Composite)
+    assert isinstance(nodes[-1].scalar_op,theano.scalar.basic.Inv)
+    assert numpy.allclose(f(val_no0),val_no0**(-16))
     
 class T_Rebroadcast(unittest.TestCase):