Optimizer in progress. This commit is just to allow Fred to give some feedback.

theano/sandbox/cuda/opt.py

@@ -63,6 +63,7 @@ from theano.tensor import nlinalg
 from theano.tensor import slinalg
 
 from theano.tensor.nnet.Conv3D import Conv3D
+from theano.tests.breakpoint import PdbBreakpoint
 
 try:
     # We need to be able to import this file even if cuda isn't avail.
@@ -1140,6 +1141,43 @@ def local_gpu_print_op(node):
             return [host_from_gpu(new_op(gpu_x))]
     return False
 
+@register_opt()
+@local_optimizer([PdbBreakpoint])
+def local_gpu_pdbbreakpoint_op(node):
+    if isinstance(node.op, PdbBreakpoint):
+
+        old_inputs = node.inputs
+
+        # Obtain the inputs to the new op. The condition (first input) should
+        # be left on the host but the other inputs can be taken from the GPU.
+        new_inputs = old_inputs[:1]
+        for inp in old_inputs[1:]:
+            if inp.owner and isinstance(inp.owner.op, HostFromGpu):
+                # Take the input directly from the gpu
+                new_inputs.append(inp.owner.inputs[0])
+            else:
+                new_inputs.append(inp)
+                
+        # Only proceed further if one of the outputs to the op was a
+        # HostFromGpu
+        if new_inputs[1:] == old_inputs[1:]:
+            return False
+            
+        # Apply the op on the new inputs
+        new_outputs = node.op(*new_inputs)
+
+        # For every output of the new op for which we took the corresponding
+        # input from the GPU instead of the host, we need to transfer the
+        # output back to the host before returning it.
+        for i in range(len(new_outputs)):
+            inp = old_inputs[i + 1]
+            if (inp.owner and isinstance(inp.owner.op, HostFromGpu)):
+                new_outputs[i] = host_from_gpu(new_outputs[i])
+
+        return new_outputs
+
+    return False
+
 
 def cast(x, dtype):
     stype = scal.Scalar(dtype)

theano/tests/test_breakpoint.py

@@ -59,6 +59,21 @@ class TestPdbBreakpoint(utt.InferShapeTester):
 
         for i in range(len(gradients)):
             numpy.testing.assert_allclose(gradients[i], expected_gradients[i])
+            
+    def test_fprop(self):
+        
+        input1_value = numpy.arange(9).reshape(3,3).astype("float32")
+        input2_value = 10.0
+        fct = theano.function([self.input1, self.input2],
+                              [self.monitored_input1, self.monitored_input2])
+
+        output = fct(input1_value, input2_value)[:-1]
+        
+        import pdb
+        pdb.set_trace()
+
+        expected_output = [numpy.ones((3, 3), dtype="float32"),
+                           numpy.array(1., dtype="float32")]
 
     def test_connection_pattern(self):