Merge pull request #616 from nouiz/gpu_conv_faster

theano/sandbox/cuda/__init__.py

@@ -309,8 +309,15 @@ def use(device,
             device = 0
         try:
             if device != 'gpu':
+                assert isinstance(device, int)
                 gpu_init(device)
-            use.device_number = device
+                use.device_number = device
+            else:
+                # This mean we let the driver select the GPU.
+                # But default it is always number 0.
+                # If the driver is in exclusive mode, it will always show
+                # device 0 event if it use something else.
+                use.device_number = 0
             if test_driver:
                 import theano.sandbox.cuda.tests.test_driver
                 theano.sandbox.cuda.tests.test_driver.test_nvidia_driver1()

theano/sandbox/cuda/blas.py

@@ -718,10 +718,11 @@ class GpuConv(GpuOp):
         node_ = copy.copy(node)
         assert node.op is node_.op
         if node_.op.max_threads_dim0 is None:
-            op = copy.copy(node_.op)
-            device_id = theano.sandbox.cuda.use.device_number[3:]
-            if device_id == '':
-                device_id = 0
+            cuda = theano.sandbox.cuda
+            device_id = cuda.use.device_number
+            if device_id is None:
+                cuda.use("gpu", False, False, False, False, True)
+                device_id = cuda.use.device_number
             cuda_ndarray = theano.sandbox.cuda.cuda_ndarray.cuda_ndarray
             prop = cuda_ndarray.device_properties(device_id)
             node_.op.max_threads_dim0 = prop['maxThreadsDim0']