Add :cudaOn() and TransferGPU module.

With cutorch auto-mode, multi-gpu models can be run without explicitly
updating the device. However, copies must be explicit, so a
TransferGPU module is necessary to make the copy.

Here's an example multi-GPU module (added as a test):

```lua
cutorch.setDevice(0)

net = nn.Sequential()
for i = 1,3 do
  net:add(nn.Linear(1000,1000):cudaOn(i))
  net:add(nn.SoftMax():cudaOn(i))
  net:add(nn.TransferGPU(i, i+1))
end

local input = torch.CudaTensorOn(1, 1000)
local output = net:forward(input)
print(output:getDevice())

local gradOutput = output/100
local gradInput = net:backward(input, gradOutput)
print(gradInput:getDevice())
```

1 files changed, 10 insertions, 0 deletions

diff --git a/init.lua b/init.lua
index 0145570..0f2e2ab 100644
--- a/init.lua
+++ b/init.lua
@@ -3,5 +3,15 @@ require "nn"
 require "libcunn"
 
 include('test.lua')
+include('utils.lua')
 
 include('DataParallelTable.lua')
+include('TransferGPU.lua')
+
+function nn.Module:cudaOn(device)
+  return nn.utils.recursiveCudaOn(self, device)
+end
+
+function nn.Criterion:cudaOn(device)
+  return nn.utils.recursiveCudaOn(self, device)
+end