Merge remote-tracking branch 'upstream/master' into 17.03-devel

Conflicts:
	init.lua

3 files changed, 20 insertions, 5 deletions

diff --git a/find.lua b/find.lua
index f1c783f..045ed86 100644
--- a/find.lua
+++ b/find.lua
@@ -110,7 +110,7 @@ local function fallbackWarning(layer, msg)
             "\n *** Falling back to 32-bit math for: " .. convDataString(layer))
       print(" *** [ Set cudnn.find.verboseFallback to false to disable this message ] *** ")
       print(" *** [ Alternatively, you may force CUDNN to always operate on CudaHalfTensors via 32-bit float conversion, in Lua: ] ***\n"
-               .." *** [ cudnn.configureMath({ ['torch.CudaHalfTensor']   = 'CUDNN_DATA_FLOAT'} ] ***")
+               .." ***  cudnn.configureMath({ ['torch.CudaHalfTensor']   = 'CUDNN_DATA_FLOAT'} ) ***")
       print(" *** [ Note: result may be faster or slower than native FP16, depending on your GPU and CUDNN operations ] *** ")
    end
 end
diff --git a/init.lua b/init.lua
index 20fa154..cbd434b 100644
--- a/init.lua
+++ b/init.lua
@@ -216,11 +216,21 @@ function cudnn.setConvolutionDescriptor(data, desc)
    local myDesc = desc or cudnn.createDescriptors(
       1, 'struct cudnnConvolutionStruct*[?]',
       'cudnnCreateConvolutionDescriptor', 'cudnnDestroyConvolutionDescriptor')
+   -- make sure we have references to these tensors so gc doesn't clean them up
+   local padATensor = torch.IntTensor(data.padA)
+   local filterStrideATensor = torch.IntTensor(data.filterStrideA)
+   local upscaleATensor = torch.IntTensor(data.upscaleA)
    errcheck('cudnnSetConvolutionNdDescriptor', myDesc[0],
             data.arrayLength,
+<<<<<<< HEAD
             torch.IntTensor(data.padA):data(),
             torch.IntTensor(data.filterStrideA):data(),
             torch.IntTensor(data.dilationA):data(),
+=======
+            padATensor:data(),
+            filterStrideATensor:data(),
+            upscaleATensor:data(),
+>>>>>>> upstream/master
             data.mode,
             data.dataType)
    return myDesc
@@ -231,9 +241,11 @@ function cudnn.setFilterDescriptor(data, filterDesc)
       1, 'struct cudnnFilterStruct*[?]',
       'cudnnCreateFilterDescriptor', 'cudnnDestroyFilterDescriptor')
    local dims = data.nbDims or #data.filterDimA
+   -- make sure we have references to these tensors so gc doesn't clean them up
+   local filterDimATensor = torch.IntTensor(data.filterDimA)
    errcheck('cudnnSetFilterNdDescriptor', myDesc[0],
             data.dataType, data.format or 'CUDNN_TENSOR_NCHW',
-            dims, torch.IntTensor(data.filterDimA):data());
+            dims, filterDimATensor:data());
    return myDesc
 end
 
diff --git a/test/test.lua b/test/test.lua
index c641f29..ffe6cc3 100644
--- a/test/test.lua
+++ b/test/test.lua
@@ -77,9 +77,10 @@ local function testLayer(nnlayer, cudnnlayer, input, gradOutput, scale,
    end
 
    local res = {} -- result
-   res.output = cudnnlayer:forward(cast(input))
+   inputcudnn = cast(input):clone() -- for inplace layers
+   res.output = cudnnlayer:forward(inputcudnn)
    cudnnlayer:zeroGradParameters()
-   res.gradInput = cudnnlayer:backward(cast(input), cast(gradOutput), scale)
+   res.gradInput = cudnnlayer:backward(inputcudnn, cast(gradOutput), scale)
    if parametric then
       res.gradWeight = cudnnlayer.gradWeight
       res.gradBias = cudnnlayer.gradBias
@@ -550,9 +551,11 @@ function cudnntest.ReLU()
    nonlin('ReLU', false) -- out of place
 end
 function cudnntest.Tanh()
+   nonlin('Tanh', true) -- inplace
    nonlin('Tanh', false) -- out of place
 end
 function cudnntest.Sigmoid()
+   nonlin('Sigmoid', true) -- inplace
    nonlin('Sigmoid', false) -- out of place
 end
 
@@ -1021,7 +1024,7 @@ cudnn.find.verbose=false
 -- todo: put it back for release to demo 16->32 bit float fallback
 cudnn.find.verboseFallback=false
 cudnn.useFindEx=false
-
+cudnn.configureMath({ ['torch.CudaHalfTensor']   = 'CUDNN_DATA_FLOAT'} )
 for i = 1, 1 do -- cutorch.getDeviceCount() do
 
    for _, benchmark, fast in ipairs({true, false}) do