Improved existing 16->32 fallback. Added performance-based fallback.

soumith · Nov 10, 2016 · f746b79 · f746b79
1 parent d6e0ce5
commit f746b79
Show file tree

Hide file tree

Showing 10 changed files with 174 additions and 145 deletions.
diff --git a/SpatialConvolution.lua b/SpatialConvolution.lua
@@ -127,12 +127,13 @@ function SpatialConvolution:createIODescriptors(input)
         self.pad = {self.padH, self.padW}
         self.stride = {self.dH, self.dW}
 
-        self.convDesc = cudnn.setConvolutionDescriptor(
-           { padA = self.pad,
+        self.convDescData = { padA = self.pad,
              filterStrideA = self.stride,
              upscaleA = {1,1},
              dataType = cudnn.configmap(torch.type(self.weight))
-           })
+        }
+
+        self.convDesc = cudnn.setConvolutionDescriptor(self.convDescData)
 
         -- get output shape, resize output
         local oSize = torch.IntTensor(4)

diff --git a/SpatialFullConvolution.lua b/SpatialFullConvolution.lua
@@ -47,11 +47,11 @@ function SpatialFullConvolution:createIODescriptors(input)
         self.pad = {self.padH, self.padW}
         self.stride = {self.dH, self.dW}
 
-        self.convDesc = cudnn.setConvolutionDescriptor(
-           { padA = self.pad,
-             filterStrideA = self.stride,
-             dataType = cudnn.configmap(torch.type(self.weight))
-           })
+        self.convDescData = { padA = self.pad,
+                              filterStrideA = self.stride,
+                              dataType = cudnn.configmap(torch.type(self.weight))
+        }
+        self.convDesc = cudnn.setConvolutionDescriptor(self.convDescData)
 
         -- get output shape, resize output
         local iwidth = input:size(4)

diff --git a/TemporalConvolution.lua b/TemporalConvolution.lua
@@ -37,7 +37,7 @@ function TemporalConvolution:createIODescriptors(input)
 end
 
 function TemporalConvolution:fastest(mode)
-    self = cudnn.SpatialConvolution.fastest(self,mode)
+    cudnn.SpatialConvolution.fastest(self,mode)
     return self
 end
 

diff --git a/VolumetricConvolution.lua b/VolumetricConvolution.lua
@@ -43,10 +43,9 @@ function VolumetricConvolution:createIODescriptors(input)
          if mathtype == 'CUDNN_DATA_HALF' then
             mathtype = 'CUDNN_DATA_FLOAT'
          end
-         self.convDesc = cudnn.setConvolutionDescriptor(
-            { padA = self.pad, filterStrideA = self.stride,
-              dataType = mathtype
-            })
+         self.convDescData = { padA = self.pad, filterStrideA = self.stride,
+                               dataType = mathtype }
+         self.convDesc = cudnn.setConvolutionDescriptor(self.convDescData)
 
          local oSize = torch.IntTensor(5)
          errcheck('cudnnGetConvolutionNdForwardOutputDim',

diff --git a/VolumetricFullConvolution.lua b/VolumetricFullConvolution.lua
@@ -45,10 +45,9 @@ function VolumetricFullConvolution:createIODescriptors(input)
          -- create conv descriptor
          self.pad = {self.padT, self.padH, self.padW}
          self.stride = {self.dT, self.dH, self.dW}
-         self.convDesc = cudnn.setConvolutionDescriptor(
-            { padA = self.pad, filterStrideA = self.stride,
-              dataType = cudnn.configmap(torch.type(self.weight))
-            })
+         self.convDescData = { padA = self.pad, filterStrideA = self.stride,
+                               dataType = cudnn.configmap(torch.type(self.weight))}
+         self.convDesc = cudnn.setConvolutionDescriptor(self.convDescData)
 
         -- get output shape, resize output
         local iwidth = input:size(5)

diff --git a/ffi.lua b/ffi.lua
@@ -1614,10 +1614,10 @@ end
 
 -- check cuDNN version
 cudnn.version = tonumber(cudnn.C.cudnnGetVersion())
-if cudnn.version < 5005 then
-  error('These bindings are for version 5005 or above, '
+if cudnn.version < 5005 or cudnn.version >= 6000 then
+   error('These bindings are for CUDNN 5.x (5005 <= cudnn.version > 6000) , '
         .. 'while the loaded CuDNN is version: ' .. cudnn.version
-           .. '  \nAre you using an older version of CuDNN?')
+           .. '  \nAre you using an older or newer version of CuDNN?')
 end
 
 -- check GPU driver version