Merge pull request #3572 from carriepl/fix_cudnn_index_errors

theano/sandbox/cuda/dnn_fwd.c

@@ -188,12 +188,12 @@ APPLY_SPECIFIC(conv_fwd)(CudaNdarray *input, CudaNdarray *kerns,
       }
 
       // Extract the spatial size of the filters
-      int filter_h = CudaNdarray_HOST_DIMS(kerns)[3];
-      int filter_w = CudaNdarray_HOST_DIMS(kerns)[4];
+      int filter_h = CudaNdarray_HOST_DIMS(kerns)[2];
+      int filter_w = CudaNdarray_HOST_DIMS(kerns)[3];
 
       // Extract the spatial size of the input
-      int input_h = CudaNdarray_HOST_DIMS(input)[3];
-      int input_w = CudaNdarray_HOST_DIMS(input)[4];
+      int input_h = CudaNdarray_HOST_DIMS(input)[2];
+      int input_w = CudaNdarray_HOST_DIMS(input)[3];
 
       // Ensure that the selected implementation supports the requested
       // convolution. Fall back to a safe implementation otherwise.

theano/sandbox/cuda/dnn_gi.c

@@ -183,12 +183,12 @@ APPLY_SPECIFIC(conv_gi)(CudaNdarray *kerns, CudaNdarray *output,
       }
 
       // Extract the spatial size of the filters
-      int filter_h = CudaNdarray_HOST_DIMS(kerns)[3];
-      int filter_w = CudaNdarray_HOST_DIMS(kerns)[4];
+      int filter_h = CudaNdarray_HOST_DIMS(kerns)[2];
+      int filter_w = CudaNdarray_HOST_DIMS(kerns)[3];
 
       // Extract the spatial size of the input
-      int input_h = CudaNdarray_HOST_DIMS(*input)[3];
-      int input_w = CudaNdarray_HOST_DIMS(*input)[4];
+      int input_h = CudaNdarray_HOST_DIMS(*input)[2];
+      int input_w = CudaNdarray_HOST_DIMS(*input)[3];
 
       // Ensure that the selected implementation supports the requested
       // convolution. Fall back to a safe implementation otherwise.

theano/sandbox/cuda/dnn_gw.c

@@ -183,12 +183,12 @@ APPLY_SPECIFIC(conv_gw)(CudaNdarray *input, CudaNdarray *output,
       }
 
       // Extract the spatial size of the filters
-      int filter_h = CudaNdarray_HOST_DIMS(*kerns)[3];
-      int filter_w = CudaNdarray_HOST_DIMS(*kerns)[4];
+      int filter_h = CudaNdarray_HOST_DIMS(*kerns)[2];
+      int filter_w = CudaNdarray_HOST_DIMS(*kerns)[3];
 
       // Extract the spatial size of the input
-      int input_h = CudaNdarray_HOST_DIMS(input)[3];
-      int input_w = CudaNdarray_HOST_DIMS(input)[4];
+      int input_h = CudaNdarray_HOST_DIMS(input)[2];
+      int input_w = CudaNdarray_HOST_DIMS(input)[3];
 
       // Ensure that the selected implementation supports the requested
       // convolution. Fall back to a safe implementation otherwise.

theano/sandbox/gpuarray/dnn_fwd.c

@@ -154,8 +154,8 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
     }
 
     if (stride[0] != 1 || stride[1] != 1 ||
-        PyGpuArray_DIM(input, 0) > 1024 || PyGpuArray_DIM(input, 1) > 1024 ||
-        (PyGpuArray_DIM(kerns, 0) == 1 && PyGpuArray_DIM(kerns, 1) == 1)) {
+        PyGpuArray_DIM(input, 2) > 1024 || PyGpuArray_DIM(input, 3) > 1024 ||
+        (PyGpuArray_DIM(kerns, 2) == 1 && PyGpuArray_DIM(kerns, 3) == 1)) {
       algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM;
     }
   }

theano/sandbox/gpuarray/dnn_gi.c

@@ -146,8 +146,8 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
     }
 
     if (stride[0] != 1 || stride[1] != 1 ||
-        PyGpuArray_DIM(*input, 0) > 1024 || PyGpuArray_DIM(*input, 1) > 1024 ||
-        (PyGpuArray_DIM(kerns, 0) == 1 && PyGpuArray_DIM(kerns, 1) == 1)) {
+        PyGpuArray_DIM(*input, 2) > 1024 || PyGpuArray_DIM(*input, 3) > 1024 ||
+        (PyGpuArray_DIM(kerns, 2) == 1 && PyGpuArray_DIM(kerns, 3) == 1)) {
       algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_0;
     }
   }

theano/sandbox/gpuarray/dnn_gw.c

@@ -148,8 +148,8 @@ APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
     }
 
     if (stride[0] != 1 || stride[1] != 1 ||
-        PyGpuArray_DIM(input, 0) > 1024 || PyGpuArray_DIM(input, 1) > 1024 ||
-        (PyGpuArray_DIM(*kerns, 0) == 1 && PyGpuArray_DIM(*kerns, 1) == 1)) {
+        PyGpuArray_DIM(input, 2) > 1024 || PyGpuArray_DIM(input, 3) > 1024 ||
+        (PyGpuArray_DIM(*kerns, 2) == 1 && PyGpuArray_DIM(*kerns, 3) == 1)) {
       algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0;
     }
   }