Move the autoselection variables to be per-thunk rather than per-program since…

theano/sandbox/gpuarray/dnn_fwd.c

+#section init_code_struct
+
+#ifdef CHOOSE_ALGO
+reuse_algo = 0;
+prev_algo = CONV_ALGO;
+#ifndef CHOOSE_ONCE
+memset(prev_img_dims, 0, sizeof(prev_img_dims));
+memset(prev_kern_dims, 0, sizeof(prev_kern_dims));
+#endif
+#endif
+
 #section support_code_struct
 
+#ifdef CHOOSE_ALGO
+int reuse_algo;
+cudnnConvolutionFwdAlgo_t prev_algo;
+#ifndef CHOOSE_ONCE
+size_t prev_img_dims[5];
+size_t prev_kern_dims[5];
+#endif
+#endif
+
 int
 APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
                          PyGpuArrayObject *om,
@@ -57,15 +77,7 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
 
   cuda_enter(c->ctx);
 #ifdef CHOOSE_ALGO
-  /* Static variables are only initialized once so this will not
-   * reset the previous algo every time */
-  static int reuse_algo = 0;
-  static cudnnConvolutionFwdAlgo_t prev_algo = CONV_ALGO;
-
 #ifndef CHOOSE_ONCE
-  static size_t prev_img_dims[5] = {0};
-  static size_t prev_kern_dims[5] = {0};
-
   reuse_algo = 1;
   for (unsigned int i = 0; i < PyGpuArray_NDIM(input); i++) {
     reuse_algo = (reuse_algo &&

theano/sandbox/gpuarray/dnn_gi.c

+#section init_code_struct
+
+#ifdef CHOOSE_ALGO
+reuse_algo = 0;
+prev_algo = CONV_ALGO;
+#ifndef CHOOSE_ONCE
+memset(prev_kern_dims, 0, sizeof(prev_kern_dims));
+memset(prev_top_dims, 0, sizeof(prev_top_dims));
+#endif
+#endif
+
 #section support_code_struct
 
+#ifdef CHOOSE_ALGO
+int reuse_algo = 0;
+cudnnConvolutionBwdDataAlgo_t prev_algo = CONV_ALGO;
+#ifndef CHOOSE_ONCE
+size_t prev_kern_dims[5] = {0};
+size_t prev_top_dims[5] = {0};
+#endif
+#endif
+
 int
 APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
                         PyGpuArrayObject *im,
@@ -57,13 +77,7 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
   cuda_enter(c->ctx);
 
 #ifdef CHOOSE_ALGO
-  static int reuse_algo = 0;
-  static cudnnConvolutionBwdDataAlgo_t prev_algo = CONV_ALGO;
-
 #ifndef CHOOSE_ONCE
-  static size_t prev_kern_dims[5] = {0};
-  static size_t prev_top_dims[5] = {0};
-
   reuse_algo = 1;
   for (unsigned int i = 0; i < PyGpuArray_NDIM(kerns); i++) {
     reuse_algo = (reuse_algo &&

theano/sandbox/gpuarray/dnn_gw.c

+#section init_code_struct
+
+#ifdef CHOOSE_ALGO
+reuse_algo = 0;
+prev_algo = CONV_ALGO;
+#ifndef CHOOSE_ONCE
+memset(prev_img_dims, 0, sizeof(prev_img_dims));
+memset(prev_top_dims, 0, sizeof(prev_top_dims));
+#endif
+#endif
+
 #section support_code_struct
 
+#ifdef CHOOSE_ALGO
+int reuse_algo;
+cudnnConvolutionBwdFilterAlgo_t prev_algo;
+#ifndef CHOOSE_ONCE
+size_t prev_img_dims[5];
+size_t prev_top_dims[5];
+#endif
+#endif
+
 int
 APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
                         PyGpuArrayObject *km,
@@ -57,13 +77,7 @@ APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
   cuda_enter(c->ctx);
 
 #ifdef CHOOSE_ALGO
-  static int reuse_algo = 0;
-  static cudnnConvolutionBwdFilterAlgo_t prev_algo = CONV_ALGO;
-
 #ifndef CHOOSE_ONCE
-  static size_t prev_img_dims[5] = {0};
-  static size_t prev_top_dims[5] = {0};
-
   reuse_algo = 1;
   for (unsigned int i = 0; i < PyGpuArray_NDIM(input); i++) {
     reuse_algo = (reuse_algo &&