Fix elusive memory-corruption bug in Scan.

NEWS.txt

@@ -2,9 +2,13 @@
 
 Updates in the Trunk since the last release:
 
-Sparse Sandbox bugfix
- * Fix the grad of theano.sparse.sandbox.sp.row_scale. It didn't
-   returned the right number of element. (Frederic B.)
+Bug fixes
+ * Outputs of Scan nodes could contain corrupted values: some parts of the
+   output would be repeated a second time, instead of the correct values.
+   It happened randomly, and quite infrequently, but the bug has been present
+   (both in Python and Cython) since April 2011. (Pascal L.)
+ * In Sparse sandbox, fix the grad of theano.sparse.sandbox.sp.row_scale.
+   It did not return the right number of elements. (Frederic B.)
 
 Documentation
  * Added in the tutorial documentation on how to extend Theano.

theano/scan_module/scan_op.py

@@ -809,7 +809,6 @@ class Scan(PureOp):
         ############## THE MAIN LOOP #########################
         #for i in xrange(n_steps):
         while (i < n_steps) and cond:
-
             # sequences over which scan iterates
             # 3. collect input slices
             for idx in xrange(self.n_seqs):
@@ -955,12 +954,17 @@ class Scan(PureOp):
         begin = self.n_mit_mot
         end = self.n_outs + self.n_nit_sot
         for idx in xrange(begin, end):
-            min_tap = self.mintaps[idx]
             if (store_steps[idx] < i - self.mintaps[idx] and
                 pos[idx] < store_steps[idx]):
 
                 pdx = pos[idx]
-                if pdx < store_steps[idx] // 2:
+                if pdx >= store_steps[idx] // 2:
+                    # It seems inefficient to copy the bigger part of the
+                    # array over, and back, but it is the only way that
+                    # there is no overlap in the areas of out[idx][0] that
+                    # are read and written.
+                    # This way, there will be no information overwritten
+                    # before it is read (as it used to happen).
                     shape = (pdx,) + outs[idx][0].shape[1:]
                     if cuda.cuda_available and isinstance(outs[idx][0],
                                                           cuda.CudaNdarray):

theano/scan_module/scan_perform.pyx

@@ -63,7 +63,7 @@ from theano.sandbox import cuda
 
 
 def get_version():
-    return 0.265
+    return 0.266
 
 @cython.boundscheck(False)
 def perform(
@@ -418,7 +418,13 @@ def perform(
             pos[idx] < store_steps[idx] ):
 
             pdx = pos[idx]
-            if pdx < store_steps[idx]//2 :
+            if pdx >= store_steps[idx]//2 :
+                # It seems inefficient to copy the bigger part of the
+                # array over, and back, but it is the only way that
+                # there is no overlap in the areas of out[idx][0] that
+                # are read and written.
+                # This way, there will be no information overwritten
+                # before it is read (as it used to happen).
                 shape = (pdx,)+ outs[idx][0].shape[1:]
 
                 if cuda.cuda_available and isinstance( outs[idx][0],

theano/scan_module/scan_perform_ext.py

@@ -14,7 +14,7 @@ logging.basicConfig(level=logging.DEBUG)
 if config.compiledir not in sys.path:
     sys.path.append(config.compiledir)
 
-version = 0.265 # must match constant returned in function get_version()
+version = 0.266  # must match constant returned in function get_version()
 
 need_reload = False
 try: