Document cnmem

doc/acknowledgement.txt

@@ -24,4 +24,4 @@ Acknowledgements
     P. L'Ecuyer and R. Touzin, `Fast Combined Multiple Recursive Generators with Multipliers of the form a = +/- 2^d +/- 2^e <http://www.informs-sim.org/wsc00papers/090.PDF>`_, Proceedings of the 2000 Winter Simulation Conference, Dec. 2000, 683--689.
 
   We were authorized by Pierre L'Ecuyer to copy/modify his Java implementation in the `SSJ <http://www.iro.umontreal.ca/~simardr/ssj/>`_ software and to relicense it under BSD 3-Clauses in Theano.
-* A better CUDA memory allocator `CnMEM <https://github.com/NVIDIA/cnmem>`_ is included in Theano. It have the same license.
+* A better GPU memory allocator :attr:`CNMeM <config.lib.cnmem>` is included in Theano. It have the same license.

doc/faq.txt

@@ -72,13 +72,18 @@ and use directly the optimized graph from the pickled file.
 Faster Theano function
 ----------------------
 
-You can set the Theano flag ``allow_gc`` to ``False`` to get a speed-up by using
+You can set the Theano flag :attr:`allow_gc <config.allow_gc>` to ``False`` to get a speed-up by using
 more memory. By default, Theano frees intermediate results when we don't need
 them anymore. Doing so prevents us from reusing this memory. So disabling the
 garbage collection will keep all intermediate results' memory space to allow to
 reuse them during the next call to the same Theano function, if they are of the
 correct shape. The shape could change if the shapes of the inputs change.
 
+.. note::
+
+   With :attr:`CNMeM <config.lib.cnmem>`, this isn't very useful with GPU
+   anymore.
+
 .. _unsafe_optimization:
 
 Unsafe optimization

doc/index.txt

@@ -21,6 +21,9 @@ Montreal).
 News
 ====
 
+* We added support for :attr:`CNMeM <config.lib.cnmem>` to speed up
+  the GPU memory allocation.
+
 * Theano 0.7 was released 26th March 2015. Everybody is encouraged to update.
 
 * We support `cuDNN <http://deeplearning.net/software/theano/library/sandbox/cuda/dnn.html>`_ if it is installed by the user.

doc/library/config.txt

@@ -370,6 +370,33 @@ import theano and print the config variable, as in:
     `amdlibm <http://developer.amd.com/cpu/libraries/libm/>`__
     library, which is faster than the standard libm.
 
+.. attribute:: lib.cnmem
+
+    Float value: >= 0
+
+    Do we enable `CNMeM <https://github.com/NVIDIA/cnmem>`_ or not (a
+    faster CUDA memory allocator). In Theano dev version until 0.7.1
+    is released.
+
+    That library is included in Theano, you do not need to install it.
+
+    The value represent the start size (in MB or % of
+    total GPU memory) of the memory pool:
+
+        * 0: not enabled.
+        * 0 < N <= 1: % of the total GPU memory (clipped to .985 for driver memory)
+        * > 0: use that number of MB of memory.
+
+
+    Default 0 (but should change later)
+
+    .. note::
+
+        This could cause memory fragmentation. So if you have an
+        memory error while using cnmem, try to allocate more memory at
+        the start or without it. If you try this, report your result
+        on :ref`theano-dev`.
+
 .. attribute:: linker
 
     String value: 'c|py', 'py', 'c', 'c|py_nogc'