Update the version of CNMeM

theano/sandbox/cuda/cnmem.h

@@ -57,7 +57,6 @@ typedef enum
   CNMEM_STATUS_SUCCESS = 0,
   CNMEM_STATUS_CUDA_ERROR,
   CNMEM_STATUS_INVALID_ARGUMENT,
-  CNMEM_STATUS_MEMORY_LEAK,
   CNMEM_STATUS_NOT_INITIALIZED,
   CNMEM_STATUS_OUT_OF_MEMORY,
   CNMEM_STATUS_UNKNOWN_ERROR
@@ -109,29 +108,58 @@ typedef struct cnmemDevice_t_
 cnmemStatus_t CNMEM_API cnmemInit(int numDevices, const cnmemDevice_t *devices, unsigned flags);
 
 /**
- * \brief Add a new stream to the pool of managed streams on a device.
+ * \brief Release all the allocated memory. 
+ * 
+ * This function must be called by a single thread and after all threads that called 
+ * cnmemMalloc/cnmemFree have joined. This function is not thread-safe.
  *
- * This function registers a new stream into a device memory manager. It is thread-safe.
+ * \return 
+ * CNMEM_STATUS_SUCCESS,          if everything goes fine,
+ * CNMEM_STATUS_NOT_INITIALIZED,  if the ::cnmemInit function has not been called,
+ * CNMEM_STATUS_CUDA_ERROR,       if an error happens in one of the CUDA functions.
+ */
+cnmemStatus_t CNMEM_API cnmemFinalize();
+
+/**
+ * \brief Increase the internal reference counter of the context object.
+ * 
+ * This function increases the internal reference counter of the library. The purpose of that
+ * reference counting mechanism is to give more control to the user over the lifetime of the 
+ * library. It is useful with scoped memory allocation which may be destroyed in a final 
+ * memory collection after the end of main(). That function is thread-safe.
  *
  * \return 
  * CNMEM_STATUS_SUCCESS,          if everything goes fine,
- * CNMEM_STATUS_INVALID_ARGUMENT, if one of the argument is invalid,
+ * CNMEM_STATUS_NOT_INITIALIZED,  if the ::cnmemInit function has not been called,
  */
-cnmemStatus_t CNMEM_API cnmemRegisterStream(cudaStream_t stream);
+cnmemStatus_t CNMEM_API cnmemRetain();
 
 /**
- * \brief Release all the allocated memory. 
+ * \brief Decrease the internal reference counter of the context object.
  * 
- * This function must be called by a single thread and after all threads that called 
- * cnmemMalloc/cnmemFree have joined. This function is not thread-safe.
+ * This function decreases the internal reference counter of the library. The purpose of that
+ * reference counting mechanism is to give more control to the user over the lifetime of the 
+ * library. It is useful with scoped memory allocation which may be destroyed in a final 
+ * memory collection after the end of main(). That function is thread-safe.
+ *
+ * You can use \c cnmemRelease to explicitly finalize the library.
  *
  * \return 
  * CNMEM_STATUS_SUCCESS,          if everything goes fine,
  * CNMEM_STATUS_NOT_INITIALIZED,  if the ::cnmemInit function has not been called,
- * CNMEM_STATUS_MEMORY_LEAK,      if there are unreleased blocks in the memory queues,
- * CNMEM_STATUS_CUDA_ERROR,       if an error happens in one of the CUDA functions.
  */
-cnmemStatus_t CNMEM_API cnmemFinalize();
+cnmemStatus_t CNMEM_API cnmemRelease();
+
+/**
+ * \brief Add a new stream to the pool of managed streams on a device.
+ *
+ * This function registers a new stream into a device memory manager. It is thread-safe.
+ *
+ * \return 
+ * CNMEM_STATUS_SUCCESS,          if everything goes fine,
+ * CNMEM_STATUS_INVALID_ARGUMENT, if one of the argument is invalid,
+ */
+cnmemStatus_t CNMEM_API cnmemRegisterStream(cudaStream_t stream);
 
 /**
  * \brief Allocate memory. 

theano/sandbox/cuda/cuda_ndarray.cu

@@ -138,9 +138,7 @@ void * device_malloc(size_t size, int verbose)
     ///@TODO: thejaswi: support for multiple-streams?
     if(g_use_cnmem) {
         cnmemStatus_t status = CNMEM_STATUS_SUCCESS;
-        if( size != 0 ) {
-            status = cnmemMalloc(&rval, size, NULL);
-        }
+        status = cnmemMalloc(&rval, size, NULL);
         if(status != CNMEM_STATUS_SUCCESS) {
             PyErr_Format(PyExc_MemoryError,
                          "Error allocating %zd bytes of device memory (%s).",