Merge pull request #3966 from nouiz/dnn

doc/library/config.txt

@@ -567,6 +567,20 @@ import theano and print the config variable, as in:
 
     A directory with bin/, lib/, include/ folders containing cuda utilities.
 
+.. attribute:: dnn.enabled
+
+  String value: ``auto``, ``True``, ``False``
+
+  Default ``auto``
+
+  If ``auto``, auto detect if cudnn is there, if not, do not use it,
+  but do not raise an error.
+
+  If ``True``, if we can't use cudnn, raise an error.
+
+  If ``False``, do not test if cudnn is there and don't use it.
+  Useful to disable it completely when the auto detection crashes.
+
 .. attribute:: config.dnn.conv.workmem
 
    Deprecated, use dnn.conv.algo_fwd.

doc/library/sandbox/cuda/dnn.txt

@@ -38,7 +38,10 @@ By default, Theano will detect if it can use cuDNN. If so, it will use
 it.  If not, Theano optimizations will not introduce cuDNN ops. So
 Theano will still work if the user did not introduce them manually.
 
-To get an error if Theano can not use cuDNN, use this Theano flag:
+The recently added Theano flag :attr:`dnn.enabled
+<config.dnn.enabled>` allows to change the default behavior to force
+it or disable it. Older Theano version do not support this flag. To
+get an error when CuDNN can not be used with them, use this flag:
 ``optimizer_including=cudnn``.
 
 .. note::
@@ -63,6 +66,8 @@ To get an error if Theano can not use cuDNN, use this Theano flag:
    * ``large`` : use a sometimes faster implementation with large memory usage
    * ``fft`` : use the Fast Fourrier Transform implementation of convolution
      (very high memory usage)
+   * ``fft_tiling`` : use the Fast Fourrier Transform implementation of convolution
+     with tiling (high memory usage, but less then fft)
    * ``guess_once`` : the first time a convolution is executed, the
      implementation to use is chosen according to CuDNN's heuristics and reused
      for every subsequent execution of the convolution.
@@ -76,9 +81,10 @@ To get an error if Theano can not use cuDNN, use this Theano flag:
      implementation selected every time the shapes of the inputs and kernels
      don't match the shapes from the last execution.
 
-   The Theano flag ``dnn.conv.algo_bwd`` allows to specify the CuDNN
-   convolution implementation that Theano should use for gradient convolutions.
-   Possible values include :
+   The Theano flag ``dnn.conv.algo_bwd_filter`` and
+   ``dnn.conv.algo_bwd_data`` allows to specify the CuDNN
+   convolution implementation that Theano should use for gradient
+   convolutions.  Possible values include :
 
    * ``none`` (default) : use the default non-deterministic convolution
      implementation
@@ -98,6 +104,13 @@ To get an error if Theano can not use cuDNN, use this Theano flag:
      implementation selected every time the shapes of the inputs and kernels
      don't match the shapes from the last execution.
 
+   * (algo_bwd_data only) ``fft_tiling`` : use the Fast Fourrier
+     Transform implementation of convolution with tiling (high memory
+     usage, but less then fft)
+
+   * (algo_bwd_data only) ``small`` : use a convolution implementation
+     with small memory usage
+
    ``guess_*`` and ``time_*`` flag values take into account the amount of
    available memory when selecting an implementation. This means that slower
    implementations might be selected if not enough memory is available for the

theano/configdefaults.py

@@ -329,6 +329,14 @@ AddConfigVar('dnn.library_path',
              "Location of the cudnn header (defaults to the cuda root)",
              StrParam(default_dnn_path('lib64')))
 
+AddConfigVar('dnn.enabled',
+             "'auto', use CuDNN if available, but silently fall back"
+             " to not using it if not present."
+             " If True and CuDNN can not be used, raise an error."
+             " If False, disable cudnn",
+             StrParam("auto", "True", "False"),
+             in_c_key=False)
+
 # This flag determines whether or not to raise error/warning message if
 # there is a CPU Op in the computational graph.
 AddConfigVar(

theano/sandbox/cuda/dnn.py

@@ -36,11 +36,13 @@ from theano.tensor.nnet.abstract_conv import (AbstractConv2d,
 
 
 def dnn_available():
-    if dnn_available.avail is None:
-        if not theano.sandbox.cuda.cuda_available:
+    if config.dnn.enabled == "False":
+        dnn_available.avail = False
+        dnn_available.msg = "disabled by dnn.enabled flag"
+    if dnn_available.avail is None and not theano.sandbox.cuda.cuda_available:
         dnn_available.msg = "CUDA not available"
         dnn_available.avail = False
-            return False
+    elif dnn_available.avail is None:
         dev = theano.sandbox.cuda.active_device_number()
         if theano.sandbox.cuda.device_properties(dev)['major'] < 3:
             dnn_available.msg = "Device not supported by cuDNN"
@@ -100,7 +102,11 @@ if ((err = cudnnCreate(&_handle)) != CUDNN_STATUS_SUCCESS) {
                         "candidate) that isn't supported.  Please update to "
                         "at least v3 final version.")
                     raise RuntimeError(dnn_available.msg)
-
+    if config.dnn.enabled == "True":
+        if not dnn_available.avail:
+            raise RuntimeError(
+                "You enabled CuDNN, but we aren't able to use it: %s" %
+                dnn_available.msg)
     return dnn_available.avail
 
 

theano/sandbox/gpuarray/dnn.py

@@ -86,6 +86,9 @@ def _dnn_check_version():
 def dnn_present():
     if dnn_present.avail is not None:
         return dnn_present.avail
+    if config.dnn.enabled == "False":
+        dnn_present.msg = "disabled by dnn.enabled flag"
+        dnn_present.avail = False
 
     if pygpu is None:
         dnn_present.msg = "PyGPU not available"
@@ -98,6 +101,12 @@ def dnn_present():
         if not dnn_present.avail:
             raise RuntimeError(dnn_present.msg)
 
+    if config.dnn.enabled == "True":
+        if not dnn_present.avail:
+            raise RuntimeError(
+                "You enabled CuDNN, but we aren't able to use it: %s" %
+                dnn_present.msg)
+
     return dnn_present.avail
 
 dnn_present.avail = None