GpuDnnConv3dGradI algo deterministic for v4 and v5. fft_tiling for v5

theano/sandbox/cuda/dnn.py

@@ -880,6 +880,7 @@ class GpuDnnConvGradI(DnnBase, COp):
                 alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT'
             elif self.algo == 'fft_tiling':
                 # need v4, big workspace, but less then fft
+                # need v5, for conv3d.
                 alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT_TILING'
             elif self.algo in ['guess_once', 'guess_on_shape_change']:
                 # The convolution implementation should be chosen according
@@ -939,7 +940,8 @@ class GpuDnnConv3dGradI(GpuDnnConvGradI):
     :param descr: the convolution descriptor
     :param workmem:
         *deprecated*, use parameter algo instead.
-    :param algo: ['none', 'guess_once', 'guess_on_shape_change',
+    :param algo: ['none', 'deterministic, 'fft_tiling',
+                  'guess_once', 'guess_on_shape_change',
                   'time_once', 'time_on_shape_change']
 
         Default is the value of :attr:`config.dnn.conv.algo_bwd_data`.
@@ -955,11 +957,20 @@ class GpuDnnConv3dGradI(GpuDnnConvGradI):
                            "deprecated. Use 'algo' instead."), stacklevel=3)
             assert algo is None
             algo = workmem
-
-        super(GpuDnnConv3dGradI, self).__init__(inplace=inplace,
-                                                algo="none")
-        assert self.algo in ['none', 'guess_once', 'guess_on_shape_change',
+        good_algo = ['none', 'deterministic', 'fft_tiling',
+                     'guess_once', 'guess_on_shape_change',
                      'time_once', 'time_on_shape_change']
+        if algo is None and config.dnn.conv.algo_bwd_data not in good_algo:
+            algo = 'guess_once'
+        elif algo is not None and algo not in good_algo:
+            algo = 'guess_once'
+        super(GpuDnnConv3dGradI, self).__init__(inplace=inplace,
+                                                algo=algo)
+        assert self.algo in good_algo
+        if version() < (5000, 5000):
+            if self.algo == 'fft_tiling':
+                raise RuntimeError("CuDNN 3d tiled-FFT convolution requires "
+                                   "CuDNN v5 or more recent")
 
     def grad(self, inp, grads):
         kerns, top, output, desc, alpha, beta = inp