Merge pull request #5050 from nouiz/conv3d

theano/gpuarray/dnn.py

@@ -461,9 +461,9 @@ class GpuDnnConv(DnnBase):
             defs.append(('CONV_INPLACE', '1'))
 
         alg = 'CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM'
-        if self.algo == 'none':
+        if self.algo == 'none':  # 3d (at least in v4)
             alg = 'CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM'
-        elif self.algo == 'small':
+        elif self.algo == 'small':  # 3d (at least in v4)
             alg = 'CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM'
         elif self.algo == 'large':
             alg = 'CUDNN_CONVOLUTION_FWD_ALGO_GEMM'
@@ -471,7 +471,7 @@ class GpuDnnConv(DnnBase):
             alg = 'CUDNN_CONVOLUTION_FWD_ALGO_DIRECT'
         elif self.algo == 'fft':
             alg = 'CUDNN_CONVOLUTION_FWD_ALGO_FFT'
-        elif self.algo == 'fft_tiling':
+        elif self.algo == 'fft_tiling':  # 3d (not in v4, in v5)
             alg = 'CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING'
         elif self.algo == 'winograd':
             # need v5
@@ -505,10 +505,13 @@ class GpuDnnConv(DnnBase):
             raise TypeError("The number of dimensions of "
                             "img, kern and output must match")
 
-        if (img.type.ndim == 5 and
-                self.algo in ['small', 'large', 'fft', 'fft_tiling']):
+        if img.type.ndim == 5 and self.algo in ['large', 'fft']:
             raise ValueError("convolution algo %s can't be used for "
                              "3d convolutions", (self.algo,))
+        if (img.type.ndim == 5 and
+                self.algo in ['fft_tiling'] and
+                version() < 5000):
+            raise ValueError("3d convolution algo fft_tiling need cudnn v5")
 
         if (not isinstance(desc.type, CDataType) or
                 desc.type.ctype != 'cudnnConvolutionDescriptor_t'):
@@ -634,13 +637,13 @@ class GpuDnnConvGradW(DnnBase):
             defs.append(('CONV_INPLACE', '1'))
 
         alg = 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0'
-        if self.algo == 'none':
+        if self.algo == 'none':  # 3d in at least v4
             alg = 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0'
         if self.algo == 'deterministic':
             alg = 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1'
         if self.algo == 'fft':
             alg = 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_FFT'
-        if self.algo == 'small':
+        if self.algo == 'small':  # 3d in at least v4
             # non-deterministic, small workspace
             alg = 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_3'
         if self.algo in ['guess_once', 'guess_on_shape_change',
@@ -673,7 +676,7 @@ class GpuDnnConvGradW(DnnBase):
                             "img, topgrad and output must match")
 
         if (img.type.ndim == 5 and
-                self.algo in ['fft', 'deterministic', 'small']):
+                self.algo in ['fft', 'deterministic']):
             raise ValueError("convolution algo %s can't be used for "
                              "3d convolutions", (self.algo,))
 
@@ -766,13 +769,13 @@ class GpuDnnConvGradI(DnnBase):
             defs.append(('CONV_INPLACE', '1'))
 
         alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_0'
-        if self.algo == 'none':
+        if self.algo == 'none':  # 3d at least v4
             alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_0'
-        elif self.algo == 'deterministic':
+        elif self.algo == 'deterministic':  # 3d at least v4
             alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_1'
         elif self.algo == 'fft':
             alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT'
-        elif self.algo == 'fft_tiling':
+        elif self.algo == 'fft_tiling':  # 3d not v4, since v5
             # big workspace but less than fft
             alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT_TILING'
         elif self.algo == 'winograd':
@@ -808,10 +811,13 @@ class GpuDnnConvGradI(DnnBase):
             raise TypeError("The number of dimensions of "
                             "kern, topgrad and output must match")
 
-        if (kern.type.ndim == 5 and
-                self.algo in ['fft', 'deterministic', 'fft_tiling']):
+        if kern.type.ndim == 5 and self.algo in ['fft']:
             raise ValueError("convolution algo %s can't be used for "
                              "3d convolutions", (self.algo,))
+        if (kern.type.ndim == 5 and
+                self.algo == 'fft_tiling' and
+                version() < 5000):
+            raise ValueError("3d convolution algo fft_tiling need cudnn v5")
 
         if (not isinstance(desc.type, CDataType) or
                 desc.type.ctype != 'cudnnConvolutionDescriptor_t'):