Speed up GpuDnnSoftmaxGrad by using the best mapping.

theano/gpuarray/dnn.py

@@ -1634,8 +1634,8 @@ def local_softmax_dnn_grad(node, ctx_name):
         n = as_gpuarray_variable(n, ctx_name)
         if n.ndim != 2:
             return
-        ins.append(n.dimshuffle(0, 1, 'x', 'x'))
+        ins.append(n.dimshuffle(0, 'x', 1, 'x'))
 
-    out = GpuDnnSoftmaxGrad('accurate', 'channel')(
+    out = GpuDnnSoftmaxGrad('accurate', 'instance')(
         gpu_contiguous(ins[0]), gpu_contiguous(ins[1]))
-    return [out.dimshuffle(0, 1)]
+    return [out.dimshuffle(0, 2)]

theano/gpuarray/tests/test_dnn.py

@@ -836,6 +836,8 @@ class test_SoftMax(test_nnet.test_SoftMax):
             mode=mode_with_gpu
         )
         sorted_f = f.maker.fgraph.toposort()
+        val = numpy.random.rand(5).astype('float32')
+        out_dnn = f(val)
         assert(len([i
                     for i in sorted_f
                     if isinstance(
@@ -860,6 +862,8 @@ class test_SoftMax(test_nnet.test_SoftMax):
             mode=mode_wo_cudnn
         )
         sorted_f = f.maker.fgraph.toposort()
+        out_cpu = f(val)
+        utt.assert_allclose(out_dnn, out_cpu)
         assert(len([i
                     for i in sorted_f
                     if isinstance(

theano/sandbox/cuda/dnn.py

@@ -2590,17 +2590,17 @@ if True:
                     n = n.owner.inputs[0]
                 if n.ndim != 2:
                     return
-                ins.append(n.dimshuffle(0, 1, 'x', 'x'))
+                ins.append(n.dimshuffle(0, 'x', 1, 'x'))
 
             out = GpuDnnSoftmaxGrad(
                 'bc01',
                 'accurate',
-                'channel'
+                'instance',
             )(
                 gpu_contiguous(ins[0]),
                 gpu_contiguous(ins[1])
             )
-            return [out.dimshuffle(0, 1)]
+            return [out.dimshuffle(0, 2)]
 
 
 # AbstractConv Optimizations

theano/sandbox/cuda/tests/test_dnn.py

@@ -585,6 +585,8 @@ class test_DnnSoftMax(test_nnet.test_SoftMax):
             mode=mode_with_gpu
         )
         sorted_f = f.maker.fgraph.toposort()
+        val = numpy.random.rand(5).astype('float32')
+        out_dnn = f(val)
         assert(len([i
                     for i in sorted_f
                     if isinstance(
@@ -608,6 +610,8 @@ class test_DnnSoftMax(test_nnet.test_SoftMax):
             mode=mode_wo_cudnn
         )
         sorted_f = f.maker.fgraph.toposort()
+        out_cpu = f(val)
+        utt.assert_allclose(out_dnn, out_cpu)
         assert(len([i
                     for i in sorted_f
                     if isinstance(