Corrected bug in GpuJoin and reactivated the optimization.

theano/sandbox/cuda/basic_ops.py

@@ -1480,7 +1480,8 @@ class GpuJoin(tensor.Join):
         # except for 'axis'
 
         def construct_slices(curlen):
-            slices = [slice(None,None,None) for i in range(len(cndas))]
+            slices = [slice(None,None,None) for i in \
+                            range(len(template_shape))]
             slices[axis] = slice(curpos,curpos+curlen,None)
             return tuple(slices)
 

theano/sandbox/cuda/opt.py

@@ -499,7 +499,7 @@ def local_gpu_join(node):
     not implemented yet.
 
     """
-    if False and isinstance(node.op, tensor.Join):
+    if isinstance(node.op, tensor.Join):
         # optimizing this case:
         # join(host_from_gpu) -> host_from_gpu(gpu_join)
 

theano/sandbox/cuda/tests/test_opt.py

@@ -72,9 +72,7 @@ def test_softmax_with_bias():
     bv=numpy.random.rand(8)
     assert numpy.allclose(f(xv,bv),f2(xv,bv))
 
-
-
-def test_opt_gpujoin_joinvectors_elemwise_than_minusone():
+def test_opt_gpujoin_onlyajoin():
     # from a bug in normal sampling
     _a = numpy.asarray([[1,2],[3,4]],dtype='float32')
     _b = numpy.asarray([[5,6,7],[8,9,10]],dtype='float32')
@@ -82,11 +80,13 @@ def test_opt_gpujoin_joinvectors_elemwise_than_minusone():
     b = theano.shared(_b)
 
     c = tensor.join(1,a,b)
-
+    
     f = theano.function([], c)
 
     #theano.printing.debugprint(f)
 
+    f()
+
     graph_nodes = f.maker.env.toposort()
 
     assert isinstance(graph_nodes[-1].op, cuda.HostFromGpu)
@@ -96,6 +96,38 @@ def test_opt_gpujoin_joinvectors_elemwise_than_minusone():
 
 
 
+def test_opt_gpujoin_joinvectors_elemwise_then_minusone():
+    # from a bug in gpu normal sampling
+    _a = numpy.asarray([1,2,3,4],dtype='float32')
+    _b = numpy.asarray([5,6,7,8],dtype='float32')
+    a = theano.shared(_a)
+    b = theano.shared(_b)
+
+    a_prime = tensor.cos(a)
+    b_prime = tensor.sin(b)
+
+    c = tensor.join(0,a_prime,b_prime)
+    
+    d = c[:-1]
+
+    f = theano.function([], d)
+
+    #theano.printing.debugprint(f)
+
+    graph_nodes = f.maker.env.toposort()
+
+    assert isinstance(graph_nodes[-1].op, cuda.HostFromGpu)
+    assert isinstance(graph_nodes[-2].op, cuda.GpuSubtensor)
+    assert isinstance(graph_nodes[-3].op, cuda.GpuJoin)
+
+    concat = numpy.concatenate([numpy.cos(_a),numpy.sin(_b)],axis=1)
+    concat = concat[:-1]
+
+    assert numpy.allclose(numpy.asarray(f()), concat)
+
+
+
 if __name__ == '__main__':
     test_opt_gpujoin_onlyajoin()
-    test_opt_gpujoin_joinvectors_elemwise_than_minusone()
+    test_opt_gpujoin_joinvectors_elemwise_then_minusone()
+

theano/sandbox/rng_mrg.py

@@ -808,7 +808,6 @@ def test_normal0():
         print prefix, 'samples/sec', steps*sample_size[0]*sample_size[1] / dt
 
     sample_size = (999,100)
-    '''
     print ''
     print 'ON CPU:'
 
@@ -820,11 +819,9 @@ def test_normal0():
     basictest(f, 50, -5.0, 2.0, prefix='mrg ')
 
     sys.stdout.flush()
-    '''
 
     # now with odd number of samples
-    sample_size = (10,10)
-    #sample_size = (999,99)
+    sample_size = (999,99)
 
 
     print ''
@@ -844,7 +841,6 @@ def test_normal0():
     basictest(f, 50, -5.0, 2.0, prefix='gpu mrg ')
 
 
-    '''
     print ''
     print 'ON CPU w NUMPY:'
     RR = theano.tensor.shared_randomstreams.RandomStreams(234)
@@ -853,7 +849,6 @@ def test_normal0():
     ff = theano.function([], nn)
 
     basictest(ff, 50, -5.0, 2.0, prefix='numpy ')
-    '''
 
 
 if __name__ == '__main__':