more stack trace copying

theano/gpuarray/opt.py

@@ -1323,6 +1323,7 @@ def local_gpua_gemm(op, context_name, inputs, outputs):
 def local_gpua_gemmbatch(op, context_name, inputs, outputs):
     if inputs[0].dtype not in ['float16', 'float32', 'float64']:
         return
+    with inherit_stack_trace(outputs):
         a, b = inputs
         # Since GpuGemmBatch only supports 3D inputs and output,
         # we need to add broadcastable dims to the inputs, and drop
@@ -2599,6 +2600,7 @@ def local_gpu_solve(op, context_name, inputs, outputs):
 @local_optimizer([GpuCusolverSolve], inplace=True)
 def local_inplace_gpu_solve(node):
     if isinstance(node.op, GpuCusolverSolve) and not node.op.inplace:
+        with inherit_stack_trace(node.outputs):
             return [GpuCusolverSolve(A_structure=node.op.A_structure, trans=node.op.trans,
                                      inplace=True)(*node.inputs)]
 
@@ -2638,6 +2640,7 @@ register_opt2([slinalg.Solve], 'fast_compile', name='matrix_ops_db2')(matrix_ops
 @local_optimizer([GpuCholesky], inplace=True)
 def local_inplace_gpu_cholesky(node):
     if isinstance(node.op, GpuCholesky) and not node.op.inplace:
+        with inherit_stack_trace(node.outputs):
             return [node.op.clone_inplace()(*node.inputs)]
 
 
@@ -2721,6 +2724,7 @@ def local_gpu_magma_matrix_inverse(op, context_name, inputs, outputs):
 @local_optimizer([GpuMagmaMatrixInverse])
 def local_inplace_gpu_magma_matrix_inverse(node):
     if isinstance(node.op, GpuMagmaMatrixInverse) and not node.op.inplace:
+        with inherit_stack_trace(node.outputs):
             return [node.op.clone_inplace()(*node.inputs)]
 
 

theano/gpuarray/tests/test_opt.py

@@ -41,6 +41,7 @@ def _check_stack_trace(thing):
                                    theano.ifelse.IfElse,
                                    GpuFromHost, HostFromGpu,
                                    GpuCAReduceCuda,
+                                   basic_ops.GpuContiguous,
                                    GpuElemwise,
                                    theano.printing.Print,
                                    PdbBreakpoint,