Merge pull request #2479 from carriepl/prevent_output_from_inplace

theano/compile/mode.py

@@ -159,6 +159,18 @@ class AddDestroyHandler(gof.Optimizer):
         fgraph.attach_feature(gof.DestroyHandler())
 
 
+class AddNoOutputFromInplace(gof.Optimizer):
+    """This optimizer adds to the fgraph a feature that will prevent outputs
+    of a fgraph to be created by performing inplace operations on intermediary
+    variables. This is useful when the outputs of the fgraph are preallocated
+    to prevent useless copying of the data. Currently, scan preallocates its
+    outputs
+    """
+    def add_requirements(self, fgraph):
+        super(AddNoOutputFromInplace, self).add_requirements(fgraph)
+        fgraph.attach_feature(gof.NoOutputFromInplace())
+
+
 class PrintCurrentFunctionGraph(gof.Optimizer):
     """This optimizer is for debugging.
 
@@ -211,6 +223,9 @@ optdb.register('specialize_device', gof.EquilibriumDB(),
 optdb.register('merge2', gof.MergeOptimizer(),
         49, 'fast_run', 'merge')
 
+optdb.register('add_no_output_from_inplace', AddNoOutputFromInplace(),
+        49.4)
+
 optdb.register('add_destroy_handler', AddDestroyHandler(),
         49.5, 'fast_run', 'inplace')
 

theano/compile/tests/test_mode.py

+import theano
+from theano.compile.mode import Mode
+import theano.tensor as T
+
+def test_no_output_from_implace():
+
+    x = T.matrix()
+    y = T.matrix()
+    a = T.dot(x, y)
+    b = T.tanh(a)
+
+    # Ensure that the elemwise op that produces the output is inplace when
+    # using a mode that does not include the optimization
+    fct_no_opt = theano.function([x,y], b, mode="FAST_RUN")
+    op = fct_no_opt.maker.fgraph.outputs[0].owner.op
+    assert (hasattr(op, 'destroy_map') and 0 in op.destroy_map)
+
+    # Ensure that the elemwise op that produces the output is not inplace when
+    # using a mode that includes the optimization
+    mode_opt = Mode(linker="cvm", optimizer="fast_run")
+    mode_opt = mode_opt.including("add_no_output_from_inplace")
+
+    fct_opt = theano.function([x,y], b, mode=mode_opt)
+    op = fct_opt.maker.fgraph.outputs[0].owner.op
+    assert (not hasattr(op, 'destroy_map') or 0 not in op.destroy_map)

theano/gof/__init__.py

@@ -74,7 +74,7 @@ from theano.gof.optdb import \
 from theano.gof.toolbox import \
     Feature, \
     Bookkeeper, History, Validator, ReplaceValidate, NodeFinder,\
-    PrintListener, ReplacementDidntRemovedError
+    PrintListener, ReplacementDidntRemovedError, NoOutputFromInplace
 
 from theano.gof.type import \
     Type, Generic, generic

theano/gof/toolbox.py

 import sys
 import time
 
+import theano
 from theano import config
 from theano.gof.python25 import partial
 from theano.gof.python25 import OrderedDict
@@ -394,3 +395,26 @@ class PreserveNames(Feature):
             new_r.name = r.name
 
 
+class NoOutputFromInplace(Feature):
+
+    def validate(self, fgraph):
+        if not hasattr(fgraph, 'destroyers'):
+            return True
+        for out in list(fgraph.outputs):
+
+            if out.owner is None:
+                continue
+
+            # Validate that the node that produces the output does not produce
+            # it by modifying something else inplace.
+            node = out.owner
+            op = node.op
+            out_idx = node.outputs.index(out)
+            if hasattr(op, 'destroy_map') and out_idx in op.destroy_map.keys():
+                raise theano.gof.InconsistencyError(
+                    "A function graph Feature has requested (probably for ",
+                    "efficiency reasons for scan) that outputs of the graph",
+                    "be prevented from being the result of inplace ",
+                    "operations. This has prevented output ", out, " from ",
+                    "being computed by modifying another variable ",
+                    "inplace.")

theano/sandbox/test_rng_mrg.py

@@ -90,7 +90,7 @@ def test_consistency_randomstreams():
     for use_cuda in test_use_cuda:
         #print 'use_cuda =', use_cuda
         samples = []
-        rng = MRG_RandomStreams(seed=seed, use_cuda=False)
+        rng = MRG_RandomStreams(seed=seed, use_cuda=use_cuda)
         for i in range(n_streams):
             stream_samples = []
             u = rng.uniform(size=(n_substreams,), nstreams=n_substreams)

theano/scan_module/numpy_api_changes.diff

-diff --git a/theano/scan_module/scan_perform.c b/theano/scan_module/scan_perform.c
-index aaebb43..2d06b29 100644
---- a/theano/scan_module/scan_perform.c
-+++ b/theano/scan_module/scan_perform.c
-@@ -5595,7 +5595,7 @@ static int __pyx_pf_5numpy_7ndarray___getbuffer__(PyArrayObject *__pyx_v_self, P
+@@ -5597,7 +5597,7 @@
   *             cdef list stack
   *             cdef int offset
   */
@@ -11,29 +7,29 @@ index aaebb43..2d06b29 100644
    __Pyx_INCREF(__pyx_t_4);
    __pyx_v_descr = ((PyArray_Descr *)__pyx_t_4);
    __pyx_t_4 = 0;
-@@ -7147,7 +7147,7 @@ static CYTHON_INLINE void __pyx_f_5numpy_set_array_base(PyArrayObject *__pyx_v_a
+@@ -7126,7 +7126,7 @@
   *      arr.base = baseptr
   * 
   */
 -  Py_XDECREF(__pyx_v_arr->base);
 +  Py_XDECREF(PyArray_BASE(__pyx_v_arr));
  
-   /* "/home/anakha/.local/lib/python2.7/site-packages/Cython/Includes/numpy/__init__.pxd":974
+   /* "numpy.pxd":973
   *          baseptr = <PyObject*>base
-@@ -7156,7 +7156,11 @@ static CYTHON_INLINE void __pyx_f_5numpy_set_array_base(PyArrayObject *__pyx_v_a
+@@ -7135,7 +7135,11 @@
   * 
   * cdef inline object get_array_base(ndarray arr):
   */
 -  __pyx_v_arr->base = __pyx_v_baseptr;
-+#if NPY_API_VERSION < 0x00000007
-+  PyArray_BASE(__pyx_v_arr) = __pyx_v_baseptr;
-+#else
-+  PyArray_SetBaseObject(__pyx_v_arr, __pyx_v_baseptr);
-+#endif
++  #if NPY_API_VERSION < 0x00000007
++      PyArray_BASE(__pyx_v_arr) = __pyx_v_baseptr;
++  #else
++      PyArray_SetBaseObject(__pyx_v_arr, __pyx_v_baseptr);
++  #endif
  
-   /* "/home/anakha/.local/lib/python2.7/site-packages/Cython/Includes/numpy/__init__.pxd":966
-  * 
-@@ -7191,7 +7195,7 @@ static CYTHON_INLINE PyObject *__pyx_f_5numpy_get_array_base(PyArrayObject *__py
+   __Pyx_RefNannyFinishContext();
+ }
+@@ -7161,7 +7165,7 @@
   *         return None
   *     else:
   */
@@ -41,8 +37,8 @@ index aaebb43..2d06b29 100644
 +  __pyx_t_1 = ((PyArray_BASE(__pyx_v_arr) == NULL) != 0);
    if (__pyx_t_1) {
  
-     /* "/home/anakha/.local/lib/python2.7/site-packages/Cython/Includes/numpy/__init__.pxd":978
-@@ -7214,8 +7218,8 @@ static CYTHON_INLINE PyObject *__pyx_f_5numpy_get_array_base(PyArrayObject *__py
+     /* "numpy.pxd":977
+@@ -7185,8 +7189,8 @@
   *         return <object>arr.base             # <<<<<<<<<<<<<<
   */
      __Pyx_XDECREF(__pyx_r);
@@ -52,4 +48,4 @@ index aaebb43..2d06b29 100644
 +    __pyx_r = ((PyObject *)PyArray_BASE(__pyx_v_arr));
      goto __pyx_L0;
    }
- 
+   __pyx_L3:;

theano/scan_module/scan_op.py

@@ -535,7 +535,7 @@ class Scan(PureOp):
                   self.n_sit_sot +
                   self.n_nit_sot)
         wrapped_inputs = [Param(x, borrow=True) for x in self.inputs]
-        wrapped_outputs = [Out(x, borrow=False) for x in
+        wrapped_outputs = [Out(x, borrow=(x not in self.inputs)) for x in
                            self.outputs[:slices]]
         wrapped_outputs += self.outputs[slices:]
         profile = None
@@ -927,11 +927,14 @@ class Scan(PureOp):
                     offset += 1
 
             # 4. collecting slices where the output should be stored
+
+            # 4.1. Collect slices for mitmots
             for idx in xrange(self.n_mit_mot_outs):
                 output_storage[idx].storage[0] = None
 
+            # 4.2. Collect slices for mitsots, sitsots and nitsots
             offset = self.n_mit_mot_outs
-            if i != 0 and self.n_nit_sot > 0:
+            if i != 0:
                 for idx in xrange(self.n_outs + self.n_nit_sot -
                                   self.n_mit_mot):
                     if (store_steps[idx + self.n_mit_mot] == 1 or
@@ -946,15 +949,24 @@ class Scan(PureOp):
                                   self.n_mit_mot):
                     output_storage[idx + offset].storage[0] = None
 
+            # 4.3. Collect slices for shared outputs
             offset += self.n_outs + self.n_nit_sot - self.n_mit_mot
             for idx in xrange(self.n_shared_outs):
                 output_storage[idx + offset].storage[0] = None
-            # If condition add it to the mix
+
+            # 4.4. If there is a condition add it to the mix
             if self.as_while:
                 pdx = offset + self.n_shared_outs
                 output_storage[pdx].storage[0] = None
+
+            # 4.5. Keep a reference to the variables currently in the
+            # output_storage to be able to compare them with the actual
+            # outputs of the inner function after its execution
+            old_output_storage = [o.storage[0] for o in output_storage]
+
             # 5. compute outputs
             t0_fn = time.time()
+
             try:
                 fn()
             except Exception:
@@ -974,11 +986,18 @@ class Scan(PureOp):
                 else:
                     # old-style linkers raise their own exceptions
                     raise
+
             dt_fn = time.time() - t0_fn
             if self.as_while:
                 pdx = offset + self.n_shared_outs
                 cond = output_storage[pdx].storage[0] == 0
 
+            # Check which of the pre-allocated outputs (if applicable) have
+            # been reused by the inner function
+            output_reused = [old_output_storage[o] is
+                             output_storage[o].storage[0]
+                             for o in range(len(output_storage))]
+
             t_fn += dt_fn
             offset_out = 0
             # 5.1 Copy over the values for mit_mot outputs
@@ -995,8 +1014,7 @@ class Scan(PureOp):
 
             for j in xrange(begin, end):
                 if (store_steps[j] == 1 or self.vector_outs[j] or
-                    outs[j][0][pos[j]] is not
-                      output_storage[offset_out + j].storage[0]):
+                    not output_reused[offset_out + j]):
                     outs[j][0][pos[j]] = \
                             output_storage[offset_out + j].storage[0]
 
@@ -1020,8 +1038,7 @@ class Scan(PureOp):
                         outs[j][0] = outs[j][0][:store_steps[j]]
                     outs[j][0][pos[j]] = output_storage[jout].storage[0]
                 elif (store_steps[j] == 1 or self.vector_outs[j] or
-                      outs[j][0][pos[j]] is not
-                      output_storage[j + offset_out].storage[0]):
+                      not output_reused[offset_out + j]):
                     outs[j][0][pos[j]] = \
                             output_storage[j + offset_out].storage[0]
 

theano/scan_module/scan_perform.pyx

@@ -62,7 +62,7 @@ import copy
 
 
 def get_version():
-    return 0.284
+    return 0.285
 
 @cython.boundscheck(False)
 def perform(
@@ -191,6 +191,10 @@ def perform(
     cdef unsigned int begin
     cdef unsigned int end
     cdef int cond
+    cdef unsigned int len_output_storage = (n_mit_mot_outs + n_mit_sot +
+                                            n_sit_sot + n_nit_sot +
+                                            n_shared_outs)
+    cdef int output_reused[500] # max 500 outputs
 
 
     if n_steps < 0:
@@ -304,11 +308,14 @@ def perform(
                 offset += 1
 
         # 4. collecting slices where the output should be stored
+
+        # 4.1. Collect slices for mitmots
         for idx in range(n_mit_mot_outs):
             output_storage[idx].storage[0] = None
 
+        # 4.2. Collect slices for mitsots, sitsots and nitsots
         offset = n_mit_mot_outs
-        if i !=0 and n_nit_sot >0:
+        if i != 0:
             for idx in range(n_outs + n_nit_sot - n_mit_mot):
                 if ( store_steps[<unsigned int>(idx+n_mit_mot)] == 1 or
                     vector_outs[<unsigned int>(idx+n_mit_mot)] == 1):
@@ -321,12 +328,21 @@ def perform(
             for idx in range(n_outs + n_nit_sot - n_mit_mot):
                 output_storage[<unsigned int>(idx+offset)].storage[0] = None
 
+        # 4.3. Collect slices for shared outputs
         offset += n_outs+n_nit_sot - n_mit_mot
         for idx in range(n_shared_outs):
             output_storage[<unsigned int>(idx+offset)].storage[0] = None
+
+        # 4.4. If there is a condition add it to the mix
         if as_while:
             pdx = offset + n_shared_outs
             output_storage[<unsigned int>pdx].storage[0] = None
+
+        # 4.5. Keep a reference to the variables currently in the
+        # output_storage to be able to compare them with the actual
+        # outputs of the inner function after its execution
+        old_output_storage = [o.storage[0] for o in output_storage]
+
         # 5. compute outputs
         t0_fn = time.time()
 
@@ -348,6 +364,11 @@ def perform(
             pdx = offset + n_shared_outs
             cond = output_storage[pdx].storage[0] == 0
 
+        # Check which of the pre-allocated outputs (if applicable) have
+        # been reused by the inner function
+        for j in range(len_output_storage):
+            output_reused[j] = (old_output_storage[j] is
+                                output_storage[j].storage[0])
 
         offset_out = 0
         # 5.1 Copy over the values for mit_mot outputs
@@ -363,8 +384,8 @@ def perform(
         offset_out -= n_mit_mot
 
         for j in range(begin, end):
-            if ( store_steps[j] == 1 or vector_outs[j] ==1 or
-                outs[j][0][pos[j]] is not output_storage[<unsigned int>(offset_out+j)].storage[0]):
+            if (store_steps[j] == 1 or vector_outs[j] == 1 or
+                not output_reused[<unsigned int>(offset_out+j)]):
 
                 outs[j][0][pos[j]] = output_storage[<unsigned int>(offset_out+j)].storage[0]
 
@@ -387,7 +408,7 @@ def perform(
                     outs[j][0] = outs[j][0][:store_steps[j]]
                 outs[j][0][pos[j]] = output_storage[jout].storage[0]
             elif (store_steps[j] == 1 or vector_outs[j] == 1 or
-                  outs[j][0][pos[j]] is not output_storage[j+offset_out].storage[0]):
+                  not output_reused[<unsigned int>(offset_out+j)]):
                 outs[j][0][pos[j]] = output_storage[j+offset_out].storage[0]
 
 

theano/scan_module/scan_perform_ext.py

@@ -16,7 +16,7 @@ from theano.gof import cmodule
 _logger = logging.getLogger('theano.scan_module.scan_perform')
 
 
-version = 0.284  # must match constant returned in function get_version()
+version = 0.285  # must match constant returned in function get_version()
 
 need_reload = False