Merge pull request #309 from thouis/malloc_hooks

ENH: expose PyDataMem_NEW/FREE/RENEW as numpy API functions with an event hook.

12 files changed, 872 insertions, 14 deletions

diff --git a/numpy/core/code_generators/cversions.txt b/numpy/core/code_generators/cversions.txt
index 99ea072ca..5bdb33c94 100644
--- a/numpy/core/code_generators/cversions.txt
+++ b/numpy/core/code_generators/cversions.txt
@@ -10,4 +10,4 @@
 # PyArray_CountNonzero, PyArray_NewLikeArray and PyArray_MatrixProduct2.
 0x00000006 = e61d5dc51fa1c6459328266e215d6987
 # Version 7 (NumPy 1.7) improved datetime64, misc utilities.
-0x00000007 = 1768b6c404a3d5a2a6bfe7c68f89e3aa
+0x00000007 = e396ba3912dcf052eaee1b0b203a7724
diff --git a/numpy/core/code_generators/numpy_api.py b/numpy/core/code_generators/numpy_api.py
index fd2b9628e..c49c3c346 100644
--- a/numpy/core/code_generators/numpy_api.py
+++ b/numpy/core/code_generators/numpy_api.py
@@ -324,6 +324,10 @@ multiarray_funcs_api = {
     'PyArray_DebugPrint':                   285,
     'PyArray_FailUnlessWriteable':          286,
     'PyArray_SetUpdateIfCopyBase':          287,
+    'PyDataMem_NEW':                        288,
+    'PyDataMem_FREE':                       289,
+    'PyDataMem_RENEW':                      290,
+    'PyDataMem_SetEventHook':               291,
 }
 
 ufunc_types_api = {
diff --git a/numpy/core/include/numpy/ndarraytypes.h b/numpy/core/include/numpy/ndarraytypes.h
index 9605c20f8..13483396b 100644
--- a/numpy/core/include/numpy/ndarraytypes.h
+++ b/numpy/core/include/numpy/ndarraytypes.h
@@ -327,10 +327,7 @@ struct NpyAuxData_tag {
    * allocated.
    */
 
-  /* Data buffer */
-#define PyDataMem_NEW(size) ((char *)malloc(size))
-#define PyDataMem_FREE(ptr)  free(ptr)
-#define PyDataMem_RENEW(ptr,size) ((char *)realloc(ptr,size))
+  /* Data buffer - PyDataMem_NEW/FREE/RENEW are in multiarraymodule.c */
 
 #define NPY_USE_PYMEM 1
 
@@ -1714,6 +1711,13 @@ typedef struct {
                            */
 } PyArrayInterface;
 
+/*
+ * This is a function for hooking into the PyDataMem_NEW/FREE/RENEW functions.
+ * See the documentation for PyDataMem_SetEventHook.
+ */
+typedef void (PyDataMem_EventHookFunc)(void *inp, void *outp, size_t size,
+                                       void *user_data);
+
 #if !(defined(NPY_NO_DEPRECATED_API) && (NPY_API_VERSION <= NPY_NO_DEPRECATED_API))
 #include "npy_deprecated_api.h"
 #endif
diff --git a/numpy/core/src/multiarray/multiarray_tests.c.src b/numpy/core/src/multiarray/multiarray_tests.c.src
index 15d4c871e..a0f70aabc 100644
--- a/numpy/core/src/multiarray/multiarray_tests.c.src
+++ b/numpy/core/src/multiarray/multiarray_tests.c.src
@@ -375,6 +375,59 @@ clean_ax:
     return NULL;
 }
 
+/* PyDataMem_SetHook tests */
+static int malloc_free_counts[2];
+static PyDataMem_EventHookFunc *old_hook = NULL;
+static void *old_data;
+
+static void test_hook(void *old, void *new, size_t size, void *user_data)
+{
+    int* counters = (int *) user_data;
+    if (old == NULL) {
+        counters[0]++; /* malloc counter */
+    }
+    if (size == 0) {
+        counters[1]++; /* free counter */
+    }
+}
+
+static PyObject*
+test_pydatamem_seteventhook_start(PyObject* NPY_UNUSED(self), PyObject* NPY_UNUSED(args))
+{
+    malloc_free_counts[0] = malloc_free_counts[1] = 0;
+    old_hook = PyDataMem_SetEventHook(test_hook, (void *) malloc_free_counts, &old_data);
+    Py_INCREF(Py_None);
+    return Py_None;
+}
+
+static PyObject*
+test_pydatamem_seteventhook_end(PyObject* NPY_UNUSED(self), PyObject* NPY_UNUSED(args))
+{
+    PyDataMem_EventHookFunc *my_hook;
+    void *my_data;
+
+    my_hook = PyDataMem_SetEventHook(old_hook, old_data, &my_data);
+    if ((my_hook != test_hook) || (my_data != (void *) malloc_free_counts)) {
+        PyErr_SetString(PyExc_ValueError,
+                        "hook/data was not the expected test hook");
+        return NULL;
+    }
+
+    if (malloc_free_counts[0] == 0) {
+        PyErr_SetString(PyExc_ValueError,
+                        "malloc count is zero after test");
+        return NULL;
+    }
+    if (malloc_free_counts[1] == 0) {
+        PyErr_SetString(PyExc_ValueError,
+                        "free count is zero after test");
+        return NULL;
+    }
+
+    Py_INCREF(Py_None);
+    return Py_None;
+}
+
 static PyMethodDef Multiarray_TestsMethods[] = {
     {"test_neighborhood_iterator",
         test_neighborhood_iterator,
@@ -382,6 +435,12 @@ static PyMethodDef Multiarray_TestsMethods[] = {
     {"test_neighborhood_iterator_oob",
         test_neighborhood_iterator_oob,
         METH_VARARGS, NULL},
+    {"test_pydatamem_seteventhook_start",
+        test_pydatamem_seteventhook_start,
+        METH_NOARGS, NULL},
+    {"test_pydatamem_seteventhook_end",
+        test_pydatamem_seteventhook_end,
+        METH_NOARGS, NULL},
     {NULL, NULL, 0, NULL}        /* Sentinel */
 };
 
diff --git a/numpy/core/src/multiarray/multiarraymodule.c b/numpy/core/src/multiarray/multiarraymodule.c
index 1ab7823ad..0180b1f89 100644
--- a/numpy/core/src/multiarray/multiarraymodule.c
+++ b/numpy/core/src/multiarray/multiarraymodule.c
@@ -3430,6 +3430,102 @@ test_interrupt(PyObject *NPY_UNUSED(self), PyObject *args)
     return PyInt_FromLong(a);
 }
 
+/* malloc/free/realloc hook */
+NPY_NO_EXPORT PyDataMem_EventHookFunc *_PyDataMem_eventhook;
+NPY_NO_EXPORT void *_PyDataMem_eventhook_user_data;
+
+/*NUMPY_API
+ * Sets the allocation event hook for numpy array data.
+ * Takes a PyDataMem_EventHookFunc *, which has the signature:
+ *        void hook(void *old, void *new, size_t size, void *user_data).
+ *   Also takes a void *user_data, and void **old_data.
+ *
+ * Returns a pointer to the previous hook or NULL.  If old_data is
+ * non-NULL, the previous user_data pointer will be copied to it.
+ *
+ * If not NULL, hook will be called at the end of each PyDataMem_NEW/FREE/RENEW:
+ *   result = PyDataMem_NEW(size)        -> (*hook)(NULL, result, size, user_data)
+ *   PyDataMem_FREE(ptr)                 -> (*hook)(ptr, NULL, 0, user_data)
+ *   result = PyDataMem_RENEW(ptr, size) -> (*hook)(ptr, result, size, user_data)
+ *
+ * When the hook is called, the GIL will be held by the calling
+ * thread.  The hook should be written to be reentrant, if it performs
+ * operations that might cause new allocation events (such as the
+ * creation/descruction numpy objects, or creating/destroying Python
+ * objects which might cause a gc)
+ */
+NPY_NO_EXPORT PyDataMem_EventHookFunc *
+PyDataMem_SetEventHook(PyDataMem_EventHookFunc *newhook,
+                       void *user_data, void **old_data)
+{
+    PyGILState_STATE gilstate = PyGILState_Ensure();
+    PyDataMem_EventHookFunc *temp = _PyDataMem_eventhook;
+    _PyDataMem_eventhook = newhook;
+    if (old_data != NULL) {
+        *old_data = _PyDataMem_eventhook_user_data;
+    }
+    _PyDataMem_eventhook_user_data = user_data;
+    PyGILState_Release(gilstate);
+    return temp;
+}
+
+/*NUMPY_API
+ * Allocates memory for array data.
+ */
+NPY_NO_EXPORT void *
+PyDataMem_NEW(size_t size)
+{
+    void *result;
+
+    result = malloc(size);
+    if (_PyDataMem_eventhook != NULL) {
+        PyGILState_STATE gilstate = PyGILState_Ensure();
+        if (_PyDataMem_eventhook != NULL) {
+            (*_PyDataMem_eventhook)(NULL, result, size,
+                                    _PyDataMem_eventhook_user_data);
+        }
+        PyGILState_Release(gilstate);
+    }
+    return (char *)result;
+}
+
+/*NUMPY_API
+ * Free memory for array data.
+ */
+NPY_NO_EXPORT void
+PyDataMem_FREE(void *ptr)
+{
+    free(ptr);
+    if (_PyDataMem_eventhook != NULL) {
+        PyGILState_STATE gilstate = PyGILState_Ensure();
+        if (_PyDataMem_eventhook != NULL) {
+            (*_PyDataMem_eventhook)(ptr, NULL, 0,
+                                    _PyDataMem_eventhook_user_data);
+        }
+        PyGILState_Release(gilstate);
+    }
+}
+
+/*NUMPY_API
+ * Reallocate/resize memory for array data.
+ */
+NPY_NO_EXPORT void *
+PyDataMem_RENEW(void *ptr, size_t size)
+{
+    void *result;
+
+    result = realloc(ptr, size);
+    if (_PyDataMem_eventhook != NULL) {
+        PyGILState_STATE gilstate = PyGILState_Ensure();
+        if (_PyDataMem_eventhook != NULL) {
+            (*_PyDataMem_eventhook)(ptr, result, size,
+                                    _PyDataMem_eventhook_user_data);
+        }
+        PyGILState_Release(gilstate);
+    }
+    return (char *)result;
+}
+
 static struct PyMethodDef array_module_methods[] = {
     {"_get_ndarray_c_version",
         (PyCFunction)array__get_ndarray_c_version,
diff --git a/numpy/core/src/npysort/sort.c.src b/numpy/core/src/npysort/sort.c.src
index 618259f96..514131748 100644
--- a/numpy/core/src/npysort/sort.c.src
+++ b/numpy/core/src/npysort/sort.c.src
@@ -38,6 +38,12 @@
 #define SMALL_MERGESORT 20
 #define SMALL_STRING 16
 
+/* These should be changed to PyDataMem_NEW/FREE if npysort is moved
+ *  to the multiarray directory.
+ */
+#define NpySortArray_malloc(size) ((char *)malloc(size))
+#define NpySortArray_free(ptr)  free(ptr)
+
 
 /*
  *****************************************************************************
@@ -335,14 +341,14 @@ mergesort_@suff@(@type@ *start, npy_intp num, void *NOT_USED)
 
     pl = start;
     pr = pl + num;
-    pw = (@type@ *) PyDataMem_NEW((num/2)*sizeof(@type@));
+    pw = (@type@ *) NpySortArray_malloc((num/2)*sizeof(@type@));
     if (!pw) {
         PyErr_NoMemory();
         return -1;
     }
     mergesort0_@suff@(pl, pr, pw);
 
-    PyDataMem_FREE(pw);
+    NpySortArray_free(pw);
     return 0;
 }
 
@@ -476,13 +482,13 @@ mergesort_@suff@(@type@ *start, npy_intp num, PyArrayObject *arr)
 
     pl = start;
     pr = pl + num*len;
-    pw = (@type@ *) PyDataMem_NEW((num/2)*elsize);
+    pw = (@type@ *) NpySortArray_malloc((num/2)*elsize);
     if (!pw) {
         PyErr_NoMemory();
         err = -1;
         goto fail_0;
     }
-    vp = (@type@ *) PyDataMem_NEW(elsize);
+    vp = (@type@ *) NpySortArray_malloc(elsize);
     if (!vp) {
         PyErr_NoMemory();
         err = -1;
@@ -490,9 +496,9 @@ mergesort_@suff@(@type@ *start, npy_intp num, PyArrayObject *arr)
     }
     mergesort0_@suff@(pl, pr, pw, vp, len);
 
-    PyDataMem_FREE(vp);
+    NpySortArray_free(vp);
 fail_1:
-    PyDataMem_FREE(pw);
+    NpySortArray_free(pw);
 fail_0:
     return err;
 }
diff --git a/numpy/core/tests/test_multiarray.py b/numpy/core/tests/test_multiarray.py
index 82ead8958..73abc81b9 100644
--- a/numpy/core/tests/test_multiarray.py
+++ b/numpy/core/tests/test_multiarray.py
@@ -10,7 +10,8 @@ from numpy.testing.utils import WarningManager
 from numpy.compat import asbytes, getexception, strchar
 from test_print import in_foreign_locale
 from numpy.core.multiarray_tests import (
-        test_neighborhood_iterator, test_neighborhood_iterator_oob
+        test_neighborhood_iterator, test_neighborhood_iterator_oob,
+        test_pydatamem_seteventhook_start, test_pydatamem_seteventhook_end,
         )
 from numpy.testing import (
         TestCase, run_module_suite, assert_, assert_raises,
@@ -2649,6 +2650,16 @@ def test_flat_element_deletion():
     except:
         raise AssertionError
 
+class TestMemEventHook(TestCase):
+    def test_mem_seteventhook(self):
+        # The actual tests are within the C code in
+        # multiarray/multiarray_tests.c.src
+        test_pydatamem_seteventhook_start()
+        # force an allocation and free of a numpy array
+        a = np.zeros(10)
+        del a
+        test_pydatamem_seteventhook_end()
+
 
 if __name__ == "__main__":
     run_module_suite()
diff --git a/numpy/lib/src/_compiled_base.c b/numpy/lib/src/_compiled_base.c
index c31ee5cd8..9bb8a613d 100644
--- a/numpy/lib/src/_compiled_base.c
+++ b/numpy/lib/src/_compiled_base.c
@@ -670,7 +670,10 @@ arr_interp(PyObject *NPY_UNUSED(self), PyObject *args, PyObject *kwdict)
 
     /* only pre-calculate slopes if there are relatively few of them. */
     if (lenxp <= lenx) {
-        slopes = (double *) PyDataMem_NEW((lenxp - 1)*sizeof(double));
+        slopes = (double *) PyArray_malloc((lenxp - 1)*sizeof(double));
+        if (! slopes) {
+            goto fail;
+        }
         NPY_BEGIN_ALLOW_THREADS;
         for (i = 0; i < lenxp - 1; i++) {
             slopes[i] = (dy[i + 1] - dy[i])/(dx[i + 1] - dx[i]);
@@ -692,7 +695,7 @@ arr_interp(PyObject *NPY_UNUSED(self), PyObject *args, PyObject *kwdict)
             }
         }
         NPY_END_ALLOW_THREADS;
-        PyDataMem_FREE(slopes);
+        PyArray_free(slopes);
     }
     else {
         NPY_BEGIN_ALLOW_THREADS;
diff --git a/tools/allocation_tracking/alloc_hook.pyx b/tools/allocation_tracking/alloc_hook.pyx
new file mode 100644
index 000000000..d1e656f90
--- /dev/null
+++ b/tools/allocation_tracking/alloc_hook.pyx
@@ -0,0 +1,42 @@
+# A cython wrapper for using python functions as callbacks for
+# PyDataMem_SetEventHook.
+
+cimport numpy as np
+
+cdef extern from "Python.h":
+    object PyLong_FromVoidPtr(void *)
+    void *PyLong_AsVoidPtr(object)
+
+ctypedef void PyDataMem_EventHookFunc(void *inp, void *outp, size_t size,
+                                      void *user_data)
+cdef extern from "numpy/arrayobject.h":
+    PyDataMem_EventHookFunc * \
+        PyDataMem_SetEventHook(PyDataMem_EventHookFunc *newhook,
+                               void *user_data, void **old_data)
+
+np.import_array()
+
+cdef void pyhook(void *old, void *new, size_t size, void *user_data):
+    cdef object pyfunc = <object> user_data
+    pyfunc(PyLong_FromVoidPtr(old),
+           PyLong_FromVoidPtr(new),
+           size)
+
+class NumpyAllocHook(object):
+    def __init__(self, callback):
+        self.callback = callback
+
+    def __enter__(self):
+        cdef void *old_hook, *old_data
+        old_hook = <void *> \
+            PyDataMem_SetEventHook(<PyDataMem_EventHookFunc *> pyhook,
+                                    <void *> self.callback,
+                                    <void **> &old_data)
+        self.old_hook = PyLong_FromVoidPtr(old_hook)
+        self.old_data = PyLong_FromVoidPtr(old_data)
+
+    def __exit__(self):
+        PyDataMem_SetEventHook(<PyDataMem_EventHookFunc *> \
+                                    PyLong_AsVoidPtr(self.old_hook),
+                                <void *> PyLong_AsVoidPtr(self.old_data),
+                                <void **> 0)
diff --git a/tools/allocation_tracking/setup.py b/tools/allocation_tracking/setup.py
new file mode 100644
index 000000000..4462f9f4e
--- /dev/null
+++ b/tools/allocation_tracking/setup.py
@@ -0,0 +1,9 @@
+from distutils.core import setup
+from distutils.extension import Extension
+from Cython.Distutils import build_ext
+import numpy
+
+setup(
+    cmdclass = {'build_ext': build_ext},
+    ext_modules = [Extension("alloc_hook", ["alloc_hook.pyx"],
+                             include_dirs=[numpy.get_include()])])
diff --git a/tools/allocation_tracking/sorttable.js b/tools/allocation_tracking/sorttable.js
new file mode 100644
index 000000000..25bccb2b6
--- /dev/null
+++ b/tools/allocation_tracking/sorttable.js
@@ -0,0 +1,493 @@
+/*
+  SortTable
+  version 2
+  7th April 2007
+  Stuart Langridge, http://www.kryogenix.org/code/browser/sorttable/
+  
+  Instructions:
+  Download this file
+  Add <script src="sorttable.js"></script> to your HTML
+  Add class="sortable" to any table you'd like to make sortable
+  Click on the headers to sort
+  
+  Thanks to many, many people for contributions and suggestions.
+  Licenced as X11: http://www.kryogenix.org/code/browser/licence.html
+  This basically means: do what you want with it.
+*/
+
+ 
+var stIsIE = /*@cc_on!@*/false;
+
+sorttable = {
+  init: function() {
+    // quit if this function has already been called
+    if (arguments.callee.done) return;
+    // flag this function so we don't do the same thing twice
+    arguments.callee.done = true;
+    // kill the timer
+    if (_timer) clearInterval(_timer);
+    
+    if (!document.createElement || !document.getElementsByTagName) return;
+    
+    sorttable.DATE_RE = /^(\d\d?)[\/\.-](\d\d?)[\/\.-]((\d\d)?\d\d)$/;
+    
+    forEach(document.getElementsByTagName('table'), function(table) {
+      if (table.className.search(/\bsortable\b/) != -1) {
+        sorttable.makeSortable(table);
+      }
+    });
+    
+  },
+  
+  makeSortable: function(table) {
+    if (table.getElementsByTagName('thead').length == 0) {
+      // table doesn't have a tHead. Since it should have, create one and
+      // put the first table row in it.
+      the = document.createElement('thead');
+      the.appendChild(table.rows[0]);
+      table.insertBefore(the,table.firstChild);
+    }
+    // Safari doesn't support table.tHead, sigh
+    if (table.tHead == null) table.tHead = table.getElementsByTagName('thead')[0];
+    
+    if (table.tHead.rows.length != 1) return; // can't cope with two header rows
+    
+    // Sorttable v1 put rows with a class of "sortbottom" at the bottom (as
+    // "total" rows, for example). This is B&R, since what you're supposed
+    // to do is put them in a tfoot. So, if there are sortbottom rows,
+    // for backwards compatibility, move them to tfoot (creating it if needed).
+    sortbottomrows = [];
+    for (var i=0; i<table.rows.length; i++) {
+      if (table.rows[i].className.search(/\bsortbottom\b/) != -1) {
+        sortbottomrows[sortbottomrows.length] = table.rows[i];
+      }
+    }
+    if (sortbottomrows) {
+      if (table.tFoot == null) {
+        // table doesn't have a tfoot. Create one.
+        tfo = document.createElement('tfoot');
+        table.appendChild(tfo);
+      }
+      for (var i=0; i<sortbottomrows.length; i++) {
+        tfo.appendChild(sortbottomrows[i]);
+      }
+      delete sortbottomrows;
+    }
+    
+    // work through each column and calculate its type
+    headrow = table.tHead.rows[0].cells;
+    for (var i=0; i<headrow.length; i++) {
+      // manually override the type with a sorttable_type attribute
+      if (!headrow[i].className.match(/\bsorttable_nosort\b/)) { // skip this col
+        mtch = headrow[i].className.match(/\bsorttable_([a-z0-9]+)\b/);
+        if (mtch) { override = mtch[1]; }
+	      if (mtch && typeof sorttable["sort_"+override] == 'function') {
+	        headrow[i].sorttable_sortfunction = sorttable["sort_"+override];
+	      } else {
+	        headrow[i].sorttable_sortfunction = sorttable.guessType(table,i);
+	      }
+	      // make it clickable to sort
+	      headrow[i].sorttable_columnindex = i;
+	      headrow[i].sorttable_tbody = table.tBodies[0];
+	      dean_addEvent(headrow[i],"click", function(e) {
+
+          if (this.className.search(/\bsorttable_sorted\b/) != -1) {
+            // if we're already sorted by this column, just 
+            // reverse the table, which is quicker
+            sorttable.reverse(this.sorttable_tbody);
+            this.className = this.className.replace('sorttable_sorted',
+                                                    'sorttable_sorted_reverse');
+            this.removeChild(document.getElementById('sorttable_sortfwdind'));
+            sortrevind = document.createElement('span');
+            sortrevind.id = "sorttable_sortrevind";
+            sortrevind.innerHTML = stIsIE ? '&nbsp<font face="webdings">5</font>' : '&nbsp;&#x25B4;';
+            this.appendChild(sortrevind);
+            return;
+          }
+          if (this.className.search(/\bsorttable_sorted_reverse\b/) != -1) {
+            // if we're already sorted by this column in reverse, just 
+            // re-reverse the table, which is quicker
+            sorttable.reverse(this.sorttable_tbody);
+            this.className = this.className.replace('sorttable_sorted_reverse',
+                                                    'sorttable_sorted');
+            this.removeChild(document.getElementById('sorttable_sortrevind'));
+            sortfwdind = document.createElement('span');
+            sortfwdind.id = "sorttable_sortfwdind";
+            sortfwdind.innerHTML = stIsIE ? '&nbsp<font face="webdings">6</font>' : '&nbsp;&#x25BE;';
+            this.appendChild(sortfwdind);
+            return;
+          }
+          
+          // remove sorttable_sorted classes
+          theadrow = this.parentNode;
+          forEach(theadrow.childNodes, function(cell) {
+            if (cell.nodeType == 1) { // an element
+              cell.className = cell.className.replace('sorttable_sorted_reverse','');
+              cell.className = cell.className.replace('sorttable_sorted','');
+            }
+          });
+          sortfwdind = document.getElementById('sorttable_sortfwdind');
+          if (sortfwdind) { sortfwdind.parentNode.removeChild(sortfwdind); }
+          sortrevind = document.getElementById('sorttable_sortrevind');
+          if (sortrevind) { sortrevind.parentNode.removeChild(sortrevind); }
+          
+          this.className += ' sorttable_sorted';
+          sortfwdind = document.createElement('span');
+          sortfwdind.id = "sorttable_sortfwdind";
+          sortfwdind.innerHTML = stIsIE ? '&nbsp<font face="webdings">6</font>' : '&nbsp;&#x25BE;';
+          this.appendChild(sortfwdind);
+
+	        // build an array to sort. This is a Schwartzian transform thing,
+	        // i.e., we "decorate" each row with the actual sort key,
+	        // sort based on the sort keys, and then put the rows back in order
+	        // which is a lot faster because you only do getInnerText once per row
+	        row_array = [];
+	        col = this.sorttable_columnindex;
+	        rows = this.sorttable_tbody.rows;
+	        for (var j=0; j<rows.length; j++) {
+	          row_array[row_array.length] = [sorttable.getInnerText(rows[j].cells[col]), rows[j]];
+	        }
+	        /* If you want a stable sort, uncomment the following line */
+	        //sorttable.shaker_sort(row_array, this.sorttable_sortfunction);
+	        /* and comment out this one */
+	        row_array.sort(this.sorttable_sortfunction);
+	        
+	        tb = this.sorttable_tbody;
+	        for (var j=0; j<row_array.length; j++) {
+	          tb.appendChild(row_array[j][1]);
+	        }
+	        
+	        delete row_array;
+	      });
+	    }
+    }
+  },
+  
+  guessType: function(table, column) {
+    // guess the type of a column based on its first non-blank row
+    sortfn = sorttable.sort_alpha;
+    for (var i=0; i<table.tBodies[0].rows.length; i++) {
+      text = sorttable.getInnerText(table.tBodies[0].rows[i].cells[column]);
+      if (text != '') {
+        if (text.match(/^-?[£$¤]?[\d,.]+%?$/)) {
+          return sorttable.sort_numeric;
+        }
+        // check for a date: dd/mm/yyyy or dd/mm/yy 
+        // can have / or . or - as separator
+        // can be mm/dd as well
+        possdate = text.match(sorttable.DATE_RE)
+        if (possdate) {
+          // looks like a date
+          first = parseInt(possdate[1]);
+          second = parseInt(possdate[2]);
+          if (first > 12) {
+            // definitely dd/mm
+            return sorttable.sort_ddmm;
+          } else if (second > 12) {
+            return sorttable.sort_mmdd;
+          } else {
+            // looks like a date, but we can't tell which, so assume
+            // that it's dd/mm (English imperialism!) and keep looking
+            sortfn = sorttable.sort_ddmm;
+          }
+        }
+      }
+    }
+    return sortfn;
+  },
+  
+  getInnerText: function(node) {
+    // gets the text we want to use for sorting for a cell.
+    // strips leading and trailing whitespace.
+    // this is *not* a generic getInnerText function; it's special to sorttable.
+    // for example, you can override the cell text with a customkey attribute.
+    // it also gets .value for <input> fields.
+    
+    hasInputs = (typeof node.getElementsByTagName == 'function') &&
+                 node.getElementsByTagName('input').length;
+    
+    if (node.getAttribute("sorttable_customkey") != null) {
+      return node.getAttribute("sorttable_customkey");
+    }
+    else if (typeof node.textContent != 'undefined' && !hasInputs) {
+      return node.textContent.replace(/^\s+|\s+$/g, '');
+    }
+    else if (typeof node.innerText != 'undefined' && !hasInputs) {
+      return node.innerText.replace(/^\s+|\s+$/g, '');
+    }
+    else if (typeof node.text != 'undefined' && !hasInputs) {
+      return node.text.replace(/^\s+|\s+$/g, '');
+    }
+    else {
+      switch (node.nodeType) {
+        case 3:
+          if (node.nodeName.toLowerCase() == 'input') {
+            return node.value.replace(/^\s+|\s+$/g, '');
+          }
+        case 4:
+          return node.nodeValue.replace(/^\s+|\s+$/g, '');
+          break;
+        case 1:
+        case 11:
+          var innerText = '';
+          for (var i = 0; i < node.childNodes.length; i++) {
+            innerText += sorttable.getInnerText(node.childNodes[i]);
+          }
+          return innerText.replace(/^\s+|\s+$/g, '');
+          break;
+        default:
+          return '';
+      }
+    }
+  },
+  
+  reverse: function(tbody) {
+    // reverse the rows in a tbody
+    newrows = [];
+    for (var i=0; i<tbody.rows.length; i++) {
+      newrows[newrows.length] = tbody.rows[i];
+    }
+    for (var i=newrows.length-1; i>=0; i--) {
+       tbody.appendChild(newrows[i]);
+    }
+    delete newrows;
+  },
+  
+  /* sort functions
+     each sort function takes two parameters, a and b
+     you are comparing a[0] and b[0] */
+  sort_numeric: function(a,b) {
+    aa = parseFloat(a[0].replace(/[^0-9.-]/g,''));
+    if (isNaN(aa)) aa = 0;
+    bb = parseFloat(b[0].replace(/[^0-9.-]/g,'')); 
+    if (isNaN(bb)) bb = 0;
+    return aa-bb;
+  },
+  sort_alpha: function(a,b) {
+    if (a[0]==b[0]) return 0;
+    if (a[0]<b[0]) return -1;
+    return 1;
+  },
+  sort_ddmm: function(a,b) {
+    mtch = a[0].match(sorttable.DATE_RE);
+    y = mtch[3]; m = mtch[2]; d = mtch[1];
+    if (m.length == 1) m = '0'+m;
+    if (d.length == 1) d = '0'+d;
+    dt1 = y+m+d;
+    mtch = b[0].match(sorttable.DATE_RE);
+    y = mtch[3]; m = mtch[2]; d = mtch[1];
+    if (m.length == 1) m = '0'+m;
+    if (d.length == 1) d = '0'+d;
+    dt2 = y+m+d;
+    if (dt1==dt2) return 0;
+    if (dt1<dt2) return -1;
+    return 1;
+  },
+  sort_mmdd: function(a,b) {
+    mtch = a[0].match(sorttable.DATE_RE);
+    y = mtch[3]; d = mtch[2]; m = mtch[1];
+    if (m.length == 1) m = '0'+m;
+    if (d.length == 1) d = '0'+d;
+    dt1 = y+m+d;
+    mtch = b[0].match(sorttable.DATE_RE);
+    y = mtch[3]; d = mtch[2]; m = mtch[1];
+    if (m.length == 1) m = '0'+m;
+    if (d.length == 1) d = '0'+d;
+    dt2 = y+m+d;
+    if (dt1==dt2) return 0;
+    if (dt1<dt2) return -1;
+    return 1;
+  },
+  
+  shaker_sort: function(list, comp_func) {
+    // A stable sort function to allow multi-level sorting of data
+    // see: http://en.wikipedia.org/wiki/Cocktail_sort
+    // thanks to Joseph Nahmias
+    var b = 0;
+    var t = list.length - 1;
+    var swap = true;
+
+    while(swap) {
+        swap = false;
+        for(var i = b; i < t; ++i) {
+            if ( comp_func(list[i], list[i+1]) > 0 ) {
+                var q = list[i]; list[i] = list[i+1]; list[i+1] = q;
+                swap = true;
+            }
+        } // for
+        t--;
+
+        if (!swap) break;
+
+        for(var i = t; i > b; --i) {
+            if ( comp_func(list[i], list[i-1]) < 0 ) {
+                var q = list[i]; list[i] = list[i-1]; list[i-1] = q;
+                swap = true;
+            }
+        } // for
+        b++;
+
+    } // while(swap)
+  }  
+}
+
+/* ******************************************************************
+   Supporting functions: bundled here to avoid depending on a library
+   ****************************************************************** */
+
+// Dean Edwards/Matthias Miller/John Resig
+
+/* for Mozilla/Opera9 */
+if (document.addEventListener) {
+    document.addEventListener("DOMContentLoaded", sorttable.init, false);
+}
+
+/* for Internet Explorer */
+/*@cc_on @*/
+/*@if (@_win32)
+    document.write("<script id=__ie_onload defer src=javascript:void(0)><\/script>");
+    var script = document.getElementById("__ie_onload");
+    script.onreadystatechange = function() {
+        if (this.readyState == "complete") {
+            sorttable.init(); // call the onload handler
+        }
+    };
+/*@end @*/
+
+/* for Safari */
+if (/WebKit/i.test(navigator.userAgent)) { // sniff
+    var _timer = setInterval(function() {
+        if (/loaded|complete/.test(document.readyState)) {
+            sorttable.init(); // call the onload handler
+        }
+    }, 10);
+}
+
+/* for other browsers */
+window.onload = sorttable.init;
+
+// written by Dean Edwards, 2005
+// with input from Tino Zijdel, Matthias Miller, Diego Perini
+
+// http://dean.edwards.name/weblog/2005/10/add-event/
+
+function dean_addEvent(element, type, handler) {
+	if (element.addEventListener) {
+		element.addEventListener(type, handler, false);
+	} else {
+		// assign each event handler a unique ID
+		if (!handler.$$guid) handler.$$guid = dean_addEvent.guid++;
+		// create a hash table of event types for the element
+		if (!element.events) element.events = {};
+		// create a hash table of event handlers for each element/event pair
+		var handlers = element.events[type];
+		if (!handlers) {
+			handlers = element.events[type] = {};
+			// store the existing event handler (if there is one)
+			if (element["on" + type]) {
+				handlers[0] = element["on" + type];
+			}
+		}
+		// store the event handler in the hash table
+		handlers[handler.$$guid] = handler;
+		// assign a global event handler to do all the work
+		element["on" + type] = handleEvent;
+	}
+};
+// a counter used to create unique IDs
+dean_addEvent.guid = 1;
+
+function removeEvent(element, type, handler) {
+	if (element.removeEventListener) {
+		element.removeEventListener(type, handler, false);
+	} else {
+		// delete the event handler from the hash table
+		if (element.events && element.events[type]) {
+			delete element.events[type][handler.$$guid];
+		}
+	}
+};
+
+function handleEvent(event) {
+	var returnValue = true;
+	// grab the event object (IE uses a global event object)
+	event = event || fixEvent(((this.ownerDocument || this.document || this).parentWindow || window).event);
+	// get a reference to the hash table of event handlers
+	var handlers = this.events[event.type];
+	// execute each event handler
+	for (var i in handlers) {
+		this.$$handleEvent = handlers[i];
+		if (this.$$handleEvent(event) === false) {
+			returnValue = false;
+		}
+	}
+	return returnValue;
+};
+
+function fixEvent(event) {
+	// add W3C standard event methods
+	event.preventDefault = fixEvent.preventDefault;
+	event.stopPropagation = fixEvent.stopPropagation;
+	return event;
+};
+fixEvent.preventDefault = function() {
+	this.returnValue = false;
+};
+fixEvent.stopPropagation = function() {
+  this.cancelBubble = true;
+}
+
+// Dean's forEach: http://dean.edwards.name/base/forEach.js
+/*
+	forEach, version 1.0
+	Copyright 2006, Dean Edwards
+	License: http://www.opensource.org/licenses/mit-license.php
+*/
+
+// array-like enumeration
+if (!Array.forEach) { // mozilla already supports this
+	Array.forEach = function(array, block, context) {
+		for (var i = 0; i < array.length; i++) {
+			block.call(context, array[i], i, array);
+		}
+	};
+}
+
+// generic enumeration
+Function.prototype.forEach = function(object, block, context) {
+	for (var key in object) {
+		if (typeof this.prototype[key] == "undefined") {
+			block.call(context, object[key], key, object);
+		}
+	}
+};
+
+// character enumeration
+String.forEach = function(string, block, context) {
+	Array.forEach(string.split(""), function(chr, index) {
+		block.call(context, chr, index, string);
+	});
+};
+
+// globally resolve forEach enumeration
+var forEach = function(object, block, context) {
+	if (object) {
+		var resolve = Object; // default
+		if (object instanceof Function) {
+			// functions have a "length" property
+			resolve = Function;
+		} else if (object.forEach instanceof Function) {
+			// the object implements a custom forEach method so use that
+			object.forEach(block, context);
+			return;
+		} else if (typeof object == "string") {
+			// the object is a string
+			resolve = String;
+		} else if (typeof object.length == "number") {
+			// the object is array-like
+			resolve = Array;
+		}
+		resolve.forEach(object, block, context);
+	}
+};
+
diff --git a/tools/allocation_tracking/track_allocations.py b/tools/allocation_tracking/track_allocations.py
new file mode 100644
index 000000000..e9d937817
--- /dev/null
+++ b/tools/allocation_tracking/track_allocations.py
@@ -0,0 +1,131 @@
+import numpy as np
+import inspect
+from alloc_hook import NumpyAllocHook
+
+class AllocationTracker(object):
+    def __init__(self, threshold=0):
+        '''track numpy allocations of size threshold bytes or more.'''
+
+        self.threshold = threshold
+
+        # The total number of bytes currently allocated with size above
+        # threshold
+        self.total_bytes = 0
+
+        # We buffer requests line by line and move them into the allocation
+        # trace when a new line occurs
+        self.current_line = None
+        self.pending_allocations = []
+
+        self.blocksizes = {}
+
+        # list of (lineinfo, bytes allocated, bytes freed, # allocations, #
+        # frees, maximum memory usage, long-lived bytes allocated)
+        self.allocation_trace = []
+
+        self.numpy_hook = NumpyAllocHook(self.hook)
+
+    def __enter__(self):
+        self.numpy_hook.__enter__()
+
+    def __exit__(self, type, value, traceback):
+        self.check_line_changed()  # forces pending events to be handled
+        self.numpy_hook.__exit__()
+
+    def hook(self, inptr, outptr, size):
+        if outptr == 0:  # it's a free
+            self.free_cb(inptr)
+        elif inptr != 0:  # realloc
+            self.realloc_cb(inptr, outptr, size)
+        else:  # malloc
+            self.alloc_cb(outptr, size)
+
+    def alloc_cb(self, ptr, size):
+        if size >= self.threshold:
+            self.check_line_changed()
+            self.blocksizes[ptr] = size
+            self.pending_allocations.append(size)
+
+    def free_cb(self, ptr):
+        size = self.blocksizes.pop(ptr, 0)
+        if size:
+            self.check_line_changed()
+            self.pending_allocations.append(-size)
+
+    def realloc_cb(self, newptr, oldptr, size):
+        if (size >= self.threshold) or (oldptr in self.blocksizes):
+            self.check_line_changed()
+            oldsize = self.blocksizes.pop(oldptr, 0)
+            self.pending_allocations.append(size - oldsize)
+            self.blocksizes[newptr] = size
+
+    def get_code_line(self):
+        # first frame is this line, then check_line_changed(), then 2 callbacks,
+        # then actual code.
+        try:
+            return inspect.stack()[4][1:]
+        except:
+            return inspect.stack()[0][1:]
+
+    def check_line_changed(self):
+        line = self.get_code_line()
+        if line != self.current_line and (self.current_line is not None):
+            # move pending events into the allocation_trace
+            max_size = self.total_bytes
+            bytes_allocated = 0
+            bytes_freed = 0
+            num_allocations = 0
+            num_frees = 0
+            before_size = self.total_bytes
+            for allocation in self.pending_allocations:
+                self.total_bytes += allocation
+                if allocation > 0:
+                    bytes_allocated += allocation
+                    num_allocations += 1
+                else:
+                    bytes_freed += -allocation
+                    num_frees += 1
+                max_size = max(max_size, self.total_bytes)
+            long_lived = max(self.total_bytes - before_size, 0)
+            self.allocation_trace.append((self.current_line, bytes_allocated,
+                                          bytes_freed, num_allocations,
+                                          num_frees, max_size, long_lived))
+            # clear pending allocations
+            self.pending_allocations = []
+        # move to the new line
+        self.current_line = line
+
+    def write_html(self, filename):
+        f = open(filename, "w")
+        f.write('<HTML><HEAD><script src="sorttable.js"></script></HEAD><BODY>\n')
+        f.write('<TABLE class="sortable" width=100%>\n')
+        f.write("<TR>\n")
+        cols = "event#,lineinfo,bytes allocated,bytes freed,#allocations,#frees,max memory usage,long lived bytes".split(',')
+        for header in cols:
+            f.write("  <TH>{0}</TH>".format(header))
+        f.write("\n</TR>\n")
+        for idx, event in enumerate(self.allocation_trace):
+            f.write("<TR>\n")
+            event = [idx] + list(event)
+            for col, val in zip(cols, event):
+                if col == 'lineinfo':
+                    # special handling
+                    try:
+                        filename, line, module, code, index = val
+                        val = "{0}({1}): {2}".format(filename, line, code[index])
+                    except:
+                        # sometimes this info is not available (from eval()?)
+                        val = str(val)
+                f.write("  <TD>{0}</TD>".format(val))
+            f.write("\n</TR>\n")
+        f.write("</TABLE></BODY></HTML>\n")
+        f.close()
+
+
+if __name__ == '__main__':
+    tracker = AllocationTracker(1000)
+    with tracker:
+        for i in range(100):
+            np.zeros(i * 100)
+            np.zeros(i * 200)
+    tracker.write_html("allocations.html")