Fix so USE_USE_DEFAULTS code works in multi-threaded case. Speed up 1-d array indexing by an integer.

4 files changed, 99 insertions, 96 deletions

diff --git a/numpy/core/numeric.py b/numpy/core/numeric.py
index f57527990..557c8e7f4 100644
--- a/numpy/core/numeric.py
+++ b/numpy/core/numeric.py
@@ -395,13 +395,6 @@ def allclose (a, b, rtol=1.e-5, atol=1.e-8):
     return d.ravel().all()
 
 
-class ufunc_values_obj(object):
-    def __init__(self, obj):
-        self._val_obj = obj
-    def __del__(self):
-        umath.seterrobj(self._val_obj)
-        del self._val_obj
-        
 
 _errdict = {"ignore":ERR_IGNORE,
             "warn":ERR_WARN,
@@ -413,11 +406,11 @@ for key in _errdict.keys():
     _errdict_rev[_errdict[key]] = key
 del key
 
-def seterr(divide=None, over=None, under=None, invalid=None):

-

+def seterr(divide=None, over=None, under=None, invalid=None):
+
     pyvals = umath.geterrobj()
-    old = geterr()

-

+    old = geterr()
+
     if divide is None: divide = old['divide']
     if over is None: over = old['over']
     if under is None: under = old['under']
diff --git a/numpy/core/src/arrayobject.c b/numpy/core/src/arrayobject.c
index 0bd14c251..d4b4bb83f 100644
--- a/numpy/core/src/arrayobject.c
+++ b/numpy/core/src/arrayobject.c
@@ -1468,10 +1468,22 @@ array_big_item(PyArrayObject *self, intp i)
 	return (PyObject *)r;
 }
 
+/* contains optimization for 1-d arrays */
 static PyObject *
 array_item_nice(PyArrayObject *self, _int_or_ssize_t i)
 {
-	return PyArray_Return((PyArrayObject *)array_big_item(self, (intp) i));
+	if (self->nd == 1) {
+		char *item;
+                if (i < 0) {
+			i += self->dimensions[0];
+		}
+		if ((item = index2ptr(self, i)) == NULL) return NULL;
+		return PyArray_Scalar(item, self->descr, (PyObject *)self);
+	}
+	else {
+		return PyArray_Return((PyArrayObject *)\
+				      array_big_item(self, (intp) i));
+	}
 }
 
 static int
@@ -2344,7 +2356,22 @@ array_subscript_nice(PyArrayObject *self, PyObject *op)
 	   implementation may be possible by refactoring
 	   array_subscript */
 
-	PyArrayObject *mp = (PyArrayObject *)array_subscript(self, op);
+	PyArrayObject *mp;
+
+	/* optimization for integer select and 1-d */
+	if (self->nd == 1 && (PyInt_Check(op) || PyLong_Check(op))) {
+                intp value;
+		char *item;
+                value = PyArray_PyIntAsIntp(op);
+		if (PyErr_Occurred()) 
+			return NULL;
+                else if (value < 0) {
+			value += self->dimensions[0];
+		}
+		if ((item = index2ptr(self, value)) == NULL) return NULL;
+		return PyArray_Scalar(item, self->descr, (PyObject *)self);
+	}
+	mp = (PyArrayObject *)array_subscript(self, op);
 
 	if (mp == NULL) return NULL;
 
diff --git a/numpy/core/src/ufuncobject.c b/numpy/core/src/ufuncobject.c
index bc439b61c..40cf40735 100644
--- a/numpy/core/src/ufuncobject.c
+++ b/numpy/core/src/ufuncobject.c
@@ -35,9 +35,9 @@ typedef void (CdoubleBinaryFunc)(cdouble *x, cdouble *y, cdouble *res);
 typedef void (CfloatBinaryFunc)(cfloat *x, cfloat *y, cfloat *res);
 typedef void (ClongdoubleBinaryFunc)(clongdouble *x, clongdouble *y, \
 				     clongdouble *res);
-

-#define USE_USE_DEFAULTS 0

-

+
+#define USE_USE_DEFAULTS 1
+
 
 /*UFUNC_API*/
 static void
@@ -699,11 +699,11 @@ select_types(PyUFuncObject *self, int *arg_types,
 
 	return 0;
 }
-

-#if USE_USE_DEFAULTS
-static int PyUFunc_USEDEFAULTS=0;

+
+#if USE_USE_DEFAULTS==1
+static int PyUFunc_NUM_NODEFAULTS=0;
 #endif
-static PyObject *PyUFunc_PYVALS_NAME=NULL;

+static PyObject *PyUFunc_PYVALS_NAME=NULL;
 
 
 /*UFUNC_API*/
@@ -713,20 +713,21 @@ PyUFunc_GetPyValues(char *name, int *bufsize, int *errmask, PyObject **errobj)
         PyObject *thedict;
         PyObject *ref=NULL;
 	PyObject *retval;
-

-        #if USE_USE_DEFAULTS
-	if (!PyUFunc_USEDEFAULTS) {

+
+        #if USE_USE_DEFAULTS==1
+	if (PyUFunc_NUM_NODEFAULTS != 0) {
         #endif
 		if (PyUFunc_PYVALS_NAME == NULL) {
-			PyUFunc_PYVALS_NAME = PyString_InternFromString(UFUNC_PYVALS_NAME);
+			PyUFunc_PYVALS_NAME = \
+				PyString_InternFromString(UFUNC_PYVALS_NAME);
 		}
 		thedict = PyThreadState_GetDict();
 		if (thedict == NULL) {
 			thedict = PyEval_GetBuiltins();
 		}
-		ref = PyDict_GetItem(thedict, PyUFunc_PYVALS_NAME);

-        #if USE_USE_DEFAULTS
-	}

+		ref = PyDict_GetItem(thedict, PyUFunc_PYVALS_NAME);
+        #if USE_USE_DEFAULTS==1
+	}
         #endif
 	if (ref == NULL) {
 		*errmask = UFUNC_ERR_DEFAULT;
@@ -2768,7 +2769,6 @@ ufunc_geterr(PyObject *dummy, PyObject *args)
 		return res;
 	}
 	/* Construct list of defaults */
-	fprintf(stderr, "Nothing found... return defaults.\n");
 	res = PyList_New(3);
 	if (res == NULL) return NULL;
 	PyList_SET_ITEM(res, 0, PyInt_FromLong(PyArray_BUFSIZE));
@@ -2776,43 +2776,39 @@ ufunc_geterr(PyObject *dummy, PyObject *args)
 	PyList_SET_ITEM(res, 2, Py_None); Py_INCREF(Py_None);
 	return res;
 }
-

-#if USE_USE_DEFAULTS

-/* 

-This doesn't look it will work in the presence of threads. It updates 

-PyUFunc_USEDEFAULTS based on the current thread. If some other thread is 

-around, it will see an incorrect value for use_defaults.

-

-I think the following strategy would fix this:

-    1. Change PyUFunc_USEDEFAULTS to PyUFunc_NONDEFAULTCOUNT or similar

-    2. Increment PyUFunc_NONDEFAULTCOUNT whenever a value is set to a nondefault

-       value

-    3. Only use defaults when PyUFunc_NONDEFAULTCOUNT is nonzero.

-

-However, I'm not sure that it's worth the trouble. I've done a few small 

-benchmarks and I see at most marginal speed improvements with the

-default values. So, for the time being, I'm simply ifdefing out the 

-nonworking code and not worrying about it. If those benchmarks hold up, we

-should go ahead and rip the code out so as not to confuse future generations.

-

+
+#if USE_USE_DEFAULTS==1
+/* 
+This is a strategy to buy a little speed up and avoid the dictionary
+look-up in the default case.  It should work in the presence of
+threads.  If it is deemed too complicated or it doesn't actually work
+it could be taken out.
 */
 static int 
 ufunc_update_use_defaults(void)
 {
 	PyObject *errobj;
 	int errmask, bufsize;
-
-	PyUFunc_USEDEFAULTS = 0;
-	if (PyUFunc_GetPyValues("test", &bufsize, &errmask, &errobj) < 0) return -1;
+	int res;
 	
-	if ((errmask == UFUNC_ERR_DEFAULT) &&		\
-	    (bufsize == PyArray_BUFSIZE) &&		\
-	    (PyTuple_GET_ITEM(errobj, 1) == Py_None)) {
-		PyUFunc_USEDEFAULTS = 1;
+	PyUFunc_NUM_NODEFAULTS += 1;
+        res = PyUFunc_GetPyValues("test", &bufsize, &errmask, 
+				  &errobj);
+	PyUFunc_NUM_NODEFAULTS -= 1;
+	
+	if (res < 0) return -1;
+	
+	if ((errmask != UFUNC_ERR_DEFAULT) ||		\
+	    (bufsize != PyArray_BUFSIZE) ||		\
+	    (PyTuple_GET_ITEM(errobj, 1) != Py_None)) {
+		PyUFunc_NUM_NODEFAULTS += 1;
+	}
+	else if (PyUFunc_NUM_NODEFAULTS > 0) {
+		PyUFunc_NUM_NODEFAULTS -= 1;
 	}
 	return 0;
 }
-#endif

+#endif
 
 static PyObject *
 ufunc_seterr(PyObject *dummy, PyObject *args)
@@ -2823,21 +2819,9 @@ ufunc_seterr(PyObject *dummy, PyObject *args)
 	static char *msg = "Error object must be a list of length 3";
 	
 	if (!PyArg_ParseTuple(args, "O", &val)) return NULL;
-
-	if (!PyList_CheckExact(val)) {
-		PyObject *new;
-		new = PyObject_GetAttrString(val, "_val_obj");
-		if (new == NULL) {
-			PyErr_SetString(PyExc_ValueError, msg);
-			return NULL;
-		}
-		val = new;
-	}
-	else Py_INCREF(val);
 	
 	if (!PyList_CheckExact(val) || PyList_GET_SIZE(val) != 3) {
 		PyErr_SetString(PyExc_ValueError, msg);
-		Py_DECREF(val);
 		return NULL;
 	}
 	if (PyUFunc_PYVALS_NAME == NULL) {
@@ -2848,10 +2832,9 @@ ufunc_seterr(PyObject *dummy, PyObject *args)
 		thedict = PyEval_GetBuiltins();
 	}
 	res = PyDict_SetItem(thedict, PyUFunc_PYVALS_NAME, val);
-	Py_DECREF(val);
-	if (res < 0) return NULL;

-#if USE_USE_DEFAULTS
-	if (ufunc_update_use_defaults() < 0) return NULL;

+	if (res < 0) return NULL;
+#if USE_USE_DEFAULTS==1
+	if (ufunc_update_use_defaults() < 0) return NULL;
 #endif
 	Py_INCREF(Py_None);
 	return Py_None;
diff --git a/numpy/core/tests/test_numeric.py b/numpy/core/tests/test_numeric.py
index 45204e3bf..000879b96 100644
--- a/numpy/core/tests/test_numeric.py
+++ b/numpy/core/tests/test_numeric.py
@@ -134,31 +134,31 @@ class test_bool_scalar(ScipyTestCase):
         self.failUnless((t ^ t) is f)
         self.failUnless((f ^ t) is t)
         self.failUnless((t ^ f) is t)
-        self.failUnless((f ^ f) is f)

-

-
-class test_seterr(ScipyTestCase):

-    def test_set(self):

-        err = seterr()

-        old = seterr(divide='warn')

-        self.failUnless(err == old)

-        new = seterr()

-        self.failUnless(new['divide'] == 'warn')

-        seterr(over='raise')

-        self.failUnless(geterr()['over'] == 'raise')

-        self.failUnless(new['divide'] == 'warn')

-        seterr(**old)

-        self.failUnless(geterr() == old)

-    def test_divideerr(self):

-        seterr(divide='raise')

-        try:

-            array([1.]) / array([0.])

-        except FloatingPointError:

-            pass

-        else:

-            self.fail()

-        seterr(divide='ignore')

+        self.failUnless((f ^ f) is f)
+
+
+class test_seterr(ScipyTestCase):
+    def test_set(self):
+        err = seterr()
+        old = seterr(divide='warn')
+        self.failUnless(err == old)
+        new = seterr()
+        self.failUnless(new['divide'] == 'warn')
+        seterr(over='raise')
+        self.failUnless(geterr()['over'] == 'raise')
+        self.failUnless(new['divide'] == 'warn')
+        seterr(**old)
+        self.failUnless(geterr() == old)
+    def test_divideerr(self):
+        seterr(divide='raise')
+        try:
+            array([1.]) / array([0.])
+        except FloatingPointError:
+            pass
+        else:
+            self.fail()
+        seterr(divide='ignore')
         array([1.]) / array([0.])
 
-if __name__ == '__main__':

+if __name__ == '__main__':
     NumpyTest().run()
\ No newline at end of file