Merge pull request numpy#26348 from ngoldbaum/nogil-ufunc-caches

NOGIL: Make loop data cache and dispatch cache thread-safe in nogil build

Author: seberg

numpy/_core/src/common/npy_hashtable.c

@@ -29,6 +29,33 @@
 #define _NpyHASH_XXROTATE(x) ((x << 13) | (x >> 19))  /* Rotate left 13 bits */
 #endif
 
+#ifdef Py_GIL_DISABLED
+// TODO: replace with PyMutex when it is public
+#define LOCK_TABLE(tb)                                      \
+    if (!PyThread_acquire_lock(tb->mutex, NOWAIT_LOCK)) {   \
+        PyThread_acquire_lock(tb->mutex, WAIT_LOCK);        \
+    }
+#define UNLOCK_TABLE(tb) PyThread_release_lock(tb->mutex);
+#define INITIALIZE_LOCK(tb)                     \
+    tb->mutex = PyThread_allocate_lock();       \
+    if (tb->mutex == NULL) {                    \
+        PyErr_NoMemory();                       \
+        PyMem_Free(res);                        \
+        return NULL;                            \
+    }
+#define FREE_LOCK(tb)                           \
+    if (tb->mutex != NULL) {                    \
+        PyThread_free_lock(tb->mutex);          \
+    }
+#else
+// the GIL serializes access to the table so no need
+// for locking if it is enabled
+#define LOCK_TABLE(tb)
+#define UNLOCK_TABLE(tb)
+#define INITIALIZE_LOCK(tb)
+#define FREE_LOCK(tb)
+#endif
+
 /*
  * This hashing function is basically the Python tuple hash with the type
  * identity hash inlined. The tuple hash itself is a reduced version of xxHash.
@@ -100,6 +127,8 @@ PyArrayIdentityHash_New(int key_len)
     res->size = 4;  /* Start with a size of 4 */
     res->nelem = 0;
 
+    INITIALIZE_LOCK(res);
+
     res->buckets = PyMem_Calloc(4 * (key_len + 1), sizeof(PyObject *));
     if (res->buckets == NULL) {
         PyErr_NoMemory();
@@ -114,6 +143,7 @@ NPY_NO_EXPORT void
 PyArrayIdentityHash_Dealloc(PyArrayIdentityHash *tb)
 {
     PyMem_Free(tb->buckets);
+    FREE_LOCK(tb);
     PyMem_Free(tb);
 }
 
@@ -160,8 +190,9 @@ _resize_if_necessary(PyArrayIdentityHash *tb)
     for (npy_intp i = 0; i < prev_size; i++) {
         PyObject **item = &old_table[i * (tb->key_len + 1)];
         if (item[0] != NULL) {
-            tb->nelem -= 1;  /* Decrement, setitem will increment again */
-            PyArrayIdentityHash_SetItem(tb, item+1, item[0], 1);
+            PyObject **tb_item = find_item(tb, item + 1);
+            tb_item[0] = item[0];
+            memcpy(tb_item+1, item+1, tb->key_len * sizeof(PyObject *));
         }
     }
     PyMem_Free(old_table);
@@ -188,14 +219,17 @@ NPY_NO_EXPORT int
 PyArrayIdentityHash_SetItem(PyArrayIdentityHash *tb,
         PyObject *const *key, PyObject *value, int replace)
 {
+    LOCK_TABLE(tb);
     if (value != NULL && _resize_if_necessary(tb) < 0) {
         /* Shrink, only if a new value is added. */
+        UNLOCK_TABLE(tb);
         return -1;
     }
 
     PyObject **tb_item = find_item(tb, key);
     if (value != NULL) {
         if (tb_item[0] != NULL && !replace) {
+            UNLOCK_TABLE(tb);
             PyErr_SetString(PyExc_RuntimeError,
                     "Identity cache already includes the item.");
             return -1;
@@ -209,12 +243,16 @@ PyArrayIdentityHash_SetItem(PyArrayIdentityHash *tb,
         memset(tb_item, 0, (tb->key_len + 1) * sizeof(PyObject *));
     }
 
+    UNLOCK_TABLE(tb);
     return 0;
 }
 
 
 NPY_NO_EXPORT PyObject *
 PyArrayIdentityHash_GetItem(PyArrayIdentityHash const *tb, PyObject *const *key)
 {
-    return find_item(tb, key)[0];
+    LOCK_TABLE(tb);
+    PyObject *res = find_item(tb, key)[0];
+    UNLOCK_TABLE(tb);
+    return res;
 }

numpy/_core/src/common/npy_hashtable.h

@@ -13,6 +13,9 @@ typedef struct {
     PyObject **buckets;
     npy_intp size;  /* current size */
     npy_intp nelem;  /* number of elements */
+#ifdef Py_GIL_DISABLED
+    PyThread_type_lock *mutex;
+#endif
 } PyArrayIdentityHash;
 
 

numpy/_core/src/umath/legacy_array_method.c

@@ -33,37 +33,43 @@ typedef struct {
 
 
 /* Use a free list, since we should normally only need one at a time */
+#ifndef Py_GIL_DISABLED
 #define NPY_LOOP_DATA_CACHE_SIZE 5
 static int loop_data_num_cached = 0;
 static  legacy_array_method_auxdata *loop_data_cache[NPY_LOOP_DATA_CACHE_SIZE];
-
+#else
+#define NPY_LOOP_DATA_CACHE_SIZE 0
+#endif
 
 static void
 legacy_array_method_auxdata_free(NpyAuxData *data)
 {
+#if NPY_LOOP_DATA_CACHE_SIZE > 0
     if (loop_data_num_cached < NPY_LOOP_DATA_CACHE_SIZE) {
         loop_data_cache[loop_data_num_cached] = (
                 (legacy_array_method_auxdata *)data);
         loop_data_num_cached++;
     }
-    else {
+    else
+#endif
+    {
         PyMem_Free(data);
     }
 }
 
-#undef NPY_LOOP_DATA_CACHE_SIZE
-
-
 NpyAuxData *
 get_new_loop_data(
         PyUFuncGenericFunction loop, void *user_data, int pyerr_check)
 {
     legacy_array_method_auxdata *data;
+#if NPY_LOOP_DATA_CACHE_SIZE > 0
     if (NPY_LIKELY(loop_data_num_cached > 0)) {
         loop_data_num_cached--;
         data = loop_data_cache[loop_data_num_cached];
     }
-    else {
+    else
+#endif
+    {
         data = PyMem_Malloc(sizeof(legacy_array_method_auxdata));
         if (data == NULL) {
             return NULL;
@@ -77,6 +83,7 @@ get_new_loop_data(
     return (NpyAuxData *)data;
 }
 
+#undef NPY_LOOP_DATA_CACHE_SIZE
 
 /*
  * This is a thin wrapper around the legacy loop signature.

numpy/_core/tests/test_multithreading.py

@@ -9,13 +9,30 @@
     pytest.skip(allow_module_level=True, reason="no threading support in wasm")
 
 
-def test_parallel_errstate_creation():
+def run_threaded(func, iters, pass_count=False):
+    with concurrent.futures.ThreadPoolExecutor(max_workers=8) as tpe:
+        if pass_count:
+            futures = [tpe.submit(func, i) for i in range(iters)]
+        else:
+            futures = [tpe.submit(func) for _ in range(iters)]
+        for f in futures:
+            f.result()
+
+
+def test_parallel_randomstate_creation():
     # if the coercion cache is enabled and not thread-safe, creating
     # RandomState instances simultaneously leads to a data race
     def func(seed):
         np.random.RandomState(seed)
 
-    with concurrent.futures.ThreadPoolExecutor(max_workers=8) as tpe:
-        futures = [tpe.submit(func, i) for i in range(500)]
-        for f in futures:
-            f.result()
+    run_threaded(func, 500, pass_count=True)
+
+def test_parallel_ufunc_execution():
+    # if the loop data cache or dispatch cache are not thread-safe
+    # computing ufuncs simultaneously in multiple threads leads
+    # to a data race
+    def func():
+        arr = np.random.random((25,))
+        np.isnan(arr)
+
+    run_threaded(func, 500)