MNT: move unpack_bits LUT into global static struct

Author: ngoldbaum

numpy/_core/src/multiarray/compiled_base.c

@@ -1747,15 +1747,6 @@ pack_bits(PyObject *input, int axis, char order)
 static PyObject *
 unpack_bits(PyObject *input, int axis, PyObject *count_obj, char order)
 {
-    static int unpack_init = 0;
-    /*
-     * lookuptable for bitorder big as it has been around longer
-     * bitorder little is handled via byteswapping in the loop
-     */
-    static union {
-        npy_uint8  bytes[8];
-        npy_uint64 uint64;
-    } unpack_lookup_big[256];
     PyArrayObject *inp;
     PyArrayObject *new = NULL;
     PyArrayObject *out = NULL;
@@ -1841,22 +1832,6 @@ unpack_bits(PyObject *input, int axis, PyObject *count_obj, char order)
         goto fail;
     }
 
-    /*
-     * setup lookup table under GIL, 256 8 byte blocks representing 8 bits
-     * expanded to 1/0 bytes
-     */
-    if (unpack_init == 0) {
-        npy_intp j;
-        for (j=0; j < 256; j++) {
-            npy_intp k;
-            for (k=0; k < 8; k++) {
-                npy_uint8 v = (j & (1 << k)) == (1 << k);
-                unpack_lookup_big[j].bytes[7 - k] = v;
-            }
-        }
-        unpack_init = 1;
-    }
-
     count = PyArray_DIM(new, axis) * 8;
     if (outdims[axis] > count) {
         in_n = count / 8;
@@ -1883,15 +1858,15 @@ unpack_bits(PyObject *input, int axis, PyObject *count_obj, char order)
             /* for unity stride we can just copy out of the lookup table */
             if (order == 'b') {
                 for (index = 0; index < in_n; index++) {
-                    npy_uint64 v = unpack_lookup_big[*inptr].uint64;
+                    npy_uint64 v = npy_ma_global_data->unpack_lookup_big[*inptr].uint64;
                     memcpy(outptr, &v, 8);
                     outptr += 8;
                     inptr += in_stride;
                 }
             }
             else {
                 for (index = 0; index < in_n; index++) {
-                    npy_uint64 v = unpack_lookup_big[*inptr].uint64;
+                    npy_uint64 v = npy_ma_global_data->unpack_lookup_big[*inptr].uint64;
                     if (order != 'b') {
                         v = npy_bswap8(v);
                     }
@@ -1902,7 +1877,7 @@ unpack_bits(PyObject *input, int axis, PyObject *count_obj, char order)
             }
             /* Clean up the tail portion */
             if (in_tail) {
-                npy_uint64 v = unpack_lookup_big[*inptr].uint64;
+                npy_uint64 v = npy_ma_global_data->unpack_lookup_big[*inptr].uint64;
                 if (order != 'b') {
                     v = npy_bswap8(v);
                 }

numpy/_core/src/multiarray/multiarraymodule.c

@@ -4990,6 +4990,24 @@ initialize_static_globals(void)
     }
     npy_ma_global_data->optimize = PyLong_AsLong(level);
     Py_DECREF(level);
+
+    /*
+     * see unpack_bits for how this table is used.
+     *
+     * LUT for bigendian bitorder, littleendian is handled via
+     * byteswapping in the loop.
+     *
+     * 256 8 byte blocks representing 8 bits expanded to 1 or 0 bytes
+     */
+    npy_intp j;
+    for (j=0; j < 256; j++) {
+        npy_intp k;
+        for (k=0; k < 8; k++) {
+            npy_uint8 v = (j & (1 << k)) == (1 << k);
+            npy_ma_global_data->unpack_lookup_big[j].bytes[7 - k] = v;
+        }
+    }
+
     return 0;
 }
 

numpy/_core/src/multiarray/multiarraymodule.h

@@ -74,6 +74,14 @@ typedef struct npy_ma_global_data_struct {
      */
     long optimize;
 
+    /*
+     * LUT used by unpack_bits
+     */
+    union {
+        npy_uint8  bytes[8];
+        npy_uint64 uint64;
+    } unpack_lookup_big[256];
+
     /*
      * Used for CPU feature detection and dispatch
      *