Merge pull request #14 from static-frame/2/uint-hash

Improved uint hashing

Author: flexatone

arraymap.c

@@ -145,11 +145,15 @@ char_get_end_p(char* p, Py_ssize_t dt_size) {
 }
 
 
+// This masks the input with INT64_MAX, which removes the MSB; we then cast to an int64; the range is now between 0 and INT64_MAX. We then use the MSB of the original value; if set, we negate the number, producing negative values for the upper half of the uint64 range. Note that we only need to check for hash -1 in this branch.
 static inline Py_hash_t
 uint_to_hash(npy_uint64 v) {
-    Py_hash_t hash = (Py_hash_t)(v >> 1); // half unsigned fits in signed
-    if (hash == -1) {
-        return -2;
+    Py_hash_t hash = (Py_hash_t)(v & INT64_MAX);
+    if (v >> 63) {
+        hash = -hash;
+        if (hash == -1) {
+            return -2;
+        }
     }
     return hash;
 }

doc/articles/npy-opt.py

@@ -309,9 +309,20 @@ def get_array(size: int) -> np.ndarray:
         return array
 
 
+class FFUInt32(FixtureFactory):
+    NAME = "uint32"
+    SORT = 3
+
+    @staticmethod
+    def get_array(size: int) -> np.ndarray:
+        array = np.arange(INT_START, INT_START + size, dtype=np.uint32)
+        array.flags.writeable = False
+        return array
+
+
 class FFFloat64(FixtureFactory):
     NAME = "float64"
-    SORT = 3
+    SORT = 4
 
     @staticmethod
     def get_array(size: int) -> np.ndarray:
@@ -322,7 +333,7 @@ def get_array(size: int) -> np.ndarray:
 
 class FFFloat32(FixtureFactory):
     NAME = "float32"
-    SORT = 4
+    SORT = 5
 
     @staticmethod
     def get_array(size: int) -> np.ndarray:
@@ -392,17 +403,18 @@ def get_versions() -> str:
 CLS_FF = (
     FFInt32,
     FFInt64,
+    FFUInt32,
     FFUInt64,
     FFFloat64,
     FFString,
-    FFString4x,
+    # FFString4x,
     FFBytes,
     # FFObject,
 )
 FF_ORDER = [f.NAME for f in sorted(CLS_FF, key=lambda ff: ff.SORT)]
 
 # -------------------------------------------------------------------------------
-NUMBER = 2
+NUMBER = 20
 
 from itertools import product