Intermediate state

spacy/tokens/doc.pxd

@@ -9,7 +9,10 @@ from ..attrs cimport attr_id_t
 
 cdef attr_t get_token_attr(const TokenC* token, attr_id_t feat_name) nogil
 cdef attr_t get_token_attr_for_matcher(const TokenC* token, attr_id_t feat_name) nogil
-
+cdef const unsigned char[:] get_utf16_memoryview(str unicode_string, bint check_2_bytes)
+cdef bint check_utf16_char(unsigned short utf16_char, const unsigned char[:] scs)
+cdef void set_scs(const unsigned char[:] searched_string, const unsigned char[:] scs, bytearray working_array, bint suffs_not_prefs)
+    
 
 ctypedef const LexemeC* const_Lexeme_ptr
 ctypedef const TokenC* const_TokenC_ptr
@@ -69,4 +72,3 @@ cdef class Doc:
     cdef int push_back(self, LexemeOrToken lex_or_tok, bint has_space) except -1
 
     cpdef np.ndarray to_array(self, object features)
-

spacy/tokens/doc.pyx

@@ -16,7 +16,7 @@ import srsly
 from thinc.api import get_array_module, get_current_ops
 from thinc.util import copy_array
 import warnings
-import murmurhash.mrmr
+from murmurhash.mrmr cimport hash32
 
 from .span cimport Span
 from .token cimport MISSING_DEP
@@ -94,6 +94,50 @@ cdef attr_t get_token_attr_for_matcher(const TokenC* token, attr_id_t feat_name)
     else:
         return get_token_attr(token, feat_name)
 
+    
+cdef const unsigned char[:] get_utf16_memoryview(str unicode_string, bint check_2_bytes):
+    """
+    Returns a memory view of the UTF-16 representation of a string with the default endianness of the platform.
+    Throws a ValueError if *check_2_bytes == True* and one or more characters in the UTF-16 representation
+    occupy four bytes rather than two.
+    """
+    cdef const unsigned char[:] view = unicode_string.encode("UTF-16")[2:] # first two bytes are endianness
+    cdef unsigned int us_len = len(unicode_string), view_len = len(view)
+    if check_2_bytes and us_len * 2 != view_len:
+        raise ValueError(Errors.E1046)
+    return view
+
+cdef bint check_utf16_char(unsigned short utf16_char, const unsigned char[:] scs):
+    cdef unsigned int scs_idx = 0, scs_len = len(scs)
+    while scs_idx < scs_len:
+        if utf16_char == (<unsigned short*> &scs[scs_idx])[0]:
+            return True
+        scs_idx += 2
+    return False
+
+cdef void set_scs(const unsigned char[:] searched_string, const unsigned char[:] scs, bytearray working_array, bint suffs_not_prefs):
+    cdef unsigned int wa_idx = 0, wa_len = len(working_array), ss_len = len(searched_string)
+    cdef unsigned int ss_idx = ss_len - 2 if suffs_not_prefs else 0
+    cdef unsigned short working_utf16_char, SPACE = 32
+
+    while wa_idx < wa_len:
+        working_utf16_char = (<unsigned short*> &searched_string[ss_idx])[0]
+        if check_utf16_char(working_utf16_char, scs):
+            working_array[wa_idx] = working_utf16_char
+            wa_idx += 2
+        if suffs_not_prefs:
+            if ss_idx == 0:
+                break
+            ss_idx -= 2
+        else:
+            ss_idx += 2
+            if ss_idx == ss_len:
+                break
+    
+    while wa_idx < wa_len:
+        working_array[wa_idx] = SPACE
+        wa_idx += 2
+
 
 class SetEntsDefault(str, Enum):
     blocked = "blocked"
@@ -1741,23 +1785,33 @@ cdef class Doc:
         TODO
         """
         cdef unsigned int token_index, norm_hash_index, spec_hash_index
-        cdef str token_string, specials_string, working_substring
+        cdef const attr_t* num_token_attr
+        cdef const unsigned char[:] token_string, working_substring, scs = get_utf16_memoryview(special_chars, True)
         cdef unsigned int num_norm_hashes = len_end - len_start, num_spec_hashes = sc_len_end - sc_len_start, num_tokens = len(self)
+        cdef unsigned int len_token_string
         cdef np.ndarray[np.int64_t, ndim=2] output = numpy.empty((num_tokens, num_norm_hashes + num_spec_hashes), dtype="int64")
+        cdef bytearray working_scs_buffer = bytearray(sc_len_end * 2)
+        cdef unsigned int working_start, working_len
 
         for token_index in range(num_tokens):
-            token_string = self[token_index].orth_ if case_sensitive else self[token_index].lower_
-            if suffs_not_prefs:
-                token_string = token_string[::-1]
+            num_token_attr = &self.c[token_index].lex.orth if case_sensitive else &self.c[token_index].lex.lower
+            token_string = get_utf16_memoryview(self.vocab.strings[num_token_attr[0]], False)
+            if not suffs_not_prefs:
+                len_token_string = len(token_string)
 
             for norm_hash_index in range(num_norm_hashes):
-                working_substring = token_string[: len_start + norm_hash_index]
-                output[token_index, norm_hash_index] = murmurhash.mrmr.hash(working_substring)
+                if suffs_not_prefs:
+                    working_start = 0
+                    working_len = (len_start + norm_hash_index) * 2
+                else:
+                    working_len = (len_start + norm_hash_index) * 2
+                    working_start = len_token_string - working_len
+                output[token_index, norm_hash_index] = hash32(<char*> &token_string[working_start], working_len, 0)
 
-            specials_string = "".join([c for c in token_string if c in special_chars]) + " " * sc_len_end
+            set_scs(token_string, scs, working_scs_buffer, suffs_not_prefs)
             for spec_hash_index in range(num_spec_hashes):
-                working_substring = specials_string[: sc_len_start + spec_hash_index]
-                output[token_index, num_norm_hashes + spec_hash_index] = murmurhash.mrmr.hash(working_substring)
+                working_len = (sc_len_start + spec_hash_index) * 2
+                output[token_index, num_norm_hashes + spec_hash_index] = hash32(<char*> &working_scs_buffer[0], working_len, 0)
 
         return output.astype("uint64", casting="unsafe", copy=False)