REF: Split random-kit core rng into two functions to impreve performance

Splitting the function into the small commonly used part and the rarely
used larger part (typically 1 in 624 calls) improves inline performance

Author: Sheppard, Kevin

randomstate/src/random-kit/random-kit.c

@@ -84,5 +84,22 @@ extern void init_by_array(rk_state *state, unsigned long init_key[], int key_len
     mt[0] = 0x80000000UL; /* MSB is 1; assuring non-zero initial array */
 }
 
+void rk_gen(rk_state *state)
+{
+        uint32_t y;
+        int i;
 
+        for (i = 0; i < N - M; i++) {
+            y = (state->key[i] & UPPER_MASK) | (state->key[i+1] & LOWER_MASK);
+            state->key[i] = state->key[i+M] ^ (y>>1) ^ (-(y & 1) & MATRIX_A);
+        }
+        for (; i < N - 1; i++) {
+            y = (state->key[i] & UPPER_MASK) | (state->key[i+1] & LOWER_MASK);
+            state->key[i] = state->key[i+(M-N)] ^ (y>>1) ^ (-(y & 1) & MATRIX_A);
+        }
+        y = (state->key[N - 1] & UPPER_MASK) | (state->key[0] & LOWER_MASK);
+        state->key[N - 1] = state->key[M - 1] ^ (y >> 1) ^ (-(y & 1) & MATRIX_A);
+
+        state->pos = 0;
+}
 

randomstate/src/random-kit/random-kit.h

@@ -27,26 +27,16 @@ rk_state;
 
 extern void rk_seed(rk_state *state, uint32_t seed);
 
+extern void rk_gen(rk_state *state);
+
 /* Slightly optimized reference implementation of the Mersenne Twister */
 inline uint32_t rk_random(rk_state *state)
 {
     uint32_t y;
 
     if (state->pos == RK_STATE_LEN) {
-        int i;
-
-        for (i = 0; i < N - M; i++) {
-            y = (state->key[i] & UPPER_MASK) | (state->key[i+1] & LOWER_MASK);
-            state->key[i] = state->key[i+M] ^ (y>>1) ^ (-(y & 1) & MATRIX_A);
-        }
-        for (; i < N - 1; i++) {
-            y = (state->key[i] & UPPER_MASK) | (state->key[i+1] & LOWER_MASK);
-            state->key[i] = state->key[i+(M-N)] ^ (y>>1) ^ (-(y & 1) & MATRIX_A);
-        }
-        y = (state->key[N - 1] & UPPER_MASK) | (state->key[0] & LOWER_MASK);
-        state->key[N - 1] = state->key[M - 1] ^ (y >> 1) ^ (-(y & 1) & MATRIX_A);
-
-        state->pos = 0;
+        // Move to function to help inlining
+        rk_gen(state);
     }
     y = state->key[state->pos++];
 

setup.py

@@ -161,5 +161,5 @@ def write_config(file_name, config):
 # Clean up generated files
 for config in configs:
     os.unlink(join(mod_dir, config['file_name'] + '.pyx'))
-   #  os.unlink(join(mod_dir, config['file_name'] + '-config.pxi'))
-   #  os.unlink(join(mod_dir, config['file_name'] + '.c'))
+    os.unlink(join(mod_dir, config['file_name'] + '-config.pxi'))
+    os.unlink(join(mod_dir, config['file_name'] + '.c'))