Perf plot

flopcounts/x86/variant_inlining_8_SoA_AVX512_polyexp_fastinvsqrt.py

@@ -0,0 +1,49 @@
+# const int n_p_pad = (n_p + 7) & ~7;
+
+# ADDs: n_iter * n_p * 33 +
+#        n_iter * n_p * (n_p_pad / 8) * 232 +
+#        n_iter * n_p * board.n_obstacles * 10
+
+# MULTs: n_iter * n_p * 40 +
+#        n_iter * n_p * (n_p_pad / 8) * 320 +
+#        n_iter * n_p * board.n_obstacles * 13
+
+# DIVs: n_iter * n_p * (n_p_pad / 8) * 8 +
+#        n_iter * n_p * board.n_obstacles
+
+# SQRTs: n_iter * n_p +
+#        n_iter * n_p * (n_p_pad / 8) * 24 +
+#        n_iter * n_p * board.n_obstacles
+
+# EXPs: n_iter * n_p * board.n_obstacles
+
+
+def adds(n_iter: int, n_p: int, n_obs: int = 2) -> int:
+  n_p_pad = (n_p + 7) & ~7
+  return int(n_iter * n_p * (33 + (n_p_pad // 8) * 232 + n_obs * 10))
+
+
+def mults(n_iter: int, n_p: int, n_obs: int = 2) -> int:
+  n_p_pad = (n_p + 7) & ~7
+  return int(n_iter * n_p * (40 + (n_p_pad // 8) * 320 + n_obs * 13))
+
+
+def divs(n_iter: int, n_p: int, n_obs: int = 2) -> int:
+  n_p_pad = (n_p + 7) & ~7
+  return int(n_iter * n_p * ((n_p_pad // 8) * 8 + n_obs))
+
+
+def sqrts(n_iter: int, n_p: int, n_obs: int = 2) -> int:
+  n_p_pad = (n_p + 7) & ~7
+  return int(n_iter * n_p * (1 + (n_p_pad // 8) * 24 + n_obs))
+
+
+def exps(n_iter: int, n_p: int, n_obs: int = 2) -> int:
+  n_p_pad = (n_p + 7) & ~7
+  return int(n_iter * n_p * n_obs)
+
+
+def total(n_iter: int, n_p: int, n_obs: int = 2) -> int:
+  return adds(n_iter, n_p, n_obs) + mults(n_iter, n_p, n_obs) + divs(
+    n_iter, n_p, n_obs) + sqrts(n_iter, n_p,
+                                n_obs) + exps(n_iter, n_p, n_obs) * 30