additional plot, and short script to do a nonvectorized benchmark (like they...

analysis/plots.py

@@ -59,7 +59,7 @@ plt.savefig("analysis/fig_tracks.png", dpi=170, bbox_inches="tight")
 bars = [
  ("baseline", "variant_baseline"),
  ("AVX2 (double)", "variant_inlining_7_SoA_AVX2"),
- ("AVX-512 (double)", "variant_avx512_1"),
+ ("AVX-512 (double)", "variant_AVX512_1"),
  ("+ polyexp", "variant_inlining_8_SoA_AVX512_polyexp_moreavx"),
  ("+ float", "variant_float_inlining_SoA_AVX512_polyexp_moreavx"),
  ("+ rsqrt & ILP", "variant_float_inlining_SoA_AVX512_polyexp_moreavx_ilp_2_rsqrt_morersqrt"),
@@ -101,7 +101,7 @@ for pk,t in [(SCALAR,"scalar (4)"),(PEAK_D,"double (16)"),(PEAK_F,"float (32)")]
 I5000 = df[(df.variant=="variant_baseline") & (df.n==5000)]["I"].values[0]
 pts = [
  ("variant_baseline", "#7a7a7a"),
- ("variant_avx512_1", "#2f9e8f"),
+ ("variant_AVX512_1", "#2f9e8f"),
  ("variant_inlining_8_SoA_AVX512_polyexp_moreavx", "#e08a1e"),
  ("variant_float_inlining_SoA_AVX512_polyexp_moreavx", "#c0436a"),
  ("variant_float_inlining_SoA_AVX512_polyexp_moreavx_ilp_2_rsqrt_morersqrt", maroon),
@@ -140,5 +140,46 @@ ax.legend(frameon=False, fontsize=10)
 plt.tight_layout()
 plt.savefig("analysis/fig_exp.png", dpi=170, bbox_inches="tight")
 
+# plot 5: the milestone ladder as lines (this is plot0 + line1, line2, ... for the doc).
+# one line per relevant optimization step across n, baseline at the bottom.
+# unlike fig_tracks this includes the scalar steps so you can see where each jump comes from.
+milestones = [
+ ("variant_baseline",                                                       "baseline",                       "#7a7a7a", "-"),
+ ("variant_cutoff_callsite",                                                "+ cutoff (best scalar)",         "#9e8f2f", "-"),
+ ("variant_inlining_7_SoA_AVX2",                                            "AVX2 (double)",                  "#3b6fb0", "--"),
+ ("variant_inlining_7_SoA_AVX2_polyexp",                                    "+ polyexp (double)",             "#2f9e8f", "--"),
+ ("variant_inlining_8_SoA_AVX512_polyexp_moreavx",                          "AVX-512 (double)",               "#2f5d9e", "-"),
+ ("variant_float_inlining_SoA_AVX2_polyexp",                                "float AVX2 + polyexp",           "#e58aa6", "--"),
+ ("variant_float_inlining_SoA_AVX2_polyexp_rsqrt_loopopt",                  "float AVX2 + rsqrt",             "#e08a1e", "--"),
+ ("variant_float_inlining_SoA_AVX512_polyexp_moreavx",                      "float AVX-512",                  "#c0436a", "-"),
+ ("variant_float_inlining_SoA_AVX512_polyexp_moreavx_ilp_2_rsqrt_morersqrt","float AVX-512 + rsqrt + ILP (best)", maroon, "-"),
+]
+
+fig, ax = plt.subplots(figsize=(9.5,5.2))
+for v,lab,c,ls in milestones:
+    x,y = getline(v)
+    if len(x) == 0:        # variant not in results.csv -> skip instead of crashing
+        print("skip (not in csv):", v)
+        continue
+    ax.plot(x, y, ls, color=c, marker="o", ms=4, lw=2.4 if c==maroon else 1.7, label=lab)
+
+for yv,t in [(PEAK_D,"double peak (16)"), (PEAK_F,"float peak (32)")]:
+    ax.axhline(yv, ls=":", color="#999", lw=1.2)
+    ax.text(52, yv*1.03, t, fontsize=9.5, color="#888")
+
+ax.set_xscale("log"); ax.set_yscale("log")
+ax.set_xticks([50,100,200,500,1000,2000,5000,10000])
+ax.set_xticklabels(["50","100","200","500","1k","2k","5k","10k"])
+ax.set_yticks([1,2,4,8,16,32]); ax.set_yticklabels(["1","2","4","8","16","32"])
+ax.set_xlim(46,13000); ax.set_ylim(0.8,40)
+ax.set_xlabel("number of pedestrians  n")
+ax.set_ylabel("performance [flops/cycle]")
+ax.grid(color="#e9e9e9")
+ax.spines["top"].set_visible(False); ax.spines["right"].set_visible(False)
+ax.legend(fontsize=8.5, loc="upper left", ncol=2, framealpha=.95)
+plt.tight_layout()
+plt.savefig("analysis/fig_milestones.png", dpi=170, bbox_inches="tight")
+
+
 #plt.show()
 print("done")

scripts/test_novec.sh

@@ -0,0 +1,67 @@
+#!/bin/bash
+# milestone 10: find the best NON-vectorized scalar variant.
+# we compile each scalar (double, no-intrinsics) variant with vectorization
+# turned off (-fno-tree-vectorize -fno-tree-slp-vectorize) so -march=native
+# can't sneak SIMD in, then time them on a small input.
+#
+# small on purpose: N=500, 200 iters -> each build+run is a couple seconds.
+# compare T(n) [cycles] to pick the fastest. flops/cycle is printed too but
+# don't rank by it: the symmetry variants do half the work, so their
+# flops/cycle looks low even when the runtime is good.
+#
+# run from the repo root:  bash scripts/test_novec.sh
+
+set -e
+
+CXX="${CXX:-$(command -v g++ || command -v clang++)}"
+N=500
+ITER=200
+NOVEC="-fno-tree-vectorize -fno-tree-slp-vectorize"
+FLAGS="-O3 -march=native -std=c++23 -flto $NOVEC -DBENCHMARK=1 -DFLOPCOUNT=0 -DRANDOM_SEED=0 -DROOFLINE=0"
+
+# variant:main_file . symmetry variants need main-symmetry.cpp (halved flop formula),
+# everything else uses main.cpp. only double, no-AVX variants here.
+CANDIDATES=(
+  "baseline:main.cpp"
+  "inlining_7:main.cpp"
+  "inlining_8:main.cpp"
+  "cutoff:main.cpp"
+  "cutoff_callsite:main.cpp"
+  "branchless:main.cpp"
+  "branchless_sqrtfree:main.cpp"
+  "lut_exp:main.cpp"
+  "inlining_7_SoA:main.cpp"
+  "inlining_7_symmetry_2:main-symmetry.cpp"
+  "inlining_7_symmetry_2_ilp_2:main-symmetry.cpp"
+)
+
+echo "compiler: $CXX"
+echo "flags:    $FLAGS"
+echo "input:    N=$N ITER=$ITER"
+echo ""
+printf "%-34s %18s %14s\n" "variant" "cycles" "flops/cycle"
+printf "%-34s %18s %14s\n" "-------" "------" "-----------"
+
+for entry in "${CANDIDATES[@]}"; do
+  v="${entry%%:*}"
+  main="${entry##*:}"
+  src="variants/variant_${v}.cpp"
+
+  if [ ! -f "$src" ]; then
+    printf "%-34s %18s\n" "$v" "(no source, skipped)"
+    continue
+  fi
+
+  bin="novec_${v}"
+  $CXX "$main" "$src" -o "$bin" $FLAGS 2>/dev/null
+
+  out="$(./"$bin" "$N" "$ITER")"
+  cyc="$(echo "$out" | grep 'T(n):' | grep -oE '[0-9.]+')"
+  fpc="$(echo "$out" | grep 'P(n):' | grep -oE '[0-9.]+')"
+
+  printf "%-34s %18s %14s\n" "$v" "$cyc" "$fpc"
+  rm -f "$bin"
+done
+
+echo ""
+echo "lowest cycles = best non-vectorized variant. report that number for milestone 10."