#!/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."