|
| 1 | +#!/bin/bash |
| 2 | +# Per-phase timing profile across program sizes. |
| 3 | +# Shows how each proving phase and sub-operation scales with program size. |
| 4 | +# |
| 5 | +# Usage: bench_timing_profile.sh [--no-build] [--programs "1M 2M 4M 8M"] |
| 6 | +# |
| 7 | +# Requires: instruments feature. |
| 8 | +# Timing is deterministic with instruments, so 1 run per size is enough. |
| 9 | + |
| 10 | +set -euo pipefail |
| 11 | + |
| 12 | +SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" |
| 13 | +ROOT_DIR="$(cd "$SCRIPT_DIR/.." && pwd)" |
| 14 | +TMP_DIR="/tmp/bench_timing_profile" |
| 15 | +ELF_DIR="$ROOT_DIR/executor/program_artifacts/asm" |
| 16 | + |
| 17 | +GREEN='\033[0;32m' |
| 18 | +BOLD='\033[1m' |
| 19 | +NC='\033[0m' |
| 20 | + |
| 21 | +BUILD=true |
| 22 | +PROGRAMS="500k 1M 2M 4M" |
| 23 | + |
| 24 | +while [[ $# -gt 0 ]]; do |
| 25 | + case $1 in |
| 26 | + --no-build) BUILD=false; shift ;; |
| 27 | + --programs) PROGRAMS="$2"; shift 2 ;; |
| 28 | + *) echo "Unknown arg: $1"; exit 1 ;; |
| 29 | + esac |
| 30 | +done |
| 31 | + |
| 32 | +suffix_to_steps() { |
| 33 | + case $1 in |
| 34 | + 160k) echo 160000 ;; 250k) echo 250000 ;; 372k) echo 372000 ;; |
| 35 | + 500k) echo 500000 ;; 1M) echo 1000000 ;; 1200k) echo 1200000 ;; |
| 36 | + 2M) echo 2000000 ;; 4M) echo 4000000 ;; 8M) echo 8000000 ;; |
| 37 | + 16M) echo 16000000 ;; 32M) echo 32000000 ;; 64M) echo 64000000 ;; 128M) echo 128000000 ;; |
| 38 | + *) echo "Unknown: $1" >&2; exit 1 ;; |
| 39 | + esac |
| 40 | +} |
| 41 | + |
| 42 | +rm -rf "$TMP_DIR" && mkdir -p "$TMP_DIR" |
| 43 | + |
| 44 | +if $BUILD; then |
| 45 | + echo -e "${GREEN}Building CLI with instruments...${NC}" |
| 46 | + cargo build --release -p cli --features instruments \ |
| 47 | + --manifest-path "$ROOT_DIR/Cargo.toml" 2>&1 | tail -1 |
| 48 | +fi |
| 49 | +CLI="$ROOT_DIR/target/release/cli" |
| 50 | + |
| 51 | +# Parse instruments stderr into key=value pairs. |
| 52 | +# Uses occurrence counting to disambiguate expand_pool_to_lde and commit (Merkle). |
| 53 | +parse_timing() { |
| 54 | + awk ' |
| 55 | + BEGIN { lde_n = 0; merkle_n = 0 } |
| 56 | + function secs( s) { |
| 57 | + if (match($0, /[0-9]+\.[0-9]+s/)) { |
| 58 | + s = substr($0, RSTART, RLENGTH - 1) |
| 59 | + return s |
| 60 | + } |
| 61 | + return "" |
| 62 | + } |
| 63 | + /^ Execute / { v = secs(); if (v) print "execute=" v } |
| 64 | + /^ Trace build/ { v = secs(); if (v) print "trace_build=" v } |
| 65 | + /^ AIR construction/ { v = secs(); if (v) print "air=" v } |
| 66 | + /^ Pre-pass/ { v = secs(); if (v) print "prepass=" v } |
| 67 | + /^ Round 1 / { v = secs(); if (v) print "round1=" v } |
| 68 | + /Main trace commits/ { v = secs(); if (v) print "main_commits=" v } |
| 69 | + /Aux trace build/ { v = secs(); if (v) print "aux_build=" v } |
| 70 | + /Aux trace commit/ { v = secs(); if (v) print "aux_commit=" v } |
| 71 | + /Rounds 2/ { v = secs(); if (v) print "rounds24=" v } |
| 72 | + /expand_pool_to_lde/ && !/R1 expand/ { |
| 73 | + lde_n++; v = secs() |
| 74 | + if (v && lde_n == 1) print "main_lde=" v |
| 75 | + if (v && lde_n == 2) print "aux_lde=" v |
| 76 | + } |
| 77 | + /commit \(Merkle\)/ { |
| 78 | + merkle_n++; v = secs() |
| 79 | + if (v && merkle_n == 1) print "main_merkle=" v |
| 80 | + if (v && merkle_n == 2) print "aux_merkle=" v |
| 81 | + } |
| 82 | + /R1 expand_pool_to_lde/ { v = secs(); if (v) print "r1_lde=" v } |
| 83 | + /R2 evaluate/ { v = secs(); if (v) print "r2_evaluate=" v } |
| 84 | + /R2 decompose/ { v = secs(); if (v) print "r2_decompose=" v } |
| 85 | + /R2 commit_comp/ { v = secs(); if (v) print "r2_commit_comp=" v } |
| 86 | + /R3 OOD/ { v = secs(); if (v) print "r3_ood=" v } |
| 87 | + /R4 deep_comp/ { v = secs(); if (v) print "r4_deep_comp=" v } |
| 88 | + /R4 interpolate/ { v = secs(); if (v) print "r4_interp=" v } |
| 89 | + /R4 fri::commit/ { v = secs(); if (v) print "r4_fri_commit=" v } |
| 90 | + /R4 queries/ { v = secs(); if (v) print "r4_queries=" v } |
| 91 | + /^ Total FFT/ { v = secs(); if (v) print "total_fft=" v } |
| 92 | + /^ Total Merkle/ { v = secs(); if (v) print "total_merkle=" v } |
| 93 | + /^ TOTAL / { v = secs(); if (v) print "total=" v } |
| 94 | + ' "$1" |
| 95 | +} |
| 96 | + |
| 97 | +for size in $PROGRAMS; do |
| 98 | + ELF="$ELF_DIR/fib_iterative_${size}.elf" |
| 99 | + [ -f "$ELF" ] || { echo "Missing: $ELF"; continue; } |
| 100 | + steps=$(suffix_to_steps "$size") |
| 101 | + echo -e "${GREEN}Running fib_iterative_${size}...${NC}" |
| 102 | + |
| 103 | + STDERR="$TMP_DIR/${size}_stderr.txt" |
| 104 | + "$CLI" prove "$ELF" -o "$TMP_DIR/proof.bin" 2>"$STDERR" >/dev/null |
| 105 | + rm -f "$TMP_DIR/proof.bin" |
| 106 | + |
| 107 | + echo "steps=$steps" > "$TMP_DIR/${size}_data.txt" |
| 108 | + parse_timing "$STDERR" >> "$TMP_DIR/${size}_data.txt" |
| 109 | +done |
| 110 | + |
| 111 | +# --- Display --------------------------------------------------------------- |
| 112 | + |
| 113 | +get_val() { |
| 114 | + grep "^${2}=" "$1" 2>/dev/null | cut -d= -f2 |
| 115 | +} |
| 116 | + |
| 117 | +print_section() { |
| 118 | + local title=$1; shift |
| 119 | + |
| 120 | + echo "" |
| 121 | + echo -e " ${BOLD}${title}${NC}" |
| 122 | + printf " %-30s" "" |
| 123 | + for size in $PROGRAMS; do printf " %9s" "$size"; done |
| 124 | + echo "" |
| 125 | + printf " %-30s" "" |
| 126 | + for size in $PROGRAMS; do printf " %9s" "─────────"; done |
| 127 | + echo "" |
| 128 | + |
| 129 | + for spec in "$@"; do |
| 130 | + local label="${spec%%:*}" |
| 131 | + local key="${spec#*:}" |
| 132 | + printf " %-30s" "$label" |
| 133 | + for size in $PROGRAMS; do |
| 134 | + local DATA="$TMP_DIR/${size}_data.txt" |
| 135 | + local val |
| 136 | + val=$(get_val "$DATA" "$key") |
| 137 | + if [ -n "$val" ]; then |
| 138 | + printf " %8.2fs" "$val" |
| 139 | + else |
| 140 | + printf " %9s" "-" |
| 141 | + fi |
| 142 | + done |
| 143 | + echo "" |
| 144 | + done |
| 145 | +} |
| 146 | + |
| 147 | +echo "" |
| 148 | +echo -e "${BOLD}=== TIMING PROFILE ACROSS SIZES ===${NC}" |
| 149 | + |
| 150 | +TABLE_K="${TABLE_PARALLELISM:-default (cores/3)}" |
| 151 | +echo -e " TABLE_PARALLELISM=$TABLE_K" |
| 152 | + |
| 153 | +print_section "Phase (wall time)" \ |
| 154 | + "TOTAL:total" \ |
| 155 | + "Execute:execute" \ |
| 156 | + "Trace build:trace_build" \ |
| 157 | + "AIR construction:air" \ |
| 158 | + "Pre-pass:prepass" \ |
| 159 | + "Round 1:round1" \ |
| 160 | + "Rounds 2-4:rounds24" |
| 161 | + |
| 162 | +print_section "Round 1 breakdown (CPU sum)" \ |
| 163 | + "Main LDE:main_lde" \ |
| 164 | + "Main Merkle:main_merkle" \ |
| 165 | + "Aux build (wall):aux_build" \ |
| 166 | + "Aux LDE:aux_lde" \ |
| 167 | + "Aux Merkle:aux_merkle" |
| 168 | + |
| 169 | +print_section "Rounds 2-4 sub-ops (CPU sum)" \ |
| 170 | + "R1 LDE (reconstruct):r1_lde" \ |
| 171 | + "R2 decompose+extend:r2_decompose" \ |
| 172 | + "R2 evaluate:r2_evaluate" \ |
| 173 | + "R2 commit comp poly:r2_commit_comp" \ |
| 174 | + "R3 OOD evaluation:r3_ood" \ |
| 175 | + "R4 deep comp evals:r4_deep_comp" \ |
| 176 | + "R4 interpolate+evaluate:r4_interp" \ |
| 177 | + "R4 FRI commit:r4_fri_commit" \ |
| 178 | + "R4 queries & openings:r4_queries" |
| 179 | + |
| 180 | +print_section "Cross-round totals (CPU sum)" \ |
| 181 | + "Total FFT:total_fft" \ |
| 182 | + "Total Merkle:total_merkle" |
| 183 | + |
| 184 | +# --- Growth rate ----------------------------------------------------------- |
| 185 | + |
| 186 | +echo "" |
| 187 | +echo -e "${BOLD}=== GROWTH RATE (per 1M steps) ===${NC}" |
| 188 | +echo "" |
| 189 | + |
| 190 | +for spec in \ |
| 191 | + "TOTAL:total" \ |
| 192 | + "Trace build:trace_build" \ |
| 193 | + "Round 1:round1" \ |
| 194 | + "Rounds 2-4:rounds24" \ |
| 195 | + "Main LDE:main_lde" \ |
| 196 | + "Aux LDE:aux_lde" \ |
| 197 | + "R1 LDE (reconstruct):r1_lde" \ |
| 198 | + "R2 decompose+extend:r2_decompose" \ |
| 199 | + "R3 OOD evaluation:r3_ood" \ |
| 200 | + "Total FFT:total_fft" \ |
| 201 | + "Total Merkle:total_merkle" |
| 202 | +do |
| 203 | + label="${spec%%:*}" |
| 204 | + key="${spec#*:}" |
| 205 | + |
| 206 | + PAIRS="" |
| 207 | + for size in $PROGRAMS; do |
| 208 | + DATA="$TMP_DIR/${size}_data.txt" |
| 209 | + [ -f "$DATA" ] || continue |
| 210 | + steps=$(get_val "$DATA" "steps") |
| 211 | + val=$(get_val "$DATA" "$key") |
| 212 | + [ -z "$val" ] && continue |
| 213 | + steps_m=$(awk "BEGIN {printf \"%.2f\", $steps / 1000000}") |
| 214 | + PAIRS="$PAIRS $steps_m $val" |
| 215 | + done |
| 216 | + |
| 217 | + echo "$PAIRS" | awk -v label="$label" '{ |
| 218 | + n = NF / 2 |
| 219 | + if (n < 2) { printf " %-30s (insufficient data)\n", label; next } |
| 220 | + for (i = 0; i < n; i++) { |
| 221 | + x[i] = $(2*i+1); y[i] = $(2*i+2) |
| 222 | + } |
| 223 | + sx=0; sy=0; sxx=0; sxy=0 |
| 224 | + for (i=0;i<n;i++) { sx+=x[i]; sy+=y[i]; sxx+=x[i]*x[i]; sxy+=x[i]*y[i] } |
| 225 | + denom = n*sxx - sx*sx |
| 226 | + if (denom == 0) { printf " %-30s (constant)\n", label; next } |
| 227 | + b = (n*sxy - sx*sy) / denom |
| 228 | + a = (sy - b*sx) / n |
| 229 | + ym = sy/n; ss_tot=0; ss_res=0 |
| 230 | + for (i=0;i<n;i++) { p=a+b*x[i]; ss_res+=(y[i]-p)^2; ss_tot+=(y[i]-ym)^2 } |
| 231 | + r2 = (ss_tot>0) ? 1 - ss_res/ss_tot : 1 |
| 232 | + printf " %-30s %+.2fs/M steps (base: %.2fs, R\xc2\xb2=%.3f)\n", label, b, a, r2 |
| 233 | + }' |
| 234 | +done |
| 235 | + |
| 236 | +echo "" |
| 237 | +echo "Raw data in $TMP_DIR/" |
0 commit comments