From b2f646d9f99dd272f3b3a9d045b5039e6fc1dc50 Mon Sep 17 00:00:00 2001
From: Peter Fors <peter.fors@mindkiller.com>
Date: Sat, 25 Oct 2025 23:07:35 +0200
Subject: Refactor benchmarking to self-contained C implementation

- Add mknes_bench.c with direct PMC access via perf_event_open()
- Remove dependency on external perf/awk for statistics
- Add RT priority and CPU affinity control in C code
- Use static BSS allocation (page-aligned) instead of malloc
- Add stalled-cycles-backend counter (gracefully handles AMD unavailability)
- Add throughput metrics (MIPS, Mcycles/sec)
- Optimize Bench.sh to only regenerate profile data when needed
- Add -n and -f flags for configurable runs and frames
- Suppress mapper messages during benchmark
- ~6x faster benchmark workflow (20s first run, 16s subsequent)
---
 Bench.sh | 50 +++++++++-----------------------------------------
 1 file changed, 9 insertions(+), 41 deletions(-)

(limited to 'Bench.sh')

diff --git a/Bench.sh b/Bench.sh
index 157ad57..8be485a 100755
--- a/Bench.sh
+++ b/Bench.sh
@@ -1,45 +1,13 @@
 #!/usr/bin/env bash
-./build.sh clean
-./build.sh profile
-./mknes
-./build.sh profile_release
-
-runs=10
-frames=4096
-events="cycles,instructions,task-clock"
-tmp=$(mktemp)
-
-taskset -c 1 ./mknes
-
-> "$tmp"
-for i in $(seq 1 $runs); do
-  taskset -c 1 chrt -f 99 perf stat -x, -e $events -- ./mknes 2>>"$tmp"
-done
-
 
-awk -F, -v F="$frames" '
-  $3=="cycles"        { c[++nc]=$1/F }
-  $3=="instructions"  { i[++ni]=$1/F }
-#   $3=="task-clock"    { t[++nt]=$1 }  # milliseconds NOTE(peter): changed to nanoseconds...
-  $3=="task-clock" { t[++nt]=$1/1000000 }
+# Only rebuild profile data if it doesn't exist
+if [ ! -f mknes.gcda ]; then
+	./build.sh clean
+	./build.sh profile
+	./mknes -n 1 -f 1024  # Quick single run for profile generation
+fi
 
-  END {
-    for(k=1;k<=nc;k++) sumc+=c[k]; mc=sumc/nc
-    for(k=1;k<=ni;k++) sumi+=i[k]; mi=sumi/ni
-    for(k=1;k<=nt;k++) sumt+=t[k]; mt=sumt/nt
-
-    for(k=1;k<=nc;k++) sdc+=(c[k]-mc)^2; sdc=sqrt(sdc/(nc-1))
-    for(k=1;k<=ni;k++) sdi+=(i[k]-mi)^2; sdi=sqrt(sdi/(ni-1))
-    for(k=1;k<=nt;k++) sdt+=(t[k]-mt)^2; sdt=sqrt(sdt/(nt-1))
-
-    ms_per_frame = mt / F
-    fps = F / (mt / 1000)
-
-    printf "IPC (insn/cycle) = %.3f\n", mi/mc
-    printf "cycles/frame  mean=%.0f  sd=%.0f  relSD=%.3f%%  n=%d\n", mc, sdc, 100*sdc/mc, nc
-    printf "insn/frame    mean=%.0f  sd=%.0f  relSD=%.3f%%  n=%d\n", mi, sdi, 100*sdi/mi, ni
-    printf "time (ms)     mean=%.3f  sd=%.3f  relSD=%.3f%%  n=%d\n", mt, sdt, 100*sdt/mt, nt
-    printf "FPS (frames/second) = %.2f\n", fps
-    printf "ms/frame     = %.6f\n", ms_per_frame
-  }' "$tmp"
+./build.sh profile_release
 
+# Run full benchmark
+./mknes
-- 
cgit v1.2.3