launch_turbo.sh

#!/bin/bash
# launch_turbo.sh — Launch ch1mera with all CPU optimizations
#
# First run:  chmod +x launch_turbo.sh
# Usage:      ./launch_turbo.sh [train_hyper.py args...]
# Example:    ./launch_turbo.sh --growlength --reservoir
# Full:       ./launch_turbo.sh --scale tiny --seq_len 128 --max_steps 300 --batch_size 32 --growlength --reservoir

set -euo pipefail
trap 'echo "[TURBO] ✗ Failed at line $LINENO (exit $?)" >&2' ERR

# ── Detect core topology ──
# On Raptor Lake i7-14700T: 8 P-cores (0-7) + 12 E-cores (8-19)
# P-cores have AVX-512 BF16; E-cores do NOT.
# Mixing P+E in OpenMP parallel regions causes barrier stalls:
# P-cores finish tiles in X cycles, E-cores in ~2.5X → P-cores idle.
# Solution: pin to P-cores only.
PHYS_CORES=$(lscpu -p | grep -v '^#' | sort -t, -k 2 -un | wc -l)
PCORE_THREADS=${CHIMERA_PCORE_THREADS:-8}  # Override for non-hybrid CPUs
echo "[TURBO] Physical cores: $PHYS_CORES → P-core threads: $PCORE_THREADS"

# ── Threading: P-cores only ──
export OMP_NUM_THREADS=$PCORE_THREADS
export MKL_NUM_THREADS=$PCORE_THREADS
# Explicit proclist pins threads 0-7 to physical P-cores.
# On Raptor Lake, P-cores are typically CPU 0-7 in lscpu topology.
# Verify with: lscpu -p | head -20
export KMP_AFFINITY="granularity=fine,proclist=[0,1,2,3,4,5,6,7],explicit"
export KMP_BLOCKTIME=0  # Zero blocktime: yield immediately after parallel region
                        # Eliminates busy-wait power waste on 35W TDP chip

# ── tcmalloc (if available — non-debug only) ──
# IMPORTANT: tcmalloc_debug / tcmalloc_minimal_debug cause silent aborts.
# Only load the release version.
TCMALLOC_LIB=""
for candidate in $(ldconfig -p 2>/dev/null | grep -oP '/\S*libtcmalloc(|_minimal)\.so\S*' | grep -v debug || true); do
    if [ -f "$candidate" ]; then
        TCMALLOC_LIB="$candidate"
        break
    fi
done

if [ -n "$TCMALLOC_LIB" ]; then
    echo "[TURBO] tcmalloc: $TCMALLOC_LIB"
    export LD_PRELOAD="$TCMALLOC_LIB${LD_PRELOAD:+:$LD_PRELOAD}"
else
    echo "[TURBO] ⚠ tcmalloc (non-debug) not found — expect 5-8% throughput loss."
    echo "[TURBO]   Install: sudo apt install libgoogle-perftools4"
fi

# ── IOMP (Intel OpenMP, if available) ──
IOMP_LIB=""
IOMP_LIB=$(python3 -c "
import intel_extension_for_pytorch, os
print(os.path.join(os.path.dirname(intel_extension_for_pytorch.__file__), '..', 'libiomp5.so'))
" 2>/dev/null) || true

if [ -n "$IOMP_LIB" ] && [ -f "$IOMP_LIB" ]; then
    echo "[TURBO] libiomp5: $IOMP_LIB"
    export LD_PRELOAD="$IOMP_LIB${LD_PRELOAD:+:$LD_PRELOAD}"
fi

# ── NUMA pinning (if numactl available) ──
NUMA_PREFIX=""
if command -v numactl &>/dev/null; then
    echo "[TURBO] NUMA: pinning to node 0"
    NUMA_PREFIX="numactl --cpunodebind=0 --membind=0"
fi

# ── Launch ──
echo "[TURBO] Launching: python3 train_hyper.py $*"
echo "═══════════════════════════════════════════════════"

$NUMA_PREFIX python3 train_hyper.py "$@"