#!/bin/bash
# Final Performance Benchmarks - Wave QAT Validation
# Run all critical performance tests and validate against targets

set -e

RESULTS_FILE="FINAL_PERFORMANCE_BENCHMARKS.txt"
echo "=== FOXHUNT FINAL PERFORMANCE BENCHMARKS ===" > $RESULTS_FILE
echo "Date: $(date)" >> $RESULTS_FILE
echo "System: $(uname -a)" >> $RESULTS_FILE
echo "GPU: $(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null || echo 'N/A')" >> $RESULTS_FILE
echo "" >> $RESULTS_FILE

echo "Running final performance benchmarks..."

# 1. DQN Memory Benchmark (target: <10MB actual, <150MB with overhead)
echo "=== 1. DQN MEMORY BENCHMARK ===" | tee -a $RESULTS_FILE
cargo run -p ml --example measure_dqn_memory --features cuda 2>&1 | tee -a $RESULTS_FILE
echo "" >> $RESULTS_FILE

# 2. TFT INT8 Inference Benchmark (target: <5ms)
echo "=== 2. TFT INT8 INFERENCE BENCHMARK ===" | tee -a $RESULTS_FILE
cd ml && timeout 300 cargo bench --bench tft_int8_inference_bench --features cuda 2>&1 | grep -A 20 "time:" | tee -a ../$RESULTS_FILE
cd ..
echo "" >> $RESULTS_FILE

# 3. Wave D Feature Extraction Benchmark (target: <50μs/bar)
echo "=== 3. WAVE D FEATURE EXTRACTION BENCHMARK ===" | tee -a $RESULTS_FILE
cd ml && timeout 120 cargo bench --bench wave_d_features_bench --features cuda 2>&1 | grep -E "(time:|thrpt:)" | head -30 | tee -a ../$RESULTS_FILE
cd ..
echo "" >> $RESULTS_FILE

# 4. ML Strategy Inference Benchmark (target: <1ms)
echo "=== 4. ML STRATEGY INFERENCE BENCHMARK ===" | tee -a $RESULTS_FILE
cd common && timeout 120 cargo bench --bench ml_strategy_bench 2>&1 | grep -E "(time:|thrpt:)" | head -20 | tee -a ../$RESULTS_FILE
cd ..
echo "" >> $RESULTS_FILE

echo "Benchmark suite completed. Results saved to: $RESULTS_FILE"