feat(bench): fix find_optimal_batch to reduce bench time

Author: SouchonTheo

tfhe-benchmark/benches/integer/bench.rs

@@ -3559,27 +3559,27 @@ criterion_group!(
     mul_parallelized,
     div_rem_parallelized,
     bitand_parallelized,
-    bitnot,
-    left_shift_parallelized,
-    rotate_left_parallelized,
-    max_parallelized,
-    eq_parallelized,
-    gt_parallelized,
-    if_then_else_parallelized,
-    flip_parallelized,
-    neg_parallelized,
-    leading_zeros_parallelized,
-    ilog2_parallelized,
-    scalar_add_parallelized,
-    scalar_mul_parallelized,
-    scalar_div_parallelized,
-    scalar_rem_parallelized,
-    scalar_bitand_parallelized,
-    scalar_left_shift_parallelized,
-    scalar_rotate_left_parallelized,
-    scalar_max_parallelized,
-    scalar_eq_parallelized,
-    scalar_gt_parallelized,
+    // bitnot,
+    // left_shift_parallelized,
+    // rotate_left_parallelized,
+    // max_parallelized,
+    // eq_parallelized,
+    // gt_parallelized,
+    // if_then_else_parallelized,
+    // flip_parallelized,
+    // neg_parallelized,
+    // leading_zeros_parallelized,
+    // ilog2_parallelized,
+    // scalar_add_parallelized,
+    // scalar_mul_parallelized,
+    // scalar_div_parallelized,
+    // scalar_rem_parallelized,
+    // scalar_bitand_parallelized,
+    // scalar_left_shift_parallelized,
+    // scalar_rotate_left_parallelized,
+    // scalar_max_parallelized,
+    // scalar_eq_parallelized,
+    // scalar_gt_parallelized,
 );
 
 criterion_group!(

tfhe-benchmark/src/find_optimal_batch.rs

@@ -4,19 +4,25 @@ use std::time::{Duration, Instant};
 struct MeasureConfig {
     pub target_ratio: f64,
     pub long_target_ratio: f64,
+    pub min_improvement_ratio: f64,
     pub starting_batch_size: usize,
     pub minimum_time_per_batch: Duration,
     pub duration_threshold_for_long_test: Duration,
+    pub max_time_per_batch: Duration,
+    pub max_time_overshoot_ratio: f64,
 }
 
 impl MeasureConfig {
     pub fn default() -> Self {
         Self {
             target_ratio: 0.95,
             long_target_ratio: 0.80,
+            min_improvement_ratio: 0.02,
             starting_batch_size: 1,
             minimum_time_per_batch: Duration::from_secs(3),
             duration_threshold_for_long_test: Duration::from_secs(30),
+            max_time_per_batch: Duration::from_secs(200),
+            max_time_overshoot_ratio: 1.3,
         }
     }
 }
@@ -63,7 +69,7 @@ where
     let target = cores * measure_config.target_ratio;
     let long_target = cores * measure_config.long_target_ratio;
 
-    let mut low;
+    let mut low = measure_config.starting_batch_size;
     let mut high = measure_config.starting_batch_size;
     let mut last_usage = 0.0;
 
@@ -86,12 +92,23 @@ where
             return high;
         }
 
+        if duration >= measure_config.max_time_per_batch {
+            if duration
+                <= measure_config
+                    .max_time_per_batch
+                    .mul_f64(measure_config.max_time_overshoot_ratio)
+            {
+                return high;
+            }
+            return low;
+        }
+
         let improvement = (usage - last_usage).abs() / last_usage;
-        if improvement < 0.05
+        if improvement < measure_config.min_improvement_ratio
             && duration >= measure_config.duration_threshold_for_long_test
             && last_usage > long_target
         {
-            return high;
+            return low;
         }
 
         low = high;