Merge branch 'master' into evaluator-trait

primenumber · May 4, 2024 · 407fa75 · 407fa75
2 parents 23fc861 + 5ce34d2
commit 407fa75
Show file tree

Hide file tree

Showing 14 changed files with 128 additions and 74 deletions.
diff --git a/README.md b/README.md
@@ -30,60 +30,60 @@ $ cargo run --release -- ffobench
 
 ## Benchmark result
 
-- Date: 2024/02/21
+- Date: 2024/03/29
 - Hardware: AMD Ryzen 9 7950X3D, DDR5-4800 64GB
-- Environment: Linux 6.5.0-18-generic, Ubuntu 22.04.4, rustc 1.78.0-nightly
+- Environment: Linux 6.5.0-26-generic, Ubuntu 22.04.4, rustc 1.78.0-nightly
 
 FFO 40-59
 
 |No.|empties|result|answer|move|nodes|time|NPS|
 |---:|---:|---:|---:|---:|---:|:--:|---:|
-|40|20|+38|+38|A2|90.4M|   0.054s|1644M/s|
-|41|22| +0| +0|H4| 117M|   0.092s|1267M/s|
-|42|22| +6| +6|G2| 287M|   0.177s|1617M/s|
-|43|23|-12|-12|C7| 159M|   0.121s|1303M/s|
-|44|23|-14|-14|D2| 111M|   0.083s|1322M/s|
-|45|24| +6| +6|B2|1.57G|   1.010s|1561M/s|
-|46|24| -8| -8|B3| 494M|   0.349s|1412M/s|
-|47|25| +4| +4|G2| 167M|   0.123s|1350M/s|
-|48|25|+28|+28|F6| 901M|   0.680s|1323M/s|
-|49|26|+16|+16|E1|3.10G|   1.988s|1559M/s|
-|50|26|+10|+10|D8|3.56G|   2.773s|1285M/s|
-|51|27| +6| +6|E2|1.47G|   1.240s|1192M/s|
-|52|27| +0| +0|A3|1.31G|   1.080s|1215M/s|
-|53|28| -2| -2|D8|5.77G|   4.829s|1196M/s|
-|54|28| -2| -2|C7|15.6G|  11.420s|1368M/s|
-|55|29| +0| +0|G6|29.2G|  28.391s|1030M/s|
-|56|29| +2| +2|H5|4.56G|   4.727s|965M/s|
-|57|30|-10|-10|A6|19.8G|  18.593s|1066M/s|
-|58|30| +4| +4|G1|4.83G|   4.961s|973M/s|
-|59|34|+64|+64|G8|1.66k|   0.034s|0M/s|
-
-[Total] elapsed: 82738454us, node count: 93254297417, NPS: 1127097412nodes/sec
+|40|20|+38|+38|A2|76.2M|   0.062s|1209M/s|
+|41|22| +0| +0|H4|87.2M|   0.080s|1076M/s|
+|42|22| +6| +6|G2| 270M|   0.178s|1511M/s|
+|43|23|-12|-12|C7| 164M|   0.146s|1122M/s|
+|44|23|-14|-14|D2|84.3M|   0.081s|1029M/s|
+|45|24| +6| +6|B2|1.51G|   0.978s|1547M/s|
+|46|24| -8| -8|B3| 443M|   0.329s|1343M/s|
+|47|25| +4| +4|G2| 133M|   0.122s|1084M/s|
+|48|25|+28|+28|F6| 874M|   0.676s|1291M/s|
+|49|26|+16|+16|E1|3.14G|   2.047s|1536M/s|
+|50|26|+10|+10|D8|3.24G|   2.543s|1276M/s|
+|51|27| +6| +6|E2|1.41G|   1.188s|1191M/s|
+|52|27| +0| +0|A3|1.42G|   1.175s|1208M/s|
+|53|28| -2| -2|D8|4.91G|   4.038s|1216M/s|
+|54|28| -2| -2|C7|13.6G|   9.923s|1375M/s|
+|55|29| +0| +0|G6|27.3G|  25.529s|1070M/s|
+|56|29| +2| +2|H5|4.26G|   4.629s|921M/s|
+|57|30|-10|-10|A6|19.6G|  18.350s|1071M/s|
+|58|30| +4| +4|G1|4.48G|   4.760s|942M/s|
+|59|34|+64|+64|G8|1.26k|   0.029s|0M/s|
+
+[Total] elapsed: 76872921us, node count: 87193095283, NPS: 1134249800nodes/sec
 
 FFO 60-79
 
 |No.|empties|result|answer|move|nodes|time|NPS|
 |---:|---:|---:|---:|---:|---:|:--:|---:|
-|60|24|+20|+20|C2| 216M|   0.164s|1313M/s|
-|61|25|-14|-14|G1| 339M|   0.318s|1062M/s|
-|62|27|+28|+28|E8|8.51G|   7.369s|1154M/s|
-|63|27| -2| -2|F2|2.88G|   2.393s|1203M/s|
-|64|27|+20|+20|B4|11.2G|   9.350s|1207M/s|
-|65|28|+10|+10|G1|29.2G|  20.527s|1422M/s|
-|66|28|+30|+30|H3|21.1G|  16.079s|1314M/s|
-|67|28|+22|+22|H3|28.2G|  20.363s|1386M/s|
-|68|30|+28|+28|E8| 139G| 107.049s|1301M/s|
-|69|30| +0| +0|H3|15.7G|  14.081s|1121M/s|
-|70|30|-24|-24|E3|14.4G|  13.097s|1106M/s|
-|71|31|+20|+20|D2|20.1G|  19.281s|1044M/s|
-|72|31|+24|+24|E1| 258G| 298.006s|867M/s|
-|73|31| -4| -4|G4|29.4G|  34.672s|849M/s|
-|74|31|-30|-30|F1| 611G| 563.534s|1084M/s|
-|75|32|+14|+14|D2| 299G| 225.346s|1327M/s|
-|76|32|+32|+32|A3|2.12T|1869.331s|1134M/s|
-|77|34|+34|+34|B7|1.13T|1015.754s|1121M/s|
-|78|34| +8| +8|F1| 673G| 844.518s|797M/s|
-|79|36|+64|+64|D7|56.8G|  43.134s|1319M/s|
-
-[Total] elapsed: 5124378900us, node count: 5480838886501, NPS: 1069561598nodes/sec
+|60|24|+20|+20|C2| 213M|   0.175s|1213M/s|
+|61|25|-14|-14|G1| 344M|   0.325s|1058M/s|
+|62|27|+28|+28|E8|7.61G|   6.587s|1155M/s|
+|63|27| -2| -2|F2|2.89G|   2.383s|1215M/s|
+|64|27|+20|+20|B4|10.4G|   8.675s|1204M/s|
+|65|28|+10|+10|G1|26.6G|  18.953s|1405M/s|
+|66|28|+30|+30|H3|19.7G|  15.215s|1299M/s|
+|67|28|+22|+22|H3|25.8G|  18.823s|1375M/s|
+|68|30|+28|+28|E8| 116G|  90.898s|1287M/s|
+|69|30| +0| +0|H3|14.5G|  13.385s|1086M/s|
+|70|30|-24|-24|E3|13.0G|  11.413s|1146M/s|
+|71|31|+20|+20|D2|21.3G|  21.350s|999M/s|
+|72|31|+24|+24|E1| 178G| 197.643s|902M/s|
+|73|31| -4| -4|G4|25.8G|  31.323s|824M/s|
+|74|31|-30|-30|F1| 599G| 580.678s|1031M/s|
+|75|32|+14|+14|D2| 234G| 180.377s|1300M/s|
+|76|32|+32|+32|A3|1.56T|1441.196s|1087M/s|
+|77|34|+34|+34|B7|1.23T|1122.816s|1100M/s|
+|78|34| +8| +8|F1| 571G| 831.075s|687M/s|
+|79|36|+64|+64|D7|14.6G|  12.269s|1191M/s|
+
+[Total] elapsed: 4605575684us, node count: 4685891940872, NPS: 1017438918nodes/sec
diff --git a/src/book.rs b/src/book.rs
@@ -178,7 +178,7 @@ fn search<Eval: Evaluator>(
     solve_obj.cache_gen += 1;
     if board.empty().count_ones() <= 18 {
         let mut solve_obj = solve_obj.clone();
-        solve_with_move(board, &mut solve_obj, &sub_solver.clone())
+        solve_with_move(board, &mut solve_obj, &sub_solver.clone(), None)
     } else {
         let start = Instant::now();
         let timer = Timer {

diff --git a/src/engine/bits.rs b/src/engine/bits.rs
@@ -7,12 +7,12 @@ pub trait BitManip {
 }
 
 impl BitManip for u64 {
-    #[cfg(target_feature = "avx2")]
+    #[cfg(all(target_feature = "bmi2", not(slow_pext)))]
     fn pext(&self, mask: u64) -> u64 {
         unsafe { _pext_u64(*self, mask) }
     }
 
-    #[cfg(not(target_feature = "avx2"))]
+    #[cfg(not(all(target_feature = "bmi2", not(slow_pext))))]
     fn pext(&self, mut mask: u64) -> u64 {
         let mut x = *self;
         x = x & mask;

diff --git a/src/engine/board.rs b/src/engine/board.rs
@@ -1,11 +1,11 @@
 #[cfg(test)]
 mod test;
-#[cfg(target_feature = "neon")]
-use std::arch::aarch64::*;
 use crate::engine::bits::*;
 use crate::engine::hand::*;
 use anyhow::Result;
 use clap::ArgMatches;
+#[cfg(target_feature = "neon")]
+use std::arch::aarch64::*;
 use std::cmp::min;
 use std::fmt;
 use std::io::{BufWriter, Write};
@@ -55,7 +55,10 @@ fn smart_upper_bit(x: u64x4) -> u64x4 {
     }
 }
 
-#[cfg(not(any(target_feature = "neon", all(target_feature = "avx512cd", target_feature = "avx512vl"))))]
+#[cfg(not(any(
+    target_feature = "neon",
+    all(target_feature = "avx512cd", target_feature = "avx512vl")
+)))]
 fn smart_upper_bit(mut x: u64x4) -> u64x4 {
     x |= x >> u64x4::from_array([8, 1, 7, 9]);
     x |= x >> u64x4::from_array([16, 2, 14, 18]);

diff --git a/src/engine/endgame.rs b/src/engine/endgame.rs
@@ -7,7 +7,9 @@ use crate::engine::hand::*;
 use crate::engine::search::*;
 use crate::engine::table::*;
 use arrayvec::ArrayVec;
+use crc64::Crc64;
 use std::cmp::max;
+use std::io::Write;
 
 fn near_leaf<Eval: Evaluator>(solve_obj: &mut SolveObj<Eval>, board: Board) -> (i8, SolveStat) {
     let (score, node_count) = solve_obj.last_cache.solve_last(board);
@@ -211,8 +213,30 @@ pub fn solve_inner<Eval: Evaluator>(
                 CutType::LessThanAlpha(v) => return (v, SolveStat::one_stcut()),
             }
         }
-        if rem < solve_obj.params.res_cache_limit {
+        if rem < solve_obj.params.local_res_cache_limit {
             fastest_first(solve_obj, board, (alpha, beta), passed)
+        } else if rem < solve_obj.params.res_cache_limit {
+            let mut crc64 = Crc64::new();
+            crc64.write(&board.player.to_le_bytes()).unwrap();
+            crc64.write(&board.opponent.to_le_bytes()).unwrap();
+            let hash = crc64.get();
+            let res_cache = solve_obj.local_res_cache.get(board, hash);
+            let lookup_result = make_lookup_result(res_cache, (&mut alpha, &mut beta));
+            let (lower, upper) = match lookup_result {
+                CacheLookupResult::Cut(v) => return (v, SolveStat::zero()),
+                CacheLookupResult::NoCut(l, u, _) => (l, u),
+            };
+            let (res, stat) = fastest_first(solve_obj, board, (alpha, beta), passed);
+            let record = make_record(
+                solve_obj.local_cache_gen,
+                board,
+                res,
+                None,
+                (alpha, beta),
+                (lower, upper),
+            );
+            solve_obj.local_res_cache.update(&record, hash);
+            (res, stat)
         } else if rem < solve_obj.params.eval_ordering_limit {
             let (lower, upper) = match lookup_table(solve_obj, board, (&mut alpha, &mut beta)) {
                 CacheLookupResult::Cut(v) => return (v, SolveStat::zero()),

diff --git a/src/engine/last_cache.rs b/src/engine/last_cache.rs
@@ -63,23 +63,23 @@ impl LastCache {
         }
     }
 
-    #[cfg(target_feature = "bmi2")]
+    #[cfg(all(target_feature = "bmi2", not(slow_pext)))]
     fn get_col_bits(bits: u64, mask: u64, _col: usize) -> u64 {
         bits.pext(mask)
     }
 
-    #[cfg(not(target_feature = "bmi2"))]
+    #[cfg(not(all(target_feature = "bmi2", not(slow_pext))))]
     fn get_col_bits(mut bits: u64, mask: u64, col: usize) -> u64 {
         bits &= mask;
         ((bits >> col).wrapping_mul(0x0002_0408_1020_4081) >> 49) & 0xff
     }
 
-    #[cfg(target_feature = "bmi2")]
+    #[cfg(all(target_feature = "bmi2", not(slow_pext)))]
     fn get_diag1_bits(bits: u64, mask: u64, _row: usize, _col: usize) -> u64 {
         bits.pext(mask)
     }
 
-    #[cfg(not(target_feature = "bmi2"))]
+    #[cfg(not(all(target_feature = "bmi2", not(slow_pext))))]
     fn get_diag1_bits(mut bits: u64, mask: u64, row: usize, col: usize) -> u64 {
         bits &= mask;
         let width = if row >= col {
@@ -92,12 +92,12 @@ impl LastCache {
         (bits.wrapping_mul(0x0101_0101_0101_0101) >> 56) & ((1 << width) - 1)
     }
 
-    #[cfg(target_feature = "bmi2")]
+    #[cfg(all(target_feature = "bmi2", not(slow_pext)))]
     fn get_diag2_bits(bits: u64, mask: u64, _row: usize, _col: usize) -> u64 {
         bits.pext(mask)
     }
 
-    #[cfg(not(target_feature = "bmi2"))]
+    #[cfg(not(all(target_feature = "bmi2", not(slow_pext))))]
     fn get_diag2_bits(mut bits: u64, mask: u64, row: usize, col: usize) -> u64 {
         bits &= mask;
         let width = if row + col >= 7 {

diff --git a/src/engine/midgame.rs b/src/engine/midgame.rs
@@ -100,6 +100,7 @@ fn simplified_abdada_intro<Eval: Evaluator>(
     if depth >= ctx.solve_obj.params.parallel_depth_limit || rem < ctx.solve_obj.params.parallel_empties_limit {
         let (res, stat) = solve_inner(&mut ctx.solve_obj, board, (alpha, beta), passed);
         ctx.stats.merge(stat);
+        ctx.solve_obj.local_cache_gen += 1;
         return Some((res, None));
     }
     ctx.stats.merge(SolveStat::one());
@@ -148,12 +149,14 @@ pub fn simplified_abdada<Eval: Evaluator>(
     (alpha, beta): (i8, i8),
     passed: bool,
     depth: i8,
+    num_threads: Option<usize>,
 ) -> (i8, Option<Hand>, SolveStat) {
     thread::scope(|s| {
         let mut handles = Vec::new();
         let cs_hash = Arc::new(DashSet::new());
         let finished = Arc::new(AtomicBool::new(false));
-        for _ in 0..num_cpus::get() {
+        let num_threads = num_threads.unwrap_or(num_cpus::get());
+        for _ in 0..num_threads {
             let solve_obj = solve_obj.clone();
             let cs_hash = cs_hash.clone();
             let finished = finished.clone();

diff --git a/src/engine/search.rs b/src/engine/search.rs
@@ -40,6 +40,7 @@ pub struct SearchParams {
     pub parallel_empties_limit: i8,
     pub eval_ordering_limit: i8,
     pub res_cache_limit: i8,
+    pub local_res_cache_limit: i8,
     pub stability_cut_limit: i8,
     pub ffs_ordering_limit: i8,
     pub static_ordering_limit: i8,
@@ -48,21 +49,25 @@ pub struct SearchParams {
 pub struct SolveObj<Eval: Evaluator> {
     pub res_cache: Arc<ResCacheTable>,
     pub eval_cache: Arc<EvalCacheTable>,
+    pub local_res_cache: CacheArray<ResCache>,
     pub evaluator: Arc<Eval>,
     pub last_cache: Arc<LastCache>,
     pub params: SearchParams,
     pub cache_gen: u32,
+    pub local_cache_gen: u32,
 }
 
 impl<Eval: Evaluator> Clone for SolveObj<Eval> {
     fn clone(&self) -> Self {
         SolveObj::<Eval> {
             res_cache: self.res_cache.clone(),
             eval_cache: self.eval_cache.clone(),
+            local_res_cache: self.local_res_cache.clone(),
             evaluator: self.evaluator.clone(),
             last_cache: self.last_cache.clone(),
             params: self.params.clone(),
             cache_gen: self.cache_gen.clone(),
+            local_cache_gen: self.local_cache_gen.clone(),
         }
     }
 }
@@ -78,10 +83,12 @@ impl<Eval: Evaluator> SolveObj<Eval> {
         SolveObj {
             res_cache,
             eval_cache,
+            local_res_cache: CacheArray::<ResCache>::new(65536),
             evaluator,
             last_cache: Arc::new(LastCache::new()),
             params,
             cache_gen,
+            local_cache_gen: 0,
         }
     }
 }
@@ -294,21 +301,25 @@ pub fn solve<Eval: Evaluator>(
     (alpha, beta): (i8, i8),
     passed: bool,
     depth: i8,
+    num_threads: Option<usize>,
 ) -> (i8, Option<Hand>, SolveStat) {
-    simplified_abdada(solve_obj, board, (alpha, beta), passed, depth)
+    simplified_abdada(solve_obj, board, (alpha, beta), passed, depth, num_threads)
 }
 
+// num_threads: number of searching threads, use number of cpus when None
 pub fn solve_with_move<Eval: Evaluator>(
     board: Board,
     solve_obj: &mut SolveObj<Eval>,
     _sub_solver: &Arc<SubSolver>,
+    num_threads: Option<usize>,
 ) -> Hand {
     if let Some(best) = simplified_abdada(
         solve_obj,
         board,
         (-(BOARD_SIZE as i8), BOARD_SIZE as i8),
         false,
         0,
+        num_threads,
     )
     .1
     {
@@ -318,7 +329,15 @@ pub fn solve_with_move<Eval: Evaluator>(
     let mut result = -65;
     for pos in board.mobility() {
         let next = board.play(pos).unwrap();
-        let res = -simplified_abdada(solve_obj, next, (-(BOARD_SIZE as i8), -result), false, 0).0;
+        let res = -simplified_abdada(
+            solve_obj,
+            next,
+            (-(BOARD_SIZE as i8), -result),
+            false,
+            0,
+            num_threads,
+        )
+        .0;
         if res > result {
             result = res;
             best_pos = Some(pos);