Add our own blas-like routines that use stockfish's thread pool for parallelization.

src/Makefile

@@ -47,6 +47,7 @@ PGOGENSFEN = ./$(EXE) gensfen depth 3 loop 1000 output_file_name $(PGO_TRAINING_
 SRCS = benchmark.cpp bitbase.cpp bitboard.cpp endgame.cpp evaluate.cpp main.cpp \
 	material.cpp misc.cpp movegen.cpp movepick.cpp pawns.cpp position.cpp psqt.cpp \
 	search.cpp thread.cpp timeman.cpp tt.cpp uci.cpp ucioption.cpp tune.cpp syzygy/tbprobe.cpp \
+	extra/stockfish_blas.cpp \
 	nnue/evaluate_nnue.cpp \
 	nnue/evaluate_nnue_learner.cpp \
 	nnue/features/half_kp.cpp \

src/extra/stockfish_blas.h

@@ -0,0 +1,130 @@
+#ifndef _STOCKFISH_BLAS_H_
+#define _STOCKFISH_BLAS_H_
+
+struct ThreadPool;
+
+#if defined (_MSC_VER)
+#define SF_BLAS_RESTRICT __restrict
+#elif defined (__INTEL_COMPILER)
+#define SF_BLAS_RESTRICT restrict
+#elif defined (__clang__)
+#define SF_BLAS_RESTRICT __restrict__
+#elif defined (__GNUC__)
+#define SF_BLAS_RESTRICT __restrict__
+#endif
+
+namespace Blas {
+
+    enum struct MatrixLayout {
+        RowMajor = 101,
+        ColMajor = 102
+    };
+
+    enum struct MatrixTranspose {
+        NoTrans = 111,
+        Trans = 112
+    };
+
+    void scopy(
+        const int N,
+        const float * SF_BLAS_RESTRICT X,
+        float * SF_BLAS_RESTRICT Y
+    );
+
+    void scopy(
+        const int N,
+        const float * SF_BLAS_RESTRICT X, const int incX,
+        float * SF_BLAS_RESTRICT Y, const int incY
+    );
+
+    void scopy(
+        ThreadPool& thread_pool,
+        const int N,
+        const float * SF_BLAS_RESTRICT X,
+        float * SF_BLAS_RESTRICT Y
+    );
+
+    void scopy(
+        ThreadPool& thread_pool,
+        const int N,
+        const float * SF_BLAS_RESTRICT X, const int incX,
+        float * SF_BLAS_RESTRICT Y, const int incY
+    );
+
+    void sscal(
+        const int N,
+        const float alpha,
+        float * SF_BLAS_RESTRICT X
+    );
+
+    void sscal(
+        const int N,
+        const float alpha,
+        float * SF_BLAS_RESTRICT X, const int incX
+    );
+
+    void sscal(
+        ThreadPool& thread_pool,
+        const int N,
+        const float alpha,
+        float * SF_BLAS_RESTRICT X
+    );
+
+    void sscal(
+        ThreadPool& thread_pool,
+        const int N,
+        const float alpha,
+        float * SF_BLAS_RESTRICT X, const int incX
+    );
+
+    void saxpy(
+        const int N,
+        const float alpha,
+        const float * SF_BLAS_RESTRICT X,
+        float * SF_BLAS_RESTRICT Y
+    );
+
+    void saxpy(
+        const int N,
+        const float alpha,
+        const float * SF_BLAS_RESTRICT X, const int incX,
+        float * SF_BLAS_RESTRICT Y, const int incY
+    );
+
+    void saxpy(
+        ThreadPool& thread_pool,
+        const int N,
+        const float alpha,
+        const float * SF_BLAS_RESTRICT X,
+        float * SF_BLAS_RESTRICT Y
+    );
+
+    void saxpy(
+        ThreadPool& thread_pool,
+        const int N,
+        const float alpha,
+        const float * SF_BLAS_RESTRICT X, const int incX,
+        float * SF_BLAS_RESTRICT Y, const int incY
+    );
+
+    void sgemm(
+        ThreadPool& thread_pool,
+        MatrixLayout layout, MatrixTranspose TransA, MatrixTranspose TransB,
+        const int M, const int N, const int K,
+        const float alpha,
+        const float * SF_BLAS_RESTRICT A, const int lda,
+        const float * SF_BLAS_RESTRICT B, const int ldb,
+        const float beta,
+        float * SF_BLAS_RESTRICT C, const int ldc
+    );
+
+    void test(
+        ThreadPool& thread_pool
+    );
+
+    void bench(
+        ThreadPool& thread_pool
+    );
+}
+
+#endif

src/thread.h

@@ -39,6 +39,15 @@
 /// pointer to an entry its life time is unlimited and we don't have
 /// to care about someone changing the entry under our feet.
 
+namespace Detail {
+
+  template <typename T>
+  struct TypeIdentity {
+    using Type = T;
+  };
+
+}
+
 class Thread {
 
   std::mutex mutex;
@@ -120,6 +129,26 @@ struct ThreadPool : public std::vector<Thread*> {
   // to the state of the `worker` function object.
   void execute_with_workers(const std::function<void(Thread&)>& worker);
 
+  template <typename IndexT, typename FuncT>
+  void for_each_index_with_workers(
+    IndexT begin,
+    typename Detail::TypeIdentity<IndexT>::Type end,
+    FuncT func)
+  {
+    std::atomic<IndexT> i_atomic = begin;
+
+    execute_with_workers(
+      [&i_atomic, end, func](Thread& th) mutable {
+        for(;;) {
+          const auto i = i_atomic.fetch_add(1);
+          if (i >= end)
+            break;
+
+          func(th, i);
+        }
+      });
+  }
+
   void start_thinking(Position&, StateListPtr&, const Search::LimitsType&, bool = false);
   void clear();
   void set(size_t);

src/uci.cpp

@@ -22,6 +22,7 @@
 #include <sstream>
 #include <string>
 
+#include "extra/stockfish_blas.h"
 #include "nnue/evaluate_nnue.h"
 #include "evaluate.h"
 #include "movegen.h"
@@ -354,6 +355,14 @@ void UCI::loop(int argc, char* argv[]) {
           std::cout << th.thread_idx() << '\n';
         });
       }
+      else if (token == "blastest")
+      {
+        Blas::test(Threads);
+      }
+      else if (token == "blasbench")
+      {
+        Blas::bench(Threads);
+      }
 
       // test command
       else if (token == "test") test_cmd(pos, is);