bench 10561225

Makefile

@@ -3,7 +3,7 @@ OBJ_DIR=build
 MOLY_DIR=src
 CXX=clang++
 
-DEFAULT_NET_NAME=moly_20240526.nnue
+DEFAULT_NET_NAME=params.bin
 
 DEFAULT_EXE = $(OBJ_DIR)/Molybdenum
 DATAGEN_EXE = $(OBJ_DIR)/Datagen

src/nnue.cpp

@@ -26,10 +26,16 @@ INCBIN(network, EVALFILE);
 const Weights defaultWeights = *reinterpret_cast<const Weights*>(gnetworkData);
 
 void Net::loadDefaultNet() {
-    weights0 = defaultWeights.weights0;
+    for (size_t i = 0; i < defaultWeights.weights0.size(); i++)
+        weights0[i] = int16_t(defaultWeights.weights0[i] * 255);
+
+    for (size_t i = 0; i < defaultWeights.bias0.size(); i++)
+        bias0[i] = int16_t(defaultWeights.bias0[i] * 255);
+
     weights1 = defaultWeights.weights1;
-    bias0 = defaultWeights.bias0;
-    bias1 = defaultWeights.bias1;
+    bias1    = defaultWeights.bias1;
+    weights2 = defaultWeights.weights2;
+    bias2    = defaultWeights.bias2;
 }
 
 void Net::initAccumulator(std::array<u64, 13> &bitboards) {
@@ -52,12 +58,18 @@ void Net::initAccumulator(std::array<u64, 13> &bitboards) {
 }
 
 int Net::calculate(Color c) {
-    int output = 0;
+    float out = bias2[0];
+    std::array<float, 8> output = bias1;
 
-    for (int n = 0; n != L1_SIZE; n++) {
-         output += screlu(accumulator[ c][n]) * weights1[n          ];
-         output += screlu(accumulator[!c][n]) * weights1[n + L1_SIZE];
+    for (int n = 0; n < L1_SIZE; n++) {
+        for (int m = 0; m < L2_SIZE; m++) {
+            output[m] += screlu(float(accumulator[ c][n]) / 255.0f) * weights1[n * L2_SIZE + m                    ];
+            output[m] += screlu(float(accumulator[!c][n]) / 255.0f) * weights1[n * L2_SIZE + m + L1_SIZE * L2_SIZE];
+        }
     }
 
-    return ((output / 255) + bias1[0]) * 133 / (64 * 255);
+    for (int n = 0; n < L2_SIZE; n++)
+        out += screlu(output[n]) * weights2[n];
+
+    return int(out * 133.0f);
 }

src/nnue.h

@@ -14,23 +14,26 @@ enum Toggle {
 
 static const int INPUT_SIZE = 12 * 64;
 static const int L1_SIZE = 256;
-static const int OUTPUT_SIZE = 1;
-static const int NET_SIZE = 3;
-static const std::array<int, NET_SIZE> LAYER_SIZE = {INPUT_SIZE, L1_SIZE, OUTPUT_SIZE};
+static const int L2_SIZE = 8;
+static const int OUT_SIZE = 1;
 
 struct Weights {
-    std::array<int16_t , L1_SIZE * INPUT_SIZE> weights0{};
-    std::array<int16_t, L1_SIZE> bias0{};
-    std::array<int16_t, L1_SIZE * OUTPUT_SIZE * 2> weights1{};
-    std::array<int16_t, OUTPUT_SIZE> bias1{};
+    std::array<float, L1_SIZE * INPUT_SIZE> weights0{};
+    std::array<float, L1_SIZE> bias0{};
+    std::array<float, L1_SIZE * L2_SIZE * 2> weights1{};
+    std::array<float, L2_SIZE> bias1{};
+    std::array<float, L2_SIZE * OUT_SIZE> weights2{};
+    std::array<float, OUT_SIZE> bias2{};
 };
 
 class Net {
 public:
     std::array<int16_t , L1_SIZE * INPUT_SIZE> weights0{};
     std::array<int16_t, L1_SIZE> bias0{};
-    std::array<int16_t, L1_SIZE * OUTPUT_SIZE * 2> weights1{};
-    std::array<int16_t, OUTPUT_SIZE> bias1{};
+    std::array<float, L1_SIZE * L2_SIZE * 2> weights1{};
+    std::array<float, L2_SIZE> bias1{};
+    std::array<float, L2_SIZE * OUT_SIZE> weights2{};
+    std::array<float, OUT_SIZE> bias2{};
     std::array<std::array<int16_t, L1_SIZE>, 2> accumulator{};
     Stack<std::array<std::array<int16_t, L1_SIZE>, 2>, MAXDEPTH> accumulatorStack;
 
@@ -50,6 +53,11 @@ inline int screlu(int16_t input) {
     return clamped * clamped;
 }
 
+inline float screlu(float input) {
+    float clamped = std::clamp(input, 0.0f, 1.0f);
+    return clamped * clamped;
+}
+
 template<Color C> inline
 int index(int pc, int sq) {
     int square = C ? sq : sq ^ 56;