2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2023 The Stockfish developers (see AUTHORS file)
5 Stockfish is free software: you can redistribute it and/or modify
6 it under the terms of the GNU General Public License as published by
7 the Free Software Foundation, either version 3 of the License, or
8 (at your option) any later version.
10 Stockfish is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 GNU General Public License for more details.
15 You should have received a copy of the GNU General Public License
16 along with this program. If not, see <http://www.gnu.org/licenses/>.
19 // Input features and network structure used in NNUE evaluation function
21 #ifndef NNUE_ARCHITECTURE_H_INCLUDED
22 #define NNUE_ARCHITECTURE_H_INCLUDED
26 #include "nnue_common.h"
28 #include "features/half_ka_v2_hm.h"
30 #include "layers/affine_transform.h"
31 #include "layers/clipped_relu.h"
32 #include "layers/sqr_clipped_relu.h"
36 namespace Stockfish::Eval::NNUE {
38 // Input features used in evaluation function
39 using FeatureSet = Features::HalfKAv2_hm;
41 // Number of input feature dimensions after conversion
42 constexpr IndexType TransformedFeatureDimensions = 1024;
43 constexpr IndexType PSQTBuckets = 8;
44 constexpr IndexType LayerStacks = 8;
48 static constexpr int FC_0_OUTPUTS = 15;
49 static constexpr int FC_1_OUTPUTS = 32;
51 Layers::AffineTransform<TransformedFeatureDimensions, FC_0_OUTPUTS + 1> fc_0;
52 Layers::SqrClippedReLU<FC_0_OUTPUTS + 1> ac_sqr_0;
53 Layers::ClippedReLU<FC_0_OUTPUTS + 1> ac_0;
54 Layers::AffineTransform<FC_0_OUTPUTS * 2, FC_1_OUTPUTS> fc_1;
55 Layers::ClippedReLU<FC_1_OUTPUTS> ac_1;
56 Layers::AffineTransform<FC_1_OUTPUTS, 1> fc_2;
58 // Hash value embedded in the evaluation file
59 static constexpr std::uint32_t get_hash_value() {
61 std::uint32_t hashValue = 0xEC42E90Du;
62 hashValue ^= TransformedFeatureDimensions * 2;
64 hashValue = decltype(fc_0)::get_hash_value(hashValue);
65 hashValue = decltype(ac_0)::get_hash_value(hashValue);
66 hashValue = decltype(fc_1)::get_hash_value(hashValue);
67 hashValue = decltype(ac_1)::get_hash_value(hashValue);
68 hashValue = decltype(fc_2)::get_hash_value(hashValue);
73 // Read network parameters
74 bool read_parameters(std::istream& stream) {
75 return fc_0.read_parameters(stream)
76 && ac_0.read_parameters(stream)
77 && fc_1.read_parameters(stream)
78 && ac_1.read_parameters(stream)
79 && fc_2.read_parameters(stream);
82 // Write network parameters
83 bool write_parameters(std::ostream& stream) const {
84 return fc_0.write_parameters(stream)
85 && ac_0.write_parameters(stream)
86 && fc_1.write_parameters(stream)
87 && ac_1.write_parameters(stream)
88 && fc_2.write_parameters(stream);
91 std::int32_t propagate(const TransformedFeatureType* transformedFeatures)
93 struct alignas(CacheLineSize) Buffer
95 alignas(CacheLineSize) decltype(fc_0)::OutputBuffer fc_0_out;
96 alignas(CacheLineSize) decltype(ac_sqr_0)::OutputType ac_sqr_0_out[ceil_to_multiple<IndexType>(FC_0_OUTPUTS * 2, 32)];
97 alignas(CacheLineSize) decltype(ac_0)::OutputBuffer ac_0_out;
98 alignas(CacheLineSize) decltype(fc_1)::OutputBuffer fc_1_out;
99 alignas(CacheLineSize) decltype(ac_1)::OutputBuffer ac_1_out;
100 alignas(CacheLineSize) decltype(fc_2)::OutputBuffer fc_2_out;
104 std::memset(this, 0, sizeof(*this));
108 #if defined(__clang__) && (__APPLE__)
109 // workaround for a bug reported with xcode 12
110 static thread_local auto tlsBuffer = std::make_unique<Buffer>();
111 // Access TLS only once, cache result.
112 Buffer& buffer = *tlsBuffer;
114 alignas(CacheLineSize) static thread_local Buffer buffer;
117 fc_0.propagate(transformedFeatures, buffer.fc_0_out);
118 ac_sqr_0.propagate(buffer.fc_0_out, buffer.ac_sqr_0_out);
119 ac_0.propagate(buffer.fc_0_out, buffer.ac_0_out);
120 std::memcpy(buffer.ac_sqr_0_out + FC_0_OUTPUTS, buffer.ac_0_out, FC_0_OUTPUTS * sizeof(decltype(ac_0)::OutputType));
121 fc_1.propagate(buffer.ac_sqr_0_out, buffer.fc_1_out);
122 ac_1.propagate(buffer.fc_1_out, buffer.ac_1_out);
123 fc_2.propagate(buffer.ac_1_out, buffer.fc_2_out);
125 // buffer.fc_0_out[FC_0_OUTPUTS] is such that 1.0 is equal to 127*(1<<WeightScaleBits) in quantized form
126 // but we want 1.0 to be equal to 600*OutputScale
127 std::int32_t fwdOut = int(buffer.fc_0_out[FC_0_OUTPUTS]) * (600*OutputScale) / (127*(1<<WeightScaleBits));
128 std::int32_t outputValue = buffer.fc_2_out[0] + fwdOut;
134 } // namespace Stockfish::Eval::NNUE
136 #endif // #ifndef NNUE_ARCHITECTURE_H_INCLUDED