Skip to content

Commit 0f4300b

Browse files
Sopel97Sopel97
committed
Update architecture to "SF15a".
Update network to nn-6877cd24400e.nnue. Bench: 4919707
1 parent b0b3155 commit 0f4300b

File tree

7 files changed

+241
-306
lines changed

7 files changed

+241
-306
lines changed

‎src/evaluate.h‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -39,7 +39,7 @@ namespace Eval {
3939
// The default net name MUST follow the format nn-[SHA256 first 12 digits].nnue
4040
// for the build process (profile-build and fishtest) to work. Do not change the
4141
// name of the macro, as it is used in the Makefile.
42-
#define EvalFileDefaultName "nn-ac07bd334b62.nnue"
42+
#define EvalFileDefaultName "nn-6877cd24400e.nnue"
4343

4444
namespace NNUE {
4545

‎src/nnue/evaluate_nnue.cpp‎

Lines changed: 5 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -146,24 +146,20 @@ namespace Stockfish::Eval::NNUE {
146146
int delta = 7;
147147

148148
#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
149-
TransformedFeatureType transformedFeaturesUnaligned[
149+
TransformedFeatureType transformed_features_unaligned[
150150
FeatureTransformer::BufferSize + alignment / sizeof(TransformedFeatureType)];
151-
char bufferUnaligned[Network::BufferSize + alignment];
152151

153-
auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
154-
auto* buffer = align_ptr_up<alignment>(&bufferUnaligned[0]);
152+
auto* transformedFeatures = align_ptr_up<alignment>(&transformed_features_unaligned[0]);
155153
#else
156154
alignas(alignment)
157155
TransformedFeatureType transformedFeatures[FeatureTransformer::BufferSize];
158-
alignas(alignment) char buffer[Network::BufferSize];
159156
#endif
160157

161158
ASSERT_ALIGNED(transformedFeatures, alignment);
162-
ASSERT_ALIGNED(buffer, alignment);
163159

164160
const std::size_t bucket = (pos.count<ALL_PIECES>() - 1) / 4;
165161
const auto psqt = featureTransformer->transform(pos, transformedFeatures, bucket);
166-
const auto positional = network[bucket]->propagate(transformedFeatures, buffer)[0];
162+
const auto positional = network[bucket]->propagate(transformedFeatures);
167163

168164
// Give more value to positional evaluation when adjusted flag is set
169165
if (adjusted)
@@ -190,27 +186,20 @@ namespace Stockfish::Eval::NNUE {
190186
#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
191187
TransformedFeatureType transformedFeaturesUnaligned[
192188
FeatureTransformer::BufferSize + alignment / sizeof(TransformedFeatureType)];
193-
char bufferUnaligned[Network::BufferSize + alignment];
194189

195190
auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
196-
auto* buffer = align_ptr_up<alignment>(&bufferUnaligned[0]);
197191
#else
198192
alignas(alignment)
199193
TransformedFeatureType transformedFeatures[FeatureTransformer::BufferSize];
200-
alignas(alignment) char buffer[Network::BufferSize];
201194
#endif
202195

203196
ASSERT_ALIGNED(transformedFeatures, alignment);
204-
ASSERT_ALIGNED(buffer, alignment);
205197

206198
NnueEvalTrace t{};
207199
t.correctBucket = (pos.count<ALL_PIECES>() - 1) / 4;
208200
for (std::size_t bucket = 0; bucket < LayerStacks; ++bucket) {
209-
const auto psqt = featureTransformer->transform(pos, transformedFeatures, bucket);
210-
const auto output = network[bucket]->propagate(transformedFeatures, buffer);
211-
212-
int materialist = psqt;
213-
int positional = output[0];
201+
const auto materialist = featureTransformer->transform(pos, transformedFeatures, bucket);
202+
const auto positional = network[bucket]->propagate(transformedFeatures);
214203

215204
t.psqt[bucket] = static_cast<Value>( materialist / OutputScale );
216205
t.positional[bucket] = static_cast<Value>( positional / OutputScale );

‎src/nnue/layers/affine_transform.h‎

Lines changed: 32 additions & 59 deletions
Original file line numberDiff line numberDiff line change
@@ -63,19 +63,17 @@ namespace Stockfish::Eval::NNUE::Layers {
6363
{
6464
# if defined(USE_SSE2)
6565
// At least a multiple of 16, with SSE2.
66-
static_assert(PaddedInputDimensions % 16 == 0);
67-
constexpr IndexType NumChunks = PaddedInputDimensions / 16;
66+
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
6867
const __m128i Zeros = _mm_setzero_si128();
6968
const auto inputVector = reinterpret_cast<const __m128i*>(input);
7069

7170
# elif defined(USE_MMX)
72-
static_assert(InputDimensions % 8 == 0);
73-
constexpr IndexType NumChunks = InputDimensions / 8;
71+
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / 8;
7472
const __m64 Zeros = _mm_setzero_si64();
7573
const auto inputVector = reinterpret_cast<const __m64*>(input);
7674

7775
# elif defined(USE_NEON)
78-
constexpr IndexType NumChunks = (InputDimensions + 15) / 16;
76+
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
7977
const auto inputVector = reinterpret_cast<const int8x8_t*>(input);
8078
# endif
8179

@@ -150,24 +148,27 @@ namespace Stockfish::Eval::NNUE::Layers {
150148
}
151149
#endif
152150

153-
template <typename PreviousLayer, IndexType OutDims, typename Enabled = void>
151+
template <IndexType InDims, IndexType OutDims, typename Enabled = void>
154152
class AffineTransform;
155153

156154
// A specialization for large inputs.
157-
template <typename PreviousLayer, IndexType OutDims>
158-
class AffineTransform<PreviousLayer, OutDims, std::enable_if_t<(PreviousLayer::OutputDimensions >= 2*64-1)>> {
155+
template <IndexType InDims, IndexType OutDims>
156+
class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) >= 2*64)>> {
159157
public:
160158
// Input/output type
161-
using InputType = typename PreviousLayer::OutputType;
159+
using InputType = std::uint8_t;
162160
using OutputType = std::int32_t;
163-
static_assert(std::is_same<InputType, std::uint8_t>::value, "");
164161

165162
// Number of input/output dimensions
166-
static constexpr IndexType InputDimensions = PreviousLayer::OutputDimensions;
163+
static constexpr IndexType InputDimensions = InDims;
167164
static constexpr IndexType OutputDimensions = OutDims;
168165

169166
static constexpr IndexType PaddedInputDimensions =
170167
ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
168+
static constexpr IndexType PaddedOutputDimensions =
169+
ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);
170+
171+
using OutputBuffer = OutputType[PaddedOutputDimensions];
171172

172173
static_assert(PaddedInputDimensions >= 128, "Something went wrong. This specialization should not have been chosen.");
173174

@@ -202,20 +203,12 @@ namespace Stockfish::Eval::NNUE::Layers {
202203

203204
static_assert(OutputDimensions % NumOutputRegs == 0);
204205

205-
// Size of forward propagation buffer used in this layer
206-
static constexpr std::size_t SelfBufferSize =
207-
ceil_to_multiple(OutputDimensions * sizeof(OutputType), CacheLineSize);
208-
209-
// Size of the forward propagation buffer used from the input layer to this layer
210-
static constexpr std::size_t BufferSize =
211-
PreviousLayer::BufferSize + SelfBufferSize;
212-
213206
// Hash value embedded in the evaluation file
214-
static constexpr std::uint32_t get_hash_value() {
207+
static constexpr std::uint32_t get_hash_value(std::uint32_t prevhash) {
215208
std::uint32_t hashValue = 0xCC03DAE4u;
216209
hashValue += OutputDimensions;
217-
hashValue ^= PreviousLayer::get_hash_value() >> 1;
218-
hashValue ^= PreviousLayer::get_hash_value() << 31;
210+
hashValue ^= prevhash >> 1;
211+
hashValue ^= prevhash << 31;
219212
return hashValue;
220213
}
221214

@@ -242,7 +235,6 @@ namespace Stockfish::Eval::NNUE::Layers {
242235

243236
// Read network parameters
244237
bool read_parameters(std::istream& stream) {
245-
if (!previousLayer.read_parameters(stream)) return false;
246238
for (std::size_t i = 0; i < OutputDimensions; ++i)
247239
biases[i] = read_little_endian<BiasType>(stream);
248240

@@ -254,7 +246,6 @@ namespace Stockfish::Eval::NNUE::Layers {
254246

255247
// Write network parameters
256248
bool write_parameters(std::ostream& stream) const {
257-
if (!previousLayer.write_parameters(stream)) return false;
258249
for (std::size_t i = 0; i < OutputDimensions; ++i)
259250
write_little_endian<BiasType>(stream, biases[i]);
260251

@@ -266,10 +257,7 @@ namespace Stockfish::Eval::NNUE::Layers {
266257

267258
// Forward propagation
268259
const OutputType* propagate(
269-
const TransformedFeatureType* transformedFeatures, char* buffer) const {
270-
const auto input = previousLayer.propagate(
271-
transformedFeatures, buffer + SelfBufferSize);
272-
OutputType* output = reinterpret_cast<OutputType*>(buffer);
260+
const InputType* input, OutputType* output) const {
273261

274262
#if defined (USE_AVX512)
275263
using acc_vec_t = __m512i;
@@ -312,7 +300,6 @@ namespace Stockfish::Eval::NNUE::Layers {
312300
#if defined (USE_SSSE3) || defined (USE_NEON)
313301
const in_vec_t* invec = reinterpret_cast<const in_vec_t*>(input);
314302

315-
316303
// Perform accumulation to registers for each big block
317304
for (IndexType bigBlock = 0; bigBlock < NumBigBlocks; ++bigBlock)
318305
{
@@ -377,26 +364,28 @@ namespace Stockfish::Eval::NNUE::Layers {
377364
using BiasType = OutputType;
378365
using WeightType = std::int8_t;
379366

380-
PreviousLayer previousLayer;
381-
382367
alignas(CacheLineSize) BiasType biases[OutputDimensions];
383368
alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
384369
};
385370

386-
template <typename PreviousLayer, IndexType OutDims>
387-
class AffineTransform<PreviousLayer, OutDims, std::enable_if_t<(PreviousLayer::OutputDimensions < 2*64-1)>> {
371+
template <IndexType InDims, IndexType OutDims>
372+
class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) < 2*64)>> {
388373
public:
389374
// Input/output type
390-
using InputType = typename PreviousLayer::OutputType;
375+
// Input/output type
376+
using InputType = std::uint8_t;
391377
using OutputType = std::int32_t;
392-
static_assert(std::is_same<InputType, std::uint8_t>::value, "");
393378

394379
// Number of input/output dimensions
395-
static constexpr IndexType InputDimensions =
396-
PreviousLayer::OutputDimensions;
380+
static constexpr IndexType InputDimensions = InDims;
397381
static constexpr IndexType OutputDimensions = OutDims;
382+
398383
static constexpr IndexType PaddedInputDimensions =
399-
ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
384+
ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
385+
static constexpr IndexType PaddedOutputDimensions =
386+
ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);
387+
388+
using OutputBuffer = OutputType[PaddedOutputDimensions];
400389

401390
static_assert(PaddedInputDimensions < 128, "Something went wrong. This specialization should not have been chosen.");
402391

@@ -405,20 +394,12 @@ namespace Stockfish::Eval::NNUE::Layers {
405394
static constexpr const IndexType InputSimdWidth = SimdWidth;
406395
#endif
407396

408-
// Size of forward propagation buffer used in this layer
409-
static constexpr std::size_t SelfBufferSize =
410-
ceil_to_multiple(OutputDimensions * sizeof(OutputType), CacheLineSize);
411-
412-
// Size of the forward propagation buffer used from the input layer to this layer
413-
static constexpr std::size_t BufferSize =
414-
PreviousLayer::BufferSize + SelfBufferSize;
415-
416397
// Hash value embedded in the evaluation file
417-
static constexpr std::uint32_t get_hash_value() {
398+
static constexpr std::uint32_t get_hash_value(std::uint32_t prevhash) {
418399
std::uint32_t hashValue = 0xCC03DAE4u;
419400
hashValue += OutputDimensions;
420-
hashValue ^= PreviousLayer::get_hash_value() >> 1;
421-
hashValue ^= PreviousLayer::get_hash_value() << 31;
401+
hashValue ^= prevhash >> 1;
402+
hashValue ^= prevhash << 31;
422403
return hashValue;
423404
}
424405

@@ -441,7 +422,6 @@ namespace Stockfish::Eval::NNUE::Layers {
441422

442423
// Read network parameters
443424
bool read_parameters(std::istream& stream) {
444-
if (!previousLayer.read_parameters(stream)) return false;
445425
for (std::size_t i = 0; i < OutputDimensions; ++i)
446426
biases[i] = read_little_endian<BiasType>(stream);
447427
for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
@@ -452,7 +432,6 @@ namespace Stockfish::Eval::NNUE::Layers {
452432

453433
// Write network parameters
454434
bool write_parameters(std::ostream& stream) const {
455-
if (!previousLayer.write_parameters(stream)) return false;
456435
for (std::size_t i = 0; i < OutputDimensions; ++i)
457436
write_little_endian<BiasType>(stream, biases[i]);
458437

@@ -463,10 +442,7 @@ namespace Stockfish::Eval::NNUE::Layers {
463442
}
464443
// Forward propagation
465444
const OutputType* propagate(
466-
const TransformedFeatureType* transformedFeatures, char* buffer) const {
467-
const auto input = previousLayer.propagate(
468-
transformedFeatures, buffer + SelfBufferSize);
469-
const auto output = reinterpret_cast<OutputType*>(buffer);
445+
const InputType* input, OutputType* output) const {
470446

471447
#if defined (USE_AVX2)
472448
using vec_t = __m256i;
@@ -491,12 +467,11 @@ namespace Stockfish::Eval::NNUE::Layers {
491467
#if defined (USE_SSSE3)
492468
const auto inputVector = reinterpret_cast<const vec_t*>(input);
493469

494-
static_assert(InputDimensions % 8 == 0);
495470
static_assert(OutputDimensions % OutputSimdWidth == 0 || OutputDimensions == 1);
496471

497472
if constexpr (OutputDimensions % OutputSimdWidth == 0)
498473
{
499-
constexpr IndexType NumChunks = InputDimensions / 4;
474+
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / 4;
500475
constexpr IndexType NumRegs = OutputDimensions / OutputSimdWidth;
501476

502477
const auto input32 = reinterpret_cast<const std::int32_t*>(input);
@@ -555,8 +530,6 @@ namespace Stockfish::Eval::NNUE::Layers {
555530
using BiasType = OutputType;
556531
using WeightType = std::int8_t;
557532

558-
PreviousLayer previousLayer;
559-
560533
alignas(CacheLineSize) BiasType biases[OutputDimensions];
561534
alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
562535
};

‎src/nnue/layers/clipped_relu.h‎

Lines changed: 13 additions & 26 deletions
Original file line numberDiff line numberDiff line change
@@ -26,51 +26,41 @@
2626
namespace Stockfish::Eval::NNUE::Layers {
2727

2828
// Clipped ReLU
29-
template <typename PreviousLayer>
29+
template <IndexType InDims>
3030
class ClippedReLU {
3131
public:
3232
// Input/output type
33-
using InputType = typename PreviousLayer::OutputType;
33+
using InputType = std::int32_t;
3434
using OutputType = std::uint8_t;
35-
static_assert(std::is_same<InputType, std::int32_t>::value, "");
3635

3736
// Number of input/output dimensions
38-
static constexpr IndexType InputDimensions = PreviousLayer::OutputDimensions;
37+
static constexpr IndexType InputDimensions = InDims;
3938
static constexpr IndexType OutputDimensions = InputDimensions;
4039
static constexpr IndexType PaddedOutputDimensions =
4140
ceil_to_multiple<IndexType>(OutputDimensions, 32);
4241

43-
// Size of forward propagation buffer used in this layer
44-
static constexpr std::size_t SelfBufferSize =
45-
ceil_to_multiple(OutputDimensions * sizeof(OutputType), CacheLineSize);
46-
47-
// Size of the forward propagation buffer used from the input layer to this layer
48-
static constexpr std::size_t BufferSize =
49-
PreviousLayer::BufferSize + SelfBufferSize;
42+
using OutputBuffer = OutputType[PaddedOutputDimensions];
5043

5144
// Hash value embedded in the evaluation file
52-
static constexpr std::uint32_t get_hash_value() {
53-
std::uint32_t hashValue = 0x538D24C7u;
54-
hashValue += PreviousLayer::get_hash_value();
55-
return hashValue;
45+
static constexpr std::uint32_t get_hash_value(std::uint32_t prevhash) {
46+
std::uint32_t hash_value = 0x538D24C7u;
47+
hash_value += prevhash;
48+
return hash_value;
5649
}
5750

5851
// Read network parameters
59-
bool read_parameters(std::istream& stream) {
60-
return previousLayer.read_parameters(stream);
52+
bool read_parameters(std::istream&) {
53+
return true;
6154
}
6255

6356
// Write network parameters
64-
bool write_parameters(std::ostream& stream) const {
65-
return previousLayer.write_parameters(stream);
57+
bool write_parameters(std::ostream&) const {
58+
return true;
6659
}
6760

6861
// Forward propagation
6962
const OutputType* propagate(
70-
const TransformedFeatureType* transformedFeatures, char* buffer) const {
71-
const auto input = previousLayer.propagate(
72-
transformedFeatures, buffer + SelfBufferSize);
73-
const auto output = reinterpret_cast<OutputType*>(buffer);
63+
const InputType* input, OutputType* output) const {
7464

7565
#if defined(USE_AVX2)
7666
if constexpr (InputDimensions % SimdWidth == 0) {
@@ -191,9 +181,6 @@ namespace Stockfish::Eval::NNUE::Layers {
191181

192182
return output;
193183
}
194-
195-
private:
196-
PreviousLayer previousLayer;
197184
};
198185

199186
} // namespace Stockfish::Eval::NNUE::Layers

0 commit comments

Comments
 (0)