/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
- Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file)
+ Copyright (C) 2004-2022 The Stockfish developers (see AUTHORS file)
Stockfish is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
namespace Stockfish::Eval::NNUE::Layers {
// Clipped ReLU
- template <typename PreviousLayer>
+ template <IndexType InDims>
class ClippedReLU {
public:
// Input/output type
- using InputType = typename PreviousLayer::OutputType;
+ using InputType = std::int32_t;
using OutputType = std::uint8_t;
- static_assert(std::is_same<InputType, std::int32_t>::value, "");
// Number of input/output dimensions
- static constexpr IndexType InputDimensions =
- PreviousLayer::OutputDimensions;
+ static constexpr IndexType InputDimensions = InDims;
static constexpr IndexType OutputDimensions = InputDimensions;
+ static constexpr IndexType PaddedOutputDimensions =
+ ceil_to_multiple<IndexType>(OutputDimensions, 32);
- // Size of forward propagation buffer used in this layer
- static constexpr std::size_t SelfBufferSize =
- ceil_to_multiple(OutputDimensions * sizeof(OutputType), CacheLineSize);
-
- // Size of the forward propagation buffer used from the input layer to this layer
- static constexpr std::size_t BufferSize =
- PreviousLayer::BufferSize + SelfBufferSize;
+ using OutputBuffer = OutputType[PaddedOutputDimensions];
// Hash value embedded in the evaluation file
- static constexpr std::uint32_t get_hash_value() {
+ static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
std::uint32_t hashValue = 0x538D24C7u;
- hashValue += PreviousLayer::get_hash_value();
+ hashValue += prevHash;
return hashValue;
}
// Read network parameters
- bool read_parameters(std::istream& stream) {
- return previousLayer.read_parameters(stream);
+ bool read_parameters(std::istream&) {
+ return true;
+ }
+
+ // Write network parameters
+ bool write_parameters(std::ostream&) const {
+ return true;
}
// Forward propagation
const OutputType* propagate(
- const TransformedFeatureType* transformedFeatures, char* buffer) const {
- const auto input = previousLayer.propagate(
- transformedFeatures, buffer + SelfBufferSize);
- const auto output = reinterpret_cast<OutputType*>(buffer);
+ const InputType* input, OutputType* output) const {
#if defined(USE_AVX2)
- constexpr IndexType NumChunks = InputDimensions / SimdWidth;
- const __m256i Zero = _mm256_setzero_si256();
- const __m256i Offsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
- const auto in = reinterpret_cast<const __m256i*>(input);
- const auto out = reinterpret_cast<__m256i*>(output);
- for (IndexType i = 0; i < NumChunks; ++i) {
- const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
- _mm256_load_si256(&in[i * 4 + 0]),
- _mm256_load_si256(&in[i * 4 + 1])), WeightScaleBits);
- const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
- _mm256_load_si256(&in[i * 4 + 2]),
- _mm256_load_si256(&in[i * 4 + 3])), WeightScaleBits);
- _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8(
- _mm256_packs_epi16(words0, words1), Zero), Offsets));
+ if constexpr (InputDimensions % SimdWidth == 0) {
+ constexpr IndexType NumChunks = InputDimensions / SimdWidth;
+ const __m256i Zero = _mm256_setzero_si256();
+ const __m256i Offsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
+ const auto in = reinterpret_cast<const __m256i*>(input);
+ const auto out = reinterpret_cast<__m256i*>(output);
+ for (IndexType i = 0; i < NumChunks; ++i) {
+ const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
+ _mm256_load_si256(&in[i * 4 + 0]),
+ _mm256_load_si256(&in[i * 4 + 1])), WeightScaleBits);
+ const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
+ _mm256_load_si256(&in[i * 4 + 2]),
+ _mm256_load_si256(&in[i * 4 + 3])), WeightScaleBits);
+ _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8(
+ _mm256_packs_epi16(words0, words1), Zero), Offsets));
+ }
+ } else {
+ constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
+ const __m128i Zero = _mm_setzero_si128();
+ const auto in = reinterpret_cast<const __m128i*>(input);
+ const auto out = reinterpret_cast<__m128i*>(output);
+ for (IndexType i = 0; i < NumChunks; ++i) {
+ const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
+ _mm_load_si128(&in[i * 4 + 0]),
+ _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits);
+ const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
+ _mm_load_si128(&in[i * 4 + 2]),
+ _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits);
+ const __m128i packedbytes = _mm_packs_epi16(words0, words1);
+ _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero));
+ }
}
- constexpr IndexType Start = NumChunks * SimdWidth;
+ constexpr IndexType Start =
+ InputDimensions % SimdWidth == 0
+ ? InputDimensions / SimdWidth * SimdWidth
+ : InputDimensions / (SimdWidth / 2) * (SimdWidth / 2);
#elif defined(USE_SSE2)
constexpr IndexType NumChunks = InputDimensions / SimdWidth;
output[i] = static_cast<OutputType>(
std::max(0, std::min(127, input[i] >> WeightScaleBits)));
}
+
+ // Affine transform layers expect that there is at least
+ // ceil_to_multiple(OutputDimensions, 32) initialized values.
+ // We cannot do this in the affine transform because it requires
+ // preallocating space here.
+ for (IndexType i = OutputDimensions; i < PaddedOutputDimensions; ++i) {
+ output[i] = 0;
+ }
+
return output;
}
-
- private:
- PreviousLayer previousLayer;
};
} // namespace Stockfish::Eval::NNUE::Layers