Line data    Source code

       1              : #pragma once
       2              : 
       3              : #include "definition.hpp"
       4              : 
       5              : #ifdef WITH_NNUE
       6              : 
       7              : #include "quantization.hpp"
       8              : #include "stackVector.hpp"
       9              : #include "weightReader.hpp"
      10              : 
      11              : constexpr size_t inputLayerSize      = FeatureIdx::major * FeatureIdx::minor;
      12              : constexpr size_t firstInnerLayerSize = 384;
      13              : 
      14              : template<typename NT, size_t dim0, size_t dim1, bool Q> 
      15              : struct InputLayer {
      16              :    static constexpr size_t nbW = dim0 * dim1;
      17              :    static constexpr size_t nbB = dim1;
      18              : 
      19              :    using BIT = typename Quantization<Q>::BIT;
      20              :    using WIT = typename Quantization<Q>::WIT;
      21              : 
      22              :    // often too big to be statically allocated (see CTOR/DTOR for dynamic alloc)
      23              :    typename Quantization<Q>::WIT* W {nullptr};
      24              : 
      25              :    // bias can be statically allocated 
      26              :    alignas(NNUEALIGNMENT) BIT b[nbB];
      27              : 
      28              :    FORCE_FINLINE void insertIdx(const size_t idx, StackVector<BIT, nbB, Q>& x) const {
      29    324993060 :       const WIT* wPtr = W + idx * dim1;
      30    517499686 :       x.add_(wPtr);
      31              :    }
      32              : 
      33              :    FORCE_FINLINE void eraseIdx(const size_t idx, StackVector<BIT, nbB, Q>& x) const {
      34     17744168 :       const WIT* wPtr = W + idx * dim1;
      35    342737228 :       x.sub_(wPtr);
      36              :    }
      37              : 
      38           44 :    InputLayer<NT, dim0, dim1, Q>& load_(WeightsReader<NT>& ws) {
      39           44 :       ws.template streamWI<WIT, Q>(W, nbW)
      40           44 :         .template streamBI<BIT, Q>(b, nbB);
      41           44 :       return *this;
      42              :    }
      43              : 
      44              :    // non copyable
      45              :    InputLayer<NT, dim0, dim1, Q>& operator=(const InputLayer<NT, dim0, dim1, Q>& other) = delete;
      46              :    InputLayer<NT, dim0, dim1, Q>& operator=(InputLayer<NT, dim0, dim1, Q>&& other) = delete;
      47              :    InputLayer(const InputLayer<NT, dim0, dim1, Q>& other) = delete;
      48              :    InputLayer(InputLayer<NT, dim0, dim1, Q>&& other) = delete;
      49              : 
      50              :    InputLayer() { 
      51              :        W = static_cast<WIT*>(operator new[](sizeof(WIT) * nbW, NNUEALIGNMENT_STD));
      52              :     }
      53              : 
      54              :    ~InputLayer() {
      55              :        ::operator delete(W, NNUEALIGNMENT_STD);
      56              :    }
      57              : };
      58              : 
      59              : #endif // WITH_NNUE