/*
  This file is part of ctak.

  This program is free software: you can redistribute it and/or modify
  it under the terms of the GNU General Public License as published by
  the Free Software Foundation, either version 3 of the License, or
  (at your option) any later version.

  This program is distributed in the hope that it will be useful, but
  WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  GNU General Public License for more details.

  You should have received a copy of the GNU General Public License
  along with ctak. If not, see <https://www.gnu.org/licenses/>.
*/

#include "nn.h"

#include "weights_5.h"
#include "weights_6.h"

// ===================================================================
// Helpers
// ===================================================================

#define RELU(x) ((x) = ((x) < 0) ? 0 : (x))

static float fast_tanh(const float x) {
  // tanh via a degree-11 odd polynomial, least-squares fit of tanh on [-3, 3]
  //  (max absolute error ~4e-3), saturating to +-1 outside that interval.
  if (x > 3.0f)
    return 1.0f;
  if (x < -3.0f)
    return -1.0f;
  const float x2 = x * x;
  return x * (0.997136032f + x2 * (-0.305215124f + x2 *
                                   (0.0832466371f + x2 *
                                    (-0.0139947924f + x2 *
                                     (0.00123106314f + x2 * -4.29748644e-05f)))));
}
// ===================================================================
// Input encoding
// ===================================================================

inline uint8_t cell_idx_at(const tak_state_p state, const uint8_t loc) {
  const uint8_t d = state->celldat[loc];
  if ((d & NUM_MASK) == 0)
    return 0; // empty
  return (uint8_t)(1u + 3u * (state->colours[loc] & 1u) + (d & STONE_MASK));
}

inline void nn_cell_idxs(uint8_t *out, tak_state_p state) {
  const unsigned int n_cells = state->board_size * state->board_size;
  for (unsigned int loc = 0; loc < n_cells; loc++)
    out[loc] = cell_idx_at(state, loc);
}

inline void nn_prepare_head(float out_head[3], tak_state_p state) {
  const float inv_max_flats =
    (state->board_size == 6) ? (1.0f / 30.0f) : (1.0f / 21.0f);
  out_head[0] = (state->ply & 1) ? 1 : -1;
  out_head[1] = (float)(state->white_count & 127) * inv_max_flats;
  out_head[2] = (float)(state->black_count & 127) * inv_max_flats;
}

void nn_prepare_input(float *cur_board, tak_state_p state,
                      const float embedding[NN_N_EMB]) {
  const unsigned int n_cells = state->board_size * state->board_size;
  nn_prepare_head(cur_board, state);
  uint8_t idxs[36];
  nn_cell_idxs(idxs, state);
  for (unsigned int loc = 0; loc < n_cells; loc++)
    cur_board[3 + loc] = embedding[idxs[loc]];
}

// ===================================================================
// Implementation of small neural networks
// ===================================================================

float eval_net(const float *board, const unsigned int n_inp,
               const float *w, const float *biases,
               const float *logit_w, const float logit_b,
               const unsigned int n_dense) {
  float dense1[FIVE_DENSE_NUM]; // FIVE_DENSE_NUM == SIX_DENSE_NUM
  float z = logit_b;

  for (unsigned int d1 = 0; d1 < n_dense; d1++) {
    float h = biases[d1];
    const float *wd = w + d1 * n_inp;
    for (unsigned int fl = 0; fl < n_inp; fl++)
      h += board[fl] * wd[fl];
    RELU(h);
    dense1[d1] = h;
  }
  for (unsigned int d1 = 0; d1 < n_dense; d1++)
    z += dense1[d1] * logit_w[d1];

  return fast_tanh(z);
}

// ===================================================================
// Interfaces to the two built-in evaluators
// ===================================================================

float five_nn_fn(tak_state_p state) {
  static float cur_board[FIVE_INP_NUM] NN_ALIGNED;

  nn_prepare_input(cur_board, state, five_embedding);
  return eval_net(cur_board, FIVE_INP_NUM, &five_dense_weights[0][0],
                  five_dense_biases, five_logit_weights, five_logit_bias,
                  FIVE_DENSE_NUM);
}

float six_nn_fn(tak_state_p state) {
  static float cur_board[SIX_INP_NUM] NN_ALIGNED;

  nn_prepare_input(cur_board, state, six_embedding);
  return eval_net(cur_board, SIX_INP_NUM, &six_dense_weights[0][0],
                  six_dense_biases, six_logit_weights, six_logit_bias,
                  SIX_DENSE_NUM);
}