/*****************************************************************************
 * This file is part of Kvazaar HEVC encoder.
 *
 * Copyright (C) 2013-2015 Tampere University of Technology and others (see
 * COPYING file).
 *
 * Kvazaar is free software: you can redistribute it and/or modify it under
 * the terms of the GNU Lesser General Public License as published by the
 * Free Software Foundation; either version 2.1 of the License, or (at your
 * option) any later version.
 *
 * Kvazaar 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 Lesser General Public License for
 * more details.
 *
 * You should have received a copy of the GNU General Public License along
 * with Kvazaar.  If not, see <http://www.gnu.org/licenses/>.
 ****************************************************************************/

/*
 * \file
 */

#include <stdlib.h>

#include "intra-avx2.h"
#include "strategyselector.h"

#if COMPILE_INTEL_AVX2
#include <immintrin.h>

 /**
 * \brief Generage angular predictions.
 * \param log2_width    Log2 of width, range 2..5.
 * \param intra_mode    Angular mode in range 2..34.
 * \param in_ref_above  Pointer to -1 index of above reference, length=width*2+1.
 * \param in_ref_left   Pointer to -1 index of left reference, length=width*2+1.
 * \param dst           Buffer of size width*width.
 */
static void kvz_angular_pred_avx2(
  const int_fast8_t log2_width,
  const int_fast8_t intra_mode,
  const kvz_pixel *const in_ref_above,
  const kvz_pixel *const in_ref_left,
  kvz_pixel *const dst)
{
  assert(log2_width >= 2 && log2_width <= 5);
  assert(intra_mode >= 2 && intra_mode <= 34);

  static const int8_t modedisp2sampledisp[9] = { 0, 2, 5, 9, 13, 17, 21, 26, 32 };
  static const int16_t modedisp2invsampledisp[9] = { 0, 4096, 1638, 910, 630, 482, 390, 315, 256 }; // (256 * 32) / sampledisp

                                                    // Temporary buffer for modes 11-25.
                                                    // It only needs to be big enough to hold indices from -width to width-1.
  kvz_pixel tmp_ref[2 * 32];
  const int_fast8_t width = 1 << log2_width;

  // Whether to swap references to always project on the left reference row.
  const bool vertical_mode = intra_mode >= 18;
  // Modes distance to horizontal or vertical mode.
  const int_fast8_t mode_disp = vertical_mode ? intra_mode - 26 : 10 - intra_mode;
  // Sample displacement per column in fractions of 32.
  const int_fast8_t sample_disp = (mode_disp < 0 ? -1 : 1) * modedisp2sampledisp[abs(mode_disp)];

  // Pointer for the reference we are interpolating from.
  const kvz_pixel *ref_main;
  // Pointer for the other reference.
  const kvz_pixel *ref_side;

  // Set ref_main and ref_side such that, when indexed with 0, they point to
  // index 0 in block coordinates.
  if (sample_disp < 0) {
    // Negative sample_disp means, we need to use both references.

    ref_side = (vertical_mode ? in_ref_left : in_ref_above) + 1;
    ref_main = (vertical_mode ? in_ref_above : in_ref_left) + 1;

    // Move the reference pixels to start from the middle to the later half of
    // the tmp_ref, so there is room for negative indices.
    for (int_fast8_t x = -1; x < width; ++x) {
      tmp_ref[x + width] = ref_main[x];
    }
    // Get a pointer to block index 0 in tmp_ref.
    ref_main = &tmp_ref[width];

    // Extend the side reference to the negative indices of main reference.
    int_fast32_t col_sample_disp = 128; // rounding for the ">> 8"
    int_fast16_t inv_abs_sample_disp = modedisp2invsampledisp[abs(mode_disp)];
    int_fast8_t most_negative_index = (width * sample_disp) >> 5;
    for (int_fast8_t x = -2; x >= most_negative_index; --x) {
      col_sample_disp += inv_abs_sample_disp;
      int_fast8_t side_index = col_sample_disp >> 8;
      tmp_ref[x + width] = ref_side[side_index - 1];
    }
  }
  else {
    // sample_disp >= 0 means we don't need to refer to negative indices,
    // which means we can just use the references as is.
    ref_main = (vertical_mode ? in_ref_above : in_ref_left) + 1;
    ref_side = (vertical_mode ? in_ref_left : in_ref_above) + 1;
  }

  if (sample_disp != 0) {
    // The mode is not horizontal or vertical, we have to do interpolation.

    int_fast16_t delta_pos = 0;
    for (int_fast8_t y = 0; y < width; ++y) {
      delta_pos += sample_disp;
      int_fast8_t delta_int = delta_pos >> 5;
      int_fast8_t delta_fract = delta_pos & (32 - 1);

      if (delta_fract) {
        // Do linear filtering
        if (width < 8) {
          for (int_fast8_t x = 0; x < width; ++x) {
            kvz_pixel ref1 = ref_main[x + delta_int];
            kvz_pixel ref2 = ref_main[x + delta_int + 1];
            dst[y * width + x] = ((32 - delta_fract) * ref1 + delta_fract * ref2 + 16) >> 5;
          }
        } else {
          struct { uint8_t w1; uint8_t w2; } packed_weights = { 32 - delta_fract, delta_fract };
          __m128i v_weights = _mm_set1_epi16(*(int16_t*)&packed_weights);

          for (int_fast8_t x = 0; x < width; x += 8) {
            __m128i v_ref1 = _mm_loadl_epi64((__m128i*)&(ref_main[x + delta_int]));
            __m128i v_ref2 = _mm_loadl_epi64((__m128i*)&(ref_main[x + delta_int + 1]));
            __m128i v_refs = _mm_unpacklo_epi8(v_ref1, v_ref2);  
            __m128i v_tmp = _mm_maddubs_epi16(v_refs, v_weights);
            v_tmp = _mm_add_epi16(v_tmp, _mm_set1_epi16(16));
            v_tmp = _mm_srli_epi16(v_tmp, 5);
            v_tmp = _mm_packus_epi16(v_tmp, v_tmp);
            _mm_storel_epi64((__m128i*)(dst + y * width + x), v_tmp);
          }
        }
      }
      else {
        // Just copy the integer samples
        for (int_fast8_t x = 0; x < width; x+=4) {
          *(int32_t*)(&dst[y * width + x]) = *(int32_t*)(&ref_main[x + delta_int]);
        }
      }
    }
  }
  else {
    // Mode is horizontal or vertical, just copy the pixels.

    for (int_fast8_t y = 0; y < width; ++y) {
      for (int_fast8_t x = 0; x < width; x+=4) {
        *(int32_t*)&(dst[y * width + x]) = *(int32_t*)&(ref_main[x]);
      }
    }
  }

  // Flip the block if this is was a horizontal mode.
  if (!vertical_mode) {
    for (int_fast8_t y = 0; y < width - 1; ++y) {
      for (int_fast8_t x = y + 1; x < width; ++x) {
        SWAP(dst[y * width + x], dst[x * width + y], kvz_pixel);
      }
    }
  }
}

#endif //COMPILE_INTEL_AVX2

int kvz_strategy_register_intra_avx2(void* opaque, uint8_t bitdepth)
{
  bool success = true;
#if COMPILE_INTEL_AVX2
  if (bitdepth == 8) {
    success &= kvz_strategyselector_register(opaque, "angular_pred", "avx2", 40, &kvz_angular_pred_avx2);
  }
#endif //COMPILE_INTEL_AVX2
  return success;
}