/*
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <assert.h>
#include <limits.h>
#include <stdio.h>

#include "./vp10_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "./vpx_scale_rtcd.h"

#include "vpx_mem/vpx_mem.h"
#include "vpx_ports/system_state.h"
#include "vpx_ports/vpx_once.h"
#include "vpx_ports/vpx_timer.h"
#include "vpx_scale/vpx_scale.h"
#include "vpx_util/vpx_thread.h"

#include "vp10/common/alloccommon.h"
#include "vp10/common/loopfilter.h"
#include "vp10/common/onyxc_int.h"
#if CONFIG_VP9_POSTPROC
#include "vp10/common/postproc.h"
#endif
#include "vp10/common/quant_common.h"
#include "vp10/common/reconinter.h"
#include "vp10/common/reconintra.h"

#include "vp10/decoder/decodeframe.h"
#include "vp10/decoder/decoder.h"
#include "vp10/decoder/detokenize.h"

static void initialize_dec(void) {
  static volatile int init_done = 0;

  if (!init_done) {
    vp10_rtcd();
    vpx_dsp_rtcd();
    vpx_scale_rtcd();
    vp10_init_intra_predictors();
#if CONFIG_EXT_INTER
    vp10_init_wedge_masks();
#endif  // CONFIG_EXT_INTER
    init_done = 1;
  }
}

static void vp10_dec_setup_mi(VP10_COMMON *cm) {
  cm->mi = cm->mip + cm->mi_stride + 1;
  cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
  memset(cm->mi_grid_base, 0,
         cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
}

static int vp10_dec_alloc_mi(VP10_COMMON *cm, int mi_size) {
  cm->mip = vpx_calloc(mi_size, sizeof(*cm->mip));
  if (!cm->mip)
    return 1;
  cm->mi_alloc_size = mi_size;
  cm->mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
  if (!cm->mi_grid_base)
    return 1;
  return 0;
}

static void vp10_dec_free_mi(VP10_COMMON *cm) {
  vpx_free(cm->mip);
  cm->mip = NULL;
  vpx_free(cm->mi_grid_base);
  cm->mi_grid_base = NULL;
}

VP10Decoder *vp10_decoder_create(BufferPool *const pool) {
  VP10Decoder *volatile const pbi = vpx_memalign(32, sizeof(*pbi));
  VP10_COMMON *volatile const cm = pbi ? &pbi->common : NULL;

  if (!cm)
    return NULL;

  vp10_zero(*pbi);

  if (setjmp(cm->error.jmp)) {
    cm->error.setjmp = 0;
    vp10_decoder_remove(pbi);
    return NULL;
  }

  cm->error.setjmp = 1;

  CHECK_MEM_ERROR(cm, cm->fc,
                  (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc)));
  CHECK_MEM_ERROR(cm, cm->frame_contexts,
                  (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS,
                  sizeof(*cm->frame_contexts)));

  pbi->need_resync = 1;
  once(initialize_dec);

  // Initialize the references to not point to any frame buffers.
  memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
  memset(&cm->next_ref_frame_map, -1, sizeof(cm->next_ref_frame_map));

  cm->current_video_frame = 0;
  pbi->ready_for_new_data = 1;
  pbi->common.buffer_pool = pool;

  cm->bit_depth = VPX_BITS_8;
  cm->dequant_bit_depth = VPX_BITS_8;

  cm->alloc_mi = vp10_dec_alloc_mi;
  cm->free_mi = vp10_dec_free_mi;
  cm->setup_mi = vp10_dec_setup_mi;

  vp10_loop_filter_init(cm);
#if CONFIG_LOOP_RESTORATION
  vp10_loop_restoration_precal();
#endif  // CONFIG_LOOP_RESTORATION

  cm->error.setjmp = 0;

  vpx_get_worker_interface()->init(&pbi->lf_worker);

  return pbi;
}

void vp10_decoder_remove(VP10Decoder *pbi) {
  int i;

  if (!pbi)
    return;

  vpx_get_worker_interface()->end(&pbi->lf_worker);
  vpx_free(pbi->lf_worker.data1);
  vpx_free(pbi->tile_data);
  for (i = 0; i < pbi->num_tile_workers; ++i) {
    VPxWorker *const worker = &pbi->tile_workers[i];
    vpx_get_worker_interface()->end(worker);
  }
  vpx_free(pbi->tile_worker_data);
  vpx_free(pbi->tile_worker_info);
  vpx_free(pbi->tile_workers);

  if (pbi->num_tile_workers > 0) {
    vp10_loop_filter_dealloc(&pbi->lf_row_sync);
  }

  vpx_free(pbi);
}

static int equal_dimensions(const YV12_BUFFER_CONFIG *a,
                            const YV12_BUFFER_CONFIG *b) {
    return a->y_height == b->y_height && a->y_width == b->y_width &&
           a->uv_height == b->uv_height && a->uv_width == b->uv_width;
}

vpx_codec_err_t vp10_copy_reference_dec(VP10Decoder *pbi,
                                       VP9_REFFRAME ref_frame_flag,
                                       YV12_BUFFER_CONFIG *sd) {
  VP10_COMMON *cm = &pbi->common;

  /* TODO(jkoleszar): The decoder doesn't have any real knowledge of what the
   * encoder is using the frame buffers for. This is just a stub to keep the
   * vpxenc --test-decode functionality working, and will be replaced in a
   * later commit that adds VP9-specific controls for this functionality.
   */
  if (ref_frame_flag == VP9_LAST_FLAG) {
    const YV12_BUFFER_CONFIG *const cfg = get_ref_frame(cm, 0);
    if (cfg == NULL) {
      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                         "No 'last' reference frame");
      return VPX_CODEC_ERROR;
    }
    if (!equal_dimensions(cfg, sd))
      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                         "Incorrect buffer dimensions");
    else
      vp8_yv12_copy_frame(cfg, sd);
  } else {
    vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                       "Invalid reference frame");
  }

  return cm->error.error_code;
}


vpx_codec_err_t vp10_set_reference_dec(VP10_COMMON *cm,
                                      VP9_REFFRAME ref_frame_flag,
                                      YV12_BUFFER_CONFIG *sd) {
  int idx;
  YV12_BUFFER_CONFIG *ref_buf = NULL;

  // TODO(jkoleszar): The decoder doesn't have any real knowledge of what the
  // encoder is using the frame buffers for. This is just a stub to keep the
  // vpxenc --test-decode functionality working, and will be replaced in a
  // later commit that adds VP9-specific controls for this functionality.

  // (Yunqing) The set_reference control depends on the following setting in
  // encoder.
  //   cpi->lst_fb_idx = 0;
  // #if CONFIG_EXT_REFS
  //   cpi->lst2_fb_idx = 1;
  //   cpi->lst3_fb_idx = 2;
  //   cpi->lst4_fb_idx = 3;
  //   cpi->gld_fb_idx = 4;
  //   cpi->alt_fb_idx = 5;
  // #else  // CONFIG_EXT_REFS
  //   cpi->gld_fb_idx = 1;
  //   cpi->alt_fb_idx = 2;

  // TODO(zoeliu): To revisit following code and reconsider what assumption we
  // may take on the reference frame buffer virtual indexes
  if (ref_frame_flag == VP9_LAST_FLAG) {
    idx = cm->ref_frame_map[0];
#if CONFIG_EXT_REFS
  } else if (ref_frame_flag == VP9_LAST2_FLAG) {
    idx = cm->ref_frame_map[1];
  } else if (ref_frame_flag == VP9_LAST3_FLAG) {
    idx = cm->ref_frame_map[2];
  } else if (ref_frame_flag == VP9_LAST4_FLAG) {
    idx = cm->ref_frame_map[3];
  } else if (ref_frame_flag == VP9_GOLD_FLAG) {
    idx = cm->ref_frame_map[4];
  } else if (ref_frame_flag == VP9_ALT_FLAG) {
    idx = cm->ref_frame_map[5];
#else  // CONFIG_EXT_REFS
  } else if (ref_frame_flag == VP9_GOLD_FLAG) {
    idx = cm->ref_frame_map[1];
#if CONFIG_BIDIR_PRED
  } else if (ref_frame_flag == VP9_BWD_FLAG) {
    idx = cm->ref_frame_map[2];
  } else if (ref_frame_flag == VP9_ALT_FLAG) {
    idx = cm->ref_frame_map[3];
#else  // CONFIG_BIDIR_PRED
  } else if (ref_frame_flag == VP9_ALT_FLAG) {
    idx = cm->ref_frame_map[2];
#endif  // CONFIG_BIDIR_PRED
#endif  // CONFIG_EXT_REFS
  } else {
    vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                       "Invalid reference frame");
    return cm->error.error_code;
  }

  if (idx < 0 || idx >= FRAME_BUFFERS) {
    vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                       "Invalid reference frame map");
    return cm->error.error_code;
  }

  // Get the destination reference buffer.
  ref_buf = &cm->buffer_pool->frame_bufs[idx].buf;

  if (!equal_dimensions(ref_buf, sd)) {
    vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                       "Incorrect buffer dimensions");
  } else {
    // Overwrite the reference frame buffer.
    vp8_yv12_copy_frame(sd, ref_buf);
  }

  return cm->error.error_code;
}

/* If any buffer updating is signaled it should be done here. */
static void swap_frame_buffers(VP10Decoder *pbi) {
  int ref_index = 0, mask;
  VP10_COMMON *const cm = &pbi->common;
  BufferPool *const pool = cm->buffer_pool;
  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;

  lock_buffer_pool(pool);
  for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
    const int old_idx = cm->ref_frame_map[ref_index];
    // Current thread releases the holding of reference frame.
    decrease_ref_count(old_idx, frame_bufs, pool);

    // Release the reference frame holding in the reference map for the decoding
    // of the next frame.
    if (mask & 1)
      decrease_ref_count(old_idx, frame_bufs, pool);
    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
    ++ref_index;
  }

  // Current thread releases the holding of reference frame.
#if !CONFIG_EXT_REFS && CONFIG_BIDIR_PRED
  for (; ref_index < REF_FRAMES; ++ref_index) {
    const int old_idx = cm->ref_frame_map[ref_index];
    decrease_ref_count(old_idx, frame_bufs, pool);
    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
  }
#else
  for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
    const int old_idx = cm->ref_frame_map[ref_index];
    decrease_ref_count(old_idx, frame_bufs, pool);
    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
  }
#endif  // !CONFIG_EXT_REFS && CONFIG_BIDIR_PRED
  unlock_buffer_pool(pool);
  pbi->hold_ref_buf = 0;
  cm->frame_to_show = get_frame_new_buffer(cm);

  // TODO(zoeliu): To fix the ref frame buffer update for the scenario of
  //               cm->frame_parellel_decode == 1
  if (!cm->frame_parallel_decode || !cm->show_frame) {
    lock_buffer_pool(pool);
    --frame_bufs[cm->new_fb_idx].ref_count;
    unlock_buffer_pool(pool);
  }

  // Invalidate these references until the next frame starts.
  for (ref_index = 0; ref_index < REFS_PER_FRAME; ref_index++) {
    cm->frame_refs[ref_index].idx = INVALID_IDX;
    cm->frame_refs[ref_index].buf = NULL;
  }
}

int vp10_receive_compressed_data(VP10Decoder *pbi,
                                size_t size, const uint8_t **psource) {
  VP10_COMMON *volatile const cm = &pbi->common;
  BufferPool *volatile const pool = cm->buffer_pool;
  RefCntBuffer *volatile const frame_bufs = cm->buffer_pool->frame_bufs;
  const uint8_t *source = *psource;
  int retcode = 0;
  cm->error.error_code = VPX_CODEC_OK;

  if (size == 0) {
    // This is used to signal that we are missing frames.
    // We do not know if the missing frame(s) was supposed to update
    // any of the reference buffers, but we act conservative and
    // mark only the last buffer as corrupted.
    //
    // TODO(jkoleszar): Error concealment is undefined and non-normative
    // at this point, but if it becomes so, [0] may not always be the correct
    // thing to do here.
    if (cm->frame_refs[0].idx > 0) {
      assert(cm->frame_refs[0].buf != NULL);
      cm->frame_refs[0].buf->corrupted = 1;
    }
  }

  pbi->ready_for_new_data = 0;

  // Find a free buffer for the new frame, releasing the reference previously
  // held.

  // Check if the previous frame was a frame without any references to it.
  // Release frame buffer if not decoding in frame parallel mode.
  if (!cm->frame_parallel_decode && cm->new_fb_idx >= 0
      && frame_bufs[cm->new_fb_idx].ref_count == 0)
    pool->release_fb_cb(pool->cb_priv,
                        &frame_bufs[cm->new_fb_idx].raw_frame_buffer);

  // Find a free frame buffer. Return error if can not find any.
  cm->new_fb_idx = get_free_fb(cm);
  if (cm->new_fb_idx == INVALID_IDX)
    return VPX_CODEC_MEM_ERROR;

  // Assign a MV array to the frame buffer.
  cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];

  pbi->hold_ref_buf = 0;
  if (cm->frame_parallel_decode) {
    VPxWorker *const worker = pbi->frame_worker_owner;
    vp10_frameworker_lock_stats(worker);
    frame_bufs[cm->new_fb_idx].frame_worker_owner = worker;
    // Reset decoding progress.
    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
    pbi->cur_buf->row = -1;
    pbi->cur_buf->col = -1;
    vp10_frameworker_unlock_stats(worker);
  } else {
    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
  }

  if (setjmp(cm->error.jmp)) {
    const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
    int i;

    cm->error.setjmp = 0;
    pbi->ready_for_new_data = 1;

    // Synchronize all threads immediately as a subsequent decode call may
    // cause a resize invalidating some allocations.
    winterface->sync(&pbi->lf_worker);
    for (i = 0; i < pbi->num_tile_workers; ++i) {
      winterface->sync(&pbi->tile_workers[i]);
    }

    lock_buffer_pool(pool);
    // Release all the reference buffers if worker thread is holding them.
    if (pbi->hold_ref_buf == 1) {
      int ref_index = 0, mask;
      for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
        const int old_idx = cm->ref_frame_map[ref_index];
        // Current thread releases the holding of reference frame.
        decrease_ref_count(old_idx, frame_bufs, pool);

       // Release the reference frame holding in the reference map for the
       // decoding of the next frame.
       if (mask & 1)
          decrease_ref_count(old_idx, frame_bufs, pool);
        ++ref_index;
      }

      // Current thread releases the holding of reference frame.
#if !CONFIG_EXT_REFS && CONFIG_BIDIR_PRED
      for (; ref_index < REF_FRAMES; ++ref_index) {
        const int old_idx = cm->ref_frame_map[ref_index];
        decrease_ref_count(old_idx, frame_bufs, pool);
      }
#else
      for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
        const int old_idx = cm->ref_frame_map[ref_index];
        decrease_ref_count(old_idx, frame_bufs, pool);
      }
#endif  // !CONFIG_EXT_REFS && CONFIG_BIDIR_PRED
      pbi->hold_ref_buf = 0;
    }
    // Release current frame.
    decrease_ref_count(cm->new_fb_idx, frame_bufs, pool);
    unlock_buffer_pool(pool);

    vpx_clear_system_state();
    return -1;
  }

  cm->error.setjmp = 1;
  vp10_decode_frame(pbi, source, source + size, psource);

  swap_frame_buffers(pbi);

#if CONFIG_EXT_TILE
  // For now, we only extend the frame borders when the whole frame is decoded.
  // Later, if needed, extend the border for the decoded tile on the frame
  // border.
  if (pbi->dec_tile_row == -1 && pbi->dec_tile_col == -1)
#endif  // CONFIG_EXT_TILE
    vpx_extend_frame_inner_borders(cm->frame_to_show);

  vpx_clear_system_state();

  if (!cm->show_existing_frame) {
    cm->last_show_frame = cm->show_frame;

#if !CONFIG_EXT_REFS && CONFIG_BIDIR_PRED
    // NOTE: It is not supposed to ref to any frame not used as reference
    if (cm->is_reference_frame)
#endif  // !CONFIG_EXT_REFS && CONFIG_BIDIR_PRED
      cm->prev_frame = cm->cur_frame;

    if (cm->seg.enabled && !cm->frame_parallel_decode)
      vp10_swap_current_and_last_seg_map(cm);
  }

  // Update progress in frame parallel decode.
  if (cm->frame_parallel_decode) {
    // Need to lock the mutex here as another thread may
    // be accessing this buffer.
    VPxWorker *const worker = pbi->frame_worker_owner;
    FrameWorkerData *const frame_worker_data = worker->data1;
    vp10_frameworker_lock_stats(worker);

    if (cm->show_frame) {
      cm->current_video_frame++;
    }
    frame_worker_data->frame_decoded = 1;
    frame_worker_data->frame_context_ready = 1;
    vp10_frameworker_signal_stats(worker);
    vp10_frameworker_unlock_stats(worker);
  } else {
    cm->last_width = cm->width;
    cm->last_height = cm->height;
    if (cm->show_frame) {
      cm->current_video_frame++;
    }
  }

  cm->error.setjmp = 0;
  return retcode;
}

int vp10_get_raw_frame(VP10Decoder *pbi, YV12_BUFFER_CONFIG *sd,
                      vp10_ppflags_t *flags) {
  VP10_COMMON *const cm = &pbi->common;
  int ret = -1;
#if !CONFIG_VP9_POSTPROC
  (void)*flags;
#endif

  if (pbi->ready_for_new_data == 1)
    return ret;

  pbi->ready_for_new_data = 1;

  /* no raw frame to show!!! */
  if (!cm->show_frame)
    return ret;

  pbi->ready_for_new_data = 1;

#if CONFIG_VP9_POSTPROC
  if (!cm->show_existing_frame) {
    ret = vp10_post_proc_frame(cm, sd, flags);
  } else {
    *sd = *cm->frame_to_show;
    ret = 0;
  }
#else
  *sd = *cm->frame_to_show;
  ret = 0;
#endif /*!CONFIG_POSTPROC*/
  vpx_clear_system_state();
  return ret;
}

int vp10_get_frame_to_show(VP10Decoder *pbi,
                           YV12_BUFFER_CONFIG *frame) {
  VP10_COMMON *const cm = &pbi->common;

  if (!cm->show_frame || !cm->frame_to_show)
    return -1;

  *frame = *cm->frame_to_show;
  return 0;
}

vpx_codec_err_t vp10_parse_superframe_index(const uint8_t *data,
                                           size_t data_sz,
                                           uint32_t sizes[8], int *count,
                                           vpx_decrypt_cb decrypt_cb,
                                           void *decrypt_state) {
  // A chunk ending with a byte matching 0xc0 is an invalid chunk unless
  // it is a super frame index. If the last byte of real video compression
  // data is 0xc0 the encoder must add a 0 byte. If we have the marker but
  // not the associated matching marker byte at the front of the index we have
  // an invalid bitstream and need to return an error.

  uint8_t marker;
  size_t frame_sz_sum = 0;

  assert(data_sz);
  marker = read_marker(decrypt_cb, decrypt_state, data + data_sz - 1);
  *count = 0;

  if ((marker & 0xe0) == 0xc0) {
    const uint32_t frames = (marker & 0x7) + 1;
    const uint32_t mag = ((marker >> 3) & 0x3) + 1;
    const size_t index_sz = 2 + mag * (frames - 1);

    // This chunk is marked as having a superframe index but doesn't have
    // enough data for it, thus it's an invalid superframe index.
    if (data_sz < index_sz)
      return VPX_CODEC_CORRUPT_FRAME;

    {
      const uint8_t marker2 = read_marker(decrypt_cb, decrypt_state,
                                          data + data_sz - index_sz);

      // This chunk is marked as having a superframe index but doesn't have
      // the matching marker byte at the front of the index therefore it's an
      // invalid chunk.
      if (marker != marker2)
        return VPX_CODEC_CORRUPT_FRAME;
    }

    {
      // Found a valid superframe index.
      uint32_t i, j;
      const uint8_t *x = &data[data_sz - index_sz + 1];

      // Frames has a maximum of 8 and mag has a maximum of 4.
      uint8_t clear_buffer[32];
      assert(sizeof(clear_buffer) >= frames * mag);
      if (decrypt_cb) {
        decrypt_cb(decrypt_state, x, clear_buffer, frames * mag);
        x = clear_buffer;
      }

      for (i = 0; i < frames - 1; ++i) {
        uint32_t this_sz = 0;

        for (j = 0; j < mag; ++j)
          this_sz |= (*x++) << (j * 8);
        this_sz += 1;
        sizes[i] = this_sz;
        frame_sz_sum += this_sz;
      }
      sizes[i] = (uint32_t)(data_sz - index_sz - frame_sz_sum);
      *count = frames;
    }
  }
  return VPX_CODEC_OK;
}