gstreamer/sys/va/gstvah265dec.c

/* GStreamer
 * Copyright (C) 2020 Igalia, S.L.
 *     Author: Víctor Jáquez <vjaquez@igalia.com>
 * Copyright (C) 2020 Collabora
 *     Author: Nicolas Dufresne <nicolas.dufresne@collabora.com>
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Library General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This library 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
 * Library General Public License for more details.
 *
 * You should have received a copy of the GNU Library General Public
 * License along with this library; if not, write to the0
 * Free Software Foundation, Inc., 51 Franklin St, Fifth Floor,
 * Boston, MA 02110-1301, USA.
 */

/**
 * SECTION:element-vah265dec
 * @title: vah265dec
 * @short_description: A VA-API based H265 video decoder
 *
 * vah265dec decodes H265 bitstreams to VA surfaces using the
 * installed and chosen [VA-API](https://01.org/linuxmedia/vaapi)
 * driver.
 *
 * The decoding surfaces can be mapped onto main memory as video
 * frames.
 *
 * ## Example launch line
 * ```
 * gst-launch-1.0 filesrc location=big_buck_bunny.mov ! parsebin ! vah265dec ! autovideosink
 * ```
 *
 * Since: 1.20
 *
 */

/* ToDo:
 *
 * + interlaced streams
 * + mutiview and stereo profiles
 * + SCC extension buffer
 * + Add 10bit support
 */

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include "gstvah265dec.h"

#include "gstvabasedec.h"

GST_DEBUG_CATEGORY_STATIC (gst_va_h265dec_debug);
#ifndef GST_DISABLE_GST_DEBUG
#define GST_CAT_DEFAULT gst_va_h265dec_debug
#else
#define GST_CAT_DEFAULT NULL
#endif

#define GST_VA_H265_DEC(obj)           ((GstVaH265Dec *) obj)
#define GST_VA_H265_DEC_GET_CLASS(obj) (G_TYPE_INSTANCE_GET_CLASS ((obj), G_TYPE_FROM_INSTANCE (obj), GstVaH265DecClass))
#define GST_VA_H265_DEC_CLASS(klass)   ((GstVaH265DecClass *) klass)

typedef struct _GstVaH265Dec GstVaH265Dec;
typedef struct _GstVaH265DecClass GstVaH265DecClass;

struct _GstVaH265DecClass
{
  GstVaBaseDecClass parent_class;
};

struct _GstVaH265Dec
{
  GstVaBaseDec parent;

  GstFlowReturn last_ret;

  gint coded_width;
  gint coded_height;
  gint dpb_size;

  VAPictureParameterBufferHEVC pic_param;
  gint32 WpOffsetHalfRangeC;

  gboolean need_negotiation;
};

#define parent_class gst_va_base_dec_parent_class
extern gpointer gst_va_base_dec_parent_class;

/* *INDENT-OFF* */
static const gchar *src_caps_str = GST_VIDEO_CAPS_MAKE_WITH_FEATURES ("memory:VAMemory",
            "{ NV12, P010_10LE }") " ;" GST_VIDEO_CAPS_MAKE ("{ NV12, P010_10LE }");
/* *INDENT-ON* */

static const gchar *sink_caps_str = "video/x-h265";

static gboolean
gst_va_h265_dec_end_picture (GstH265Decoder * decoder, GstH265Picture * picture)
{
  GstVaBaseDec *base = GST_VA_BASE_DEC (decoder);
  GstVaDecodePicture *va_pic;

  GST_LOG_OBJECT (base, "end picture %p, (poc %d)",
      picture, picture->pic_order_cnt);

  va_pic = gst_h265_picture_get_user_data (picture);

  return gst_va_decoder_decode (base->decoder, va_pic);
}

static GstFlowReturn
gst_va_h265_dec_output_picture (GstH265Decoder * decoder,
    GstVideoCodecFrame * frame, GstH265Picture * picture)
{
  GstVaBaseDec *base = GST_VA_BASE_DEC (decoder);
  GstVaH265Dec *self = GST_VA_H265_DEC (decoder);

  GST_LOG_OBJECT (self,
      "Outputting picture %p (poc %d)", picture, picture->pic_order_cnt);

  if (self->last_ret != GST_FLOW_OK) {
    gst_h265_picture_unref (picture);
    gst_video_decoder_drop_frame (GST_VIDEO_DECODER (self), frame);
    return self->last_ret;
  }

  if (base->copy_frames)
    gst_va_base_dec_copy_output_buffer (base, frame);

  gst_h265_picture_unref (picture);

  return gst_video_decoder_finish_frame (GST_VIDEO_DECODER (self), frame);
}

static void
_init_vaapi_pic (VAPictureHEVC * va_picture)
{
  va_picture->picture_id = VA_INVALID_ID;
  va_picture->flags = VA_PICTURE_HEVC_INVALID;
  va_picture->pic_order_cnt = 0;
}

static gint
_find_frame_rps_type (GstH265Decoder * decoder, GstH265Picture * ref_pic)
{
  gint i;

  for (i = 0; i < G_N_ELEMENTS (decoder->RefPicSetStCurrBefore); i++) {
    if (ref_pic == decoder->RefPicSetStCurrBefore[i])
      return VA_PICTURE_HEVC_RPS_ST_CURR_BEFORE;
  }

  for (i = 0; i < G_N_ELEMENTS (decoder->RefPicSetStCurrAfter); i++) {
    if (ref_pic == decoder->RefPicSetStCurrAfter[i])
      return VA_PICTURE_HEVC_RPS_ST_CURR_AFTER;
  }

  for (i = 0; i < G_N_ELEMENTS (decoder->RefPicSetLtCurr); i++) {
    if (ref_pic == decoder->RefPicSetLtCurr[i])
      return VA_PICTURE_HEVC_RPS_LT_CURR;
  }

  return 0;
}


static void
_fill_vaapi_pic (GstH265Decoder * decoder, VAPictureHEVC * va_picture,
    GstH265Picture * picture)
{
  GstVaDecodePicture *va_pic;

  va_pic = gst_h265_picture_get_user_data (picture);

  if (!va_pic) {
    _init_vaapi_pic (va_picture);
    return;
  }

  va_picture->picture_id = gst_va_decode_picture_get_surface (va_pic);
  va_picture->pic_order_cnt = picture->pic_order_cnt;
  va_picture->flags = 0;

  if (picture->ref && picture->long_term)
    va_picture->flags |= VA_PICTURE_HEVC_LONG_TERM_REFERENCE;

  va_picture->flags |= _find_frame_rps_type (decoder, picture);

  switch (picture->field) {
    case GST_H265_PICTURE_FIELD_FRAME:
      break;
    case GST_H265_PICTURE_FILED_TOP_FIELD:
      va_picture->flags |= VA_PICTURE_HEVC_FIELD_PIC;
      break;
    case GST_H265_PICTURE_FIELD_BOTTOM_FIELD:
      va_picture->flags |= VA_PICTURE_HEVC_FIELD_PIC;
      va_picture->flags |= VA_PICTURE_HEVC_BOTTOM_FIELD;
      break;
    default:
      break;
  }
}

static guint8
_get_reference_index (GstH265Decoder * decoder, GstH265Picture * picture)
{
  GstVaH265Dec *self = GST_VA_H265_DEC (decoder);
  guint8 i;

  for (i = 0; i < 15; i++) {
    VAPictureHEVC *ref_va_pic = &self->pic_param.ReferenceFrames[i];

    if (ref_va_pic->picture_id == VA_INVALID_ID)
      break;

    if (ref_va_pic->pic_order_cnt == picture->pic_order_cnt)
      return i;
  }

  return 0xFF;
}

/* fill the VA API reference picture lists from the GstCodec reference
 * picture list */
static void
_fill_ref_pic_list (GstH265Decoder * decoder, GstH265Picture * cur_pic,
    guint8 va_reflist[15], GArray * reflist)
{
  guint i;

  for (i = 0; i < reflist->len && i < 15; i++) {
    GstH265Picture *picture = g_array_index (reflist, GstH265Picture *, i);
    va_reflist[i] = _get_reference_index (decoder, picture);
  }

  for (; i < 15; i++)
    va_reflist[i] = 0xFF;
}

static void
_fill_pred_weight_table (GstVaH265Dec * self, GstH265SliceHdr * header,
    VASliceParameterBufferHEVC * slice_param)
{
  gint chroma_weight, chroma_log2_weight_denom;
  gint i, j;
  GstH265PPS *pps = header->pps;

  if (GST_H265_IS_I_SLICE (header) ||
      (!pps->weighted_pred_flag && GST_H265_IS_P_SLICE (header)) ||
      (!pps->weighted_bipred_flag && GST_H265_IS_B_SLICE (header)))
    return;

  slice_param->luma_log2_weight_denom =
      header->pred_weight_table.luma_log2_weight_denom;

  if (pps->sps->chroma_array_type != 0)
    slice_param->delta_chroma_log2_weight_denom =
        header->pred_weight_table.delta_chroma_log2_weight_denom;

  for (i = 0; i <= header->num_ref_idx_l0_active_minus1; i++) {
    if (!header->pred_weight_table.luma_weight_l0_flag[i])
      continue;

    slice_param->delta_luma_weight_l0[i] =
        header->pred_weight_table.delta_luma_weight_l0[i];
    slice_param->luma_offset_l0[i] =
        header->pred_weight_table.luma_offset_l0[i];
  }

  chroma_log2_weight_denom = slice_param->luma_log2_weight_denom +
      slice_param->delta_chroma_log2_weight_denom;

  for (i = 0; i <= header->num_ref_idx_l0_active_minus1; i++) {
    if (!header->pred_weight_table.chroma_weight_l0_flag[i])
      continue;

    for (j = 0; j < 2; j++) {
      gint8 delta_chroma_offset_l0 =
          header->pred_weight_table.delta_chroma_offset_l0[i][j];

      slice_param->delta_chroma_weight_l0[i][j] =
          header->pred_weight_table.delta_chroma_weight_l0[i][j];

      /* Find  ChromaWeightL0 */
      chroma_weight = (1 << chroma_log2_weight_denom) +
          header->pred_weight_table.delta_chroma_weight_l0[i][j];

      /* 7-56 */
      slice_param->ChromaOffsetL0[i][j] = CLAMP (
          (self->WpOffsetHalfRangeC + delta_chroma_offset_l0 -
              ((self->WpOffsetHalfRangeC *
                      chroma_weight) >> chroma_log2_weight_denom)),
          -self->WpOffsetHalfRangeC, self->WpOffsetHalfRangeC - 1);
    }
  }

  /* Skip l1 if this is not a B-Frame. */
  if (!GST_H265_IS_B_SLICE (header))
    return;

  for (i = 0; i <= header->num_ref_idx_l1_active_minus1; i++) {
    if (!header->pred_weight_table.luma_weight_l1_flag[i])
      continue;

    slice_param->delta_luma_weight_l1[i] =
        header->pred_weight_table.delta_luma_weight_l1[i];
    slice_param->luma_offset_l1[i] =
        header->pred_weight_table.luma_offset_l1[i];
  }

  for (i = 0; i <= header->num_ref_idx_l1_active_minus1; i++) {
    if (!header->pred_weight_table.chroma_weight_l1_flag[i])
      continue;

    for (j = 0; j < 2; j++) {
      gint8 delta_chroma_offset_l1 =
          header->pred_weight_table.delta_chroma_offset_l1[i][j];

      slice_param->delta_chroma_weight_l1[i][j] =
          header->pred_weight_table.delta_chroma_weight_l1[i][j];

      /* Find  ChromaWeightL1 */
      chroma_weight = (1 << chroma_log2_weight_denom) +
          header->pred_weight_table.delta_chroma_weight_l1[i][j];

      /* 7-56 */
      slice_param->ChromaOffsetL1[i][j] = CLAMP (
          (self->WpOffsetHalfRangeC + delta_chroma_offset_l1 -
              ((self->WpOffsetHalfRangeC *
                      chroma_weight) >> chroma_log2_weight_denom)),
          -self->WpOffsetHalfRangeC, self->WpOffsetHalfRangeC - 1);
    }
  }
}

static inline guint
_get_slice_data_byte_offset (GstH265SliceHdr * slice_hdr,
    guint nal_header_bytes)
{
  guint epb_count;

  epb_count = slice_hdr->n_emulation_prevention_bytes;
  return nal_header_bytes + (slice_hdr->header_size + 7) / 8 - epb_count;
}

static gboolean
gst_va_h265_dec_decode_slice (GstH265Decoder * decoder,
    GstH265Picture * picture, GstH265Slice * slice, GArray * ref_pic_list0,
    GArray * ref_pic_list1)
{
  GstH265SliceHdr *header = &slice->header;
  GstH265NalUnit *nalu = &slice->nalu;
  GstVaBaseDec *base = GST_VA_BASE_DEC (decoder);
  GstVaDecodePicture *va_pic;
  VASliceParameterBufferHEVC slice_param;
  gboolean ret;

  /* *INDENT-OFF* */
  slice_param = (VASliceParameterBufferHEVC) {
    .slice_data_size = nalu->size,
    .slice_data_offset = 0,
    .slice_data_flag = VA_SLICE_DATA_FLAG_ALL,
    .slice_data_byte_offset = _get_slice_data_byte_offset (header, nalu->header_bytes),
    .slice_segment_address = header->segment_address,
    .collocated_ref_idx = header->temporal_mvp_enabled_flag ? header->collocated_ref_idx : 0xFF,
    .num_ref_idx_l0_active_minus1 = header->num_ref_idx_l0_active_minus1,
    .num_ref_idx_l1_active_minus1 = header->num_ref_idx_l1_active_minus1,
    .slice_qp_delta = header->qp_delta,
    .slice_cb_qp_offset = header->cb_qp_offset,
    .slice_cr_qp_offset = header->cr_qp_offset,
    .slice_beta_offset_div2 = header->beta_offset_div2,
    .slice_tc_offset_div2 = header->tc_offset_div2,
    .five_minus_max_num_merge_cand = header->five_minus_max_num_merge_cand,
    .num_entry_point_offsets = header->num_entry_point_offsets,
    .entry_offset_to_subset_array = 0, /* does not exist in spec */
    .slice_data_num_emu_prevn_bytes = header->n_emulation_prevention_bytes,
  };
  /* *INDENT-ON* */

  /* FIXME to set this right, we'd need to delay writing this until we get
   * called on end_picture(), which is all silly */
  slice_param.LongSliceFlags.fields.LastSliceOfPic = 0;
  slice_param.LongSliceFlags.fields.dependent_slice_segment_flag =
      header->dependent_slice_segment_flag;
  slice_param.LongSliceFlags.fields.slice_type = header->type;
  slice_param.LongSliceFlags.fields.color_plane_id = header->colour_plane_id;
  slice_param.LongSliceFlags.fields.slice_sao_luma_flag = header->sao_luma_flag;
  slice_param.LongSliceFlags.fields.slice_sao_chroma_flag =
      header->sao_chroma_flag;
  slice_param.LongSliceFlags.fields.mvd_l1_zero_flag = header->mvd_l1_zero_flag;
  slice_param.LongSliceFlags.fields.cabac_init_flag = header->cabac_init_flag;
  slice_param.LongSliceFlags.fields.slice_temporal_mvp_enabled_flag =
      header->temporal_mvp_enabled_flag;
  slice_param.LongSliceFlags.fields.slice_deblocking_filter_disabled_flag =
      header->deblocking_filter_disabled_flag;
  slice_param.LongSliceFlags.fields.collocated_from_l0_flag =
      header->collocated_from_l0_flag;
  slice_param.LongSliceFlags.
      fields.slice_loop_filter_across_slices_enabled_flag =
      header->loop_filter_across_slices_enabled_flag;

  _fill_ref_pic_list (decoder, picture, slice_param.RefPicList[0],
      ref_pic_list0);
  _fill_ref_pic_list (decoder, picture, slice_param.RefPicList[1],
      ref_pic_list1);

  _fill_pred_weight_table (GST_VA_H265_DEC (decoder), header, &slice_param);

  va_pic = gst_h265_picture_get_user_data (picture);

  ret = gst_va_decoder_add_slice_buffer (base->decoder, va_pic, &slice_param,
      sizeof (slice_param), slice->nalu.data + slice->nalu.offset,
      slice->nalu.size);
  if (!ret) {
    gst_va_decoder_destroy_buffers (base->decoder, va_pic);
    return FALSE;
  }

  return TRUE;
}

static gboolean
gst_va_h265_dec_start_picture (GstH265Decoder * decoder,
    GstH265Picture * picture, GstH265Slice * slice, GstH265Dpb * dpb)
{
  GstH265PPS *pps;
  GstH265SPS *sps;
  GstVaH265Dec *self = GST_VA_H265_DEC (decoder);
  GstVaBaseDec *base = &self->parent;
  GstVaDecodePicture *va_pic;
  GstH265ScalingList *scaling_list = NULL;
  VAIQMatrixBufferHEVC iq_matrix = { 0, };
  VAPictureParameterBufferHEVC *pic_param = &self->pic_param;
  guint i;

  va_pic = gst_h265_picture_get_user_data (picture);

  pps = slice->header.pps;
  sps = pps->sps;

  /* *INDENT-OFF* */
  *pic_param = (VAPictureParameterBufferHEVC) {
    .pic_width_in_luma_samples = sps->pic_width_in_luma_samples,
    .pic_height_in_luma_samples = sps->pic_height_in_luma_samples,
    .sps_max_dec_pic_buffering_minus1 = sps->max_dec_pic_buffering_minus1[sps->max_sub_layers_minus1],
    .bit_depth_luma_minus8 = sps->bit_depth_luma_minus8,
    .bit_depth_chroma_minus8 = sps->bit_depth_chroma_minus8,
    .pcm_sample_bit_depth_luma_minus1 = sps->pcm_sample_bit_depth_luma_minus1,
    .pcm_sample_bit_depth_chroma_minus1 = sps->pcm_sample_bit_depth_chroma_minus1,
    .log2_min_luma_coding_block_size_minus3 = sps->log2_min_luma_coding_block_size_minus3,
    .log2_diff_max_min_luma_coding_block_size = sps->log2_diff_max_min_luma_coding_block_size,
    .log2_min_transform_block_size_minus2 = sps->log2_min_transform_block_size_minus2,
    .log2_diff_max_min_transform_block_size = sps->log2_diff_max_min_transform_block_size,
    .log2_min_pcm_luma_coding_block_size_minus3 = sps->log2_min_pcm_luma_coding_block_size_minus3,
    .log2_diff_max_min_pcm_luma_coding_block_size = sps->log2_diff_max_min_pcm_luma_coding_block_size,
    .max_transform_hierarchy_depth_intra = sps->max_transform_hierarchy_depth_intra,
    .max_transform_hierarchy_depth_inter = sps->max_transform_hierarchy_depth_inter,
    .init_qp_minus26 = pps->init_qp_minus26,
    .diff_cu_qp_delta_depth = pps->diff_cu_qp_delta_depth,
    .pps_cb_qp_offset = pps->cb_qp_offset,
    .pps_cr_qp_offset = pps->cr_qp_offset,
    .log2_parallel_merge_level_minus2 = pps->log2_parallel_merge_level_minus2,
    .num_tile_columns_minus1 = pps->num_tile_columns_minus1,
    .num_tile_rows_minus1 = pps->num_tile_rows_minus1,
    .log2_max_pic_order_cnt_lsb_minus4 = sps->log2_max_pic_order_cnt_lsb_minus4,
    .num_short_term_ref_pic_sets = sps->num_short_term_ref_pic_sets,
    .num_long_term_ref_pic_sps = sps->num_long_term_ref_pics_sps,
    .num_ref_idx_l0_default_active_minus1 = pps->num_ref_idx_l0_default_active_minus1,
    .num_ref_idx_l1_default_active_minus1 = pps->num_ref_idx_l1_default_active_minus1,
    .pps_beta_offset_div2 = pps->beta_offset_div2,
    .pps_tc_offset_div2 = pps->tc_offset_div2,
    .num_extra_slice_header_bits = pps->num_extra_slice_header_bits,
    .st_rps_bits = slice->header.short_term_ref_pic_set_size, /* FIXME missing emulation bits removal */
    .pic_fields.bits = {
      .chroma_format_idc = sps->chroma_format_idc,
      .separate_colour_plane_flag = sps->separate_colour_plane_flag,
      .pcm_enabled_flag = sps->pcm_enabled_flag,
      .scaling_list_enabled_flag = sps->scaling_list_enabled_flag,
      .transform_skip_enabled_flag = pps->transform_skip_enabled_flag,
      .amp_enabled_flag = sps->amp_enabled_flag,
      .strong_intra_smoothing_enabled_flag = sps->strong_intra_smoothing_enabled_flag,
      .sign_data_hiding_enabled_flag = pps->sign_data_hiding_enabled_flag,
      .constrained_intra_pred_flag = pps->constrained_intra_pred_flag,
      .cu_qp_delta_enabled_flag = pps->cu_qp_delta_enabled_flag,
      .weighted_pred_flag = pps->weighted_pred_flag,
      .weighted_bipred_flag = pps->weighted_bipred_flag,
      .transquant_bypass_enabled_flag = pps->transquant_bypass_enabled_flag,
      .tiles_enabled_flag = pps->tiles_enabled_flag,
      .entropy_coding_sync_enabled_flag = pps->entropy_coding_sync_enabled_flag,
      .pps_loop_filter_across_slices_enabled_flag = pps->loop_filter_across_slices_enabled_flag,
      .loop_filter_across_tiles_enabled_flag = pps->loop_filter_across_tiles_enabled_flag,
      .pcm_loop_filter_disabled_flag = sps->pcm_loop_filter_disabled_flag,
      /* Not set by FFMPEG either */
      .NoPicReorderingFlag = 0,
      .NoBiPredFlag = 0,
    },
    .slice_parsing_fields.bits = {
      .lists_modification_present_flag = pps->lists_modification_present_flag,
      .long_term_ref_pics_present_flag = sps->long_term_ref_pics_present_flag,
      .sps_temporal_mvp_enabled_flag = sps->temporal_mvp_enabled_flag,
      .cabac_init_present_flag = pps->cabac_init_present_flag,
      .output_flag_present_flag = pps->output_flag_present_flag,
      .dependent_slice_segments_enabled_flag = pps->dependent_slice_segments_enabled_flag,
      .pps_slice_chroma_qp_offsets_present_flag = pps->slice_chroma_qp_offsets_present_flag,
      .sample_adaptive_offset_enabled_flag = sps->sample_adaptive_offset_enabled_flag,
      .deblocking_filter_override_enabled_flag = pps->deblocking_filter_override_enabled_flag,
      .pps_disable_deblocking_filter_flag = pps->deblocking_filter_disabled_flag,
      .slice_segment_header_extension_present_flag = pps->slice_segment_header_extension_present_flag,
      .RapPicFlag = picture->RapPicFlag,
      .IdrPicFlag = GST_H265_IS_NAL_TYPE_IDR (slice->nalu.type),
      .IntraPicFlag = GST_H265_IS_NAL_TYPE_IRAP (slice->nalu.type),
    },
  };
  /* *INDENT-ON* */

  for (i = 0; i <= pps->num_tile_columns_minus1; i++)
    pic_param->column_width_minus1[i] = pps->column_width_minus1[i];

  for (i = 0; i <= pps->num_tile_rows_minus1; i++)
    pic_param->row_height_minus1[i] = pps->row_height_minus1[i];

  _fill_vaapi_pic (decoder, &pic_param->CurrPic, picture);

  /* reference frames */
  {
    GArray *ref_list = gst_h265_dpb_get_pictures_all (dpb);
    for (i = 0; i < 15 && i < ref_list->len; i++) {
      GstH265Picture *pic = g_array_index (ref_list, GstH265Picture *, i);
      _fill_vaapi_pic (decoder, &pic_param->ReferenceFrames[i], pic);
    }
    g_array_unref (ref_list);

    for (; i < 15; i++)
      _init_vaapi_pic (&pic_param->ReferenceFrames[i]);
  }

  if (!gst_va_decoder_add_param_buffer (base->decoder, va_pic,
          VAPictureParameterBufferType, pic_param, sizeof (*pic_param)))
    goto fail;

  if (pps->scaling_list_data_present_flag ||
      (sps->scaling_list_enabled_flag
          && !sps->scaling_list_data_present_flag)) {
    scaling_list = &pps->scaling_list;
    GST_DEBUG_OBJECT (decoder, "Passing scaling list from PPS");
  } else if (sps->scaling_list_enabled_flag &&
      sps->scaling_list_data_present_flag) {
    scaling_list = &sps->scaling_list;
    GST_DEBUG_OBJECT (decoder, "Passing scaling list from SPS");
  }

  if (scaling_list) {
    for (i = 0; i < G_N_ELEMENTS (iq_matrix.ScalingList4x4); i++)
      gst_h265_quant_matrix_4x4_get_raster_from_uprightdiagonal
          (iq_matrix.ScalingList4x4[i], scaling_list->scaling_lists_4x4[i]);

    for (i = 0; i < G_N_ELEMENTS (iq_matrix.ScalingList8x8); i++)
      gst_h265_quant_matrix_8x8_get_raster_from_uprightdiagonal
          (iq_matrix.ScalingList8x8[i], scaling_list->scaling_lists_8x8[i]);

    for (i = 0; i < G_N_ELEMENTS (iq_matrix.ScalingList16x16); i++)
      gst_h265_quant_matrix_16x16_get_raster_from_uprightdiagonal
          (iq_matrix.ScalingList16x16[i], scaling_list->scaling_lists_16x16[i]);

    for (i = 0; i < G_N_ELEMENTS (iq_matrix.ScalingList32x32); i++)
      gst_h265_quant_matrix_32x32_get_raster_from_uprightdiagonal
          (iq_matrix.ScalingList32x32[i], scaling_list->scaling_lists_32x32[i]);

    for (i = 0; i < 6; i++)
      iq_matrix.ScalingListDC16x16[i] =
          scaling_list->scaling_list_dc_coef_minus8_16x16[i] + 8;

    for (i = 0; i < 2; i++)
      iq_matrix.ScalingListDC32x32[i] =
          scaling_list->scaling_list_dc_coef_minus8_32x32[i] + 8;

    if (!gst_va_decoder_add_param_buffer (base->decoder, va_pic,
            VAIQMatrixBufferType, &iq_matrix, sizeof (iq_matrix)))
      goto fail;
  }

  return TRUE;

fail:
  {
    gst_va_decoder_destroy_buffers (base->decoder, va_pic);
    return FALSE;
  }
}

static gboolean
gst_va_h265_dec_new_picture (GstH265Decoder * decoder,
    GstVideoCodecFrame * frame, GstH265Picture * picture)
{
  GstVaH265Dec *self = GST_VA_H265_DEC (decoder);
  GstVaDecodePicture *pic;
  GstVideoDecoder *vdec = GST_VIDEO_DECODER (decoder);

  self->last_ret = gst_video_decoder_allocate_output_frame (vdec, frame);
  if (self->last_ret != GST_FLOW_OK)
    goto error;

  pic = gst_va_decode_picture_new (frame->output_buffer);

  gst_h265_picture_set_user_data (picture, pic,
      (GDestroyNotify) gst_va_decode_picture_free);

  GST_LOG_OBJECT (self, "New va decode picture %p - %#x", pic,
      gst_va_decode_picture_get_surface (pic));

  return TRUE;

error:
  {
    GST_WARNING_OBJECT (self,
        "Failed to allocated output buffer, return %s",
        gst_flow_get_name (self->last_ret));
    return FALSE;
  }
}

static guint
_get_rtformat (GstVaH265Dec * self, guint8 bit_depth_luma,
    guint8 chroma_format_idc)
{
  switch (bit_depth_luma) {
    case 10:
      if (chroma_format_idc == 3)
        return VA_RT_FORMAT_YUV444_10;
      if (chroma_format_idc == 2)
        return VA_RT_FORMAT_YUV422_10;
      else
        return VA_RT_FORMAT_YUV420_10;
      break;
    case 8:
      if (chroma_format_idc == 3)
        return VA_RT_FORMAT_YUV444;
      if (chroma_format_idc == 2)
        return VA_RT_FORMAT_YUV422;
      else
        return VA_RT_FORMAT_YUV420;
      break;
    default:
      GST_ERROR_OBJECT (self, "Unsupported chroma format: %d "
          "(with depth luma: %d)", chroma_format_idc, bit_depth_luma);
      return 0;
  }
}

/* *INDENT-OFF* */
static const struct
{
  GstH265Profile profile;
  VAProfile va_profile;
} profile_map[] = {
#define P(idc, va) { G_PASTE (GST_H265_PROFILE_, idc), G_PASTE (VAProfileHEVC, va) }
  P (MAIN, Main),
  P (MAIN_10, Main10),
  /*P (MAIN_STILL_PICTURE, ),
  P (MONOCHROME, ),
  P (MONOCHROME_12, ),
  P (MONOCHROME_16, ),*/
  P (MAIN_12, Main12),
  P (MAIN_422_10, Main422_10),
  P (MAIN_422_12, Main422_12),
  P (MAIN_444, Main444),
  P (MAIN_444_10, Main444_10),
  P (MAIN_444_12, Main444_12),
  /*P (MAIN_INTRA, ),
  P (MAIN_10_INTRA, ),
  P (MAIN_12_INTRA, ),
  P (MAIN_422_10_INTRA, ),
  P (MAIN_422_12_INTRA, ),
  P (MAIN_444_INTRA, ),
  P (MAIN_444_10_INTRA, ),
  P (MAIN_444_12_INTRA, ),
  P (MAIN_444_16_INTRA, ),
  P (MAIN_444_STILL_PICTURE, ),
  P (MAIN_444_16_STILL_PICTURE, ),
  P (MONOCHROME_10, ),
  P (HIGH_THROUGHPUT_444, ),
  P (HIGH_THROUGHPUT_444_10, ),
  P (HIGH_THROUGHPUT_444_14, ),
  P (HIGH_THROUGHPUT_444_16_INTRA, ),*/
  P (SCREEN_EXTENDED_MAIN, SccMain),
  P (SCREEN_EXTENDED_MAIN_10, SccMain10),
  P (SCREEN_EXTENDED_MAIN_444, SccMain444),
  /*P (SCREEN_EXTENDED_MAIN_444_10, ),
  P (SCREEN_EXTENDED_HIGH_THROUGHPUT_444, ),
  P (SCREEN_EXTENDED_HIGH_THROUGHPUT_444_10, ),
  P (SCREEN_EXTENDED_HIGH_THROUGHPUT_444_14, ),
  P (MULTIVIEW_MAIN, ),
  P (SCALABLE_MAIN, ),
  P (SCALABLE_MAIN_10, ),
  P (SCALABLE_MONOCHROME, ),
  P (SCALABLE_MONOCHROME_12, ),
  P (SCALABLE_MONOCHROME_16, ),
  P (SCALABLE_MAIN_444, ),
  P (3D_MAIN, ),*/
#undef P
};
/* *INDENT-ON* */

static VAProfile
_get_profile (GstVaH265Dec * self, const GstH265SPS * sps, gint max_dpb_size)
{
  GstVaBaseDec *base = GST_VA_BASE_DEC (self);
  GstH265Profile profile =
      gst_h265_profile_tier_level_get_profile (&sps->profile_tier_level);
  VAProfile profiles[4];
  gint i = 0, j;

  for (j = 0; j < G_N_ELEMENTS (profile_map); j++) {
    if (profile_map[j].profile == profile) {
      profiles[i++] = profile_map[j].va_profile;
      break;
    }
  }

  /* TODO Special cases here */

  for (j = 0; j < i && j < G_N_ELEMENTS (profiles); j++) {
    if (gst_va_decoder_has_profile (base->decoder, profiles[j]))
      return profiles[j];
  }

  GST_ERROR_OBJECT (self, "Unsupported profile: %d", profile);

  return VAProfileNone;
}

static gboolean
gst_va_h265_dec_new_sequence (GstH265Decoder * decoder, const GstH265SPS * sps,
    gint max_dpb_size)
{
  GstVaBaseDec *base = GST_VA_BASE_DEC (decoder);
  GstVaH265Dec *self = GST_VA_H265_DEC (decoder);
  VAProfile profile;
  gint display_width;
  gint display_height;
  guint rt_format;
  gboolean negotiation_needed = FALSE;

  if (self->dpb_size < max_dpb_size)
    self->dpb_size = max_dpb_size;

  if (sps->conformance_window_flag) {
    display_width = sps->crop_rect_width;
    display_height = sps->crop_rect_height;
  } else {
    display_width = sps->width;
    display_height = sps->height;
  }

  profile = _get_profile (self, sps, max_dpb_size);
  if (profile == VAProfileNone)
    return FALSE;

  rt_format = _get_rtformat (self, sps->bit_depth_luma_minus8 + 8,
      sps->chroma_format_idc);
  if (rt_format == 0)
    return FALSE;

  if (gst_va_decoder_format_changed (base->decoder, profile,
          rt_format, sps->width, sps->height)) {
    base->profile = profile;
    base->rt_format = rt_format;
    self->coded_width = sps->width;
    self->coded_height = sps->height;

    negotiation_needed = TRUE;
    GST_INFO_OBJECT (self, "Format changed to %s [%x] (%dx%d)",
        gst_va_profile_name (profile), rt_format, self->coded_width,
        self->coded_height);
  }

  if (base->width != display_width || base->height != display_height) {
    base->width = display_width;
    base->height = display_height;

    negotiation_needed = TRUE;
    GST_INFO_OBJECT (self, "Resolution changed to %dx%d", base->width,
        base->height);
  }

  base->need_valign = base->width < self->coded_width
      || base->height < self->coded_height;
  if (base->need_valign) {
    /* *INDENT-OFF* */
    base->valign = (GstVideoAlignment) {
      .padding_bottom = self->coded_height - base->height,
      .padding_left = self->coded_width - base->width,
    };
    /* *INDENT-ON* */
  }

  base->min_buffers = self->dpb_size + 4;       /* dpb size + scratch surfaces */

  if (negotiation_needed) {
    self->need_negotiation = TRUE;
    if (!gst_video_decoder_negotiate (GST_VIDEO_DECODER (self))) {
      GST_ERROR_OBJECT (self, "Failed to negotiate with downstream");
      return FALSE;
    }
  }

  {
    /* FIXME: We don't have parser API for sps_range_extension, so
     * assuming high_precision_offsets_enabled_flag as zero */
    guint high_precision_offsets_enabled_flag = 0, bitdepthC = 0;

    /* Calculate WpOffsetHalfRangeC: (7-34) */
    bitdepthC = sps->bit_depth_chroma_minus8 + 8;
    self->WpOffsetHalfRangeC =
        1 << (high_precision_offsets_enabled_flag ? (bitdepthC - 1) : 7);
  }

  return TRUE;
}

static GstCaps *
_complete_sink_caps (GstCaps * sinkcaps)
{
  GstCaps *caps = gst_caps_copy (sinkcaps);
  GValue val = G_VALUE_INIT;
  const gchar *streamformat[] = { "hvc", "hev1", "byte-stream" };
  gint i;

  g_value_init (&val, G_TYPE_STRING);
  g_value_set_string (&val, "au");
  gst_caps_set_value (caps, "alignment", &val);
  g_value_unset (&val);

  gst_value_list_init (&val, G_N_ELEMENTS (streamformat));
  for (i = 0; i < G_N_ELEMENTS (streamformat); i++) {
    GValue v = G_VALUE_INIT;

    g_value_init (&v, G_TYPE_STRING);
    g_value_set_string (&v, streamformat[i]);
    gst_value_list_append_value (&val, &v);
    g_value_unset (&v);
  }
  gst_caps_set_value (caps, "stream-format", &val);
  g_value_unset (&val);

  return caps;
}

static GstCaps *
gst_va_h265_dec_getcaps (GstVideoDecoder * decoder, GstCaps * filter)
{
  GstCaps *sinkcaps, *caps = NULL, *tmp;
  GstVaBaseDec *base = GST_VA_BASE_DEC (decoder);

  if (base->decoder)
    caps = gst_va_decoder_get_sinkpad_caps (base->decoder);

  if (caps) {
    sinkcaps = _complete_sink_caps (caps);
    gst_caps_unref (caps);
    if (filter) {
      tmp = gst_caps_intersect_full (filter, sinkcaps,
          GST_CAPS_INTERSECT_FIRST);
      gst_caps_unref (sinkcaps);
      caps = tmp;
    } else {
      caps = sinkcaps;
    }
    GST_LOG_OBJECT (base, "Returning caps %" GST_PTR_FORMAT, caps);
  } else if (!caps) {
    caps = gst_video_decoder_proxy_getcaps (decoder, NULL, filter);
  }

  return caps;
}

static gboolean
gst_va_h265_dec_negotiate (GstVideoDecoder * decoder)
{
  GstVaBaseDec *base = GST_VA_BASE_DEC (decoder);
  GstVaH265Dec *self = GST_VA_H265_DEC (decoder);
  GstVideoFormat format = GST_VIDEO_FORMAT_UNKNOWN;
  GstCapsFeatures *capsfeatures = NULL;
  GstH265Decoder *h265dec = GST_H265_DECODER (decoder);

  /* Ignore downstream renegotiation request. */
  if (!self->need_negotiation)
    return TRUE;

  self->need_negotiation = FALSE;

  if (gst_va_decoder_is_open (base->decoder)
      && !gst_va_decoder_close (base->decoder))
    return FALSE;

  if (!gst_va_decoder_open (base->decoder, base->profile, base->rt_format))
    return FALSE;

  if (!gst_va_decoder_set_format (base->decoder, self->coded_width,
          self->coded_height, NULL))
    return FALSE;

  if (base->output_state)
    gst_video_codec_state_unref (base->output_state);

  gst_va_base_dec_get_preferred_format_and_caps_features (base, &format,
      &capsfeatures);

  base->output_state =
      gst_video_decoder_set_output_state (decoder, format,
      base->width, base->height, h265dec->input_state);

  base->output_state->caps = gst_video_info_to_caps (&base->output_state->info);
  if (capsfeatures)
    gst_caps_set_features_simple (base->output_state->caps, capsfeatures);

  GST_INFO_OBJECT (self, "Negotiated caps %" GST_PTR_FORMAT,
      base->output_state->caps);

  return GST_VIDEO_DECODER_CLASS (parent_class)->negotiate (decoder);
}

static void
gst_va_h265_dec_dispose (GObject * object)
{
  gst_va_base_dec_close (GST_VIDEO_DECODER (object));
  G_OBJECT_CLASS (parent_class)->dispose (object);
}

static void
gst_va_h265_dec_class_init (gpointer g_class, gpointer class_data)
{
  GstCaps *src_doc_caps, *sink_doc_caps;
  GObjectClass *gobject_class = G_OBJECT_CLASS (g_class);
  GstElementClass *element_class = GST_ELEMENT_CLASS (g_class);
  GstH265DecoderClass *h265decoder_class = GST_H265_DECODER_CLASS (g_class);
  GstVideoDecoderClass *decoder_class = GST_VIDEO_DECODER_CLASS (g_class);
  struct CData *cdata = class_data;
  gchar *long_name;

  if (cdata->description) {
    long_name = g_strdup_printf ("VA-API H.265 Decoder in %s",
        cdata->description);
  } else {
    long_name = g_strdup ("VA-API H.265 Decoder");
  }

  gst_element_class_set_metadata (element_class, long_name,
      "Codec/Decoder/Video/Hardware",
      "VA-API based H.265 video decoder",
      "Nicolas Dufresne <nicolas.dufresne@collabora.com>");

  sink_doc_caps = gst_caps_from_string (sink_caps_str);
  src_doc_caps = gst_caps_from_string (src_caps_str);

  gst_va_base_dec_class_init (GST_VA_BASE_DEC_CLASS (g_class), HEVC,
      cdata->render_device_path, cdata->sink_caps, cdata->src_caps,
      src_doc_caps, sink_doc_caps);

  gobject_class->dispose = gst_va_h265_dec_dispose;

  decoder_class->getcaps = GST_DEBUG_FUNCPTR (gst_va_h265_dec_getcaps);
  decoder_class->negotiate = GST_DEBUG_FUNCPTR (gst_va_h265_dec_negotiate);

  h265decoder_class->new_sequence =
      GST_DEBUG_FUNCPTR (gst_va_h265_dec_new_sequence);
  h265decoder_class->decode_slice =
      GST_DEBUG_FUNCPTR (gst_va_h265_dec_decode_slice);

  h265decoder_class->new_picture =
      GST_DEBUG_FUNCPTR (gst_va_h265_dec_new_picture);
  h265decoder_class->output_picture =
      GST_DEBUG_FUNCPTR (gst_va_h265_dec_output_picture);
  h265decoder_class->start_picture =
      GST_DEBUG_FUNCPTR (gst_va_h265_dec_start_picture);
  h265decoder_class->end_picture =
      GST_DEBUG_FUNCPTR (gst_va_h265_dec_end_picture);

  g_free (long_name);
  g_free (cdata->description);
  g_free (cdata->render_device_path);
  gst_caps_unref (cdata->src_caps);
  gst_caps_unref (cdata->sink_caps);
  g_free (cdata);
}

static void
gst_va_h265_dec_init (GTypeInstance * instance, gpointer g_class)
{
  gst_va_base_dec_init (GST_VA_BASE_DEC (instance), GST_CAT_DEFAULT);
  gst_h265_decoder_set_process_ref_pic_lists (GST_H265_DECODER (instance),
      TRUE);
}

static gpointer
_register_debug_category (gpointer data)
{
  GST_DEBUG_CATEGORY_INIT (gst_va_h265dec_debug, "vah265dec", 0,
      "VA H265 decoder");

  return NULL;
}

gboolean
gst_va_h265_dec_register (GstPlugin * plugin, GstVaDevice * device,
    GstCaps * sink_caps, GstCaps * src_caps, guint rank)
{
  static GOnce debug_once = G_ONCE_INIT;
  GType type;
  GTypeInfo type_info = {
    .class_size = sizeof (GstVaH265DecClass),
    .class_init = gst_va_h265_dec_class_init,
    .instance_size = sizeof (GstVaH265Dec),
    .instance_init = gst_va_h265_dec_init,
  };
  struct CData *cdata;
  gboolean ret;
  gchar *type_name, *feature_name;

  g_return_val_if_fail (GST_IS_PLUGIN (plugin), FALSE);
  g_return_val_if_fail (GST_IS_VA_DEVICE (device), FALSE);
  g_return_val_if_fail (GST_IS_CAPS (sink_caps), FALSE);
  g_return_val_if_fail (GST_IS_CAPS (src_caps), FALSE);

  cdata = g_new (struct CData, 1);
  cdata->description = NULL;
  cdata->render_device_path = g_strdup (device->render_device_path);
  cdata->sink_caps = _complete_sink_caps (sink_caps);
  cdata->src_caps = gst_caps_ref (src_caps);

  /* class data will be leaked if the element never gets instantiated */
  GST_MINI_OBJECT_FLAG_SET (cdata->sink_caps,
      GST_MINI_OBJECT_FLAG_MAY_BE_LEAKED);
  GST_MINI_OBJECT_FLAG_SET (src_caps, GST_MINI_OBJECT_FLAG_MAY_BE_LEAKED);

  type_info.class_data = cdata;

  type_name = g_strdup ("GstVaH265Dec");
  feature_name = g_strdup ("vah265dec");

  /* The first decoder to be registered should use a constant name,
   * like vah265dec, for any additional decoders, we create unique
   * names, using inserting the render device name. */
  if (g_type_from_name (type_name)) {
    gchar *basename = g_path_get_basename (device->render_device_path);
    g_free (type_name);
    g_free (feature_name);
    type_name = g_strdup_printf ("GstVa%sH265Dec", basename);
    feature_name = g_strdup_printf ("va%sh265dec", basename);
    cdata->description = basename;

    /* lower rank for non-first device */
    if (rank > 0)
      rank--;
  }

  g_once (&debug_once, _register_debug_category, NULL);

  type = g_type_register_static (GST_TYPE_H265_DECODER,
      type_name, &type_info, 0);

  ret = gst_element_register (plugin, feature_name, rank, type);

  g_free (type_name);
  g_free (feature_name);

  return ret;
}