git.sesse.net Git - ffmpeg/blob - libavcodec/nvdec_av1.c

   1 /*
   2  * AV1 HW decode acceleration through NVDEC
   3  *
   4  * Copyright (c) 2020 Timo Rothenpieler
   5  *
   6  * This file is part of FFmpeg.
   7  *
   8  * FFmpeg is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * FFmpeg is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with FFmpeg; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #include "avcodec.h"
  24 #include "nvdec.h"
  25 #include "decode.h"
  26 #include "internal.h"
  27 #include "av1dec.h"
  28
  29
  30 static int get_bit_depth_from_seq(const AV1RawSequenceHeader *seq)
  31 {
  32     if (seq->seq_profile == 2 && seq->color_config.high_bitdepth)
  33         return seq->color_config.twelve_bit ? 12 : 10;
  34     else if (seq->seq_profile <= 2 && seq->color_config.high_bitdepth)
  35         return 10;
  36     else
  37         return 8;
  38 }
  39
  40 static int nvdec_av1_start_frame(AVCodecContext *avctx, const uint8_t *buffer, uint32_t size)
  41 {
  42     const AV1DecContext *s = avctx->priv_data;
  43     const AV1RawSequenceHeader *seq = s->raw_seq;
  44     const AV1RawFrameHeader *frame_header = s->raw_frame_header;
  45     const AV1RawFilmGrainParams *film_grain = &s->cur_frame.film_grain;
  46
  47     NVDECContext      *ctx = avctx->internal->hwaccel_priv_data;
  48     CUVIDPICPARAMS     *pp = &ctx->pic_params;
  49     CUVIDAV1PICPARAMS *ppc = &pp->CodecSpecific.av1;
  50     FrameDecodeData *fdd;
  51     NVDECFrame *cf;
  52     AVFrame *cur_frame = s->cur_frame.tf.f;
  53
  54     unsigned char remap_lr_type[4] = { AV1_RESTORE_NONE, AV1_RESTORE_SWITCHABLE, AV1_RESTORE_WIENER, AV1_RESTORE_SGRPROJ };
  55
  56     int ret, i, j;
  57
  58     ret = ff_nvdec_start_frame_sep_ref(avctx, cur_frame, film_grain->apply_grain);
  59     if (ret < 0)
  60         return ret;
  61
  62     fdd = (FrameDecodeData*)cur_frame->private_ref->data;
  63     cf  = (NVDECFrame*)fdd->hwaccel_priv;
  64
  65     *pp = (CUVIDPICPARAMS) {
  66         .PicWidthInMbs     = (cur_frame->width  + 15) / 16,
  67         .FrameHeightInMbs  = (cur_frame->height + 15) / 16,
  68         .CurrPicIdx        = cf->idx,
  69         .ref_pic_flag      = !!frame_header->refresh_frame_flags,
  70         .intra_pic_flag    = frame_header->frame_type == AV1_FRAME_INTRA_ONLY ||
  71                              frame_header->frame_type == AV1_FRAME_KEY,
  72
  73         .CodecSpecific.av1 = {
  74             .width  = cur_frame->width,
  75             .height = cur_frame->height,
  76
  77             .frame_offset = frame_header->order_hint,
  78             .decodePicIdx = cf->ref_idx,
  79
  80             /* Sequence Header */
  81             .profile                    = seq->seq_profile,
  82             .use_128x128_superblock     = seq->use_128x128_superblock,
  83             .subsampling_x              = seq->color_config.subsampling_x,
  84             .subsampling_y              = seq->color_config.subsampling_y,
  85             .mono_chrome                = seq->color_config.mono_chrome,
  86             .bit_depth_minus8           = get_bit_depth_from_seq(seq) - 8,
  87             .enable_filter_intra        = seq->enable_filter_intra,
  88             .enable_intra_edge_filter   = seq->enable_intra_edge_filter,
  89             .enable_interintra_compound = seq->enable_interintra_compound,
  90             .enable_masked_compound     = seq->enable_masked_compound,
  91             .enable_dual_filter         = seq->enable_dual_filter,
  92             .enable_order_hint          = seq->enable_order_hint,
  93             .order_hint_bits_minus1     = seq->order_hint_bits_minus_1,
  94             .enable_jnt_comp            = seq->enable_jnt_comp,
  95             .enable_superres            = seq->enable_superres,
  96             .enable_cdef                = seq->enable_cdef,
  97             .enable_restoration         = seq->enable_restoration,
  98             .enable_fgs                 = seq->film_grain_params_present,
  99
 100             /* Frame Header */
 101             .frame_type                   = frame_header->frame_type,
 102             .show_frame                   = frame_header->show_frame,
 103             .disable_cdf_update           = frame_header->disable_cdf_update,
 104             .allow_screen_content_tools   = frame_header->allow_screen_content_tools,
 105             .force_integer_mv             = frame_header->force_integer_mv ||
 106                                             frame_header->frame_type == AV1_FRAME_INTRA_ONLY ||
 107                                             frame_header->frame_type == AV1_FRAME_KEY,
 108             .coded_denom                  = frame_header->coded_denom,
 109             .allow_intrabc                = frame_header->allow_intrabc,
 110             .allow_high_precision_mv      = frame_header->allow_high_precision_mv,
 111             .interp_filter                = frame_header->interpolation_filter,
 112             .switchable_motion_mode       = frame_header->is_motion_mode_switchable,
 113             .use_ref_frame_mvs            = frame_header->use_ref_frame_mvs,
 114             .disable_frame_end_update_cdf = frame_header->disable_frame_end_update_cdf,
 115             .delta_q_present              = frame_header->delta_q_present,
 116             .delta_q_res                  = frame_header->delta_q_res,
 117             .using_qmatrix                = frame_header->using_qmatrix,
 118             .coded_lossless               = s->cur_frame.coded_lossless,
 119             .use_superres                 = frame_header->use_superres,
 120             .tx_mode                      = frame_header->tx_mode,
 121             .reference_mode               = frame_header->reference_select,
 122             .allow_warped_motion          = frame_header->allow_warped_motion,
 123             .reduced_tx_set               = frame_header->reduced_tx_set,
 124             .skip_mode                    = frame_header->skip_mode_present,
 125
 126             /* Tiling Info */
 127             .num_tile_cols          = frame_header->tile_cols,
 128             .num_tile_rows          = frame_header->tile_rows,
 129             .context_update_tile_id = frame_header->context_update_tile_id,
 130
 131             /* CDEF */
 132             .cdef_damping_minus_3 = frame_header->cdef_damping_minus_3,
 133             .cdef_bits            = frame_header->cdef_bits,
 134
 135             /* SkipModeFrames */
 136             .SkipModeFrame0 = frame_header->skip_mode_present ?
 137                               s->cur_frame.skip_mode_frame_idx[0] : 0,
 138             .SkipModeFrame1 = frame_header->skip_mode_present ?
 139                               s->cur_frame.skip_mode_frame_idx[1] : 0,
 140
 141             /* QP Information */
 142             .base_qindex     = frame_header->base_q_idx,
 143             .qp_y_dc_delta_q = frame_header->delta_q_y_dc,
 144             .qp_u_dc_delta_q = frame_header->delta_q_u_dc,
 145             .qp_v_dc_delta_q = frame_header->delta_q_v_dc,
 146             .qp_u_ac_delta_q = frame_header->delta_q_u_ac,
 147             .qp_v_ac_delta_q = frame_header->delta_q_v_ac,
 148             .qm_y            = frame_header->qm_y,
 149             .qm_u            = frame_header->qm_u,
 150             .qm_v            = frame_header->qm_v,
 151
 152             /* Segmentation */
 153             .segmentation_enabled         = frame_header->segmentation_enabled,
 154             .segmentation_update_map      = frame_header->segmentation_update_map,
 155             .segmentation_update_data     = frame_header->segmentation_update_data,
 156             .segmentation_temporal_update = frame_header->segmentation_temporal_update,
 157
 158             /* Loopfilter */
 159             .loop_filter_level[0]       = frame_header->loop_filter_level[0],
 160             .loop_filter_level[1]       = frame_header->loop_filter_level[1],
 161             .loop_filter_level_u        = frame_header->loop_filter_level[2],
 162             .loop_filter_level_v        = frame_header->loop_filter_level[3],
 163             .loop_filter_sharpness      = frame_header->loop_filter_sharpness,
 164             .loop_filter_delta_enabled  = frame_header->loop_filter_delta_enabled,
 165             .loop_filter_delta_update   = frame_header->loop_filter_delta_update,
 166             .loop_filter_mode_deltas[0] = frame_header->loop_filter_mode_deltas[0],
 167             .loop_filter_mode_deltas[1] = frame_header->loop_filter_mode_deltas[1],
 168             .delta_lf_present           = frame_header->delta_lf_present,
 169             .delta_lf_res               = frame_header->delta_lf_res,
 170             .delta_lf_multi             = frame_header->delta_lf_multi,
 171
 172             /* Restoration */
 173             .lr_type[0]      = remap_lr_type[frame_header->lr_type[0]],
 174             .lr_type[1]      = remap_lr_type[frame_header->lr_type[1]],
 175             .lr_type[2]      = remap_lr_type[frame_header->lr_type[2]],
 176             .lr_unit_size[0] = 1 + frame_header->lr_unit_shift,
 177             .lr_unit_size[1] = 1 + frame_header->lr_unit_shift - frame_header->lr_uv_shift,
 178             .lr_unit_size[2] = 1 + frame_header->lr_unit_shift - frame_header->lr_uv_shift,
 179
 180             /* Reference Frames */
 181             .temporal_layer_id = s->cur_frame.temporal_id,
 182             .spatial_layer_id  = s->cur_frame.spatial_id,
 183
 184             /* Film Grain Params */
 185             .apply_grain              = film_grain->apply_grain,
 186             .overlap_flag             = film_grain->overlap_flag,
 187             .scaling_shift_minus8     = film_grain->grain_scaling_minus_8,
 188             .chroma_scaling_from_luma = film_grain->chroma_scaling_from_luma,
 189             .ar_coeff_lag             = film_grain->ar_coeff_lag,
 190             .ar_coeff_shift_minus6    = film_grain->ar_coeff_shift_minus_6,
 191             .grain_scale_shift        = film_grain->grain_scale_shift,
 192             .clip_to_restricted_range = film_grain->clip_to_restricted_range,
 193             .num_y_points             = film_grain->num_y_points,
 194             .num_cb_points            = film_grain->num_cb_points,
 195             .num_cr_points            = film_grain->num_cr_points,
 196             .random_seed              = film_grain->grain_seed,
 197             .cb_mult                  = film_grain->cb_mult,
 198             .cb_luma_mult             = film_grain->cb_luma_mult,
 199             .cb_offset                = film_grain->cb_offset,
 200             .cr_mult                  = film_grain->cr_mult,
 201             .cr_luma_mult             = film_grain->cr_luma_mult,
 202             .cr_offset                = film_grain->cr_offset
 203         }
 204     };
 205
 206     /* Tiling Info */
 207     for (i = 0; i < frame_header->tile_cols; ++i) {
 208         ppc->tile_widths[i] = frame_header->width_in_sbs_minus_1[i] + 1;
 209     }
 210     for (i = 0; i < frame_header->tile_rows; ++i) {
 211         ppc->tile_heights[i] = frame_header->height_in_sbs_minus_1[i] + 1;
 212     }
 213
 214     /* CDEF */
 215     for (i = 0; i < (1 << frame_header->cdef_bits); ++i) {
 216         ppc->cdef_y_strength[i] = (frame_header->cdef_y_pri_strength[i] & 0x0F) | (frame_header->cdef_y_sec_strength[i] << 4);
 217         ppc->cdef_uv_strength[i] = (frame_header->cdef_uv_pri_strength[i] & 0x0F) | (frame_header->cdef_uv_sec_strength[i] << 4);
 218     }
 219
 220     /* Segmentation */
 221     for (i = 0; i < AV1_MAX_SEGMENTS; ++i) {
 222         ppc->segmentation_feature_mask[i] = 0;
 223         for (j = 0; j < AV1_SEG_LVL_MAX; ++j) {
 224             ppc->segmentation_feature_mask[i] |= frame_header->feature_enabled[i][j] << j;
 225             ppc->segmentation_feature_data[i][j] = frame_header->feature_value[i][j];
 226         }
 227     }
 228
 229     for (i = 0; i < AV1_NUM_REF_FRAMES; ++i) {
 230         /* Loopfilter */
 231         ppc->loop_filter_ref_deltas[i] = frame_header->loop_filter_ref_deltas[i];
 232
 233         /* Reference Frames */
 234         ppc->ref_frame_map[i] = ff_nvdec_get_ref_idx(s->ref[i].tf.f);
 235     }
 236
 237     if (frame_header->primary_ref_frame == AV1_PRIMARY_REF_NONE) {
 238         ppc->primary_ref_frame = -1;
 239     } else {
 240         int8_t pri_ref_idx = frame_header->ref_frame_idx[frame_header->primary_ref_frame];
 241         ppc->primary_ref_frame = ppc->ref_frame_map[pri_ref_idx];
 242     }
 243
 244     for (i = 0; i < AV1_REFS_PER_FRAME; ++i) {
 245         /* Ref Frame List */
 246         int8_t ref_idx = frame_header->ref_frame_idx[i];
 247         AVFrame *ref_frame = s->ref[ref_idx].tf.f;
 248
 249         ppc->ref_frame[i].index = ppc->ref_frame_map[ref_idx];
 250         ppc->ref_frame[i].width = ref_frame->width;
 251         ppc->ref_frame[i].height = ref_frame->height;
 252
 253         /* Global Motion */
 254         ppc->global_motion[i].invalid = !frame_header->is_global[AV1_REF_FRAME_LAST + i];
 255         ppc->global_motion[i].wmtype = s->cur_frame.gm_type[AV1_REF_FRAME_LAST + i];
 256         for (j = 0; j < 6; ++j) {
 257             ppc->global_motion[i].wmmat[j] = s->cur_frame.gm_params[AV1_REF_FRAME_LAST + i][j];
 258         }
 259     }
 260
 261     /* Film Grain Params */
 262     if (film_grain->apply_grain) {
 263         for (i = 0; i < 14; ++i) {
 264             ppc->scaling_points_y[i][0] = film_grain->point_y_value[i];
 265             ppc->scaling_points_y[i][1] = film_grain->point_y_scaling[i];
 266         }
 267         for (i = 0; i < 10; ++i) {
 268             ppc->scaling_points_cb[i][0] = film_grain->point_cb_value[i];
 269             ppc->scaling_points_cb[i][1] = film_grain->point_cb_scaling[i];
 270             ppc->scaling_points_cr[i][0] = film_grain->point_cr_value[i];
 271             ppc->scaling_points_cr[i][1] = film_grain->point_cr_scaling[i];
 272         }
 273         for (i = 0; i < 24; ++i) {
 274             ppc->ar_coeffs_y[i] = (short)film_grain->ar_coeffs_y_plus_128[i] - 128;
 275         }
 276         for (i = 0; i < 25; ++i) {
 277             ppc->ar_coeffs_cb[i] = (short)film_grain->ar_coeffs_cb_plus_128[i] - 128;
 278             ppc->ar_coeffs_cr[i] = (short)film_grain->ar_coeffs_cr_plus_128[i] - 128;
 279         }
 280     }
 281
 282     return 0;
 283 }
 284
 285 static int nvdec_av1_decode_slice(AVCodecContext *avctx, const uint8_t *buffer, uint32_t size)
 286 {
 287     const AV1DecContext *s = avctx->priv_data;
 288     const AV1RawFrameHeader *frame_header = s->raw_frame_header;
 289     NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
 290     void *tmp;
 291
 292     ctx->nb_slices = frame_header->tile_cols * frame_header->tile_rows;
 293
 294     tmp = av_fast_realloc(ctx->slice_offsets, &ctx->slice_offsets_allocated,
 295                           ctx->nb_slices * 2 * sizeof(*ctx->slice_offsets));
 296     if (!tmp) {
 297         return AVERROR(ENOMEM);
 298     }
 299     ctx->slice_offsets = tmp;
 300
 301     /* Shortcut if all tiles are in the same buffer */
 302     if (ctx->nb_slices == s->tg_end - s->tg_start + 1) {
 303         ctx->bitstream = (uint8_t*)buffer;
 304         ctx->bitstream_len = size;
 305
 306         for (int i = 0; i < ctx->nb_slices; ++i) {
 307             ctx->slice_offsets[i*2    ] = s->tile_group_info[i].tile_offset;
 308             ctx->slice_offsets[i*2 + 1] = ctx->slice_offsets[i*2] + s->tile_group_info[i].tile_size;
 309         }
 310
 311         return 0;
 312     }
 313
 314     tmp = av_fast_realloc(ctx->bitstream_internal, &ctx->bitstream_allocated,
 315                           ctx->bitstream_len + size);
 316     if (!tmp) {
 317         return AVERROR(ENOMEM);
 318     }
 319     ctx->bitstream = ctx->bitstream_internal = tmp;
 320
 321     memcpy(ctx->bitstream + ctx->bitstream_len, buffer, size);
 322
 323     for (uint32_t tile_num = s->tg_start; tile_num <= s->tg_end; ++tile_num) {
 324         ctx->slice_offsets[tile_num*2    ] = ctx->bitstream_len + s->tile_group_info[tile_num].tile_offset;
 325         ctx->slice_offsets[tile_num*2 + 1] = ctx->slice_offsets[tile_num*2] + s->tile_group_info[tile_num].tile_size;
 326     }
 327     ctx->bitstream_len += size;
 328
 329     return 0;
 330 }
 331
 332 static int nvdec_av1_frame_params(AVCodecContext *avctx, AVBufferRef *hw_frames_ctx)
 333 {
 334     /* Maximum of 8 reference frames, but potentially stored twice due to film grain */
 335     return ff_nvdec_frame_params(avctx, hw_frames_ctx, 8 * 2, 0);
 336 }
 337
 338 const AVHWAccel ff_av1_nvdec_hwaccel = {
 339     .name                 = "av1_nvdec",
 340     .type                 = AVMEDIA_TYPE_VIDEO,
 341     .id                   = AV_CODEC_ID_AV1,
 342     .pix_fmt              = AV_PIX_FMT_CUDA,
 343     .start_frame          = nvdec_av1_start_frame,
 344     .end_frame            = ff_nvdec_simple_end_frame,
 345     .decode_slice         = nvdec_av1_decode_slice,
 346     .frame_params         = nvdec_av1_frame_params,
 347     .init                 = ff_nvdec_decode_init,
 348     .uninit               = ff_nvdec_decode_uninit,
 349     .priv_data_size       = sizeof(NVDECContext),
 350 };