git.sesse.net Git - ffmpeg/blob - libavcodec/nvdec_h264.c

   1 /*
   2  * MPEG-4 Part 10 / AVC / H.264 HW decode acceleration through NVDEC
   3  *
   4  * Copyright (c) 2016 Anton Khirnov
   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 <stdint.h>
  24 #include <string.h>
  25
  26 #include "avcodec.h"
  27 #include "nvdec.h"
  28 #include "decode.h"
  29 #include "internal.h"
  30 #include "h264dec.h"
  31
  32 static void dpb_add(const H264Context *h, CUVIDH264DPBENTRY *dst, const H264Picture *src,
  33                     int frame_idx)
  34 {
  35     FrameDecodeData *fdd = (FrameDecodeData*)src->f->private_ref->data;
  36     const NVDECFrame *cf = fdd->hwaccel_priv;
  37
  38     dst->PicIdx             = cf ? cf->idx : -1;
  39     dst->FrameIdx           = frame_idx;
  40     dst->is_long_term       = src->long_ref;
  41     dst->not_existing       = 0;
  42     dst->used_for_reference = src->reference & 3;
  43     dst->FieldOrderCnt[0]   = src->field_poc[0];
  44     dst->FieldOrderCnt[1]   = src->field_poc[1];
  45 }
  46
  47 static int nvdec_h264_start_frame(AVCodecContext *avctx,
  48                                   const uint8_t *buffer, uint32_t size)
  49 {
  50     const H264Context *h = avctx->priv_data;
  51     const PPS *pps = h->ps.pps;
  52     const SPS *sps = h->ps.sps;
  53
  54     NVDECContext       *ctx = avctx->internal->hwaccel_priv_data;
  55     CUVIDPICPARAMS      *pp = &ctx->pic_params;
  56     CUVIDH264PICPARAMS *ppc = &pp->CodecSpecific.h264;
  57     FrameDecodeData *fdd;
  58     NVDECFrame *cf;
  59
  60     int i, dpb_size, ret;
  61
  62     ret = ff_nvdec_start_frame(avctx, h->cur_pic_ptr->f);
  63     if (ret < 0)
  64         return ret;
  65
  66     fdd = (FrameDecodeData*)h->cur_pic_ptr->f->private_ref->data;
  67     cf  = (NVDECFrame*)fdd->hwaccel_priv;
  68
  69     *pp = (CUVIDPICPARAMS) {
  70         .PicWidthInMbs     = h->mb_width,
  71         .FrameHeightInMbs  = h->mb_height,
  72         .CurrPicIdx        = cf->idx,
  73         .field_pic_flag    = FIELD_PICTURE(h),
  74         .bottom_field_flag = h->picture_structure == PICT_BOTTOM_FIELD,
  75         .second_field      = FIELD_PICTURE(h) && !h->first_field,
  76         .ref_pic_flag      = h->nal_ref_idc != 0,
  77         .intra_pic_flag    = 0,
  78
  79         .CodecSpecific.h264 = {
  80             .log2_max_frame_num_minus4            = sps->log2_max_frame_num - 4,
  81             .pic_order_cnt_type                   = sps->poc_type,
  82             .log2_max_pic_order_cnt_lsb_minus4    = FFMAX(sps->log2_max_poc_lsb - 4, 0),
  83             .delta_pic_order_always_zero_flag     = sps->delta_pic_order_always_zero_flag,
  84             .frame_mbs_only_flag                  = sps->frame_mbs_only_flag,
  85             .direct_8x8_inference_flag            = sps->direct_8x8_inference_flag,
  86             .num_ref_frames                       = sps->ref_frame_count,
  87             .residual_colour_transform_flag       = sps->residual_color_transform_flag,
  88             .bit_depth_luma_minus8                = sps->bit_depth_luma - 8,
  89             .bit_depth_chroma_minus8              = sps->bit_depth_chroma - 8,
  90             .qpprime_y_zero_transform_bypass_flag = sps->transform_bypass,
  91
  92             .entropy_coding_mode_flag               = pps->cabac,
  93             .pic_order_present_flag                 = pps->pic_order_present,
  94             .num_ref_idx_l0_active_minus1           = pps->ref_count[0] - 1,
  95             .num_ref_idx_l1_active_minus1           = pps->ref_count[1] - 1,
  96             .weighted_pred_flag                     = pps->weighted_pred,
  97             .weighted_bipred_idc                    = pps->weighted_bipred_idc,
  98             .pic_init_qp_minus26                    = pps->init_qp - 26,
  99             .deblocking_filter_control_present_flag = pps->deblocking_filter_parameters_present,
 100             .redundant_pic_cnt_present_flag         = pps->redundant_pic_cnt_present,
 101             .transform_8x8_mode_flag                = pps->transform_8x8_mode,
 102             .MbaffFrameFlag                         = sps->mb_aff && !FIELD_PICTURE(h),
 103             .constrained_intra_pred_flag            = pps->constrained_intra_pred,
 104             .chroma_qp_index_offset                 = pps->chroma_qp_index_offset[0],
 105             .second_chroma_qp_index_offset          = pps->chroma_qp_index_offset[1],
 106             .ref_pic_flag                           = h->nal_ref_idc != 0,
 107             .frame_num                              = h->poc.frame_num,
 108             .CurrFieldOrderCnt[0]                   = h->cur_pic_ptr->field_poc[0],
 109             .CurrFieldOrderCnt[1]                   = h->cur_pic_ptr->field_poc[1],
 110         },
 111     };
 112
 113     memcpy(ppc->WeightScale4x4,    pps->scaling_matrix4,    sizeof(ppc->WeightScale4x4));
 114     memcpy(ppc->WeightScale8x8[0], pps->scaling_matrix8[0], sizeof(ppc->WeightScale8x8[0]));
 115     memcpy(ppc->WeightScale8x8[1], pps->scaling_matrix8[3], sizeof(ppc->WeightScale8x8[0]));
 116
 117     dpb_size = 0;
 118     for (i = 0; i < h->short_ref_count; i++)
 119         dpb_add(h, &ppc->dpb[dpb_size++], h->short_ref[i], h->short_ref[i]->frame_num);
 120     for (i = 0; i < 16; i++) {
 121         if (h->long_ref[i])
 122             dpb_add(h, &ppc->dpb[dpb_size++], h->long_ref[i], i);
 123     }
 124
 125     for (i = dpb_size; i < FF_ARRAY_ELEMS(ppc->dpb); i++)
 126         ppc->dpb[i].PicIdx = -1;
 127
 128     return 0;
 129 }
 130
 131 static int nvdec_h264_decode_slice(AVCodecContext *avctx, const uint8_t *buffer,
 132                                    uint32_t size)
 133 {
 134     NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
 135     void *tmp;
 136
 137     tmp = av_fast_realloc(ctx->bitstream, &ctx->bitstream_allocated,
 138                           ctx->bitstream_len + size + 3);
 139     if (!tmp)
 140         return AVERROR(ENOMEM);
 141     ctx->bitstream = tmp;
 142
 143     tmp = av_fast_realloc(ctx->slice_offsets, &ctx->slice_offsets_allocated,
 144                           (ctx->nb_slices + 1) * sizeof(*ctx->slice_offsets));
 145     if (!tmp)
 146         return AVERROR(ENOMEM);
 147     ctx->slice_offsets = tmp;
 148
 149     AV_WB24(ctx->bitstream + ctx->bitstream_len, 1);
 150     memcpy(ctx->bitstream + ctx->bitstream_len + 3, buffer, size);
 151     ctx->slice_offsets[ctx->nb_slices] = ctx->bitstream_len ;
 152     ctx->bitstream_len += size + 3;
 153     ctx->nb_slices++;
 154
 155     return 0;
 156 }
 157
 158 static int nvdec_h264_frame_params(AVCodecContext *avctx,
 159                                    AVBufferRef *hw_frames_ctx)
 160 {
 161     const H264Context *h = avctx->priv_data;
 162     const SPS       *sps = h->ps.sps;
 163     return ff_nvdec_frame_params(avctx, hw_frames_ctx, sps->ref_frame_count + sps->num_reorder_frames);
 164 }
 165
 166 AVHWAccel ff_h264_nvdec_hwaccel = {
 167     .name                 = "h264_nvdec",
 168     .type                 = AVMEDIA_TYPE_VIDEO,
 169     .id                   = AV_CODEC_ID_H264,
 170     .pix_fmt              = AV_PIX_FMT_CUDA,
 171     .start_frame          = nvdec_h264_start_frame,
 172     .end_frame            = ff_nvdec_end_frame,
 173     .decode_slice         = nvdec_h264_decode_slice,
 174     .frame_params         = nvdec_h264_frame_params,
 175     .init                 = ff_nvdec_decode_init,
 176     .uninit               = ff_nvdec_decode_uninit,
 177     .priv_data_size       = sizeof(NVDECContext),
 178 };