git.sesse.net Git - ffmpeg/blob - libavcodec/vaapi_h264.c

   1 /*
   2  * H.264 HW decode acceleration through VA API
   3  *
   4  * Copyright (C) 2008-2009 Splitted-Desktop Systems
   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 "h264dec.h"
  24 #include "h264_ps.h"
  25 #include "hwaccel.h"
  26 #include "vaapi_decode.h"
  27
  28 /**
  29  * @file
  30  * This file implements the glue code between FFmpeg's and VA API's
  31  * structures for H.264 decoding.
  32  */
  33
  34 /**
  35  * Initialize an empty VA API picture.
  36  *
  37  * VA API requires a fixed-size reference picture array.
  38  */
  39 static void init_vaapi_pic(VAPictureH264 *va_pic)
  40 {
  41     va_pic->picture_id          = VA_INVALID_ID;
  42     va_pic->flags               = VA_PICTURE_H264_INVALID;
  43     va_pic->TopFieldOrderCnt    = 0;
  44     va_pic->BottomFieldOrderCnt = 0;
  45 }
  46
  47 /**
  48  * Translate an FFmpeg Picture into its VA API form.
  49  *
  50  * @param[out] va_pic          A pointer to VA API's own picture struct
  51  * @param[in]  pic             A pointer to the FFmpeg picture struct to convert
  52  * @param[in]  pic_structure   The picture field type (as defined in mpegvideo.h),
  53  *                             supersedes pic's field type if nonzero.
  54  */
  55 static void fill_vaapi_pic(VAPictureH264 *va_pic,
  56                            const H264Picture *pic,
  57                            int            pic_structure)
  58 {
  59     if (pic_structure == 0)
  60         pic_structure = pic->reference;
  61     pic_structure &= PICT_FRAME; /* PICT_TOP_FIELD|PICT_BOTTOM_FIELD */
  62
  63     va_pic->picture_id = ff_vaapi_get_surface_id(pic->f);
  64     va_pic->frame_idx  = pic->long_ref ? pic->pic_id : pic->frame_num;
  65
  66     va_pic->flags      = 0;
  67     if (pic_structure != PICT_FRAME)
  68         va_pic->flags |= (pic_structure & PICT_TOP_FIELD) ? VA_PICTURE_H264_TOP_FIELD : VA_PICTURE_H264_BOTTOM_FIELD;
  69     if (pic->reference)
  70         va_pic->flags |= pic->long_ref ? VA_PICTURE_H264_LONG_TERM_REFERENCE : VA_PICTURE_H264_SHORT_TERM_REFERENCE;
  71
  72     va_pic->TopFieldOrderCnt = 0;
  73     if (pic->field_poc[0] != INT_MAX)
  74         va_pic->TopFieldOrderCnt = pic->field_poc[0];
  75
  76     va_pic->BottomFieldOrderCnt = 0;
  77     if (pic->field_poc[1] != INT_MAX)
  78         va_pic->BottomFieldOrderCnt = pic->field_poc[1];
  79 }
  80
  81 /** Decoded Picture Buffer (DPB). */
  82 typedef struct DPB {
  83     int            size;        ///< Current number of reference frames in the DPB
  84     int            max_size;    ///< Max number of reference frames. This is FF_ARRAY_ELEMS(VAPictureParameterBufferH264.ReferenceFrames)
  85     VAPictureH264 *va_pics;     ///< Pointer to VAPictureParameterBufferH264.ReferenceFrames array
  86 } DPB;
  87
  88 /**
  89  * Append picture to the decoded picture buffer, in a VA API form that
  90  * merges the second field picture attributes with the first, if
  91  * available.  The decoded picture buffer's size must be large enough
  92  * to receive the new VA API picture object.
  93  */
  94 static int dpb_add(DPB *dpb, const H264Picture *pic)
  95 {
  96     int i;
  97
  98     if (dpb->size >= dpb->max_size)
  99         return -1;
 100
 101     for (i = 0; i < dpb->size; i++) {
 102         VAPictureH264 * const va_pic = &dpb->va_pics[i];
 103         if (va_pic->picture_id == ff_vaapi_get_surface_id(pic->f)) {
 104             VAPictureH264 temp_va_pic;
 105             fill_vaapi_pic(&temp_va_pic, pic, 0);
 106
 107             if ((temp_va_pic.flags ^ va_pic->flags) & (VA_PICTURE_H264_TOP_FIELD | VA_PICTURE_H264_BOTTOM_FIELD)) {
 108                 va_pic->flags |= temp_va_pic.flags & (VA_PICTURE_H264_TOP_FIELD | VA_PICTURE_H264_BOTTOM_FIELD);
 109                 /* Merge second field */
 110                 if (temp_va_pic.flags & VA_PICTURE_H264_TOP_FIELD) {
 111                     va_pic->TopFieldOrderCnt    = temp_va_pic.TopFieldOrderCnt;
 112                 } else {
 113                     va_pic->BottomFieldOrderCnt = temp_va_pic.BottomFieldOrderCnt;
 114                 }
 115             }
 116             return 0;
 117         }
 118     }
 119
 120     fill_vaapi_pic(&dpb->va_pics[dpb->size++], pic, 0);
 121     return 0;
 122 }
 123
 124 /** Fill in VA API reference frames array. */
 125 static int fill_vaapi_ReferenceFrames(VAPictureParameterBufferH264 *pic_param,
 126                                       const H264Context            *h)
 127 {
 128     DPB dpb;
 129     int i;
 130
 131     dpb.size     = 0;
 132     dpb.max_size = FF_ARRAY_ELEMS(pic_param->ReferenceFrames);
 133     dpb.va_pics  = pic_param->ReferenceFrames;
 134     for (i = 0; i < dpb.max_size; i++)
 135         init_vaapi_pic(&dpb.va_pics[i]);
 136
 137     for (i = 0; i < h->short_ref_count; i++) {
 138         const H264Picture *pic = h->short_ref[i];
 139         if (pic && pic->reference && dpb_add(&dpb, pic) < 0)
 140             return -1;
 141     }
 142
 143     for (i = 0; i < 16; i++) {
 144         const H264Picture *pic = h->long_ref[i];
 145         if (pic && pic->reference && dpb_add(&dpb, pic) < 0)
 146             return -1;
 147     }
 148     return 0;
 149 }
 150
 151 /**
 152  * Fill in VA API reference picture lists from the FFmpeg reference
 153  * picture list.
 154  *
 155  * @param[out] RefPicList  VA API internal reference picture list
 156  * @param[in]  ref_list    A pointer to the FFmpeg reference list
 157  * @param[in]  ref_count   The number of reference pictures in ref_list
 158  */
 159 static void fill_vaapi_RefPicList(VAPictureH264 RefPicList[32],
 160                                   const H264Ref *ref_list,
 161                                   unsigned int  ref_count)
 162 {
 163     unsigned int i, n = 0;
 164     for (i = 0; i < ref_count; i++)
 165         if (ref_list[i].reference)
 166             fill_vaapi_pic(&RefPicList[n++], ref_list[i].parent,
 167                            ref_list[i].reference);
 168
 169     for (; n < 32; n++)
 170         init_vaapi_pic(&RefPicList[n]);
 171 }
 172
 173 /**
 174  * Fill in prediction weight table.
 175  *
 176  * VA API requires a plain prediction weight table as it does not infer
 177  * any value.
 178  *
 179  * @param[in]  h                   A pointer to the current H.264 context
 180  * @param[in]  list                The reference frame list index to use
 181  * @param[out] luma_weight_flag    VA API plain luma weight flag
 182  * @param[out] luma_weight         VA API plain luma weight table
 183  * @param[out] luma_offset         VA API plain luma offset table
 184  * @param[out] chroma_weight_flag  VA API plain chroma weight flag
 185  * @param[out] chroma_weight       VA API plain chroma weight table
 186  * @param[out] chroma_offset       VA API plain chroma offset table
 187  */
 188 static void fill_vaapi_plain_pred_weight_table(const H264Context *h,
 189                                                int            list,
 190                                                unsigned char *luma_weight_flag,
 191                                                short          luma_weight[32],
 192                                                short          luma_offset[32],
 193                                                unsigned char *chroma_weight_flag,
 194                                                short          chroma_weight[32][2],
 195                                                short          chroma_offset[32][2])
 196 {
 197     const H264SliceContext *sl = &h->slice_ctx[0];
 198     unsigned int i, j;
 199
 200     *luma_weight_flag    = sl->pwt.luma_weight_flag[list];
 201     *chroma_weight_flag  = sl->pwt.chroma_weight_flag[list];
 202
 203     for (i = 0; i < sl->ref_count[list]; i++) {
 204         /* VA API also wants the inferred (default) values, not
 205            only what is available in the bitstream (7.4.3.2). */
 206         if (sl->pwt.luma_weight_flag[list]) {
 207             luma_weight[i] = sl->pwt.luma_weight[i][list][0];
 208             luma_offset[i] = sl->pwt.luma_weight[i][list][1];
 209         } else {
 210             luma_weight[i] = 1 << sl->pwt.luma_log2_weight_denom;
 211             luma_offset[i] = 0;
 212         }
 213         for (j = 0; j < 2; j++) {
 214             if (sl->pwt.chroma_weight_flag[list]) {
 215                 chroma_weight[i][j] = sl->pwt.chroma_weight[i][list][j][0];
 216                 chroma_offset[i][j] = sl->pwt.chroma_weight[i][list][j][1];
 217             } else {
 218                 chroma_weight[i][j] = 1 << sl->pwt.chroma_log2_weight_denom;
 219                 chroma_offset[i][j] = 0;
 220             }
 221         }
 222     }
 223 }
 224
 225 /** Initialize and start decoding a frame with VA API. */
 226 static int vaapi_h264_start_frame(AVCodecContext          *avctx,
 227                                   av_unused const uint8_t *buffer,
 228                                   av_unused uint32_t       size)
 229 {
 230     const H264Context *h = avctx->priv_data;
 231     VAAPIDecodePicture *pic = h->cur_pic_ptr->hwaccel_picture_private;
 232     const PPS *pps = h->ps.pps;
 233     const SPS *sps = h->ps.sps;
 234     VAPictureParameterBufferH264 pic_param;
 235     VAIQMatrixBufferH264 iq_matrix;
 236     int err;
 237
 238     pic->output_surface = ff_vaapi_get_surface_id(h->cur_pic_ptr->f);
 239
 240     pic_param = (VAPictureParameterBufferH264) {
 241         .picture_width_in_mbs_minus1                = h->mb_width - 1,
 242         .picture_height_in_mbs_minus1               = h->mb_height - 1,
 243         .bit_depth_luma_minus8                      = sps->bit_depth_luma - 8,
 244         .bit_depth_chroma_minus8                    = sps->bit_depth_chroma - 8,
 245         .num_ref_frames                             = sps->ref_frame_count,
 246         .seq_fields.bits = {
 247             .chroma_format_idc                      = sps->chroma_format_idc,
 248             .residual_colour_transform_flag         = sps->residual_color_transform_flag,
 249             .gaps_in_frame_num_value_allowed_flag   = sps->gaps_in_frame_num_allowed_flag,
 250             .frame_mbs_only_flag                    = sps->frame_mbs_only_flag,
 251             .mb_adaptive_frame_field_flag           = sps->mb_aff,
 252             .direct_8x8_inference_flag              = sps->direct_8x8_inference_flag,
 253             .MinLumaBiPredSize8x8                   = sps->level_idc >= 31, /* A.3.3.2 */
 254             .log2_max_frame_num_minus4              = sps->log2_max_frame_num - 4,
 255             .pic_order_cnt_type                     = sps->poc_type,
 256             .log2_max_pic_order_cnt_lsb_minus4      = sps->log2_max_poc_lsb - 4,
 257             .delta_pic_order_always_zero_flag       = sps->delta_pic_order_always_zero_flag,
 258         },
 259         .num_slice_groups_minus1                    = pps->slice_group_count - 1,
 260         .slice_group_map_type                       = pps->mb_slice_group_map_type,
 261         .slice_group_change_rate_minus1             = 0, /* FMO is not implemented */
 262         .pic_init_qp_minus26                        = pps->init_qp - 26,
 263         .pic_init_qs_minus26                        = pps->init_qs - 26,
 264         .chroma_qp_index_offset                     = pps->chroma_qp_index_offset[0],
 265         .second_chroma_qp_index_offset              = pps->chroma_qp_index_offset[1],
 266         .pic_fields.bits = {
 267             .entropy_coding_mode_flag               = pps->cabac,
 268             .weighted_pred_flag                     = pps->weighted_pred,
 269             .weighted_bipred_idc                    = pps->weighted_bipred_idc,
 270             .transform_8x8_mode_flag                = pps->transform_8x8_mode,
 271             .field_pic_flag                         = h->picture_structure != PICT_FRAME,
 272             .constrained_intra_pred_flag            = pps->constrained_intra_pred,
 273             .pic_order_present_flag                 = pps->pic_order_present,
 274             .deblocking_filter_control_present_flag = pps->deblocking_filter_parameters_present,
 275             .redundant_pic_cnt_present_flag         = pps->redundant_pic_cnt_present,
 276             .reference_pic_flag                     = h->nal_ref_idc != 0,
 277         },
 278         .frame_num                                  = h->poc.frame_num,
 279     };
 280
 281     fill_vaapi_pic(&pic_param.CurrPic, h->cur_pic_ptr, h->picture_structure);
 282     err = fill_vaapi_ReferenceFrames(&pic_param, h);
 283     if (err < 0)
 284         goto fail;
 285
 286     err = ff_vaapi_decode_make_param_buffer(avctx, pic,
 287                                             VAPictureParameterBufferType,
 288                                             &pic_param, sizeof(pic_param));
 289     if (err < 0)
 290         goto fail;
 291
 292     memcpy(iq_matrix.ScalingList4x4,
 293            pps->scaling_matrix4, sizeof(iq_matrix.ScalingList4x4));
 294     memcpy(iq_matrix.ScalingList8x8[0],
 295            pps->scaling_matrix8[0], sizeof(iq_matrix.ScalingList8x8[0]));
 296     memcpy(iq_matrix.ScalingList8x8[1],
 297            pps->scaling_matrix8[3], sizeof(iq_matrix.ScalingList8x8[0]));
 298
 299     err = ff_vaapi_decode_make_param_buffer(avctx, pic,
 300                                             VAIQMatrixBufferType,
 301                                             &iq_matrix, sizeof(iq_matrix));
 302     if (err < 0)
 303         goto fail;
 304
 305     return 0;
 306
 307 fail:
 308     ff_vaapi_decode_cancel(avctx, pic);
 309     return err;
 310 }
 311
 312 /** End a hardware decoding based frame. */
 313 static int vaapi_h264_end_frame(AVCodecContext *avctx)
 314 {
 315     const H264Context *h = avctx->priv_data;
 316     VAAPIDecodePicture *pic = h->cur_pic_ptr->hwaccel_picture_private;
 317     H264SliceContext *sl = &h->slice_ctx[0];
 318     int ret;
 319
 320     ret = ff_vaapi_decode_issue(avctx, pic);
 321     if (ret < 0)
 322         goto finish;
 323
 324     ff_h264_draw_horiz_band(h, sl, 0, h->avctx->height);
 325
 326 finish:
 327     return ret;
 328 }
 329
 330 /** Decode the given H.264 slice with VA API. */
 331 static int vaapi_h264_decode_slice(AVCodecContext *avctx,
 332                                    const uint8_t  *buffer,
 333                                    uint32_t        size)
 334 {
 335     const H264Context *h = avctx->priv_data;
 336     VAAPIDecodePicture *pic = h->cur_pic_ptr->hwaccel_picture_private;
 337     const H264SliceContext *sl  = &h->slice_ctx[0];
 338     VASliceParameterBufferH264 slice_param;
 339     int err;
 340
 341     slice_param = (VASliceParameterBufferH264) {
 342         .slice_data_size               = size,
 343         .slice_data_offset             = 0,
 344         .slice_data_flag               = VA_SLICE_DATA_FLAG_ALL,
 345         .slice_data_bit_offset         = get_bits_count(&sl->gb),
 346         .first_mb_in_slice             = (sl->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + sl->mb_x,
 347         .slice_type                    = ff_h264_get_slice_type(sl),
 348         .direct_spatial_mv_pred_flag   = sl->slice_type == AV_PICTURE_TYPE_B ? sl->direct_spatial_mv_pred : 0,
 349         .num_ref_idx_l0_active_minus1  = sl->list_count > 0 ? sl->ref_count[0] - 1 : 0,
 350         .num_ref_idx_l1_active_minus1  = sl->list_count > 1 ? sl->ref_count[1] - 1 : 0,
 351         .cabac_init_idc                = sl->cabac_init_idc,
 352         .slice_qp_delta                = sl->qscale - h->ps.pps->init_qp,
 353         .disable_deblocking_filter_idc = sl->deblocking_filter < 2 ? !sl->deblocking_filter : sl->deblocking_filter,
 354         .slice_alpha_c0_offset_div2    = sl->slice_alpha_c0_offset / 2,
 355         .slice_beta_offset_div2        = sl->slice_beta_offset     / 2,
 356         .luma_log2_weight_denom        = sl->pwt.luma_log2_weight_denom,
 357         .chroma_log2_weight_denom      = sl->pwt.chroma_log2_weight_denom,
 358     };
 359
 360     fill_vaapi_RefPicList(slice_param.RefPicList0, sl->ref_list[0],
 361                           sl->list_count > 0 ? sl->ref_count[0] : 0);
 362     fill_vaapi_RefPicList(slice_param.RefPicList1, sl->ref_list[1],
 363                           sl->list_count > 1 ? sl->ref_count[1] : 0);
 364
 365     fill_vaapi_plain_pred_weight_table(h, 0,
 366                                        &slice_param.luma_weight_l0_flag,
 367                                        slice_param.luma_weight_l0,
 368                                        slice_param.luma_offset_l0,
 369                                        &slice_param.chroma_weight_l0_flag,
 370                                        slice_param.chroma_weight_l0,
 371                                        slice_param.chroma_offset_l0);
 372     fill_vaapi_plain_pred_weight_table(h, 1,
 373                                        &slice_param.luma_weight_l1_flag,
 374                                        slice_param.luma_weight_l1,
 375                                        slice_param.luma_offset_l1,
 376                                        &slice_param.chroma_weight_l1_flag,
 377                                        slice_param.chroma_weight_l1,
 378                                        slice_param.chroma_offset_l1);
 379
 380     err = ff_vaapi_decode_make_slice_buffer(avctx, pic,
 381                                             &slice_param, sizeof(slice_param),
 382                                             buffer, size);
 383     if (err) {
 384         ff_vaapi_decode_cancel(avctx, pic);
 385         return err;
 386     }
 387
 388     return 0;
 389 }
 390
 391 const AVHWAccel ff_h264_vaapi_hwaccel = {
 392     .name                 = "h264_vaapi",
 393     .type                 = AVMEDIA_TYPE_VIDEO,
 394     .id                   = AV_CODEC_ID_H264,
 395     .pix_fmt              = AV_PIX_FMT_VAAPI,
 396     .start_frame          = &vaapi_h264_start_frame,
 397     .end_frame            = &vaapi_h264_end_frame,
 398     .decode_slice         = &vaapi_h264_decode_slice,
 399     .frame_priv_data_size = sizeof(VAAPIDecodePicture),
 400     .init                 = &ff_vaapi_decode_init,
 401     .uninit               = &ff_vaapi_decode_uninit,
 402     .frame_params         = &ff_vaapi_common_frame_params,
 403     .priv_data_size       = sizeof(VAAPIDecodeContext),
 404     .caps_internal        = HWACCEL_CAP_ASYNC_SAFE,
 405 };