X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fvaapi_h264.c;h=97d4387d717fe88d90cd5084febbccdb00354f0e;hb=f6171471e6cfbd472f7ab71c12aa87510a07d4fd;hp=6823f6d7c6491d920bbb1427489c67eb2fbc8e41;hpb=066aafced4dc6c7c9e7b37082635472249f1e93e;p=ffmpeg diff --git a/libavcodec/vaapi_h264.c b/libavcodec/vaapi_h264.c index 6823f6d7c64..97d4387d717 100644 --- a/libavcodec/vaapi_h264.c +++ b/libavcodec/vaapi_h264.c @@ -20,9 +20,10 @@ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA */ -#include "vaapi_internal.h" -#include "h264.h" -#include "mpegutils.h" +#include "h264dec.h" +#include "h264_ps.h" +#include "hwaccel.h" +#include "vaapi_decode.h" /** * @file @@ -52,14 +53,14 @@ static void init_vaapi_pic(VAPictureH264 *va_pic) * supersedes pic's field type if nonzero. */ static void fill_vaapi_pic(VAPictureH264 *va_pic, - H264Picture *pic, + const H264Picture *pic, int pic_structure) { if (pic_structure == 0) pic_structure = pic->reference; pic_structure &= PICT_FRAME; /* PICT_TOP_FIELD|PICT_BOTTOM_FIELD */ - va_pic->picture_id = ff_vaapi_get_surface_id(&pic->f); + va_pic->picture_id = ff_vaapi_get_surface_id(pic->f); va_pic->frame_idx = pic->long_ref ? pic->pic_id : pic->frame_num; va_pic->flags = 0; @@ -90,7 +91,7 @@ typedef struct DPB { * available. The decoded picture buffer's size must be large enough * to receive the new VA API picture object. */ -static int dpb_add(DPB *dpb, H264Picture *pic) +static int dpb_add(DPB *dpb, const H264Picture *pic) { int i; @@ -99,7 +100,7 @@ static int dpb_add(DPB *dpb, H264Picture *pic) for (i = 0; i < dpb->size; i++) { VAPictureH264 * const va_pic = &dpb->va_pics[i]; - if (va_pic->picture_id == ff_vaapi_get_surface_id(&pic->f)) { + if (va_pic->picture_id == ff_vaapi_get_surface_id(pic->f)) { VAPictureH264 temp_va_pic; fill_vaapi_pic(&temp_va_pic, pic, 0); @@ -122,7 +123,7 @@ static int dpb_add(DPB *dpb, H264Picture *pic) /** Fill in VA API reference frames array. */ static int fill_vaapi_ReferenceFrames(VAPictureParameterBufferH264 *pic_param, - H264Context *h) + const H264Context *h) { DPB dpb; int i; @@ -134,13 +135,13 @@ static int fill_vaapi_ReferenceFrames(VAPictureParameterBufferH264 *pic_param, init_vaapi_pic(&dpb.va_pics[i]); for (i = 0; i < h->short_ref_count; i++) { - H264Picture * const pic = h->short_ref[i]; + const H264Picture *pic = h->short_ref[i]; if (pic && pic->reference && dpb_add(&dpb, pic) < 0) return -1; } for (i = 0; i < 16; i++) { - H264Picture * const pic = h->long_ref[i]; + const H264Picture *pic = h->long_ref[i]; if (pic && pic->reference && dpb_add(&dpb, pic) < 0) return -1; } @@ -156,13 +157,14 @@ static int fill_vaapi_ReferenceFrames(VAPictureParameterBufferH264 *pic_param, * @param[in] ref_count The number of reference pictures in ref_list */ static void fill_vaapi_RefPicList(VAPictureH264 RefPicList[32], - H264Picture *ref_list, + const H264Ref *ref_list, unsigned int ref_count) { unsigned int i, n = 0; for (i = 0; i < ref_count; i++) if (ref_list[i].reference) - fill_vaapi_pic(&RefPicList[n++], &ref_list[i], 0); + fill_vaapi_pic(&RefPicList[n++], ref_list[i].parent, + ref_list[i].reference); for (; n < 32; n++) init_vaapi_pic(&RefPicList[n]); @@ -183,7 +185,7 @@ static void fill_vaapi_RefPicList(VAPictureH264 RefPicList[32], * @param[out] chroma_weight VA API plain chroma weight table * @param[out] chroma_offset VA API plain chroma offset table */ -static void fill_vaapi_plain_pred_weight_table(H264Context *h, +static void fill_vaapi_plain_pred_weight_table(const H264Context *h, int list, unsigned char *luma_weight_flag, short luma_weight[32], @@ -192,28 +194,28 @@ static void fill_vaapi_plain_pred_weight_table(H264Context *h, short chroma_weight[32][2], short chroma_offset[32][2]) { - H264SliceContext *sl = &h->slice_ctx[0]; + const H264SliceContext *sl = &h->slice_ctx[0]; unsigned int i, j; - *luma_weight_flag = sl->luma_weight_flag[list]; - *chroma_weight_flag = sl->chroma_weight_flag[list]; + *luma_weight_flag = sl->pwt.luma_weight_flag[list]; + *chroma_weight_flag = sl->pwt.chroma_weight_flag[list]; - for (i = 0; i < h->ref_count[list]; i++) { + for (i = 0; i < sl->ref_count[list]; i++) { /* VA API also wants the inferred (default) values, not only what is available in the bitstream (7.4.3.2). */ - if (sl->luma_weight_flag[list]) { - luma_weight[i] = sl->luma_weight[i][list][0]; - luma_offset[i] = sl->luma_weight[i][list][1]; + if (sl->pwt.luma_weight_flag[list]) { + luma_weight[i] = sl->pwt.luma_weight[i][list][0]; + luma_offset[i] = sl->pwt.luma_weight[i][list][1]; } else { - luma_weight[i] = 1 << sl->luma_log2_weight_denom; + luma_weight[i] = 1 << sl->pwt.luma_log2_weight_denom; luma_offset[i] = 0; } for (j = 0; j < 2; j++) { - if (sl->chroma_weight_flag[list]) { - chroma_weight[i][j] = sl->chroma_weight[i][list][j][0]; - chroma_offset[i][j] = sl->chroma_weight[i][list][j][1]; + if (sl->pwt.chroma_weight_flag[list]) { + chroma_weight[i][j] = sl->pwt.chroma_weight[i][list][j][0]; + chroma_offset[i][j] = sl->pwt.chroma_weight[i][list][j][1]; } else { - chroma_weight[i][j] = 1 << sl->chroma_log2_weight_denom; + chroma_weight[i][j] = 1 << sl->pwt.chroma_log2_weight_denom; chroma_offset[i][j] = 0; } } @@ -225,89 +227,103 @@ static int vaapi_h264_start_frame(AVCodecContext *avctx, av_unused const uint8_t *buffer, av_unused uint32_t size) { - H264Context * const h = avctx->priv_data; - struct vaapi_context * const vactx = avctx->hwaccel_context; - VAPictureParameterBufferH264 *pic_param; - VAIQMatrixBufferH264 *iq_matrix; - - av_dlog(avctx, "vaapi_h264_start_frame()\n"); + const H264Context *h = avctx->priv_data; + VAAPIDecodePicture *pic = h->cur_pic_ptr->hwaccel_picture_private; + const PPS *pps = h->ps.pps; + const SPS *sps = h->ps.sps; + VAPictureParameterBufferH264 pic_param; + VAIQMatrixBufferH264 iq_matrix; + int err; + + pic->output_surface = ff_vaapi_get_surface_id(h->cur_pic_ptr->f); + + pic_param = (VAPictureParameterBufferH264) { + .picture_width_in_mbs_minus1 = h->mb_width - 1, + .picture_height_in_mbs_minus1 = h->mb_height - 1, + .bit_depth_luma_minus8 = sps->bit_depth_luma - 8, + .bit_depth_chroma_minus8 = sps->bit_depth_chroma - 8, + .num_ref_frames = sps->ref_frame_count, + .seq_fields.bits = { + .chroma_format_idc = sps->chroma_format_idc, + .residual_colour_transform_flag = sps->residual_color_transform_flag, + .gaps_in_frame_num_value_allowed_flag = sps->gaps_in_frame_num_allowed_flag, + .frame_mbs_only_flag = sps->frame_mbs_only_flag, + .mb_adaptive_frame_field_flag = sps->mb_aff, + .direct_8x8_inference_flag = sps->direct_8x8_inference_flag, + .MinLumaBiPredSize8x8 = sps->level_idc >= 31, /* A.3.3.2 */ + .log2_max_frame_num_minus4 = sps->log2_max_frame_num - 4, + .pic_order_cnt_type = sps->poc_type, + .log2_max_pic_order_cnt_lsb_minus4 = sps->log2_max_poc_lsb - 4, + .delta_pic_order_always_zero_flag = sps->delta_pic_order_always_zero_flag, + }, + .num_slice_groups_minus1 = pps->slice_group_count - 1, + .slice_group_map_type = pps->mb_slice_group_map_type, + .slice_group_change_rate_minus1 = 0, /* FMO is not implemented */ + .pic_init_qp_minus26 = pps->init_qp - 26, + .pic_init_qs_minus26 = pps->init_qs - 26, + .chroma_qp_index_offset = pps->chroma_qp_index_offset[0], + .second_chroma_qp_index_offset = pps->chroma_qp_index_offset[1], + .pic_fields.bits = { + .entropy_coding_mode_flag = pps->cabac, + .weighted_pred_flag = pps->weighted_pred, + .weighted_bipred_idc = pps->weighted_bipred_idc, + .transform_8x8_mode_flag = pps->transform_8x8_mode, + .field_pic_flag = h->picture_structure != PICT_FRAME, + .constrained_intra_pred_flag = pps->constrained_intra_pred, + .pic_order_present_flag = pps->pic_order_present, + .deblocking_filter_control_present_flag = pps->deblocking_filter_parameters_present, + .redundant_pic_cnt_present_flag = pps->redundant_pic_cnt_present, + .reference_pic_flag = h->nal_ref_idc != 0, + }, + .frame_num = h->poc.frame_num, + }; + + fill_vaapi_pic(&pic_param.CurrPic, h->cur_pic_ptr, h->picture_structure); + err = fill_vaapi_ReferenceFrames(&pic_param, h); + if (err < 0) + goto fail; + + err = ff_vaapi_decode_make_param_buffer(avctx, pic, + VAPictureParameterBufferType, + &pic_param, sizeof(pic_param)); + if (err < 0) + goto fail; + + memcpy(iq_matrix.ScalingList4x4, + pps->scaling_matrix4, sizeof(iq_matrix.ScalingList4x4)); + memcpy(iq_matrix.ScalingList8x8[0], + pps->scaling_matrix8[0], sizeof(iq_matrix.ScalingList8x8[0])); + memcpy(iq_matrix.ScalingList8x8[1], + pps->scaling_matrix8[3], sizeof(iq_matrix.ScalingList8x8[0])); + + err = ff_vaapi_decode_make_param_buffer(avctx, pic, + VAIQMatrixBufferType, + &iq_matrix, sizeof(iq_matrix)); + if (err < 0) + goto fail; - vactx->slice_param_size = sizeof(VASliceParameterBufferH264); - - /* Fill in VAPictureParameterBufferH264. */ - pic_param = ff_vaapi_alloc_pic_param(vactx, sizeof(VAPictureParameterBufferH264)); - if (!pic_param) - return -1; - fill_vaapi_pic(&pic_param->CurrPic, h->cur_pic_ptr, h->picture_structure); - if (fill_vaapi_ReferenceFrames(pic_param, h) < 0) - return -1; - pic_param->picture_width_in_mbs_minus1 = h->mb_width - 1; - pic_param->picture_height_in_mbs_minus1 = h->mb_height - 1; - pic_param->bit_depth_luma_minus8 = h->sps.bit_depth_luma - 8; - pic_param->bit_depth_chroma_minus8 = h->sps.bit_depth_chroma - 8; - pic_param->num_ref_frames = h->sps.ref_frame_count; - pic_param->seq_fields.value = 0; /* reset all bits */ - pic_param->seq_fields.bits.chroma_format_idc = h->sps.chroma_format_idc; - pic_param->seq_fields.bits.residual_colour_transform_flag = h->sps.residual_color_transform_flag; /* XXX: only for 4:4:4 high profile? */ - pic_param->seq_fields.bits.gaps_in_frame_num_value_allowed_flag = h->sps.gaps_in_frame_num_allowed_flag; - pic_param->seq_fields.bits.frame_mbs_only_flag = h->sps.frame_mbs_only_flag; - pic_param->seq_fields.bits.mb_adaptive_frame_field_flag = h->sps.mb_aff; - pic_param->seq_fields.bits.direct_8x8_inference_flag = h->sps.direct_8x8_inference_flag; - pic_param->seq_fields.bits.MinLumaBiPredSize8x8 = h->sps.level_idc >= 31; /* A.3.3.2 */ - pic_param->seq_fields.bits.log2_max_frame_num_minus4 = h->sps.log2_max_frame_num - 4; - pic_param->seq_fields.bits.pic_order_cnt_type = h->sps.poc_type; - pic_param->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = h->sps.log2_max_poc_lsb - 4; - pic_param->seq_fields.bits.delta_pic_order_always_zero_flag = h->sps.delta_pic_order_always_zero_flag; - pic_param->num_slice_groups_minus1 = h->pps.slice_group_count - 1; - pic_param->slice_group_map_type = h->pps.mb_slice_group_map_type; - pic_param->slice_group_change_rate_minus1 = 0; /* XXX: unimplemented in Libav */ - pic_param->pic_init_qp_minus26 = h->pps.init_qp - 26; - pic_param->pic_init_qs_minus26 = h->pps.init_qs - 26; - pic_param->chroma_qp_index_offset = h->pps.chroma_qp_index_offset[0]; - pic_param->second_chroma_qp_index_offset = h->pps.chroma_qp_index_offset[1]; - pic_param->pic_fields.value = 0; /* reset all bits */ - pic_param->pic_fields.bits.entropy_coding_mode_flag = h->pps.cabac; - pic_param->pic_fields.bits.weighted_pred_flag = h->pps.weighted_pred; - pic_param->pic_fields.bits.weighted_bipred_idc = h->pps.weighted_bipred_idc; - pic_param->pic_fields.bits.transform_8x8_mode_flag = h->pps.transform_8x8_mode; - pic_param->pic_fields.bits.field_pic_flag = h->picture_structure != PICT_FRAME; - pic_param->pic_fields.bits.constrained_intra_pred_flag = h->pps.constrained_intra_pred; - pic_param->pic_fields.bits.pic_order_present_flag = h->pps.pic_order_present; - pic_param->pic_fields.bits.deblocking_filter_control_present_flag = h->pps.deblocking_filter_parameters_present; - pic_param->pic_fields.bits.redundant_pic_cnt_present_flag = h->pps.redundant_pic_cnt_present; - pic_param->pic_fields.bits.reference_pic_flag = h->nal_ref_idc != 0; - pic_param->frame_num = h->frame_num; - - /* Fill in VAIQMatrixBufferH264. */ - iq_matrix = ff_vaapi_alloc_iq_matrix(vactx, sizeof(VAIQMatrixBufferH264)); - if (!iq_matrix) - return -1; - memcpy(iq_matrix->ScalingList4x4, h->pps.scaling_matrix4, sizeof(iq_matrix->ScalingList4x4)); - memcpy(iq_matrix->ScalingList8x8[0], h->pps.scaling_matrix8[0], sizeof(iq_matrix->ScalingList8x8[0])); - memcpy(iq_matrix->ScalingList8x8[1], h->pps.scaling_matrix8[3], sizeof(iq_matrix->ScalingList8x8[0])); return 0; + +fail: + ff_vaapi_decode_cancel(avctx, pic); + return err; } /** End a hardware decoding based frame. */ static int vaapi_h264_end_frame(AVCodecContext *avctx) { - struct vaapi_context * const vactx = avctx->hwaccel_context; - H264Context * const h = avctx->priv_data; + const H264Context *h = avctx->priv_data; + VAAPIDecodePicture *pic = h->cur_pic_ptr->hwaccel_picture_private; + H264SliceContext *sl = &h->slice_ctx[0]; int ret; - av_dlog(avctx, "vaapi_h264_end_frame()\n"); - ret = ff_vaapi_commit_slices(vactx); + ret = ff_vaapi_decode_issue(avctx, pic); if (ret < 0) goto finish; - ret = ff_vaapi_render_picture(vactx, ff_vaapi_get_surface_id(&h->cur_pic_ptr->f)); - if (ret < 0) - goto finish; - - ff_h264_draw_horiz_band(h, 0, h->avctx->height); + ff_h264_draw_horiz_band(h, sl, 0, h->avctx->height); finish: - ff_vaapi_common_end_frame(avctx); return ret; } @@ -316,49 +332,74 @@ static int vaapi_h264_decode_slice(AVCodecContext *avctx, const uint8_t *buffer, uint32_t size) { - H264Context * const h = avctx->priv_data; - H264SliceContext *sl = &h->slice_ctx[0]; - VASliceParameterBufferH264 *slice_param; - - av_dlog(avctx, "vaapi_h264_decode_slice(): buffer %p, size %d\n", - buffer, size); - - /* Fill in VASliceParameterBufferH264. */ - slice_param = (VASliceParameterBufferH264 *)ff_vaapi_alloc_slice(avctx->hwaccel_context, buffer, size); - if (!slice_param) - return -1; - slice_param->slice_data_bit_offset = get_bits_count(&h->gb) + 8; /* bit buffer started beyond nal_unit_type */ - slice_param->first_mb_in_slice = (h->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + h->mb_x; - slice_param->slice_type = ff_h264_get_slice_type(sl); - slice_param->direct_spatial_mv_pred_flag = sl->slice_type == AV_PICTURE_TYPE_B ? sl->direct_spatial_mv_pred : 0; - slice_param->num_ref_idx_l0_active_minus1 = h->list_count > 0 ? h->ref_count[0] - 1 : 0; - slice_param->num_ref_idx_l1_active_minus1 = h->list_count > 1 ? h->ref_count[1] - 1 : 0; - slice_param->cabac_init_idc = h->cabac_init_idc; - slice_param->slice_qp_delta = sl->qscale - h->pps.init_qp; - slice_param->disable_deblocking_filter_idc = h->deblocking_filter < 2 ? !h->deblocking_filter : h->deblocking_filter; - slice_param->slice_alpha_c0_offset_div2 = h->slice_alpha_c0_offset / 2; - slice_param->slice_beta_offset_div2 = h->slice_beta_offset / 2; - slice_param->luma_log2_weight_denom = sl->luma_log2_weight_denom; - slice_param->chroma_log2_weight_denom = sl->chroma_log2_weight_denom; - - fill_vaapi_RefPicList(slice_param->RefPicList0, h->ref_list[0], h->list_count > 0 ? h->ref_count[0] : 0); - fill_vaapi_RefPicList(slice_param->RefPicList1, h->ref_list[1], h->list_count > 1 ? h->ref_count[1] : 0); + const H264Context *h = avctx->priv_data; + VAAPIDecodePicture *pic = h->cur_pic_ptr->hwaccel_picture_private; + const H264SliceContext *sl = &h->slice_ctx[0]; + VASliceParameterBufferH264 slice_param; + int err; + + slice_param = (VASliceParameterBufferH264) { + .slice_data_size = size, + .slice_data_offset = 0, + .slice_data_flag = VA_SLICE_DATA_FLAG_ALL, + .slice_data_bit_offset = get_bits_count(&sl->gb), + .first_mb_in_slice = (sl->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + sl->mb_x, + .slice_type = ff_h264_get_slice_type(sl), + .direct_spatial_mv_pred_flag = sl->slice_type == AV_PICTURE_TYPE_B ? sl->direct_spatial_mv_pred : 0, + .num_ref_idx_l0_active_minus1 = sl->list_count > 0 ? sl->ref_count[0] - 1 : 0, + .num_ref_idx_l1_active_minus1 = sl->list_count > 1 ? sl->ref_count[1] - 1 : 0, + .cabac_init_idc = sl->cabac_init_idc, + .slice_qp_delta = sl->qscale - h->ps.pps->init_qp, + .disable_deblocking_filter_idc = sl->deblocking_filter < 2 ? !sl->deblocking_filter : sl->deblocking_filter, + .slice_alpha_c0_offset_div2 = sl->slice_alpha_c0_offset / 2, + .slice_beta_offset_div2 = sl->slice_beta_offset / 2, + .luma_log2_weight_denom = sl->pwt.luma_log2_weight_denom, + .chroma_log2_weight_denom = sl->pwt.chroma_log2_weight_denom, + }; + + fill_vaapi_RefPicList(slice_param.RefPicList0, sl->ref_list[0], + sl->list_count > 0 ? sl->ref_count[0] : 0); + fill_vaapi_RefPicList(slice_param.RefPicList1, sl->ref_list[1], + sl->list_count > 1 ? sl->ref_count[1] : 0); fill_vaapi_plain_pred_weight_table(h, 0, - &slice_param->luma_weight_l0_flag, slice_param->luma_weight_l0, slice_param->luma_offset_l0, - &slice_param->chroma_weight_l0_flag, slice_param->chroma_weight_l0, slice_param->chroma_offset_l0); + &slice_param.luma_weight_l0_flag, + slice_param.luma_weight_l0, + slice_param.luma_offset_l0, + &slice_param.chroma_weight_l0_flag, + slice_param.chroma_weight_l0, + slice_param.chroma_offset_l0); fill_vaapi_plain_pred_weight_table(h, 1, - &slice_param->luma_weight_l1_flag, slice_param->luma_weight_l1, slice_param->luma_offset_l1, - &slice_param->chroma_weight_l1_flag, slice_param->chroma_weight_l1, slice_param->chroma_offset_l1); + &slice_param.luma_weight_l1_flag, + slice_param.luma_weight_l1, + slice_param.luma_offset_l1, + &slice_param.chroma_weight_l1_flag, + slice_param.chroma_weight_l1, + slice_param.chroma_offset_l1); + + err = ff_vaapi_decode_make_slice_buffer(avctx, pic, + &slice_param, sizeof(slice_param), + buffer, size); + if (err) { + ff_vaapi_decode_cancel(avctx, pic); + return err; + } + return 0; } -AVHWAccel ff_h264_vaapi_hwaccel = { - .name = "h264_vaapi", - .type = AVMEDIA_TYPE_VIDEO, - .id = AV_CODEC_ID_H264, - .pix_fmt = AV_PIX_FMT_VAAPI_VLD, - .start_frame = vaapi_h264_start_frame, - .end_frame = vaapi_h264_end_frame, - .decode_slice = vaapi_h264_decode_slice, +const AVHWAccel ff_h264_vaapi_hwaccel = { + .name = "h264_vaapi", + .type = AVMEDIA_TYPE_VIDEO, + .id = AV_CODEC_ID_H264, + .pix_fmt = AV_PIX_FMT_VAAPI, + .start_frame = &vaapi_h264_start_frame, + .end_frame = &vaapi_h264_end_frame, + .decode_slice = &vaapi_h264_decode_slice, + .frame_priv_data_size = sizeof(VAAPIDecodePicture), + .init = &ff_vaapi_decode_init, + .uninit = &ff_vaapi_decode_uninit, + .frame_params = &ff_vaapi_common_frame_params, + .priv_data_size = sizeof(VAAPIDecodeContext), + .caps_internal = HWACCEL_CAP_ASYNC_SAFE, };