avcodec/idctdsp: Clear idct/idct_add for studio profile
[ffmpeg.git] / libavcodec / nvdec_vp9.c
1 /*
2  * VP9 HW decode acceleration through NVDEC
3  *
4  * Copyright (c) 2016 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 "libavutil/pixdesc.h"
24
25 #include "avcodec.h"
26 #include "nvdec.h"
27 #include "decode.h"
28 #include "internal.h"
29 #include "vp9shared.h"
30
31 static int nvdec_vp9_start_frame(AVCodecContext *avctx, const uint8_t *buffer, uint32_t size)
32 {
33     VP9SharedContext *h = avctx->priv_data;
34     const AVPixFmtDescriptor *pixdesc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
35
36     NVDECContext      *ctx = avctx->internal->hwaccel_priv_data;
37     CUVIDPICPARAMS     *pp = &ctx->pic_params;
38     CUVIDVP9PICPARAMS *ppc = &pp->CodecSpecific.vp9;
39     FrameDecodeData *fdd;
40     NVDECFrame *cf;
41     AVFrame *cur_frame = h->frames[CUR_FRAME].tf.f;
42
43     int ret, i;
44
45     ret = ff_nvdec_start_frame(avctx, cur_frame);
46     if (ret < 0)
47         return ret;
48
49     fdd = (FrameDecodeData*)cur_frame->private_ref->data;
50     cf  = (NVDECFrame*)fdd->hwaccel_priv;
51
52     *pp = (CUVIDPICPARAMS) {
53         .PicWidthInMbs     = (cur_frame->width  + 15) / 16,
54         .FrameHeightInMbs  = (cur_frame->height + 15) / 16,
55         .CurrPicIdx        = cf->idx,
56
57         .CodecSpecific.vp9 = {
58             .width                    = cur_frame->width,
59             .height                   = cur_frame->height,
60
61             .LastRefIdx               = ff_nvdec_get_ref_idx(h->refs[h->h.refidx[0]].f),
62             .GoldenRefIdx             = ff_nvdec_get_ref_idx(h->refs[h->h.refidx[1]].f),
63             .AltRefIdx                = ff_nvdec_get_ref_idx(h->refs[h->h.refidx[2]].f),
64
65             .profile                  = h->h.profile,
66             .frameContextIdx          = h->h.framectxid,
67             .frameType                = !h->h.keyframe,
68             .showFrame                = !h->h.invisible,
69             .errorResilient           = h->h.errorres,
70             .frameParallelDecoding    = h->h.parallelmode,
71             .subSamplingX             = pixdesc->log2_chroma_w,
72             .subSamplingY             = pixdesc->log2_chroma_h,
73             .intraOnly                = h->h.intraonly,
74             .allow_high_precision_mv  = h->h.keyframe ? 0 : h->h.highprecisionmvs,
75             .refreshEntropyProbs      = h->h.refreshctx,
76
77             .bitDepthMinus8Luma       = pixdesc->comp[0].depth - 8,
78             .bitDepthMinus8Chroma     = pixdesc->comp[1].depth - 8,
79
80             .loopFilterLevel          = h->h.filter.level,
81             .loopFilterSharpness      = h->h.filter.sharpness,
82             .modeRefLfEnabled         = h->h.lf_delta.enabled,
83
84             .log2_tile_columns        = h->h.tiling.log2_tile_cols,
85             .log2_tile_rows           = h->h.tiling.log2_tile_rows,
86
87             .segmentEnabled           = h->h.segmentation.enabled,
88             .segmentMapUpdate         = h->h.segmentation.update_map,
89             .segmentMapTemporalUpdate = h->h.segmentation.temporal,
90             .segmentFeatureMode       = h->h.segmentation.absolute_vals,
91
92             .qpYAc                    = h->h.yac_qi,
93             .qpYDc                    = h->h.ydc_qdelta,
94             .qpChDc                   = h->h.uvdc_qdelta,
95             .qpChAc                   = h->h.uvac_qdelta,
96
97             .resetFrameContext        = h->h.resetctx,
98             .mcomp_filter_type        = h->h.filtermode ^ (h->h.filtermode <= 1),
99
100             .frameTagSize             = h->h.uncompressed_header_size,
101             .offsetToDctParts         = h->h.compressed_header_size,
102
103             .refFrameSignBias[0]      = 0,
104         }
105     };
106
107     for (i = 0; i < 2; i++)
108         ppc->mbModeLfDelta[i] = h->h.lf_delta.mode[i];
109
110     for (i = 0; i < 4; i++)
111         ppc->mbRefLfDelta[i] = h->h.lf_delta.ref[i];
112
113     for (i = 0; i < 7; i++)
114         ppc->mb_segment_tree_probs[i] = h->h.segmentation.prob[i];
115
116     for (i = 0; i < 3; i++) {
117         ppc->activeRefIdx[i] = h->h.refidx[i];
118         ppc->segment_pred_probs[i] = h->h.segmentation.pred_prob[i];
119         ppc->refFrameSignBias[i + 1] = h->h.signbias[i];
120     }
121
122     for (i = 0; i < 8; i++) {
123         ppc->segmentFeatureEnable[i][0] = h->h.segmentation.feat[i].q_enabled;
124         ppc->segmentFeatureEnable[i][1] = h->h.segmentation.feat[i].lf_enabled;
125         ppc->segmentFeatureEnable[i][2] = h->h.segmentation.feat[i].ref_enabled;
126         ppc->segmentFeatureEnable[i][3] = h->h.segmentation.feat[i].skip_enabled;
127
128         ppc->segmentFeatureData[i][0] = h->h.segmentation.feat[i].q_val;
129         ppc->segmentFeatureData[i][1] = h->h.segmentation.feat[i].lf_val;
130         ppc->segmentFeatureData[i][2] = h->h.segmentation.feat[i].ref_val;
131         ppc->segmentFeatureData[i][3] = 0;
132     }
133
134     switch (avctx->colorspace) {
135     default:
136     case AVCOL_SPC_UNSPECIFIED:
137         ppc->colorSpace = 0;
138         break;
139     case AVCOL_SPC_BT470BG:
140         ppc->colorSpace = 1;
141         break;
142     case AVCOL_SPC_BT709:
143         ppc->colorSpace = 2;
144         break;
145     case AVCOL_SPC_SMPTE170M:
146         ppc->colorSpace = 3;
147         break;
148     case AVCOL_SPC_SMPTE240M:
149         ppc->colorSpace = 4;
150         break;
151     case AVCOL_SPC_BT2020_NCL:
152         ppc->colorSpace = 5;
153         break;
154     case AVCOL_SPC_RESERVED:
155         ppc->colorSpace = 6;
156         break;
157     case AVCOL_SPC_RGB:
158         ppc->colorSpace = 7;
159         break;
160     }
161
162     return 0;
163 }
164
165 static int nvdec_vp9_frame_params(AVCodecContext *avctx,
166                                   AVBufferRef *hw_frames_ctx)
167 {
168     // VP9 uses a fixed size pool of 8 possible reference frames
169     return ff_nvdec_frame_params(avctx, hw_frames_ctx, 8);
170 }
171
172 const AVHWAccel ff_vp9_nvdec_hwaccel = {
173     .name                 = "vp9_nvdec",
174     .type                 = AVMEDIA_TYPE_VIDEO,
175     .id                   = AV_CODEC_ID_VP9,
176     .pix_fmt              = AV_PIX_FMT_CUDA,
177     .start_frame          = nvdec_vp9_start_frame,
178     .end_frame            = ff_nvdec_simple_end_frame,
179     .decode_slice         = ff_nvdec_simple_decode_slice,
180     .frame_params         = nvdec_vp9_frame_params,
181     .init                 = ff_nvdec_decode_init,
182     .uninit               = ff_nvdec_decode_uninit,
183     .priv_data_size       = sizeof(NVDECContext),
184 };