]> git.sesse.net Git - ffmpeg/blob - libavcodec/vaapi_h264.c
fft-test: add option to set cpuflag mask
[ffmpeg] / 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 Libav.
7  *
8  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22
23 #include "vaapi_internal.h"
24 #include "h264.h"
25
26 /**
27  * @file
28  * This file implements the glue code between Libav's and VA API's
29  * structures for H.264 decoding.
30  */
31
32 /**
33  * Initialize an empty VA API picture.
34  *
35  * VA API requires a fixed-size reference picture array.
36  */
37 static void init_vaapi_pic(VAPictureH264 *va_pic)
38 {
39     va_pic->picture_id          = VA_INVALID_ID;
40     va_pic->flags               = VA_PICTURE_H264_INVALID;
41     va_pic->TopFieldOrderCnt    = 0;
42     va_pic->BottomFieldOrderCnt = 0;
43 }
44
45 /**
46  * Translate an Libav Picture into its VA API form.
47  *
48  * @param[out] va_pic          A pointer to VA API's own picture struct
49  * @param[in]  pic             A pointer to the Libav picture struct to convert
50  * @param[in]  pic_structure   The picture field type (as defined in mpegvideo.h),
51  *                             supersedes pic's field type if nonzero.
52  */
53 static void fill_vaapi_pic(VAPictureH264 *va_pic,
54                            Picture       *pic,
55                            int            pic_structure)
56 {
57     if (pic_structure == 0)
58         pic_structure = pic->f.reference;
59     pic_structure &= PICT_FRAME; /* PICT_TOP_FIELD|PICT_BOTTOM_FIELD */
60
61     va_pic->picture_id = ff_vaapi_get_surface_id(pic);
62     va_pic->frame_idx  = pic->long_ref ? pic->pic_id : pic->frame_num;
63
64     va_pic->flags      = 0;
65     if (pic_structure != PICT_FRAME)
66         va_pic->flags |= (pic_structure & PICT_TOP_FIELD) ? VA_PICTURE_H264_TOP_FIELD : VA_PICTURE_H264_BOTTOM_FIELD;
67     if (pic->f.reference)
68         va_pic->flags |= pic->long_ref ? VA_PICTURE_H264_LONG_TERM_REFERENCE : VA_PICTURE_H264_SHORT_TERM_REFERENCE;
69
70     va_pic->TopFieldOrderCnt = 0;
71     if (pic->field_poc[0] != INT_MAX)
72         va_pic->TopFieldOrderCnt = pic->field_poc[0];
73
74     va_pic->BottomFieldOrderCnt = 0;
75     if (pic->field_poc[1] != INT_MAX)
76         va_pic->BottomFieldOrderCnt = pic->field_poc[1];
77 }
78
79 /** Decoded Picture Buffer (DPB). */
80 typedef struct DPB {
81     int            size;        ///< Current number of reference frames in the DPB
82     int            max_size;    ///< Max number of reference frames. This is FF_ARRAY_ELEMS(VAPictureParameterBufferH264.ReferenceFrames)
83     VAPictureH264 *va_pics;     ///< Pointer to VAPictureParameterBufferH264.ReferenceFrames array
84 } DPB;
85
86 /**
87  * Append picture to the decoded picture buffer, in a VA API form that
88  * merges the second field picture attributes with the first, if
89  * available.  The decoded picture buffer's size must be large enough
90  * to receive the new VA API picture object.
91  */
92 static int dpb_add(DPB *dpb, Picture *pic)
93 {
94     int i;
95
96     if (dpb->size >= dpb->max_size)
97         return -1;
98
99     for (i = 0; i < dpb->size; i++) {
100         VAPictureH264 * const va_pic = &dpb->va_pics[i];
101         if (va_pic->picture_id == ff_vaapi_get_surface_id(pic)) {
102             VAPictureH264 temp_va_pic;
103             fill_vaapi_pic(&temp_va_pic, pic, 0);
104
105             if ((temp_va_pic.flags ^ va_pic->flags) & (VA_PICTURE_H264_TOP_FIELD | VA_PICTURE_H264_BOTTOM_FIELD)) {
106                 va_pic->flags |= temp_va_pic.flags & (VA_PICTURE_H264_TOP_FIELD | VA_PICTURE_H264_BOTTOM_FIELD);
107                 /* Merge second field */
108                 if (temp_va_pic.flags & VA_PICTURE_H264_TOP_FIELD) {
109                     va_pic->TopFieldOrderCnt    = temp_va_pic.TopFieldOrderCnt;
110                 } else {
111                     va_pic->BottomFieldOrderCnt = temp_va_pic.BottomFieldOrderCnt;
112                 }
113             }
114             return 0;
115         }
116     }
117
118     fill_vaapi_pic(&dpb->va_pics[dpb->size++], pic, 0);
119     return 0;
120 }
121
122 /** Fill in VA API reference frames array. */
123 static int fill_vaapi_ReferenceFrames(VAPictureParameterBufferH264 *pic_param,
124                                       H264Context                  *h)
125 {
126     DPB dpb;
127     int i;
128
129     dpb.size     = 0;
130     dpb.max_size = FF_ARRAY_ELEMS(pic_param->ReferenceFrames);
131     dpb.va_pics  = pic_param->ReferenceFrames;
132     for (i = 0; i < dpb.max_size; i++)
133         init_vaapi_pic(&dpb.va_pics[i]);
134
135     for (i = 0; i < h->short_ref_count; i++) {
136         Picture * const pic = h->short_ref[i];
137         if (pic && pic->f.reference && dpb_add(&dpb, pic) < 0)
138             return -1;
139     }
140
141     for (i = 0; i < 16; i++) {
142         Picture * const pic = h->long_ref[i];
143         if (pic && pic->f.reference && dpb_add(&dpb, pic) < 0)
144             return -1;
145     }
146     return 0;
147 }
148
149 /**
150  * Fill in VA API reference picture lists from the Libav reference
151  * picture list.
152  *
153  * @param[out] RefPicList  VA API internal reference picture list
154  * @param[in]  ref_list    A pointer to the Libav reference list
155  * @param[in]  ref_count   The number of reference pictures in ref_list
156  */
157 static void fill_vaapi_RefPicList(VAPictureH264 RefPicList[32],
158                                   Picture      *ref_list,
159                                   unsigned int  ref_count)
160 {
161     unsigned int i, n = 0;
162     for (i = 0; i < ref_count; i++)
163         if (ref_list[i].f.reference)
164             fill_vaapi_pic(&RefPicList[n++], &ref_list[i], 0);
165
166     for (; n < 32; n++)
167         init_vaapi_pic(&RefPicList[n]);
168 }
169
170 /**
171  * Fill in prediction weight table.
172  *
173  * VA API requires a plain prediction weight table as it does not infer
174  * any value.
175  *
176  * @param[in]  h                   A pointer to the current H.264 context
177  * @param[in]  list                The reference frame list index to use
178  * @param[out] luma_weight_flag    VA API plain luma weight flag
179  * @param[out] luma_weight         VA API plain luma weight table
180  * @param[out] luma_offset         VA API plain luma offset table
181  * @param[out] chroma_weight_flag  VA API plain chroma weight flag
182  * @param[out] chroma_weight       VA API plain chroma weight table
183  * @param[out] chroma_offset       VA API plain chroma offset table
184  */
185 static void fill_vaapi_plain_pred_weight_table(H264Context   *h,
186                                                int            list,
187                                                unsigned char *luma_weight_flag,
188                                                short          luma_weight[32],
189                                                short          luma_offset[32],
190                                                unsigned char *chroma_weight_flag,
191                                                short          chroma_weight[32][2],
192                                                short          chroma_offset[32][2])
193 {
194     unsigned int i, j;
195
196     *luma_weight_flag    = h->luma_weight_flag[list];
197     *chroma_weight_flag  = h->chroma_weight_flag[list];
198
199     for (i = 0; i < h->ref_count[list]; i++) {
200         /* VA API also wants the inferred (default) values, not
201            only what is available in the bitstream (7.4.3.2). */
202         if (h->luma_weight_flag[list]) {
203             luma_weight[i] = h->luma_weight[i][list][0];
204             luma_offset[i] = h->luma_weight[i][list][1];
205         } else {
206             luma_weight[i] = 1 << h->luma_log2_weight_denom;
207             luma_offset[i] = 0;
208         }
209         for (j = 0; j < 2; j++) {
210             if (h->chroma_weight_flag[list]) {
211                 chroma_weight[i][j] = h->chroma_weight[i][list][j][0];
212                 chroma_offset[i][j] = h->chroma_weight[i][list][j][1];
213             } else {
214                 chroma_weight[i][j] = 1 << h->chroma_log2_weight_denom;
215                 chroma_offset[i][j] = 0;
216             }
217         }
218     }
219 }
220
221 /** Initialize and start decoding a frame with VA API. */
222 static int start_frame(AVCodecContext          *avctx,
223                        av_unused const uint8_t *buffer,
224                        av_unused uint32_t       size)
225 {
226     H264Context * const h = avctx->priv_data;
227     MpegEncContext * const s = &h->s;
228     struct vaapi_context * const vactx = avctx->hwaccel_context;
229     VAPictureParameterBufferH264 *pic_param;
230     VAIQMatrixBufferH264 *iq_matrix;
231
232     av_dlog(avctx, "start_frame()\n");
233
234     vactx->slice_param_size = sizeof(VASliceParameterBufferH264);
235
236     /* Fill in VAPictureParameterBufferH264. */
237     pic_param = ff_vaapi_alloc_pic_param(vactx, sizeof(VAPictureParameterBufferH264));
238     if (!pic_param)
239         return -1;
240     fill_vaapi_pic(&pic_param->CurrPic, s->current_picture_ptr, s->picture_structure);
241     if (fill_vaapi_ReferenceFrames(pic_param, h) < 0)
242         return -1;
243     pic_param->picture_width_in_mbs_minus1                      = s->mb_width - 1;
244     pic_param->picture_height_in_mbs_minus1                     = s->mb_height - 1;
245     pic_param->bit_depth_luma_minus8                            = h->sps.bit_depth_luma - 8;
246     pic_param->bit_depth_chroma_minus8                          = h->sps.bit_depth_chroma - 8;
247     pic_param->num_ref_frames                                   = h->sps.ref_frame_count;
248     pic_param->seq_fields.value                                 = 0; /* reset all bits */
249     pic_param->seq_fields.bits.chroma_format_idc                = h->sps.chroma_format_idc;
250     pic_param->seq_fields.bits.residual_colour_transform_flag   = h->sps.residual_color_transform_flag; /* XXX: only for 4:4:4 high profile? */
251     pic_param->seq_fields.bits.gaps_in_frame_num_value_allowed_flag = h->sps.gaps_in_frame_num_allowed_flag;
252     pic_param->seq_fields.bits.frame_mbs_only_flag              = h->sps.frame_mbs_only_flag;
253     pic_param->seq_fields.bits.mb_adaptive_frame_field_flag     = h->sps.mb_aff;
254     pic_param->seq_fields.bits.direct_8x8_inference_flag        = h->sps.direct_8x8_inference_flag;
255     pic_param->seq_fields.bits.MinLumaBiPredSize8x8             = h->sps.level_idc >= 31; /* A.3.3.2 */
256     pic_param->seq_fields.bits.log2_max_frame_num_minus4        = h->sps.log2_max_frame_num - 4;
257     pic_param->seq_fields.bits.pic_order_cnt_type               = h->sps.poc_type;
258     pic_param->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = h->sps.log2_max_poc_lsb - 4;
259     pic_param->seq_fields.bits.delta_pic_order_always_zero_flag = h->sps.delta_pic_order_always_zero_flag;
260     pic_param->num_slice_groups_minus1                          = h->pps.slice_group_count - 1;
261     pic_param->slice_group_map_type                             = h->pps.mb_slice_group_map_type;
262     pic_param->slice_group_change_rate_minus1                   = 0; /* XXX: unimplemented in Libav */
263     pic_param->pic_init_qp_minus26                              = h->pps.init_qp - 26;
264     pic_param->pic_init_qs_minus26                              = h->pps.init_qs - 26;
265     pic_param->chroma_qp_index_offset                           = h->pps.chroma_qp_index_offset[0];
266     pic_param->second_chroma_qp_index_offset                    = h->pps.chroma_qp_index_offset[1];
267     pic_param->pic_fields.value                                 = 0; /* reset all bits */
268     pic_param->pic_fields.bits.entropy_coding_mode_flag         = h->pps.cabac;
269     pic_param->pic_fields.bits.weighted_pred_flag               = h->pps.weighted_pred;
270     pic_param->pic_fields.bits.weighted_bipred_idc              = h->pps.weighted_bipred_idc;
271     pic_param->pic_fields.bits.transform_8x8_mode_flag          = h->pps.transform_8x8_mode;
272     pic_param->pic_fields.bits.field_pic_flag                   = s->picture_structure != PICT_FRAME;
273     pic_param->pic_fields.bits.constrained_intra_pred_flag      = h->pps.constrained_intra_pred;
274     pic_param->pic_fields.bits.pic_order_present_flag           = h->pps.pic_order_present;
275     pic_param->pic_fields.bits.deblocking_filter_control_present_flag = h->pps.deblocking_filter_parameters_present;
276     pic_param->pic_fields.bits.redundant_pic_cnt_present_flag   = h->pps.redundant_pic_cnt_present;
277     pic_param->pic_fields.bits.reference_pic_flag               = h->nal_ref_idc != 0;
278     pic_param->frame_num                                        = h->frame_num;
279
280     /* Fill in VAIQMatrixBufferH264. */
281     iq_matrix = ff_vaapi_alloc_iq_matrix(vactx, sizeof(VAIQMatrixBufferH264));
282     if (!iq_matrix)
283         return -1;
284     memcpy(iq_matrix->ScalingList4x4, h->pps.scaling_matrix4, sizeof(iq_matrix->ScalingList4x4));
285     memcpy(iq_matrix->ScalingList8x8, h->pps.scaling_matrix8, sizeof(iq_matrix->ScalingList8x8));
286     return 0;
287 }
288
289 /** End a hardware decoding based frame. */
290 static int end_frame(AVCodecContext *avctx)
291 {
292     H264Context * const h = avctx->priv_data;
293
294     av_dlog(avctx, "end_frame()\n");
295     return ff_vaapi_common_end_frame(&h->s);
296 }
297
298 /** Decode the given H.264 slice with VA API. */
299 static int decode_slice(AVCodecContext *avctx,
300                         const uint8_t  *buffer,
301                         uint32_t        size)
302 {
303     H264Context * const h = avctx->priv_data;
304     MpegEncContext * const s = &h->s;
305     VASliceParameterBufferH264 *slice_param;
306
307     av_dlog(avctx, "decode_slice(): buffer %p, size %d\n", buffer, size);
308
309     /* Fill in VASliceParameterBufferH264. */
310     slice_param = (VASliceParameterBufferH264 *)ff_vaapi_alloc_slice(avctx->hwaccel_context, buffer, size);
311     if (!slice_param)
312         return -1;
313     slice_param->slice_data_bit_offset          = get_bits_count(&h->s.gb) + 8; /* bit buffer started beyond nal_unit_type */
314     slice_param->first_mb_in_slice              = (s->mb_y >> FIELD_OR_MBAFF_PICTURE) * s->mb_width + s->mb_x;
315     slice_param->slice_type                     = ff_h264_get_slice_type(h);
316     slice_param->direct_spatial_mv_pred_flag    = h->slice_type == AV_PICTURE_TYPE_B ? h->direct_spatial_mv_pred : 0;
317     slice_param->num_ref_idx_l0_active_minus1   = h->list_count > 0 ? h->ref_count[0] - 1 : 0;
318     slice_param->num_ref_idx_l1_active_minus1   = h->list_count > 1 ? h->ref_count[1] - 1 : 0;
319     slice_param->cabac_init_idc                 = h->cabac_init_idc;
320     slice_param->slice_qp_delta                 = s->qscale - h->pps.init_qp;
321     slice_param->disable_deblocking_filter_idc  = h->deblocking_filter < 2 ? !h->deblocking_filter : h->deblocking_filter;
322     slice_param->slice_alpha_c0_offset_div2     = h->slice_alpha_c0_offset / 2 - 26;
323     slice_param->slice_beta_offset_div2         = h->slice_beta_offset     / 2 - 26;
324     slice_param->luma_log2_weight_denom         = h->luma_log2_weight_denom;
325     slice_param->chroma_log2_weight_denom       = h->chroma_log2_weight_denom;
326
327     fill_vaapi_RefPicList(slice_param->RefPicList0, h->ref_list[0], h->list_count > 0 ? h->ref_count[0] : 0);
328     fill_vaapi_RefPicList(slice_param->RefPicList1, h->ref_list[1], h->list_count > 1 ? h->ref_count[1] : 0);
329
330     fill_vaapi_plain_pred_weight_table(h, 0,
331                                        &slice_param->luma_weight_l0_flag,   slice_param->luma_weight_l0,   slice_param->luma_offset_l0,
332                                        &slice_param->chroma_weight_l0_flag, slice_param->chroma_weight_l0, slice_param->chroma_offset_l0);
333     fill_vaapi_plain_pred_weight_table(h, 1,
334                                        &slice_param->luma_weight_l1_flag,   slice_param->luma_weight_l1,   slice_param->luma_offset_l1,
335                                        &slice_param->chroma_weight_l1_flag, slice_param->chroma_weight_l1, slice_param->chroma_offset_l1);
336     return 0;
337 }
338
339 AVHWAccel ff_h264_vaapi_hwaccel = {
340     .name           = "h264_vaapi",
341     .type           = AVMEDIA_TYPE_VIDEO,
342     .id             = CODEC_ID_H264,
343     .pix_fmt        = PIX_FMT_VAAPI_VLD,
344     .start_frame    = start_frame,
345     .end_frame      = end_frame,
346     .decode_slice   = decode_slice,
347 };