88d47823b86d9753d39731c543ab3caab1364f06
[libav.git] / libavcodec / dxva2_hevc.c
1 /*
2 * DXVA2 HEVC HW acceleration.
3 *
4 * copyright (c) 2014 - 2015 Hendrik Leppkes
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 "libavutil/avassert.h"
24
25 #include "dxva2_internal.h"
26 #include "hevc.h"
27
28 #define MAX_SLICES 256
29
30 struct hevc_dxva2_picture_context {
31 DXVA_PicParams_HEVC pp;
32 DXVA_Qmatrix_HEVC qm;
33 unsigned slice_count;
34 DXVA_Slice_HEVC_Short slice_short[MAX_SLICES];
35 const uint8_t *bitstream;
36 unsigned bitstream_size;
37 };
38
39 static void fill_picture_entry(DXVA_PicEntry_HEVC *pic,
40 unsigned index, unsigned flag)
41 {
42 av_assert0((index & 0x7f) == index && (flag & 0x01) == flag);
43 pic->bPicEntry = index | (flag << 7);
44 }
45
46 static int get_refpic_index(const DXVA_PicParams_HEVC *pp, int surface_index)
47 {
48 int i;
49 for (i = 0; i < FF_ARRAY_ELEMS(pp->RefPicList); i++) {
50 if ((pp->RefPicList[i].bPicEntry & 0x7f) == surface_index)
51 return i;
52 }
53 return 0xff;
54 }
55
56 static void fill_picture_parameters(const AVCodecContext *avctx, AVDXVAContext *ctx, const HEVCContext *h,
57 DXVA_PicParams_HEVC *pp)
58 {
59 const HEVCFrame *current_picture = h->ref;
60 const HEVCSPS *sps = h->ps.sps;
61 const HEVCPPS *pps = h->ps.pps;
62 int i, j;
63
64 memset(pp, 0, sizeof(*pp));
65
66 pp->PicWidthInMinCbsY = sps->min_cb_width;
67 pp->PicHeightInMinCbsY = sps->min_cb_height;
68
69 pp->wFormatAndSequenceInfoFlags = (sps->chroma_format_idc << 0) |
70 (sps->separate_colour_plane_flag << 2) |
71 ((sps->bit_depth - 8) << 3) |
72 ((sps->bit_depth - 8) << 6) |
73 ((sps->log2_max_poc_lsb - 4) << 9) |
74 (0 << 13) |
75 (0 << 14) |
76 (0 << 15);
77
78 fill_picture_entry(&pp->CurrPic, ff_dxva2_get_surface_index(avctx, ctx, current_picture->frame), 0);
79
80 pp->sps_max_dec_pic_buffering_minus1 = sps->temporal_layer[sps->max_sub_layers - 1].max_dec_pic_buffering - 1;
81 pp->log2_min_luma_coding_block_size_minus3 = sps->log2_min_cb_size - 3;
82 pp->log2_diff_max_min_luma_coding_block_size = sps->log2_diff_max_min_coding_block_size;
83 pp->log2_min_transform_block_size_minus2 = sps->log2_min_tb_size - 2;
84 pp->log2_diff_max_min_transform_block_size = sps->log2_max_trafo_size - sps->log2_min_tb_size;
85 pp->max_transform_hierarchy_depth_inter = sps->max_transform_hierarchy_depth_inter;
86 pp->max_transform_hierarchy_depth_intra = sps->max_transform_hierarchy_depth_intra;
87 pp->num_short_term_ref_pic_sets = sps->nb_st_rps;
88 pp->num_long_term_ref_pics_sps = sps->num_long_term_ref_pics_sps;
89
90 pp->num_ref_idx_l0_default_active_minus1 = pps->num_ref_idx_l0_default_active - 1;
91 pp->num_ref_idx_l1_default_active_minus1 = pps->num_ref_idx_l1_default_active - 1;
92 pp->init_qp_minus26 = pps->pic_init_qp_minus26;
93
94 if (h->sh.short_term_ref_pic_set_sps_flag == 0 && h->sh.short_term_rps) {
95 pp->ucNumDeltaPocsOfRefRpsIdx = h->sh.short_term_rps->num_delta_pocs;
96 pp->wNumBitsForShortTermRPSInSlice = h->sh.short_term_ref_pic_set_size;
97 }
98
99 pp->dwCodingParamToolFlags = (sps->scaling_list_enable_flag << 0) |
100 (sps->amp_enabled_flag << 1) |
101 (sps->sao_enabled << 2) |
102 (sps->pcm_enabled_flag << 3) |
103 ((sps->pcm_enabled_flag ? (sps->pcm.bit_depth - 1) : 0) << 4) |
104 ((sps->pcm_enabled_flag ? (sps->pcm.bit_depth_chroma - 1) : 0) << 8) |
105 ((sps->pcm_enabled_flag ? (sps->pcm.log2_min_pcm_cb_size - 3) : 0) << 12) |
106 ((sps->pcm_enabled_flag ? (sps->pcm.log2_max_pcm_cb_size - sps->pcm.log2_min_pcm_cb_size) : 0) << 14) |
107 (sps->pcm.loop_filter_disable_flag << 16) |
108 (sps->long_term_ref_pics_present_flag << 17) |
109 (sps->sps_temporal_mvp_enabled_flag << 18) |
110 (sps->sps_strong_intra_smoothing_enable_flag << 19) |
111 (pps->dependent_slice_segments_enabled_flag << 20) |
112 (pps->output_flag_present_flag << 21) |
113 (pps->num_extra_slice_header_bits << 22) |
114 (pps->sign_data_hiding_flag << 25) |
115 (pps->cabac_init_present_flag << 26) |
116 (0 << 27);
117
118 pp->dwCodingSettingPicturePropertyFlags = (pps->constrained_intra_pred_flag << 0) |
119 (pps->transform_skip_enabled_flag << 1) |
120 (pps->cu_qp_delta_enabled_flag << 2) |
121 (pps->pic_slice_level_chroma_qp_offsets_present_flag << 3) |
122 (pps->weighted_pred_flag << 4) |
123 (pps->weighted_bipred_flag << 5) |
124 (pps->transquant_bypass_enable_flag << 6) |
125 (pps->tiles_enabled_flag << 7) |
126 (pps->entropy_coding_sync_enabled_flag << 8) |
127 (pps->uniform_spacing_flag << 9) |
128 ((pps->tiles_enabled_flag ? pps->loop_filter_across_tiles_enabled_flag : 0) << 10) |
129 (pps->seq_loop_filter_across_slices_enabled_flag << 11) |
130 (pps->deblocking_filter_override_enabled_flag << 12) |
131 (pps->disable_dbf << 13) |
132 (pps->lists_modification_present_flag << 14) |
133 (pps->slice_header_extension_present_flag << 15) |
134 (IS_IRAP(h) << 16) |
135 (IS_IDR(h) << 17) |
136 /* IntraPicFlag */
137 (IS_IRAP(h) << 18) |
138 (0 << 19);
139 pp->pps_cb_qp_offset = pps->cb_qp_offset;
140 pp->pps_cr_qp_offset = pps->cr_qp_offset;
141 if (pps->tiles_enabled_flag) {
142 pp->num_tile_columns_minus1 = pps->num_tile_columns - 1;
143 pp->num_tile_rows_minus1 = pps->num_tile_rows - 1;
144
145 if (!pps->uniform_spacing_flag) {
146 for (i = 0; i < pps->num_tile_columns; i++)
147 pp->column_width_minus1[i] = pps->column_width[i] - 1;
148
149 for (i = 0; i < pps->num_tile_rows; i++)
150 pp->row_height_minus1[i] = pps->row_height[i] - 1;
151 }
152 }
153
154 pp->diff_cu_qp_delta_depth = pps->diff_cu_qp_delta_depth;
155 pp->pps_beta_offset_div2 = pps->beta_offset / 2;
156 pp->pps_tc_offset_div2 = pps->tc_offset / 2;
157 pp->log2_parallel_merge_level_minus2 = pps->log2_parallel_merge_level - 2;
158 pp->CurrPicOrderCntVal = h->poc;
159
160 // fill RefPicList from the DPB
161 for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->RefPicList); i++) {
162 const HEVCFrame *frame = NULL;
163 while (!frame && j < FF_ARRAY_ELEMS(h->DPB)) {
164 if (&h->DPB[j] != current_picture && (h->DPB[j].flags & (HEVC_FRAME_FLAG_LONG_REF | HEVC_FRAME_FLAG_SHORT_REF)))
165 frame = &h->DPB[j];
166 j++;
167 }
168
169 if (frame) {
170 fill_picture_entry(&pp->RefPicList[i], ff_dxva2_get_surface_index(avctx, ctx, frame->frame), !!(frame->flags & HEVC_FRAME_FLAG_LONG_REF));
171 pp->PicOrderCntValList[i] = frame->poc;
172 } else {
173 pp->RefPicList[i].bPicEntry = 0xff;
174 pp->PicOrderCntValList[i] = 0;
175 }
176 }
177
178 #define DO_REF_LIST(ref_idx, ref_list) { \
179 const RefPicList *rpl = &h->rps[ref_idx]; \
180 for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->ref_list); i++) { \
181 const HEVCFrame *frame = NULL; \
182 while (!frame && j < rpl->nb_refs) \
183 frame = rpl->ref[j++]; \
184 if (frame) \
185 pp->ref_list[i] = get_refpic_index(pp, ff_dxva2_get_surface_index(avctx, ctx, frame->frame)); \
186 else \
187 pp->ref_list[i] = 0xff; \
188 } \
189 }
190
191 // Fill short term and long term lists
192 DO_REF_LIST(ST_CURR_BEF, RefPicSetStCurrBefore);
193 DO_REF_LIST(ST_CURR_AFT, RefPicSetStCurrAfter);
194 DO_REF_LIST(LT_CURR, RefPicSetLtCurr);
195
196 pp->StatusReportFeedbackNumber = 1 + DXVA_CONTEXT_REPORT_ID(avctx, ctx)++;
197 }
198
199 static void fill_scaling_lists(AVDXVAContext *ctx, const HEVCContext *h, DXVA_Qmatrix_HEVC *qm)
200 {
201 unsigned i, j, pos;
202 const ScalingList *sl = h->ps.pps->scaling_list_data_present_flag ?
203 &h->ps.pps->scaling_list : &h->ps.sps->scaling_list;
204
205 memset(qm, 0, sizeof(*qm));
206 for (i = 0; i < 6; i++) {
207 for (j = 0; j < 16; j++) {
208 pos = 4 * ff_hevc_diag_scan4x4_y[j] + ff_hevc_diag_scan4x4_x[j];
209 qm->ucScalingLists0[i][j] = sl->sl[0][i][pos];
210 }
211
212 for (j = 0; j < 64; j++) {
213 pos = 8 * ff_hevc_diag_scan8x8_y[j] + ff_hevc_diag_scan8x8_x[j];
214 qm->ucScalingLists1[i][j] = sl->sl[1][i][pos];
215 qm->ucScalingLists2[i][j] = sl->sl[2][i][pos];
216
217 if (i < 2)
218 qm->ucScalingLists3[i][j] = sl->sl[3][i * 3][pos];
219 }
220
221 qm->ucScalingListDCCoefSizeID2[i] = sl->sl_dc[0][i];
222 if (i < 2)
223 qm->ucScalingListDCCoefSizeID3[i] = sl->sl_dc[1][i * 3];
224 }
225 }
226
227 static void fill_slice_short(DXVA_Slice_HEVC_Short *slice,
228 unsigned position, unsigned size)
229 {
230 memset(slice, 0, sizeof(*slice));
231 slice->BSNALunitDataLocation = position;
232 slice->SliceBytesInBuffer = size;
233 slice->wBadSliceChopping = 0;
234 }
235
236 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
237 DECODER_BUFFER_DESC *bs,
238 DECODER_BUFFER_DESC *sc)
239 {
240 const HEVCContext *h = avctx->priv_data;
241 AVDXVAContext *ctx = avctx->hwaccel_context;
242 const HEVCFrame *current_picture = h->ref;
243 struct hevc_dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
244 DXVA_Slice_HEVC_Short *slice = NULL;
245 void *dxva_data_ptr;
246 uint8_t *dxva_data, *current, *end;
247 unsigned dxva_size;
248 void *slice_data;
249 unsigned slice_size;
250 unsigned padding;
251 unsigned i;
252 unsigned type;
253
254 /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
255 #if CONFIG_D3D11VA
256 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
257 type = D3D11_VIDEO_DECODER_BUFFER_BITSTREAM;
258 if (FAILED(ID3D11VideoContext_GetDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context,
259 D3D11VA_CONTEXT(ctx)->decoder,
260 type,
261 &dxva_size, &dxva_data_ptr)))
262 return -1;
263 }
264 #endif
265 #if CONFIG_DXVA2
266 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
267 type = DXVA2_BitStreamDateBufferType;
268 if (FAILED(IDirectXVideoDecoder_GetBuffer(DXVA2_CONTEXT(ctx)->decoder,
269 type,
270 &dxva_data_ptr, &dxva_size)))
271 return -1;
272 }
273 #endif
274
275 dxva_data = dxva_data_ptr;
276 current = dxva_data;
277 end = dxva_data + dxva_size;
278
279 for (i = 0; i < ctx_pic->slice_count; i++) {
280 static const uint8_t start_code[] = { 0, 0, 1 };
281 static const unsigned start_code_size = sizeof(start_code);
282 unsigned position, size;
283
284 slice = &ctx_pic->slice_short[i];
285
286 position = slice->BSNALunitDataLocation;
287 size = slice->SliceBytesInBuffer;
288 if (start_code_size + size > end - current) {
289 av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
290 break;
291 }
292
293 slice->BSNALunitDataLocation = current - dxva_data;
294 slice->SliceBytesInBuffer = start_code_size + size;
295
296 memcpy(current, start_code, start_code_size);
297 current += start_code_size;
298
299 memcpy(current, &ctx_pic->bitstream[position], size);
300 current += size;
301 }
302 padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
303 if (slice && padding > 0) {
304 memset(current, 0, padding);
305 current += padding;
306
307 slice->SliceBytesInBuffer += padding;
308 }
309 #if CONFIG_D3D11VA
310 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD)
311 if (FAILED(ID3D11VideoContext_ReleaseDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder, type)))
312 return -1;
313 #endif
314 #if CONFIG_DXVA2
315 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
316 if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(DXVA2_CONTEXT(ctx)->decoder, type)))
317 return -1;
318 #endif
319 if (i < ctx_pic->slice_count)
320 return -1;
321
322 #if CONFIG_D3D11VA
323 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
324 D3D11_VIDEO_DECODER_BUFFER_DESC *dsc11 = bs;
325 memset(dsc11, 0, sizeof(*dsc11));
326 dsc11->BufferType = type;
327 dsc11->DataSize = current - dxva_data;
328 dsc11->NumMBsInBuffer = 0;
329
330 type = D3D11_VIDEO_DECODER_BUFFER_SLICE_CONTROL;
331 }
332 #endif
333 #if CONFIG_DXVA2
334 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
335 DXVA2_DecodeBufferDesc *dsc2 = bs;
336 memset(dsc2, 0, sizeof(*dsc2));
337 dsc2->CompressedBufferType = type;
338 dsc2->DataSize = current - dxva_data;
339 dsc2->NumMBsInBuffer = 0;
340
341 type = DXVA2_SliceControlBufferType;
342 }
343 #endif
344
345 slice_data = ctx_pic->slice_short;
346 slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
347
348 av_assert0(((current - dxva_data) & 127) == 0);
349 return ff_dxva2_commit_buffer(avctx, ctx, sc,
350 type,
351 slice_data, slice_size, 0);
352 }
353
354
355 static int dxva2_hevc_start_frame(AVCodecContext *avctx,
356 av_unused const uint8_t *buffer,
357 av_unused uint32_t size)
358 {
359 const HEVCContext *h = avctx->priv_data;
360 AVDXVAContext *ctx = avctx->hwaccel_context;
361 struct hevc_dxva2_picture_context *ctx_pic = h->ref->hwaccel_picture_private;
362
363 if (DXVA_CONTEXT_DECODER(avctx, ctx) == NULL ||
364 DXVA_CONTEXT_CFG(avctx, ctx) == NULL ||
365 DXVA_CONTEXT_COUNT(avctx, ctx) <= 0)
366 return -1;
367 av_assert0(ctx_pic);
368
369 /* Fill up DXVA_PicParams_HEVC */
370 fill_picture_parameters(avctx, ctx, h, &ctx_pic->pp);
371
372 /* Fill up DXVA_Qmatrix_HEVC */
373 fill_scaling_lists(ctx, h, &ctx_pic->qm);
374
375 ctx_pic->slice_count = 0;
376 ctx_pic->bitstream_size = 0;
377 ctx_pic->bitstream = NULL;
378 return 0;
379 }
380
381 static int dxva2_hevc_decode_slice(AVCodecContext *avctx,
382 const uint8_t *buffer,
383 uint32_t size)
384 {
385 const HEVCContext *h = avctx->priv_data;
386 const HEVCFrame *current_picture = h->ref;
387 struct hevc_dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
388 unsigned position;
389
390 if (ctx_pic->slice_count >= MAX_SLICES)
391 return -1;
392
393 if (!ctx_pic->bitstream)
394 ctx_pic->bitstream = buffer;
395 ctx_pic->bitstream_size += size;
396
397 position = buffer - ctx_pic->bitstream;
398 fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count], position, size);
399 ctx_pic->slice_count++;
400
401 return 0;
402 }
403
404 static int dxva2_hevc_end_frame(AVCodecContext *avctx)
405 {
406 HEVCContext *h = avctx->priv_data;
407 struct hevc_dxva2_picture_context *ctx_pic = h->ref->hwaccel_picture_private;
408 int scale = ctx_pic->pp.dwCodingParamToolFlags & 1;
409 int ret;
410
411 if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
412 return -1;
413
414 ret = ff_dxva2_common_end_frame(avctx, h->ref->frame,
415 &ctx_pic->pp, sizeof(ctx_pic->pp),
416 scale ? &ctx_pic->qm : NULL, scale ? sizeof(ctx_pic->qm) : 0,
417 commit_bitstream_and_slice_buffer);
418 return ret;
419 }
420
421 #if CONFIG_HEVC_DXVA2_HWACCEL
422 AVHWAccel ff_hevc_dxva2_hwaccel = {
423 .name = "hevc_dxva2",
424 .type = AVMEDIA_TYPE_VIDEO,
425 .id = AV_CODEC_ID_HEVC,
426 .pix_fmt = AV_PIX_FMT_DXVA2_VLD,
427 .start_frame = dxva2_hevc_start_frame,
428 .decode_slice = dxva2_hevc_decode_slice,
429 .end_frame = dxva2_hevc_end_frame,
430 .frame_priv_data_size = sizeof(struct hevc_dxva2_picture_context),
431 };
432 #endif
433
434 #if CONFIG_HEVC_D3D11VA_HWACCEL
435 AVHWAccel ff_hevc_d3d11va_hwaccel = {
436 .name = "hevc_d3d11va",
437 .type = AVMEDIA_TYPE_VIDEO,
438 .id = AV_CODEC_ID_HEVC,
439 .pix_fmt = AV_PIX_FMT_D3D11VA_VLD,
440 .start_frame = dxva2_hevc_start_frame,
441 .decode_slice = dxva2_hevc_decode_slice,
442 .end_frame = dxva2_hevc_end_frame,
443 .frame_priv_data_size = sizeof(struct hevc_dxva2_picture_context),
444 };
445 #endif