2 * DXVA2 HEVC HW acceleration.
4 * copyright (c) 2014 - 2015 Hendrik Leppkes
6 * This file is part of FFmpeg.
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.
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.
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
23 #include "libavutil/avassert.h"
27 // The headers above may include w32threads.h, which uses the original
28 // _WIN32_WINNT define, while dxva2_internal.h redefines it to target a
29 // potentially newer version.
30 #include "dxva2_internal.h"
32 #define MAX_SLICES 256
34 struct hevc_dxva2_picture_context {
35 DXVA_PicParams_HEVC pp;
38 DXVA_Slice_HEVC_Short slice_short[MAX_SLICES];
39 const uint8_t *bitstream;
40 unsigned bitstream_size;
43 static void fill_picture_entry(DXVA_PicEntry_HEVC *pic,
44 unsigned index, unsigned flag)
46 av_assert0((index & 0x7f) == index && (flag & 0x01) == flag);
47 pic->bPicEntry = index | (flag << 7);
50 static int get_refpic_index(const DXVA_PicParams_HEVC *pp, int surface_index)
53 for (i = 0; i < FF_ARRAY_ELEMS(pp->RefPicList); i++) {
54 if ((pp->RefPicList[i].bPicEntry & 0x7f) == surface_index)
60 static void fill_picture_parameters(const AVCodecContext *avctx, AVDXVAContext *ctx, const HEVCContext *h,
61 DXVA_PicParams_HEVC *pp)
63 const HEVCFrame *current_picture = h->ref;
64 const HEVCSPS *sps = h->ps.sps;
65 const HEVCPPS *pps = h->ps.pps;
68 memset(pp, 0, sizeof(*pp));
70 pp->PicWidthInMinCbsY = sps->min_cb_width;
71 pp->PicHeightInMinCbsY = sps->min_cb_height;
73 pp->wFormatAndSequenceInfoFlags = (sps->chroma_format_idc << 0) |
74 (sps->separate_colour_plane_flag << 2) |
75 ((sps->bit_depth - 8) << 3) |
76 ((sps->bit_depth - 8) << 6) |
77 ((sps->log2_max_poc_lsb - 4) << 9) |
82 fill_picture_entry(&pp->CurrPic, ff_dxva2_get_surface_index(avctx, ctx, current_picture->frame), 0);
84 pp->sps_max_dec_pic_buffering_minus1 = sps->temporal_layer[sps->max_sub_layers - 1].max_dec_pic_buffering - 1;
85 pp->log2_min_luma_coding_block_size_minus3 = sps->log2_min_cb_size - 3;
86 pp->log2_diff_max_min_luma_coding_block_size = sps->log2_diff_max_min_coding_block_size;
87 pp->log2_min_transform_block_size_minus2 = sps->log2_min_tb_size - 2;
88 pp->log2_diff_max_min_transform_block_size = sps->log2_max_trafo_size - sps->log2_min_tb_size;
89 pp->max_transform_hierarchy_depth_inter = sps->max_transform_hierarchy_depth_inter;
90 pp->max_transform_hierarchy_depth_intra = sps->max_transform_hierarchy_depth_intra;
91 pp->num_short_term_ref_pic_sets = sps->nb_st_rps;
92 pp->num_long_term_ref_pics_sps = sps->num_long_term_ref_pics_sps;
94 pp->num_ref_idx_l0_default_active_minus1 = pps->num_ref_idx_l0_default_active - 1;
95 pp->num_ref_idx_l1_default_active_minus1 = pps->num_ref_idx_l1_default_active - 1;
96 pp->init_qp_minus26 = pps->pic_init_qp_minus26;
98 if (h->sh.short_term_ref_pic_set_sps_flag == 0 && h->sh.short_term_rps) {
99 pp->ucNumDeltaPocsOfRefRpsIdx = h->sh.short_term_rps->rps_idx_num_delta_pocs;
100 pp->wNumBitsForShortTermRPSInSlice = h->sh.short_term_ref_pic_set_size;
103 pp->dwCodingParamToolFlags = (sps->scaling_list_enable_flag << 0) |
104 (sps->amp_enabled_flag << 1) |
105 (sps->sao_enabled << 2) |
106 (sps->pcm_enabled_flag << 3) |
107 ((sps->pcm_enabled_flag ? (sps->pcm.bit_depth - 1) : 0) << 4) |
108 ((sps->pcm_enabled_flag ? (sps->pcm.bit_depth_chroma - 1) : 0) << 8) |
109 ((sps->pcm_enabled_flag ? (sps->pcm.log2_min_pcm_cb_size - 3) : 0) << 12) |
110 ((sps->pcm_enabled_flag ? (sps->pcm.log2_max_pcm_cb_size - sps->pcm.log2_min_pcm_cb_size) : 0) << 14) |
111 (sps->pcm.loop_filter_disable_flag << 16) |
112 (sps->long_term_ref_pics_present_flag << 17) |
113 (sps->sps_temporal_mvp_enabled_flag << 18) |
114 (sps->sps_strong_intra_smoothing_enable_flag << 19) |
115 (pps->dependent_slice_segments_enabled_flag << 20) |
116 (pps->output_flag_present_flag << 21) |
117 (pps->num_extra_slice_header_bits << 22) |
118 (pps->sign_data_hiding_flag << 25) |
119 (pps->cabac_init_present_flag << 26) |
122 pp->dwCodingSettingPicturePropertyFlags = (pps->constrained_intra_pred_flag << 0) |
123 (pps->transform_skip_enabled_flag << 1) |
124 (pps->cu_qp_delta_enabled_flag << 2) |
125 (pps->pic_slice_level_chroma_qp_offsets_present_flag << 3) |
126 (pps->weighted_pred_flag << 4) |
127 (pps->weighted_bipred_flag << 5) |
128 (pps->transquant_bypass_enable_flag << 6) |
129 (pps->tiles_enabled_flag << 7) |
130 (pps->entropy_coding_sync_enabled_flag << 8) |
131 (pps->uniform_spacing_flag << 9) |
132 ((pps->tiles_enabled_flag ? pps->loop_filter_across_tiles_enabled_flag : 0) << 10) |
133 (pps->seq_loop_filter_across_slices_enabled_flag << 11) |
134 (pps->deblocking_filter_override_enabled_flag << 12) |
135 (pps->disable_dbf << 13) |
136 (pps->lists_modification_present_flag << 14) |
137 (pps->slice_header_extension_present_flag << 15) |
143 pp->pps_cb_qp_offset = pps->cb_qp_offset;
144 pp->pps_cr_qp_offset = pps->cr_qp_offset;
145 if (pps->tiles_enabled_flag) {
146 pp->num_tile_columns_minus1 = pps->num_tile_columns - 1;
147 pp->num_tile_rows_minus1 = pps->num_tile_rows - 1;
149 if (!pps->uniform_spacing_flag) {
150 for (i = 0; i < pps->num_tile_columns; i++)
151 pp->column_width_minus1[i] = pps->column_width[i] - 1;
153 for (i = 0; i < pps->num_tile_rows; i++)
154 pp->row_height_minus1[i] = pps->row_height[i] - 1;
158 pp->diff_cu_qp_delta_depth = pps->diff_cu_qp_delta_depth;
159 pp->pps_beta_offset_div2 = pps->beta_offset / 2;
160 pp->pps_tc_offset_div2 = pps->tc_offset / 2;
161 pp->log2_parallel_merge_level_minus2 = pps->log2_parallel_merge_level - 2;
162 pp->CurrPicOrderCntVal = h->poc;
164 // fill RefPicList from the DPB
165 for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->RefPicList); i++) {
166 const HEVCFrame *frame = NULL;
167 while (!frame && j < FF_ARRAY_ELEMS(h->DPB)) {
168 if (&h->DPB[j] != current_picture && (h->DPB[j].flags & (HEVC_FRAME_FLAG_LONG_REF | HEVC_FRAME_FLAG_SHORT_REF)))
174 fill_picture_entry(&pp->RefPicList[i], ff_dxva2_get_surface_index(avctx, ctx, frame->frame), !!(frame->flags & HEVC_FRAME_FLAG_LONG_REF));
175 pp->PicOrderCntValList[i] = frame->poc;
177 pp->RefPicList[i].bPicEntry = 0xff;
178 pp->PicOrderCntValList[i] = 0;
182 #define DO_REF_LIST(ref_idx, ref_list) { \
183 const RefPicList *rpl = &h->rps[ref_idx]; \
184 for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->ref_list); i++) { \
185 const HEVCFrame *frame = NULL; \
186 while (!frame && j < rpl->nb_refs) \
187 frame = rpl->ref[j++]; \
189 pp->ref_list[i] = get_refpic_index(pp, ff_dxva2_get_surface_index(avctx, ctx, frame->frame)); \
191 pp->ref_list[i] = 0xff; \
195 // Fill short term and long term lists
196 DO_REF_LIST(ST_CURR_BEF, RefPicSetStCurrBefore);
197 DO_REF_LIST(ST_CURR_AFT, RefPicSetStCurrAfter);
198 DO_REF_LIST(LT_CURR, RefPicSetLtCurr);
200 pp->StatusReportFeedbackNumber = 1 + DXVA_CONTEXT_REPORT_ID(avctx, ctx)++;
203 static void fill_scaling_lists(AVDXVAContext *ctx, const HEVCContext *h, DXVA_Qmatrix_HEVC *qm)
206 const ScalingList *sl = h->ps.pps->scaling_list_data_present_flag ?
207 &h->ps.pps->scaling_list : &h->ps.sps->scaling_list;
209 memset(qm, 0, sizeof(*qm));
210 for (i = 0; i < 6; i++) {
211 for (j = 0; j < 16; j++) {
212 pos = 4 * ff_hevc_diag_scan4x4_y[j] + ff_hevc_diag_scan4x4_x[j];
213 qm->ucScalingLists0[i][j] = sl->sl[0][i][pos];
216 for (j = 0; j < 64; j++) {
217 pos = 8 * ff_hevc_diag_scan8x8_y[j] + ff_hevc_diag_scan8x8_x[j];
218 qm->ucScalingLists1[i][j] = sl->sl[1][i][pos];
219 qm->ucScalingLists2[i][j] = sl->sl[2][i][pos];
222 qm->ucScalingLists3[i][j] = sl->sl[3][i * 3][pos];
225 qm->ucScalingListDCCoefSizeID2[i] = sl->sl_dc[0][i];
227 qm->ucScalingListDCCoefSizeID3[i] = sl->sl_dc[1][i * 3];
231 static void fill_slice_short(DXVA_Slice_HEVC_Short *slice,
232 unsigned position, unsigned size)
234 memset(slice, 0, sizeof(*slice));
235 slice->BSNALunitDataLocation = position;
236 slice->SliceBytesInBuffer = size;
237 slice->wBadSliceChopping = 0;
240 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
241 DECODER_BUFFER_DESC *bs,
242 DECODER_BUFFER_DESC *sc)
244 const HEVCContext *h = avctx->priv_data;
245 AVDXVAContext *ctx = avctx->hwaccel_context;
246 const HEVCFrame *current_picture = h->ref;
247 struct hevc_dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
248 DXVA_Slice_HEVC_Short *slice = NULL;
250 uint8_t *dxva_data, *current, *end;
258 /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
260 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
261 type = D3D11_VIDEO_DECODER_BUFFER_BITSTREAM;
262 if (FAILED(ID3D11VideoContext_GetDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context,
263 D3D11VA_CONTEXT(ctx)->decoder,
265 &dxva_size, &dxva_data_ptr)))
270 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
271 type = DXVA2_BitStreamDateBufferType;
272 if (FAILED(IDirectXVideoDecoder_GetBuffer(DXVA2_CONTEXT(ctx)->decoder,
274 &dxva_data_ptr, &dxva_size)))
279 dxva_data = dxva_data_ptr;
281 end = dxva_data + dxva_size;
283 for (i = 0; i < ctx_pic->slice_count; i++) {
284 static const uint8_t start_code[] = { 0, 0, 1 };
285 static const unsigned start_code_size = sizeof(start_code);
286 unsigned position, size;
288 slice = &ctx_pic->slice_short[i];
290 position = slice->BSNALunitDataLocation;
291 size = slice->SliceBytesInBuffer;
292 if (start_code_size + size > end - current) {
293 av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
297 slice->BSNALunitDataLocation = current - dxva_data;
298 slice->SliceBytesInBuffer = start_code_size + size;
300 memcpy(current, start_code, start_code_size);
301 current += start_code_size;
303 memcpy(current, &ctx_pic->bitstream[position], size);
306 padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
307 if (slice && padding > 0) {
308 memset(current, 0, padding);
311 slice->SliceBytesInBuffer += padding;
314 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD)
315 if (FAILED(ID3D11VideoContext_ReleaseDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder, type)))
319 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
320 if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(DXVA2_CONTEXT(ctx)->decoder, type)))
323 if (i < ctx_pic->slice_count)
327 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
328 D3D11_VIDEO_DECODER_BUFFER_DESC *dsc11 = bs;
329 memset(dsc11, 0, sizeof(*dsc11));
330 dsc11->BufferType = type;
331 dsc11->DataSize = current - dxva_data;
332 dsc11->NumMBsInBuffer = 0;
334 type = D3D11_VIDEO_DECODER_BUFFER_SLICE_CONTROL;
338 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
339 DXVA2_DecodeBufferDesc *dsc2 = bs;
340 memset(dsc2, 0, sizeof(*dsc2));
341 dsc2->CompressedBufferType = type;
342 dsc2->DataSize = current - dxva_data;
343 dsc2->NumMBsInBuffer = 0;
345 type = DXVA2_SliceControlBufferType;
349 slice_data = ctx_pic->slice_short;
350 slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
352 av_assert0(((current - dxva_data) & 127) == 0);
353 return ff_dxva2_commit_buffer(avctx, ctx, sc,
355 slice_data, slice_size, 0);
359 static int dxva2_hevc_start_frame(AVCodecContext *avctx,
360 av_unused const uint8_t *buffer,
361 av_unused uint32_t size)
363 const HEVCContext *h = avctx->priv_data;
364 AVDXVAContext *ctx = avctx->hwaccel_context;
365 struct hevc_dxva2_picture_context *ctx_pic = h->ref->hwaccel_picture_private;
367 if (!DXVA_CONTEXT_VALID(avctx, ctx))
371 /* Fill up DXVA_PicParams_HEVC */
372 fill_picture_parameters(avctx, ctx, h, &ctx_pic->pp);
374 /* Fill up DXVA_Qmatrix_HEVC */
375 fill_scaling_lists(ctx, h, &ctx_pic->qm);
377 ctx_pic->slice_count = 0;
378 ctx_pic->bitstream_size = 0;
379 ctx_pic->bitstream = NULL;
383 static int dxva2_hevc_decode_slice(AVCodecContext *avctx,
384 const uint8_t *buffer,
387 const HEVCContext *h = avctx->priv_data;
388 const HEVCFrame *current_picture = h->ref;
389 struct hevc_dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
392 if (ctx_pic->slice_count >= MAX_SLICES)
395 if (!ctx_pic->bitstream)
396 ctx_pic->bitstream = buffer;
397 ctx_pic->bitstream_size += size;
399 position = buffer - ctx_pic->bitstream;
400 fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count], position, size);
401 ctx_pic->slice_count++;
406 static int dxva2_hevc_end_frame(AVCodecContext *avctx)
408 HEVCContext *h = avctx->priv_data;
409 struct hevc_dxva2_picture_context *ctx_pic = h->ref->hwaccel_picture_private;
410 int scale = ctx_pic->pp.dwCodingParamToolFlags & 1;
413 if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
416 ret = ff_dxva2_common_end_frame(avctx, h->ref->frame,
417 &ctx_pic->pp, sizeof(ctx_pic->pp),
418 scale ? &ctx_pic->qm : NULL, scale ? sizeof(ctx_pic->qm) : 0,
419 commit_bitstream_and_slice_buffer);
423 #if CONFIG_HEVC_DXVA2_HWACCEL
424 AVHWAccel ff_hevc_dxva2_hwaccel = {
425 .name = "hevc_dxva2",
426 .type = AVMEDIA_TYPE_VIDEO,
427 .id = AV_CODEC_ID_HEVC,
428 .pix_fmt = AV_PIX_FMT_DXVA2_VLD,
429 .start_frame = dxva2_hevc_start_frame,
430 .decode_slice = dxva2_hevc_decode_slice,
431 .end_frame = dxva2_hevc_end_frame,
432 .frame_priv_data_size = sizeof(struct hevc_dxva2_picture_context),
436 #if CONFIG_HEVC_D3D11VA_HWACCEL
437 AVHWAccel ff_hevc_d3d11va_hwaccel = {
438 .name = "hevc_d3d11va",
439 .type = AVMEDIA_TYPE_VIDEO,
440 .id = AV_CODEC_ID_HEVC,
441 .pix_fmt = AV_PIX_FMT_D3D11VA_VLD,
442 .start_frame = dxva2_hevc_start_frame,
443 .decode_slice = dxva2_hevc_decode_slice,
444 .end_frame = dxva2_hevc_end_frame,
445 .frame_priv_data_size = sizeof(struct hevc_dxva2_picture_context),