2 * DXVA2 H.264 HW acceleration.
4 * copyright (c) 2009 Laurent Aimar
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 #include "mpegutils.h"
29 // The headers above may include w32threads.h, which uses the original
30 // _WIN32_WINNT define, while dxva2_internal.h redefines it to target a
31 // potentially newer version.
32 #include "dxva2_internal.h"
34 struct dxva2_picture_context {
35 DXVA_PicParams_H264 pp;
38 DXVA_Slice_H264_Short slice_short[MAX_SLICES];
39 DXVA_Slice_H264_Long slice_long[MAX_SLICES];
40 const uint8_t *bitstream;
41 unsigned bitstream_size;
44 static void fill_picture_entry(DXVA_PicEntry_H264 *pic,
45 unsigned index, unsigned flag)
47 assert((index&0x7f) == index && (flag&0x01) == flag);
48 pic->bPicEntry = index | (flag << 7);
51 static void fill_picture_parameters(const AVCodecContext *avctx, AVDXVAContext *ctx, const H264Context *h,
52 DXVA_PicParams_H264 *pp)
54 const H264Picture *current_picture = h->cur_pic_ptr;
55 const SPS *sps = h->ps.sps;
56 const PPS *pps = h->ps.pps;
59 memset(pp, 0, sizeof(*pp));
60 /* Configure current picture */
61 fill_picture_entry(&pp->CurrPic,
62 ff_dxva2_get_surface_index(avctx, ctx, current_picture->f),
63 h->picture_structure == PICT_BOTTOM_FIELD);
64 /* Configure the set of references */
65 pp->UsedForReferenceFlags = 0;
66 pp->NonExistingFrameFlags = 0;
67 for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
69 if (j < h->short_ref_count) {
70 r = h->short_ref[j++];
73 while (!r && j < h->short_ref_count + 16)
74 r = h->long_ref[j++ - h->short_ref_count];
77 fill_picture_entry(&pp->RefFrameList[i],
78 ff_dxva2_get_surface_index(avctx, ctx, r->f),
81 if ((r->reference & PICT_TOP_FIELD) && r->field_poc[0] != INT_MAX)
82 pp->FieldOrderCntList[i][0] = r->field_poc[0];
83 if ((r->reference & PICT_BOTTOM_FIELD) && r->field_poc[1] != INT_MAX)
84 pp->FieldOrderCntList[i][1] = r->field_poc[1];
86 pp->FrameNumList[i] = r->long_ref ? r->pic_id : r->frame_num;
87 if (r->reference & PICT_TOP_FIELD)
88 pp->UsedForReferenceFlags |= 1 << (2*i + 0);
89 if (r->reference & PICT_BOTTOM_FIELD)
90 pp->UsedForReferenceFlags |= 1 << (2*i + 1);
92 pp->RefFrameList[i].bPicEntry = 0xff;
93 pp->FieldOrderCntList[i][0] = 0;
94 pp->FieldOrderCntList[i][1] = 0;
95 pp->FrameNumList[i] = 0;
99 pp->wFrameWidthInMbsMinus1 = h->mb_width - 1;
100 pp->wFrameHeightInMbsMinus1 = h->mb_height - 1;
101 pp->num_ref_frames = sps->ref_frame_count;
103 pp->wBitFields = ((h->picture_structure != PICT_FRAME) << 0) |
105 (h->picture_structure == PICT_FRAME)) << 1) |
106 (sps->residual_color_transform_flag << 2) |
107 /* sp_for_switch_flag (not implemented by FFmpeg) */
109 (sps->chroma_format_idc << 4) |
110 ((h->nal_ref_idc != 0) << 6) |
111 (pps->constrained_intra_pred << 7) |
112 (pps->weighted_pred << 8) |
113 (pps->weighted_bipred_idc << 9) |
114 /* MbsConsecutiveFlag */
116 (sps->frame_mbs_only_flag << 12) |
117 (pps->transform_8x8_mode << 13) |
118 ((sps->level_idc >= 31) << 14) |
119 /* IntraPicFlag (Modified if we detect a non
120 * intra slice in dxva2_h264_decode_slice) */
123 pp->bit_depth_luma_minus8 = sps->bit_depth_luma - 8;
124 pp->bit_depth_chroma_minus8 = sps->bit_depth_chroma - 8;
125 if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG)
126 pp->Reserved16Bits = 0;
127 else if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
128 pp->Reserved16Bits = 0x34c;
130 pp->Reserved16Bits = 3; /* FIXME is there a way to detect the right mode ? */
131 pp->StatusReportFeedbackNumber = 1 + DXVA_CONTEXT_REPORT_ID(avctx, ctx)++;
132 pp->CurrFieldOrderCnt[0] = 0;
133 if ((h->picture_structure & PICT_TOP_FIELD) &&
134 current_picture->field_poc[0] != INT_MAX)
135 pp->CurrFieldOrderCnt[0] = current_picture->field_poc[0];
136 pp->CurrFieldOrderCnt[1] = 0;
137 if ((h->picture_structure & PICT_BOTTOM_FIELD) &&
138 current_picture->field_poc[1] != INT_MAX)
139 pp->CurrFieldOrderCnt[1] = current_picture->field_poc[1];
140 pp->pic_init_qs_minus26 = pps->init_qs - 26;
141 pp->chroma_qp_index_offset = pps->chroma_qp_index_offset[0];
142 pp->second_chroma_qp_index_offset = pps->chroma_qp_index_offset[1];
143 pp->ContinuationFlag = 1;
144 pp->pic_init_qp_minus26 = pps->init_qp - 26;
145 pp->num_ref_idx_l0_active_minus1 = pps->ref_count[0] - 1;
146 pp->num_ref_idx_l1_active_minus1 = pps->ref_count[1] - 1;
147 pp->Reserved8BitsA = 0;
148 pp->frame_num = h->poc.frame_num;
149 pp->log2_max_frame_num_minus4 = sps->log2_max_frame_num - 4;
150 pp->pic_order_cnt_type = sps->poc_type;
151 if (sps->poc_type == 0)
152 pp->log2_max_pic_order_cnt_lsb_minus4 = sps->log2_max_poc_lsb - 4;
153 else if (sps->poc_type == 1)
154 pp->delta_pic_order_always_zero_flag = sps->delta_pic_order_always_zero_flag;
155 pp->direct_8x8_inference_flag = sps->direct_8x8_inference_flag;
156 pp->entropy_coding_mode_flag = pps->cabac;
157 pp->pic_order_present_flag = pps->pic_order_present;
158 pp->num_slice_groups_minus1 = pps->slice_group_count - 1;
159 pp->slice_group_map_type = pps->mb_slice_group_map_type;
160 pp->deblocking_filter_control_present_flag = pps->deblocking_filter_parameters_present;
161 pp->redundant_pic_cnt_present_flag= pps->redundant_pic_cnt_present;
162 pp->Reserved8BitsB = 0;
163 pp->slice_group_change_rate_minus1= 0; /* XXX not implemented by FFmpeg */
164 //pp->SliceGroupMap[810]; /* XXX not implemented by FFmpeg */
167 static void fill_scaling_lists(const AVCodecContext *avctx, AVDXVAContext *ctx, const H264Context *h, DXVA_Qmatrix_H264 *qm)
169 const PPS *pps = h->ps.pps;
171 memset(qm, 0, sizeof(*qm));
172 if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG) {
173 for (i = 0; i < 6; i++)
174 for (j = 0; j < 16; j++)
175 qm->bScalingLists4x4[i][j] = pps->scaling_matrix4[i][j];
177 for (i = 0; i < 64; i++) {
178 qm->bScalingLists8x8[0][i] = pps->scaling_matrix8[0][i];
179 qm->bScalingLists8x8[1][i] = pps->scaling_matrix8[3][i];
182 for (i = 0; i < 6; i++)
183 for (j = 0; j < 16; j++)
184 qm->bScalingLists4x4[i][j] = pps->scaling_matrix4[i][ff_zigzag_scan[j]];
186 for (i = 0; i < 64; i++) {
187 qm->bScalingLists8x8[0][i] = pps->scaling_matrix8[0][ff_zigzag_direct[i]];
188 qm->bScalingLists8x8[1][i] = pps->scaling_matrix8[3][ff_zigzag_direct[i]];
193 static int is_slice_short(const AVCodecContext *avctx, AVDXVAContext *ctx)
195 assert(DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 1 ||
196 DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 2);
197 return DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 2;
200 static void fill_slice_short(DXVA_Slice_H264_Short *slice,
201 unsigned position, unsigned size)
203 memset(slice, 0, sizeof(*slice));
204 slice->BSNALunitDataLocation = position;
205 slice->SliceBytesInBuffer = size;
206 slice->wBadSliceChopping = 0;
209 static int get_refpic_index(const DXVA_PicParams_H264 *pp, int surface_index)
212 for (i = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
213 if ((pp->RefFrameList[i].bPicEntry & 0x7f) == surface_index)
219 static void fill_slice_long(AVCodecContext *avctx, DXVA_Slice_H264_Long *slice,
220 const DXVA_PicParams_H264 *pp, unsigned position, unsigned size)
222 const H264Context *h = avctx->priv_data;
223 H264SliceContext *sl = &h->slice_ctx[0];
224 AVDXVAContext *ctx = avctx->hwaccel_context;
227 memset(slice, 0, sizeof(*slice));
228 slice->BSNALunitDataLocation = position;
229 slice->SliceBytesInBuffer = size;
230 slice->wBadSliceChopping = 0;
232 slice->first_mb_in_slice = (sl->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + sl->mb_x;
233 slice->NumMbsForSlice = 0; /* XXX it is set once we have all slices */
234 slice->BitOffsetToSliceData = get_bits_count(&sl->gb) - 8;
235 slice->slice_type = ff_h264_get_slice_type(sl);
236 if (sl->slice_type_fixed)
237 slice->slice_type += 5;
238 slice->luma_log2_weight_denom = sl->pwt.luma_log2_weight_denom;
239 slice->chroma_log2_weight_denom = sl->pwt.chroma_log2_weight_denom;
240 if (sl->list_count > 0)
241 slice->num_ref_idx_l0_active_minus1 = sl->ref_count[0] - 1;
242 if (sl->list_count > 1)
243 slice->num_ref_idx_l1_active_minus1 = sl->ref_count[1] - 1;
244 slice->slice_alpha_c0_offset_div2 = sl->slice_alpha_c0_offset / 2;
245 slice->slice_beta_offset_div2 = sl->slice_beta_offset / 2;
246 slice->Reserved8Bits = 0;
248 for (list = 0; list < 2; list++) {
250 for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
251 if (list < sl->list_count && i < sl->ref_count[list]) {
252 const H264Picture *r = sl->ref_list[list][i].parent;
255 if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
256 index = ff_dxva2_get_surface_index(avctx, ctx, r->f);
258 index = get_refpic_index(pp, ff_dxva2_get_surface_index(avctx, ctx, r->f));
259 fill_picture_entry(&slice->RefPicList[list][i], index,
260 sl->ref_list[list][i].reference == PICT_BOTTOM_FIELD);
261 for (plane = 0; plane < 3; plane++) {
263 if (plane == 0 && sl->pwt.luma_weight_flag[list]) {
264 w = sl->pwt.luma_weight[i][list][0];
265 o = sl->pwt.luma_weight[i][list][1];
266 } else if (plane >= 1 && sl->pwt.chroma_weight_flag[list]) {
267 w = sl->pwt.chroma_weight[i][list][plane-1][0];
268 o = sl->pwt.chroma_weight[i][list][plane-1][1];
270 w = 1 << (plane == 0 ? sl->pwt.luma_log2_weight_denom :
271 sl->pwt.chroma_log2_weight_denom);
274 slice->Weights[list][i][plane][0] = w;
275 slice->Weights[list][i][plane][1] = o;
279 slice->RefPicList[list][i].bPicEntry = 0xff;
280 for (plane = 0; plane < 3; plane++) {
281 slice->Weights[list][i][plane][0] = 0;
282 slice->Weights[list][i][plane][1] = 0;
287 slice->slice_qs_delta = 0; /* XXX not implemented by FFmpeg */
288 slice->slice_qp_delta = sl->qscale - h->ps.pps->init_qp;
289 slice->redundant_pic_cnt = sl->redundant_pic_count;
290 if (sl->slice_type == AV_PICTURE_TYPE_B)
291 slice->direct_spatial_mv_pred_flag = sl->direct_spatial_mv_pred;
292 slice->cabac_init_idc = h->ps.pps->cabac ? sl->cabac_init_idc : 0;
293 if (sl->deblocking_filter < 2)
294 slice->disable_deblocking_filter_idc = 1 - sl->deblocking_filter;
296 slice->disable_deblocking_filter_idc = sl->deblocking_filter;
297 slice->slice_id = h->current_slice - 1;
300 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
301 DECODER_BUFFER_DESC *bs,
302 DECODER_BUFFER_DESC *sc)
304 const H264Context *h = avctx->priv_data;
305 const unsigned mb_count = h->mb_width * h->mb_height;
306 AVDXVAContext *ctx = avctx->hwaccel_context;
307 const H264Picture *current_picture = h->cur_pic_ptr;
308 struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
309 DXVA_Slice_H264_Short *slice = NULL;
310 void *dxva_data_ptr = NULL;
311 uint8_t *dxva_data, *current, *end;
312 unsigned dxva_size = 0;
319 /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
321 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
322 type = D3D11_VIDEO_DECODER_BUFFER_BITSTREAM;
323 if (FAILED(ID3D11VideoContext_GetDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context,
324 D3D11VA_CONTEXT(ctx)->decoder,
326 &dxva_size, &dxva_data_ptr)))
331 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
332 type = DXVA2_BitStreamDateBufferType;
333 if (FAILED(IDirectXVideoDecoder_GetBuffer(DXVA2_CONTEXT(ctx)->decoder,
335 &dxva_data_ptr, &dxva_size)))
340 dxva_data = dxva_data_ptr;
342 end = dxva_data + dxva_size;
344 for (i = 0; i < ctx_pic->slice_count; i++) {
345 static const uint8_t start_code[] = { 0, 0, 1 };
346 static const unsigned start_code_size = sizeof(start_code);
347 unsigned position, size;
349 assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
350 offsetof(DXVA_Slice_H264_Long, BSNALunitDataLocation));
351 assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
352 offsetof(DXVA_Slice_H264_Long, SliceBytesInBuffer));
354 if (is_slice_short(avctx, ctx))
355 slice = &ctx_pic->slice_short[i];
357 slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
359 position = slice->BSNALunitDataLocation;
360 size = slice->SliceBytesInBuffer;
361 if (start_code_size + size > end - current) {
362 av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
366 slice->BSNALunitDataLocation = current - dxva_data;
367 slice->SliceBytesInBuffer = start_code_size + size;
369 if (!is_slice_short(avctx, ctx)) {
370 DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
371 if (i < ctx_pic->slice_count - 1)
372 slice_long->NumMbsForSlice =
373 slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
375 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
378 memcpy(current, start_code, start_code_size);
379 current += start_code_size;
381 memcpy(current, &ctx_pic->bitstream[position], size);
384 padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
385 if (slice && padding > 0) {
386 memset(current, 0, padding);
389 slice->SliceBytesInBuffer += padding;
392 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD)
393 if (FAILED(ID3D11VideoContext_ReleaseDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder, type)))
397 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
398 if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(DXVA2_CONTEXT(ctx)->decoder, type)))
401 if (i < ctx_pic->slice_count)
405 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
406 D3D11_VIDEO_DECODER_BUFFER_DESC *dsc11 = bs;
407 memset(dsc11, 0, sizeof(*dsc11));
408 dsc11->BufferType = type;
409 dsc11->DataSize = current - dxva_data;
410 dsc11->NumMBsInBuffer = mb_count;
412 type = D3D11_VIDEO_DECODER_BUFFER_SLICE_CONTROL;
414 av_assert0((dsc11->DataSize & 127) == 0);
418 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
419 DXVA2_DecodeBufferDesc *dsc2 = bs;
420 memset(dsc2, 0, sizeof(*dsc2));
421 dsc2->CompressedBufferType = type;
422 dsc2->DataSize = current - dxva_data;
423 dsc2->NumMBsInBuffer = mb_count;
425 type = DXVA2_SliceControlBufferType;
427 av_assert0((dsc2->DataSize & 127) == 0);
431 if (is_slice_short(avctx, ctx)) {
432 slice_data = ctx_pic->slice_short;
433 slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
435 slice_data = ctx_pic->slice_long;
436 slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
438 return ff_dxva2_commit_buffer(avctx, ctx, sc,
440 slice_data, slice_size, mb_count);
444 static int dxva2_h264_start_frame(AVCodecContext *avctx,
445 av_unused const uint8_t *buffer,
446 av_unused uint32_t size)
448 const H264Context *h = avctx->priv_data;
449 AVDXVAContext *ctx = avctx->hwaccel_context;
450 struct dxva2_picture_context *ctx_pic = h->cur_pic_ptr->hwaccel_picture_private;
452 if (DXVA_CONTEXT_DECODER(avctx, ctx) == NULL ||
453 DXVA_CONTEXT_CFG(avctx, ctx) == NULL ||
454 DXVA_CONTEXT_COUNT(avctx, ctx) <= 0)
458 /* Fill up DXVA_PicParams_H264 */
459 fill_picture_parameters(avctx, ctx, h, &ctx_pic->pp);
461 /* Fill up DXVA_Qmatrix_H264 */
462 fill_scaling_lists(avctx, ctx, h, &ctx_pic->qm);
464 ctx_pic->slice_count = 0;
465 ctx_pic->bitstream_size = 0;
466 ctx_pic->bitstream = NULL;
470 static int dxva2_h264_decode_slice(AVCodecContext *avctx,
471 const uint8_t *buffer,
474 const H264Context *h = avctx->priv_data;
475 const H264SliceContext *sl = &h->slice_ctx[0];
476 AVDXVAContext *ctx = avctx->hwaccel_context;
477 const H264Picture *current_picture = h->cur_pic_ptr;
478 struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
481 if (ctx_pic->slice_count >= MAX_SLICES)
484 if (!ctx_pic->bitstream)
485 ctx_pic->bitstream = buffer;
486 ctx_pic->bitstream_size += size;
488 position = buffer - ctx_pic->bitstream;
489 if (is_slice_short(avctx, ctx))
490 fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
493 fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
494 &ctx_pic->pp, position, size);
495 ctx_pic->slice_count++;
497 if (sl->slice_type != AV_PICTURE_TYPE_I && sl->slice_type != AV_PICTURE_TYPE_SI)
498 ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
502 static int dxva2_h264_end_frame(AVCodecContext *avctx)
504 H264Context *h = avctx->priv_data;
505 H264SliceContext *sl = &h->slice_ctx[0];
506 struct dxva2_picture_context *ctx_pic =
507 h->cur_pic_ptr->hwaccel_picture_private;
510 if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
512 ret = ff_dxva2_common_end_frame(avctx, h->cur_pic_ptr->f,
513 &ctx_pic->pp, sizeof(ctx_pic->pp),
514 &ctx_pic->qm, sizeof(ctx_pic->qm),
515 commit_bitstream_and_slice_buffer);
517 ff_h264_draw_horiz_band(h, sl, 0, h->avctx->height);
521 #if CONFIG_H264_DXVA2_HWACCEL
522 AVHWAccel ff_h264_dxva2_hwaccel = {
523 .name = "h264_dxva2",
524 .type = AVMEDIA_TYPE_VIDEO,
525 .id = AV_CODEC_ID_H264,
526 .pix_fmt = AV_PIX_FMT_DXVA2_VLD,
527 .start_frame = dxva2_h264_start_frame,
528 .decode_slice = dxva2_h264_decode_slice,
529 .end_frame = dxva2_h264_end_frame,
530 .frame_priv_data_size = sizeof(struct dxva2_picture_context),
534 #if CONFIG_H264_D3D11VA_HWACCEL
535 AVHWAccel ff_h264_d3d11va_hwaccel = {
536 .name = "h264_d3d11va",
537 .type = AVMEDIA_TYPE_VIDEO,
538 .id = AV_CODEC_ID_H264,
539 .pix_fmt = AV_PIX_FMT_D3D11VA_VLD,
540 .start_frame = dxva2_h264_start_frame,
541 .decode_slice = dxva2_h264_decode_slice,
542 .end_frame = dxva2_h264_end_frame,
543 .frame_priv_data_size = sizeof(struct dxva2_picture_context),