2 * DXVA2 H264 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 "dxva2_internal.h"
26 #include "mpegutils.h"
28 struct dxva2_picture_context {
29 DXVA_PicParams_H264 pp;
32 DXVA_Slice_H264_Short slice_short[MAX_SLICES];
33 DXVA_Slice_H264_Long slice_long[MAX_SLICES];
34 const uint8_t *bitstream;
35 unsigned bitstream_size;
38 static void fill_picture_entry(DXVA_PicEntry_H264 *pic,
39 unsigned index, unsigned flag)
41 assert((index&0x7f) == index && (flag&0x01) == flag);
42 pic->bPicEntry = index | (flag << 7);
45 static void fill_picture_parameters(const AVCodecContext *avctx, AVDXVAContext *ctx, const H264Context *h,
46 DXVA_PicParams_H264 *pp)
48 const H264Picture *current_picture = h->cur_pic_ptr;
51 memset(pp, 0, sizeof(*pp));
52 /* Configure current picture */
53 fill_picture_entry(&pp->CurrPic,
54 ff_dxva2_get_surface_index(avctx, ctx, current_picture->f),
55 h->picture_structure == PICT_BOTTOM_FIELD);
56 /* Configure the set of references */
57 pp->UsedForReferenceFlags = 0;
58 pp->NonExistingFrameFlags = 0;
59 for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
61 if (j < h->short_ref_count) {
62 r = h->short_ref[j++];
65 while (!r && j < h->short_ref_count + 16)
66 r = h->long_ref[j++ - h->short_ref_count];
69 fill_picture_entry(&pp->RefFrameList[i],
70 ff_dxva2_get_surface_index(avctx, ctx, r->f),
73 if ((r->reference & PICT_TOP_FIELD) && r->field_poc[0] != INT_MAX)
74 pp->FieldOrderCntList[i][0] = r->field_poc[0];
75 if ((r->reference & PICT_BOTTOM_FIELD) && r->field_poc[1] != INT_MAX)
76 pp->FieldOrderCntList[i][1] = r->field_poc[1];
78 pp->FrameNumList[i] = r->long_ref ? r->pic_id : r->frame_num;
79 if (r->reference & PICT_TOP_FIELD)
80 pp->UsedForReferenceFlags |= 1 << (2*i + 0);
81 if (r->reference & PICT_BOTTOM_FIELD)
82 pp->UsedForReferenceFlags |= 1 << (2*i + 1);
84 pp->RefFrameList[i].bPicEntry = 0xff;
85 pp->FieldOrderCntList[i][0] = 0;
86 pp->FieldOrderCntList[i][1] = 0;
87 pp->FrameNumList[i] = 0;
91 pp->wFrameWidthInMbsMinus1 = h->mb_width - 1;
92 pp->wFrameHeightInMbsMinus1 = h->mb_height - 1;
93 pp->num_ref_frames = h->sps.ref_frame_count;
95 pp->wBitFields = ((h->picture_structure != PICT_FRAME) << 0) |
97 (h->picture_structure == PICT_FRAME)) << 1) |
98 (h->sps.residual_color_transform_flag << 2) |
99 /* sp_for_switch_flag (not implemented by FFmpeg) */
101 (h->sps.chroma_format_idc << 4) |
102 ((h->nal_ref_idc != 0) << 6) |
103 (h->pps.constrained_intra_pred << 7) |
104 (h->pps.weighted_pred << 8) |
105 (h->pps.weighted_bipred_idc << 9) |
106 /* MbsConsecutiveFlag */
108 (h->sps.frame_mbs_only_flag << 12) |
109 (h->pps.transform_8x8_mode << 13) |
110 ((h->sps.level_idc >= 31) << 14) |
111 /* IntraPicFlag (Modified if we detect a non
112 * intra slice in dxva2_h264_decode_slice) */
115 pp->bit_depth_luma_minus8 = h->sps.bit_depth_luma - 8;
116 pp->bit_depth_chroma_minus8 = h->sps.bit_depth_chroma - 8;
117 if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG)
118 pp->Reserved16Bits = 0;
119 else if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
120 pp->Reserved16Bits = 0x34c;
122 pp->Reserved16Bits = 3; /* FIXME is there a way to detect the right mode ? */
123 pp->StatusReportFeedbackNumber = 1 + DXVA_CONTEXT_REPORT_ID(avctx, ctx)++;
124 pp->CurrFieldOrderCnt[0] = 0;
125 if ((h->picture_structure & PICT_TOP_FIELD) &&
126 current_picture->field_poc[0] != INT_MAX)
127 pp->CurrFieldOrderCnt[0] = current_picture->field_poc[0];
128 pp->CurrFieldOrderCnt[1] = 0;
129 if ((h->picture_structure & PICT_BOTTOM_FIELD) &&
130 current_picture->field_poc[1] != INT_MAX)
131 pp->CurrFieldOrderCnt[1] = current_picture->field_poc[1];
132 pp->pic_init_qs_minus26 = h->pps.init_qs - 26;
133 pp->chroma_qp_index_offset = h->pps.chroma_qp_index_offset[0];
134 pp->second_chroma_qp_index_offset = h->pps.chroma_qp_index_offset[1];
135 pp->ContinuationFlag = 1;
136 pp->pic_init_qp_minus26 = h->pps.init_qp - 26;
137 pp->num_ref_idx_l0_active_minus1 = h->pps.ref_count[0] - 1;
138 pp->num_ref_idx_l1_active_minus1 = h->pps.ref_count[1] - 1;
139 pp->Reserved8BitsA = 0;
140 pp->frame_num = h->frame_num;
141 pp->log2_max_frame_num_minus4 = h->sps.log2_max_frame_num - 4;
142 pp->pic_order_cnt_type = h->sps.poc_type;
143 if (h->sps.poc_type == 0)
144 pp->log2_max_pic_order_cnt_lsb_minus4 = h->sps.log2_max_poc_lsb - 4;
145 else if (h->sps.poc_type == 1)
146 pp->delta_pic_order_always_zero_flag = h->sps.delta_pic_order_always_zero_flag;
147 pp->direct_8x8_inference_flag = h->sps.direct_8x8_inference_flag;
148 pp->entropy_coding_mode_flag = h->pps.cabac;
149 pp->pic_order_present_flag = h->pps.pic_order_present;
150 pp->num_slice_groups_minus1 = h->pps.slice_group_count - 1;
151 pp->slice_group_map_type = h->pps.mb_slice_group_map_type;
152 pp->deblocking_filter_control_present_flag = h->pps.deblocking_filter_parameters_present;
153 pp->redundant_pic_cnt_present_flag= h->pps.redundant_pic_cnt_present;
154 pp->Reserved8BitsB = 0;
155 pp->slice_group_change_rate_minus1= 0; /* XXX not implemented by FFmpeg */
156 //pp->SliceGroupMap[810]; /* XXX not implemented by FFmpeg */
159 static void fill_scaling_lists(const AVCodecContext *avctx, AVDXVAContext *ctx, const H264Context *h, DXVA_Qmatrix_H264 *qm)
162 memset(qm, 0, sizeof(*qm));
163 if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG) {
164 for (i = 0; i < 6; i++)
165 for (j = 0; j < 16; j++)
166 qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][j];
168 for (i = 0; i < 64; i++) {
169 qm->bScalingLists8x8[0][i] = h->pps.scaling_matrix8[0][i];
170 qm->bScalingLists8x8[1][i] = h->pps.scaling_matrix8[3][i];
173 for (i = 0; i < 6; i++)
174 for (j = 0; j < 16; j++)
175 qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][zigzag_scan[j]];
177 for (i = 0; i < 64; i++) {
178 qm->bScalingLists8x8[0][i] = h->pps.scaling_matrix8[0][ff_zigzag_direct[i]];
179 qm->bScalingLists8x8[1][i] = h->pps.scaling_matrix8[3][ff_zigzag_direct[i]];
184 static int is_slice_short(const AVCodecContext *avctx, AVDXVAContext *ctx)
186 assert(DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 1 ||
187 DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 2);
188 return DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 2;
191 static void fill_slice_short(DXVA_Slice_H264_Short *slice,
192 unsigned position, unsigned size)
194 memset(slice, 0, sizeof(*slice));
195 slice->BSNALunitDataLocation = position;
196 slice->SliceBytesInBuffer = size;
197 slice->wBadSliceChopping = 0;
200 static int get_refpic_index(const DXVA_PicParams_H264 *pp, int surface_index)
203 for (i = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
204 if ((pp->RefFrameList[i].bPicEntry & 0x7f) == surface_index)
210 static void fill_slice_long(AVCodecContext *avctx, DXVA_Slice_H264_Long *slice,
211 const DXVA_PicParams_H264 *pp, unsigned position, unsigned size)
213 const H264Context *h = avctx->priv_data;
214 H264SliceContext *sl = &h->slice_ctx[0];
215 AVDXVAContext *ctx = avctx->hwaccel_context;
218 memset(slice, 0, sizeof(*slice));
219 slice->BSNALunitDataLocation = position;
220 slice->SliceBytesInBuffer = size;
221 slice->wBadSliceChopping = 0;
223 slice->first_mb_in_slice = (sl->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + sl->mb_x;
224 slice->NumMbsForSlice = 0; /* XXX it is set once we have all slices */
225 slice->BitOffsetToSliceData = get_bits_count(&sl->gb);
226 slice->slice_type = ff_h264_get_slice_type(sl);
227 if (sl->slice_type_fixed)
228 slice->slice_type += 5;
229 slice->luma_log2_weight_denom = sl->luma_log2_weight_denom;
230 slice->chroma_log2_weight_denom = sl->chroma_log2_weight_denom;
231 if (sl->list_count > 0)
232 slice->num_ref_idx_l0_active_minus1 = sl->ref_count[0] - 1;
233 if (sl->list_count > 1)
234 slice->num_ref_idx_l1_active_minus1 = sl->ref_count[1] - 1;
235 slice->slice_alpha_c0_offset_div2 = sl->slice_alpha_c0_offset / 2;
236 slice->slice_beta_offset_div2 = sl->slice_beta_offset / 2;
237 slice->Reserved8Bits = 0;
239 for (list = 0; list < 2; list++) {
241 for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
242 if (list < sl->list_count && i < sl->ref_count[list]) {
243 const H264Picture *r = sl->ref_list[list][i].parent;
246 if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
247 index = ff_dxva2_get_surface_index(avctx, ctx, r->f);
249 index = get_refpic_index(pp, ff_dxva2_get_surface_index(avctx, ctx, r->f));
250 fill_picture_entry(&slice->RefPicList[list][i], index,
251 r->reference == PICT_BOTTOM_FIELD);
252 for (plane = 0; plane < 3; plane++) {
254 if (plane == 0 && sl->luma_weight_flag[list]) {
255 w = sl->luma_weight[i][list][0];
256 o = sl->luma_weight[i][list][1];
257 } else if (plane >= 1 && sl->chroma_weight_flag[list]) {
258 w = sl->chroma_weight[i][list][plane-1][0];
259 o = sl->chroma_weight[i][list][plane-1][1];
261 w = 1 << (plane == 0 ? sl->luma_log2_weight_denom :
262 sl->chroma_log2_weight_denom);
265 slice->Weights[list][i][plane][0] = w;
266 slice->Weights[list][i][plane][1] = o;
270 slice->RefPicList[list][i].bPicEntry = 0xff;
271 for (plane = 0; plane < 3; plane++) {
272 slice->Weights[list][i][plane][0] = 0;
273 slice->Weights[list][i][plane][1] = 0;
278 slice->slice_qs_delta = 0; /* XXX not implemented by FFmpeg */
279 slice->slice_qp_delta = sl->qscale - h->pps.init_qp;
280 slice->redundant_pic_cnt = sl->redundant_pic_count;
281 if (sl->slice_type == AV_PICTURE_TYPE_B)
282 slice->direct_spatial_mv_pred_flag = sl->direct_spatial_mv_pred;
283 slice->cabac_init_idc = h->pps.cabac ? sl->cabac_init_idc : 0;
284 if (sl->deblocking_filter < 2)
285 slice->disable_deblocking_filter_idc = 1 - sl->deblocking_filter;
287 slice->disable_deblocking_filter_idc = sl->deblocking_filter;
288 slice->slice_id = h->current_slice - 1;
291 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
292 DECODER_BUFFER_DESC *bs,
293 DECODER_BUFFER_DESC *sc)
295 const H264Context *h = avctx->priv_data;
296 const unsigned mb_count = h->mb_width * h->mb_height;
297 AVDXVAContext *ctx = avctx->hwaccel_context;
298 const H264Picture *current_picture = h->cur_pic_ptr;
299 struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
300 DXVA_Slice_H264_Short *slice = NULL;
301 void *dxva_data_ptr = NULL;
302 uint8_t *dxva_data, *current, *end;
303 unsigned dxva_size = 0;
310 /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
312 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
313 type = D3D11_VIDEO_DECODER_BUFFER_BITSTREAM;
314 if (FAILED(ID3D11VideoContext_GetDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context,
315 D3D11VA_CONTEXT(ctx)->decoder,
317 &dxva_size, &dxva_data_ptr)))
322 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
323 type = DXVA2_BitStreamDateBufferType;
324 if (FAILED(IDirectXVideoDecoder_GetBuffer(DXVA2_CONTEXT(ctx)->decoder,
326 &dxva_data_ptr, &dxva_size)))
331 dxva_data = dxva_data_ptr;
333 end = dxva_data + dxva_size;
335 for (i = 0; i < ctx_pic->slice_count; i++) {
336 static const uint8_t start_code[] = { 0, 0, 1 };
337 static const unsigned start_code_size = sizeof(start_code);
338 unsigned position, size;
340 assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
341 offsetof(DXVA_Slice_H264_Long, BSNALunitDataLocation));
342 assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
343 offsetof(DXVA_Slice_H264_Long, SliceBytesInBuffer));
345 if (is_slice_short(avctx, ctx))
346 slice = &ctx_pic->slice_short[i];
348 slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
350 position = slice->BSNALunitDataLocation;
351 size = slice->SliceBytesInBuffer;
352 if (start_code_size + size > end - current) {
353 av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
357 slice->BSNALunitDataLocation = current - dxva_data;
358 slice->SliceBytesInBuffer = start_code_size + size;
360 if (!is_slice_short(avctx, ctx)) {
361 DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
362 if (i < ctx_pic->slice_count - 1)
363 slice_long->NumMbsForSlice =
364 slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
366 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
369 memcpy(current, start_code, start_code_size);
370 current += start_code_size;
372 memcpy(current, &ctx_pic->bitstream[position], size);
375 padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
376 if (slice && padding > 0) {
377 memset(current, 0, padding);
380 slice->SliceBytesInBuffer += padding;
383 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD)
384 if (FAILED(ID3D11VideoContext_ReleaseDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder, type)))
388 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
389 if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(DXVA2_CONTEXT(ctx)->decoder, type)))
392 if (i < ctx_pic->slice_count)
396 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
397 D3D11_VIDEO_DECODER_BUFFER_DESC *dsc11 = bs;
398 memset(dsc11, 0, sizeof(*dsc11));
399 dsc11->BufferType = type;
400 dsc11->DataSize = current - dxva_data;
401 dsc11->NumMBsInBuffer = mb_count;
403 type = D3D11_VIDEO_DECODER_BUFFER_SLICE_CONTROL;
407 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
408 DXVA2_DecodeBufferDesc *dsc2 = bs;
409 memset(dsc2, 0, sizeof(*dsc2));
410 dsc2->CompressedBufferType = type;
411 dsc2->DataSize = current - dxva_data;
412 dsc2->NumMBsInBuffer = mb_count;
414 type = DXVA2_SliceControlBufferType;
418 if (is_slice_short(avctx, ctx)) {
419 slice_data = ctx_pic->slice_short;
420 slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
422 slice_data = ctx_pic->slice_long;
423 slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
425 assert((bs->DataSize & 127) == 0);
426 return ff_dxva2_commit_buffer(avctx, ctx, sc,
428 slice_data, slice_size, mb_count);
432 static int dxva2_h264_start_frame(AVCodecContext *avctx,
433 av_unused const uint8_t *buffer,
434 av_unused uint32_t size)
436 const H264Context *h = avctx->priv_data;
437 AVDXVAContext *ctx = avctx->hwaccel_context;
438 struct dxva2_picture_context *ctx_pic = h->cur_pic_ptr->hwaccel_picture_private;
440 if (DXVA_CONTEXT_DECODER(avctx, ctx) == NULL ||
441 DXVA_CONTEXT_CFG(avctx, ctx) == NULL ||
442 DXVA_CONTEXT_COUNT(avctx, ctx) <= 0)
446 /* Fill up DXVA_PicParams_H264 */
447 fill_picture_parameters(avctx, ctx, h, &ctx_pic->pp);
449 /* Fill up DXVA_Qmatrix_H264 */
450 fill_scaling_lists(avctx, ctx, h, &ctx_pic->qm);
452 ctx_pic->slice_count = 0;
453 ctx_pic->bitstream_size = 0;
454 ctx_pic->bitstream = NULL;
458 static int dxva2_h264_decode_slice(AVCodecContext *avctx,
459 const uint8_t *buffer,
462 const H264Context *h = avctx->priv_data;
463 const H264SliceContext *sl = &h->slice_ctx[0];
464 AVDXVAContext *ctx = avctx->hwaccel_context;
465 const H264Picture *current_picture = h->cur_pic_ptr;
466 struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
469 if (ctx_pic->slice_count >= MAX_SLICES)
472 if (!ctx_pic->bitstream)
473 ctx_pic->bitstream = buffer;
474 ctx_pic->bitstream_size += size;
476 position = buffer - ctx_pic->bitstream;
477 if (is_slice_short(avctx, ctx))
478 fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
481 fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
482 &ctx_pic->pp, position, size);
483 ctx_pic->slice_count++;
485 if (sl->slice_type != AV_PICTURE_TYPE_I && sl->slice_type != AV_PICTURE_TYPE_SI)
486 ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
490 static int dxva2_h264_end_frame(AVCodecContext *avctx)
492 H264Context *h = avctx->priv_data;
493 H264SliceContext *sl = &h->slice_ctx[0];
494 struct dxva2_picture_context *ctx_pic =
495 h->cur_pic_ptr->hwaccel_picture_private;
498 if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
500 ret = ff_dxva2_common_end_frame(avctx, h->cur_pic_ptr->f,
501 &ctx_pic->pp, sizeof(ctx_pic->pp),
502 &ctx_pic->qm, sizeof(ctx_pic->qm),
503 commit_bitstream_and_slice_buffer);
505 ff_h264_draw_horiz_band(h, sl, 0, h->avctx->height);
509 #if CONFIG_H264_DXVA2_HWACCEL
510 AVHWAccel ff_h264_dxva2_hwaccel = {
511 .name = "h264_dxva2",
512 .type = AVMEDIA_TYPE_VIDEO,
513 .id = AV_CODEC_ID_H264,
514 .pix_fmt = AV_PIX_FMT_DXVA2_VLD,
515 .start_frame = dxva2_h264_start_frame,
516 .decode_slice = dxva2_h264_decode_slice,
517 .end_frame = dxva2_h264_end_frame,
518 .frame_priv_data_size = sizeof(struct dxva2_picture_context),
522 #if CONFIG_H264_D3D11VA_HWACCEL
523 AVHWAccel ff_h264_d3d11va_hwaccel = {
524 .name = "h264_d3d11va",
525 .type = AVMEDIA_TYPE_VIDEO,
526 .id = AV_CODEC_ID_H264,
527 .pix_fmt = AV_PIX_FMT_D3D11VA_VLD,
528 .start_frame = dxva2_h264_start_frame,
529 .decode_slice = dxva2_h264_decode_slice,
530 .end_frame = dxva2_h264_end_frame,
531 .frame_priv_data_size = sizeof(struct dxva2_picture_context),