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 "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;
57 memset(pp, 0, sizeof(*pp));
58 /* Configure current picture */
59 fill_picture_entry(&pp->CurrPic,
60 ff_dxva2_get_surface_index(avctx, ctx, current_picture->f),
61 h->picture_structure == PICT_BOTTOM_FIELD);
62 /* Configure the set of references */
63 pp->UsedForReferenceFlags = 0;
64 pp->NonExistingFrameFlags = 0;
65 for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
67 if (j < h->short_ref_count) {
68 r = h->short_ref[j++];
71 while (!r && j < h->short_ref_count + 16)
72 r = h->long_ref[j++ - h->short_ref_count];
75 fill_picture_entry(&pp->RefFrameList[i],
76 ff_dxva2_get_surface_index(avctx, ctx, r->f),
79 if ((r->reference & PICT_TOP_FIELD) && r->field_poc[0] != INT_MAX)
80 pp->FieldOrderCntList[i][0] = r->field_poc[0];
81 if ((r->reference & PICT_BOTTOM_FIELD) && r->field_poc[1] != INT_MAX)
82 pp->FieldOrderCntList[i][1] = r->field_poc[1];
84 pp->FrameNumList[i] = r->long_ref ? r->pic_id : r->frame_num;
85 if (r->reference & PICT_TOP_FIELD)
86 pp->UsedForReferenceFlags |= 1 << (2*i + 0);
87 if (r->reference & PICT_BOTTOM_FIELD)
88 pp->UsedForReferenceFlags |= 1 << (2*i + 1);
90 pp->RefFrameList[i].bPicEntry = 0xff;
91 pp->FieldOrderCntList[i][0] = 0;
92 pp->FieldOrderCntList[i][1] = 0;
93 pp->FrameNumList[i] = 0;
97 pp->wFrameWidthInMbsMinus1 = h->mb_width - 1;
98 pp->wFrameHeightInMbsMinus1 = h->mb_height - 1;
99 pp->num_ref_frames = h->sps.ref_frame_count;
101 pp->wBitFields = ((h->picture_structure != PICT_FRAME) << 0) |
103 (h->picture_structure == PICT_FRAME)) << 1) |
104 (h->sps.residual_color_transform_flag << 2) |
105 /* sp_for_switch_flag (not implemented by FFmpeg) */
107 (h->sps.chroma_format_idc << 4) |
108 ((h->nal_ref_idc != 0) << 6) |
109 (h->pps.constrained_intra_pred << 7) |
110 (h->pps.weighted_pred << 8) |
111 (h->pps.weighted_bipred_idc << 9) |
112 /* MbsConsecutiveFlag */
114 (h->sps.frame_mbs_only_flag << 12) |
115 (h->pps.transform_8x8_mode << 13) |
116 ((h->sps.level_idc >= 31) << 14) |
117 /* IntraPicFlag (Modified if we detect a non
118 * intra slice in dxva2_h264_decode_slice) */
121 pp->bit_depth_luma_minus8 = h->sps.bit_depth_luma - 8;
122 pp->bit_depth_chroma_minus8 = h->sps.bit_depth_chroma - 8;
123 if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG)
124 pp->Reserved16Bits = 0;
125 else if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
126 pp->Reserved16Bits = 0x34c;
128 pp->Reserved16Bits = 3; /* FIXME is there a way to detect the right mode ? */
129 pp->StatusReportFeedbackNumber = 1 + DXVA_CONTEXT_REPORT_ID(avctx, ctx)++;
130 pp->CurrFieldOrderCnt[0] = 0;
131 if ((h->picture_structure & PICT_TOP_FIELD) &&
132 current_picture->field_poc[0] != INT_MAX)
133 pp->CurrFieldOrderCnt[0] = current_picture->field_poc[0];
134 pp->CurrFieldOrderCnt[1] = 0;
135 if ((h->picture_structure & PICT_BOTTOM_FIELD) &&
136 current_picture->field_poc[1] != INT_MAX)
137 pp->CurrFieldOrderCnt[1] = current_picture->field_poc[1];
138 pp->pic_init_qs_minus26 = h->pps.init_qs - 26;
139 pp->chroma_qp_index_offset = h->pps.chroma_qp_index_offset[0];
140 pp->second_chroma_qp_index_offset = h->pps.chroma_qp_index_offset[1];
141 pp->ContinuationFlag = 1;
142 pp->pic_init_qp_minus26 = h->pps.init_qp - 26;
143 pp->num_ref_idx_l0_active_minus1 = h->pps.ref_count[0] - 1;
144 pp->num_ref_idx_l1_active_minus1 = h->pps.ref_count[1] - 1;
145 pp->Reserved8BitsA = 0;
146 pp->frame_num = h->frame_num;
147 pp->log2_max_frame_num_minus4 = h->sps.log2_max_frame_num - 4;
148 pp->pic_order_cnt_type = h->sps.poc_type;
149 if (h->sps.poc_type == 0)
150 pp->log2_max_pic_order_cnt_lsb_minus4 = h->sps.log2_max_poc_lsb - 4;
151 else if (h->sps.poc_type == 1)
152 pp->delta_pic_order_always_zero_flag = h->sps.delta_pic_order_always_zero_flag;
153 pp->direct_8x8_inference_flag = h->sps.direct_8x8_inference_flag;
154 pp->entropy_coding_mode_flag = h->pps.cabac;
155 pp->pic_order_present_flag = h->pps.pic_order_present;
156 pp->num_slice_groups_minus1 = h->pps.slice_group_count - 1;
157 pp->slice_group_map_type = h->pps.mb_slice_group_map_type;
158 pp->deblocking_filter_control_present_flag = h->pps.deblocking_filter_parameters_present;
159 pp->redundant_pic_cnt_present_flag= h->pps.redundant_pic_cnt_present;
160 pp->Reserved8BitsB = 0;
161 pp->slice_group_change_rate_minus1= 0; /* XXX not implemented by FFmpeg */
162 //pp->SliceGroupMap[810]; /* XXX not implemented by FFmpeg */
165 static void fill_scaling_lists(const AVCodecContext *avctx, AVDXVAContext *ctx, const H264Context *h, DXVA_Qmatrix_H264 *qm)
168 memset(qm, 0, sizeof(*qm));
169 if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG) {
170 for (i = 0; i < 6; i++)
171 for (j = 0; j < 16; j++)
172 qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][j];
174 for (i = 0; i < 64; i++) {
175 qm->bScalingLists8x8[0][i] = h->pps.scaling_matrix8[0][i];
176 qm->bScalingLists8x8[1][i] = h->pps.scaling_matrix8[3][i];
179 for (i = 0; i < 6; i++)
180 for (j = 0; j < 16; j++)
181 qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][ff_zigzag_scan[j]];
183 for (i = 0; i < 64; i++) {
184 qm->bScalingLists8x8[0][i] = h->pps.scaling_matrix8[0][ff_zigzag_direct[i]];
185 qm->bScalingLists8x8[1][i] = h->pps.scaling_matrix8[3][ff_zigzag_direct[i]];
190 static int is_slice_short(const AVCodecContext *avctx, AVDXVAContext *ctx)
192 assert(DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 1 ||
193 DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 2);
194 return DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 2;
197 static void fill_slice_short(DXVA_Slice_H264_Short *slice,
198 unsigned position, unsigned size)
200 memset(slice, 0, sizeof(*slice));
201 slice->BSNALunitDataLocation = position;
202 slice->SliceBytesInBuffer = size;
203 slice->wBadSliceChopping = 0;
206 static int get_refpic_index(const DXVA_PicParams_H264 *pp, int surface_index)
209 for (i = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
210 if ((pp->RefFrameList[i].bPicEntry & 0x7f) == surface_index)
216 static void fill_slice_long(AVCodecContext *avctx, DXVA_Slice_H264_Long *slice,
217 const DXVA_PicParams_H264 *pp, unsigned position, unsigned size)
219 const H264Context *h = avctx->priv_data;
220 H264SliceContext *sl = &h->slice_ctx[0];
221 AVDXVAContext *ctx = avctx->hwaccel_context;
224 memset(slice, 0, sizeof(*slice));
225 slice->BSNALunitDataLocation = position;
226 slice->SliceBytesInBuffer = size;
227 slice->wBadSliceChopping = 0;
229 slice->first_mb_in_slice = (sl->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + sl->mb_x;
230 slice->NumMbsForSlice = 0; /* XXX it is set once we have all slices */
231 slice->BitOffsetToSliceData = get_bits_count(&sl->gb);
232 slice->slice_type = ff_h264_get_slice_type(sl);
233 if (sl->slice_type_fixed)
234 slice->slice_type += 5;
235 slice->luma_log2_weight_denom = sl->luma_log2_weight_denom;
236 slice->chroma_log2_weight_denom = sl->chroma_log2_weight_denom;
237 if (sl->list_count > 0)
238 slice->num_ref_idx_l0_active_minus1 = sl->ref_count[0] - 1;
239 if (sl->list_count > 1)
240 slice->num_ref_idx_l1_active_minus1 = sl->ref_count[1] - 1;
241 slice->slice_alpha_c0_offset_div2 = sl->slice_alpha_c0_offset / 2;
242 slice->slice_beta_offset_div2 = sl->slice_beta_offset / 2;
243 slice->Reserved8Bits = 0;
245 for (list = 0; list < 2; list++) {
247 for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
248 if (list < sl->list_count && i < sl->ref_count[list]) {
249 const H264Picture *r = sl->ref_list[list][i].parent;
252 if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
253 index = ff_dxva2_get_surface_index(avctx, ctx, r->f);
255 index = get_refpic_index(pp, ff_dxva2_get_surface_index(avctx, ctx, r->f));
256 fill_picture_entry(&slice->RefPicList[list][i], index,
257 sl->ref_list[list][i].reference == PICT_BOTTOM_FIELD);
258 for (plane = 0; plane < 3; plane++) {
260 if (plane == 0 && sl->luma_weight_flag[list]) {
261 w = sl->luma_weight[i][list][0];
262 o = sl->luma_weight[i][list][1];
263 } else if (plane >= 1 && sl->chroma_weight_flag[list]) {
264 w = sl->chroma_weight[i][list][plane-1][0];
265 o = sl->chroma_weight[i][list][plane-1][1];
267 w = 1 << (plane == 0 ? sl->luma_log2_weight_denom :
268 sl->chroma_log2_weight_denom);
271 slice->Weights[list][i][plane][0] = w;
272 slice->Weights[list][i][plane][1] = o;
276 slice->RefPicList[list][i].bPicEntry = 0xff;
277 for (plane = 0; plane < 3; plane++) {
278 slice->Weights[list][i][plane][0] = 0;
279 slice->Weights[list][i][plane][1] = 0;
284 slice->slice_qs_delta = 0; /* XXX not implemented by FFmpeg */
285 slice->slice_qp_delta = sl->qscale - h->pps.init_qp;
286 slice->redundant_pic_cnt = sl->redundant_pic_count;
287 if (sl->slice_type == AV_PICTURE_TYPE_B)
288 slice->direct_spatial_mv_pred_flag = sl->direct_spatial_mv_pred;
289 slice->cabac_init_idc = h->pps.cabac ? sl->cabac_init_idc : 0;
290 if (sl->deblocking_filter < 2)
291 slice->disable_deblocking_filter_idc = 1 - sl->deblocking_filter;
293 slice->disable_deblocking_filter_idc = sl->deblocking_filter;
294 slice->slice_id = h->current_slice - 1;
297 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
298 DECODER_BUFFER_DESC *bs,
299 DECODER_BUFFER_DESC *sc)
301 const H264Context *h = avctx->priv_data;
302 const unsigned mb_count = h->mb_width * h->mb_height;
303 AVDXVAContext *ctx = avctx->hwaccel_context;
304 const H264Picture *current_picture = h->cur_pic_ptr;
305 struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
306 DXVA_Slice_H264_Short *slice = NULL;
307 void *dxva_data_ptr = NULL;
308 uint8_t *dxva_data, *current, *end;
309 unsigned dxva_size = 0;
316 /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
318 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
319 type = D3D11_VIDEO_DECODER_BUFFER_BITSTREAM;
320 if (FAILED(ID3D11VideoContext_GetDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context,
321 D3D11VA_CONTEXT(ctx)->decoder,
323 &dxva_size, &dxva_data_ptr)))
328 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
329 type = DXVA2_BitStreamDateBufferType;
330 if (FAILED(IDirectXVideoDecoder_GetBuffer(DXVA2_CONTEXT(ctx)->decoder,
332 &dxva_data_ptr, &dxva_size)))
337 dxva_data = dxva_data_ptr;
339 end = dxva_data + dxva_size;
341 for (i = 0; i < ctx_pic->slice_count; i++) {
342 static const uint8_t start_code[] = { 0, 0, 1 };
343 static const unsigned start_code_size = sizeof(start_code);
344 unsigned position, size;
346 assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
347 offsetof(DXVA_Slice_H264_Long, BSNALunitDataLocation));
348 assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
349 offsetof(DXVA_Slice_H264_Long, SliceBytesInBuffer));
351 if (is_slice_short(avctx, ctx))
352 slice = &ctx_pic->slice_short[i];
354 slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
356 position = slice->BSNALunitDataLocation;
357 size = slice->SliceBytesInBuffer;
358 if (start_code_size + size > end - current) {
359 av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
363 slice->BSNALunitDataLocation = current - dxva_data;
364 slice->SliceBytesInBuffer = start_code_size + size;
366 if (!is_slice_short(avctx, ctx)) {
367 DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
368 if (i < ctx_pic->slice_count - 1)
369 slice_long->NumMbsForSlice =
370 slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
372 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
375 memcpy(current, start_code, start_code_size);
376 current += start_code_size;
378 memcpy(current, &ctx_pic->bitstream[position], size);
381 padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
382 if (slice && padding > 0) {
383 memset(current, 0, padding);
386 slice->SliceBytesInBuffer += padding;
389 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD)
390 if (FAILED(ID3D11VideoContext_ReleaseDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder, type)))
394 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
395 if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(DXVA2_CONTEXT(ctx)->decoder, type)))
398 if (i < ctx_pic->slice_count)
402 if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
403 D3D11_VIDEO_DECODER_BUFFER_DESC *dsc11 = bs;
404 memset(dsc11, 0, sizeof(*dsc11));
405 dsc11->BufferType = type;
406 dsc11->DataSize = current - dxva_data;
407 dsc11->NumMBsInBuffer = mb_count;
409 type = D3D11_VIDEO_DECODER_BUFFER_SLICE_CONTROL;
411 av_assert0((dsc11->DataSize & 127) == 0);
415 if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
416 DXVA2_DecodeBufferDesc *dsc2 = bs;
417 memset(dsc2, 0, sizeof(*dsc2));
418 dsc2->CompressedBufferType = type;
419 dsc2->DataSize = current - dxva_data;
420 dsc2->NumMBsInBuffer = mb_count;
422 type = DXVA2_SliceControlBufferType;
424 av_assert0((dsc2->DataSize & 127) == 0);
428 if (is_slice_short(avctx, ctx)) {
429 slice_data = ctx_pic->slice_short;
430 slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
432 slice_data = ctx_pic->slice_long;
433 slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
435 return ff_dxva2_commit_buffer(avctx, ctx, sc,
437 slice_data, slice_size, mb_count);
441 static int dxva2_h264_start_frame(AVCodecContext *avctx,
442 av_unused const uint8_t *buffer,
443 av_unused uint32_t size)
445 const H264Context *h = avctx->priv_data;
446 AVDXVAContext *ctx = avctx->hwaccel_context;
447 struct dxva2_picture_context *ctx_pic = h->cur_pic_ptr->hwaccel_picture_private;
449 if (DXVA_CONTEXT_DECODER(avctx, ctx) == NULL ||
450 DXVA_CONTEXT_CFG(avctx, ctx) == NULL ||
451 DXVA_CONTEXT_COUNT(avctx, ctx) <= 0)
455 /* Fill up DXVA_PicParams_H264 */
456 fill_picture_parameters(avctx, ctx, h, &ctx_pic->pp);
458 /* Fill up DXVA_Qmatrix_H264 */
459 fill_scaling_lists(avctx, ctx, h, &ctx_pic->qm);
461 ctx_pic->slice_count = 0;
462 ctx_pic->bitstream_size = 0;
463 ctx_pic->bitstream = NULL;
467 static int dxva2_h264_decode_slice(AVCodecContext *avctx,
468 const uint8_t *buffer,
471 const H264Context *h = avctx->priv_data;
472 const H264SliceContext *sl = &h->slice_ctx[0];
473 AVDXVAContext *ctx = avctx->hwaccel_context;
474 const H264Picture *current_picture = h->cur_pic_ptr;
475 struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
478 if (ctx_pic->slice_count >= MAX_SLICES)
481 if (!ctx_pic->bitstream)
482 ctx_pic->bitstream = buffer;
483 ctx_pic->bitstream_size += size;
485 position = buffer - ctx_pic->bitstream;
486 if (is_slice_short(avctx, ctx))
487 fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
490 fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
491 &ctx_pic->pp, position, size);
492 ctx_pic->slice_count++;
494 if (sl->slice_type != AV_PICTURE_TYPE_I && sl->slice_type != AV_PICTURE_TYPE_SI)
495 ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
499 static int dxva2_h264_end_frame(AVCodecContext *avctx)
501 H264Context *h = avctx->priv_data;
502 H264SliceContext *sl = &h->slice_ctx[0];
503 struct dxva2_picture_context *ctx_pic =
504 h->cur_pic_ptr->hwaccel_picture_private;
507 if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
509 ret = ff_dxva2_common_end_frame(avctx, h->cur_pic_ptr->f,
510 &ctx_pic->pp, sizeof(ctx_pic->pp),
511 &ctx_pic->qm, sizeof(ctx_pic->qm),
512 commit_bitstream_and_slice_buffer);
514 ff_h264_draw_horiz_band(h, sl, 0, h->avctx->height);
518 #if CONFIG_H264_DXVA2_HWACCEL
519 AVHWAccel ff_h264_dxva2_hwaccel = {
520 .name = "h264_dxva2",
521 .type = AVMEDIA_TYPE_VIDEO,
522 .id = AV_CODEC_ID_H264,
523 .pix_fmt = AV_PIX_FMT_DXVA2_VLD,
524 .start_frame = dxva2_h264_start_frame,
525 .decode_slice = dxva2_h264_decode_slice,
526 .end_frame = dxva2_h264_end_frame,
527 .frame_priv_data_size = sizeof(struct dxva2_picture_context),
531 #if CONFIG_H264_D3D11VA_HWACCEL
532 AVHWAccel ff_h264_d3d11va_hwaccel = {
533 .name = "h264_d3d11va",
534 .type = AVMEDIA_TYPE_VIDEO,
535 .id = AV_CODEC_ID_H264,
536 .pix_fmt = AV_PIX_FMT_D3D11VA_VLD,
537 .start_frame = dxva2_h264_start_frame,
538 .decode_slice = dxva2_h264_decode_slice,
539 .end_frame = dxva2_h264_end_frame,
540 .frame_priv_data_size = sizeof(struct dxva2_picture_context),