2 * This file is part of Libav.
4 * Libav is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU Lesser General Public
6 * License as published by the Free Software Foundation; either
7 * version 2.1 of the License, or (at your option) any later version.
9 * Libav is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * Lesser General Public License for more details.
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with Libav; if not, write to the Free Software
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 #include "libavutil/avassert.h"
23 #include "libavutil/common.h"
24 #include "libavutil/log.h"
25 #include "libavutil/pixdesc.h"
27 #include "vaapi_encode.h"
30 static const char * const picture_type_name[] = { "IDR", "I", "P", "B" };
32 static int vaapi_encode_make_packed_header(AVCodecContext *avctx,
33 VAAPIEncodePicture *pic,
34 int type, char *data, size_t bit_len)
36 VAAPIEncodeContext *ctx = avctx->priv_data;
38 VABufferID param_buffer, data_buffer;
39 VAEncPackedHeaderParameterBuffer params = {
41 .bit_length = bit_len,
42 .has_emulation_bytes = 1,
45 av_assert0(pic->nb_param_buffers + 2 <= MAX_PARAM_BUFFERS);
47 vas = vaCreateBuffer(ctx->hwctx->display, ctx->va_context,
48 VAEncPackedHeaderParameterBufferType,
49 sizeof(params), 1, ¶ms, ¶m_buffer);
50 if (vas != VA_STATUS_SUCCESS) {
51 av_log(avctx, AV_LOG_ERROR, "Failed to create parameter buffer "
52 "for packed header (type %d): %d (%s).\n",
53 type, vas, vaErrorStr(vas));
56 pic->param_buffers[pic->nb_param_buffers++] = param_buffer;
58 vas = vaCreateBuffer(ctx->hwctx->display, ctx->va_context,
59 VAEncPackedHeaderDataBufferType,
60 (bit_len + 7) / 8, 1, data, &data_buffer);
61 if (vas != VA_STATUS_SUCCESS) {
62 av_log(avctx, AV_LOG_ERROR, "Failed to create data buffer "
63 "for packed header (type %d): %d (%s).\n",
64 type, vas, vaErrorStr(vas));
67 pic->param_buffers[pic->nb_param_buffers++] = data_buffer;
69 av_log(avctx, AV_LOG_DEBUG, "Packed header buffer (%d) is %#x/%#x "
70 "(%zu bits).\n", type, param_buffer, data_buffer, bit_len);
74 static int vaapi_encode_make_param_buffer(AVCodecContext *avctx,
75 VAAPIEncodePicture *pic,
76 int type, char *data, size_t len)
78 VAAPIEncodeContext *ctx = avctx->priv_data;
82 av_assert0(pic->nb_param_buffers + 1 <= MAX_PARAM_BUFFERS);
84 vas = vaCreateBuffer(ctx->hwctx->display, ctx->va_context,
85 type, len, 1, data, &buffer);
86 if (vas != VA_STATUS_SUCCESS) {
87 av_log(avctx, AV_LOG_ERROR, "Failed to create parameter buffer "
88 "(type %d): %d (%s).\n", type, vas, vaErrorStr(vas));
91 pic->param_buffers[pic->nb_param_buffers++] = buffer;
93 av_log(avctx, AV_LOG_DEBUG, "Param buffer (%d) is %#x.\n",
98 static int vaapi_encode_wait(AVCodecContext *avctx,
99 VAAPIEncodePicture *pic)
101 VAAPIEncodeContext *ctx = avctx->priv_data;
104 av_assert0(pic->encode_issued);
106 if (pic->encode_complete) {
107 // Already waited for this picture.
111 av_log(avctx, AV_LOG_DEBUG, "Sync to pic %"PRId64"/%"PRId64" "
112 "(input surface %#x).\n", pic->display_order,
113 pic->encode_order, pic->input_surface);
115 vas = vaSyncSurface(ctx->hwctx->display, pic->input_surface);
116 if (vas != VA_STATUS_SUCCESS) {
117 av_log(avctx, AV_LOG_ERROR, "Failed to sync to picture completion: "
118 "%d (%s).\n", vas, vaErrorStr(vas));
122 // Input is definitely finished with now.
123 av_frame_free(&pic->input_image);
125 pic->encode_complete = 1;
129 static int vaapi_encode_issue(AVCodecContext *avctx,
130 VAAPIEncodePicture *pic)
132 VAAPIEncodeContext *ctx = avctx->priv_data;
133 VAAPIEncodeSlice *slice;
136 char data[MAX_PARAM_BUFFER_SIZE];
139 av_log(avctx, AV_LOG_DEBUG, "Issuing encode for pic %"PRId64"/%"PRId64" "
140 "as type %s.\n", pic->display_order, pic->encode_order,
141 picture_type_name[pic->type]);
142 if (pic->nb_refs == 0) {
143 av_log(avctx, AV_LOG_DEBUG, "No reference pictures.\n");
145 av_log(avctx, AV_LOG_DEBUG, "Refers to:");
146 for (i = 0; i < pic->nb_refs; i++) {
147 av_log(avctx, AV_LOG_DEBUG, " %"PRId64"/%"PRId64,
148 pic->refs[i]->display_order, pic->refs[i]->encode_order);
150 av_log(avctx, AV_LOG_DEBUG, ".\n");
153 av_assert0(pic->input_available && !pic->encode_issued);
154 for (i = 0; i < pic->nb_refs; i++) {
155 av_assert0(pic->refs[i]);
156 // If we are serialised then the references must have already
157 // completed. If not, they must have been issued but need not
158 // have completed yet.
159 if (ctx->issue_mode == ISSUE_MODE_SERIALISE_EVERYTHING)
160 av_assert0(pic->refs[i]->encode_complete);
162 av_assert0(pic->refs[i]->encode_issued);
165 av_log(avctx, AV_LOG_DEBUG, "Input surface is %#x.\n", pic->input_surface);
167 pic->recon_image = av_frame_alloc();
168 if (!pic->recon_image) {
169 err = AVERROR(ENOMEM);
173 err = av_hwframe_get_buffer(ctx->recon_frames_ref, pic->recon_image, 0);
175 err = AVERROR(ENOMEM);
178 pic->recon_surface = (VASurfaceID)(uintptr_t)pic->recon_image->data[3];
179 av_log(avctx, AV_LOG_DEBUG, "Recon surface is %#x.\n", pic->recon_surface);
181 pic->output_buffer_ref = av_buffer_pool_get(ctx->output_buffer_pool);
182 if (!pic->output_buffer_ref) {
183 err = AVERROR(ENOMEM);
186 pic->output_buffer = (VABufferID)(uintptr_t)pic->output_buffer_ref->data;
187 av_log(avctx, AV_LOG_DEBUG, "Output buffer is %#x.\n",
190 if (ctx->codec->picture_params_size > 0) {
191 pic->codec_picture_params = av_malloc(ctx->codec->picture_params_size);
192 if (!pic->codec_picture_params)
194 memcpy(pic->codec_picture_params, ctx->codec_picture_params,
195 ctx->codec->picture_params_size);
197 av_assert0(!ctx->codec_picture_params);
200 pic->nb_param_buffers = 0;
202 if (pic->encode_order == 0) {
203 // Global parameter buffers are set on the first picture only.
205 for (i = 0; i < ctx->nb_global_params; i++) {
206 err = vaapi_encode_make_param_buffer(avctx, pic,
207 VAEncMiscParameterBufferType,
208 (char*)ctx->global_params[i],
209 ctx->global_params_size[i]);
215 if (pic->type == PICTURE_TYPE_IDR && ctx->codec->init_sequence_params) {
216 err = vaapi_encode_make_param_buffer(avctx, pic,
217 VAEncSequenceParameterBufferType,
218 ctx->codec_sequence_params,
219 ctx->codec->sequence_params_size);
224 if (ctx->codec->init_picture_params) {
225 err = ctx->codec->init_picture_params(avctx, pic);
227 av_log(avctx, AV_LOG_ERROR, "Failed to initialise picture "
228 "parameters: %d.\n", err);
231 err = vaapi_encode_make_param_buffer(avctx, pic,
232 VAEncPictureParameterBufferType,
233 pic->codec_picture_params,
234 ctx->codec->picture_params_size);
239 if (pic->type == PICTURE_TYPE_IDR) {
240 if (ctx->va_packed_headers & VA_ENC_PACKED_HEADER_SEQUENCE &&
241 ctx->codec->write_sequence_header) {
242 bit_len = 8 * sizeof(data);
243 err = ctx->codec->write_sequence_header(avctx, data, &bit_len);
245 av_log(avctx, AV_LOG_ERROR, "Failed to write per-sequence "
246 "header: %d.\n", err);
249 err = vaapi_encode_make_packed_header(avctx, pic,
250 ctx->codec->sequence_header_type,
257 if (ctx->va_packed_headers & VA_ENC_PACKED_HEADER_PICTURE &&
258 ctx->codec->write_picture_header) {
259 bit_len = 8 * sizeof(data);
260 err = ctx->codec->write_picture_header(avctx, pic, data, &bit_len);
262 av_log(avctx, AV_LOG_ERROR, "Failed to write per-picture "
263 "header: %d.\n", err);
266 err = vaapi_encode_make_packed_header(avctx, pic,
267 ctx->codec->picture_header_type,
273 if (ctx->codec->write_extra_buffer) {
275 size_t len = sizeof(data);
277 err = ctx->codec->write_extra_buffer(avctx, pic, i, &type,
279 if (err == AVERROR_EOF)
282 av_log(avctx, AV_LOG_ERROR, "Failed to write extra "
283 "buffer %d: %d.\n", i, err);
287 err = vaapi_encode_make_param_buffer(avctx, pic, type,
294 if (ctx->va_packed_headers & VA_ENC_PACKED_HEADER_MISC &&
295 ctx->codec->write_extra_header) {
298 bit_len = 8 * sizeof(data);
299 err = ctx->codec->write_extra_header(avctx, pic, i, &type,
301 if (err == AVERROR_EOF)
304 av_log(avctx, AV_LOG_ERROR, "Failed to write extra "
305 "header %d: %d.\n", i, err);
309 err = vaapi_encode_make_packed_header(avctx, pic, type,
316 av_assert0(pic->nb_slices <= MAX_PICTURE_SLICES);
317 for (i = 0; i < pic->nb_slices; i++) {
318 slice = av_mallocz(sizeof(*slice));
320 err = AVERROR(ENOMEM);
324 pic->slices[i] = slice;
326 if (ctx->codec->slice_params_size > 0) {
327 slice->codec_slice_params = av_mallocz(ctx->codec->slice_params_size);
328 if (!slice->codec_slice_params) {
329 err = AVERROR(ENOMEM);
334 if (ctx->codec->init_slice_params) {
335 err = ctx->codec->init_slice_params(avctx, pic, slice);
337 av_log(avctx, AV_LOG_ERROR, "Failed to initalise slice "
338 "parameters: %d.\n", err);
343 if (ctx->va_packed_headers & VA_ENC_PACKED_HEADER_SLICE &&
344 ctx->codec->write_slice_header) {
345 bit_len = 8 * sizeof(data);
346 err = ctx->codec->write_slice_header(avctx, pic, slice,
349 av_log(avctx, AV_LOG_ERROR, "Failed to write per-slice "
350 "header: %d.\n", err);
353 err = vaapi_encode_make_packed_header(avctx, pic,
354 ctx->codec->slice_header_type,
360 if (ctx->codec->init_slice_params) {
361 err = vaapi_encode_make_param_buffer(avctx, pic,
362 VAEncSliceParameterBufferType,
363 slice->codec_slice_params,
364 ctx->codec->slice_params_size);
370 vas = vaBeginPicture(ctx->hwctx->display, ctx->va_context,
372 if (vas != VA_STATUS_SUCCESS) {
373 av_log(avctx, AV_LOG_ERROR, "Failed to begin picture encode issue: "
374 "%d (%s).\n", vas, vaErrorStr(vas));
376 goto fail_with_picture;
379 vas = vaRenderPicture(ctx->hwctx->display, ctx->va_context,
380 pic->param_buffers, pic->nb_param_buffers);
381 if (vas != VA_STATUS_SUCCESS) {
382 av_log(avctx, AV_LOG_ERROR, "Failed to upload encode parameters: "
383 "%d (%s).\n", vas, vaErrorStr(vas));
385 goto fail_with_picture;
388 vas = vaEndPicture(ctx->hwctx->display, ctx->va_context);
389 if (vas != VA_STATUS_SUCCESS) {
390 av_log(avctx, AV_LOG_ERROR, "Failed to end picture encode issue: "
391 "%d (%s).\n", vas, vaErrorStr(vas));
393 // vaRenderPicture() has been called here, so we should not destroy
394 // the parameter buffers unless separate destruction is required.
395 if (ctx->hwctx->driver_quirks &
396 AV_VAAPI_DRIVER_QUIRK_RENDER_PARAM_BUFFERS)
402 if (ctx->hwctx->driver_quirks &
403 AV_VAAPI_DRIVER_QUIRK_RENDER_PARAM_BUFFERS) {
404 for (i = 0; i < pic->nb_param_buffers; i++) {
405 vas = vaDestroyBuffer(ctx->hwctx->display,
406 pic->param_buffers[i]);
407 if (vas != VA_STATUS_SUCCESS) {
408 av_log(avctx, AV_LOG_ERROR, "Failed to destroy "
409 "param buffer %#x: %d (%s).\n",
410 pic->param_buffers[i], vas, vaErrorStr(vas));
416 pic->encode_issued = 1;
418 if (ctx->issue_mode == ISSUE_MODE_SERIALISE_EVERYTHING)
419 return vaapi_encode_wait(avctx, pic);
424 vaEndPicture(ctx->hwctx->display, ctx->va_context);
426 for(i = 0; i < pic->nb_param_buffers; i++)
427 vaDestroyBuffer(ctx->hwctx->display, pic->param_buffers[i]);
429 av_freep(&pic->codec_picture_params);
430 av_frame_free(&pic->recon_image);
431 av_buffer_unref(&pic->output_buffer_ref);
432 pic->output_buffer = VA_INVALID_ID;
436 static int vaapi_encode_output(AVCodecContext *avctx,
437 VAAPIEncodePicture *pic, AVPacket *pkt)
439 VAAPIEncodeContext *ctx = avctx->priv_data;
440 VACodedBufferSegment *buf_list, *buf;
444 err = vaapi_encode_wait(avctx, pic);
449 vas = vaMapBuffer(ctx->hwctx->display, pic->output_buffer,
451 if (vas != VA_STATUS_SUCCESS) {
452 av_log(avctx, AV_LOG_ERROR, "Failed to map output buffers: "
453 "%d (%s).\n", vas, vaErrorStr(vas));
458 for (buf = buf_list; buf; buf = buf->next) {
459 av_log(avctx, AV_LOG_DEBUG, "Output buffer: %u bytes "
460 "(status %08x).\n", buf->size, buf->status);
462 err = av_new_packet(pkt, buf->size);
466 memcpy(pkt->data, buf->buf, buf->size);
469 if (pic->type == PICTURE_TYPE_IDR)
470 pkt->flags |= AV_PKT_FLAG_KEY;
474 vas = vaUnmapBuffer(ctx->hwctx->display, pic->output_buffer);
475 if (vas != VA_STATUS_SUCCESS) {
476 av_log(avctx, AV_LOG_ERROR, "Failed to unmap output buffers: "
477 "%d (%s).\n", vas, vaErrorStr(vas));
482 av_buffer_unref(&pic->output_buffer_ref);
483 pic->output_buffer = VA_INVALID_ID;
485 av_log(avctx, AV_LOG_DEBUG, "Output read for pic %"PRId64"/%"PRId64".\n",
486 pic->display_order, pic->encode_order);
490 vaUnmapBuffer(ctx->hwctx->display, pic->output_buffer);
492 av_buffer_unref(&pic->output_buffer_ref);
493 pic->output_buffer = VA_INVALID_ID;
497 static int vaapi_encode_discard(AVCodecContext *avctx,
498 VAAPIEncodePicture *pic)
500 vaapi_encode_wait(avctx, pic);
502 if (pic->output_buffer_ref) {
503 av_log(avctx, AV_LOG_DEBUG, "Discard output for pic "
504 "%"PRId64"/%"PRId64".\n",
505 pic->display_order, pic->encode_order);
507 av_buffer_unref(&pic->output_buffer_ref);
508 pic->output_buffer = VA_INVALID_ID;
514 static VAAPIEncodePicture *vaapi_encode_alloc(void)
516 VAAPIEncodePicture *pic;
518 pic = av_mallocz(sizeof(*pic));
522 pic->input_surface = VA_INVALID_ID;
523 pic->recon_surface = VA_INVALID_ID;
524 pic->output_buffer = VA_INVALID_ID;
529 static int vaapi_encode_free(AVCodecContext *avctx,
530 VAAPIEncodePicture *pic)
534 if (pic->encode_issued)
535 vaapi_encode_discard(avctx, pic);
537 for (i = 0; i < pic->nb_slices; i++) {
538 av_freep(&pic->slices[i]->priv_data);
539 av_freep(&pic->slices[i]->codec_slice_params);
540 av_freep(&pic->slices[i]);
542 av_freep(&pic->codec_picture_params);
544 av_frame_free(&pic->input_image);
545 av_frame_free(&pic->recon_image);
547 // Output buffer should already be destroyed.
548 av_assert0(pic->output_buffer == VA_INVALID_ID);
550 av_freep(&pic->priv_data);
551 av_freep(&pic->codec_picture_params);
558 static int vaapi_encode_step(AVCodecContext *avctx,
559 VAAPIEncodePicture *target)
561 VAAPIEncodeContext *ctx = avctx->priv_data;
562 VAAPIEncodePicture *pic;
565 if (ctx->issue_mode == ISSUE_MODE_SERIALISE_EVERYTHING ||
566 ctx->issue_mode == ISSUE_MODE_MINIMISE_LATENCY) {
567 // These two modes are equivalent, except that we wait for
568 // immediate completion on each operation if serialised.
571 // No target, nothing to do yet.
575 if (target->encode_complete) {
581 for (i = 0; i < pic->nb_refs; i++) {
582 if (!pic->refs[i]->encode_complete) {
583 err = vaapi_encode_step(avctx, pic->refs[i]);
589 err = vaapi_encode_issue(avctx, pic);
593 } else if (ctx->issue_mode == ISSUE_MODE_MAXIMISE_THROUGHPUT) {
596 // Run through the list of all available pictures repeatedly
597 // and issue the first one found which has all dependencies
598 // available (including previously-issued but not necessarily
599 // completed pictures).
602 for (pic = ctx->pic_start; pic; pic = pic->next) {
603 if (!pic->input_available || pic->encode_issued)
605 for (i = 0; i < pic->nb_refs; i++) {
606 if (!pic->refs[i]->encode_issued)
609 if (i < pic->nb_refs)
611 err = vaapi_encode_issue(avctx, pic);
615 // Start again from the beginning of the list,
616 // because issuing this picture may have satisfied
617 // forward dependencies of earlier ones.
622 // If we had a defined target for this step then it will
623 // always have been issued by now.
625 av_assert0(target->encode_issued && "broken dependencies?");
635 static int vaapi_encode_get_next(AVCodecContext *avctx,
636 VAAPIEncodePicture **pic_out)
638 VAAPIEncodeContext *ctx = avctx->priv_data;
639 VAAPIEncodePicture *start, *end, *pic;
642 for (pic = ctx->pic_start; pic; pic = pic->next) {
644 av_assert0(pic->display_order + 1 == pic->next->display_order);
645 if (pic->display_order == ctx->input_order) {
651 pic = vaapi_encode_alloc();
653 return AVERROR(ENOMEM);
655 if (ctx->input_order == 0 || ctx->force_idr ||
656 ctx->gop_counter >= avctx->gop_size) {
657 pic->type = PICTURE_TYPE_IDR;
659 ctx->gop_counter = 1;
661 } else if (ctx->p_counter >= ctx->p_per_i) {
662 pic->type = PICTURE_TYPE_I;
666 pic->type = PICTURE_TYPE_P;
667 pic->refs[0] = ctx->pic_end;
674 if (pic->type != PICTURE_TYPE_IDR) {
675 // If that was not an IDR frame, add B-frames display-before and
676 // encode-after it, but not exceeding the GOP size.
678 for (i = 0; i < ctx->b_per_p &&
679 ctx->gop_counter < avctx->gop_size; i++) {
680 pic = vaapi_encode_alloc();
684 pic->type = PICTURE_TYPE_B;
685 pic->refs[0] = ctx->pic_end;
690 pic->display_order = ctx->input_order + ctx->b_per_p - i - 1;
691 pic->encode_order = pic->display_order + 1;
698 if (ctx->input_order == 0) {
699 pic->display_order = 0;
700 pic->encode_order = 0;
702 ctx->pic_start = ctx->pic_end = pic;
705 for (i = 0, pic = start; pic; i++, pic = pic->next) {
706 pic->display_order = ctx->input_order + i;
707 if (end->type == PICTURE_TYPE_IDR)
708 pic->encode_order = ctx->input_order + i;
710 pic->encode_order = ctx->input_order;
712 pic->encode_order = ctx->input_order + i + 1;
715 av_assert0(ctx->pic_end);
716 ctx->pic_end->next = start;
721 av_log(avctx, AV_LOG_DEBUG, "Pictures:");
722 for (pic = ctx->pic_start; pic; pic = pic->next) {
723 av_log(avctx, AV_LOG_DEBUG, " %s (%"PRId64"/%"PRId64")",
724 picture_type_name[pic->type],
725 pic->display_order, pic->encode_order);
727 av_log(avctx, AV_LOG_DEBUG, "\n");
734 vaapi_encode_free(avctx, start);
737 return AVERROR(ENOMEM);
740 static int vaapi_encode_truncate_gop(AVCodecContext *avctx)
742 VAAPIEncodeContext *ctx = avctx->priv_data;
743 VAAPIEncodePicture *pic, *last_pic, *next;
745 // Find the last picture we actually have input for.
746 for (pic = ctx->pic_start; pic; pic = pic->next) {
747 if (!pic->input_available)
753 av_assert0(last_pic);
755 if (last_pic->type == PICTURE_TYPE_B) {
756 // Some fixing up is required. Change the type of this
757 // picture to P, then modify preceding B references which
758 // point beyond it to point at it instead.
760 last_pic->type = PICTURE_TYPE_P;
761 last_pic->encode_order = last_pic->refs[1]->encode_order;
763 for (pic = ctx->pic_start; pic != last_pic; pic = pic->next) {
764 if (pic->type == PICTURE_TYPE_B &&
765 pic->refs[1] == last_pic->refs[1])
766 pic->refs[1] = last_pic;
769 last_pic->nb_refs = 1;
770 last_pic->refs[1] = NULL;
772 // We can use the current structure (no references point
773 // beyond the end), but there are unused pics to discard.
776 // Discard all following pics, they will never be used.
777 for (pic = last_pic->next; pic; pic = next) {
779 vaapi_encode_free(avctx, pic);
782 last_pic->next = NULL;
783 ctx->pic_end = last_pic;
786 // Input is available for all pictures, so we don't need to
790 av_log(avctx, AV_LOG_DEBUG, "Pictures ending truncated GOP:");
791 for (pic = ctx->pic_start; pic; pic = pic->next) {
792 av_log(avctx, AV_LOG_DEBUG, " %s (%"PRId64"/%"PRId64")",
793 picture_type_name[pic->type],
794 pic->display_order, pic->encode_order);
796 av_log(avctx, AV_LOG_DEBUG, "\n");
801 static int vaapi_encode_clear_old(AVCodecContext *avctx)
803 VAAPIEncodeContext *ctx = avctx->priv_data;
804 VAAPIEncodePicture *pic, *old;
807 while (ctx->pic_start != ctx->pic_end) {
808 old = ctx->pic_start;
809 if (old->encode_order > ctx->output_order)
812 for (pic = old->next; pic; pic = pic->next) {
813 if (pic->encode_complete)
815 for (i = 0; i < pic->nb_refs; i++) {
816 if (pic->refs[i] == old) {
817 // We still need this picture because it's referred to
818 // directly by a later one, so it and all following
819 // pictures have to stay.
825 pic = ctx->pic_start;
826 ctx->pic_start = pic->next;
827 vaapi_encode_free(avctx, pic);
833 int ff_vaapi_encode2(AVCodecContext *avctx, AVPacket *pkt,
834 const AVFrame *input_image, int *got_packet)
836 VAAPIEncodeContext *ctx = avctx->priv_data;
837 VAAPIEncodePicture *pic;
841 av_log(avctx, AV_LOG_DEBUG, "Encode frame: %ux%u (%"PRId64").\n",
842 input_image->width, input_image->height, input_image->pts);
844 if (input_image->pict_type == AV_PICTURE_TYPE_I) {
845 err = vaapi_encode_truncate_gop(avctx);
851 err = vaapi_encode_get_next(avctx, &pic);
853 av_log(avctx, AV_LOG_ERROR, "Input setup failed: %d.\n", err);
857 pic->input_image = av_frame_alloc();
858 if (!pic->input_image) {
859 err = AVERROR(ENOMEM);
862 err = av_frame_ref(pic->input_image, input_image);
865 pic->input_surface = (VASurfaceID)(uintptr_t)input_image->data[3];
866 pic->pts = input_image->pts;
868 if (ctx->input_order == 0)
869 ctx->first_pts = pic->pts;
870 if (ctx->input_order == ctx->decode_delay)
871 ctx->dts_pts_diff = pic->pts - ctx->first_pts;
872 if (ctx->output_delay > 0)
873 ctx->ts_ring[ctx->input_order % (3 * ctx->output_delay)] = pic->pts;
875 pic->input_available = 1;
878 if (!ctx->end_of_stream) {
879 err = vaapi_encode_truncate_gop(avctx);
882 ctx->end_of_stream = 1;
888 av_assert0(ctx->output_order + ctx->output_delay + 1 == ctx->input_order);
890 for (pic = ctx->pic_start; pic; pic = pic->next)
891 if (pic->encode_order == ctx->output_order)
894 // pic can be null here if we don't have a specific target in this
895 // iteration. We might still issue encodes if things can be overlapped,
896 // even though we don't intend to output anything.
898 err = vaapi_encode_step(avctx, pic);
900 av_log(avctx, AV_LOG_ERROR, "Encode failed: %d.\n", err);
907 err = vaapi_encode_output(avctx, pic, pkt);
909 av_log(avctx, AV_LOG_ERROR, "Output failed: %d.\n", err);
913 if (ctx->output_delay == 0) {
915 } else if (ctx->output_order < ctx->decode_delay) {
916 if (ctx->ts_ring[ctx->output_order] < INT64_MIN + ctx->dts_pts_diff)
917 pkt->dts = INT64_MIN;
919 pkt->dts = ctx->ts_ring[ctx->output_order] - ctx->dts_pts_diff;
921 pkt->dts = ctx->ts_ring[(ctx->output_order - ctx->decode_delay) %
922 (3 * ctx->output_delay)];
928 err = vaapi_encode_clear_old(avctx);
930 av_log(avctx, AV_LOG_ERROR, "List clearing failed: %d.\n", err);
937 // Unclear what to clean up on failure. There are probably some things we
938 // could do usefully clean up here, but for now just leave them for uninit()
943 static av_cold int vaapi_encode_config_attributes(AVCodecContext *avctx)
945 VAAPIEncodeContext *ctx = avctx->priv_data;
948 VAProfile *profiles = NULL;
949 VAEntrypoint *entrypoints = NULL;
950 VAConfigAttrib attr[] = {
951 { VAConfigAttribRTFormat },
952 { VAConfigAttribRateControl },
953 { VAConfigAttribEncMaxRefFrames },
954 { VAConfigAttribEncPackedHeaders },
957 n = vaMaxNumProfiles(ctx->hwctx->display);
958 profiles = av_malloc_array(n, sizeof(VAProfile));
960 err = AVERROR(ENOMEM);
963 vas = vaQueryConfigProfiles(ctx->hwctx->display, profiles, &n);
964 if (vas != VA_STATUS_SUCCESS) {
965 av_log(ctx, AV_LOG_ERROR, "Failed to query profiles: %d (%s).\n",
966 vas, vaErrorStr(vas));
967 err = AVERROR(ENOSYS);
970 for (i = 0; i < n; i++) {
971 if (profiles[i] == ctx->va_profile)
975 av_log(ctx, AV_LOG_ERROR, "Encoding profile not found (%d).\n",
977 err = AVERROR(ENOSYS);
981 n = vaMaxNumEntrypoints(ctx->hwctx->display);
982 entrypoints = av_malloc_array(n, sizeof(VAEntrypoint));
984 err = AVERROR(ENOMEM);
987 vas = vaQueryConfigEntrypoints(ctx->hwctx->display, ctx->va_profile,
989 if (vas != VA_STATUS_SUCCESS) {
990 av_log(ctx, AV_LOG_ERROR, "Failed to query entrypoints for "
991 "profile %u: %d (%s).\n", ctx->va_profile,
992 vas, vaErrorStr(vas));
993 err = AVERROR(ENOSYS);
996 for (i = 0; i < n; i++) {
997 if (entrypoints[i] == ctx->va_entrypoint)
1001 av_log(ctx, AV_LOG_ERROR, "Encoding entrypoint not found "
1002 "(%d / %d).\n", ctx->va_profile, ctx->va_entrypoint);
1003 err = AVERROR(ENOSYS);
1007 vas = vaGetConfigAttributes(ctx->hwctx->display,
1008 ctx->va_profile, ctx->va_entrypoint,
1009 attr, FF_ARRAY_ELEMS(attr));
1010 if (vas != VA_STATUS_SUCCESS) {
1011 av_log(avctx, AV_LOG_ERROR, "Failed to fetch config "
1012 "attributes: %d (%s).\n", vas, vaErrorStr(vas));
1013 return AVERROR(EINVAL);
1016 for (i = 0; i < FF_ARRAY_ELEMS(attr); i++) {
1017 if (attr[i].value == VA_ATTRIB_NOT_SUPPORTED) {
1018 // Unfortunately we have to treat this as "don't know" and hope
1019 // for the best, because the Intel MJPEG encoder returns this
1020 // for all the interesting attributes.
1023 switch (attr[i].type) {
1024 case VAConfigAttribRTFormat:
1025 if (!(ctx->va_rt_format & attr[i].value)) {
1026 av_log(avctx, AV_LOG_ERROR, "Surface RT format %#x "
1027 "is not supported (mask %#x).\n",
1028 ctx->va_rt_format, attr[i].value);
1029 err = AVERROR(EINVAL);
1032 ctx->config_attributes[ctx->nb_config_attributes++] =
1034 .type = VAConfigAttribRTFormat,
1035 .value = ctx->va_rt_format,
1038 case VAConfigAttribRateControl:
1039 // Hack for backward compatibility: CBR was the only
1040 // usable RC mode for a long time, so old drivers will
1041 // only have it. Normal default options may now choose
1042 // VBR and then fail, however, so override it here with
1043 // CBR if that is the only supported mode.
1044 if (ctx->va_rc_mode == VA_RC_VBR &&
1045 !(attr[i].value & VA_RC_VBR) &&
1046 (attr[i].value & VA_RC_CBR)) {
1047 av_log(avctx, AV_LOG_WARNING, "VBR rate control is "
1048 "not supported with this driver version; "
1049 "using CBR instead.\n");
1050 ctx->va_rc_mode = VA_RC_CBR;
1052 if (!(ctx->va_rc_mode & attr[i].value)) {
1053 av_log(avctx, AV_LOG_ERROR, "Rate control mode %#x "
1054 "is not supported (mask: %#x).\n",
1055 ctx->va_rc_mode, attr[i].value);
1056 err = AVERROR(EINVAL);
1059 ctx->config_attributes[ctx->nb_config_attributes++] =
1061 .type = VAConfigAttribRateControl,
1062 .value = ctx->va_rc_mode,
1065 case VAConfigAttribEncMaxRefFrames:
1067 unsigned int ref_l0 = attr[i].value & 0xffff;
1068 unsigned int ref_l1 = (attr[i].value >> 16) & 0xffff;
1070 if (avctx->gop_size > 1 && ref_l0 < 1) {
1071 av_log(avctx, AV_LOG_ERROR, "P frames are not "
1072 "supported (%#x).\n", attr[i].value);
1073 err = AVERROR(EINVAL);
1076 if (avctx->max_b_frames > 0 && ref_l1 < 1) {
1077 av_log(avctx, AV_LOG_ERROR, "B frames are not "
1078 "supported (%#x).\n", attr[i].value);
1079 err = AVERROR(EINVAL);
1084 case VAConfigAttribEncPackedHeaders:
1085 if (ctx->va_packed_headers & ~attr[i].value) {
1086 // This isn't fatal, but packed headers are always
1087 // preferable because they are under our control.
1088 // When absent, the driver is generating them and some
1089 // features may not work (e.g. VUI or SEI in H.264).
1090 av_log(avctx, AV_LOG_WARNING, "Warning: some packed "
1091 "headers are not supported (want %#x, got %#x).\n",
1092 ctx->va_packed_headers, attr[i].value);
1093 ctx->va_packed_headers &= attr[i].value;
1095 ctx->config_attributes[ctx->nb_config_attributes++] =
1097 .type = VAConfigAttribEncPackedHeaders,
1098 .value = ctx->va_packed_headers,
1102 av_assert0(0 && "Unexpected config attribute.");
1108 av_freep(&profiles);
1109 av_freep(&entrypoints);
1113 static av_cold int vaapi_encode_init_rate_control(AVCodecContext *avctx)
1115 VAAPIEncodeContext *ctx = avctx->priv_data;
1116 int rc_bits_per_second;
1117 int rc_target_percentage;
1119 int hrd_buffer_size;
1120 int hrd_initial_buffer_fullness;
1123 if (avctx->rc_buffer_size)
1124 hrd_buffer_size = avctx->rc_buffer_size;
1126 hrd_buffer_size = avctx->bit_rate;
1127 if (avctx->rc_initial_buffer_occupancy)
1128 hrd_initial_buffer_fullness = avctx->rc_initial_buffer_occupancy;
1130 hrd_initial_buffer_fullness = hrd_buffer_size * 3 / 4;
1132 if (ctx->va_rc_mode == VA_RC_CBR) {
1133 rc_bits_per_second = avctx->bit_rate;
1134 rc_target_percentage = 100;
1135 rc_window_size = 1000;
1137 if (avctx->rc_max_rate < avctx->bit_rate) {
1138 // Max rate is unset or invalid, just use the normal bitrate.
1139 rc_bits_per_second = avctx->bit_rate;
1140 rc_target_percentage = 100;
1142 rc_bits_per_second = avctx->rc_max_rate;
1143 rc_target_percentage = (avctx->bit_rate * 100) / rc_bits_per_second;
1145 rc_window_size = (hrd_buffer_size * 1000) / avctx->bit_rate;
1148 ctx->rc_params.misc.type = VAEncMiscParameterTypeRateControl;
1149 ctx->rc_params.rc = (VAEncMiscParameterRateControl) {
1150 .bits_per_second = rc_bits_per_second,
1151 .target_percentage = rc_target_percentage,
1152 .window_size = rc_window_size,
1154 .min_qp = (avctx->qmin > 0 ? avctx->qmin : 0),
1155 .basic_unit_size = 0,
1157 ctx->global_params[ctx->nb_global_params] =
1158 &ctx->rc_params.misc;
1159 ctx->global_params_size[ctx->nb_global_params++] =
1160 sizeof(ctx->rc_params);
1162 ctx->hrd_params.misc.type = VAEncMiscParameterTypeHRD;
1163 ctx->hrd_params.hrd = (VAEncMiscParameterHRD) {
1164 .initial_buffer_fullness = hrd_initial_buffer_fullness,
1165 .buffer_size = hrd_buffer_size,
1167 ctx->global_params[ctx->nb_global_params] =
1168 &ctx->hrd_params.misc;
1169 ctx->global_params_size[ctx->nb_global_params++] =
1170 sizeof(ctx->hrd_params);
1172 if (avctx->framerate.num > 0 && avctx->framerate.den > 0)
1173 av_reduce(&fr_num, &fr_den,
1174 avctx->framerate.num, avctx->framerate.den, 65535);
1176 av_reduce(&fr_num, &fr_den,
1177 avctx->time_base.den, avctx->time_base.num, 65535);
1179 ctx->fr_params.misc.type = VAEncMiscParameterTypeFrameRate;
1180 ctx->fr_params.fr.framerate = (unsigned int)fr_den << 16 | fr_num;
1182 #if VA_CHECK_VERSION(0, 40, 0)
1183 ctx->global_params[ctx->nb_global_params] =
1184 &ctx->fr_params.misc;
1185 ctx->global_params_size[ctx->nb_global_params++] =
1186 sizeof(ctx->fr_params);
1192 static void vaapi_encode_free_output_buffer(void *opaque,
1195 AVCodecContext *avctx = opaque;
1196 VAAPIEncodeContext *ctx = avctx->priv_data;
1197 VABufferID buffer_id;
1199 buffer_id = (VABufferID)(uintptr_t)data;
1201 vaDestroyBuffer(ctx->hwctx->display, buffer_id);
1203 av_log(avctx, AV_LOG_DEBUG, "Freed output buffer %#x\n", buffer_id);
1206 static AVBufferRef *vaapi_encode_alloc_output_buffer(void *opaque,
1209 AVCodecContext *avctx = opaque;
1210 VAAPIEncodeContext *ctx = avctx->priv_data;
1211 VABufferID buffer_id;
1215 // The output buffer size is fixed, so it needs to be large enough
1216 // to hold the largest possible compressed frame. We assume here
1217 // that the uncompressed frame plus some header data is an upper
1219 vas = vaCreateBuffer(ctx->hwctx->display, ctx->va_context,
1220 VAEncCodedBufferType,
1221 3 * ctx->surface_width * ctx->surface_height +
1222 (1 << 16), 1, 0, &buffer_id);
1223 if (vas != VA_STATUS_SUCCESS) {
1224 av_log(avctx, AV_LOG_ERROR, "Failed to create bitstream "
1225 "output buffer: %d (%s).\n", vas, vaErrorStr(vas));
1229 av_log(avctx, AV_LOG_DEBUG, "Allocated output buffer %#x\n", buffer_id);
1231 ref = av_buffer_create((uint8_t*)(uintptr_t)buffer_id,
1233 &vaapi_encode_free_output_buffer,
1234 avctx, AV_BUFFER_FLAG_READONLY);
1236 vaDestroyBuffer(ctx->hwctx->display, buffer_id);
1243 static av_cold int vaapi_encode_create_recon_frames(AVCodecContext *avctx)
1245 VAAPIEncodeContext *ctx = avctx->priv_data;
1246 AVVAAPIHWConfig *hwconfig = NULL;
1247 AVHWFramesConstraints *constraints = NULL;
1248 enum AVPixelFormat recon_format;
1251 hwconfig = av_hwdevice_hwconfig_alloc(ctx->device_ref);
1253 err = AVERROR(ENOMEM);
1256 hwconfig->config_id = ctx->va_config;
1258 constraints = av_hwdevice_get_hwframe_constraints(ctx->device_ref,
1261 err = AVERROR(ENOMEM);
1265 // Probably we can use the input surface format as the surface format
1266 // of the reconstructed frames. If not, we just pick the first (only?)
1267 // format in the valid list and hope that it all works.
1268 recon_format = AV_PIX_FMT_NONE;
1269 if (constraints->valid_sw_formats) {
1270 for (i = 0; constraints->valid_sw_formats[i] != AV_PIX_FMT_NONE; i++) {
1271 if (ctx->input_frames->sw_format ==
1272 constraints->valid_sw_formats[i]) {
1273 recon_format = ctx->input_frames->sw_format;
1277 if (recon_format == AV_PIX_FMT_NONE) {
1278 // No match. Just use the first in the supported list and
1279 // hope for the best.
1280 recon_format = constraints->valid_sw_formats[0];
1283 // No idea what to use; copy input format.
1284 recon_format = ctx->input_frames->sw_format;
1286 av_log(avctx, AV_LOG_DEBUG, "Using %s as format of "
1287 "reconstructed frames.\n", av_get_pix_fmt_name(recon_format));
1289 if (ctx->surface_width < constraints->min_width ||
1290 ctx->surface_height < constraints->min_height ||
1291 ctx->surface_width > constraints->max_width ||
1292 ctx->surface_height > constraints->max_height) {
1293 av_log(avctx, AV_LOG_ERROR, "Hardware does not support encoding at "
1294 "size %dx%d (constraints: width %d-%d height %d-%d).\n",
1295 ctx->surface_width, ctx->surface_height,
1296 constraints->min_width, constraints->max_width,
1297 constraints->min_height, constraints->max_height);
1298 err = AVERROR(EINVAL);
1302 av_freep(&hwconfig);
1303 av_hwframe_constraints_free(&constraints);
1305 ctx->recon_frames_ref = av_hwframe_ctx_alloc(ctx->device_ref);
1306 if (!ctx->recon_frames_ref) {
1307 err = AVERROR(ENOMEM);
1310 ctx->recon_frames = (AVHWFramesContext*)ctx->recon_frames_ref->data;
1312 ctx->recon_frames->format = AV_PIX_FMT_VAAPI;
1313 ctx->recon_frames->sw_format = recon_format;
1314 ctx->recon_frames->width = ctx->surface_width;
1315 ctx->recon_frames->height = ctx->surface_height;
1316 // At most three IDR/I/P frames and two runs of B frames can be in
1317 // flight at any one time.
1318 ctx->recon_frames->initial_pool_size = 3 + 2 * avctx->max_b_frames;
1320 err = av_hwframe_ctx_init(ctx->recon_frames_ref);
1322 av_log(avctx, AV_LOG_ERROR, "Failed to initialise reconstructed "
1323 "frame context: %d.\n", err);
1329 av_freep(&hwconfig);
1330 av_hwframe_constraints_free(&constraints);
1334 av_cold int ff_vaapi_encode_init(AVCodecContext *avctx)
1336 VAAPIEncodeContext *ctx = avctx->priv_data;
1337 AVVAAPIFramesContext *recon_hwctx = NULL;
1341 if (!avctx->hw_frames_ctx) {
1342 av_log(avctx, AV_LOG_ERROR, "A hardware frames reference is "
1343 "required to associate the encoding device.\n");
1344 return AVERROR(EINVAL);
1347 ctx->codec_options = ctx->codec_options_data;
1349 ctx->va_config = VA_INVALID_ID;
1350 ctx->va_context = VA_INVALID_ID;
1352 ctx->priv_data = av_mallocz(ctx->codec->priv_data_size);
1353 if (!ctx->priv_data) {
1354 err = AVERROR(ENOMEM);
1358 ctx->input_frames_ref = av_buffer_ref(avctx->hw_frames_ctx);
1359 if (!ctx->input_frames_ref) {
1360 err = AVERROR(ENOMEM);
1363 ctx->input_frames = (AVHWFramesContext*)ctx->input_frames_ref->data;
1365 ctx->device_ref = av_buffer_ref(ctx->input_frames->device_ref);
1366 if (!ctx->device_ref) {
1367 err = AVERROR(ENOMEM);
1370 ctx->device = (AVHWDeviceContext*)ctx->device_ref->data;
1371 ctx->hwctx = ctx->device->hwctx;
1373 err = vaapi_encode_config_attributes(avctx);
1377 vas = vaCreateConfig(ctx->hwctx->display,
1378 ctx->va_profile, ctx->va_entrypoint,
1379 ctx->config_attributes, ctx->nb_config_attributes,
1381 if (vas != VA_STATUS_SUCCESS) {
1382 av_log(avctx, AV_LOG_ERROR, "Failed to create encode pipeline "
1383 "configuration: %d (%s).\n", vas, vaErrorStr(vas));
1388 err = vaapi_encode_create_recon_frames(avctx);
1392 recon_hwctx = ctx->recon_frames->hwctx;
1393 vas = vaCreateContext(ctx->hwctx->display, ctx->va_config,
1394 ctx->surface_width, ctx->surface_height,
1396 recon_hwctx->surface_ids,
1397 recon_hwctx->nb_surfaces,
1399 if (vas != VA_STATUS_SUCCESS) {
1400 av_log(avctx, AV_LOG_ERROR, "Failed to create encode pipeline "
1401 "context: %d (%s).\n", vas, vaErrorStr(vas));
1406 ctx->output_buffer_pool =
1407 av_buffer_pool_init2(sizeof(VABufferID), avctx,
1408 &vaapi_encode_alloc_output_buffer, NULL);
1409 if (!ctx->output_buffer_pool) {
1410 err = AVERROR(ENOMEM);
1414 if (ctx->va_rc_mode & ~VA_RC_CQP) {
1415 err = vaapi_encode_init_rate_control(avctx);
1420 if (ctx->codec->configure) {
1421 err = ctx->codec->configure(avctx);
1426 ctx->input_order = 0;
1427 ctx->output_delay = avctx->max_b_frames;
1428 ctx->decode_delay = 1;
1429 ctx->output_order = - ctx->output_delay - 1;
1431 // Currently we never generate I frames, only IDR.
1432 ctx->p_per_i = ((avctx->gop_size + avctx->max_b_frames) /
1433 (avctx->max_b_frames + 1));
1434 ctx->b_per_p = avctx->max_b_frames;
1436 if (ctx->codec->sequence_params_size > 0) {
1437 ctx->codec_sequence_params =
1438 av_mallocz(ctx->codec->sequence_params_size);
1439 if (!ctx->codec_sequence_params) {
1440 err = AVERROR(ENOMEM);
1444 if (ctx->codec->picture_params_size > 0) {
1445 ctx->codec_picture_params =
1446 av_mallocz(ctx->codec->picture_params_size);
1447 if (!ctx->codec_picture_params) {
1448 err = AVERROR(ENOMEM);
1453 if (ctx->codec->init_sequence_params) {
1454 err = ctx->codec->init_sequence_params(avctx);
1456 av_log(avctx, AV_LOG_ERROR, "Codec sequence initialisation "
1457 "failed: %d.\n", err);
1462 // This should be configurable somehow. (Needs testing on a machine
1463 // where it actually overlaps properly, though.)
1464 ctx->issue_mode = ISSUE_MODE_MAXIMISE_THROUGHPUT;
1466 if (ctx->va_packed_headers & VA_ENC_PACKED_HEADER_SEQUENCE &&
1467 ctx->codec->write_sequence_header) {
1468 char data[MAX_PARAM_BUFFER_SIZE];
1469 size_t bit_len = 8 * sizeof(data);
1471 err = ctx->codec->write_sequence_header(avctx, data, &bit_len);
1473 av_log(avctx, AV_LOG_ERROR, "Failed to write sequence header "
1474 "for extradata: %d.\n", err);
1477 avctx->extradata_size = (bit_len + 7) / 8;
1478 avctx->extradata = av_mallocz(avctx->extradata_size +
1479 AV_INPUT_BUFFER_PADDING_SIZE);
1480 if (!avctx->extradata) {
1481 err = AVERROR(ENOMEM);
1484 memcpy(avctx->extradata, data, avctx->extradata_size);
1491 ff_vaapi_encode_close(avctx);
1495 av_cold int ff_vaapi_encode_close(AVCodecContext *avctx)
1497 VAAPIEncodeContext *ctx = avctx->priv_data;
1498 VAAPIEncodePicture *pic, *next;
1500 for (pic = ctx->pic_start; pic; pic = next) {
1502 vaapi_encode_free(avctx, pic);
1505 if (ctx->va_context != VA_INVALID_ID) {
1506 vaDestroyContext(ctx->hwctx->display, ctx->va_context);
1507 ctx->va_context = VA_INVALID_ID;
1510 if (ctx->va_config != VA_INVALID_ID) {
1511 vaDestroyConfig(ctx->hwctx->display, ctx->va_config);
1512 ctx->va_config = VA_INVALID_ID;
1515 av_buffer_pool_uninit(&ctx->output_buffer_pool);
1517 av_freep(&ctx->codec_sequence_params);
1518 av_freep(&ctx->codec_picture_params);
1520 av_buffer_unref(&ctx->recon_frames_ref);
1521 av_buffer_unref(&ctx->input_frames_ref);
1522 av_buffer_unref(&ctx->device_ref);
1524 av_freep(&ctx->priv_data);