3 * Copyright (C) 2015 Luca Barbato
5 * This file is part of Libav.
7 * Libav is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * Libav is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with Libav; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 #include <nvEncodeAPI.h>
28 #define CUDA_LIBNAME "libcuda.so"
33 #define NVENC_LIBNAME "libnvidia-encode.so"
39 #define NVENC_LIBNAME "nvEncodeAPI64.dll"
41 #define NVENC_LIBNAME "nvEncodeAPI.dll"
44 #define dlopen(filename, flags) LoadLibrary((filename))
45 #define dlsym(handle, symbol) GetProcAddress(handle, symbol)
46 #define dlclose(handle) FreeLibrary(handle)
49 #include "libavutil/common.h"
50 #include "libavutil/imgutils.h"
51 #include "libavutil/mem.h"
56 #define NVENC_CAP 0x30
57 #define BITSTREAM_BUFFER_SIZE 1024 * 1024
59 #define LOAD_LIBRARY(l, path) \
61 if (!((l) = dlopen(path, RTLD_LAZY))) { \
62 av_log(avctx, AV_LOG_ERROR, \
65 return AVERROR_UNKNOWN; \
69 #define LOAD_SYMBOL(fun, lib, symbol) \
71 if (!((fun) = dlsym(lib, symbol))) { \
72 av_log(avctx, AV_LOG_ERROR, \
75 return AVERROR_UNKNOWN; \
79 static av_cold int nvenc_load_libraries(AVCodecContext *avctx)
81 NVENCContext *ctx = avctx->priv_data;
82 NVENCLibraryContext *nvel = &ctx->nvel;
83 PNVENCODEAPICREATEINSTANCE nvenc_create_instance;
85 LOAD_LIBRARY(nvel->cuda, CUDA_LIBNAME);
87 LOAD_SYMBOL(nvel->cu_init, nvel->cuda, "cuInit");
88 LOAD_SYMBOL(nvel->cu_device_get_count, nvel->cuda, "cuDeviceGetCount");
89 LOAD_SYMBOL(nvel->cu_device_get, nvel->cuda, "cuDeviceGet");
90 LOAD_SYMBOL(nvel->cu_device_get_name, nvel->cuda, "cuDeviceGetName");
91 LOAD_SYMBOL(nvel->cu_device_compute_capability, nvel->cuda,
92 "cuDeviceComputeCapability");
93 LOAD_SYMBOL(nvel->cu_ctx_create, nvel->cuda, "cuCtxCreate_v2");
94 LOAD_SYMBOL(nvel->cu_ctx_pop_current, nvel->cuda, "cuCtxPopCurrent_v2");
95 LOAD_SYMBOL(nvel->cu_ctx_destroy, nvel->cuda, "cuCtxDestroy_v2");
97 LOAD_LIBRARY(nvel->nvenc, NVENC_LIBNAME);
99 LOAD_SYMBOL(nvenc_create_instance, nvel->nvenc,
100 "NvEncodeAPICreateInstance");
102 nvel->nvenc_funcs.version = NV_ENCODE_API_FUNCTION_LIST_VER;
104 if ((nvenc_create_instance(&nvel->nvenc_funcs)) != NV_ENC_SUCCESS) {
105 av_log(avctx, AV_LOG_ERROR, "Cannot create the NVENC instance");
106 return AVERROR_UNKNOWN;
112 static int nvenc_open_session(AVCodecContext *avctx)
114 NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS params = { 0 };
115 NVENCContext *ctx = avctx->priv_data;
116 NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
119 params.version = NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER;
120 params.apiVersion = NVENCAPI_VERSION;
121 params.device = ctx->cu_context;
122 params.deviceType = NV_ENC_DEVICE_TYPE_CUDA;
124 ret = nv->nvEncOpenEncodeSessionEx(¶ms, &ctx->nvenc_ctx);
125 if (ret != NV_ENC_SUCCESS) {
126 ctx->nvenc_ctx = NULL;
127 av_log(avctx, AV_LOG_ERROR,
128 "Cannot open the NVENC Session\n");
129 return AVERROR_UNKNOWN;
135 static int nvenc_check_codec_support(AVCodecContext *avctx)
137 NVENCContext *ctx = avctx->priv_data;
138 NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
139 int i, ret, count = 0;
142 ret = nv->nvEncGetEncodeGUIDCount(ctx->nvenc_ctx, &count);
144 if (ret != NV_ENC_SUCCESS || !count)
145 return AVERROR(ENOSYS);
147 guids = av_malloc(count * sizeof(GUID));
149 return AVERROR(ENOMEM);
151 ret = nv->nvEncGetEncodeGUIDs(ctx->nvenc_ctx, guids, count, &count);
152 if (ret != NV_ENC_SUCCESS) {
153 ret = AVERROR(ENOSYS);
157 ret = AVERROR(ENOSYS);
158 for (i = 0; i < count; i++) {
159 if (!memcmp(&guids[i], &ctx->params.encodeGUID, sizeof(*guids))) {
171 static int nvenc_check_cap(AVCodecContext *avctx, NV_ENC_CAPS cap)
173 NVENCContext *ctx = avctx->priv_data;
174 NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
175 NV_ENC_CAPS_PARAM params = { 0 };
178 params.version = NV_ENC_CAPS_PARAM_VER;
179 params.capsToQuery = cap;
181 ret = nv->nvEncGetEncodeCaps(ctx->nvenc_ctx, ctx->params.encodeGUID, ¶ms, &val);
183 if (ret == NV_ENC_SUCCESS)
188 static int nvenc_check_capabilities(AVCodecContext *avctx)
192 ret = nvenc_check_codec_support(avctx);
194 av_log(avctx, AV_LOG_VERBOSE, "Codec not supported\n");
198 ret = nvenc_check_cap(avctx, NV_ENC_CAPS_SUPPORT_YUV444_ENCODE);
199 if (avctx->pix_fmt == AV_PIX_FMT_YUV444P && ret <= 0) {
200 av_log(avctx, AV_LOG_VERBOSE, "YUV444P not supported\n");
201 return AVERROR(ENOSYS);
204 ret = nvenc_check_cap(avctx, NV_ENC_CAPS_WIDTH_MAX);
205 if (ret < avctx->width) {
206 av_log(avctx, AV_LOG_VERBOSE, "Width %d exceeds %d\n",
208 return AVERROR(ENOSYS);
211 ret = nvenc_check_cap(avctx, NV_ENC_CAPS_HEIGHT_MAX);
212 if (ret < avctx->height) {
213 av_log(avctx, AV_LOG_VERBOSE, "Height %d exceeds %d\n",
215 return AVERROR(ENOSYS);
218 ret = nvenc_check_cap(avctx, NV_ENC_CAPS_NUM_MAX_BFRAMES);
219 if (ret < avctx->max_b_frames) {
220 av_log(avctx, AV_LOG_VERBOSE, "Max b-frames %d exceed %d\n",
221 avctx->max_b_frames, ret);
223 return AVERROR(ENOSYS);
229 static int nvenc_check_device(AVCodecContext *avctx, int idx)
231 NVENCContext *ctx = avctx->priv_data;
232 NVENCLibraryContext *nvel = &ctx->nvel;
233 char name[128] = { 0 };
234 int major, minor, ret;
237 int loglevel = AV_LOG_VERBOSE;
239 if (ctx->device == LIST_DEVICES)
240 loglevel = AV_LOG_INFO;
242 ret = nvel->cu_device_get(&cu_device, idx);
243 if (ret != CUDA_SUCCESS) {
244 av_log(avctx, AV_LOG_ERROR,
245 "Cannot access the CUDA device %d\n",
250 ret = nvel->cu_device_get_name(name, sizeof(name), cu_device);
251 if (ret != CUDA_SUCCESS)
254 ret = nvel->cu_device_compute_capability(&major, &minor, cu_device);
255 if (ret != CUDA_SUCCESS)
258 av_log(avctx, loglevel, "Device %d [%s] ", cu_device, name);
260 if (((major << 4) | minor) < NVENC_CAP)
263 ret = nvel->cu_ctx_create(&ctx->cu_context, 0, cu_device);
264 if (ret != CUDA_SUCCESS)
267 ret = nvel->cu_ctx_pop_current(&dummy);
268 if (ret != CUDA_SUCCESS)
271 if ((ret = nvenc_open_session(avctx)) < 0)
274 if ((ret = nvenc_check_capabilities(avctx)) < 0)
277 av_log(avctx, loglevel, "supports NVENC\n");
279 if (ctx->device == cu_device || ctx->device == ANY_DEVICE)
283 nvel->nvenc_funcs.nvEncDestroyEncoder(ctx->nvenc_ctx);
284 ctx->nvenc_ctx = NULL;
287 nvel->cu_ctx_destroy(ctx->cu_context);
288 ctx->cu_context = NULL;
292 av_log(avctx, loglevel, "does not support NVENC (major %d minor %d)\n",
295 return AVERROR(ENOSYS);
298 static int nvenc_setup_device(AVCodecContext *avctx)
300 NVENCContext *ctx = avctx->priv_data;
301 NVENCLibraryContext *nvel = &ctx->nvel;
302 int i, nb_devices = 0;
304 if ((nvel->cu_init(0)) != CUDA_SUCCESS) {
305 av_log(avctx, AV_LOG_ERROR,
306 "Cannot init CUDA\n");
307 return AVERROR_UNKNOWN;
310 if ((nvel->cu_device_get_count(&nb_devices)) != CUDA_SUCCESS) {
311 av_log(avctx, AV_LOG_ERROR,
312 "Cannot enumerate the CUDA devices\n");
313 return AVERROR_UNKNOWN;
316 switch (avctx->codec->id) {
317 case AV_CODEC_ID_H264:
318 ctx->params.encodeGUID = NV_ENC_CODEC_H264_GUID;
320 case AV_CODEC_ID_HEVC:
321 ctx->params.encodeGUID = NV_ENC_CODEC_HEVC_GUID;
327 for (i = 0; i < nb_devices; ++i) {
328 if ((nvenc_check_device(avctx, i)) >= 0 && ctx->device != LIST_DEVICES)
332 if (ctx->device == LIST_DEVICES)
335 return AVERROR(ENOSYS);
338 typedef struct GUIDTuple {
343 static int nvec_map_preset(NVENCContext *ctx)
345 GUIDTuple presets[] = {
346 { NV_ENC_PRESET_DEFAULT_GUID },
347 { NV_ENC_PRESET_HP_GUID },
348 { NV_ENC_PRESET_HQ_GUID },
349 { NV_ENC_PRESET_BD_GUID },
350 { NV_ENC_PRESET_LOW_LATENCY_DEFAULT_GUID, NVENC_LOWLATENCY },
351 { NV_ENC_PRESET_LOW_LATENCY_HP_GUID, NVENC_LOWLATENCY },
352 { NV_ENC_PRESET_LOW_LATENCY_HQ_GUID, NVENC_LOWLATENCY },
353 { NV_ENC_PRESET_LOSSLESS_DEFAULT_GUID, NVENC_LOSSLESS },
354 { NV_ENC_PRESET_LOSSLESS_HP_GUID, NVENC_LOSSLESS },
358 GUIDTuple *t = &presets[ctx->preset];
360 ctx->params.presetGUID = t->guid;
361 ctx->flags = t->flags;
363 return AVERROR(EINVAL);
366 static void set_constqp(AVCodecContext *avctx, NV_ENC_RC_PARAMS *rc)
368 rc->rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
369 rc->constQP.qpInterB = avctx->global_quality;
370 rc->constQP.qpInterP = avctx->global_quality;
371 rc->constQP.qpIntra = avctx->global_quality;
374 static void set_vbr(AVCodecContext *avctx, NV_ENC_RC_PARAMS *rc)
376 if (avctx->qmin >= 0) {
378 rc->minQP.qpInterB = avctx->qmin;
379 rc->minQP.qpInterP = avctx->qmin;
380 rc->minQP.qpIntra = avctx->qmin;
383 if (avctx->qmax >= 0) {
385 rc->maxQP.qpInterB = avctx->qmax;
386 rc->maxQP.qpInterP = avctx->qmax;
387 rc->maxQP.qpIntra = avctx->qmax;
391 static void nvenc_override_rate_control(AVCodecContext *avctx,
392 NV_ENC_RC_PARAMS *rc)
394 NVENCContext *ctx = avctx->priv_data;
397 case NV_ENC_PARAMS_RC_CONSTQP:
398 if (avctx->global_quality < 0) {
399 av_log(avctx, AV_LOG_WARNING,
400 "The constant quality rate-control requires "
401 "the 'global_quality' option set.\n");
404 set_constqp(avctx, rc);
406 case NV_ENC_PARAMS_RC_2_PASS_VBR:
407 case NV_ENC_PARAMS_RC_VBR:
408 if (avctx->qmin < 0 && avctx->qmax < 0) {
409 av_log(avctx, AV_LOG_WARNING,
410 "The variable bitrate rate-control requires "
411 "the 'qmin' and/or 'qmax' option set.\n");
414 case NV_ENC_PARAMS_RC_VBR_MINQP:
415 if (avctx->qmin < 0) {
416 av_log(avctx, AV_LOG_WARNING,
417 "The variable bitrate rate-control requires "
418 "the 'qmin' option set.\n");
423 case NV_ENC_PARAMS_RC_CBR:
425 case NV_ENC_PARAMS_RC_2_PASS_QUALITY:
426 case NV_ENC_PARAMS_RC_2_PASS_FRAMESIZE_CAP:
427 if (!(ctx->flags & NVENC_LOWLATENCY)) {
428 av_log(avctx, AV_LOG_WARNING,
429 "The multipass rate-control requires "
430 "a low-latency preset.\n");
435 rc->rateControlMode = ctx->rc;
438 static void nvenc_setup_rate_control(AVCodecContext *avctx)
440 NVENCContext *ctx = avctx->priv_data;
441 NV_ENC_RC_PARAMS *rc = &ctx->config.rcParams;
443 if (avctx->bit_rate > 0)
444 rc->averageBitRate = avctx->bit_rate;
446 if (avctx->rc_max_rate > 0)
447 rc->maxBitRate = avctx->rc_max_rate;
450 nvenc_override_rate_control(avctx, rc);
451 } else if (avctx->global_quality > 0) {
452 set_constqp(avctx, rc);
453 } else if (avctx->qmin >= 0 && avctx->qmax >= 0) {
454 rc->rateControlMode = NV_ENC_PARAMS_RC_VBR;
458 if (avctx->rc_buffer_size > 0)
459 rc->vbvBufferSize = avctx->rc_buffer_size;
461 if (rc->averageBitRate > 0)
462 avctx->bit_rate = rc->averageBitRate;
465 static int nvenc_setup_h264_config(AVCodecContext *avctx)
467 NVENCContext *ctx = avctx->priv_data;
468 NV_ENC_CONFIG *cc = &ctx->config;
469 NV_ENC_CONFIG_H264 *h264 = &cc->encodeCodecConfig.h264Config;
470 NV_ENC_CONFIG_H264_VUI_PARAMETERS *vui = &h264->h264VUIParameters;
472 vui->colourDescriptionPresentFlag = 1;
473 vui->videoSignalTypePresentFlag = 1;
475 vui->colourMatrix = avctx->colorspace;
476 vui->colourPrimaries = avctx->color_primaries;
477 vui->transferCharacteristics = avctx->color_trc;
479 vui->videoFullRangeFlag = avctx->color_range == AVCOL_RANGE_JPEG;
481 h264->disableSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
482 h264->repeatSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 0 : 1;
484 h264->maxNumRefFrames = avctx->refs;
485 h264->idrPeriod = cc->gopLength;
488 avctx->profile = ctx->profile;
490 if (avctx->pix_fmt == AV_PIX_FMT_YUV444P)
491 h264->chromaFormatIDC = 3;
493 h264->chromaFormatIDC = 1;
495 switch (ctx->profile) {
496 case NV_ENC_H264_PROFILE_BASELINE:
497 cc->profileGUID = NV_ENC_H264_PROFILE_BASELINE_GUID;
499 case NV_ENC_H264_PROFILE_MAIN:
500 cc->profileGUID = NV_ENC_H264_PROFILE_MAIN_GUID;
502 case NV_ENC_H264_PROFILE_HIGH:
503 cc->profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
505 case NV_ENC_H264_PROFILE_HIGH_444:
506 cc->profileGUID = NV_ENC_H264_PROFILE_HIGH_444_GUID;
508 case NV_ENC_H264_PROFILE_CONSTRAINED_HIGH:
509 cc->profileGUID = NV_ENC_H264_PROFILE_CONSTRAINED_HIGH_GUID;
513 h264->level = ctx->level;
518 static int nvenc_setup_hevc_config(AVCodecContext *avctx)
520 NVENCContext *ctx = avctx->priv_data;
521 NV_ENC_CONFIG *cc = &ctx->config;
522 NV_ENC_CONFIG_HEVC *hevc = &cc->encodeCodecConfig.hevcConfig;
524 hevc->disableSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
525 hevc->repeatSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 0 : 1;
527 hevc->maxNumRefFramesInDPB = avctx->refs;
528 hevc->idrPeriod = cc->gopLength;
530 /* No other profile is supported in the current SDK version 5 */
531 cc->profileGUID = NV_ENC_HEVC_PROFILE_MAIN_GUID;
532 avctx->profile = FF_PROFILE_HEVC_MAIN;
535 hevc->level = ctx->level;
537 hevc->level = NV_ENC_LEVEL_AUTOSELECT;
541 hevc->tier = ctx->tier;
546 static int nvenc_setup_codec_config(AVCodecContext *avctx)
548 switch (avctx->codec->id) {
549 case AV_CODEC_ID_H264:
550 return nvenc_setup_h264_config(avctx);
551 case AV_CODEC_ID_HEVC:
552 return nvenc_setup_hevc_config(avctx);
557 static int nvenc_setup_encoder(AVCodecContext *avctx)
559 NVENCContext *ctx = avctx->priv_data;
560 NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
561 NV_ENC_PRESET_CONFIG preset_cfg = { 0 };
564 ctx->params.version = NV_ENC_INITIALIZE_PARAMS_VER;
566 ctx->params.encodeHeight = avctx->height;
567 ctx->params.encodeWidth = avctx->width;
569 if (avctx->sample_aspect_ratio.num &&
570 avctx->sample_aspect_ratio.den &&
571 (avctx->sample_aspect_ratio.num != 1 ||
572 avctx->sample_aspect_ratio.den != 1)) {
573 av_reduce(&ctx->params.darWidth,
574 &ctx->params.darHeight,
575 avctx->width * avctx->sample_aspect_ratio.num,
576 avctx->height * avctx->sample_aspect_ratio.den,
579 ctx->params.darHeight = avctx->height;
580 ctx->params.darWidth = avctx->width;
583 ctx->params.frameRateNum = avctx->time_base.den;
584 ctx->params.frameRateDen = avctx->time_base.num * avctx->ticks_per_frame;
586 ctx->params.enableEncodeAsync = 0;
587 ctx->params.enablePTD = 1;
589 ctx->params.encodeConfig = &ctx->config;
591 nvec_map_preset(ctx);
593 preset_cfg.version = NV_ENC_PRESET_CONFIG_VER;
594 preset_cfg.presetCfg.version = NV_ENC_CONFIG_VER;
596 ret = nv->nvEncGetEncodePresetConfig(ctx->nvenc_ctx,
597 ctx->params.encodeGUID,
598 ctx->params.presetGUID,
600 if (ret != NV_ENC_SUCCESS) {
601 av_log(avctx, AV_LOG_ERROR,
602 "Cannot get the preset configuration\n");
603 return AVERROR_UNKNOWN;
606 memcpy(&ctx->config, &preset_cfg.presetCfg, sizeof(ctx->config));
608 ctx->config.version = NV_ENC_CONFIG_VER;
610 if (avctx->gop_size > 0) {
611 if (avctx->max_b_frames > 0) {
616 * 3 two B frames, and so on. */
617 ctx->config.frameIntervalP = avctx->max_b_frames + 1;
618 } else if (avctx->max_b_frames == 0) {
619 ctx->config.frameIntervalP = 1;
621 ctx->config.gopLength = avctx->gop_size;
622 } else if (avctx->gop_size == 0) {
623 ctx->config.frameIntervalP = 0;
624 ctx->config.gopLength = 1;
627 if (ctx->config.frameIntervalP > 1)
628 avctx->max_b_frames = ctx->config.frameIntervalP - 1;
630 nvenc_setup_rate_control(avctx);
632 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
633 ctx->config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FIELD;
635 ctx->config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FRAME;
638 if ((ret = nvenc_setup_codec_config(avctx)) < 0)
641 ret = nv->nvEncInitializeEncoder(ctx->nvenc_ctx, &ctx->params);
642 if (ret != NV_ENC_SUCCESS) {
643 av_log(avctx, AV_LOG_ERROR, "Cannot initialize the decoder");
644 return AVERROR_UNKNOWN;
650 static int nvenc_alloc_surface(AVCodecContext *avctx, int idx)
652 NVENCContext *ctx = avctx->priv_data;
653 NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
655 NV_ENC_CREATE_INPUT_BUFFER in_buffer = { 0 };
656 NV_ENC_CREATE_BITSTREAM_BUFFER out_buffer = { 0 };
658 in_buffer.version = NV_ENC_CREATE_INPUT_BUFFER_VER;
659 out_buffer.version = NV_ENC_CREATE_BITSTREAM_BUFFER_VER;
661 in_buffer.width = avctx->width;
662 in_buffer.height = avctx->height;
664 in_buffer.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_UNCACHED;
666 switch (avctx->pix_fmt) {
667 case AV_PIX_FMT_YUV420P:
668 in_buffer.bufferFmt = NV_ENC_BUFFER_FORMAT_YV12_PL;
670 case AV_PIX_FMT_NV12:
671 in_buffer.bufferFmt = NV_ENC_BUFFER_FORMAT_NV12_PL;
673 case AV_PIX_FMT_YUV444P:
674 in_buffer.bufferFmt = NV_ENC_BUFFER_FORMAT_YUV444_PL;
680 ret = nv->nvEncCreateInputBuffer(ctx->nvenc_ctx, &in_buffer);
681 if (ret != NV_ENC_SUCCESS) {
682 av_log(avctx, AV_LOG_ERROR, "CreateInputBuffer failed\n");
683 return AVERROR_UNKNOWN;
686 ctx->in[idx].in = in_buffer.inputBuffer;
687 ctx->in[idx].format = in_buffer.bufferFmt;
689 /* 1MB is large enough to hold most output frames.
690 * NVENC increases this automaticaly if it's not enough. */
691 out_buffer.size = BITSTREAM_BUFFER_SIZE;
693 out_buffer.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_UNCACHED;
695 ret = nv->nvEncCreateBitstreamBuffer(ctx->nvenc_ctx, &out_buffer);
696 if (ret != NV_ENC_SUCCESS) {
697 av_log(avctx, AV_LOG_ERROR, "CreateBitstreamBuffer failed\n");
698 return AVERROR_UNKNOWN;
701 ctx->out[idx].out = out_buffer.bitstreamBuffer;
702 ctx->out[idx].busy = 0;
707 static int nvenc_setup_surfaces(AVCodecContext *avctx)
709 NVENCContext *ctx = avctx->priv_data;
712 ctx->nb_surfaces = FFMAX(4 + avctx->max_b_frames,
715 ctx->in = av_mallocz(ctx->nb_surfaces * sizeof(*ctx->in));
717 return AVERROR(ENOMEM);
719 ctx->out = av_mallocz(ctx->nb_surfaces * sizeof(*ctx->out));
721 return AVERROR(ENOMEM);
723 ctx->timestamps = av_fifo_alloc(ctx->nb_surfaces * sizeof(int64_t));
724 if (!ctx->timestamps)
725 return AVERROR(ENOMEM);
726 ctx->pending = av_fifo_alloc(ctx->nb_surfaces * sizeof(ctx->out));
728 return AVERROR(ENOMEM);
729 ctx->ready = av_fifo_alloc(ctx->nb_surfaces * sizeof(ctx->out));
731 return AVERROR(ENOMEM);
733 for (i = 0; i < ctx->nb_surfaces; i++) {
734 if ((ret = nvenc_alloc_surface(avctx, i)) < 0)
741 #define EXTRADATA_SIZE 512
743 static int nvenc_setup_extradata(AVCodecContext *avctx)
745 NVENCContext *ctx = avctx->priv_data;
746 NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
747 NV_ENC_SEQUENCE_PARAM_PAYLOAD payload = { 0 };
750 avctx->extradata = av_mallocz(EXTRADATA_SIZE + AV_INPUT_BUFFER_PADDING_SIZE);
751 if (!avctx->extradata)
752 return AVERROR(ENOMEM);
754 payload.version = NV_ENC_SEQUENCE_PARAM_PAYLOAD_VER;
755 payload.spsppsBuffer = avctx->extradata;
756 payload.inBufferSize = EXTRADATA_SIZE;
757 payload.outSPSPPSPayloadSize = &avctx->extradata_size;
759 ret = nv->nvEncGetSequenceParams(ctx->nvenc_ctx, &payload);
760 if (ret != NV_ENC_SUCCESS) {
761 av_log(avctx, AV_LOG_ERROR, "Cannot get the extradata\n");
762 return AVERROR_UNKNOWN;
768 av_cold int ff_nvenc_encode_close(AVCodecContext *avctx)
770 NVENCContext *ctx = avctx->priv_data;
771 NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
774 av_fifo_free(ctx->timestamps);
775 av_fifo_free(ctx->pending);
776 av_fifo_free(ctx->ready);
779 for (i = 0; i < ctx->nb_surfaces; ++i) {
780 nv->nvEncDestroyInputBuffer(ctx->nvenc_ctx, ctx->in[i].in);
781 nv->nvEncDestroyBitstreamBuffer(ctx->nvenc_ctx, ctx->out[i].out);
789 nv->nvEncDestroyEncoder(ctx->nvenc_ctx);
792 ctx->nvel.cu_ctx_destroy(ctx->cu_context);
795 dlclose(ctx->nvel.nvenc);
798 dlclose(ctx->nvel.cuda);
803 av_cold int ff_nvenc_encode_init(AVCodecContext *avctx)
807 if ((ret = nvenc_load_libraries(avctx)) < 0)
810 if ((ret = nvenc_setup_device(avctx)) < 0)
813 if ((ret = nvenc_setup_encoder(avctx)) < 0)
816 if ((ret = nvenc_setup_surfaces(avctx)) < 0)
819 if (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) {
820 if ((ret = nvenc_setup_extradata(avctx)) < 0)
827 static NVENCInputSurface *get_input_surface(NVENCContext *ctx)
831 for (i = 0; i < ctx->nb_surfaces; i++) {
832 if (!ctx->in[i].locked) {
833 ctx->in[i].locked = 1;
841 static NVENCOutputSurface *get_output_surface(NVENCContext *ctx)
845 for (i = 0; i < ctx->nb_surfaces; i++) {
846 if (!ctx->out[i].busy) {
854 static int nvenc_copy_frame(NV_ENC_LOCK_INPUT_BUFFER *in, const AVFrame *frame)
856 uint8_t *buf = in->bufferDataPtr;
857 int off = frame->height * in->pitch;
859 switch (frame->format) {
860 case AV_PIX_FMT_YUV420P:
861 av_image_copy_plane(buf, in->pitch,
862 frame->data[0], frame->linesize[0],
863 frame->width, frame->height);
866 av_image_copy_plane(buf, in->pitch >> 1,
867 frame->data[2], frame->linesize[2],
868 frame->width >> 1, frame->height >> 1);
872 av_image_copy_plane(buf, in->pitch >> 1,
873 frame->data[1], frame->linesize[1],
874 frame->width >> 1, frame->height >> 1);
876 case AV_PIX_FMT_NV12:
877 av_image_copy_plane(buf, in->pitch,
878 frame->data[0], frame->linesize[0],
879 frame->width, frame->height);
882 av_image_copy_plane(buf, in->pitch,
883 frame->data[1], frame->linesize[1],
884 frame->width, frame->height >> 1);
886 case AV_PIX_FMT_YUV444P:
887 av_image_copy_plane(buf, in->pitch,
888 frame->data[0], frame->linesize[0],
889 frame->width, frame->height);
892 av_image_copy_plane(buf, in->pitch,
893 frame->data[1], frame->linesize[1],
894 frame->width, frame->height);
897 av_image_copy_plane(buf, in->pitch,
898 frame->data[2], frame->linesize[2],
899 frame->width, frame->height);
908 static int nvenc_enqueue_frame(AVCodecContext *avctx, const AVFrame *frame,
909 NVENCInputSurface **in_surf)
911 NVENCContext *ctx = avctx->priv_data;
912 NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
913 NV_ENC_LOCK_INPUT_BUFFER params = { 0 };
914 NVENCInputSurface *in = get_input_surface(ctx);
920 params.version = NV_ENC_LOCK_INPUT_BUFFER_VER;
921 params.inputBuffer = in->in;
924 ret = nv->nvEncLockInputBuffer(ctx->nvenc_ctx, ¶ms);
925 if (ret != NV_ENC_SUCCESS) {
926 av_log(avctx, AV_LOG_ERROR, "Cannot lock the buffer %p.\n",
928 return AVERROR_UNKNOWN;
931 ret = nvenc_copy_frame(¶ms, frame);
935 ret = nv->nvEncUnlockInputBuffer(ctx->nvenc_ctx, in->in);
936 if (ret != NV_ENC_SUCCESS) {
937 av_log(avctx, AV_LOG_ERROR, "Cannot unlock the buffer %p.\n",
939 return AVERROR_UNKNOWN;
947 nv->nvEncUnlockInputBuffer(ctx->nvenc_ctx, in->in);
952 static void nvenc_codec_specific_pic_params(AVCodecContext *avctx,
953 NV_ENC_PIC_PARAMS *params)
955 NVENCContext *ctx = avctx->priv_data;
957 switch (avctx->codec->id) {
958 case AV_CODEC_ID_H264:
959 params->codecPicParams.h264PicParams.sliceMode =
960 ctx->config.encodeCodecConfig.h264Config.sliceMode;
961 params->codecPicParams.h264PicParams.sliceModeData =
962 ctx->config.encodeCodecConfig.h264Config.sliceModeData;
964 case AV_CODEC_ID_HEVC:
965 params->codecPicParams.hevcPicParams.sliceMode =
966 ctx->config.encodeCodecConfig.hevcConfig.sliceMode;
967 params->codecPicParams.hevcPicParams.sliceModeData =
968 ctx->config.encodeCodecConfig.hevcConfig.sliceModeData;
973 static inline int nvenc_enqueue_timestamp(AVFifoBuffer *f, int64_t pts)
975 return av_fifo_generic_write(f, &pts, sizeof(pts), NULL);
978 static inline int nvenc_dequeue_timestamp(AVFifoBuffer *f, int64_t *pts)
980 return av_fifo_generic_read(f, pts, sizeof(*pts), NULL);
983 static inline int nvenc_enqueue_surface(AVFifoBuffer *f,
984 NVENCOutputSurface *surf)
987 return av_fifo_generic_write(f, &surf, sizeof(surf), NULL);
990 static inline int nvenc_dequeue_surface(AVFifoBuffer *f,
991 NVENCOutputSurface **surf)
993 return av_fifo_generic_read(f, surf, sizeof(*surf), NULL);
996 static int nvenc_set_timestamp(NVENCContext *ctx,
997 NV_ENC_LOCK_BITSTREAM *params,
1000 pkt->pts = params->outputTimeStamp;
1001 pkt->duration = params->outputDuration;
1003 return nvenc_dequeue_timestamp(ctx->timestamps, &pkt->dts);
1006 static int nvenc_get_frame(AVCodecContext *avctx, AVPacket *pkt)
1008 NVENCContext *ctx = avctx->priv_data;
1009 NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
1010 NV_ENC_LOCK_BITSTREAM params = { 0 };
1011 NVENCOutputSurface *out = NULL;
1014 ret = nvenc_dequeue_surface(ctx->pending, &out);
1018 params.version = NV_ENC_LOCK_BITSTREAM_VER;
1019 params.outputBitstream = out->out;
1021 ret = nv->nvEncLockBitstream(ctx->nvenc_ctx, ¶ms);
1023 return AVERROR_UNKNOWN;
1025 ret = ff_alloc_packet(pkt, params.bitstreamSizeInBytes);
1029 memcpy(pkt->data, params.bitstreamBufferPtr, pkt->size);
1031 ret = nv->nvEncUnlockBitstream(ctx->nvenc_ctx, out->out);
1033 return AVERROR_UNKNOWN;
1035 out->busy = out->in->locked = 0;
1037 ret = nvenc_set_timestamp(ctx, ¶ms, pkt);
1041 switch (params.pictureType) {
1042 case NV_ENC_PIC_TYPE_IDR:
1043 pkt->flags |= AV_PKT_FLAG_KEY;
1044 #if FF_API_CODED_FRAME
1045 FF_DISABLE_DEPRECATION_WARNINGS
1046 case NV_ENC_PIC_TYPE_INTRA_REFRESH:
1047 case NV_ENC_PIC_TYPE_I:
1048 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
1050 case NV_ENC_PIC_TYPE_P:
1051 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_P;
1053 case NV_ENC_PIC_TYPE_B:
1054 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_B;
1056 case NV_ENC_PIC_TYPE_BI:
1057 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_BI;
1059 FF_ENABLE_DEPRECATION_WARNINGS
1066 int ff_nvenc_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1067 const AVFrame *frame, int *got_packet)
1069 NVENCContext *ctx = avctx->priv_data;
1070 NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
1071 NV_ENC_PIC_PARAMS params = { 0 };
1072 NVENCInputSurface *in = NULL;
1073 NVENCOutputSurface *out = NULL;
1076 params.version = NV_ENC_PIC_PARAMS_VER;
1079 ret = nvenc_enqueue_frame(avctx, frame, &in);
1082 out = get_output_surface(ctx);
1088 params.inputBuffer = in->in;
1089 params.bufferFmt = in->format;
1090 params.inputWidth = frame->width;
1091 params.inputHeight = frame->height;
1092 params.outputBitstream = out->out;
1093 params.inputTimeStamp = frame->pts;
1095 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
1096 if (frame->top_field_first)
1097 params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_TOP_BOTTOM;
1099 params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_BOTTOM_TOP;
1101 params.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
1104 nvenc_codec_specific_pic_params(avctx, ¶ms);
1106 ret = nvenc_enqueue_timestamp(ctx->timestamps, frame->pts);
1110 params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
1113 ret = nv->nvEncEncodePicture(ctx->nvenc_ctx, ¶ms);
1115 if (ret != NV_ENC_SUCCESS &&
1116 ret != NV_ENC_ERR_NEED_MORE_INPUT) {
1118 return AVERROR_UNKNOWN;
1122 ret = nvenc_enqueue_surface(ctx->pending, out);
1127 if (ret != NV_ENC_ERR_NEED_MORE_INPUT &&
1128 av_fifo_size(ctx->pending)) {
1129 ret = nvenc_get_frame(avctx, pkt);