2 * H.264 hardware encoding using nvidia nvenc
3 * Copyright (c) 2014 Timo Rothenpieler <timo@rothenpieler.org>
5 * This file is part of FFmpeg.
7 * FFmpeg 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 * FFmpeg 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 FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
28 #include <nvEncodeAPI.h>
30 #include "libavutil/internal.h"
31 #include "libavutil/imgutils.h"
32 #include "libavutil/avassert.h"
33 #include "libavutil/opt.h"
34 #include "libavutil/mem.h"
40 #define CUDAAPI __stdcall
46 #define LOAD_FUNC(l, s) GetProcAddress(l, s)
47 #define DL_CLOSE_FUNC(l) FreeLibrary(l)
49 #define LOAD_FUNC(l, s) dlsym(l, s)
50 #define DL_CLOSE_FUNC(l) dlclose(l)
53 typedef enum cudaError_enum {
57 typedef void* CUcontext;
59 typedef CUresult(CUDAAPI *PCUINIT)(unsigned int Flags);
60 typedef CUresult(CUDAAPI *PCUDEVICEGETCOUNT)(int *count);
61 typedef CUresult(CUDAAPI *PCUDEVICEGET)(CUdevice *device, int ordinal);
62 typedef CUresult(CUDAAPI *PCUDEVICEGETNAME)(char *name, int len, CUdevice dev);
63 typedef CUresult(CUDAAPI *PCUDEVICECOMPUTECAPABILITY)(int *major, int *minor, CUdevice dev);
64 typedef CUresult(CUDAAPI *PCUCTXCREATE)(CUcontext *pctx, unsigned int flags, CUdevice dev);
65 typedef CUresult(CUDAAPI *PCUCTXPOPCURRENT)(CUcontext *pctx);
66 typedef CUresult(CUDAAPI *PCUCTXDESTROY)(CUcontext ctx);
68 typedef NVENCSTATUS (NVENCAPI* PNVENCODEAPICREATEINSTANCE)(NV_ENCODE_API_FUNCTION_LIST *functionList);
70 #if NVENCAPI_MAJOR_VERSION < 5
71 static const GUID dummy_license = { 0x0, 0x0, 0x0, { 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 } };
74 typedef struct NvencInputSurface
76 NV_ENC_INPUT_PTR input_surface;
82 NV_ENC_BUFFER_FORMAT format;
85 typedef struct NvencOutputSurface
87 NV_ENC_OUTPUT_PTR output_surface;
90 NvencInputSurface* input_surface;
95 typedef struct NvencData
99 NvencOutputSurface *surface;
103 typedef struct NvencDataList
112 typedef struct NvencDynLoadFunctions
115 PCUDEVICEGETCOUNT cu_device_get_count;
116 PCUDEVICEGET cu_device_get;
117 PCUDEVICEGETNAME cu_device_get_name;
118 PCUDEVICECOMPUTECAPABILITY cu_device_compute_capability;
119 PCUCTXCREATE cu_ctx_create;
120 PCUCTXPOPCURRENT cu_ctx_pop_current;
121 PCUCTXDESTROY cu_ctx_destroy;
123 NV_ENCODE_API_FUNCTION_LIST nvenc_funcs;
124 int nvenc_device_count;
125 CUdevice nvenc_devices[16];
134 } NvencDynLoadFunctions;
136 typedef struct NvencContext
140 NvencDynLoadFunctions nvenc_dload_funcs;
142 NV_ENC_INITIALIZE_PARAMS init_encode_params;
143 NV_ENC_CONFIG encode_config;
144 CUcontext cu_context;
146 int max_surface_count;
147 NvencInputSurface *input_surfaces;
148 NvencOutputSurface *output_surfaces;
150 NvencDataList output_surface_queue;
151 NvencDataList output_surface_ready_queue;
152 NvencDataList timestamp_list;
163 static NvencData* data_queue_dequeue(NvencDataList* queue)
169 av_assert0(queue->size);
170 av_assert0(queue->data);
175 /* Size always is a multiple of two */
176 mask = queue->size - 1;
177 read_pos = (queue->pos - queue->count) & mask;
180 return &queue->data[read_pos];
183 static int data_queue_enqueue(NvencDataList* queue, NvencData *data)
185 NvencDataList new_queue;
190 /* size always has to be a multiple of two */
195 queue->data = av_malloc(queue->size * sizeof(*(queue->data)));
199 return AVERROR(ENOMEM);
203 if (queue->count == queue->size) {
204 new_queue.size = queue->size << 1;
207 new_queue.data = av_malloc(new_queue.size * sizeof(*(queue->data)));
210 return AVERROR(ENOMEM);
212 while (tmp_data = data_queue_dequeue(queue))
213 data_queue_enqueue(&new_queue, tmp_data);
215 av_free(queue->data);
219 mask = queue->size - 1;
221 queue->data[queue->pos] = *data;
222 queue->pos = (queue->pos + 1) & mask;
228 static int out_surf_queue_enqueue(NvencDataList* queue, NvencOutputSurface* surface)
231 data.surface = surface;
233 return data_queue_enqueue(queue, &data);
236 static NvencOutputSurface* out_surf_queue_dequeue(NvencDataList* queue)
238 NvencData* res = data_queue_dequeue(queue);
246 static int timestamp_queue_enqueue(NvencDataList* queue, int64_t timestamp)
249 data.timestamp = timestamp;
251 return data_queue_enqueue(queue, &data);
254 static int64_t timestamp_queue_dequeue(NvencDataList* queue)
256 NvencData* res = data_queue_dequeue(queue);
259 return AV_NOPTS_VALUE;
261 return res->timestamp;
264 #define CHECK_LOAD_FUNC(t, f, s) \
266 (f) = (t)LOAD_FUNC(dl_fn->cuda_lib, s); \
268 av_log(avctx, AV_LOG_FATAL, "Failed loading %s from CUDA library\n", s); \
273 static av_cold int nvenc_dyload_cuda(AVCodecContext *avctx)
275 NvencContext *ctx = avctx->priv_data;
276 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
282 dl_fn->cuda_lib = LoadLibrary(TEXT("nvcuda.dll"));
284 dl_fn->cuda_lib = dlopen("libcuda.so", RTLD_LAZY);
287 if (!dl_fn->cuda_lib) {
288 av_log(avctx, AV_LOG_FATAL, "Failed loading CUDA library\n");
292 CHECK_LOAD_FUNC(PCUINIT, dl_fn->cu_init, "cuInit");
293 CHECK_LOAD_FUNC(PCUDEVICEGETCOUNT, dl_fn->cu_device_get_count, "cuDeviceGetCount");
294 CHECK_LOAD_FUNC(PCUDEVICEGET, dl_fn->cu_device_get, "cuDeviceGet");
295 CHECK_LOAD_FUNC(PCUDEVICEGETNAME, dl_fn->cu_device_get_name, "cuDeviceGetName");
296 CHECK_LOAD_FUNC(PCUDEVICECOMPUTECAPABILITY, dl_fn->cu_device_compute_capability, "cuDeviceComputeCapability");
297 CHECK_LOAD_FUNC(PCUCTXCREATE, dl_fn->cu_ctx_create, "cuCtxCreate_v2");
298 CHECK_LOAD_FUNC(PCUCTXPOPCURRENT, dl_fn->cu_ctx_pop_current, "cuCtxPopCurrent_v2");
299 CHECK_LOAD_FUNC(PCUCTXDESTROY, dl_fn->cu_ctx_destroy, "cuCtxDestroy_v2");
306 DL_CLOSE_FUNC(dl_fn->cuda_lib);
308 dl_fn->cuda_lib = NULL;
313 static av_cold int check_cuda_errors(AVCodecContext *avctx, CUresult err, const char *func)
315 if (err != CUDA_SUCCESS) {
316 av_log(avctx, AV_LOG_FATAL, ">> %s - failed with error code 0x%x\n", func, err);
321 #define check_cuda_errors(f) if (!check_cuda_errors(avctx, f, #f)) goto error
323 static av_cold int nvenc_check_cuda(AVCodecContext *avctx)
325 int device_count = 0;
326 CUdevice cu_device = 0;
328 int smminor = 0, smmajor = 0;
331 NvencContext *ctx = avctx->priv_data;
332 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
334 if (!nvenc_dyload_cuda(avctx))
337 if (dl_fn->nvenc_device_count > 0)
340 check_cuda_errors(dl_fn->cu_init(0));
342 check_cuda_errors(dl_fn->cu_device_get_count(&device_count));
345 av_log(avctx, AV_LOG_FATAL, "No CUDA capable devices found\n");
349 av_log(avctx, AV_LOG_VERBOSE, "%d CUDA capable devices found\n", device_count);
351 dl_fn->nvenc_device_count = 0;
353 for (i = 0; i < device_count; ++i) {
354 check_cuda_errors(dl_fn->cu_device_get(&cu_device, i));
355 check_cuda_errors(dl_fn->cu_device_get_name(gpu_name, sizeof(gpu_name), cu_device));
356 check_cuda_errors(dl_fn->cu_device_compute_capability(&smmajor, &smminor, cu_device));
358 smver = (smmajor << 4) | smminor;
360 av_log(avctx, AV_LOG_VERBOSE, "[ GPU #%d - < %s > has Compute SM %d.%d, NVENC %s ]\n", i, gpu_name, smmajor, smminor, (smver >= 0x30) ? "Available" : "Not Available");
363 dl_fn->nvenc_devices[dl_fn->nvenc_device_count++] = cu_device;
366 if (!dl_fn->nvenc_device_count) {
367 av_log(avctx, AV_LOG_FATAL, "No NVENC capable devices found\n");
375 dl_fn->nvenc_device_count = 0;
380 static av_cold int nvenc_dyload_nvenc(AVCodecContext *avctx)
382 PNVENCODEAPICREATEINSTANCE nvEncodeAPICreateInstance = 0;
383 NVENCSTATUS nvstatus;
385 NvencContext *ctx = avctx->priv_data;
386 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
388 if (!nvenc_check_cuda(avctx))
391 if (dl_fn->nvenc_lib)
395 if (sizeof(void*) == 8) {
396 dl_fn->nvenc_lib = LoadLibrary(TEXT("nvEncodeAPI64.dll"));
398 dl_fn->nvenc_lib = LoadLibrary(TEXT("nvEncodeAPI.dll"));
401 dl_fn->nvenc_lib = dlopen("libnvidia-encode.so.1", RTLD_LAZY);
404 if (!dl_fn->nvenc_lib) {
405 av_log(avctx, AV_LOG_FATAL, "Failed loading the nvenc library\n");
409 nvEncodeAPICreateInstance = (PNVENCODEAPICREATEINSTANCE)LOAD_FUNC(dl_fn->nvenc_lib, "NvEncodeAPICreateInstance");
411 if (!nvEncodeAPICreateInstance) {
412 av_log(avctx, AV_LOG_FATAL, "Failed to load nvenc entrypoint\n");
416 dl_fn->nvenc_funcs.version = NV_ENCODE_API_FUNCTION_LIST_VER;
418 nvstatus = nvEncodeAPICreateInstance(&dl_fn->nvenc_funcs);
420 if (nvstatus != NV_ENC_SUCCESS) {
421 av_log(avctx, AV_LOG_FATAL, "Failed to create nvenc instance\n");
425 av_log(avctx, AV_LOG_VERBOSE, "Nvenc initialized successfully\n");
430 if (dl_fn->nvenc_lib)
431 DL_CLOSE_FUNC(dl_fn->nvenc_lib);
433 dl_fn->nvenc_lib = NULL;
438 static av_cold void nvenc_unload_nvenc(AVCodecContext *avctx)
440 NvencContext *ctx = avctx->priv_data;
441 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
443 DL_CLOSE_FUNC(dl_fn->nvenc_lib);
444 dl_fn->nvenc_lib = NULL;
446 dl_fn->nvenc_device_count = 0;
448 DL_CLOSE_FUNC(dl_fn->cuda_lib);
449 dl_fn->cuda_lib = NULL;
451 dl_fn->cu_init = NULL;
452 dl_fn->cu_device_get_count = NULL;
453 dl_fn->cu_device_get = NULL;
454 dl_fn->cu_device_get_name = NULL;
455 dl_fn->cu_device_compute_capability = NULL;
456 dl_fn->cu_ctx_create = NULL;
457 dl_fn->cu_ctx_pop_current = NULL;
458 dl_fn->cu_ctx_destroy = NULL;
460 av_log(avctx, AV_LOG_VERBOSE, "Nvenc unloaded\n");
463 static av_cold int nvenc_encode_init(AVCodecContext *avctx)
465 NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS encode_session_params = { 0 };
466 NV_ENC_PRESET_CONFIG preset_config = { 0 };
467 CUcontext cu_context_curr;
469 GUID encoder_preset = NV_ENC_PRESET_HQ_GUID;
470 NVENCSTATUS nv_status = NV_ENC_SUCCESS;
471 int surfaceCount = 0;
477 #if NVENCAPI_MAJOR_VERSION < 5
478 GUID license = dummy_license;
481 NvencContext *ctx = avctx->priv_data;
482 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
483 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
485 if (!nvenc_dyload_nvenc(avctx))
486 return AVERROR_EXTERNAL;
488 avctx->coded_frame = av_frame_alloc();
489 if (!avctx->coded_frame) {
490 res = AVERROR(ENOMEM);
494 ctx->last_dts = AV_NOPTS_VALUE;
496 ctx->encode_config.version = NV_ENC_CONFIG_VER;
497 ctx->init_encode_params.version = NV_ENC_INITIALIZE_PARAMS_VER;
498 preset_config.version = NV_ENC_PRESET_CONFIG_VER;
499 preset_config.presetCfg.version = NV_ENC_CONFIG_VER;
500 encode_session_params.version = NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER;
501 encode_session_params.apiVersion = NVENCAPI_VERSION;
503 #if NVENCAPI_MAJOR_VERSION < 5
504 encode_session_params.clientKeyPtr = &license;
507 if (ctx->gpu >= dl_fn->nvenc_device_count) {
508 av_log(avctx, AV_LOG_FATAL, "Requested GPU %d, but only %d GPUs are available!\n", ctx->gpu, dl_fn->nvenc_device_count);
509 res = AVERROR(EINVAL);
513 ctx->cu_context = NULL;
514 cu_res = dl_fn->cu_ctx_create(&ctx->cu_context, 0, dl_fn->nvenc_devices[ctx->gpu]);
516 if (cu_res != CUDA_SUCCESS) {
517 av_log(avctx, AV_LOG_FATAL, "Failed creating CUDA context for NVENC: 0x%x\n", (int)cu_res);
518 res = AVERROR_EXTERNAL;
522 cu_res = dl_fn->cu_ctx_pop_current(&cu_context_curr);
524 if (cu_res != CUDA_SUCCESS) {
525 av_log(avctx, AV_LOG_FATAL, "Failed popping CUDA context: 0x%x\n", (int)cu_res);
526 res = AVERROR_EXTERNAL;
530 encode_session_params.device = ctx->cu_context;
531 encode_session_params.deviceType = NV_ENC_DEVICE_TYPE_CUDA;
533 nv_status = p_nvenc->nvEncOpenEncodeSessionEx(&encode_session_params, &ctx->nvencoder);
534 if (nv_status != NV_ENC_SUCCESS) {
535 ctx->nvencoder = NULL;
536 av_log(avctx, AV_LOG_FATAL, "OpenEncodeSessionEx failed: 0x%x - invalid license key?\n", (int)nv_status);
537 res = AVERROR_EXTERNAL;
542 if (!strcmp(ctx->preset, "hp")) {
543 encoder_preset = NV_ENC_PRESET_HP_GUID;
544 } else if (!strcmp(ctx->preset, "hq")) {
545 encoder_preset = NV_ENC_PRESET_HQ_GUID;
546 } else if (!strcmp(ctx->preset, "bd")) {
547 encoder_preset = NV_ENC_PRESET_BD_GUID;
548 } else if (!strcmp(ctx->preset, "ll")) {
549 encoder_preset = NV_ENC_PRESET_LOW_LATENCY_DEFAULT_GUID;
551 } else if (!strcmp(ctx->preset, "llhp")) {
552 encoder_preset = NV_ENC_PRESET_LOW_LATENCY_HP_GUID;
554 } else if (!strcmp(ctx->preset, "llhq")) {
555 encoder_preset = NV_ENC_PRESET_LOW_LATENCY_HQ_GUID;
557 } else if (!strcmp(ctx->preset, "default")) {
558 encoder_preset = NV_ENC_PRESET_DEFAULT_GUID;
560 av_log(avctx, AV_LOG_FATAL, "Preset \"%s\" is unknown! Supported presets: hp, hq, bd, ll, llhp, llhq, default\n", ctx->preset);
561 res = AVERROR(EINVAL);
566 nv_status = p_nvenc->nvEncGetEncodePresetConfig(ctx->nvencoder, NV_ENC_CODEC_H264_GUID, encoder_preset, &preset_config);
567 if (nv_status != NV_ENC_SUCCESS) {
568 av_log(avctx, AV_LOG_FATAL, "GetEncodePresetConfig failed: 0x%x\n", (int)nv_status);
569 res = AVERROR_EXTERNAL;
573 ctx->init_encode_params.encodeGUID = NV_ENC_CODEC_H264_GUID;
574 ctx->init_encode_params.encodeHeight = avctx->height;
575 ctx->init_encode_params.encodeWidth = avctx->width;
577 if (avctx->sample_aspect_ratio.num && avctx->sample_aspect_ratio.den &&
578 (avctx->sample_aspect_ratio.num != 1 || avctx->sample_aspect_ratio.num != 1)) {
580 avctx->width * avctx->sample_aspect_ratio.num,
581 avctx->height * avctx->sample_aspect_ratio.den,
583 ctx->init_encode_params.darHeight = dh;
584 ctx->init_encode_params.darWidth = dw;
586 ctx->init_encode_params.darHeight = avctx->height;
587 ctx->init_encode_params.darWidth = avctx->width;
590 ctx->init_encode_params.frameRateNum = avctx->time_base.den;
591 ctx->init_encode_params.frameRateDen = avctx->time_base.num * avctx->ticks_per_frame;
593 num_mbs = ((avctx->width + 15) >> 4) * ((avctx->height + 15) >> 4);
594 ctx->max_surface_count = (num_mbs >= 8160) ? 32 : 48;
596 ctx->init_encode_params.enableEncodeAsync = 0;
597 ctx->init_encode_params.enablePTD = 1;
599 ctx->init_encode_params.presetGUID = encoder_preset;
601 ctx->init_encode_params.encodeConfig = &ctx->encode_config;
602 memcpy(&ctx->encode_config, &preset_config.presetCfg, sizeof(ctx->encode_config));
603 ctx->encode_config.version = NV_ENC_CONFIG_VER;
605 if (avctx->refs >= 0) {
606 /* 0 means "let the hardware decide" */
607 ctx->encode_config.encodeCodecConfig.h264Config.maxNumRefFrames = avctx->refs;
610 if (avctx->gop_size > 0) {
611 if (avctx->max_b_frames >= 0) {
612 /* 0 is intra-only, 1 is I/P only, 2 is one B Frame, 3 two B frames, and so on. */
613 ctx->encode_config.frameIntervalP = avctx->max_b_frames + 1;
616 ctx->encode_config.gopLength = avctx->gop_size;
617 ctx->encode_config.encodeCodecConfig.h264Config.idrPeriod = avctx->gop_size;
618 } else if (avctx->gop_size == 0) {
619 ctx->encode_config.frameIntervalP = 0;
620 ctx->encode_config.gopLength = 1;
621 ctx->encode_config.encodeCodecConfig.h264Config.idrPeriod = 1;
624 /* when there're b frames, set dts offset */
625 if (ctx->encode_config.frameIntervalP >= 2)
628 if (avctx->bit_rate > 0)
629 ctx->encode_config.rcParams.averageBitRate = avctx->bit_rate;
631 if (avctx->rc_max_rate > 0)
632 ctx->encode_config.rcParams.maxBitRate = avctx->rc_max_rate;
636 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CBR;
637 } else if (ctx->twopass == 1 || isLL) {
638 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_2_PASS_QUALITY;
640 ctx->encode_config.encodeCodecConfig.h264Config.adaptiveTransformMode = NV_ENC_H264_ADAPTIVE_TRANSFORM_ENABLE;
641 ctx->encode_config.encodeCodecConfig.h264Config.fmoMode = NV_ENC_H264_FMO_DISABLE;
644 av_log(avctx, AV_LOG_WARNING, "Twopass mode is only known to work with low latency (ll, llhq, llhp) presets.\n");
646 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CBR;
648 } else if (avctx->global_quality > 0) {
649 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
650 ctx->encode_config.rcParams.constQP.qpInterB = avctx->global_quality;
651 ctx->encode_config.rcParams.constQP.qpInterP = avctx->global_quality;
652 ctx->encode_config.rcParams.constQP.qpIntra = avctx->global_quality;
656 } else if (avctx->qmin >= 0 && avctx->qmax >= 0) {
657 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_VBR;
659 ctx->encode_config.rcParams.enableMinQP = 1;
660 ctx->encode_config.rcParams.enableMaxQP = 1;
662 ctx->encode_config.rcParams.minQP.qpInterB = avctx->qmin;
663 ctx->encode_config.rcParams.minQP.qpInterP = avctx->qmin;
664 ctx->encode_config.rcParams.minQP.qpIntra = avctx->qmin;
666 ctx->encode_config.rcParams.maxQP.qpInterB = avctx->qmax;
667 ctx->encode_config.rcParams.maxQP.qpInterP = avctx->qmax;
668 ctx->encode_config.rcParams.maxQP.qpIntra = avctx->qmax;
671 if (avctx->rc_buffer_size > 0)
672 ctx->encode_config.rcParams.vbvBufferSize = avctx->rc_buffer_size;
674 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
675 ctx->encode_config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FIELD;
677 ctx->encode_config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FRAME;
680 switch (avctx->profile) {
681 case FF_PROFILE_H264_BASELINE:
682 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_BASELINE_GUID;
684 case FF_PROFILE_H264_MAIN:
685 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_MAIN_GUID;
687 case FF_PROFILE_H264_HIGH:
688 case FF_PROFILE_UNKNOWN:
689 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
692 av_log(avctx, AV_LOG_WARNING, "Unsupported h264 profile requested, falling back to high\n");
693 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
697 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourDescriptionPresentFlag = 1;
698 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.videoSignalTypePresentFlag = 1;
700 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourMatrix = avctx->colorspace;
701 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourPrimaries = avctx->color_primaries;
702 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.transferCharacteristics = avctx->color_trc;
704 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.videoFullRangeFlag = avctx->color_range == AVCOL_RANGE_JPEG;
706 ctx->encode_config.encodeCodecConfig.h264Config.disableSPSPPS = (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
707 ctx->encode_config.encodeCodecConfig.h264Config.repeatSPSPPS = (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) ? 0 : 1;
709 nv_status = p_nvenc->nvEncInitializeEncoder(ctx->nvencoder, &ctx->init_encode_params);
710 if (nv_status != NV_ENC_SUCCESS) {
711 av_log(avctx, AV_LOG_FATAL, "InitializeEncoder failed: 0x%x\n", (int)nv_status);
712 res = AVERROR_EXTERNAL;
716 ctx->input_surfaces = av_malloc(ctx->max_surface_count * sizeof(*ctx->input_surfaces));
718 if (!ctx->input_surfaces) {
719 res = AVERROR(ENOMEM);
723 ctx->output_surfaces = av_malloc(ctx->max_surface_count * sizeof(*ctx->output_surfaces));
725 if (!ctx->output_surfaces) {
726 res = AVERROR(ENOMEM);
730 for (surfaceCount = 0; surfaceCount < ctx->max_surface_count; ++surfaceCount) {
731 NV_ENC_CREATE_INPUT_BUFFER allocSurf = { 0 };
732 NV_ENC_CREATE_BITSTREAM_BUFFER allocOut = { 0 };
733 allocSurf.version = NV_ENC_CREATE_INPUT_BUFFER_VER;
734 allocOut.version = NV_ENC_CREATE_BITSTREAM_BUFFER_VER;
736 allocSurf.width = (avctx->width + 31) & ~31;
737 allocSurf.height = (avctx->height + 31) & ~31;
739 allocSurf.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_CACHED;
741 switch (avctx->pix_fmt) {
742 case AV_PIX_FMT_YUV420P:
743 allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_YV12_PL;
746 case AV_PIX_FMT_NV12:
747 allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_NV12_PL;
750 case AV_PIX_FMT_YUV444P:
751 allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_YUV444_PL;
755 av_log(avctx, AV_LOG_FATAL, "Invalid input pixel format\n");
756 res = AVERROR(EINVAL);
760 nv_status = p_nvenc->nvEncCreateInputBuffer(ctx->nvencoder, &allocSurf);
761 if (nv_status = NV_ENC_SUCCESS){
762 av_log(avctx, AV_LOG_FATAL, "CreateInputBuffer failed\n");
763 res = AVERROR_EXTERNAL;
767 ctx->input_surfaces[surfaceCount].lockCount = 0;
768 ctx->input_surfaces[surfaceCount].input_surface = allocSurf.inputBuffer;
769 ctx->input_surfaces[surfaceCount].format = allocSurf.bufferFmt;
770 ctx->input_surfaces[surfaceCount].width = allocSurf.width;
771 ctx->input_surfaces[surfaceCount].height = allocSurf.height;
773 /* 1MB is large enough to hold most output frames. NVENC increases this automaticaly if it's not enough. */
774 allocOut.size = 1024 * 1024;
776 allocOut.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_CACHED;
778 nv_status = p_nvenc->nvEncCreateBitstreamBuffer(ctx->nvencoder, &allocOut);
779 if (nv_status = NV_ENC_SUCCESS) {
780 av_log(avctx, AV_LOG_FATAL, "CreateBitstreamBuffer failed\n");
781 ctx->output_surfaces[surfaceCount++].output_surface = NULL;
782 res = AVERROR_EXTERNAL;
786 ctx->output_surfaces[surfaceCount].output_surface = allocOut.bitstreamBuffer;
787 ctx->output_surfaces[surfaceCount].size = allocOut.size;
788 ctx->output_surfaces[surfaceCount].busy = 0;
791 if (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) {
792 uint32_t outSize = 0;
794 NV_ENC_SEQUENCE_PARAM_PAYLOAD payload = { 0 };
795 payload.version = NV_ENC_SEQUENCE_PARAM_PAYLOAD_VER;
797 payload.spsppsBuffer = tmpHeader;
798 payload.inBufferSize = sizeof(tmpHeader);
799 payload.outSPSPPSPayloadSize = &outSize;
801 nv_status = p_nvenc->nvEncGetSequenceParams(ctx->nvencoder, &payload);
802 if (nv_status != NV_ENC_SUCCESS) {
803 av_log(avctx, AV_LOG_FATAL, "GetSequenceParams failed\n");
807 avctx->extradata_size = outSize;
808 avctx->extradata = av_mallocz(outSize + FF_INPUT_BUFFER_PADDING_SIZE);
810 if (!avctx->extradata) {
811 res = AVERROR(ENOMEM);
815 memcpy(avctx->extradata, tmpHeader, outSize);
818 if (ctx->encode_config.frameIntervalP > 1)
819 avctx->has_b_frames = 2;
821 if (ctx->encode_config.rcParams.averageBitRate > 0)
822 avctx->bit_rate = ctx->encode_config.rcParams.averageBitRate;
828 for (i = 0; i < surfaceCount; ++i) {
829 p_nvenc->nvEncDestroyInputBuffer(ctx->nvencoder, ctx->input_surfaces[i].input_surface);
830 if (ctx->output_surfaces[i].output_surface)
831 p_nvenc->nvEncDestroyBitstreamBuffer(ctx->nvencoder, ctx->output_surfaces[i].output_surface);
835 p_nvenc->nvEncDestroyEncoder(ctx->nvencoder);
838 dl_fn->cu_ctx_destroy(ctx->cu_context);
840 av_frame_free(&avctx->coded_frame);
842 nvenc_unload_nvenc(avctx);
844 ctx->nvencoder = NULL;
845 ctx->cu_context = NULL;
850 static av_cold int nvenc_encode_close(AVCodecContext *avctx)
852 NvencContext *ctx = avctx->priv_data;
853 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
854 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
857 av_freep(&ctx->timestamp_list.data);
858 av_freep(&ctx->output_surface_ready_queue.data);
859 av_freep(&ctx->output_surface_queue.data);
861 for (i = 0; i < ctx->max_surface_count; ++i) {
862 p_nvenc->nvEncDestroyInputBuffer(ctx->nvencoder, ctx->input_surfaces[i].input_surface);
863 p_nvenc->nvEncDestroyBitstreamBuffer(ctx->nvencoder, ctx->output_surfaces[i].output_surface);
865 ctx->max_surface_count = 0;
867 p_nvenc->nvEncDestroyEncoder(ctx->nvencoder);
868 ctx->nvencoder = NULL;
870 dl_fn->cu_ctx_destroy(ctx->cu_context);
871 ctx->cu_context = NULL;
873 nvenc_unload_nvenc(avctx);
875 av_frame_free(&avctx->coded_frame);
880 static int process_output_surface(AVCodecContext *avctx, AVPacket *pkt, AVFrame *coded_frame, NvencOutputSurface *tmpoutsurf)
882 NvencContext *ctx = avctx->priv_data;
883 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
884 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
886 uint32_t *slice_offsets = av_mallocz(ctx->encode_config.encodeCodecConfig.h264Config.sliceModeData * sizeof(*slice_offsets));
887 NV_ENC_LOCK_BITSTREAM lock_params = { 0 };
888 NVENCSTATUS nv_status;
892 return AVERROR(ENOMEM);
894 lock_params.version = NV_ENC_LOCK_BITSTREAM_VER;
896 lock_params.doNotWait = 0;
897 lock_params.outputBitstream = tmpoutsurf->output_surface;
898 lock_params.sliceOffsets = slice_offsets;
900 nv_status = p_nvenc->nvEncLockBitstream(ctx->nvencoder, &lock_params);
901 if (nv_status != NV_ENC_SUCCESS) {
902 av_log(avctx, AV_LOG_ERROR, "Failed locking bitstream buffer\n");
903 res = AVERROR_EXTERNAL;
907 if (res = ff_alloc_packet2(avctx, pkt, lock_params.bitstreamSizeInBytes)) {
908 p_nvenc->nvEncUnlockBitstream(ctx->nvencoder, tmpoutsurf->output_surface);
912 memcpy(pkt->data, lock_params.bitstreamBufferPtr, lock_params.bitstreamSizeInBytes);
914 nv_status = p_nvenc->nvEncUnlockBitstream(ctx->nvencoder, tmpoutsurf->output_surface);
915 if (nv_status != NV_ENC_SUCCESS)
916 av_log(avctx, AV_LOG_ERROR, "Failed unlocking bitstream buffer, expect the gates of mordor to open\n");
918 switch (lock_params.pictureType) {
919 case NV_ENC_PIC_TYPE_IDR:
920 pkt->flags |= AV_PKT_FLAG_KEY;
921 case NV_ENC_PIC_TYPE_I:
922 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
924 case NV_ENC_PIC_TYPE_P:
925 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_P;
927 case NV_ENC_PIC_TYPE_B:
928 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_B;
930 case NV_ENC_PIC_TYPE_BI:
931 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_BI;
934 av_log(avctx, AV_LOG_ERROR, "Unknown picture type encountered, expect the output to be broken.\n");
935 av_log(avctx, AV_LOG_ERROR, "Please report this error and include as much information on how to reproduce it as possible.\n");
936 res = AVERROR_EXTERNAL;
940 pkt->pts = lock_params.outputTimeStamp;
941 pkt->dts = timestamp_queue_dequeue(&ctx->timestamp_list);
943 /* when there're b frame(s), set dts offset */
944 if (ctx->encode_config.frameIntervalP >= 2)
947 if (pkt->dts > pkt->pts)
950 if (ctx->last_dts != AV_NOPTS_VALUE && pkt->dts <= ctx->last_dts)
951 pkt->dts = ctx->last_dts + 1;
953 ctx->last_dts = pkt->dts;
955 av_free(slice_offsets);
961 av_free(slice_offsets);
962 timestamp_queue_dequeue(&ctx->timestamp_list);
967 static int nvenc_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
968 const AVFrame *frame, int *got_packet)
970 NVENCSTATUS nv_status;
971 NvencOutputSurface *tmpoutsurf;
974 NvencContext *ctx = avctx->priv_data;
975 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
976 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
978 NV_ENC_PIC_PARAMS pic_params = { 0 };
979 pic_params.version = NV_ENC_PIC_PARAMS_VER;
982 NV_ENC_LOCK_INPUT_BUFFER lockBufferParams = { 0 };
983 NvencInputSurface *inSurf = NULL;
985 for (i = 0; i < ctx->max_surface_count; ++i) {
986 if (!ctx->input_surfaces[i].lockCount) {
987 inSurf = &ctx->input_surfaces[i];
994 inSurf->lockCount = 1;
996 lockBufferParams.version = NV_ENC_LOCK_INPUT_BUFFER_VER;
997 lockBufferParams.inputBuffer = inSurf->input_surface;
999 nv_status = p_nvenc->nvEncLockInputBuffer(ctx->nvencoder, &lockBufferParams);
1000 if (nv_status != NV_ENC_SUCCESS) {
1001 av_log(avctx, AV_LOG_ERROR, "Failed locking nvenc input buffer\n");
1005 if (avctx->pix_fmt == AV_PIX_FMT_YUV420P) {
1006 uint8_t *buf = lockBufferParams.bufferDataPtr;
1008 av_image_copy_plane(buf, lockBufferParams.pitch,
1009 frame->data[0], frame->linesize[0],
1010 avctx->width, avctx->height);
1012 buf += inSurf->height * lockBufferParams.pitch;
1014 av_image_copy_plane(buf, lockBufferParams.pitch >> 1,
1015 frame->data[2], frame->linesize[2],
1016 avctx->width >> 1, avctx->height >> 1);
1018 buf += (inSurf->height * lockBufferParams.pitch) >> 2;
1020 av_image_copy_plane(buf, lockBufferParams.pitch >> 1,
1021 frame->data[1], frame->linesize[1],
1022 avctx->width >> 1, avctx->height >> 1);
1023 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12) {
1024 uint8_t *buf = lockBufferParams.bufferDataPtr;
1026 av_image_copy_plane(buf, lockBufferParams.pitch,
1027 frame->data[0], frame->linesize[0],
1028 avctx->width, avctx->height);
1030 buf += inSurf->height * lockBufferParams.pitch;
1032 av_image_copy_plane(buf, lockBufferParams.pitch,
1033 frame->data[1], frame->linesize[1],
1034 avctx->width, avctx->height >> 1);
1035 } else if (avctx->pix_fmt == AV_PIX_FMT_YUV444P) {
1036 uint8_t *buf = lockBufferParams.bufferDataPtr;
1038 av_image_copy_plane(buf, lockBufferParams.pitch,
1039 frame->data[0], frame->linesize[0],
1040 avctx->width, avctx->height);
1042 buf += inSurf->height * lockBufferParams.pitch;
1044 av_image_copy_plane(buf, lockBufferParams.pitch,
1045 frame->data[1], frame->linesize[1],
1046 avctx->width, avctx->height);
1048 buf += inSurf->height * lockBufferParams.pitch;
1050 av_image_copy_plane(buf, lockBufferParams.pitch,
1051 frame->data[2], frame->linesize[2],
1052 avctx->width, avctx->height);
1054 av_log(avctx, AV_LOG_FATAL, "Invalid pixel format!\n");
1055 return AVERROR(EINVAL);
1058 nv_status = p_nvenc->nvEncUnlockInputBuffer(ctx->nvencoder, inSurf->input_surface);
1059 if (nv_status != NV_ENC_SUCCESS) {
1060 av_log(avctx, AV_LOG_FATAL, "Failed unlocking input buffer!\n");
1061 return AVERROR_EXTERNAL;
1064 for (i = 0; i < ctx->max_surface_count; ++i)
1065 if (!ctx->output_surfaces[i].busy)
1068 if (i == ctx->max_surface_count) {
1069 inSurf->lockCount = 0;
1070 av_log(avctx, AV_LOG_FATAL, "No free output surface found!\n");
1071 return AVERROR_EXTERNAL;
1074 ctx->output_surfaces[i].input_surface = inSurf;
1076 pic_params.inputBuffer = inSurf->input_surface;
1077 pic_params.bufferFmt = inSurf->format;
1078 pic_params.inputWidth = avctx->width;
1079 pic_params.inputHeight = avctx->height;
1080 pic_params.outputBitstream = ctx->output_surfaces[i].output_surface;
1081 pic_params.completionEvent = 0;
1083 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
1084 if (frame->top_field_first) {
1085 pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_TOP_BOTTOM;
1087 pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_BOTTOM_TOP;
1090 pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
1093 pic_params.encodePicFlags = 0;
1094 pic_params.inputTimeStamp = frame->pts;
1095 pic_params.inputDuration = 0;
1096 pic_params.codecPicParams.h264PicParams.sliceMode = ctx->encode_config.encodeCodecConfig.h264Config.sliceMode;
1097 pic_params.codecPicParams.h264PicParams.sliceModeData = ctx->encode_config.encodeCodecConfig.h264Config.sliceModeData;
1099 #if NVENCAPI_MAJOR_VERSION < 5
1100 memcpy(&pic_params.rcParams, &ctx->encode_config.rcParams, sizeof(NV_ENC_RC_PARAMS));
1103 res = timestamp_queue_enqueue(&ctx->timestamp_list, frame->pts);
1108 pic_params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
1111 nv_status = p_nvenc->nvEncEncodePicture(ctx->nvencoder, &pic_params);
1113 if (frame && nv_status == NV_ENC_ERR_NEED_MORE_INPUT) {
1114 res = out_surf_queue_enqueue(&ctx->output_surface_queue, &ctx->output_surfaces[i]);
1119 ctx->output_surfaces[i].busy = 1;
1122 if (nv_status != NV_ENC_SUCCESS && nv_status != NV_ENC_ERR_NEED_MORE_INPUT) {
1123 av_log(avctx, AV_LOG_ERROR, "EncodePicture failed!\n");
1124 return AVERROR_EXTERNAL;
1127 if (nv_status != NV_ENC_ERR_NEED_MORE_INPUT) {
1128 while (ctx->output_surface_queue.count) {
1129 tmpoutsurf = out_surf_queue_dequeue(&ctx->output_surface_queue);
1130 res = out_surf_queue_enqueue(&ctx->output_surface_ready_queue, tmpoutsurf);
1137 res = out_surf_queue_enqueue(&ctx->output_surface_ready_queue, &ctx->output_surfaces[i]);
1142 ctx->output_surfaces[i].busy = 1;
1146 if (ctx->output_surface_ready_queue.count) {
1147 tmpoutsurf = out_surf_queue_dequeue(&ctx->output_surface_ready_queue);
1149 res = process_output_surface(avctx, pkt, avctx->coded_frame, tmpoutsurf);
1154 tmpoutsurf->busy = 0;
1155 av_assert0(tmpoutsurf->input_surface->lockCount);
1156 tmpoutsurf->input_surface->lockCount--;
1166 static enum AVPixelFormat pix_fmts_nvenc[] = {
1171 #define OFFSET(x) offsetof(NvencContext, x)
1172 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1173 static const AVOption options[] = {
1174 { "preset", "Set the encoding preset (one of hq, hp, bd, ll, llhq, llhp, default)", OFFSET(preset), AV_OPT_TYPE_STRING, { .str = "hq" }, 0, 0, VE },
1175 { "cbr", "Use cbr encoding mode", OFFSET(cbr), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1176 { "2pass", "Use 2pass cbr encoding mode (low latency mode only)", OFFSET(twopass), AV_OPT_TYPE_INT, { .i64 = -1 }, -1, 1, VE },
1177 { "gpu", "Selects which NVENC capable GPU to use. First GPU is 0, second is 1, and so on.", OFFSET(gpu), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
1181 static const AVClass nvenc_class = {
1182 .class_name = "nvenc",
1183 .item_name = av_default_item_name,
1185 .version = LIBAVUTIL_VERSION_INT,
1188 static const AVCodecDefault nvenc_defaults[] = {
1198 AVCodec ff_nvenc_encoder = {
1200 .long_name = NULL_IF_CONFIG_SMALL("Nvidia NVENC h264 encoder"),
1201 .type = AVMEDIA_TYPE_VIDEO,
1202 .id = AV_CODEC_ID_H264,
1203 .priv_data_size = sizeof(NvencContext),
1204 .init = nvenc_encode_init,
1205 .encode2 = nvenc_encode_frame,
1206 .close = nvenc_encode_close,
1207 .capabilities = CODEC_CAP_DELAY,
1208 .priv_class = &nvenc_class,
1209 .defaults = nvenc_defaults,
1210 .pix_fmts = pix_fmts_nvenc,