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;
164 static NvencData* data_queue_dequeue(NvencDataList* queue)
170 av_assert0(queue->size);
171 av_assert0(queue->data);
176 /* Size always is a multiple of two */
177 mask = queue->size - 1;
178 read_pos = (queue->pos - queue->count) & mask;
181 return &queue->data[read_pos];
184 static int data_queue_enqueue(NvencDataList* queue, NvencData *data)
186 NvencDataList new_queue;
191 /* size always has to be a multiple of two */
196 queue->data = av_malloc(queue->size * sizeof(*(queue->data)));
200 return AVERROR(ENOMEM);
204 if (queue->count == queue->size) {
205 new_queue.size = queue->size << 1;
208 new_queue.data = av_malloc(new_queue.size * sizeof(*(queue->data)));
211 return AVERROR(ENOMEM);
213 while (tmp_data = data_queue_dequeue(queue))
214 data_queue_enqueue(&new_queue, tmp_data);
216 av_free(queue->data);
220 mask = queue->size - 1;
222 queue->data[queue->pos] = *data;
223 queue->pos = (queue->pos + 1) & mask;
229 static int out_surf_queue_enqueue(NvencDataList* queue, NvencOutputSurface* surface)
232 data.surface = surface;
234 return data_queue_enqueue(queue, &data);
237 static NvencOutputSurface* out_surf_queue_dequeue(NvencDataList* queue)
239 NvencData* res = data_queue_dequeue(queue);
247 static int timestamp_queue_enqueue(NvencDataList* queue, int64_t timestamp)
250 data.timestamp = timestamp;
252 return data_queue_enqueue(queue, &data);
255 static int64_t timestamp_queue_dequeue(NvencDataList* queue)
257 NvencData* res = data_queue_dequeue(queue);
260 return AV_NOPTS_VALUE;
262 return res->timestamp;
265 #define CHECK_LOAD_FUNC(t, f, s) \
267 (f) = (t)LOAD_FUNC(dl_fn->cuda_lib, s); \
269 av_log(avctx, AV_LOG_FATAL, "Failed loading %s from CUDA library\n", s); \
274 static av_cold int nvenc_dyload_cuda(AVCodecContext *avctx)
276 NvencContext *ctx = avctx->priv_data;
277 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
283 dl_fn->cuda_lib = LoadLibrary(TEXT("nvcuda.dll"));
285 dl_fn->cuda_lib = dlopen("libcuda.so", RTLD_LAZY);
288 if (!dl_fn->cuda_lib) {
289 av_log(avctx, AV_LOG_FATAL, "Failed loading CUDA library\n");
293 CHECK_LOAD_FUNC(PCUINIT, dl_fn->cu_init, "cuInit");
294 CHECK_LOAD_FUNC(PCUDEVICEGETCOUNT, dl_fn->cu_device_get_count, "cuDeviceGetCount");
295 CHECK_LOAD_FUNC(PCUDEVICEGET, dl_fn->cu_device_get, "cuDeviceGet");
296 CHECK_LOAD_FUNC(PCUDEVICEGETNAME, dl_fn->cu_device_get_name, "cuDeviceGetName");
297 CHECK_LOAD_FUNC(PCUDEVICECOMPUTECAPABILITY, dl_fn->cu_device_compute_capability, "cuDeviceComputeCapability");
298 CHECK_LOAD_FUNC(PCUCTXCREATE, dl_fn->cu_ctx_create, "cuCtxCreate_v2");
299 CHECK_LOAD_FUNC(PCUCTXPOPCURRENT, dl_fn->cu_ctx_pop_current, "cuCtxPopCurrent_v2");
300 CHECK_LOAD_FUNC(PCUCTXDESTROY, dl_fn->cu_ctx_destroy, "cuCtxDestroy_v2");
307 DL_CLOSE_FUNC(dl_fn->cuda_lib);
309 dl_fn->cuda_lib = NULL;
314 static av_cold int check_cuda_errors(AVCodecContext *avctx, CUresult err, const char *func)
316 if (err != CUDA_SUCCESS) {
317 av_log(avctx, AV_LOG_FATAL, ">> %s - failed with error code 0x%x\n", func, err);
322 #define check_cuda_errors(f) if (!check_cuda_errors(avctx, f, #f)) goto error
324 static av_cold int nvenc_check_cuda(AVCodecContext *avctx)
326 int device_count = 0;
327 CUdevice cu_device = 0;
329 int smminor = 0, smmajor = 0;
332 NvencContext *ctx = avctx->priv_data;
333 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
335 if (!nvenc_dyload_cuda(avctx))
338 if (dl_fn->nvenc_device_count > 0)
341 check_cuda_errors(dl_fn->cu_init(0));
343 check_cuda_errors(dl_fn->cu_device_get_count(&device_count));
346 av_log(avctx, AV_LOG_FATAL, "No CUDA capable devices found\n");
350 av_log(avctx, AV_LOG_VERBOSE, "%d CUDA capable devices found\n", device_count);
352 dl_fn->nvenc_device_count = 0;
354 for (i = 0; i < device_count; ++i) {
355 check_cuda_errors(dl_fn->cu_device_get(&cu_device, i));
356 check_cuda_errors(dl_fn->cu_device_get_name(gpu_name, sizeof(gpu_name), cu_device));
357 check_cuda_errors(dl_fn->cu_device_compute_capability(&smmajor, &smminor, cu_device));
359 smver = (smmajor << 4) | smminor;
361 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");
364 dl_fn->nvenc_devices[dl_fn->nvenc_device_count++] = cu_device;
367 if (!dl_fn->nvenc_device_count) {
368 av_log(avctx, AV_LOG_FATAL, "No NVENC capable devices found\n");
376 dl_fn->nvenc_device_count = 0;
381 static av_cold int nvenc_dyload_nvenc(AVCodecContext *avctx)
383 PNVENCODEAPICREATEINSTANCE nvEncodeAPICreateInstance = 0;
384 NVENCSTATUS nvstatus;
386 NvencContext *ctx = avctx->priv_data;
387 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
389 if (!nvenc_check_cuda(avctx))
392 if (dl_fn->nvenc_lib)
396 if (sizeof(void*) == 8) {
397 dl_fn->nvenc_lib = LoadLibrary(TEXT("nvEncodeAPI64.dll"));
399 dl_fn->nvenc_lib = LoadLibrary(TEXT("nvEncodeAPI.dll"));
402 dl_fn->nvenc_lib = dlopen("libnvidia-encode.so.1", RTLD_LAZY);
405 if (!dl_fn->nvenc_lib) {
406 av_log(avctx, AV_LOG_FATAL, "Failed loading the nvenc library\n");
410 nvEncodeAPICreateInstance = (PNVENCODEAPICREATEINSTANCE)LOAD_FUNC(dl_fn->nvenc_lib, "NvEncodeAPICreateInstance");
412 if (!nvEncodeAPICreateInstance) {
413 av_log(avctx, AV_LOG_FATAL, "Failed to load nvenc entrypoint\n");
417 dl_fn->nvenc_funcs.version = NV_ENCODE_API_FUNCTION_LIST_VER;
419 nvstatus = nvEncodeAPICreateInstance(&dl_fn->nvenc_funcs);
421 if (nvstatus != NV_ENC_SUCCESS) {
422 av_log(avctx, AV_LOG_FATAL, "Failed to create nvenc instance\n");
426 av_log(avctx, AV_LOG_VERBOSE, "Nvenc initialized successfully\n");
431 if (dl_fn->nvenc_lib)
432 DL_CLOSE_FUNC(dl_fn->nvenc_lib);
434 dl_fn->nvenc_lib = NULL;
439 static av_cold void nvenc_unload_nvenc(AVCodecContext *avctx)
441 NvencContext *ctx = avctx->priv_data;
442 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
444 DL_CLOSE_FUNC(dl_fn->nvenc_lib);
445 dl_fn->nvenc_lib = NULL;
447 dl_fn->nvenc_device_count = 0;
449 DL_CLOSE_FUNC(dl_fn->cuda_lib);
450 dl_fn->cuda_lib = NULL;
452 dl_fn->cu_init = NULL;
453 dl_fn->cu_device_get_count = NULL;
454 dl_fn->cu_device_get = NULL;
455 dl_fn->cu_device_get_name = NULL;
456 dl_fn->cu_device_compute_capability = NULL;
457 dl_fn->cu_ctx_create = NULL;
458 dl_fn->cu_ctx_pop_current = NULL;
459 dl_fn->cu_ctx_destroy = NULL;
461 av_log(avctx, AV_LOG_VERBOSE, "Nvenc unloaded\n");
464 static av_cold int nvenc_encode_init(AVCodecContext *avctx)
466 NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS encode_session_params = { 0 };
467 NV_ENC_PRESET_CONFIG preset_config = { 0 };
468 CUcontext cu_context_curr;
470 GUID encoder_preset = NV_ENC_PRESET_HQ_GUID;
471 NVENCSTATUS nv_status = NV_ENC_SUCCESS;
472 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;
576 ctx->init_encode_params.darHeight = avctx->height;
577 ctx->init_encode_params.darWidth = avctx->width;
578 ctx->init_encode_params.frameRateNum = avctx->time_base.den;
579 ctx->init_encode_params.frameRateDen = avctx->time_base.num * avctx->ticks_per_frame;
581 num_mbs = ((avctx->width + 15) >> 4) * ((avctx->height + 15) >> 4);
582 ctx->max_surface_count = (num_mbs >= 8160) ? 32 : 48;
584 ctx->init_encode_params.enableEncodeAsync = 0;
585 ctx->init_encode_params.enablePTD = 1;
587 ctx->init_encode_params.presetGUID = encoder_preset;
589 ctx->init_encode_params.encodeConfig = &ctx->encode_config;
590 memcpy(&ctx->encode_config, &preset_config.presetCfg, sizeof(ctx->encode_config));
591 ctx->encode_config.version = NV_ENC_CONFIG_VER;
593 if (avctx->gop_size >= 0) {
594 ctx->encode_config.gopLength = avctx->gop_size;
595 ctx->encode_config.encodeCodecConfig.h264Config.idrPeriod = avctx->gop_size;
598 if (avctx->bit_rate > 0)
599 ctx->encode_config.rcParams.averageBitRate = avctx->bit_rate;
601 if (avctx->rc_max_rate > 0)
602 ctx->encode_config.rcParams.maxBitRate = avctx->rc_max_rate;
606 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CBR;
607 } else if (ctx->twopass == 1 || isLL) {
608 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_2_PASS_QUALITY;
610 ctx->encode_config.encodeCodecConfig.h264Config.adaptiveTransformMode = NV_ENC_H264_ADAPTIVE_TRANSFORM_ENABLE;
611 ctx->encode_config.encodeCodecConfig.h264Config.fmoMode = NV_ENC_H264_FMO_DISABLE;
614 av_log(avctx, AV_LOG_WARNING, "Twopass mode is only known to work with low latency (ll, llhq, llhp) presets.\n");
616 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CBR;
618 } else if (avctx->global_quality > 0) {
619 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
620 ctx->encode_config.rcParams.constQP.qpInterB = avctx->global_quality;
621 ctx->encode_config.rcParams.constQP.qpInterP = avctx->global_quality;
622 ctx->encode_config.rcParams.constQP.qpIntra = avctx->global_quality;
626 } else if (avctx->qmin >= 0 && avctx->qmax >= 0) {
627 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_VBR;
629 ctx->encode_config.rcParams.enableMinQP = 1;
630 ctx->encode_config.rcParams.enableMaxQP = 1;
632 ctx->encode_config.rcParams.minQP.qpInterB = avctx->qmin;
633 ctx->encode_config.rcParams.minQP.qpInterP = avctx->qmin;
634 ctx->encode_config.rcParams.minQP.qpIntra = avctx->qmin;
636 ctx->encode_config.rcParams.maxQP.qpInterB = avctx->qmax;
637 ctx->encode_config.rcParams.maxQP.qpInterP = avctx->qmax;
638 ctx->encode_config.rcParams.maxQP.qpIntra = avctx->qmax;
641 if (avctx->rc_buffer_size > 0)
642 ctx->encode_config.rcParams.vbvBufferSize = avctx->rc_buffer_size;
644 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
645 ctx->encode_config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FIELD;
647 ctx->encode_config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FRAME;
650 switch (avctx->profile) {
651 case FF_PROFILE_H264_BASELINE:
652 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_BASELINE_GUID;
654 case FF_PROFILE_H264_MAIN:
655 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_MAIN_GUID;
657 case FF_PROFILE_H264_HIGH:
658 case FF_PROFILE_UNKNOWN:
659 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
662 av_log(avctx, AV_LOG_WARNING, "Unsupported h264 profile requested, falling back to high\n");
663 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
667 if (ctx->gobpattern >= 0) {
668 ctx->encode_config.frameIntervalP = ctx->gobpattern;
671 // when there're b frames, set dts offset
672 if (ctx->encode_config.frameIntervalP >= 2) {
676 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourDescriptionPresentFlag = 1;
677 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.videoSignalTypePresentFlag = 1;
679 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourMatrix = avctx->colorspace;
680 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourPrimaries = avctx->color_primaries;
681 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.transferCharacteristics = avctx->color_trc;
683 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.videoFullRangeFlag = avctx->color_range == AVCOL_RANGE_JPEG;
685 ctx->encode_config.encodeCodecConfig.h264Config.disableSPSPPS = (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
687 nv_status = p_nvenc->nvEncInitializeEncoder(ctx->nvencoder, &ctx->init_encode_params);
688 if (nv_status != NV_ENC_SUCCESS) {
689 av_log(avctx, AV_LOG_FATAL, "InitializeEncoder failed: 0x%x\n", (int)nv_status);
690 res = AVERROR_EXTERNAL;
694 ctx->input_surfaces = av_malloc(ctx->max_surface_count * sizeof(*ctx->input_surfaces));
696 if (!ctx->input_surfaces) {
697 res = AVERROR(ENOMEM);
701 ctx->output_surfaces = av_malloc(ctx->max_surface_count * sizeof(*ctx->output_surfaces));
703 if (!ctx->output_surfaces) {
704 res = AVERROR(ENOMEM);
708 for (surfaceCount = 0; surfaceCount < ctx->max_surface_count; ++surfaceCount) {
709 NV_ENC_CREATE_INPUT_BUFFER allocSurf = { 0 };
710 NV_ENC_CREATE_BITSTREAM_BUFFER allocOut = { 0 };
711 allocSurf.version = NV_ENC_CREATE_INPUT_BUFFER_VER;
712 allocOut.version = NV_ENC_CREATE_BITSTREAM_BUFFER_VER;
714 allocSurf.width = (avctx->width + 31) & ~31;
715 allocSurf.height = (avctx->height + 31) & ~31;
717 allocSurf.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_CACHED;
719 switch (avctx->pix_fmt) {
720 case AV_PIX_FMT_YUV420P:
721 allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_YV12_PL;
724 case AV_PIX_FMT_NV12:
725 allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_NV12_PL;
728 case AV_PIX_FMT_YUV444P:
729 allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_YUV444_PL;
733 av_log(avctx, AV_LOG_FATAL, "Invalid input pixel format\n");
734 res = AVERROR(EINVAL);
738 nv_status = p_nvenc->nvEncCreateInputBuffer(ctx->nvencoder, &allocSurf);
739 if (nv_status = NV_ENC_SUCCESS){
740 av_log(avctx, AV_LOG_FATAL, "CreateInputBuffer failed\n");
741 res = AVERROR_EXTERNAL;
745 ctx->input_surfaces[surfaceCount].lockCount = 0;
746 ctx->input_surfaces[surfaceCount].input_surface = allocSurf.inputBuffer;
747 ctx->input_surfaces[surfaceCount].format = allocSurf.bufferFmt;
748 ctx->input_surfaces[surfaceCount].width = allocSurf.width;
749 ctx->input_surfaces[surfaceCount].height = allocSurf.height;
751 /* 1MB is large enough to hold most output frames. NVENC increases this automaticaly if it's not enough. */
752 allocOut.size = 1024 * 1024;
754 allocOut.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_CACHED;
756 nv_status = p_nvenc->nvEncCreateBitstreamBuffer(ctx->nvencoder, &allocOut);
757 if (nv_status = NV_ENC_SUCCESS) {
758 av_log(avctx, AV_LOG_FATAL, "CreateBitstreamBuffer failed\n");
759 ctx->output_surfaces[surfaceCount++].output_surface = NULL;
760 res = AVERROR_EXTERNAL;
764 ctx->output_surfaces[surfaceCount].output_surface = allocOut.bitstreamBuffer;
765 ctx->output_surfaces[surfaceCount].size = allocOut.size;
766 ctx->output_surfaces[surfaceCount].busy = 0;
769 if (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) {
770 uint32_t outSize = 0;
772 NV_ENC_SEQUENCE_PARAM_PAYLOAD payload = { 0 };
773 payload.version = NV_ENC_SEQUENCE_PARAM_PAYLOAD_VER;
775 payload.spsppsBuffer = tmpHeader;
776 payload.inBufferSize = sizeof(tmpHeader);
777 payload.outSPSPPSPayloadSize = &outSize;
779 nv_status = p_nvenc->nvEncGetSequenceParams(ctx->nvencoder, &payload);
780 if (nv_status != NV_ENC_SUCCESS) {
781 av_log(avctx, AV_LOG_FATAL, "GetSequenceParams failed\n");
785 avctx->extradata_size = outSize;
786 avctx->extradata = av_mallocz(outSize + FF_INPUT_BUFFER_PADDING_SIZE);
788 if (!avctx->extradata) {
789 res = AVERROR(ENOMEM);
793 memcpy(avctx->extradata, tmpHeader, outSize);
796 if (ctx->encode_config.frameIntervalP > 1)
797 avctx->has_b_frames = 2;
799 if (ctx->encode_config.rcParams.averageBitRate > 0)
800 avctx->bit_rate = ctx->encode_config.rcParams.averageBitRate;
806 for (i = 0; i < surfaceCount; ++i) {
807 p_nvenc->nvEncDestroyInputBuffer(ctx->nvencoder, ctx->input_surfaces[i].input_surface);
808 if (ctx->output_surfaces[i].output_surface)
809 p_nvenc->nvEncDestroyBitstreamBuffer(ctx->nvencoder, ctx->output_surfaces[i].output_surface);
813 p_nvenc->nvEncDestroyEncoder(ctx->nvencoder);
816 dl_fn->cu_ctx_destroy(ctx->cu_context);
818 av_frame_free(&avctx->coded_frame);
820 nvenc_unload_nvenc(avctx);
822 ctx->nvencoder = NULL;
823 ctx->cu_context = NULL;
828 static av_cold int nvenc_encode_close(AVCodecContext *avctx)
830 NvencContext *ctx = avctx->priv_data;
831 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
832 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
835 av_freep(&ctx->timestamp_list.data);
836 av_freep(&ctx->output_surface_ready_queue.data);
837 av_freep(&ctx->output_surface_queue.data);
839 for (i = 0; i < ctx->max_surface_count; ++i) {
840 p_nvenc->nvEncDestroyInputBuffer(ctx->nvencoder, ctx->input_surfaces[i].input_surface);
841 p_nvenc->nvEncDestroyBitstreamBuffer(ctx->nvencoder, ctx->output_surfaces[i].output_surface);
843 ctx->max_surface_count = 0;
845 p_nvenc->nvEncDestroyEncoder(ctx->nvencoder);
846 ctx->nvencoder = NULL;
848 dl_fn->cu_ctx_destroy(ctx->cu_context);
849 ctx->cu_context = NULL;
851 nvenc_unload_nvenc(avctx);
853 av_frame_free(&avctx->coded_frame);
858 static int process_output_surface(AVCodecContext *avctx, AVPacket *pkt, AVFrame *coded_frame, NvencOutputSurface *tmpoutsurf)
860 NvencContext *ctx = avctx->priv_data;
861 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
862 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
864 uint32_t *slice_offsets = av_mallocz(ctx->encode_config.encodeCodecConfig.h264Config.sliceModeData * sizeof(*slice_offsets));
865 NV_ENC_LOCK_BITSTREAM lock_params = { 0 };
866 NVENCSTATUS nv_status;
870 return AVERROR(ENOMEM);
872 lock_params.version = NV_ENC_LOCK_BITSTREAM_VER;
874 lock_params.doNotWait = 0;
875 lock_params.outputBitstream = tmpoutsurf->output_surface;
876 lock_params.sliceOffsets = slice_offsets;
878 nv_status = p_nvenc->nvEncLockBitstream(ctx->nvencoder, &lock_params);
879 if (nv_status != NV_ENC_SUCCESS) {
880 av_log(avctx, AV_LOG_ERROR, "Failed locking bitstream buffer\n");
881 res = AVERROR_EXTERNAL;
885 if (res = ff_alloc_packet2(avctx, pkt, lock_params.bitstreamSizeInBytes)) {
886 p_nvenc->nvEncUnlockBitstream(ctx->nvencoder, tmpoutsurf->output_surface);
890 memcpy(pkt->data, lock_params.bitstreamBufferPtr, lock_params.bitstreamSizeInBytes);
892 nv_status = p_nvenc->nvEncUnlockBitstream(ctx->nvencoder, tmpoutsurf->output_surface);
893 if (nv_status != NV_ENC_SUCCESS)
894 av_log(avctx, AV_LOG_ERROR, "Failed unlocking bitstream buffer, expect the gates of mordor to open\n");
896 switch (lock_params.pictureType) {
897 case NV_ENC_PIC_TYPE_IDR:
898 pkt->flags |= AV_PKT_FLAG_KEY;
899 case NV_ENC_PIC_TYPE_I:
900 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
902 case NV_ENC_PIC_TYPE_P:
903 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_P;
905 case NV_ENC_PIC_TYPE_B:
906 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_B;
908 case NV_ENC_PIC_TYPE_BI:
909 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_BI;
912 av_log(avctx, AV_LOG_ERROR, "Unknown picture type encountered, expect the output to be broken.\n");
913 av_log(avctx, AV_LOG_ERROR, "Please report this error and include as much information on how to reproduce it as possible.\n");
914 res = AVERROR_EXTERNAL;
918 pkt->pts = lock_params.outputTimeStamp;
919 pkt->dts = timestamp_queue_dequeue(&ctx->timestamp_list);
921 // when there're b frame(s), set dts offset
922 if (ctx->encode_config.frameIntervalP >= 2)
925 if (pkt->dts > pkt->pts)
928 if (ctx->last_dts != AV_NOPTS_VALUE && pkt->dts <= ctx->last_dts)
929 pkt->dts = ctx->last_dts + 1;
931 ctx->last_dts = pkt->dts;
933 av_free(slice_offsets);
939 av_free(slice_offsets);
940 timestamp_queue_dequeue(&ctx->timestamp_list);
945 static int nvenc_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
946 const AVFrame *frame, int *got_packet)
948 NVENCSTATUS nv_status;
949 NvencOutputSurface *tmpoutsurf;
952 NvencContext *ctx = avctx->priv_data;
953 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
954 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
956 NV_ENC_PIC_PARAMS pic_params = { 0 };
957 pic_params.version = NV_ENC_PIC_PARAMS_VER;
960 NV_ENC_LOCK_INPUT_BUFFER lockBufferParams = { 0 };
961 NvencInputSurface *inSurf = NULL;
963 for (i = 0; i < ctx->max_surface_count; ++i) {
964 if (!ctx->input_surfaces[i].lockCount) {
965 inSurf = &ctx->input_surfaces[i];
972 inSurf->lockCount = 1;
974 lockBufferParams.version = NV_ENC_LOCK_INPUT_BUFFER_VER;
975 lockBufferParams.inputBuffer = inSurf->input_surface;
977 nv_status = p_nvenc->nvEncLockInputBuffer(ctx->nvencoder, &lockBufferParams);
978 if (nv_status != NV_ENC_SUCCESS) {
979 av_log(avctx, AV_LOG_ERROR, "Failed locking nvenc input buffer\n");
983 if (avctx->pix_fmt == AV_PIX_FMT_YUV420P) {
984 uint8_t *buf = lockBufferParams.bufferDataPtr;
986 av_image_copy_plane(buf, lockBufferParams.pitch,
987 frame->data[0], frame->linesize[0],
988 avctx->width, avctx->height);
990 buf += inSurf->height * lockBufferParams.pitch;
992 av_image_copy_plane(buf, lockBufferParams.pitch >> 1,
993 frame->data[2], frame->linesize[2],
994 avctx->width >> 1, avctx->height >> 1);
996 buf += (inSurf->height * lockBufferParams.pitch) >> 2;
998 av_image_copy_plane(buf, lockBufferParams.pitch >> 1,
999 frame->data[1], frame->linesize[1],
1000 avctx->width >> 1, avctx->height >> 1);
1001 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12) {
1002 uint8_t *buf = lockBufferParams.bufferDataPtr;
1004 av_image_copy_plane(buf, lockBufferParams.pitch,
1005 frame->data[0], frame->linesize[0],
1006 avctx->width, avctx->height);
1008 buf += inSurf->height * lockBufferParams.pitch;
1010 av_image_copy_plane(buf, lockBufferParams.pitch,
1011 frame->data[1], frame->linesize[1],
1012 avctx->width, avctx->height >> 1);
1013 } else if (avctx->pix_fmt == AV_PIX_FMT_YUV444P) {
1014 uint8_t *buf = lockBufferParams.bufferDataPtr;
1016 av_image_copy_plane(buf, lockBufferParams.pitch,
1017 frame->data[0], frame->linesize[0],
1018 avctx->width, avctx->height);
1020 buf += inSurf->height * lockBufferParams.pitch;
1022 av_image_copy_plane(buf, lockBufferParams.pitch,
1023 frame->data[1], frame->linesize[1],
1024 avctx->width, avctx->height);
1026 buf += inSurf->height * lockBufferParams.pitch;
1028 av_image_copy_plane(buf, lockBufferParams.pitch,
1029 frame->data[2], frame->linesize[2],
1030 avctx->width, avctx->height);
1032 av_log(avctx, AV_LOG_FATAL, "Invalid pixel format!\n");
1033 return AVERROR(EINVAL);
1036 nv_status = p_nvenc->nvEncUnlockInputBuffer(ctx->nvencoder, inSurf->input_surface);
1037 if (nv_status != NV_ENC_SUCCESS) {
1038 av_log(avctx, AV_LOG_FATAL, "Failed unlocking input buffer!\n");
1039 return AVERROR_EXTERNAL;
1042 for (i = 0; i < ctx->max_surface_count; ++i)
1043 if (!ctx->output_surfaces[i].busy)
1046 if (i == ctx->max_surface_count) {
1047 inSurf->lockCount = 0;
1048 av_log(avctx, AV_LOG_FATAL, "No free output surface found!\n");
1049 return AVERROR_EXTERNAL;
1052 ctx->output_surfaces[i].input_surface = inSurf;
1054 pic_params.inputBuffer = inSurf->input_surface;
1055 pic_params.bufferFmt = inSurf->format;
1056 pic_params.inputWidth = avctx->width;
1057 pic_params.inputHeight = avctx->height;
1058 pic_params.outputBitstream = ctx->output_surfaces[i].output_surface;
1059 pic_params.completionEvent = 0;
1061 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
1062 if (frame->top_field_first) {
1063 pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_TOP_BOTTOM;
1065 pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_BOTTOM_TOP;
1068 pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
1071 pic_params.encodePicFlags = 0;
1072 pic_params.inputTimeStamp = frame->pts;
1073 pic_params.inputDuration = 0;
1074 pic_params.codecPicParams.h264PicParams.sliceMode = ctx->encode_config.encodeCodecConfig.h264Config.sliceMode;
1075 pic_params.codecPicParams.h264PicParams.sliceModeData = ctx->encode_config.encodeCodecConfig.h264Config.sliceModeData;
1077 #if NVENCAPI_MAJOR_VERSION < 5
1078 memcpy(&pic_params.rcParams, &ctx->encode_config.rcParams, sizeof(NV_ENC_RC_PARAMS));
1081 res = timestamp_queue_enqueue(&ctx->timestamp_list, frame->pts);
1086 pic_params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
1089 nv_status = p_nvenc->nvEncEncodePicture(ctx->nvencoder, &pic_params);
1091 if (frame && nv_status == NV_ENC_ERR_NEED_MORE_INPUT) {
1092 res = out_surf_queue_enqueue(&ctx->output_surface_queue, &ctx->output_surfaces[i]);
1097 ctx->output_surfaces[i].busy = 1;
1100 if (nv_status != NV_ENC_SUCCESS && nv_status != NV_ENC_ERR_NEED_MORE_INPUT) {
1101 av_log(avctx, AV_LOG_ERROR, "EncodePicture failed!\n");
1102 return AVERROR_EXTERNAL;
1105 if (nv_status != NV_ENC_ERR_NEED_MORE_INPUT) {
1106 while (ctx->output_surface_queue.count) {
1107 tmpoutsurf = out_surf_queue_dequeue(&ctx->output_surface_queue);
1108 res = out_surf_queue_enqueue(&ctx->output_surface_ready_queue, tmpoutsurf);
1115 res = out_surf_queue_enqueue(&ctx->output_surface_ready_queue, &ctx->output_surfaces[i]);
1120 ctx->output_surfaces[i].busy = 1;
1124 if (ctx->output_surface_ready_queue.count) {
1125 tmpoutsurf = out_surf_queue_dequeue(&ctx->output_surface_ready_queue);
1127 res = process_output_surface(avctx, pkt, avctx->coded_frame, tmpoutsurf);
1132 tmpoutsurf->busy = 0;
1133 av_assert0(tmpoutsurf->input_surface->lockCount);
1134 tmpoutsurf->input_surface->lockCount--;
1144 static enum AVPixelFormat pix_fmts_nvenc[] = {
1149 #define OFFSET(x) offsetof(NvencContext, x)
1150 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1151 static const AVOption options[] = {
1152 { "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 },
1153 { "cbr", "Use cbr encoding mode", OFFSET(cbr), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1154 { "2pass", "Use 2pass cbr encoding mode (low latency mode only)", OFFSET(twopass), AV_OPT_TYPE_INT, { .i64 = -1 }, -1, 1, VE },
1155 { "goppattern", "Specifies the GOP pattern as follows: 0: I, 1: IPP, 2: IBP, 3: IBBP", OFFSET(gobpattern), AV_OPT_TYPE_INT, { .i64 = -1 }, -1, 3, VE },
1156 { "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 },
1160 static const AVClass nvenc_class = {
1161 .class_name = "nvenc",
1162 .item_name = av_default_item_name,
1164 .version = LIBAVUTIL_VERSION_INT,
1167 static const AVCodecDefault nvenc_defaults[] = {
1177 AVCodec ff_nvenc_encoder = {
1179 .long_name = NULL_IF_CONFIG_SMALL("Nvidia NVENC h264 encoder"),
1180 .type = AVMEDIA_TYPE_VIDEO,
1181 .id = AV_CODEC_ID_H264,
1182 .priv_data_size = sizeof(NvencContext),
1183 .init = nvenc_encode_init,
1184 .encode2 = nvenc_encode_frame,
1185 .close = nvenc_encode_close,
1186 .capabilities = CODEC_CAP_DELAY,
1187 .priv_class = &nvenc_class,
1188 .defaults = nvenc_defaults,
1189 .pix_fmts = pix_fmts_nvenc,