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
22 #if defined(_WIN32) && !defined(__CYGWIN__)
28 /* NVENC API is stdcall in cygwin, as it's still Windows */
29 #if defined(__CYGWIN__) && !defined(_WIN32)
34 #include <nvEncodeAPI.h>
36 #if defined(TEMP_WIN32)
41 #include "libavutil/internal.h"
42 #include "libavutil/imgutils.h"
43 #include "libavutil/avassert.h"
44 #include "libavutil/opt.h"
45 #include "libavutil/mem.h"
50 #if defined(_WIN32) || defined(__CYGWIN__)
51 #define CUDAAPI __stdcall
56 #if defined(_WIN32) && !defined(__CYGWIN__)
57 #define LOAD_FUNC(l, s) GetProcAddress(l, s)
58 #define DL_CLOSE_FUNC(l) FreeLibrary(l)
60 #define LOAD_FUNC(l, s) dlsym(l, s)
61 #define DL_CLOSE_FUNC(l) dlclose(l)
64 typedef enum cudaError_enum {
68 typedef void* CUcontext;
70 typedef CUresult(CUDAAPI *PCUINIT)(unsigned int Flags);
71 typedef CUresult(CUDAAPI *PCUDEVICEGETCOUNT)(int *count);
72 typedef CUresult(CUDAAPI *PCUDEVICEGET)(CUdevice *device, int ordinal);
73 typedef CUresult(CUDAAPI *PCUDEVICEGETNAME)(char *name, int len, CUdevice dev);
74 typedef CUresult(CUDAAPI *PCUDEVICECOMPUTECAPABILITY)(int *major, int *minor, CUdevice dev);
75 typedef CUresult(CUDAAPI *PCUCTXCREATE)(CUcontext *pctx, unsigned int flags, CUdevice dev);
76 typedef CUresult(CUDAAPI *PCUCTXPOPCURRENT)(CUcontext *pctx);
77 typedef CUresult(CUDAAPI *PCUCTXDESTROY)(CUcontext ctx);
79 typedef NVENCSTATUS (NVENCAPI* PNVENCODEAPICREATEINSTANCE)(NV_ENCODE_API_FUNCTION_LIST *functionList);
81 static const GUID dummy_license = { 0x0, 0x0, 0x0, { 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 } };
83 typedef struct NvencInputSurface
85 NV_ENC_INPUT_PTR input_surface;
91 NV_ENC_BUFFER_FORMAT format;
94 typedef struct NvencOutputSurface
96 NV_ENC_OUTPUT_PTR output_surface;
99 NvencInputSurface* input_surface;
102 } NvencOutputSurface;
104 typedef struct NvencData
108 NvencOutputSurface *surface;
112 typedef struct NvencDataList
121 typedef struct NvencDynLoadFunctions
124 PCUDEVICEGETCOUNT cu_device_get_count;
125 PCUDEVICEGET cu_device_get;
126 PCUDEVICEGETNAME cu_device_get_name;
127 PCUDEVICECOMPUTECAPABILITY cu_device_compute_capability;
128 PCUCTXCREATE cu_ctx_create;
129 PCUCTXPOPCURRENT cu_ctx_pop_current;
130 PCUCTXDESTROY cu_ctx_destroy;
132 NV_ENCODE_API_FUNCTION_LIST nvenc_funcs;
133 int nvenc_device_count;
134 CUdevice nvenc_devices[16];
136 #if defined(_WIN32) && !defined(__CYGWIN__)
143 } NvencDynLoadFunctions;
145 typedef struct NvencContext
149 NvencDynLoadFunctions nvenc_dload_funcs;
151 NV_ENC_INITIALIZE_PARAMS init_encode_params;
152 NV_ENC_CONFIG encode_config;
153 CUcontext cu_context;
155 int max_surface_count;
156 NvencInputSurface *input_surfaces;
157 NvencOutputSurface *output_surfaces;
159 NvencDataList output_surface_queue;
160 NvencDataList output_surface_ready_queue;
161 NvencDataList timestamp_list;
173 static NvencData* data_queue_dequeue(NvencDataList* queue)
179 av_assert0(queue->size);
180 av_assert0(queue->data);
185 /* Size always is a multiple of two */
186 mask = queue->size - 1;
187 read_pos = (queue->pos - queue->count) & mask;
190 return &queue->data[read_pos];
193 static int data_queue_enqueue(NvencDataList* queue, NvencData *data)
195 NvencDataList new_queue;
200 /* size always has to be a multiple of two */
205 queue->data = av_malloc(queue->size * sizeof(*(queue->data)));
209 return AVERROR(ENOMEM);
213 if (queue->count == queue->size) {
214 new_queue.size = queue->size << 1;
217 new_queue.data = av_malloc(new_queue.size * sizeof(*(queue->data)));
220 return AVERROR(ENOMEM);
222 while (tmp_data = data_queue_dequeue(queue))
223 data_queue_enqueue(&new_queue, tmp_data);
225 av_free(queue->data);
229 mask = queue->size - 1;
231 queue->data[queue->pos] = *data;
232 queue->pos = (queue->pos + 1) & mask;
238 static int out_surf_queue_enqueue(NvencDataList* queue, NvencOutputSurface* surface)
241 data.surface = surface;
243 return data_queue_enqueue(queue, &data);
246 static NvencOutputSurface* out_surf_queue_dequeue(NvencDataList* queue)
248 NvencData* res = data_queue_dequeue(queue);
256 static int timestamp_queue_enqueue(NvencDataList* queue, int64_t timestamp)
259 data.timestamp = timestamp;
261 return data_queue_enqueue(queue, &data);
264 static int64_t timestamp_queue_dequeue(NvencDataList* queue)
266 NvencData* res = data_queue_dequeue(queue);
269 return AV_NOPTS_VALUE;
271 return res->timestamp;
274 #define CHECK_LOAD_FUNC(t, f, s) \
276 (f) = (t)LOAD_FUNC(dl_fn->cuda_lib, s); \
278 av_log(avctx, AV_LOG_FATAL, "Failed loading %s from CUDA library\n", s); \
283 static av_cold int nvenc_dyload_cuda(AVCodecContext *avctx)
285 NvencContext *ctx = avctx->priv_data;
286 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
292 dl_fn->cuda_lib = LoadLibrary(TEXT("nvcuda.dll"));
293 #elif defined(__CYGWIN__)
294 dl_fn->cuda_lib = dlopen("nvcuda.dll", RTLD_LAZY);
296 dl_fn->cuda_lib = dlopen("libcuda.so", RTLD_LAZY);
299 if (!dl_fn->cuda_lib) {
300 av_log(avctx, AV_LOG_FATAL, "Failed loading CUDA library\n");
304 CHECK_LOAD_FUNC(PCUINIT, dl_fn->cu_init, "cuInit");
305 CHECK_LOAD_FUNC(PCUDEVICEGETCOUNT, dl_fn->cu_device_get_count, "cuDeviceGetCount");
306 CHECK_LOAD_FUNC(PCUDEVICEGET, dl_fn->cu_device_get, "cuDeviceGet");
307 CHECK_LOAD_FUNC(PCUDEVICEGETNAME, dl_fn->cu_device_get_name, "cuDeviceGetName");
308 CHECK_LOAD_FUNC(PCUDEVICECOMPUTECAPABILITY, dl_fn->cu_device_compute_capability, "cuDeviceComputeCapability");
309 CHECK_LOAD_FUNC(PCUCTXCREATE, dl_fn->cu_ctx_create, "cuCtxCreate_v2");
310 CHECK_LOAD_FUNC(PCUCTXPOPCURRENT, dl_fn->cu_ctx_pop_current, "cuCtxPopCurrent_v2");
311 CHECK_LOAD_FUNC(PCUCTXDESTROY, dl_fn->cu_ctx_destroy, "cuCtxDestroy_v2");
318 DL_CLOSE_FUNC(dl_fn->cuda_lib);
320 dl_fn->cuda_lib = NULL;
325 static av_cold int check_cuda_errors(AVCodecContext *avctx, CUresult err, const char *func)
327 if (err != CUDA_SUCCESS) {
328 av_log(avctx, AV_LOG_FATAL, ">> %s - failed with error code 0x%x\n", func, err);
333 #define check_cuda_errors(f) if (!check_cuda_errors(avctx, f, #f)) goto error
335 static av_cold int nvenc_check_cuda(AVCodecContext *avctx)
337 int device_count = 0;
338 CUdevice cu_device = 0;
340 int smminor = 0, smmajor = 0;
343 NvencContext *ctx = avctx->priv_data;
344 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
346 if (!nvenc_dyload_cuda(avctx))
349 if (dl_fn->nvenc_device_count > 0)
352 check_cuda_errors(dl_fn->cu_init(0));
354 check_cuda_errors(dl_fn->cu_device_get_count(&device_count));
357 av_log(avctx, AV_LOG_FATAL, "No CUDA capable devices found\n");
361 av_log(avctx, AV_LOG_VERBOSE, "%d CUDA capable devices found\n", device_count);
363 dl_fn->nvenc_device_count = 0;
365 for (i = 0; i < device_count; ++i) {
366 check_cuda_errors(dl_fn->cu_device_get(&cu_device, i));
367 check_cuda_errors(dl_fn->cu_device_get_name(gpu_name, sizeof(gpu_name), cu_device));
368 check_cuda_errors(dl_fn->cu_device_compute_capability(&smmajor, &smminor, cu_device));
370 smver = (smmajor << 4) | smminor;
372 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");
375 dl_fn->nvenc_devices[dl_fn->nvenc_device_count++] = cu_device;
378 if (!dl_fn->nvenc_device_count) {
379 av_log(avctx, AV_LOG_FATAL, "No NVENC capable devices found\n");
387 dl_fn->nvenc_device_count = 0;
392 static av_cold int nvenc_dyload_nvenc(AVCodecContext *avctx)
394 PNVENCODEAPICREATEINSTANCE nvEncodeAPICreateInstance = 0;
395 NVENCSTATUS nvstatus;
397 NvencContext *ctx = avctx->priv_data;
398 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
400 if (!nvenc_check_cuda(avctx))
403 if (dl_fn->nvenc_lib)
407 if (sizeof(void*) == 8) {
408 dl_fn->nvenc_lib = LoadLibrary(TEXT("nvEncodeAPI64.dll"));
410 dl_fn->nvenc_lib = LoadLibrary(TEXT("nvEncodeAPI.dll"));
412 #elif defined(__CYGWIN__)
413 if (sizeof(void*) == 8) {
414 dl_fn->nvenc_lib = dlopen("nvEncodeAPI64.dll", RTLD_LAZY);
416 dl_fn->nvenc_lib = dlopen("nvEncodeAPI.dll", RTLD_LAZY);
419 dl_fn->nvenc_lib = dlopen("libnvidia-encode.so.1", RTLD_LAZY);
422 if (!dl_fn->nvenc_lib) {
423 av_log(avctx, AV_LOG_FATAL, "Failed loading the nvenc library\n");
427 nvEncodeAPICreateInstance = (PNVENCODEAPICREATEINSTANCE)LOAD_FUNC(dl_fn->nvenc_lib, "NvEncodeAPICreateInstance");
429 if (!nvEncodeAPICreateInstance) {
430 av_log(avctx, AV_LOG_FATAL, "Failed to load nvenc entrypoint\n");
434 dl_fn->nvenc_funcs.version = NV_ENCODE_API_FUNCTION_LIST_VER;
436 nvstatus = nvEncodeAPICreateInstance(&dl_fn->nvenc_funcs);
438 if (nvstatus != NV_ENC_SUCCESS) {
439 av_log(avctx, AV_LOG_FATAL, "Failed to create nvenc instance\n");
443 av_log(avctx, AV_LOG_VERBOSE, "Nvenc initialized successfully\n");
448 if (dl_fn->nvenc_lib)
449 DL_CLOSE_FUNC(dl_fn->nvenc_lib);
451 dl_fn->nvenc_lib = NULL;
456 static av_cold void nvenc_unload_nvenc(AVCodecContext *avctx)
458 NvencContext *ctx = avctx->priv_data;
459 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
461 DL_CLOSE_FUNC(dl_fn->nvenc_lib);
462 dl_fn->nvenc_lib = NULL;
464 dl_fn->nvenc_device_count = 0;
466 DL_CLOSE_FUNC(dl_fn->cuda_lib);
467 dl_fn->cuda_lib = NULL;
469 dl_fn->cu_init = NULL;
470 dl_fn->cu_device_get_count = NULL;
471 dl_fn->cu_device_get = NULL;
472 dl_fn->cu_device_get_name = NULL;
473 dl_fn->cu_device_compute_capability = NULL;
474 dl_fn->cu_ctx_create = NULL;
475 dl_fn->cu_ctx_pop_current = NULL;
476 dl_fn->cu_ctx_destroy = NULL;
478 av_log(avctx, AV_LOG_VERBOSE, "Nvenc unloaded\n");
481 static av_cold int nvenc_encode_init(AVCodecContext *avctx)
483 NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS encode_session_params = { 0 };
484 NV_ENC_PRESET_CONFIG preset_config = { 0 };
485 CUcontext cu_context_curr;
487 GUID encoder_preset = NV_ENC_PRESET_HQ_GUID;
488 GUID license = dummy_license;
489 NVENCSTATUS nv_status = NV_ENC_SUCCESS;
490 int surfaceCount = 0;
495 NvencContext *ctx = avctx->priv_data;
496 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
497 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
499 if (!nvenc_dyload_nvenc(avctx))
500 return AVERROR_EXTERNAL;
502 avctx->coded_frame = av_frame_alloc();
503 if (!avctx->coded_frame) {
504 res = AVERROR(ENOMEM);
508 ctx->last_dts = AV_NOPTS_VALUE;
510 ctx->encode_config.version = NV_ENC_CONFIG_VER;
511 ctx->init_encode_params.version = NV_ENC_INITIALIZE_PARAMS_VER;
512 preset_config.version = NV_ENC_PRESET_CONFIG_VER;
513 preset_config.presetCfg.version = NV_ENC_CONFIG_VER;
514 encode_session_params.version = NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER;
515 encode_session_params.apiVersion = NVENCAPI_VERSION;
516 encode_session_params.clientKeyPtr = &license;
518 if (ctx->gpu >= dl_fn->nvenc_device_count) {
519 av_log(avctx, AV_LOG_FATAL, "Requested GPU %d, but only %d GPUs are available!\n", ctx->gpu, dl_fn->nvenc_device_count);
520 res = AVERROR(EINVAL);
524 ctx->cu_context = NULL;
525 cu_res = dl_fn->cu_ctx_create(&ctx->cu_context, 0, dl_fn->nvenc_devices[ctx->gpu]);
527 if (cu_res != CUDA_SUCCESS) {
528 av_log(avctx, AV_LOG_FATAL, "Failed creating CUDA context for NVENC: 0x%x\n", (int)cu_res);
529 res = AVERROR_EXTERNAL;
533 cu_res = dl_fn->cu_ctx_pop_current(&cu_context_curr);
535 if (cu_res != CUDA_SUCCESS) {
536 av_log(avctx, AV_LOG_FATAL, "Failed popping CUDA context: 0x%x\n", (int)cu_res);
537 res = AVERROR_EXTERNAL;
541 encode_session_params.device = ctx->cu_context;
542 encode_session_params.deviceType = NV_ENC_DEVICE_TYPE_CUDA;
544 nv_status = p_nvenc->nvEncOpenEncodeSessionEx(&encode_session_params, &ctx->nvencoder);
545 if (nv_status != NV_ENC_SUCCESS) {
546 ctx->nvencoder = NULL;
547 av_log(avctx, AV_LOG_FATAL, "OpenEncodeSessionEx failed: 0x%x - invalid license key?\n", (int)nv_status);
548 res = AVERROR_EXTERNAL;
553 if (!strcmp(ctx->preset, "hp")) {
554 encoder_preset = NV_ENC_PRESET_HP_GUID;
555 } else if (!strcmp(ctx->preset, "hq")) {
556 encoder_preset = NV_ENC_PRESET_HQ_GUID;
557 } else if (!strcmp(ctx->preset, "bd")) {
558 encoder_preset = NV_ENC_PRESET_BD_GUID;
559 } else if (!strcmp(ctx->preset, "ll")) {
560 encoder_preset = NV_ENC_PRESET_LOW_LATENCY_DEFAULT_GUID;
562 } else if (!strcmp(ctx->preset, "llhp")) {
563 encoder_preset = NV_ENC_PRESET_LOW_LATENCY_HP_GUID;
565 } else if (!strcmp(ctx->preset, "llhq")) {
566 encoder_preset = NV_ENC_PRESET_LOW_LATENCY_HQ_GUID;
568 } else if (!strcmp(ctx->preset, "default")) {
569 encoder_preset = NV_ENC_PRESET_DEFAULT_GUID;
571 av_log(avctx, AV_LOG_FATAL, "Preset \"%s\" is unknown! Supported presets: hp, hq, bd, ll, llhp, llhq, default\n", ctx->preset);
572 res = AVERROR(EINVAL);
577 nv_status = p_nvenc->nvEncGetEncodePresetConfig(ctx->nvencoder, NV_ENC_CODEC_H264_GUID, encoder_preset, &preset_config);
578 if (nv_status != NV_ENC_SUCCESS) {
579 av_log(avctx, AV_LOG_FATAL, "GetEncodePresetConfig failed: 0x%x\n", (int)nv_status);
580 res = AVERROR_EXTERNAL;
584 ctx->init_encode_params.encodeGUID = NV_ENC_CODEC_H264_GUID;
585 ctx->init_encode_params.encodeHeight = avctx->height;
586 ctx->init_encode_params.encodeWidth = avctx->width;
587 ctx->init_encode_params.darHeight = avctx->height;
588 ctx->init_encode_params.darWidth = avctx->width;
589 ctx->init_encode_params.frameRateNum = avctx->time_base.den;
590 ctx->init_encode_params.frameRateDen = avctx->time_base.num * avctx->ticks_per_frame;
592 num_mbs = ((avctx->width + 15) >> 4) * ((avctx->height + 15) >> 4);
593 ctx->max_surface_count = (num_mbs >= 8160) ? 32 : 48;
595 ctx->init_encode_params.enableEncodeAsync = 0;
596 ctx->init_encode_params.enablePTD = 1;
598 ctx->init_encode_params.presetGUID = encoder_preset;
600 ctx->init_encode_params.encodeConfig = &ctx->encode_config;
601 memcpy(&ctx->encode_config, &preset_config.presetCfg, sizeof(ctx->encode_config));
602 ctx->encode_config.version = NV_ENC_CONFIG_VER;
604 if (avctx->gop_size >= 0) {
605 ctx->encode_config.gopLength = avctx->gop_size;
606 ctx->encode_config.encodeCodecConfig.h264Config.idrPeriod = avctx->gop_size;
609 if (avctx->bit_rate > 0)
610 ctx->encode_config.rcParams.averageBitRate = avctx->bit_rate;
612 if (avctx->rc_max_rate > 0)
613 ctx->encode_config.rcParams.maxBitRate = avctx->rc_max_rate;
617 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CBR;
618 } else if (ctx->twopass == 1 || isLL) {
619 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_2_PASS_QUALITY;
621 ctx->encode_config.encodeCodecConfig.h264Config.adaptiveTransformMode = NV_ENC_H264_ADAPTIVE_TRANSFORM_ENABLE;
622 ctx->encode_config.encodeCodecConfig.h264Config.fmoMode = NV_ENC_H264_FMO_DISABLE;
625 av_log(avctx, AV_LOG_WARNING, "Twopass mode is only known to work with low latency (ll, llhq, llhp) presets.\n");
627 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CBR;
629 } else if (avctx->global_quality > 0) {
630 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
631 ctx->encode_config.rcParams.constQP.qpInterB = avctx->global_quality;
632 ctx->encode_config.rcParams.constQP.qpInterP = avctx->global_quality;
633 ctx->encode_config.rcParams.constQP.qpIntra = avctx->global_quality;
637 } else if (avctx->qmin >= 0 && avctx->qmax >= 0) {
638 ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_VBR;
640 ctx->encode_config.rcParams.enableMinQP = 1;
641 ctx->encode_config.rcParams.enableMaxQP = 1;
643 ctx->encode_config.rcParams.minQP.qpInterB = avctx->qmin;
644 ctx->encode_config.rcParams.minQP.qpInterP = avctx->qmin;
645 ctx->encode_config.rcParams.minQP.qpIntra = avctx->qmin;
647 ctx->encode_config.rcParams.maxQP.qpInterB = avctx->qmax;
648 ctx->encode_config.rcParams.maxQP.qpInterP = avctx->qmax;
649 ctx->encode_config.rcParams.maxQP.qpIntra = avctx->qmax;
652 if (avctx->rc_buffer_size > 0)
653 ctx->encode_config.rcParams.vbvBufferSize = avctx->rc_buffer_size;
655 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
656 ctx->encode_config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FIELD;
658 ctx->encode_config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FRAME;
661 switch (avctx->profile) {
662 case FF_PROFILE_H264_BASELINE:
663 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_BASELINE_GUID;
665 case FF_PROFILE_H264_MAIN:
666 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_MAIN_GUID;
668 case FF_PROFILE_H264_HIGH:
669 case FF_PROFILE_UNKNOWN:
670 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
673 av_log(avctx, AV_LOG_WARNING, "Unsupported h264 profile requested, falling back to high\n");
674 ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
678 if (ctx->gobpattern >= 0) {
679 ctx->encode_config.frameIntervalP = 1;
682 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourDescriptionPresentFlag = 1;
683 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.videoSignalTypePresentFlag = 1;
685 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourMatrix = avctx->colorspace;
686 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourPrimaries = avctx->color_primaries;
687 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.transferCharacteristics = avctx->color_trc;
689 ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.videoFullRangeFlag = avctx->color_range == AVCOL_RANGE_JPEG;
691 ctx->encode_config.encodeCodecConfig.h264Config.disableSPSPPS = (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
693 nv_status = p_nvenc->nvEncInitializeEncoder(ctx->nvencoder, &ctx->init_encode_params);
694 if (nv_status != NV_ENC_SUCCESS) {
695 av_log(avctx, AV_LOG_FATAL, "InitializeEncoder failed: 0x%x\n", (int)nv_status);
696 res = AVERROR_EXTERNAL;
700 ctx->input_surfaces = av_malloc(ctx->max_surface_count * sizeof(*ctx->input_surfaces));
702 if (!ctx->input_surfaces) {
703 res = AVERROR(ENOMEM);
707 ctx->output_surfaces = av_malloc(ctx->max_surface_count * sizeof(*ctx->output_surfaces));
709 if (!ctx->output_surfaces) {
710 res = AVERROR(ENOMEM);
714 for (surfaceCount = 0; surfaceCount < ctx->max_surface_count; ++surfaceCount) {
715 NV_ENC_CREATE_INPUT_BUFFER allocSurf = { 0 };
716 NV_ENC_CREATE_BITSTREAM_BUFFER allocOut = { 0 };
717 allocSurf.version = NV_ENC_CREATE_INPUT_BUFFER_VER;
718 allocOut.version = NV_ENC_CREATE_BITSTREAM_BUFFER_VER;
720 allocSurf.width = (avctx->width + 31) & ~31;
721 allocSurf.height = (avctx->height + 31) & ~31;
723 allocSurf.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_CACHED;
725 switch (avctx->pix_fmt) {
726 case AV_PIX_FMT_YUV420P:
727 allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_YV12_PL;
730 case AV_PIX_FMT_NV12:
731 allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_NV12_PL;
734 case AV_PIX_FMT_YUV444P:
735 allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_YUV444_PL;
739 av_log(avctx, AV_LOG_FATAL, "Invalid input pixel format\n");
740 res = AVERROR(EINVAL);
744 nv_status = p_nvenc->nvEncCreateInputBuffer(ctx->nvencoder, &allocSurf);
745 if (nv_status = NV_ENC_SUCCESS){
746 av_log(avctx, AV_LOG_FATAL, "CreateInputBuffer failed\n");
747 res = AVERROR_EXTERNAL;
751 ctx->input_surfaces[surfaceCount].lockCount = 0;
752 ctx->input_surfaces[surfaceCount].input_surface = allocSurf.inputBuffer;
753 ctx->input_surfaces[surfaceCount].format = allocSurf.bufferFmt;
754 ctx->input_surfaces[surfaceCount].width = allocSurf.width;
755 ctx->input_surfaces[surfaceCount].height = allocSurf.height;
757 /* 1MB is large enough to hold most output frames. NVENC increases this automaticaly if it's not enough. */
758 allocOut.size = 1024 * 1024;
760 allocOut.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_CACHED;
762 nv_status = p_nvenc->nvEncCreateBitstreamBuffer(ctx->nvencoder, &allocOut);
763 if (nv_status = NV_ENC_SUCCESS) {
764 av_log(avctx, AV_LOG_FATAL, "CreateBitstreamBuffer failed\n");
765 ctx->output_surfaces[surfaceCount++].output_surface = NULL;
766 res = AVERROR_EXTERNAL;
770 ctx->output_surfaces[surfaceCount].output_surface = allocOut.bitstreamBuffer;
771 ctx->output_surfaces[surfaceCount].size = allocOut.size;
772 ctx->output_surfaces[surfaceCount].busy = 0;
775 if (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) {
776 uint32_t outSize = 0;
778 NV_ENC_SEQUENCE_PARAM_PAYLOAD payload = { 0 };
779 payload.version = NV_ENC_SEQUENCE_PARAM_PAYLOAD_VER;
781 payload.spsppsBuffer = tmpHeader;
782 payload.inBufferSize = sizeof(tmpHeader);
783 payload.outSPSPPSPayloadSize = &outSize;
785 nv_status = p_nvenc->nvEncGetSequenceParams(ctx->nvencoder, &payload);
786 if (nv_status != NV_ENC_SUCCESS) {
787 av_log(avctx, AV_LOG_FATAL, "GetSequenceParams failed\n");
791 avctx->extradata_size = outSize;
792 avctx->extradata = av_mallocz(outSize + FF_INPUT_BUFFER_PADDING_SIZE);
794 if (!avctx->extradata) {
795 res = AVERROR(ENOMEM);
799 memcpy(avctx->extradata, tmpHeader, outSize);
802 if (ctx->encode_config.frameIntervalP > 1)
803 avctx->has_b_frames = 2;
805 if (ctx->encode_config.rcParams.averageBitRate > 0)
806 avctx->bit_rate = ctx->encode_config.rcParams.averageBitRate;
812 for (i = 0; i < surfaceCount; ++i) {
813 p_nvenc->nvEncDestroyInputBuffer(ctx->nvencoder, ctx->input_surfaces[i].input_surface);
814 if (ctx->output_surfaces[i].output_surface)
815 p_nvenc->nvEncDestroyBitstreamBuffer(ctx->nvencoder, ctx->output_surfaces[i].output_surface);
819 p_nvenc->nvEncDestroyEncoder(ctx->nvencoder);
822 dl_fn->cu_ctx_destroy(ctx->cu_context);
824 av_frame_free(&avctx->coded_frame);
826 nvenc_unload_nvenc(avctx);
828 ctx->nvencoder = NULL;
829 ctx->cu_context = NULL;
834 static av_cold int nvenc_encode_close(AVCodecContext *avctx)
836 NvencContext *ctx = avctx->priv_data;
837 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
838 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
841 av_freep(&ctx->timestamp_list.data);
842 av_freep(&ctx->output_surface_ready_queue.data);
843 av_freep(&ctx->output_surface_queue.data);
845 for (i = 0; i < ctx->max_surface_count; ++i) {
846 p_nvenc->nvEncDestroyInputBuffer(ctx->nvencoder, ctx->input_surfaces[i].input_surface);
847 p_nvenc->nvEncDestroyBitstreamBuffer(ctx->nvencoder, ctx->output_surfaces[i].output_surface);
849 ctx->max_surface_count = 0;
851 p_nvenc->nvEncDestroyEncoder(ctx->nvencoder);
852 ctx->nvencoder = NULL;
854 dl_fn->cu_ctx_destroy(ctx->cu_context);
855 ctx->cu_context = NULL;
857 nvenc_unload_nvenc(avctx);
859 av_frame_free(&avctx->coded_frame);
864 static int process_output_surface(AVCodecContext *avctx, AVPacket *pkt, AVFrame *coded_frame, NvencOutputSurface *tmpoutsurf)
866 NvencContext *ctx = avctx->priv_data;
867 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
868 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
870 uint32_t *slice_offsets = av_mallocz(ctx->encode_config.encodeCodecConfig.h264Config.sliceModeData * sizeof(*slice_offsets));
871 NV_ENC_LOCK_BITSTREAM lock_params = { 0 };
872 NVENCSTATUS nv_status;
876 return AVERROR(ENOMEM);
878 lock_params.version = NV_ENC_LOCK_BITSTREAM_VER;
880 lock_params.doNotWait = 0;
881 lock_params.outputBitstream = tmpoutsurf->output_surface;
882 lock_params.sliceOffsets = slice_offsets;
884 nv_status = p_nvenc->nvEncLockBitstream(ctx->nvencoder, &lock_params);
885 if (nv_status != NV_ENC_SUCCESS) {
886 av_log(avctx, AV_LOG_ERROR, "Failed locking bitstream buffer\n");
887 res = AVERROR_EXTERNAL;
891 if (res = ff_alloc_packet2(avctx, pkt, lock_params.bitstreamSizeInBytes)) {
892 p_nvenc->nvEncUnlockBitstream(ctx->nvencoder, tmpoutsurf->output_surface);
896 memcpy(pkt->data, lock_params.bitstreamBufferPtr, lock_params.bitstreamSizeInBytes);
898 nv_status = p_nvenc->nvEncUnlockBitstream(ctx->nvencoder, tmpoutsurf->output_surface);
899 if (nv_status != NV_ENC_SUCCESS)
900 av_log(avctx, AV_LOG_ERROR, "Failed unlocking bitstream buffer, expect the gates of mordor to open\n");
902 switch (lock_params.pictureType) {
903 case NV_ENC_PIC_TYPE_IDR:
904 pkt->flags |= AV_PKT_FLAG_KEY;
905 case NV_ENC_PIC_TYPE_I:
906 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
908 case NV_ENC_PIC_TYPE_P:
909 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_P;
911 case NV_ENC_PIC_TYPE_B:
912 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_B;
914 case NV_ENC_PIC_TYPE_BI:
915 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_BI;
918 av_log(avctx, AV_LOG_ERROR, "Unknown picture type encountered, expect the output to be broken.\n");
919 av_log(avctx, AV_LOG_ERROR, "Please report this error and include as much information on how to reproduce it as possible.\n");
920 res = AVERROR_EXTERNAL;
924 pkt->pts = lock_params.outputTimeStamp;
925 pkt->dts = timestamp_queue_dequeue(&ctx->timestamp_list);
927 if (pkt->dts > pkt->pts)
930 if (ctx->last_dts != AV_NOPTS_VALUE && pkt->dts <= ctx->last_dts)
931 pkt->dts = ctx->last_dts + 1;
933 ctx->last_dts = pkt->dts;
935 av_free(slice_offsets);
941 av_free(slice_offsets);
942 timestamp_queue_dequeue(&ctx->timestamp_list);
947 static int nvenc_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
948 const AVFrame *frame, int *got_packet)
950 NVENCSTATUS nv_status;
951 NvencOutputSurface *tmpoutsurf;
954 NvencContext *ctx = avctx->priv_data;
955 NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
956 NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
958 NV_ENC_PIC_PARAMS pic_params = { 0 };
959 pic_params.version = NV_ENC_PIC_PARAMS_VER;
962 NV_ENC_LOCK_INPUT_BUFFER lockBufferParams = { 0 };
963 NvencInputSurface *inSurf = NULL;
965 for (i = 0; i < ctx->max_surface_count; ++i) {
966 if (!ctx->input_surfaces[i].lockCount) {
967 inSurf = &ctx->input_surfaces[i];
974 inSurf->lockCount = 1;
976 lockBufferParams.version = NV_ENC_LOCK_INPUT_BUFFER_VER;
977 lockBufferParams.inputBuffer = inSurf->input_surface;
979 nv_status = p_nvenc->nvEncLockInputBuffer(ctx->nvencoder, &lockBufferParams);
980 if (nv_status != NV_ENC_SUCCESS) {
981 av_log(avctx, AV_LOG_ERROR, "Failed locking nvenc input buffer\n");
985 if (avctx->pix_fmt == AV_PIX_FMT_YUV420P) {
986 uint8_t *buf = lockBufferParams.bufferDataPtr;
988 av_image_copy_plane(buf, lockBufferParams.pitch,
989 frame->data[0], frame->linesize[0],
990 avctx->width, avctx->height);
992 buf += inSurf->height * lockBufferParams.pitch;
994 av_image_copy_plane(buf, lockBufferParams.pitch >> 1,
995 frame->data[2], frame->linesize[2],
996 avctx->width >> 1, avctx->height >> 1);
998 buf += (inSurf->height * lockBufferParams.pitch) >> 2;
1000 av_image_copy_plane(buf, lockBufferParams.pitch >> 1,
1001 frame->data[1], frame->linesize[1],
1002 avctx->width >> 1, avctx->height >> 1);
1003 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12) {
1004 uint8_t *buf = lockBufferParams.bufferDataPtr;
1006 av_image_copy_plane(buf, lockBufferParams.pitch,
1007 frame->data[0], frame->linesize[0],
1008 avctx->width, avctx->height);
1010 buf += inSurf->height * lockBufferParams.pitch;
1012 av_image_copy_plane(buf, lockBufferParams.pitch,
1013 frame->data[1], frame->linesize[1],
1014 avctx->width, avctx->height >> 1);
1015 } else if (avctx->pix_fmt == AV_PIX_FMT_YUV444P) {
1016 uint8_t *buf = lockBufferParams.bufferDataPtr;
1018 av_image_copy_plane(buf, lockBufferParams.pitch,
1019 frame->data[0], frame->linesize[0],
1020 avctx->width, avctx->height);
1022 buf += inSurf->height * lockBufferParams.pitch;
1024 av_image_copy_plane(buf, lockBufferParams.pitch,
1025 frame->data[1], frame->linesize[1],
1026 avctx->width, avctx->height);
1028 buf += inSurf->height * lockBufferParams.pitch;
1030 av_image_copy_plane(buf, lockBufferParams.pitch,
1031 frame->data[2], frame->linesize[2],
1032 avctx->width, avctx->height);
1034 av_log(avctx, AV_LOG_FATAL, "Invalid pixel format!\n");
1035 return AVERROR(EINVAL);
1038 nv_status = p_nvenc->nvEncUnlockInputBuffer(ctx->nvencoder, inSurf->input_surface);
1039 if (nv_status != NV_ENC_SUCCESS) {
1040 av_log(avctx, AV_LOG_FATAL, "Failed unlocking input buffer!\n");
1041 return AVERROR_EXTERNAL;
1044 for (i = 0; i < ctx->max_surface_count; ++i)
1045 if (!ctx->output_surfaces[i].busy)
1048 if (i == ctx->max_surface_count) {
1049 inSurf->lockCount = 0;
1050 av_log(avctx, AV_LOG_FATAL, "No free output surface found!\n");
1051 return AVERROR_EXTERNAL;
1054 ctx->output_surfaces[i].input_surface = inSurf;
1056 pic_params.inputBuffer = inSurf->input_surface;
1057 pic_params.bufferFmt = inSurf->format;
1058 pic_params.inputWidth = avctx->width;
1059 pic_params.inputHeight = avctx->height;
1060 pic_params.outputBitstream = ctx->output_surfaces[i].output_surface;
1061 pic_params.completionEvent = 0;
1063 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
1064 if (frame->top_field_first) {
1065 pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_TOP_BOTTOM;
1067 pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_BOTTOM_TOP;
1070 pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
1073 pic_params.encodePicFlags = 0;
1074 pic_params.inputTimeStamp = frame->pts;
1075 pic_params.inputDuration = 0;
1076 pic_params.codecPicParams.h264PicParams.sliceMode = ctx->encode_config.encodeCodecConfig.h264Config.sliceMode;
1077 pic_params.codecPicParams.h264PicParams.sliceModeData = ctx->encode_config.encodeCodecConfig.h264Config.sliceModeData;
1078 memcpy(&pic_params.rcParams, &ctx->encode_config.rcParams, sizeof(NV_ENC_RC_PARAMS));
1080 res = timestamp_queue_enqueue(&ctx->timestamp_list, frame->pts);
1085 pic_params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
1088 nv_status = p_nvenc->nvEncEncodePicture(ctx->nvencoder, &pic_params);
1090 if (frame && nv_status == NV_ENC_ERR_NEED_MORE_INPUT) {
1091 res = out_surf_queue_enqueue(&ctx->output_surface_queue, &ctx->output_surfaces[i]);
1096 ctx->output_surfaces[i].busy = 1;
1099 if (nv_status != NV_ENC_SUCCESS && nv_status != NV_ENC_ERR_NEED_MORE_INPUT) {
1100 av_log(avctx, AV_LOG_ERROR, "EncodePicture failed!\n");
1101 return AVERROR_EXTERNAL;
1104 if (nv_status != NV_ENC_ERR_NEED_MORE_INPUT) {
1105 while (ctx->output_surface_queue.count) {
1106 tmpoutsurf = out_surf_queue_dequeue(&ctx->output_surface_queue);
1107 res = out_surf_queue_enqueue(&ctx->output_surface_ready_queue, tmpoutsurf);
1114 res = out_surf_queue_enqueue(&ctx->output_surface_ready_queue, &ctx->output_surfaces[i]);
1119 ctx->output_surfaces[i].busy = 1;
1123 if (ctx->output_surface_ready_queue.count) {
1124 tmpoutsurf = out_surf_queue_dequeue(&ctx->output_surface_ready_queue);
1126 res = process_output_surface(avctx, pkt, avctx->coded_frame, tmpoutsurf);
1131 tmpoutsurf->busy = 0;
1132 av_assert0(tmpoutsurf->input_surface->lockCount);
1133 tmpoutsurf->input_surface->lockCount--;
1143 static enum AVPixelFormat pix_fmts_nvenc[] = {
1148 #define OFFSET(x) offsetof(NvencContext, x)
1149 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1150 static const AVOption options[] = {
1151 { "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 },
1152 { "cbr", "Use cbr encoding mode", OFFSET(cbr), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
1153 { "2pass", "Use 2pass cbr encoding mode (low latency mode only)", OFFSET(twopass), AV_OPT_TYPE_INT, { .i64 = -1 }, -1, 1, VE },
1154 { "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 },
1155 { "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 },
1159 static const AVClass nvenc_class = {
1160 .class_name = "nvenc",
1161 .item_name = av_default_item_name,
1163 .version = LIBAVUTIL_VERSION_INT,
1166 static const AVCodecDefault nvenc_defaults[] = {
1176 AVCodec ff_nvenc_encoder = {
1178 .long_name = NULL_IF_CONFIG_SMALL("Nvidia NVENC h264 encoder"),
1179 .type = AVMEDIA_TYPE_VIDEO,
1180 .id = AV_CODEC_ID_H264,
1181 .priv_data_size = sizeof(NvencContext),
1182 .init = nvenc_encode_init,
1183 .encode2 = nvenc_encode_frame,
1184 .close = nvenc_encode_close,
1185 .capabilities = CODEC_CAP_DELAY,
1186 .priv_class = &nvenc_class,
1187 .defaults = nvenc_defaults,
1188 .pix_fmts = pix_fmts_nvenc,