libavcodec/nvenc.c

   1 /*
   2  * NVIDIA NVENC Support
   3  * Copyright (C) 2015 Luca Barbato
   4  *
   5  * This file is part of Libav.
   6  *
   7  * Libav is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * Libav is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with Libav; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #include "config.h"
  23
  24 #include <cuda.h>
  25 #include <nvEncodeAPI.h>
  26 #include <string.h>
  27
  28 #define CUDA_LIBNAME "libcuda.so"
  29
  30 #if HAVE_DLFCN_H
  31 #include <dlfcn.h>
  32
  33 #define NVENC_LIBNAME "libnvidia-encode.so"
  34
  35 #elif HAVE_WINDOWS_H
  36 #include <windows.h>
  37
  38 #if ARCH_X86_64
  39 #define NVENC_LIBNAME "nvEncodeAPI64.dll"
  40 #else
  41 #define NVENC_LIBNAME "nvEncodeAPI.dll"
  42 #endif
  43
  44 #define dlopen(filename, flags) LoadLibrary((filename))
  45 #define dlsym(handle, symbol)   GetProcAddress(handle, symbol)
  46 #define dlclose(handle)         FreeLibrary(handle)
  47 #endif
  48
  49 #include "libavutil/common.h"
  50 #include "libavutil/imgutils.h"
  51 #include "libavutil/mem.h"
  52 #include "avcodec.h"
  53 #include "internal.h"
  54 #include "nvenc.h"
  55
  56 #define NVENC_CAP 0x30
  57 #define BITSTREAM_BUFFER_SIZE 1024 * 1024
  58
  59 #define LOAD_LIBRARY(l, path)                   \
  60     do {                                        \
  61         if (!((l) = dlopen(path, RTLD_LAZY))) { \
  62             av_log(avctx, AV_LOG_ERROR,         \
  63                    "Cannot load %s\n",          \
  64                    path);                       \
  65             return AVERROR_UNKNOWN;             \
  66         }                                       \
  67     } while (0)
  68
  69 #define LOAD_SYMBOL(fun, lib, symbol)        \
  70     do {                                     \
  71         if (!((fun) = dlsym(lib, symbol))) { \
  72             av_log(avctx, AV_LOG_ERROR,      \
  73                    "Cannot load %s\n",       \
  74                    symbol);                  \
  75             return AVERROR_UNKNOWN;          \
  76         }                                    \
  77     } while (0)
  78
  79 static av_cold int nvenc_load_libraries(AVCodecContext *avctx)
  80 {
  81     NVENCContext *ctx         = avctx->priv_data;
  82     NVENCLibraryContext *nvel = &ctx->nvel;
  83     PNVENCODEAPICREATEINSTANCE nvenc_create_instance;
  84
  85     LOAD_LIBRARY(nvel->cuda, CUDA_LIBNAME);
  86
  87     LOAD_SYMBOL(nvel->cu_init, nvel->cuda, "cuInit");
  88     LOAD_SYMBOL(nvel->cu_device_get_count, nvel->cuda, "cuDeviceGetCount");
  89     LOAD_SYMBOL(nvel->cu_device_get, nvel->cuda, "cuDeviceGet");
  90     LOAD_SYMBOL(nvel->cu_device_get_name, nvel->cuda, "cuDeviceGetName");
  91     LOAD_SYMBOL(nvel->cu_device_compute_capability, nvel->cuda,
  92                 "cuDeviceComputeCapability");
  93     LOAD_SYMBOL(nvel->cu_ctx_create, nvel->cuda, "cuCtxCreate_v2");
  94     LOAD_SYMBOL(nvel->cu_ctx_pop_current, nvel->cuda, "cuCtxPopCurrent_v2");
  95     LOAD_SYMBOL(nvel->cu_ctx_destroy, nvel->cuda, "cuCtxDestroy_v2");
  96
  97     LOAD_LIBRARY(nvel->nvenc, NVENC_LIBNAME);
  98
  99     LOAD_SYMBOL(nvenc_create_instance, nvel->nvenc,
 100                 "NvEncodeAPICreateInstance");
 101
 102     nvel->nvenc_funcs.version = NV_ENCODE_API_FUNCTION_LIST_VER;
 103
 104     if ((nvenc_create_instance(&nvel->nvenc_funcs)) != NV_ENC_SUCCESS) {
 105         av_log(avctx, AV_LOG_ERROR, "Cannot create the NVENC instance");
 106         return AVERROR_UNKNOWN;
 107     }
 108
 109     return 0;
 110 }
 111
 112 static int nvenc_open_session(AVCodecContext *avctx)
 113 {
 114     NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS params = { 0 };
 115     NVENCContext *ctx                           = avctx->priv_data;
 116     NV_ENCODE_API_FUNCTION_LIST *nv             = &ctx->nvel.nvenc_funcs;
 117     int ret;
 118
 119     params.version    = NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER;
 120     params.apiVersion = NVENCAPI_VERSION;
 121     params.device     = ctx->cu_context;
 122     params.deviceType = NV_ENC_DEVICE_TYPE_CUDA;
 123
 124     ret = nv->nvEncOpenEncodeSessionEx(&params, &ctx->nvenc_ctx);
 125     if (ret != NV_ENC_SUCCESS) {
 126         ctx->nvenc_ctx = NULL;
 127         av_log(avctx, AV_LOG_ERROR,
 128                "Cannot open the NVENC Session\n");
 129         return AVERROR_UNKNOWN;
 130     }
 131
 132     return 0;
 133 }
 134
 135 static int nvenc_check_codec_support(AVCodecContext *avctx)
 136 {
 137     NVENCContext *ctx               = avctx->priv_data;
 138     NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
 139     int i, ret, count = 0;
 140     GUID *guids = NULL;
 141
 142     ret = nv->nvEncGetEncodeGUIDCount(ctx->nvenc_ctx, &count);
 143
 144     if (ret != NV_ENC_SUCCESS || !count)
 145         return AVERROR(ENOSYS);
 146
 147     guids = av_malloc(count * sizeof(GUID));
 148     if (!guids)
 149         return AVERROR(ENOMEM);
 150
 151     ret = nv->nvEncGetEncodeGUIDs(ctx->nvenc_ctx, guids, count, &count);
 152     if (ret != NV_ENC_SUCCESS) {
 153         ret = AVERROR(ENOSYS);
 154         goto fail;
 155     }
 156
 157     ret = AVERROR(ENOSYS);
 158     for (i = 0; i < count; i++) {
 159         if (!memcmp(&guids[i], &ctx->params.encodeGUID, sizeof(*guids))) {
 160             ret = 0;
 161             break;
 162         }
 163     }
 164
 165 fail:
 166     av_free(guids);
 167
 168     return ret;
 169 }
 170
 171 static int nvenc_check_cap(AVCodecContext *avctx, NV_ENC_CAPS cap)
 172 {
 173     NVENCContext *ctx               = avctx->priv_data;
 174     NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
 175     NV_ENC_CAPS_PARAM params        = { 0 };
 176     int ret, val = 0;
 177
 178     params.version     = NV_ENC_CAPS_PARAM_VER;
 179     params.capsToQuery = cap;
 180
 181     ret = nv->nvEncGetEncodeCaps(ctx->nvenc_ctx, ctx->params.encodeGUID, &params, &val);
 182
 183     if (ret == NV_ENC_SUCCESS)
 184         return val;
 185     return 0;
 186 }
 187
 188 static int nvenc_check_capabilities(AVCodecContext *avctx)
 189 {
 190     int ret;
 191
 192     ret = nvenc_check_codec_support(avctx);
 193     if (ret < 0) {
 194         av_log(avctx, AV_LOG_VERBOSE, "Codec not supported\n");
 195         return ret;
 196     }
 197
 198     ret = nvenc_check_cap(avctx, NV_ENC_CAPS_SUPPORT_YUV444_ENCODE);
 199     if (avctx->pix_fmt == AV_PIX_FMT_YUV444P && ret <= 0) {
 200         av_log(avctx, AV_LOG_VERBOSE, "YUV444P not supported\n");
 201         return AVERROR(ENOSYS);
 202     }
 203
 204     ret = nvenc_check_cap(avctx, NV_ENC_CAPS_WIDTH_MAX);
 205     if (ret < avctx->width) {
 206         av_log(avctx, AV_LOG_VERBOSE, "Width %d exceeds %d\n",
 207                avctx->width, ret);
 208         return AVERROR(ENOSYS);
 209     }
 210
 211     ret = nvenc_check_cap(avctx, NV_ENC_CAPS_HEIGHT_MAX);
 212     if (ret < avctx->height) {
 213         av_log(avctx, AV_LOG_VERBOSE, "Height %d exceeds %d\n",
 214                avctx->height, ret);
 215         return AVERROR(ENOSYS);
 216     }
 217
 218     ret = nvenc_check_cap(avctx, NV_ENC_CAPS_NUM_MAX_BFRAMES);
 219     if (ret < avctx->max_b_frames) {
 220         av_log(avctx, AV_LOG_VERBOSE, "Max b-frames %d exceed %d\n",
 221                avctx->max_b_frames, ret);
 222
 223         return AVERROR(ENOSYS);
 224     }
 225
 226     return 0;
 227 }
 228
 229 static int nvenc_check_device(AVCodecContext *avctx, int idx)
 230 {
 231     NVENCContext *ctx               = avctx->priv_data;
 232     NVENCLibraryContext *nvel       = &ctx->nvel;
 233     char name[128]                  = { 0 };
 234     int major, minor, ret;
 235     CUdevice cu_device;
 236     CUcontext dummy;
 237     int loglevel = AV_LOG_VERBOSE;
 238
 239     if (ctx->device == LIST_DEVICES)
 240         loglevel = AV_LOG_INFO;
 241
 242     ret = nvel->cu_device_get(&cu_device, idx);
 243     if (ret != CUDA_SUCCESS) {
 244         av_log(avctx, AV_LOG_ERROR,
 245                "Cannot access the CUDA device %d\n",
 246                idx);
 247         return -1;
 248     }
 249
 250     ret = nvel->cu_device_get_name(name, sizeof(name), cu_device);
 251     if (ret != CUDA_SUCCESS)
 252         return -1;
 253
 254     ret = nvel->cu_device_compute_capability(&major, &minor, cu_device);
 255     if (ret != CUDA_SUCCESS)
 256         return -1;
 257
 258     av_log(avctx, loglevel, "Device %d [%s] ", cu_device, name);
 259
 260     if (((major << 4) | minor) < NVENC_CAP)
 261         goto fail;
 262
 263     ret = nvel->cu_ctx_create(&ctx->cu_context, 0, cu_device);
 264     if (ret != CUDA_SUCCESS)
 265         goto fail;
 266
 267     ret = nvel->cu_ctx_pop_current(&dummy);
 268     if (ret != CUDA_SUCCESS)
 269         goto fail2;
 270
 271     if ((ret = nvenc_open_session(avctx)) < 0)
 272         goto fail2;
 273
 274     if ((ret = nvenc_check_capabilities(avctx)) < 0)
 275         goto fail3;
 276
 277     av_log(avctx, loglevel, "supports NVENC\n");
 278
 279     if (ctx->device == cu_device || ctx->device == ANY_DEVICE)
 280         return 0;
 281
 282 fail3:
 283     nvel->nvenc_funcs.nvEncDestroyEncoder(ctx->nvenc_ctx);
 284     ctx->nvenc_ctx = NULL;
 285
 286 fail2:
 287     nvel->cu_ctx_destroy(ctx->cu_context);
 288     ctx->cu_context = NULL;
 289
 290 fail:
 291     if (ret != 0)
 292         av_log(avctx, loglevel, "does not support NVENC (major %d minor %d)\n",
 293                major, minor);
 294
 295     return AVERROR(ENOSYS);
 296 }
 297
 298 static int nvenc_setup_device(AVCodecContext *avctx)
 299 {
 300     NVENCContext *ctx         = avctx->priv_data;
 301     NVENCLibraryContext *nvel = &ctx->nvel;
 302     int i, nb_devices = 0;
 303
 304     if ((nvel->cu_init(0)) != CUDA_SUCCESS) {
 305         av_log(avctx, AV_LOG_ERROR,
 306                "Cannot init CUDA\n");
 307         return AVERROR_UNKNOWN;
 308     }
 309
 310     if ((nvel->cu_device_get_count(&nb_devices)) != CUDA_SUCCESS) {
 311         av_log(avctx, AV_LOG_ERROR,
 312                "Cannot enumerate the CUDA devices\n");
 313         return AVERROR_UNKNOWN;
 314     }
 315
 316     switch (avctx->codec->id) {
 317     case AV_CODEC_ID_H264:
 318         ctx->params.encodeGUID = NV_ENC_CODEC_H264_GUID;
 319         break;
 320     case AV_CODEC_ID_HEVC:
 321         ctx->params.encodeGUID = NV_ENC_CODEC_HEVC_GUID;
 322         break;
 323     default:
 324         return AVERROR_BUG;
 325     }
 326
 327     for (i = 0; i < nb_devices; ++i) {
 328         if ((nvenc_check_device(avctx, i)) >= 0 && ctx->device != LIST_DEVICES)
 329             return 0;
 330     }
 331
 332     if (ctx->device == LIST_DEVICES)
 333         return AVERROR_EXIT;
 334
 335     return AVERROR(ENOSYS);
 336 }
 337
 338 typedef struct GUIDTuple {
 339     const GUID guid;
 340     int flags;
 341 } GUIDTuple;
 342
 343 static int nvec_map_preset(NVENCContext *ctx)
 344 {
 345     GUIDTuple presets[] = {
 346         { NV_ENC_PRESET_DEFAULT_GUID },
 347         { NV_ENC_PRESET_HP_GUID },
 348         { NV_ENC_PRESET_HQ_GUID },
 349         { NV_ENC_PRESET_BD_GUID },
 350         { NV_ENC_PRESET_LOW_LATENCY_DEFAULT_GUID, NVENC_LOWLATENCY },
 351         { NV_ENC_PRESET_LOW_LATENCY_HP_GUID,      NVENC_LOWLATENCY },
 352         { NV_ENC_PRESET_LOW_LATENCY_HQ_GUID,      NVENC_LOWLATENCY },
 353         { NV_ENC_PRESET_LOSSLESS_DEFAULT_GUID,    NVENC_LOSSLESS },
 354         { NV_ENC_PRESET_LOSSLESS_HP_GUID,         NVENC_LOSSLESS },
 355         { { 0 } }
 356     };
 357
 358     GUIDTuple *t = &presets[ctx->preset];
 359
 360     ctx->params.presetGUID = t->guid;
 361     ctx->flags             = t->flags;
 362
 363     return AVERROR(EINVAL);
 364 }
 365
 366 static void set_constqp(AVCodecContext *avctx, NV_ENC_RC_PARAMS *rc)
 367 {
 368     rc->rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
 369     rc->constQP.qpInterB = avctx->global_quality;
 370     rc->constQP.qpInterP = avctx->global_quality;
 371     rc->constQP.qpIntra  = avctx->global_quality;
 372 }
 373
 374 static void set_vbr(AVCodecContext *avctx, NV_ENC_RC_PARAMS *rc)
 375 {
 376     if (avctx->qmin >= 0) {
 377         rc->enableMinQP    = 1;
 378         rc->minQP.qpInterB = avctx->qmin;
 379         rc->minQP.qpInterP = avctx->qmin;
 380         rc->minQP.qpIntra  = avctx->qmin;
 381     }
 382
 383     if (avctx->qmax >= 0) {
 384         rc->enableMaxQP = 1;
 385         rc->maxQP.qpInterB = avctx->qmax;
 386         rc->maxQP.qpInterP = avctx->qmax;
 387         rc->maxQP.qpIntra  = avctx->qmax;
 388     }
 389 }
 390
 391 static void nvenc_override_rate_control(AVCodecContext *avctx,
 392                                         NV_ENC_RC_PARAMS *rc)
 393 {
 394     NVENCContext *ctx    = avctx->priv_data;
 395
 396     switch (ctx->rc) {
 397     case NV_ENC_PARAMS_RC_CONSTQP:
 398         if (avctx->global_quality < 0) {
 399             av_log(avctx, AV_LOG_WARNING,
 400                    "The constant quality rate-control requires "
 401                    "the 'global_quality' option set.\n");
 402             return;
 403         }
 404         set_constqp(avctx, rc);
 405         return;
 406     case NV_ENC_PARAMS_RC_2_PASS_VBR:
 407     case NV_ENC_PARAMS_RC_VBR:
 408         if (avctx->qmin < 0 && avctx->qmax < 0) {
 409             av_log(avctx, AV_LOG_WARNING,
 410                    "The variable bitrate rate-control requires "
 411                    "the 'qmin' and/or 'qmax' option set.\n");
 412             return;
 413         }
 414     case NV_ENC_PARAMS_RC_VBR_MINQP:
 415         if (avctx->qmin < 0) {
 416             av_log(avctx, AV_LOG_WARNING,
 417                    "The variable bitrate rate-control requires "
 418                    "the 'qmin' option set.\n");
 419             return;
 420         }
 421         set_vbr(avctx, rc);
 422         break;
 423     case NV_ENC_PARAMS_RC_CBR:
 424         break;
 425     case NV_ENC_PARAMS_RC_2_PASS_QUALITY:
 426     case NV_ENC_PARAMS_RC_2_PASS_FRAMESIZE_CAP:
 427         if (!(ctx->flags & NVENC_LOWLATENCY)) {
 428             av_log(avctx, AV_LOG_WARNING,
 429                    "The multipass rate-control requires "
 430                    "a low-latency preset.\n");
 431             return;
 432         }
 433     }
 434
 435     rc->rateControlMode = ctx->rc;
 436 }
 437
 438 static void nvenc_setup_rate_control(AVCodecContext *avctx)
 439 {
 440     NVENCContext *ctx    = avctx->priv_data;
 441     NV_ENC_RC_PARAMS *rc = &ctx->config.rcParams;
 442
 443     if (avctx->bit_rate > 0)
 444         rc->averageBitRate = avctx->bit_rate;
 445
 446     if (avctx->rc_max_rate > 0)
 447         rc->maxBitRate = avctx->rc_max_rate;
 448
 449     if (ctx->rc > 0) {
 450         nvenc_override_rate_control(avctx, rc);
 451     } else if (avctx->global_quality > 0) {
 452         set_constqp(avctx, rc);
 453     } else if (avctx->qmin >= 0 && avctx->qmax >= 0) {
 454         rc->rateControlMode = NV_ENC_PARAMS_RC_VBR;
 455         set_vbr(avctx, rc);
 456     }
 457
 458     if (avctx->rc_buffer_size > 0)
 459         rc->vbvBufferSize = avctx->rc_buffer_size;
 460
 461     if (rc->averageBitRate > 0)
 462         avctx->bit_rate = rc->averageBitRate;
 463 }
 464
 465 static int nvenc_setup_h264_config(AVCodecContext *avctx)
 466 {
 467     NVENCContext *ctx                      = avctx->priv_data;
 468     NV_ENC_CONFIG *cc                      = &ctx->config;
 469     NV_ENC_CONFIG_H264 *h264               = &cc->encodeCodecConfig.h264Config;
 470     NV_ENC_CONFIG_H264_VUI_PARAMETERS *vui = &h264->h264VUIParameters;
 471
 472     vui->colourDescriptionPresentFlag = 1;
 473     vui->videoSignalTypePresentFlag   = 1;
 474
 475     vui->colourMatrix            = avctx->colorspace;
 476     vui->colourPrimaries         = avctx->color_primaries;
 477     vui->transferCharacteristics = avctx->color_trc;
 478
 479     vui->videoFullRangeFlag = avctx->color_range == AVCOL_RANGE_JPEG;
 480
 481     h264->disableSPSPPS = (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
 482     h264->repeatSPSPPS  = (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) ? 0 : 1;
 483
 484     h264->maxNumRefFrames = avctx->refs;
 485     h264->idrPeriod       = cc->gopLength;
 486
 487     if (ctx->profile)
 488         avctx->profile = ctx->profile;
 489
 490     if (avctx->pix_fmt == AV_PIX_FMT_YUV444P)
 491         h264->chromaFormatIDC = 3;
 492     else
 493         h264->chromaFormatIDC = 1;
 494
 495     switch (ctx->profile) {
 496     case NV_ENC_H264_PROFILE_BASELINE:
 497         cc->profileGUID = NV_ENC_H264_PROFILE_BASELINE_GUID;
 498         break;
 499     case NV_ENC_H264_PROFILE_MAIN:
 500         cc->profileGUID = NV_ENC_H264_PROFILE_MAIN_GUID;
 501         break;
 502     case NV_ENC_H264_PROFILE_HIGH:
 503         cc->profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
 504         break;
 505     case NV_ENC_H264_PROFILE_HIGH_444:
 506         cc->profileGUID = NV_ENC_H264_PROFILE_HIGH_444_GUID;
 507         break;
 508     case NV_ENC_H264_PROFILE_CONSTRAINED_HIGH:
 509         cc->profileGUID = NV_ENC_H264_PROFILE_CONSTRAINED_HIGH_GUID;
 510         break;
 511     }
 512
 513     h264->level = ctx->level;
 514
 515     return 0;
 516 }
 517
 518 static int nvenc_setup_hevc_config(AVCodecContext *avctx)
 519 {
 520     NVENCContext *ctx                      = avctx->priv_data;
 521     NV_ENC_CONFIG *cc                      = &ctx->config;
 522     NV_ENC_CONFIG_HEVC *hevc               = &cc->encodeCodecConfig.hevcConfig;
 523
 524     hevc->disableSPSPPS = (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
 525     hevc->repeatSPSPPS  = (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) ? 0 : 1;
 526
 527     hevc->maxNumRefFramesInDPB = avctx->refs;
 528     hevc->idrPeriod            = cc->gopLength;
 529
 530     /* No other profile is supported in the current SDK version 5 */
 531     cc->profileGUID = NV_ENC_HEVC_PROFILE_MAIN_GUID;
 532     avctx->profile  = FF_PROFILE_HEVC_MAIN;
 533
 534     if (ctx->level) {
 535         hevc->level = ctx->level;
 536     } else {
 537         hevc->level = NV_ENC_LEVEL_AUTOSELECT;
 538     }
 539
 540     if (ctx->tier) {
 541         hevc->tier = ctx->tier;
 542     }
 543
 544     return 0;
 545 }
 546 static int nvenc_setup_codec_config(AVCodecContext *avctx)
 547 {
 548     switch (avctx->codec->id) {
 549     case AV_CODEC_ID_H264:
 550         return nvenc_setup_h264_config(avctx);
 551     case AV_CODEC_ID_HEVC:
 552         return nvenc_setup_hevc_config(avctx);
 553     }
 554     return 0;
 555 }
 556
 557 static int nvenc_setup_encoder(AVCodecContext *avctx)
 558 {
 559     NVENCContext *ctx               = avctx->priv_data;
 560     NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
 561     NV_ENC_PRESET_CONFIG preset_cfg = { 0 };
 562     int ret;
 563
 564     ctx->params.version = NV_ENC_INITIALIZE_PARAMS_VER;
 565
 566     ctx->params.encodeHeight = avctx->height;
 567     ctx->params.encodeWidth  = avctx->width;
 568
 569     if (avctx->sample_aspect_ratio.num &&
 570         avctx->sample_aspect_ratio.den &&
 571         (avctx->sample_aspect_ratio.num != 1 ||
 572          avctx->sample_aspect_ratio.den != 1)) {
 573         av_reduce(&ctx->params.darWidth,
 574                   &ctx->params.darHeight,
 575                   avctx->width * avctx->sample_aspect_ratio.num,
 576                   avctx->height * avctx->sample_aspect_ratio.den,
 577                   INT_MAX / 8);
 578     } else {
 579         ctx->params.darHeight = avctx->height;
 580         ctx->params.darWidth  = avctx->width;
 581     }
 582
 583     ctx->params.frameRateNum = avctx->time_base.den;
 584     ctx->params.frameRateDen = avctx->time_base.num * avctx->ticks_per_frame;
 585
 586     ctx->params.enableEncodeAsync = 0;
 587     ctx->params.enablePTD         = 1;
 588
 589     ctx->params.encodeConfig = &ctx->config;
 590
 591     nvec_map_preset(ctx);
 592
 593     preset_cfg.version           = NV_ENC_PRESET_CONFIG_VER;
 594     preset_cfg.presetCfg.version = NV_ENC_CONFIG_VER;
 595
 596     ret = nv->nvEncGetEncodePresetConfig(ctx->nvenc_ctx,
 597                                          ctx->params.encodeGUID,
 598                                          ctx->params.presetGUID,
 599                                          &preset_cfg);
 600     if (ret != NV_ENC_SUCCESS) {
 601         av_log(avctx, AV_LOG_ERROR,
 602                "Cannot get the preset configuration\n");
 603         return AVERROR_UNKNOWN;
 604     }
 605
 606     memcpy(&ctx->config, &preset_cfg.presetCfg, sizeof(ctx->config));
 607
 608     ctx->config.version = NV_ENC_CONFIG_VER;
 609
 610     if (avctx->gop_size > 0) {
 611         if (avctx->max_b_frames > 0) {
 612             ctx->last_dts = -2;
 613             /* 0 is intra-only,
 614              * 1 is I/P only,
 615              * 2 is one B Frame,
 616              * 3 two B frames, and so on. */
 617             ctx->config.frameIntervalP = avctx->max_b_frames + 1;
 618         } else if (avctx->max_b_frames == 0) {
 619             ctx->config.frameIntervalP = 1;
 620         }
 621         ctx->config.gopLength = avctx->gop_size;
 622     } else if (avctx->gop_size == 0) {
 623         ctx->config.frameIntervalP = 0;
 624         ctx->config.gopLength      = 1;
 625     }
 626
 627     if (ctx->config.frameIntervalP > 1)
 628         avctx->max_b_frames = ctx->config.frameIntervalP - 1;
 629
 630     nvenc_setup_rate_control(avctx);
 631
 632     if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
 633         ctx->config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FIELD;
 634     } else {
 635         ctx->config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FRAME;
 636     }
 637
 638     if ((ret = nvenc_setup_codec_config(avctx)) < 0)
 639         return ret;
 640
 641     ret = nv->nvEncInitializeEncoder(ctx->nvenc_ctx, &ctx->params);
 642     if (ret != NV_ENC_SUCCESS) {
 643         av_log(avctx, AV_LOG_ERROR, "Cannot initialize the decoder");
 644         return AVERROR_UNKNOWN;
 645     }
 646
 647     return 0;
 648 }
 649
 650 static int nvenc_alloc_surface(AVCodecContext *avctx, int idx)
 651 {
 652     NVENCContext *ctx               = avctx->priv_data;
 653     NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
 654     int ret;
 655     NV_ENC_CREATE_INPUT_BUFFER in_buffer      = { 0 };
 656     NV_ENC_CREATE_BITSTREAM_BUFFER out_buffer = { 0 };
 657
 658     in_buffer.version  = NV_ENC_CREATE_INPUT_BUFFER_VER;
 659     out_buffer.version = NV_ENC_CREATE_BITSTREAM_BUFFER_VER;
 660
 661     in_buffer.width  = avctx->width;
 662     in_buffer.height = avctx->height;
 663
 664     in_buffer.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_UNCACHED;
 665
 666     switch (avctx->pix_fmt) {
 667     case AV_PIX_FMT_YUV420P:
 668         in_buffer.bufferFmt = NV_ENC_BUFFER_FORMAT_YV12_PL;
 669         break;
 670     case AV_PIX_FMT_NV12:
 671         in_buffer.bufferFmt = NV_ENC_BUFFER_FORMAT_NV12_PL;
 672         break;
 673     case AV_PIX_FMT_YUV444P:
 674         in_buffer.bufferFmt = NV_ENC_BUFFER_FORMAT_YUV444_PL;
 675         break;
 676     default:
 677         return AVERROR_BUG;
 678     }
 679
 680     ret = nv->nvEncCreateInputBuffer(ctx->nvenc_ctx, &in_buffer);
 681     if (ret != NV_ENC_SUCCESS) {
 682         av_log(avctx, AV_LOG_ERROR, "CreateInputBuffer failed\n");
 683         return AVERROR_UNKNOWN;
 684     }
 685
 686     ctx->in[idx].in        = in_buffer.inputBuffer;
 687     ctx->in[idx].format    = in_buffer.bufferFmt;
 688
 689     /* 1MB is large enough to hold most output frames.
 690      * NVENC increases this automaticaly if it's not enough. */
 691     out_buffer.size = BITSTREAM_BUFFER_SIZE;
 692
 693     out_buffer.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_UNCACHED;
 694
 695     ret = nv->nvEncCreateBitstreamBuffer(ctx->nvenc_ctx, &out_buffer);
 696     if (ret != NV_ENC_SUCCESS) {
 697         av_log(avctx, AV_LOG_ERROR, "CreateBitstreamBuffer failed\n");
 698         return AVERROR_UNKNOWN;
 699     }
 700
 701     ctx->out[idx].out  = out_buffer.bitstreamBuffer;
 702     ctx->out[idx].busy = 0;
 703
 704     return 0;
 705 }
 706
 707 static int nvenc_setup_surfaces(AVCodecContext *avctx)
 708 {
 709     NVENCContext *ctx = avctx->priv_data;
 710     int i, ret;
 711
 712     ctx->nb_surfaces = FFMAX(4 + avctx->max_b_frames,
 713                              ctx->nb_surfaces);
 714
 715     ctx->in = av_mallocz(ctx->nb_surfaces * sizeof(*ctx->in));
 716     if (!ctx->in)
 717         return AVERROR(ENOMEM);
 718
 719     ctx->out = av_mallocz(ctx->nb_surfaces * sizeof(*ctx->out));
 720     if (!ctx->out)
 721         return AVERROR(ENOMEM);
 722
 723     ctx->timestamps = av_fifo_alloc(ctx->nb_surfaces * sizeof(int64_t));
 724     if (!ctx->timestamps)
 725         return AVERROR(ENOMEM);
 726     ctx->pending = av_fifo_alloc(ctx->nb_surfaces * sizeof(ctx->out));
 727     if (!ctx->pending)
 728         return AVERROR(ENOMEM);
 729     ctx->ready = av_fifo_alloc(ctx->nb_surfaces * sizeof(ctx->out));
 730     if (!ctx->ready)
 731         return AVERROR(ENOMEM);
 732
 733     for (i = 0; i < ctx->nb_surfaces; i++) {
 734         if ((ret = nvenc_alloc_surface(avctx, i)) < 0)
 735             return ret;
 736     }
 737
 738     return 0;
 739 }
 740
 741 #define EXTRADATA_SIZE 512
 742
 743 static int nvenc_setup_extradata(AVCodecContext *avctx)
 744 {
 745     NVENCContext *ctx                     = avctx->priv_data;
 746     NV_ENCODE_API_FUNCTION_LIST *nv       = &ctx->nvel.nvenc_funcs;
 747     NV_ENC_SEQUENCE_PARAM_PAYLOAD payload = { 0 };
 748     int ret;
 749
 750     avctx->extradata = av_mallocz(EXTRADATA_SIZE + FF_INPUT_BUFFER_PADDING_SIZE);
 751     if (!avctx->extradata)
 752         return AVERROR(ENOMEM);
 753
 754     payload.version              = NV_ENC_SEQUENCE_PARAM_PAYLOAD_VER;
 755     payload.spsppsBuffer         = avctx->extradata;
 756     payload.inBufferSize         = EXTRADATA_SIZE;
 757     payload.outSPSPPSPayloadSize = &avctx->extradata_size;
 758
 759     ret = nv->nvEncGetSequenceParams(ctx->nvenc_ctx, &payload);
 760     if (ret != NV_ENC_SUCCESS) {
 761         av_log(avctx, AV_LOG_ERROR, "Cannot get the extradata\n");
 762         return AVERROR_UNKNOWN;
 763     }
 764
 765     return 0;
 766 }
 767
 768 av_cold int ff_nvenc_encode_close(AVCodecContext *avctx)
 769 {
 770     NVENCContext *ctx               = avctx->priv_data;
 771     NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
 772     int i;
 773
 774     if (ctx->in) {
 775         for (i = 0; i < ctx->nb_surfaces; ++i) {
 776             nv->nvEncDestroyInputBuffer(ctx->nvenc_ctx, ctx->in[i].in);
 777             nv->nvEncDestroyBitstreamBuffer(ctx->nvenc_ctx, ctx->out[i].out);
 778         }
 779     }
 780
 781     av_freep(&ctx->in);
 782     av_freep(&ctx->out);
 783
 784     if (ctx->nvenc_ctx)
 785         nv->nvEncDestroyEncoder(ctx->nvenc_ctx);
 786
 787     if (ctx->cu_context)
 788         ctx->nvel.cu_ctx_destroy(ctx->cu_context);
 789
 790     if (ctx->nvel.nvenc)
 791         dlclose(ctx->nvel.nvenc);
 792
 793     if (ctx->nvel.cuda)
 794         dlclose(ctx->nvel.cuda);
 795
 796     return 0;
 797 }
 798
 799 av_cold int ff_nvenc_encode_init(AVCodecContext *avctx)
 800 {
 801     int ret;
 802
 803     if ((ret = nvenc_load_libraries(avctx)) < 0)
 804         return ret;
 805
 806     if ((ret = nvenc_setup_device(avctx)) < 0)
 807         return ret;
 808
 809     if ((ret = nvenc_setup_encoder(avctx)) < 0)
 810         return ret;
 811
 812     if ((ret = nvenc_setup_surfaces(avctx)) < 0)
 813         return ret;
 814
 815     if (avctx->flags & CODEC_FLAG_GLOBAL_HEADER) {
 816         if ((ret = nvenc_setup_extradata(avctx)) < 0)
 817             return ret;
 818     }
 819
 820     return 0;
 821 }
 822
 823 static NVENCInputSurface *get_input_surface(NVENCContext *ctx)
 824 {
 825     int i;
 826
 827     for (i = 0; i < ctx->nb_surfaces; i++) {
 828         if (!ctx->in[i].locked) {
 829             ctx->in[i].locked = 1;
 830             return &ctx->in[i];
 831         }
 832     }
 833
 834     return NULL;
 835 }
 836
 837 static NVENCOutputSurface *get_output_surface(NVENCContext *ctx)
 838 {
 839     int i;
 840
 841     for (i = 0; i < ctx->nb_surfaces; i++) {
 842         if (!ctx->out[i].busy) {
 843             return &ctx->out[i];
 844         }
 845     }
 846
 847     return NULL;
 848 }
 849
 850 static int nvenc_copy_frame(NV_ENC_LOCK_INPUT_BUFFER *in, const AVFrame *frame)
 851 {
 852     uint8_t *buf = in->bufferDataPtr;
 853     int off      = frame->height * in->pitch;
 854
 855     switch (frame->format) {
 856     case AV_PIX_FMT_YUV420P:
 857         av_image_copy_plane(buf, in->pitch,
 858                             frame->data[0], frame->linesize[0],
 859                             frame->width, frame->height);
 860         buf += off;
 861
 862         av_image_copy_plane(buf, in->pitch >> 1,
 863                             frame->data[2], frame->linesize[2],
 864                             frame->width >> 1, frame->height >> 1);
 865
 866         buf += off >> 2;
 867
 868         av_image_copy_plane(buf, in->pitch >> 1,
 869                             frame->data[1], frame->linesize[1],
 870                             frame->width >> 1, frame->height >> 1);
 871         break;
 872     case AV_PIX_FMT_NV12:
 873         av_image_copy_plane(buf, in->pitch,
 874                             frame->data[0], frame->linesize[0],
 875                             frame->width, frame->height);
 876         buf += off;
 877
 878         av_image_copy_plane(buf, in->pitch,
 879                             frame->data[1], frame->linesize[1],
 880                             frame->width, frame->height >> 1);
 881         break;
 882     case AV_PIX_FMT_YUV444P:
 883         av_image_copy_plane(buf, in->pitch,
 884                             frame->data[0], frame->linesize[0],
 885                             frame->width, frame->height);
 886         buf += off;
 887
 888         av_image_copy_plane(buf, in->pitch,
 889                             frame->data[1], frame->linesize[1],
 890                             frame->width, frame->height);
 891         buf += off;
 892
 893         av_image_copy_plane(buf, in->pitch,
 894                             frame->data[2], frame->linesize[2],
 895                             frame->width, frame->height);
 896         break;
 897     default:
 898         return AVERROR_BUG;
 899     }
 900
 901     return 0;
 902 }
 903
 904 static int nvenc_enqueue_frame(AVCodecContext *avctx, const AVFrame *frame,
 905                                NVENCInputSurface **in_surf)
 906 {
 907     NVENCContext *ctx               = avctx->priv_data;
 908     NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
 909     NV_ENC_LOCK_INPUT_BUFFER params = { 0 };
 910     NVENCInputSurface *in           = get_input_surface(ctx);
 911     int ret;
 912
 913     if (!in)
 914         return AVERROR_BUG;
 915
 916     params.version     = NV_ENC_LOCK_INPUT_BUFFER_VER;
 917     params.inputBuffer = in->in;
 918
 919
 920     ret = nv->nvEncLockInputBuffer(ctx->nvenc_ctx, &params);
 921     if (ret != NV_ENC_SUCCESS) {
 922         av_log(avctx, AV_LOG_ERROR, "Cannot lock the buffer %p.\n",
 923                in);
 924         return AVERROR_UNKNOWN;
 925     }
 926
 927     ret = nvenc_copy_frame(&params, frame);
 928     if (ret < 0)
 929         goto fail;
 930
 931     ret = nv->nvEncUnlockInputBuffer(ctx->nvenc_ctx, in->in);
 932     if (ret != NV_ENC_SUCCESS) {
 933         av_log(avctx, AV_LOG_ERROR, "Cannot unlock the buffer %p.\n",
 934                in);
 935         return AVERROR_UNKNOWN;
 936     }
 937
 938     *in_surf = in;
 939
 940     return 0;
 941
 942 fail:
 943     nv->nvEncUnlockInputBuffer(ctx->nvenc_ctx, in->in);
 944
 945     return ret;
 946 }
 947
 948 static void nvenc_codec_specific_pic_params(AVCodecContext *avctx,
 949                                             NV_ENC_PIC_PARAMS *params)
 950 {
 951     NVENCContext *ctx = avctx->priv_data;
 952
 953     switch (avctx->codec->id) {
 954     case AV_CODEC_ID_H264:
 955         params->codecPicParams.h264PicParams.sliceMode =
 956             ctx->config.encodeCodecConfig.h264Config.sliceMode;
 957         params->codecPicParams.h264PicParams.sliceModeData =
 958             ctx->config.encodeCodecConfig.h264Config.sliceModeData;
 959         break;
 960     case AV_CODEC_ID_HEVC:
 961         params->codecPicParams.hevcPicParams.sliceMode =
 962             ctx->config.encodeCodecConfig.hevcConfig.sliceMode;
 963         params->codecPicParams.hevcPicParams.sliceModeData =
 964             ctx->config.encodeCodecConfig.hevcConfig.sliceModeData;
 965         break;
 966     }
 967 }
 968
 969 static inline int nvenc_enqueue_timestamp(AVFifoBuffer *f, int64_t pts)
 970 {
 971     return av_fifo_generic_write(f, &pts, sizeof(pts), NULL);
 972 }
 973
 974 static inline int nvenc_dequeue_timestamp(AVFifoBuffer *f, int64_t *pts)
 975 {
 976     return av_fifo_generic_read(f, pts, sizeof(*pts), NULL);
 977 }
 978
 979 static inline int nvenc_enqueue_surface(AVFifoBuffer *f,
 980                                         NVENCOutputSurface *surf)
 981 {
 982     surf->busy = 1;
 983     return av_fifo_generic_write(f, &surf, sizeof(surf), NULL);
 984 }
 985
 986 static inline int nvenc_dequeue_surface(AVFifoBuffer *f,
 987                                         NVENCOutputSurface **surf)
 988 {
 989     return av_fifo_generic_read(f, surf, sizeof(*surf), NULL);
 990 }
 991
 992 static int nvenc_set_timestamp(NVENCContext *ctx,
 993                                NV_ENC_LOCK_BITSTREAM *params,
 994                                AVPacket *pkt)
 995 {
 996     pkt->pts      = params->outputTimeStamp;
 997     pkt->duration = params->outputDuration;
 998
 999     return nvenc_dequeue_timestamp(ctx->timestamps, &pkt->dts);
1000 }
1001
1002 static int nvenc_get_frame(AVCodecContext *avctx, AVPacket *pkt)
1003 {
1004     NVENCContext *ctx               = avctx->priv_data;
1005     NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
1006     NV_ENC_LOCK_BITSTREAM params    = { 0 };
1007     NVENCOutputSurface *out         = NULL;
1008     int ret;
1009
1010     ret = nvenc_dequeue_surface(ctx->pending, &out);
1011     if (ret)
1012         return ret;
1013
1014     params.version         = NV_ENC_LOCK_BITSTREAM_VER;
1015     params.outputBitstream = out->out;
1016
1017     ret = nv->nvEncLockBitstream(ctx->nvenc_ctx, &params);
1018     if (ret < 0)
1019         return AVERROR_UNKNOWN;
1020
1021     ret = ff_alloc_packet(pkt, params.bitstreamSizeInBytes);
1022     if (ret < 0)
1023         return ret;
1024
1025     memcpy(pkt->data, params.bitstreamBufferPtr, pkt->size);
1026
1027     ret = nv->nvEncUnlockBitstream(ctx->nvenc_ctx, out->out);
1028     if (ret < 0)
1029         return AVERROR_UNKNOWN;
1030
1031     out->busy = out->in->locked = 0;
1032
1033     ret = nvenc_set_timestamp(ctx, &params, pkt);
1034     if (ret < 0)
1035         return ret;
1036
1037     switch (params.pictureType) {
1038     case NV_ENC_PIC_TYPE_IDR:
1039         pkt->flags |= AV_PKT_FLAG_KEY;
1040     case NV_ENC_PIC_TYPE_INTRA_REFRESH:
1041     case NV_ENC_PIC_TYPE_I:
1042         avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
1043         break;
1044     case NV_ENC_PIC_TYPE_P:
1045         avctx->coded_frame->pict_type = AV_PICTURE_TYPE_P;
1046         break;
1047     case NV_ENC_PIC_TYPE_B:
1048         avctx->coded_frame->pict_type = AV_PICTURE_TYPE_B;
1049         break;
1050     case NV_ENC_PIC_TYPE_BI:
1051         avctx->coded_frame->pict_type = AV_PICTURE_TYPE_BI;
1052         break;
1053     }
1054
1055     return 0;
1056 }
1057
1058 int ff_nvenc_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1059                           const AVFrame *frame, int *got_packet)
1060 {
1061     NVENCContext *ctx               = avctx->priv_data;
1062     NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
1063     NV_ENC_PIC_PARAMS params        = { 0 };
1064     NVENCInputSurface *in           = NULL;
1065     NVENCOutputSurface *out         = NULL;
1066     int ret;
1067
1068     params.version = NV_ENC_PIC_PARAMS_VER;
1069
1070     if (frame) {
1071         ret = nvenc_enqueue_frame(avctx, frame, &in);
1072         if (ret < 0)
1073             return ret;
1074         out = get_output_surface(ctx);
1075         if (!out)
1076             return AVERROR_BUG;
1077
1078         out->in = in;
1079
1080         params.inputBuffer     = in->in;
1081         params.bufferFmt       = in->format;
1082         params.inputWidth      = frame->width;
1083         params.inputHeight     = frame->height;
1084         params.outputBitstream = out->out;
1085         params.inputTimeStamp  = frame->pts;
1086
1087         if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
1088             if (frame->top_field_first)
1089                 params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_TOP_BOTTOM;
1090             else
1091                 params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_BOTTOM_TOP;
1092         } else {
1093             params.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
1094         }
1095
1096         nvenc_codec_specific_pic_params(avctx, &params);
1097
1098         ret = nvenc_enqueue_timestamp(ctx->timestamps, frame->pts);
1099         if (ret < 0)
1100             return ret;
1101     } else {
1102         params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
1103     }
1104
1105     ret = nv->nvEncEncodePicture(ctx->nvenc_ctx, &params);
1106
1107     if (ret != NV_ENC_SUCCESS &&
1108         ret != NV_ENC_ERR_NEED_MORE_INPUT) {
1109
1110         return AVERROR_UNKNOWN;
1111     }
1112
1113     if (out) {
1114         ret = nvenc_enqueue_surface(ctx->pending, out);
1115         if (ret < 0)
1116             return ret;
1117     }
1118
1119     if (ret != NV_ENC_ERR_NEED_MORE_INPUT &&
1120         av_fifo_size(ctx->pending)) {
1121         ret = nvenc_get_frame(avctx, pkt);
1122         if (ret < 0)
1123             return ret;
1124         *got_packet = 1;
1125     } else {
1126         *got_packet = 0;
1127     }
1128
1129     return 0;
1130 }