obs-nvenc.c 45 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686
  1. #include "obs-nvenc.h"
  2. #include <util/deque.h>
  3. #include <util/darray.h>
  4. #include <util/dstr.h>
  5. #include <obs-avc.h>
  6. #include <libavutil/rational.h>
  7. #define INITGUID
  8. #include <dxgi.h>
  9. #include <d3d11.h>
  10. #include <d3d11_1.h>
  11. #include <obs-hevc.h>
  12. /* ========================================================================= */
  13. /* a hack of the ages: nvenc backward compatibility */
  14. #define CONFIGURED_NVENC_MAJOR 12
  15. #define CONFIGURED_NVENC_MINOR 1
  16. #define CONFIGURED_NVENC_VER \
  17. (CONFIGURED_NVENC_MAJOR | (CONFIGURED_NVENC_MINOR << 24))
  18. /* we cannot guarantee structures haven't changed, so purposely break on
  19. * version change to force the programmer to update or remove backward
  20. * compatibility NVENC code. */
  21. #if CONFIGURED_NVENC_VER != NVENCAPI_VERSION
  22. #error NVENC version changed, update or remove NVENC compatibility code
  23. #endif
  24. #undef NVENCAPI_STRUCT_VERSION
  25. #define NVENCAPI_STRUCT_VERSION(ver) \
  26. ((uint32_t)(enc->needs_compat_ver ? NVENC_COMPAT_VER \
  27. : NVENCAPI_VERSION) | \
  28. ((ver) << 16) | (0x7 << 28))
  29. #define NV_ENC_CONFIG_COMPAT_VER (NVENCAPI_STRUCT_VERSION(7) | (1 << 31))
  30. #define NV_ENC_PIC_PARAMS_COMPAT_VER (NVENCAPI_STRUCT_VERSION(4) | (1 << 31))
  31. #define NV_ENC_LOCK_BITSTREAM_COMPAT_VER NVENCAPI_STRUCT_VERSION(1)
  32. #define NV_ENC_REGISTER_RESOURCE_COMPAT_VER NVENCAPI_STRUCT_VERSION(3)
  33. #define COMPATIBILITY_VERSION \
  34. (NVENC_COMPAT_MAJOR_VER << 4 | NVENC_COMPAT_MINOR_VER)
  35. /* ========================================================================= */
  36. #define EXTRA_BUFFERS 5
  37. #define do_log(level, format, ...) \
  38. blog(level, "[obs-nvenc: '%s'] " format, \
  39. obs_encoder_get_name(enc->encoder), ##__VA_ARGS__)
  40. #define error(format, ...) do_log(LOG_ERROR, format, ##__VA_ARGS__)
  41. #define warn(format, ...) do_log(LOG_WARNING, format, ##__VA_ARGS__)
  42. #define info(format, ...) do_log(LOG_INFO, format, ##__VA_ARGS__)
  43. #define debug(format, ...) do_log(LOG_DEBUG, format, ##__VA_ARGS__)
  44. #define error_hr(msg) error("%s: %s: 0x%08lX", __FUNCTION__, msg, (uint32_t)hr);
  45. struct nv_bitstream;
  46. struct nv_texture;
  47. struct handle_tex {
  48. uint32_t handle;
  49. ID3D11Texture2D *tex;
  50. IDXGIKeyedMutex *km;
  51. };
  52. /* ------------------------------------------------------------------------- */
  53. /* Main Implementation Structure */
  54. enum codec_type {
  55. CODEC_H264,
  56. CODEC_HEVC,
  57. CODEC_AV1,
  58. };
  59. static const char *get_codec_name(enum codec_type type)
  60. {
  61. switch (type) {
  62. case CODEC_H264:
  63. return "H264";
  64. case CODEC_HEVC:
  65. return "HEVC";
  66. case CODEC_AV1:
  67. return "AV1";
  68. }
  69. return "Unknown";
  70. }
  71. struct nvenc_data {
  72. obs_encoder_t *encoder;
  73. enum codec_type codec;
  74. GUID codec_guid;
  75. void *session;
  76. NV_ENC_INITIALIZE_PARAMS params;
  77. NV_ENC_CONFIG config;
  78. int rc_lookahead;
  79. int buf_count;
  80. int output_delay;
  81. int buffers_queued;
  82. size_t next_bitstream;
  83. size_t cur_bitstream;
  84. bool encode_started;
  85. bool first_packet;
  86. bool can_change_bitrate;
  87. bool needs_compat_ver;
  88. int32_t bframes;
  89. DARRAY(struct nv_bitstream) bitstreams;
  90. DARRAY(struct nv_texture) textures;
  91. DARRAY(struct handle_tex) input_textures;
  92. struct deque dts_list;
  93. DARRAY(uint8_t) packet_data;
  94. int64_t packet_pts;
  95. bool packet_keyframe;
  96. ID3D11Device *device;
  97. ID3D11DeviceContext *context;
  98. uint32_t cx;
  99. uint32_t cy;
  100. uint8_t *header;
  101. size_t header_size;
  102. uint8_t *sei;
  103. size_t sei_size;
  104. int8_t *roi_map;
  105. size_t roi_map_size;
  106. uint32_t roi_increment;
  107. };
  108. /* ------------------------------------------------------------------------- */
  109. /* Bitstream Buffer */
  110. struct nv_bitstream {
  111. void *ptr;
  112. };
  113. #define NV_FAIL(format, ...) nv_fail(enc->encoder, format, ##__VA_ARGS__)
  114. #define NV_FAILED(x) nv_failed(enc->encoder, x, __FUNCTION__, #x)
  115. static bool nv_bitstream_init(struct nvenc_data *enc, struct nv_bitstream *bs)
  116. {
  117. NV_ENC_CREATE_BITSTREAM_BUFFER buf = {
  118. NV_ENC_CREATE_BITSTREAM_BUFFER_VER};
  119. if (NV_FAILED(nv.nvEncCreateBitstreamBuffer(enc->session, &buf))) {
  120. return false;
  121. }
  122. bs->ptr = buf.bitstreamBuffer;
  123. return true;
  124. }
  125. static void nv_bitstream_free(struct nvenc_data *enc, struct nv_bitstream *bs)
  126. {
  127. if (bs->ptr) {
  128. nv.nvEncDestroyBitstreamBuffer(enc->session, bs->ptr);
  129. }
  130. }
  131. /* ------------------------------------------------------------------------- */
  132. /* Texture Resource */
  133. struct nv_texture {
  134. void *res;
  135. ID3D11Texture2D *tex;
  136. void *mapped_res;
  137. };
  138. static bool nv_texture_init(struct nvenc_data *enc, struct nv_texture *nvtex)
  139. {
  140. const bool p010 = obs_p010_tex_active();
  141. D3D11_TEXTURE2D_DESC desc = {0};
  142. desc.Width = enc->cx;
  143. desc.Height = enc->cy;
  144. desc.MipLevels = 1;
  145. desc.ArraySize = 1;
  146. desc.Format = p010 ? DXGI_FORMAT_P010 : DXGI_FORMAT_NV12;
  147. desc.SampleDesc.Count = 1;
  148. desc.BindFlags = D3D11_BIND_RENDER_TARGET;
  149. ID3D11Device *const device = enc->device;
  150. ID3D11Texture2D *tex;
  151. HRESULT hr = device->lpVtbl->CreateTexture2D(device, &desc, NULL, &tex);
  152. if (FAILED(hr)) {
  153. error_hr("Failed to create texture");
  154. return false;
  155. }
  156. tex->lpVtbl->SetEvictionPriority(tex, DXGI_RESOURCE_PRIORITY_MAXIMUM);
  157. uint32_t struct_ver = enc->needs_compat_ver
  158. ? NV_ENC_REGISTER_RESOURCE_COMPAT_VER
  159. : NV_ENC_REGISTER_RESOURCE_VER;
  160. NV_ENC_REGISTER_RESOURCE res = {struct_ver};
  161. res.resourceType = NV_ENC_INPUT_RESOURCE_TYPE_DIRECTX;
  162. res.resourceToRegister = tex;
  163. res.width = enc->cx;
  164. res.height = enc->cy;
  165. res.bufferFormat = p010 ? NV_ENC_BUFFER_FORMAT_YUV420_10BIT
  166. : NV_ENC_BUFFER_FORMAT_NV12;
  167. if (NV_FAILED(nv.nvEncRegisterResource(enc->session, &res))) {
  168. tex->lpVtbl->Release(tex);
  169. return false;
  170. }
  171. nvtex->res = res.registeredResource;
  172. nvtex->tex = tex;
  173. nvtex->mapped_res = NULL;
  174. return true;
  175. }
  176. static void nv_texture_free(struct nvenc_data *enc, struct nv_texture *nvtex)
  177. {
  178. if (nvtex->res) {
  179. if (nvtex->mapped_res) {
  180. nv.nvEncUnmapInputResource(enc->session,
  181. nvtex->mapped_res);
  182. }
  183. nv.nvEncUnregisterResource(enc->session, nvtex->res);
  184. nvtex->tex->lpVtbl->Release(nvtex->tex);
  185. }
  186. }
  187. /* ------------------------------------------------------------------------- */
  188. /* Implementation */
  189. static const char *h264_nvenc_get_name(void *type_data)
  190. {
  191. UNUSED_PARAMETER(type_data);
  192. return "NVIDIA NVENC H.264";
  193. }
  194. #ifdef ENABLE_HEVC
  195. static const char *hevc_nvenc_get_name(void *type_data)
  196. {
  197. UNUSED_PARAMETER(type_data);
  198. return "NVIDIA NVENC HEVC";
  199. }
  200. #endif
  201. static const char *av1_nvenc_get_name(void *type_data)
  202. {
  203. UNUSED_PARAMETER(type_data);
  204. return "NVIDIA NVENC AV1";
  205. }
  206. static inline int nv_get_cap(struct nvenc_data *enc, NV_ENC_CAPS cap)
  207. {
  208. if (!enc->session)
  209. return 0;
  210. NV_ENC_CAPS_PARAM param = {NV_ENC_CAPS_PARAM_VER};
  211. int v;
  212. param.capsToQuery = cap;
  213. nv.nvEncGetEncodeCaps(enc->session, enc->codec_guid, &param, &v);
  214. return v;
  215. }
  216. static bool nvenc_update(void *data, obs_data_t *settings)
  217. {
  218. struct nvenc_data *enc = data;
  219. /* Only support reconfiguration of CBR bitrate */
  220. if (enc->can_change_bitrate) {
  221. int bitrate = (int)obs_data_get_int(settings, "bitrate");
  222. int max_bitrate =
  223. (int)obs_data_get_int(settings, "max_bitrate");
  224. bool vbr = (enc->config.rcParams.rateControlMode ==
  225. NV_ENC_PARAMS_RC_VBR);
  226. enc->config.rcParams.averageBitRate = bitrate * 1000;
  227. enc->config.rcParams.maxBitRate = vbr ? max_bitrate * 1000
  228. : bitrate * 1000;
  229. NV_ENC_RECONFIGURE_PARAMS params = {0};
  230. params.version = NV_ENC_RECONFIGURE_PARAMS_VER;
  231. params.reInitEncodeParams = enc->params;
  232. params.resetEncoder = 1;
  233. params.forceIDR = 1;
  234. if (NV_FAILED(nv.nvEncReconfigureEncoder(enc->session,
  235. &params))) {
  236. return false;
  237. }
  238. }
  239. return true;
  240. }
  241. static HANDLE get_lib(struct nvenc_data *enc, const char *lib)
  242. {
  243. HMODULE mod = GetModuleHandleA(lib);
  244. if (mod)
  245. return mod;
  246. mod = LoadLibraryA(lib);
  247. if (!mod)
  248. error("Failed to load %s", lib);
  249. return mod;
  250. }
  251. typedef HRESULT(WINAPI *CREATEDXGIFACTORY1PROC)(REFIID, void **);
  252. static bool init_d3d11(struct nvenc_data *enc, obs_data_t *settings)
  253. {
  254. HMODULE dxgi = get_lib(enc, "DXGI.dll");
  255. HMODULE d3d11 = get_lib(enc, "D3D11.dll");
  256. CREATEDXGIFACTORY1PROC create_dxgi;
  257. PFN_D3D11_CREATE_DEVICE create_device;
  258. IDXGIFactory1 *factory;
  259. IDXGIAdapter *adapter;
  260. ID3D11Device *device;
  261. ID3D11DeviceContext *context;
  262. HRESULT hr;
  263. if (!dxgi || !d3d11) {
  264. return false;
  265. }
  266. create_dxgi = (CREATEDXGIFACTORY1PROC)GetProcAddress(
  267. dxgi, "CreateDXGIFactory1");
  268. create_device = (PFN_D3D11_CREATE_DEVICE)GetProcAddress(
  269. d3d11, "D3D11CreateDevice");
  270. if (!create_dxgi || !create_device) {
  271. error("Failed to load D3D11/DXGI procedures");
  272. return false;
  273. }
  274. hr = create_dxgi(&IID_IDXGIFactory1, &factory);
  275. if (FAILED(hr)) {
  276. error_hr("CreateDXGIFactory1 failed");
  277. return false;
  278. }
  279. hr = factory->lpVtbl->EnumAdapters(factory, 0, &adapter);
  280. factory->lpVtbl->Release(factory);
  281. if (FAILED(hr)) {
  282. error_hr("EnumAdapters failed");
  283. return false;
  284. }
  285. hr = create_device(adapter, D3D_DRIVER_TYPE_UNKNOWN, NULL, 0, NULL, 0,
  286. D3D11_SDK_VERSION, &device, NULL, &context);
  287. adapter->lpVtbl->Release(adapter);
  288. if (FAILED(hr)) {
  289. error_hr("D3D11CreateDevice failed");
  290. return false;
  291. }
  292. enc->device = device;
  293. enc->context = context;
  294. return true;
  295. }
  296. static bool init_session(struct nvenc_data *enc)
  297. {
  298. NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS params = {
  299. NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER};
  300. params.device = enc->device;
  301. params.deviceType = NV_ENC_DEVICE_TYPE_DIRECTX;
  302. params.apiVersion = enc->needs_compat_ver ? NVENC_COMPAT_VER
  303. : NVENCAPI_VERSION;
  304. if (NV_FAILED(nv.nvEncOpenEncodeSessionEx(&params, &enc->session))) {
  305. return false;
  306. }
  307. return true;
  308. }
  309. static void initialize_params(struct nvenc_data *enc, const GUID *nv_preset,
  310. NV_ENC_TUNING_INFO nv_tuning, uint32_t width,
  311. uint32_t height, uint32_t fps_num,
  312. uint32_t fps_den)
  313. {
  314. int darWidth, darHeight;
  315. av_reduce(&darWidth, &darHeight, width, height, 1024 * 1024);
  316. NV_ENC_INITIALIZE_PARAMS *params = &enc->params;
  317. memset(params, 0, sizeof(*params));
  318. params->version = NV_ENC_INITIALIZE_PARAMS_VER;
  319. params->encodeGUID = enc->codec_guid;
  320. params->presetGUID = *nv_preset;
  321. params->encodeWidth = width;
  322. params->encodeHeight = height;
  323. params->darWidth = enc->codec == CODEC_AV1 ? width : darWidth;
  324. params->darHeight = enc->codec == CODEC_AV1 ? height : darHeight;
  325. params->frameRateNum = fps_num;
  326. params->frameRateDen = fps_den;
  327. params->enableEncodeAsync = 0;
  328. params->enablePTD = 1;
  329. params->encodeConfig = &enc->config;
  330. params->tuningInfo = nv_tuning;
  331. }
  332. static inline GUID get_nv_preset2(const char *preset2)
  333. {
  334. if (astrcmpi(preset2, "p1") == 0) {
  335. return NV_ENC_PRESET_P1_GUID;
  336. } else if (astrcmpi(preset2, "p2") == 0) {
  337. return NV_ENC_PRESET_P2_GUID;
  338. } else if (astrcmpi(preset2, "p3") == 0) {
  339. return NV_ENC_PRESET_P3_GUID;
  340. } else if (astrcmpi(preset2, "p4") == 0) {
  341. return NV_ENC_PRESET_P4_GUID;
  342. } else if (astrcmpi(preset2, "p6") == 0) {
  343. return NV_ENC_PRESET_P6_GUID;
  344. } else if (astrcmpi(preset2, "p7") == 0) {
  345. return NV_ENC_PRESET_P7_GUID;
  346. } else {
  347. return NV_ENC_PRESET_P5_GUID;
  348. }
  349. }
  350. static inline NV_ENC_TUNING_INFO get_nv_tuning(const char *tuning)
  351. {
  352. if (astrcmpi(tuning, "ll") == 0) {
  353. return NV_ENC_TUNING_INFO_LOW_LATENCY;
  354. } else if (astrcmpi(tuning, "ull") == 0) {
  355. return NV_ENC_TUNING_INFO_ULTRA_LOW_LATENCY;
  356. } else {
  357. return NV_ENC_TUNING_INFO_HIGH_QUALITY;
  358. }
  359. }
  360. static inline NV_ENC_MULTI_PASS get_nv_multipass(const char *multipass)
  361. {
  362. if (astrcmpi(multipass, "qres") == 0) {
  363. return NV_ENC_TWO_PASS_QUARTER_RESOLUTION;
  364. } else if (astrcmpi(multipass, "fullres") == 0) {
  365. return NV_ENC_TWO_PASS_FULL_RESOLUTION;
  366. } else {
  367. return NV_ENC_MULTI_PASS_DISABLED;
  368. }
  369. }
  370. static bool init_encoder_base(struct nvenc_data *enc, obs_data_t *settings,
  371. int bf, bool compatibility, bool *lossless)
  372. {
  373. const char *rc = obs_data_get_string(settings, "rate_control");
  374. int bitrate = (int)obs_data_get_int(settings, "bitrate");
  375. int max_bitrate = (int)obs_data_get_int(settings, "max_bitrate");
  376. int cqp = (int)obs_data_get_int(settings, "cqp");
  377. int keyint_sec = (int)obs_data_get_int(settings, "keyint_sec");
  378. const char *preset = obs_data_get_string(settings, "preset");
  379. const char *preset2 = obs_data_get_string(settings, "preset2");
  380. const char *tuning = obs_data_get_string(settings, "tune");
  381. const char *multipass = obs_data_get_string(settings, "multipass");
  382. const char *profile = obs_data_get_string(settings, "profile");
  383. bool lookahead = obs_data_get_bool(settings, "lookahead");
  384. bool vbr = astrcmpi(rc, "VBR") == 0;
  385. bool psycho_aq = !compatibility &&
  386. obs_data_get_bool(settings, "psycho_aq");
  387. bool disable_scenecut = obs_data_get_bool(settings, "disable_scenecut");
  388. NVENCSTATUS err;
  389. video_t *video = obs_encoder_video(enc->encoder);
  390. const struct video_output_info *voi = video_output_get_info(video);
  391. enc->cx = obs_encoder_get_width(enc->encoder);
  392. enc->cy = obs_encoder_get_height(enc->encoder);
  393. /* -------------------------- */
  394. /* get preset */
  395. GUID nv_preset = get_nv_preset2(preset2);
  396. NV_ENC_TUNING_INFO nv_tuning = get_nv_tuning(tuning);
  397. NV_ENC_MULTI_PASS nv_multipass = compatibility
  398. ? NV_ENC_MULTI_PASS_DISABLED
  399. : get_nv_multipass(multipass);
  400. if (obs_data_has_user_value(settings, "preset") &&
  401. !obs_data_has_user_value(settings, "preset2") &&
  402. enc->codec == CODEC_H264) {
  403. if (astrcmpi(preset, "mq") == 0) {
  404. nv_preset = NV_ENC_PRESET_P5_GUID;
  405. nv_tuning = NV_ENC_TUNING_INFO_HIGH_QUALITY;
  406. nv_multipass = NV_ENC_TWO_PASS_QUARTER_RESOLUTION;
  407. } else if (astrcmpi(preset, "hq") == 0) {
  408. nv_preset = NV_ENC_PRESET_P5_GUID;
  409. nv_tuning = NV_ENC_TUNING_INFO_HIGH_QUALITY;
  410. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  411. } else if (astrcmpi(preset, "default") == 0) {
  412. nv_preset = NV_ENC_PRESET_P3_GUID;
  413. nv_tuning = NV_ENC_TUNING_INFO_HIGH_QUALITY;
  414. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  415. } else if (astrcmpi(preset, "hp") == 0) {
  416. nv_preset = NV_ENC_PRESET_P1_GUID;
  417. nv_tuning = NV_ENC_TUNING_INFO_HIGH_QUALITY;
  418. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  419. } else if (astrcmpi(preset, "ll") == 0) {
  420. nv_preset = NV_ENC_PRESET_P3_GUID;
  421. nv_tuning = NV_ENC_TUNING_INFO_LOW_LATENCY;
  422. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  423. } else if (astrcmpi(preset, "llhq") == 0) {
  424. nv_preset = NV_ENC_PRESET_P4_GUID;
  425. nv_tuning = NV_ENC_TUNING_INFO_LOW_LATENCY;
  426. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  427. } else if (astrcmpi(preset, "llhp") == 0) {
  428. nv_preset = NV_ENC_PRESET_P2_GUID;
  429. nv_tuning = NV_ENC_TUNING_INFO_LOW_LATENCY;
  430. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  431. }
  432. } else if (obs_data_has_user_value(settings, "preset") &&
  433. !obs_data_has_user_value(settings, "preset2") &&
  434. enc->codec == CODEC_HEVC) {
  435. if (astrcmpi(preset, "mq") == 0) {
  436. nv_preset = NV_ENC_PRESET_P6_GUID;
  437. nv_tuning = NV_ENC_TUNING_INFO_HIGH_QUALITY;
  438. nv_multipass = NV_ENC_TWO_PASS_QUARTER_RESOLUTION;
  439. } else if (astrcmpi(preset, "hq") == 0) {
  440. nv_preset = NV_ENC_PRESET_P6_GUID;
  441. nv_tuning = NV_ENC_TUNING_INFO_HIGH_QUALITY;
  442. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  443. } else if (astrcmpi(preset, "default") == 0) {
  444. nv_preset = NV_ENC_PRESET_P5_GUID;
  445. nv_tuning = NV_ENC_TUNING_INFO_HIGH_QUALITY;
  446. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  447. } else if (astrcmpi(preset, "hp") == 0) {
  448. nv_preset = NV_ENC_PRESET_P1_GUID;
  449. nv_tuning = NV_ENC_TUNING_INFO_HIGH_QUALITY;
  450. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  451. } else if (astrcmpi(preset, "ll") == 0) {
  452. nv_preset = NV_ENC_PRESET_P3_GUID;
  453. nv_tuning = NV_ENC_TUNING_INFO_LOW_LATENCY;
  454. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  455. } else if (astrcmpi(preset, "llhq") == 0) {
  456. nv_preset = NV_ENC_PRESET_P4_GUID;
  457. nv_tuning = NV_ENC_TUNING_INFO_LOW_LATENCY;
  458. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  459. } else if (astrcmpi(preset, "llhp") == 0) {
  460. nv_preset = NV_ENC_PRESET_P2_GUID;
  461. nv_tuning = NV_ENC_TUNING_INFO_LOW_LATENCY;
  462. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  463. }
  464. }
  465. const bool rc_lossless = astrcmpi(rc, "lossless") == 0;
  466. *lossless = rc_lossless;
  467. if (rc_lossless) {
  468. *lossless =
  469. nv_get_cap(enc, NV_ENC_CAPS_SUPPORT_LOSSLESS_ENCODE);
  470. if (*lossless) {
  471. nv_tuning = NV_ENC_TUNING_INFO_LOSSLESS;
  472. nv_multipass = NV_ENC_MULTI_PASS_DISABLED;
  473. } else {
  474. warn("lossless encode is not supported, ignoring");
  475. nv_preset = NV_ENC_PRESET_P5_GUID;
  476. nv_tuning = NV_ENC_TUNING_INFO_HIGH_QUALITY;
  477. nv_multipass = NV_ENC_TWO_PASS_QUARTER_RESOLUTION;
  478. }
  479. }
  480. /* -------------------------- */
  481. /* get preset default config */
  482. uint32_t config_ver = enc->needs_compat_ver ? NV_ENC_CONFIG_COMPAT_VER
  483. : NV_ENC_CONFIG_VER;
  484. NV_ENC_PRESET_CONFIG preset_config = {NV_ENC_PRESET_CONFIG_VER,
  485. {config_ver}};
  486. err = nv.nvEncGetEncodePresetConfigEx(enc->session, enc->codec_guid,
  487. nv_preset, nv_tuning,
  488. &preset_config);
  489. if (nv_failed(enc->encoder, err, __FUNCTION__,
  490. "nvEncGetEncodePresetConfig")) {
  491. return false;
  492. }
  493. /* -------------------------- */
  494. /* main configuration */
  495. enc->config = preset_config.presetCfg;
  496. uint32_t gop_size =
  497. (keyint_sec) ? keyint_sec * voi->fps_num / voi->fps_den : 250;
  498. NV_ENC_CONFIG *config = &enc->config;
  499. initialize_params(enc, &nv_preset, nv_tuning, voi->width, voi->height,
  500. voi->fps_num, voi->fps_den);
  501. config->gopLength = gop_size;
  502. config->frameIntervalP = 1 + bf;
  503. enc->bframes = bf;
  504. /* lookahead */
  505. const bool use_profile_lookahead = config->rcParams.enableLookahead;
  506. lookahead = nv_get_cap(enc, NV_ENC_CAPS_SUPPORT_LOOKAHEAD) &&
  507. (lookahead || use_profile_lookahead);
  508. if (lookahead) {
  509. enc->rc_lookahead = use_profile_lookahead
  510. ? config->rcParams.lookaheadDepth
  511. : 8;
  512. }
  513. int buf_count = max(4, config->frameIntervalP * 2 * 2);
  514. if (lookahead) {
  515. buf_count = max(buf_count, config->frameIntervalP +
  516. enc->rc_lookahead +
  517. EXTRA_BUFFERS);
  518. }
  519. buf_count = min(64, buf_count);
  520. enc->buf_count = buf_count;
  521. const int output_delay = buf_count - 1;
  522. enc->output_delay = output_delay;
  523. if (lookahead) {
  524. const int lkd_bound = output_delay - config->frameIntervalP - 4;
  525. if (lkd_bound >= 0) {
  526. config->rcParams.enableLookahead = 1;
  527. config->rcParams.lookaheadDepth =
  528. max(enc->rc_lookahead, lkd_bound);
  529. config->rcParams.disableIadapt = 0;
  530. config->rcParams.disableBadapt = 0;
  531. } else {
  532. lookahead = false;
  533. }
  534. }
  535. enc->config.rcParams.disableIadapt = disable_scenecut;
  536. /* psycho aq */
  537. if (!compatibility) {
  538. if (nv_get_cap(enc, NV_ENC_CAPS_SUPPORT_TEMPORAL_AQ)) {
  539. config->rcParams.enableAQ = psycho_aq;
  540. config->rcParams.aqStrength = 8;
  541. config->rcParams.enableTemporalAQ = psycho_aq;
  542. } else {
  543. warn("Ignoring Psycho Visual Tuning request since GPU is not capable");
  544. }
  545. }
  546. /* -------------------------- */
  547. /* rate control */
  548. enc->can_change_bitrate =
  549. nv_get_cap(enc, NV_ENC_CAPS_SUPPORT_DYN_BITRATE_CHANGE);
  550. config->rcParams.rateControlMode = NV_ENC_PARAMS_RC_VBR;
  551. if (astrcmpi(rc, "cqp") == 0 || rc_lossless) {
  552. if (*lossless)
  553. cqp = 0;
  554. int cqp_val = enc->codec == CODEC_AV1 ? cqp * 4 : cqp;
  555. config->rcParams.rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
  556. config->rcParams.constQP.qpInterP = cqp_val;
  557. config->rcParams.constQP.qpInterB = cqp_val;
  558. config->rcParams.constQP.qpIntra = cqp_val;
  559. enc->can_change_bitrate = false;
  560. bitrate = 0;
  561. max_bitrate = 0;
  562. } else if (astrcmpi(rc, "vbr") != 0) { /* CBR by default */
  563. config->rcParams.rateControlMode = NV_ENC_PARAMS_RC_CBR;
  564. }
  565. config->rcParams.averageBitRate = bitrate * 1000;
  566. config->rcParams.maxBitRate = vbr ? max_bitrate * 1000 : bitrate * 1000;
  567. config->rcParams.vbvBufferSize = bitrate * 1000;
  568. config->rcParams.multiPass = nv_multipass;
  569. config->rcParams.qpMapMode = NV_ENC_QP_MAP_DELTA;
  570. /* -------------------------- */
  571. /* initialize */
  572. info("settings:\n"
  573. "\tcodec: %s\n"
  574. "\trate_control: %s\n"
  575. "\tbitrate: %d\n"
  576. "\tcqp: %d\n"
  577. "\tkeyint: %d\n"
  578. "\tpreset: %s\n"
  579. "\ttuning: %s\n"
  580. "\tmultipass: %s\n"
  581. "\tprofile: %s\n"
  582. "\twidth: %d\n"
  583. "\theight: %d\n"
  584. "\tb-frames: %d\n"
  585. "\tlookahead: %s\n"
  586. "\tpsycho_aq: %s\n",
  587. get_codec_name(enc->codec), rc, bitrate, cqp, gop_size, preset2,
  588. tuning, multipass, profile, enc->cx, enc->cy, bf,
  589. lookahead ? "true" : "false", psycho_aq ? "true" : "false");
  590. return true;
  591. }
  592. static bool init_encoder_h264(struct nvenc_data *enc, obs_data_t *settings,
  593. int bf, bool compatibility)
  594. {
  595. const char *rc = obs_data_get_string(settings, "rate_control");
  596. int keyint_sec = (int)obs_data_get_int(settings, "keyint_sec");
  597. const char *profile = obs_data_get_string(settings, "profile");
  598. bool lossless;
  599. if (!init_encoder_base(enc, settings, bf, compatibility, &lossless)) {
  600. return false;
  601. }
  602. NV_ENC_CONFIG *config = &enc->config;
  603. NV_ENC_CONFIG_H264 *h264_config = &config->encodeCodecConfig.h264Config;
  604. NV_ENC_CONFIG_H264_VUI_PARAMETERS *vui_params =
  605. &h264_config->h264VUIParameters;
  606. video_t *video = obs_encoder_video(enc->encoder);
  607. const struct video_output_info *voi = video_output_get_info(video);
  608. uint32_t gop_size =
  609. (keyint_sec) ? keyint_sec * voi->fps_num / voi->fps_den : 250;
  610. h264_config->idrPeriod = gop_size;
  611. bool repeat_headers = obs_data_get_bool(settings, "repeat_headers");
  612. if (repeat_headers) {
  613. h264_config->repeatSPSPPS = 1;
  614. h264_config->disableSPSPPS = 0;
  615. h264_config->outputAUD = 1;
  616. }
  617. h264_config->sliceMode = 3;
  618. h264_config->sliceModeData = 1;
  619. h264_config->useBFramesAsRef = NV_ENC_BFRAME_REF_MODE_DISABLED;
  620. /* Enable CBR padding */
  621. if (config->rcParams.rateControlMode == NV_ENC_PARAMS_RC_CBR)
  622. h264_config->enableFillerDataInsertion = 1;
  623. vui_params->videoSignalTypePresentFlag = 1;
  624. vui_params->videoFullRangeFlag = (voi->range == VIDEO_RANGE_FULL);
  625. vui_params->colourDescriptionPresentFlag = 1;
  626. switch (voi->colorspace) {
  627. case VIDEO_CS_601:
  628. vui_params->colourPrimaries = 6;
  629. vui_params->transferCharacteristics = 6;
  630. vui_params->colourMatrix = 6;
  631. break;
  632. case VIDEO_CS_DEFAULT:
  633. case VIDEO_CS_709:
  634. vui_params->colourPrimaries = 1;
  635. vui_params->transferCharacteristics = 1;
  636. vui_params->colourMatrix = 1;
  637. break;
  638. case VIDEO_CS_SRGB:
  639. vui_params->colourPrimaries = 1;
  640. vui_params->transferCharacteristics = 13;
  641. vui_params->colourMatrix = 1;
  642. break;
  643. }
  644. if (astrcmpi(rc, "lossless") == 0) {
  645. h264_config->qpPrimeYZeroTransformBypassFlag = 1;
  646. } else if (astrcmpi(rc, "vbr") != 0) { /* CBR */
  647. h264_config->outputBufferingPeriodSEI = 1;
  648. }
  649. h264_config->outputPictureTimingSEI = 1;
  650. /* -------------------------- */
  651. /* profile */
  652. if (astrcmpi(profile, "main") == 0) {
  653. config->profileGUID = NV_ENC_H264_PROFILE_MAIN_GUID;
  654. } else if (astrcmpi(profile, "baseline") == 0) {
  655. config->profileGUID = NV_ENC_H264_PROFILE_BASELINE_GUID;
  656. } else if (!lossless) {
  657. config->profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
  658. }
  659. if (NV_FAILED(nv.nvEncInitializeEncoder(enc->session, &enc->params))) {
  660. return false;
  661. }
  662. return true;
  663. }
  664. static bool init_encoder_hevc(struct nvenc_data *enc, obs_data_t *settings,
  665. int bf, bool compatibility)
  666. {
  667. const char *rc = obs_data_get_string(settings, "rate_control");
  668. int keyint_sec = (int)obs_data_get_int(settings, "keyint_sec");
  669. const char *profile = obs_data_get_string(settings, "profile");
  670. bool lossless;
  671. if (!init_encoder_base(enc, settings, bf, compatibility, &lossless)) {
  672. return false;
  673. }
  674. NV_ENC_CONFIG *config = &enc->config;
  675. NV_ENC_CONFIG_HEVC *hevc_config = &config->encodeCodecConfig.hevcConfig;
  676. NV_ENC_CONFIG_HEVC_VUI_PARAMETERS *vui_params =
  677. &hevc_config->hevcVUIParameters;
  678. video_t *video = obs_encoder_video(enc->encoder);
  679. const struct video_output_info *voi = video_output_get_info(video);
  680. uint32_t gop_size =
  681. (keyint_sec) ? keyint_sec * voi->fps_num / voi->fps_den : 250;
  682. hevc_config->idrPeriod = gop_size;
  683. bool repeat_headers = obs_data_get_bool(settings, "repeat_headers");
  684. if (repeat_headers) {
  685. hevc_config->repeatSPSPPS = 1;
  686. hevc_config->disableSPSPPS = 0;
  687. hevc_config->outputAUD = 1;
  688. }
  689. hevc_config->sliceMode = 3;
  690. hevc_config->sliceModeData = 1;
  691. hevc_config->useBFramesAsRef = NV_ENC_BFRAME_REF_MODE_DISABLED;
  692. /* Enable CBR padding */
  693. if (config->rcParams.rateControlMode == NV_ENC_PARAMS_RC_CBR)
  694. hevc_config->enableFillerDataInsertion = 1;
  695. vui_params->videoSignalTypePresentFlag = 1;
  696. vui_params->videoFullRangeFlag = (voi->range == VIDEO_RANGE_FULL);
  697. vui_params->colourDescriptionPresentFlag = 1;
  698. switch (voi->colorspace) {
  699. case VIDEO_CS_601:
  700. vui_params->colourPrimaries = 6;
  701. vui_params->transferCharacteristics = 6;
  702. vui_params->colourMatrix = 6;
  703. break;
  704. case VIDEO_CS_DEFAULT:
  705. case VIDEO_CS_709:
  706. vui_params->colourPrimaries = 1;
  707. vui_params->transferCharacteristics = 1;
  708. vui_params->colourMatrix = 1;
  709. break;
  710. case VIDEO_CS_SRGB:
  711. vui_params->colourPrimaries = 1;
  712. vui_params->transferCharacteristics = 13;
  713. vui_params->colourMatrix = 1;
  714. break;
  715. case VIDEO_CS_2100_PQ:
  716. vui_params->colourPrimaries = 9;
  717. vui_params->transferCharacteristics = 16;
  718. vui_params->colourMatrix = 9;
  719. vui_params->chromaSampleLocationFlag = 1;
  720. vui_params->chromaSampleLocationTop = 2;
  721. vui_params->chromaSampleLocationBot = 2;
  722. break;
  723. case VIDEO_CS_2100_HLG:
  724. vui_params->colourPrimaries = 9;
  725. vui_params->transferCharacteristics = 18;
  726. vui_params->colourMatrix = 9;
  727. vui_params->chromaSampleLocationFlag = 1;
  728. vui_params->chromaSampleLocationTop = 2;
  729. vui_params->chromaSampleLocationBot = 2;
  730. }
  731. hevc_config->pixelBitDepthMinus8 = obs_p010_tex_active() ? 2 : 0;
  732. if (astrcmpi(rc, "cbr") == 0) {
  733. hevc_config->outputBufferingPeriodSEI = 1;
  734. }
  735. hevc_config->outputPictureTimingSEI = 1;
  736. /* -------------------------- */
  737. /* profile */
  738. if (astrcmpi(profile, "main10") == 0) {
  739. config->profileGUID = NV_ENC_HEVC_PROFILE_MAIN10_GUID;
  740. } else if (obs_p010_tex_active()) {
  741. blog(LOG_WARNING, "[obs-nvenc] Forcing main10 for P010");
  742. config->profileGUID = NV_ENC_HEVC_PROFILE_MAIN10_GUID;
  743. } else {
  744. config->profileGUID = NV_ENC_HEVC_PROFILE_MAIN_GUID;
  745. }
  746. if (NV_FAILED(nv.nvEncInitializeEncoder(enc->session, &enc->params))) {
  747. return false;
  748. }
  749. return true;
  750. }
  751. static bool init_encoder_av1(struct nvenc_data *enc, obs_data_t *settings,
  752. int bf, bool compatibility)
  753. {
  754. const char *rc = obs_data_get_string(settings, "rate_control");
  755. int keyint_sec = (int)obs_data_get_int(settings, "keyint_sec");
  756. bool lossless;
  757. if (!init_encoder_base(enc, settings, bf, compatibility, &lossless)) {
  758. return false;
  759. }
  760. NV_ENC_INITIALIZE_PARAMS *params = &enc->params;
  761. NV_ENC_CONFIG *config = &enc->config;
  762. NV_ENC_CONFIG_AV1 *av1_config = &config->encodeCodecConfig.av1Config;
  763. video_t *video = obs_encoder_video(enc->encoder);
  764. const struct video_output_info *voi = video_output_get_info(video);
  765. uint32_t gop_size =
  766. (keyint_sec) ? keyint_sec * voi->fps_num / voi->fps_den : 250;
  767. av1_config->idrPeriod = gop_size;
  768. av1_config->useBFramesAsRef = NV_ENC_BFRAME_REF_MODE_DISABLED;
  769. av1_config->colorRange = (voi->range == VIDEO_RANGE_FULL);
  770. /* Enable CBR padding */
  771. if (config->rcParams.rateControlMode == NV_ENC_PARAMS_RC_CBR)
  772. av1_config->enableBitstreamPadding = 1;
  773. switch (voi->colorspace) {
  774. case VIDEO_CS_601:
  775. av1_config->colorPrimaries = 6;
  776. av1_config->transferCharacteristics = 6;
  777. av1_config->matrixCoefficients = 6;
  778. break;
  779. case VIDEO_CS_DEFAULT:
  780. case VIDEO_CS_709:
  781. av1_config->colorPrimaries = 1;
  782. av1_config->transferCharacteristics = 1;
  783. av1_config->matrixCoefficients = 1;
  784. break;
  785. case VIDEO_CS_SRGB:
  786. av1_config->colorPrimaries = 1;
  787. av1_config->transferCharacteristics = 13;
  788. av1_config->matrixCoefficients = 1;
  789. break;
  790. case VIDEO_CS_2100_PQ:
  791. av1_config->colorPrimaries = 9;
  792. av1_config->transferCharacteristics = 16;
  793. av1_config->matrixCoefficients = 9;
  794. break;
  795. case VIDEO_CS_2100_HLG:
  796. av1_config->colorPrimaries = 9;
  797. av1_config->transferCharacteristics = 18;
  798. av1_config->matrixCoefficients = 9;
  799. }
  800. /* -------------------------- */
  801. /* profile */
  802. config->profileGUID = NV_ENC_AV1_PROFILE_MAIN_GUID;
  803. av1_config->tier = NV_ENC_TIER_AV1_0;
  804. av1_config->level = NV_ENC_LEVEL_AV1_AUTOSELECT;
  805. av1_config->chromaFormatIDC = 1;
  806. av1_config->pixelBitDepthMinus8 = obs_p010_tex_active() ? 2 : 0;
  807. av1_config->inputPixelBitDepthMinus8 = av1_config->pixelBitDepthMinus8;
  808. av1_config->numFwdRefs = 1;
  809. av1_config->numBwdRefs = 1;
  810. av1_config->repeatSeqHdr = 1;
  811. if (NV_FAILED(nv.nvEncInitializeEncoder(enc->session, &enc->params))) {
  812. return false;
  813. }
  814. return true;
  815. }
  816. static bool init_bitstreams(struct nvenc_data *enc)
  817. {
  818. da_reserve(enc->bitstreams, enc->buf_count);
  819. for (int i = 0; i < enc->buf_count; i++) {
  820. struct nv_bitstream bitstream;
  821. if (!nv_bitstream_init(enc, &bitstream)) {
  822. return false;
  823. }
  824. da_push_back(enc->bitstreams, &bitstream);
  825. }
  826. return true;
  827. }
  828. static bool init_textures(struct nvenc_data *enc)
  829. {
  830. da_reserve(enc->textures, enc->buf_count);
  831. for (int i = 0; i < enc->buf_count; i++) {
  832. struct nv_texture texture;
  833. if (!nv_texture_init(enc, &texture)) {
  834. return false;
  835. }
  836. da_push_back(enc->textures, &texture);
  837. }
  838. return true;
  839. }
  840. static void nvenc_destroy(void *data);
  841. static bool init_specific_encoder(struct nvenc_data *enc, obs_data_t *settings,
  842. int bf, bool compatibility)
  843. {
  844. switch (enc->codec) {
  845. case CODEC_HEVC:
  846. return init_encoder_hevc(enc, settings, bf, compatibility);
  847. case CODEC_H264:
  848. return init_encoder_h264(enc, settings, bf, compatibility);
  849. case CODEC_AV1:
  850. return init_encoder_av1(enc, settings, bf, compatibility);
  851. }
  852. return false;
  853. }
  854. static bool init_encoder(struct nvenc_data *enc, enum codec_type codec,
  855. obs_data_t *settings, obs_encoder_t *encoder)
  856. {
  857. int bf = (int)obs_data_get_int(settings, "bf");
  858. const bool support_10bit =
  859. nv_get_cap(enc, NV_ENC_CAPS_SUPPORT_10BIT_ENCODE);
  860. const int bf_max = nv_get_cap(enc, NV_ENC_CAPS_NUM_MAX_BFRAMES);
  861. if (obs_p010_tex_active() && !support_10bit) {
  862. NV_FAIL(obs_module_text("NVENC.10bitUnsupported"));
  863. return false;
  864. }
  865. video_t *video = obs_encoder_video(enc->encoder);
  866. const struct video_output_info *voi = video_output_get_info(video);
  867. switch (voi->format) {
  868. case VIDEO_FORMAT_I010:
  869. case VIDEO_FORMAT_P010:
  870. break;
  871. default:
  872. switch (voi->colorspace) {
  873. case VIDEO_CS_2100_PQ:
  874. case VIDEO_CS_2100_HLG:
  875. NV_FAIL(obs_module_text("NVENC.8bitUnsupportedHdr"));
  876. return false;
  877. }
  878. }
  879. if (bf > bf_max) {
  880. blog(LOG_WARNING,
  881. "[obs-nvenc] Max B-frames setting (%d) is more than encoder supports (%d).\n"
  882. "Setting B-frames to %d",
  883. bf, bf_max, bf_max);
  884. bf = bf_max;
  885. }
  886. if (!init_specific_encoder(enc, settings, bf, false)) {
  887. blog(LOG_WARNING, "[obs-nvenc] init_specific_encoder failed, "
  888. "trying again with compatibility options");
  889. nv.nvEncDestroyEncoder(enc->session);
  890. enc->session = NULL;
  891. if (!init_session(enc)) {
  892. return false;
  893. }
  894. /* try without multipass and psycho aq */
  895. if (!init_specific_encoder(enc, settings, bf, true)) {
  896. return false;
  897. }
  898. }
  899. return true;
  900. }
  901. static void *nvenc_create_internal(enum codec_type codec, obs_data_t *settings,
  902. obs_encoder_t *encoder)
  903. {
  904. struct nvenc_data *enc = bzalloc(sizeof(*enc));
  905. enc->encoder = encoder;
  906. enc->codec = codec;
  907. enc->first_packet = true;
  908. NV_ENCODE_API_FUNCTION_LIST init = {NV_ENCODE_API_FUNCTION_LIST_VER};
  909. switch (enc->codec) {
  910. case CODEC_H264:
  911. enc->codec_guid = NV_ENC_CODEC_H264_GUID;
  912. break;
  913. case CODEC_HEVC:
  914. enc->codec_guid = NV_ENC_CODEC_HEVC_GUID;
  915. break;
  916. case CODEC_AV1:
  917. enc->codec_guid = NV_ENC_CODEC_AV1_GUID;
  918. break;
  919. }
  920. if (!init_nvenc(encoder)) {
  921. goto fail;
  922. }
  923. if (NV_FAILED(nv_create_instance(&init))) {
  924. goto fail;
  925. }
  926. if (!init_d3d11(enc, settings)) {
  927. goto fail;
  928. }
  929. if (get_nvenc_ver() == COMPATIBILITY_VERSION) {
  930. enc->needs_compat_ver = true;
  931. }
  932. if (!init_session(enc)) {
  933. goto fail;
  934. }
  935. if (!init_encoder(enc, codec, settings, encoder)) {
  936. goto fail;
  937. }
  938. if (!init_bitstreams(enc)) {
  939. goto fail;
  940. }
  941. if (!init_textures(enc)) {
  942. goto fail;
  943. }
  944. #ifdef ENABLE_HEVC
  945. enc->codec = codec;
  946. #endif
  947. return enc;
  948. fail:
  949. nvenc_destroy(enc);
  950. return NULL;
  951. }
  952. static void *nvenc_create_base(enum codec_type codec, obs_data_t *settings,
  953. obs_encoder_t *encoder)
  954. {
  955. /* this encoder requires shared textures, this cannot be used on a
  956. * gpu other than the one OBS is currently running on. */
  957. const int gpu = (int)obs_data_get_int(settings, "gpu");
  958. if (gpu != 0) {
  959. blog(LOG_INFO,
  960. "[obs-nvenc] different GPU selected by user, falling back to ffmpeg");
  961. goto reroute;
  962. }
  963. if (obs_encoder_scaling_enabled(encoder)) {
  964. if (!obs_encoder_gpu_scaling_enabled(encoder)) {
  965. blog(LOG_INFO,
  966. "[obs-nvenc] CPU scaling enabled, falling back to ffmpeg");
  967. goto reroute;
  968. }
  969. blog(LOG_INFO, "[obs-nvenc] GPU scaling enabled");
  970. }
  971. if (!obs_p010_tex_active() && !obs_nv12_tex_active()) {
  972. blog(LOG_INFO,
  973. "[obs-nvenc] nv12/p010 not active, falling back to ffmpeg");
  974. goto reroute;
  975. }
  976. struct nvenc_data *enc =
  977. nvenc_create_internal(codec, settings, encoder);
  978. if (enc) {
  979. return enc;
  980. }
  981. reroute:
  982. switch (codec) {
  983. case CODEC_H264:
  984. return obs_encoder_create_rerouted(encoder, "ffmpeg_nvenc");
  985. case CODEC_HEVC:
  986. return obs_encoder_create_rerouted(encoder,
  987. "ffmpeg_hevc_nvenc");
  988. case CODEC_AV1:
  989. obs_encoder_set_last_error(
  990. encoder,
  991. obs_module_text("NVENC.NoAV1FallbackPossible"));
  992. break;
  993. }
  994. return NULL;
  995. }
  996. static void *h264_nvenc_create(obs_data_t *settings, obs_encoder_t *encoder)
  997. {
  998. return nvenc_create_base(CODEC_H264, settings, encoder);
  999. }
  1000. #ifdef ENABLE_HEVC
  1001. static void *hevc_nvenc_create(obs_data_t *settings, obs_encoder_t *encoder)
  1002. {
  1003. return nvenc_create_base(CODEC_HEVC, settings, encoder);
  1004. }
  1005. #endif
  1006. static void *av1_nvenc_create(obs_data_t *settings, obs_encoder_t *encoder)
  1007. {
  1008. return nvenc_create_base(CODEC_AV1, settings, encoder);
  1009. }
  1010. static bool get_encoded_packet(struct nvenc_data *enc, bool finalize);
  1011. static void nvenc_destroy(void *data)
  1012. {
  1013. struct nvenc_data *enc = data;
  1014. if (enc->encode_started) {
  1015. uint32_t struct_ver = enc->needs_compat_ver
  1016. ? NV_ENC_PIC_PARAMS_COMPAT_VER
  1017. : NV_ENC_PIC_PARAMS_VER;
  1018. NV_ENC_PIC_PARAMS params = {struct_ver};
  1019. params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
  1020. nv.nvEncEncodePicture(enc->session, &params);
  1021. get_encoded_packet(enc, true);
  1022. }
  1023. for (size_t i = 0; i < enc->textures.num; i++) {
  1024. nv_texture_free(enc, &enc->textures.array[i]);
  1025. }
  1026. for (size_t i = 0; i < enc->bitstreams.num; i++) {
  1027. nv_bitstream_free(enc, &enc->bitstreams.array[i]);
  1028. }
  1029. if (enc->session) {
  1030. nv.nvEncDestroyEncoder(enc->session);
  1031. }
  1032. for (size_t i = 0; i < enc->input_textures.num; i++) {
  1033. ID3D11Texture2D *tex = enc->input_textures.array[i].tex;
  1034. IDXGIKeyedMutex *km = enc->input_textures.array[i].km;
  1035. tex->lpVtbl->Release(tex);
  1036. km->lpVtbl->Release(km);
  1037. }
  1038. if (enc->context) {
  1039. enc->context->lpVtbl->Release(enc->context);
  1040. }
  1041. if (enc->device) {
  1042. enc->device->lpVtbl->Release(enc->device);
  1043. }
  1044. bfree(enc->header);
  1045. bfree(enc->sei);
  1046. deque_free(&enc->dts_list);
  1047. da_free(enc->textures);
  1048. da_free(enc->bitstreams);
  1049. da_free(enc->input_textures);
  1050. da_free(enc->packet_data);
  1051. bfree(enc->roi_map);
  1052. bfree(enc);
  1053. }
  1054. static ID3D11Texture2D *get_tex_from_handle(struct nvenc_data *enc,
  1055. uint32_t handle,
  1056. IDXGIKeyedMutex **km_out)
  1057. {
  1058. ID3D11Device *device = enc->device;
  1059. IDXGIKeyedMutex *km;
  1060. ID3D11Texture2D *input_tex;
  1061. HRESULT hr;
  1062. for (size_t i = 0; i < enc->input_textures.num; i++) {
  1063. struct handle_tex *ht = &enc->input_textures.array[i];
  1064. if (ht->handle == handle) {
  1065. *km_out = ht->km;
  1066. return ht->tex;
  1067. }
  1068. }
  1069. hr = device->lpVtbl->OpenSharedResource(device,
  1070. (HANDLE)(uintptr_t)handle,
  1071. &IID_ID3D11Texture2D,
  1072. &input_tex);
  1073. if (FAILED(hr)) {
  1074. error_hr("OpenSharedResource failed");
  1075. return NULL;
  1076. }
  1077. hr = input_tex->lpVtbl->QueryInterface(input_tex, &IID_IDXGIKeyedMutex,
  1078. &km);
  1079. if (FAILED(hr)) {
  1080. error_hr("QueryInterface(IDXGIKeyedMutex) failed");
  1081. input_tex->lpVtbl->Release(input_tex);
  1082. return NULL;
  1083. }
  1084. input_tex->lpVtbl->SetEvictionPriority(input_tex,
  1085. DXGI_RESOURCE_PRIORITY_MAXIMUM);
  1086. *km_out = km;
  1087. struct handle_tex new_ht = {handle, input_tex, km};
  1088. da_push_back(enc->input_textures, &new_ht);
  1089. return input_tex;
  1090. }
  1091. static bool get_encoded_packet(struct nvenc_data *enc, bool finalize)
  1092. {
  1093. void *s = enc->session;
  1094. da_resize(enc->packet_data, 0);
  1095. if (!enc->buffers_queued)
  1096. return true;
  1097. if (!finalize && enc->buffers_queued < enc->output_delay)
  1098. return true;
  1099. size_t count = finalize ? enc->buffers_queued : 1;
  1100. for (size_t i = 0; i < count; i++) {
  1101. size_t cur_bs_idx = enc->cur_bitstream;
  1102. struct nv_bitstream *bs = &enc->bitstreams.array[cur_bs_idx];
  1103. struct nv_texture *nvtex = &enc->textures.array[cur_bs_idx];
  1104. /* ---------------- */
  1105. uint32_t struct_ver = enc->needs_compat_ver
  1106. ? NV_ENC_LOCK_BITSTREAM_COMPAT_VER
  1107. : NV_ENC_LOCK_BITSTREAM_VER;
  1108. NV_ENC_LOCK_BITSTREAM lock = {struct_ver};
  1109. lock.outputBitstream = bs->ptr;
  1110. lock.doNotWait = false;
  1111. if (NV_FAILED(nv.nvEncLockBitstream(s, &lock))) {
  1112. return false;
  1113. }
  1114. if (enc->first_packet) {
  1115. NV_ENC_SEQUENCE_PARAM_PAYLOAD payload = {0};
  1116. uint8_t buf[256];
  1117. uint32_t size = 0;
  1118. payload.version = NV_ENC_SEQUENCE_PARAM_PAYLOAD_VER;
  1119. payload.spsppsBuffer = buf;
  1120. payload.inBufferSize = sizeof(buf);
  1121. payload.outSPSPPSPayloadSize = &size;
  1122. nv.nvEncGetSequenceParams(s, &payload);
  1123. enc->header = bmemdup(buf, size);
  1124. enc->header_size = size;
  1125. enc->first_packet = false;
  1126. }
  1127. da_copy_array(enc->packet_data, lock.bitstreamBufferPtr,
  1128. lock.bitstreamSizeInBytes);
  1129. enc->packet_pts = (int64_t)lock.outputTimeStamp;
  1130. enc->packet_keyframe = lock.pictureType == NV_ENC_PIC_TYPE_IDR;
  1131. if (NV_FAILED(nv.nvEncUnlockBitstream(s, bs->ptr))) {
  1132. return false;
  1133. }
  1134. /* ---------------- */
  1135. if (nvtex->mapped_res) {
  1136. NVENCSTATUS err;
  1137. err = nv.nvEncUnmapInputResource(s, nvtex->mapped_res);
  1138. if (nv_failed(enc->encoder, err, __FUNCTION__,
  1139. "unmap")) {
  1140. return false;
  1141. }
  1142. nvtex->mapped_res = NULL;
  1143. }
  1144. /* ---------------- */
  1145. if (++enc->cur_bitstream == enc->buf_count)
  1146. enc->cur_bitstream = 0;
  1147. enc->buffers_queued--;
  1148. }
  1149. return true;
  1150. }
  1151. struct roi_params {
  1152. uint32_t mb_width;
  1153. uint32_t mb_height;
  1154. uint32_t mb_size;
  1155. bool av1;
  1156. int8_t *map;
  1157. };
  1158. static void roi_cb(void *param, struct obs_encoder_roi *roi)
  1159. {
  1160. const struct roi_params *rp = param;
  1161. int8_t qp_val;
  1162. /* AV1 has a larger QP range than HEVC/H.264 */
  1163. if (rp->av1) {
  1164. qp_val = (int8_t)(-128.0f * roi->priority);
  1165. } else {
  1166. qp_val = (int8_t)(-51.0f * roi->priority);
  1167. }
  1168. const uint32_t roi_left = roi->left / rp->mb_size;
  1169. const uint32_t roi_top = roi->top / rp->mb_size;
  1170. const uint32_t roi_right = (roi->right - 1) / rp->mb_size;
  1171. const uint32_t roi_bottom = (roi->bottom - 1) / rp->mb_size;
  1172. for (uint32_t mb_y = 0; mb_y < rp->mb_height; mb_y++) {
  1173. if (mb_y < roi_top || mb_y > roi_bottom)
  1174. continue;
  1175. for (uint32_t mb_x = 0; mb_x < rp->mb_width; mb_x++) {
  1176. if (mb_x < roi_left || mb_x > roi_right)
  1177. continue;
  1178. rp->map[mb_y * rp->mb_width + mb_x] = qp_val;
  1179. }
  1180. }
  1181. }
  1182. static void add_roi(struct nvenc_data *enc, NV_ENC_PIC_PARAMS *params)
  1183. {
  1184. const uint32_t increment = obs_encoder_get_roi_increment(enc->encoder);
  1185. if (enc->roi_map && enc->roi_increment == increment) {
  1186. params->qpDeltaMap = enc->roi_map;
  1187. params->qpDeltaMapSize = (uint32_t)enc->roi_map_size;
  1188. return;
  1189. }
  1190. uint32_t mb_size;
  1191. switch (enc->codec) {
  1192. case CODEC_H264:
  1193. /* H.264 is always 16x16 */
  1194. mb_size = 16;
  1195. break;
  1196. case CODEC_HEVC:
  1197. /* HEVC can be 16x16, 32x32, or 64x64, but NVENC is always 32x32 */
  1198. mb_size = 32;
  1199. break;
  1200. case CODEC_AV1:
  1201. /* AV1 can be 64x64 or 128x128, but NVENC is always 64x64 */
  1202. mb_size = 64;
  1203. break;
  1204. }
  1205. const uint32_t mb_width = (enc->cx + mb_size - 1) / mb_size;
  1206. const uint32_t mb_height = (enc->cy + mb_size - 1) / mb_size;
  1207. const size_t map_size = mb_width * mb_height * sizeof(int8_t);
  1208. if (map_size != enc->roi_map_size) {
  1209. enc->roi_map = brealloc(enc->roi_map, map_size);
  1210. enc->roi_map_size = map_size;
  1211. }
  1212. memset(enc->roi_map, 0, enc->roi_map_size);
  1213. struct roi_params par = {
  1214. .mb_width = mb_width,
  1215. .mb_height = mb_height,
  1216. .mb_size = mb_size,
  1217. .av1 = enc->codec == CODEC_AV1,
  1218. .map = enc->roi_map,
  1219. };
  1220. obs_encoder_enum_roi(enc->encoder, roi_cb, &par);
  1221. enc->roi_increment = increment;
  1222. params->qpDeltaMap = enc->roi_map;
  1223. params->qpDeltaMapSize = (uint32_t)map_size;
  1224. }
  1225. static bool nvenc_encode_tex(void *data, uint32_t handle, int64_t pts,
  1226. uint64_t lock_key, uint64_t *next_key,
  1227. struct encoder_packet *packet,
  1228. bool *received_packet)
  1229. {
  1230. struct nvenc_data *enc = data;
  1231. ID3D11Device *device = enc->device;
  1232. ID3D11DeviceContext *context = enc->context;
  1233. ID3D11Texture2D *input_tex;
  1234. ID3D11Texture2D *output_tex;
  1235. IDXGIKeyedMutex *km;
  1236. struct nv_texture *nvtex;
  1237. struct nv_bitstream *bs;
  1238. NVENCSTATUS err;
  1239. if (handle == GS_INVALID_HANDLE) {
  1240. error("Encode failed: bad texture handle");
  1241. *next_key = lock_key;
  1242. return false;
  1243. }
  1244. bs = &enc->bitstreams.array[enc->next_bitstream];
  1245. nvtex = &enc->textures.array[enc->next_bitstream];
  1246. input_tex = get_tex_from_handle(enc, handle, &km);
  1247. output_tex = nvtex->tex;
  1248. if (!input_tex) {
  1249. *next_key = lock_key;
  1250. return false;
  1251. }
  1252. deque_push_back(&enc->dts_list, &pts, sizeof(pts));
  1253. /* ------------------------------------ */
  1254. /* copy to output tex */
  1255. km->lpVtbl->AcquireSync(km, lock_key, INFINITE);
  1256. context->lpVtbl->CopyResource(context, (ID3D11Resource *)output_tex,
  1257. (ID3D11Resource *)input_tex);
  1258. km->lpVtbl->ReleaseSync(km, *next_key);
  1259. /* ------------------------------------ */
  1260. /* map output tex so nvenc can use it */
  1261. NV_ENC_MAP_INPUT_RESOURCE map = {NV_ENC_MAP_INPUT_RESOURCE_VER};
  1262. map.registeredResource = nvtex->res;
  1263. if (NV_FAILED(nv.nvEncMapInputResource(enc->session, &map))) {
  1264. return false;
  1265. }
  1266. nvtex->mapped_res = map.mappedResource;
  1267. /* ------------------------------------ */
  1268. /* do actual encode call */
  1269. NV_ENC_PIC_PARAMS params = {0};
  1270. params.version = enc->needs_compat_ver ? NV_ENC_PIC_PARAMS_COMPAT_VER
  1271. : NV_ENC_PIC_PARAMS_VER;
  1272. params.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
  1273. params.inputBuffer = nvtex->mapped_res;
  1274. params.bufferFmt = obs_p010_tex_active()
  1275. ? NV_ENC_BUFFER_FORMAT_YUV420_10BIT
  1276. : NV_ENC_BUFFER_FORMAT_NV12;
  1277. params.inputTimeStamp = (uint64_t)pts;
  1278. params.inputWidth = enc->cx;
  1279. params.inputHeight = enc->cy;
  1280. params.inputPitch = enc->cx;
  1281. params.outputBitstream = bs->ptr;
  1282. /* Add ROI map if enabled */
  1283. if (obs_encoder_has_roi(enc->encoder))
  1284. add_roi(enc, &params);
  1285. err = nv.nvEncEncodePicture(enc->session, &params);
  1286. if (err != NV_ENC_SUCCESS && err != NV_ENC_ERR_NEED_MORE_INPUT) {
  1287. nv_failed(enc->encoder, err, __FUNCTION__,
  1288. "nvEncEncodePicture");
  1289. return false;
  1290. }
  1291. enc->encode_started = true;
  1292. enc->buffers_queued++;
  1293. if (++enc->next_bitstream == enc->buf_count) {
  1294. enc->next_bitstream = 0;
  1295. }
  1296. /* ------------------------------------ */
  1297. /* check for encoded packet and parse */
  1298. if (!get_encoded_packet(enc, false)) {
  1299. return false;
  1300. }
  1301. /* ------------------------------------ */
  1302. /* output encoded packet */
  1303. if (enc->packet_data.num) {
  1304. int64_t dts;
  1305. deque_pop_front(&enc->dts_list, &dts, sizeof(dts));
  1306. /* subtract bframe delay from dts */
  1307. dts -= (int64_t)enc->bframes * packet->timebase_num;
  1308. *received_packet = true;
  1309. packet->data = enc->packet_data.array;
  1310. packet->size = enc->packet_data.num;
  1311. packet->type = OBS_ENCODER_VIDEO;
  1312. packet->pts = enc->packet_pts;
  1313. packet->dts = dts;
  1314. packet->keyframe = enc->packet_keyframe;
  1315. } else {
  1316. *received_packet = false;
  1317. }
  1318. return true;
  1319. }
  1320. extern void h264_nvenc_defaults(obs_data_t *settings);
  1321. extern obs_properties_t *h264_nvenc_properties(void *unused);
  1322. #ifdef ENABLE_HEVC
  1323. extern void hevc_nvenc_defaults(obs_data_t *settings);
  1324. extern obs_properties_t *hevc_nvenc_properties(void *unused);
  1325. #endif
  1326. extern obs_properties_t *av1_nvenc_properties(void *unused);
  1327. extern void av1_nvenc_defaults(obs_data_t *settings);
  1328. static bool nvenc_extra_data(void *data, uint8_t **header, size_t *size)
  1329. {
  1330. struct nvenc_data *enc = data;
  1331. if (!enc->header) {
  1332. return false;
  1333. }
  1334. *header = enc->header;
  1335. *size = enc->header_size;
  1336. return true;
  1337. }
  1338. static bool nvenc_sei_data(void *data, uint8_t **sei, size_t *size)
  1339. {
  1340. struct nvenc_data *enc = data;
  1341. if (!enc->sei) {
  1342. return false;
  1343. }
  1344. *sei = enc->sei;
  1345. *size = enc->sei_size;
  1346. return true;
  1347. }
  1348. struct obs_encoder_info h264_nvenc_info = {
  1349. .id = "jim_nvenc",
  1350. .codec = "h264",
  1351. .type = OBS_ENCODER_VIDEO,
  1352. .caps = OBS_ENCODER_CAP_PASS_TEXTURE | OBS_ENCODER_CAP_DYN_BITRATE |
  1353. OBS_ENCODER_CAP_ROI,
  1354. .get_name = h264_nvenc_get_name,
  1355. .create = h264_nvenc_create,
  1356. .destroy = nvenc_destroy,
  1357. .update = nvenc_update,
  1358. .encode_texture = nvenc_encode_tex,
  1359. .get_defaults = h264_nvenc_defaults,
  1360. .get_properties = h264_nvenc_properties,
  1361. .get_extra_data = nvenc_extra_data,
  1362. .get_sei_data = nvenc_sei_data,
  1363. };
  1364. #ifdef ENABLE_HEVC
  1365. struct obs_encoder_info hevc_nvenc_info = {
  1366. .id = "jim_hevc_nvenc",
  1367. .codec = "hevc",
  1368. .type = OBS_ENCODER_VIDEO,
  1369. .caps = OBS_ENCODER_CAP_PASS_TEXTURE | OBS_ENCODER_CAP_DYN_BITRATE |
  1370. OBS_ENCODER_CAP_ROI,
  1371. .get_name = hevc_nvenc_get_name,
  1372. .create = hevc_nvenc_create,
  1373. .destroy = nvenc_destroy,
  1374. .update = nvenc_update,
  1375. .encode_texture = nvenc_encode_tex,
  1376. .get_defaults = hevc_nvenc_defaults,
  1377. .get_properties = hevc_nvenc_properties,
  1378. .get_extra_data = nvenc_extra_data,
  1379. .get_sei_data = nvenc_sei_data,
  1380. };
  1381. #endif
  1382. struct obs_encoder_info av1_nvenc_info = {
  1383. .id = "jim_av1_nvenc",
  1384. .codec = "av1",
  1385. .type = OBS_ENCODER_VIDEO,
  1386. .caps = OBS_ENCODER_CAP_PASS_TEXTURE | OBS_ENCODER_CAP_DYN_BITRATE |
  1387. OBS_ENCODER_CAP_ROI,
  1388. .get_name = av1_nvenc_get_name,
  1389. .create = av1_nvenc_create,
  1390. .destroy = nvenc_destroy,
  1391. .update = nvenc_update,
  1392. .encode_texture = nvenc_encode_tex,
  1393. .get_defaults = av1_nvenc_defaults,
  1394. .get_properties = av1_nvenc_properties,
  1395. .get_extra_data = nvenc_extra_data,
  1396. };