jim-nvenc.c 39 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479
  1. #include "jim-nvenc.h"
  2. #include <util/circlebuf.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. #ifdef ENABLE_HEVC
  12. #include <obs-hevc.h>
  13. #endif
  14. /* TODO: Use new preset scheme */
  15. #pragma warning(disable : 4996)
  16. /* ========================================================================= */
  17. #define EXTRA_BUFFERS 5
  18. #define do_log(level, format, ...) \
  19. blog(level, "[jim-nvenc: '%s'] " format, \
  20. obs_encoder_get_name(enc->encoder), ##__VA_ARGS__)
  21. #define error(format, ...) do_log(LOG_ERROR, format, ##__VA_ARGS__)
  22. #define warn(format, ...) do_log(LOG_WARNING, format, ##__VA_ARGS__)
  23. #define info(format, ...) do_log(LOG_INFO, format, ##__VA_ARGS__)
  24. #define debug(format, ...) do_log(LOG_DEBUG, format, ##__VA_ARGS__)
  25. #define error_hr(msg) error("%s: %s: 0x%08lX", __FUNCTION__, msg, (uint32_t)hr);
  26. struct nv_bitstream;
  27. struct nv_texture;
  28. struct handle_tex {
  29. uint32_t handle;
  30. ID3D11Texture2D *tex;
  31. IDXGIKeyedMutex *km;
  32. };
  33. /* ------------------------------------------------------------------------- */
  34. /* Main Implementation Structure */
  35. struct nvenc_data {
  36. obs_encoder_t *encoder;
  37. #ifdef ENABLE_HEVC
  38. bool hevc;
  39. #endif
  40. void *session;
  41. NV_ENC_INITIALIZE_PARAMS params;
  42. NV_ENC_CONFIG config;
  43. int rc_lookahead;
  44. int buf_count;
  45. int output_delay;
  46. int buffers_queued;
  47. size_t next_bitstream;
  48. size_t cur_bitstream;
  49. bool encode_started;
  50. bool first_packet;
  51. bool can_change_bitrate;
  52. int32_t bframes;
  53. DARRAY(struct nv_bitstream) bitstreams;
  54. DARRAY(struct nv_texture) textures;
  55. DARRAY(struct handle_tex) input_textures;
  56. struct circlebuf dts_list;
  57. DARRAY(uint8_t) packet_data;
  58. int64_t packet_pts;
  59. bool packet_keyframe;
  60. ID3D11Device *device;
  61. ID3D11DeviceContext *context;
  62. uint32_t cx;
  63. uint32_t cy;
  64. uint8_t *header;
  65. size_t header_size;
  66. uint8_t *sei;
  67. size_t sei_size;
  68. };
  69. /* ------------------------------------------------------------------------- */
  70. /* Bitstream Buffer */
  71. struct nv_bitstream {
  72. void *ptr;
  73. HANDLE event;
  74. };
  75. #define NV_FAIL(format, ...) nv_fail(enc->encoder, format, __VA_ARGS__)
  76. #define NV_FAILED(x) nv_failed(enc->encoder, x, __FUNCTION__, #x)
  77. static bool nv_bitstream_init(struct nvenc_data *enc, struct nv_bitstream *bs)
  78. {
  79. NV_ENC_CREATE_BITSTREAM_BUFFER buf = {
  80. NV_ENC_CREATE_BITSTREAM_BUFFER_VER};
  81. NV_ENC_EVENT_PARAMS params = {NV_ENC_EVENT_PARAMS_VER};
  82. HANDLE event = NULL;
  83. if (NV_FAILED(nv.nvEncCreateBitstreamBuffer(enc->session, &buf))) {
  84. return false;
  85. }
  86. event = CreateEvent(NULL, true, true, NULL);
  87. if (!event) {
  88. error("%s: %s", __FUNCTION__, "Failed to create event");
  89. goto fail;
  90. }
  91. params.completionEvent = event;
  92. if (NV_FAILED(nv.nvEncRegisterAsyncEvent(enc->session, &params))) {
  93. goto fail;
  94. }
  95. bs->ptr = buf.bitstreamBuffer;
  96. bs->event = event;
  97. return true;
  98. fail:
  99. if (event) {
  100. CloseHandle(event);
  101. }
  102. if (buf.bitstreamBuffer) {
  103. nv.nvEncDestroyBitstreamBuffer(enc->session,
  104. buf.bitstreamBuffer);
  105. }
  106. return false;
  107. }
  108. static void nv_bitstream_free(struct nvenc_data *enc, struct nv_bitstream *bs)
  109. {
  110. if (bs->ptr) {
  111. nv.nvEncDestroyBitstreamBuffer(enc->session, bs->ptr);
  112. NV_ENC_EVENT_PARAMS params = {NV_ENC_EVENT_PARAMS_VER};
  113. params.completionEvent = bs->event;
  114. nv.nvEncUnregisterAsyncEvent(enc->session, &params);
  115. CloseHandle(bs->event);
  116. }
  117. }
  118. /* ------------------------------------------------------------------------- */
  119. /* Texture Resource */
  120. struct nv_texture {
  121. void *res;
  122. ID3D11Texture2D *tex;
  123. void *mapped_res;
  124. };
  125. static bool nv_texture_init(struct nvenc_data *enc, struct nv_texture *nvtex)
  126. {
  127. const bool p010 = obs_p010_tex_active();
  128. D3D11_TEXTURE2D_DESC desc = {0};
  129. desc.Width = enc->cx;
  130. desc.Height = enc->cy;
  131. desc.MipLevels = 1;
  132. desc.ArraySize = 1;
  133. desc.Format = p010 ? DXGI_FORMAT_P010 : DXGI_FORMAT_NV12;
  134. desc.SampleDesc.Count = 1;
  135. desc.BindFlags = D3D11_BIND_RENDER_TARGET;
  136. ID3D11Device *const device = enc->device;
  137. ID3D11Texture2D *tex;
  138. HRESULT hr = device->lpVtbl->CreateTexture2D(device, &desc, NULL, &tex);
  139. if (FAILED(hr)) {
  140. error_hr("Failed to create texture");
  141. return false;
  142. }
  143. tex->lpVtbl->SetEvictionPriority(tex, DXGI_RESOURCE_PRIORITY_MAXIMUM);
  144. NV_ENC_REGISTER_RESOURCE res = {NV_ENC_REGISTER_RESOURCE_VER};
  145. res.resourceType = NV_ENC_INPUT_RESOURCE_TYPE_DIRECTX;
  146. res.resourceToRegister = tex;
  147. res.width = enc->cx;
  148. res.height = enc->cy;
  149. res.bufferFormat = p010 ? NV_ENC_BUFFER_FORMAT_YUV420_10BIT
  150. : NV_ENC_BUFFER_FORMAT_NV12;
  151. if (NV_FAILED(nv.nvEncRegisterResource(enc->session, &res))) {
  152. tex->lpVtbl->Release(tex);
  153. return false;
  154. }
  155. nvtex->res = res.registeredResource;
  156. nvtex->tex = tex;
  157. nvtex->mapped_res = NULL;
  158. return true;
  159. }
  160. static void nv_texture_free(struct nvenc_data *enc, struct nv_texture *nvtex)
  161. {
  162. if (nvtex->res) {
  163. if (nvtex->mapped_res) {
  164. nv.nvEncUnmapInputResource(enc->session,
  165. nvtex->mapped_res);
  166. }
  167. nv.nvEncUnregisterResource(enc->session, nvtex->res);
  168. nvtex->tex->lpVtbl->Release(nvtex->tex);
  169. }
  170. }
  171. /* ------------------------------------------------------------------------- */
  172. /* Implementation */
  173. static const char *h264_nvenc_get_name(void *type_data)
  174. {
  175. UNUSED_PARAMETER(type_data);
  176. return "NVIDIA NVENC H.264 (new)";
  177. }
  178. #ifdef ENABLE_HEVC
  179. static const char *hevc_nvenc_get_name(void *type_data)
  180. {
  181. UNUSED_PARAMETER(type_data);
  182. return "NVIDIA NVENC HEVC (new)";
  183. }
  184. #endif
  185. static inline int nv_get_cap_internal(struct nvenc_data *enc, NV_ENC_CAPS cap,
  186. GUID encodeGUID)
  187. {
  188. if (!enc->session)
  189. return 0;
  190. NV_ENC_CAPS_PARAM param = {NV_ENC_CAPS_PARAM_VER};
  191. int v;
  192. param.capsToQuery = cap;
  193. nv.nvEncGetEncodeCaps(enc->session, encodeGUID, &param, &v);
  194. return v;
  195. }
  196. static inline int nv_get_cap_h264(struct nvenc_data *enc, NV_ENC_CAPS cap)
  197. {
  198. return nv_get_cap_internal(enc, cap, NV_ENC_CODEC_H264_GUID);
  199. }
  200. #ifdef ENABLE_HEVC
  201. static inline int nv_get_cap_hevc(struct nvenc_data *enc, NV_ENC_CAPS cap)
  202. {
  203. return nv_get_cap_internal(enc, cap, NV_ENC_CODEC_HEVC_GUID);
  204. }
  205. #endif
  206. static bool nvenc_update(void *data, obs_data_t *settings)
  207. {
  208. struct nvenc_data *enc = data;
  209. /* Only support reconfiguration of CBR bitrate */
  210. if (enc->can_change_bitrate) {
  211. int bitrate = (int)obs_data_get_int(settings, "bitrate");
  212. enc->config.rcParams.averageBitRate = bitrate * 1000;
  213. enc->config.rcParams.maxBitRate = bitrate * 1000;
  214. NV_ENC_RECONFIGURE_PARAMS params = {0};
  215. params.version = NV_ENC_RECONFIGURE_PARAMS_VER;
  216. params.reInitEncodeParams = enc->params;
  217. params.resetEncoder = 1;
  218. params.forceIDR = 1;
  219. if (NV_FAILED(nv.nvEncReconfigureEncoder(enc->session,
  220. &params))) {
  221. return false;
  222. }
  223. }
  224. return true;
  225. }
  226. static HANDLE get_lib(struct nvenc_data *enc, const char *lib)
  227. {
  228. HMODULE mod = GetModuleHandleA(lib);
  229. if (mod)
  230. return mod;
  231. mod = LoadLibraryA(lib);
  232. if (!mod)
  233. error("Failed to load %s", lib);
  234. return mod;
  235. }
  236. typedef HRESULT(WINAPI *CREATEDXGIFACTORY1PROC)(REFIID, void **);
  237. static bool init_d3d11(struct nvenc_data *enc, obs_data_t *settings)
  238. {
  239. HMODULE dxgi = get_lib(enc, "DXGI.dll");
  240. HMODULE d3d11 = get_lib(enc, "D3D11.dll");
  241. CREATEDXGIFACTORY1PROC create_dxgi;
  242. PFN_D3D11_CREATE_DEVICE create_device;
  243. IDXGIFactory1 *factory;
  244. IDXGIAdapter *adapter;
  245. ID3D11Device *device;
  246. ID3D11DeviceContext *context;
  247. HRESULT hr;
  248. if (!dxgi || !d3d11) {
  249. return false;
  250. }
  251. create_dxgi = (CREATEDXGIFACTORY1PROC)GetProcAddress(
  252. dxgi, "CreateDXGIFactory1");
  253. create_device = (PFN_D3D11_CREATE_DEVICE)GetProcAddress(
  254. d3d11, "D3D11CreateDevice");
  255. if (!create_dxgi || !create_device) {
  256. error("Failed to load D3D11/DXGI procedures");
  257. return false;
  258. }
  259. hr = create_dxgi(&IID_IDXGIFactory1, &factory);
  260. if (FAILED(hr)) {
  261. error_hr("CreateDXGIFactory1 failed");
  262. return false;
  263. }
  264. hr = factory->lpVtbl->EnumAdapters(factory, 0, &adapter);
  265. factory->lpVtbl->Release(factory);
  266. if (FAILED(hr)) {
  267. error_hr("EnumAdapters failed");
  268. return false;
  269. }
  270. hr = create_device(adapter, D3D_DRIVER_TYPE_UNKNOWN, NULL, 0, NULL, 0,
  271. D3D11_SDK_VERSION, &device, NULL, &context);
  272. adapter->lpVtbl->Release(adapter);
  273. if (FAILED(hr)) {
  274. error_hr("D3D11CreateDevice failed");
  275. return false;
  276. }
  277. enc->device = device;
  278. enc->context = context;
  279. return true;
  280. }
  281. static bool init_session(struct nvenc_data *enc)
  282. {
  283. NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS params = {
  284. NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER};
  285. params.device = enc->device;
  286. params.deviceType = NV_ENC_DEVICE_TYPE_DIRECTX;
  287. params.apiVersion = NVENCAPI_VERSION;
  288. if (NV_FAILED(nv.nvEncOpenEncodeSessionEx(&params, &enc->session))) {
  289. return false;
  290. }
  291. return true;
  292. }
  293. static bool init_encoder_h264(struct nvenc_data *enc, obs_data_t *settings,
  294. int bf, bool psycho_aq)
  295. {
  296. const char *rc = obs_data_get_string(settings, "rate_control");
  297. int bitrate = (int)obs_data_get_int(settings, "bitrate");
  298. int max_bitrate = (int)obs_data_get_int(settings, "max_bitrate");
  299. int cqp = (int)obs_data_get_int(settings, "cqp");
  300. int keyint_sec = (int)obs_data_get_int(settings, "keyint_sec");
  301. const char *preset = obs_data_get_string(settings, "preset");
  302. const char *profile = obs_data_get_string(settings, "profile");
  303. bool lookahead = obs_data_get_bool(settings, "lookahead");
  304. bool vbr = astrcmpi(rc, "VBR") == 0;
  305. NVENCSTATUS err;
  306. video_t *video = obs_encoder_video(enc->encoder);
  307. const struct video_output_info *voi = video_output_get_info(video);
  308. enc->cx = voi->width;
  309. enc->cy = voi->height;
  310. /* -------------------------- */
  311. /* get preset */
  312. GUID nv_preset = NV_ENC_PRESET_DEFAULT_GUID;
  313. bool twopass = false;
  314. bool hp = false;
  315. bool ll = false;
  316. if (astrcmpi(preset, "hq") == 0) {
  317. nv_preset = NV_ENC_PRESET_HQ_GUID;
  318. } else if (astrcmpi(preset, "mq") == 0) {
  319. nv_preset = NV_ENC_PRESET_HQ_GUID;
  320. twopass = true;
  321. } else if (astrcmpi(preset, "hp") == 0) {
  322. nv_preset = NV_ENC_PRESET_HP_GUID;
  323. hp = true;
  324. } else if (astrcmpi(preset, "ll") == 0) {
  325. nv_preset = NV_ENC_PRESET_LOW_LATENCY_DEFAULT_GUID;
  326. ll = true;
  327. } else if (astrcmpi(preset, "llhq") == 0) {
  328. nv_preset = NV_ENC_PRESET_LOW_LATENCY_HQ_GUID;
  329. ll = true;
  330. } else if (astrcmpi(preset, "llhp") == 0) {
  331. nv_preset = NV_ENC_PRESET_LOW_LATENCY_HP_GUID;
  332. hp = true;
  333. ll = true;
  334. }
  335. const bool rc_lossless = astrcmpi(rc, "lossless") == 0;
  336. bool lossless = rc_lossless;
  337. if (rc_lossless) {
  338. lossless = nv_get_cap_h264(enc,
  339. NV_ENC_CAPS_SUPPORT_LOSSLESS_ENCODE);
  340. if (lossless) {
  341. nv_preset = hp ? NV_ENC_PRESET_LOSSLESS_HP_GUID
  342. : NV_ENC_PRESET_LOSSLESS_DEFAULT_GUID;
  343. } else {
  344. warn("lossless encode is not supported, ignoring");
  345. }
  346. }
  347. /* -------------------------- */
  348. /* get preset default config */
  349. NV_ENC_PRESET_CONFIG preset_config = {NV_ENC_PRESET_CONFIG_VER,
  350. {NV_ENC_CONFIG_VER}};
  351. err = nv.nvEncGetEncodePresetConfig(enc->session,
  352. NV_ENC_CODEC_H264_GUID, nv_preset,
  353. &preset_config);
  354. if (nv_failed(enc->encoder, err, __FUNCTION__,
  355. "nvEncGetEncodePresetConfig")) {
  356. return false;
  357. }
  358. /* -------------------------- */
  359. /* main configuration */
  360. enc->config = preset_config.presetCfg;
  361. uint32_t gop_size =
  362. (keyint_sec) ? keyint_sec * voi->fps_num / voi->fps_den : 250;
  363. NV_ENC_INITIALIZE_PARAMS *params = &enc->params;
  364. NV_ENC_CONFIG *config = &enc->config;
  365. NV_ENC_CONFIG_H264 *h264_config = &config->encodeCodecConfig.h264Config;
  366. NV_ENC_CONFIG_H264_VUI_PARAMETERS *vui_params =
  367. &h264_config->h264VUIParameters;
  368. int darWidth, darHeight;
  369. av_reduce(&darWidth, &darHeight, voi->width, voi->height, 1024 * 1024);
  370. memset(params, 0, sizeof(*params));
  371. params->version = NV_ENC_INITIALIZE_PARAMS_VER;
  372. params->encodeGUID = NV_ENC_CODEC_H264_GUID;
  373. params->presetGUID = nv_preset;
  374. params->encodeWidth = voi->width;
  375. params->encodeHeight = voi->height;
  376. params->darWidth = darWidth;
  377. params->darHeight = darHeight;
  378. params->frameRateNum = voi->fps_num;
  379. params->frameRateDen = voi->fps_den;
  380. params->enableEncodeAsync = 1;
  381. params->enablePTD = 1;
  382. params->encodeConfig = &enc->config;
  383. config->gopLength = gop_size;
  384. config->frameIntervalP = 1 + bf;
  385. h264_config->idrPeriod = gop_size;
  386. bool repeat_headers = obs_data_get_bool(settings, "repeat_headers");
  387. if (repeat_headers) {
  388. h264_config->repeatSPSPPS = 1;
  389. h264_config->disableSPSPPS = 0;
  390. h264_config->outputAUD = 1;
  391. }
  392. h264_config->sliceMode = 3;
  393. h264_config->sliceModeData = 1;
  394. h264_config->useBFramesAsRef = NV_ENC_BFRAME_REF_MODE_DISABLED;
  395. vui_params->videoSignalTypePresentFlag = 1;
  396. vui_params->videoFullRangeFlag = (voi->range == VIDEO_RANGE_FULL);
  397. vui_params->colourDescriptionPresentFlag = 1;
  398. switch (voi->colorspace) {
  399. case VIDEO_CS_601:
  400. vui_params->colourPrimaries = 6;
  401. vui_params->transferCharacteristics = 6;
  402. vui_params->colourMatrix = 6;
  403. break;
  404. case VIDEO_CS_DEFAULT:
  405. case VIDEO_CS_709:
  406. vui_params->colourPrimaries = 1;
  407. vui_params->transferCharacteristics = 1;
  408. vui_params->colourMatrix = 1;
  409. break;
  410. case VIDEO_CS_SRGB:
  411. vui_params->colourPrimaries = 1;
  412. vui_params->transferCharacteristics = 13;
  413. vui_params->colourMatrix = 1;
  414. break;
  415. }
  416. enc->bframes = bf;
  417. /* lookahead */
  418. const bool use_profile_lookahead = config->rcParams.enableLookahead;
  419. lookahead = nv_get_cap_h264(enc, NV_ENC_CAPS_SUPPORT_LOOKAHEAD) &&
  420. (lookahead || use_profile_lookahead);
  421. if (lookahead) {
  422. enc->rc_lookahead = use_profile_lookahead
  423. ? config->rcParams.lookaheadDepth
  424. : 8;
  425. }
  426. int buf_count = max(4, config->frameIntervalP * 2 * 2);
  427. if (lookahead) {
  428. buf_count = max(buf_count, config->frameIntervalP +
  429. enc->rc_lookahead +
  430. EXTRA_BUFFERS);
  431. }
  432. buf_count = min(64, buf_count);
  433. enc->buf_count = buf_count;
  434. const int output_delay = buf_count - 1;
  435. enc->output_delay = output_delay;
  436. if (lookahead) {
  437. const int lkd_bound = output_delay - config->frameIntervalP - 4;
  438. if (lkd_bound >= 0) {
  439. config->rcParams.enableLookahead = 1;
  440. config->rcParams.lookaheadDepth =
  441. max(enc->rc_lookahead, lkd_bound);
  442. config->rcParams.disableIadapt = 0;
  443. config->rcParams.disableBadapt = 0;
  444. } else {
  445. lookahead = false;
  446. }
  447. }
  448. /* psycho aq */
  449. if (nv_get_cap_h264(enc, NV_ENC_CAPS_SUPPORT_TEMPORAL_AQ)) {
  450. config->rcParams.enableAQ = psycho_aq;
  451. config->rcParams.aqStrength = 8;
  452. config->rcParams.enableTemporalAQ = psycho_aq;
  453. } else if (psycho_aq) {
  454. warn("Ignoring Psycho Visual Tuning request since GPU is not capable");
  455. }
  456. /* -------------------------- */
  457. /* rate control */
  458. enc->can_change_bitrate =
  459. nv_get_cap_h264(enc, NV_ENC_CAPS_SUPPORT_DYN_BITRATE_CHANGE) &&
  460. !lookahead;
  461. config->rcParams.rateControlMode = twopass ? NV_ENC_PARAMS_RC_VBR_HQ
  462. : NV_ENC_PARAMS_RC_VBR;
  463. if (astrcmpi(rc, "cqp") == 0 || rc_lossless) {
  464. if (lossless) {
  465. h264_config->qpPrimeYZeroTransformBypassFlag = 1;
  466. cqp = 0;
  467. }
  468. config->rcParams.rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
  469. config->rcParams.constQP.qpInterP = cqp;
  470. config->rcParams.constQP.qpInterB = cqp;
  471. config->rcParams.constQP.qpIntra = cqp;
  472. enc->can_change_bitrate = false;
  473. bitrate = 0;
  474. max_bitrate = 0;
  475. } else if (astrcmpi(rc, "vbr") != 0) { /* CBR by default */
  476. h264_config->outputBufferingPeriodSEI = 1;
  477. config->rcParams.rateControlMode =
  478. twopass ? NV_ENC_PARAMS_RC_2_PASS_QUALITY
  479. : NV_ENC_PARAMS_RC_CBR;
  480. }
  481. h264_config->outputPictureTimingSEI = 1;
  482. config->rcParams.averageBitRate = bitrate * 1000;
  483. config->rcParams.maxBitRate = vbr ? max_bitrate * 1000 : bitrate * 1000;
  484. config->rcParams.vbvBufferSize = bitrate * 1000;
  485. /* -------------------------- */
  486. /* profile */
  487. if (astrcmpi(profile, "main") == 0) {
  488. config->profileGUID = NV_ENC_H264_PROFILE_MAIN_GUID;
  489. } else if (astrcmpi(profile, "baseline") == 0) {
  490. config->profileGUID = NV_ENC_H264_PROFILE_BASELINE_GUID;
  491. } else if (!lossless) {
  492. config->profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
  493. }
  494. /* -------------------------- */
  495. /* initialize */
  496. if (NV_FAILED(nv.nvEncInitializeEncoder(enc->session, params))) {
  497. return false;
  498. }
  499. info("settings:\n"
  500. "\trate_control: %s\n"
  501. "\tbitrate: %d\n"
  502. "\tcqp: %d\n"
  503. "\tkeyint: %d\n"
  504. "\tpreset: %s\n"
  505. "\tprofile: %s\n"
  506. "\twidth: %d\n"
  507. "\theight: %d\n"
  508. "\t2-pass: %s\n"
  509. "\tb-frames: %d\n"
  510. "\tlookahead: %s\n"
  511. "\tpsycho_aq: %s\n",
  512. rc, bitrate, cqp, gop_size, preset, profile, enc->cx, enc->cy,
  513. twopass ? "true" : "false", bf, lookahead ? "true" : "false",
  514. psycho_aq ? "true" : "false");
  515. return true;
  516. }
  517. #ifdef ENABLE_HEVC
  518. static bool init_encoder_hevc(struct nvenc_data *enc, obs_data_t *settings,
  519. int bf, bool psycho_aq)
  520. {
  521. const char *rc = obs_data_get_string(settings, "rate_control");
  522. int bitrate = (int)obs_data_get_int(settings, "bitrate");
  523. int max_bitrate = (int)obs_data_get_int(settings, "max_bitrate");
  524. int cqp = (int)obs_data_get_int(settings, "cqp");
  525. int keyint_sec = (int)obs_data_get_int(settings, "keyint_sec");
  526. const char *preset = obs_data_get_string(settings, "preset");
  527. const char *profile = obs_data_get_string(settings, "profile");
  528. bool lookahead = obs_data_get_bool(settings, "lookahead");
  529. bool vbr = astrcmpi(rc, "VBR") == 0;
  530. NVENCSTATUS err;
  531. video_t *video = obs_encoder_video(enc->encoder);
  532. const struct video_output_info *voi = video_output_get_info(video);
  533. enc->cx = voi->width;
  534. enc->cy = voi->height;
  535. /* -------------------------- */
  536. /* get preset */
  537. GUID nv_preset = NV_ENC_PRESET_DEFAULT_GUID;
  538. bool twopass = false;
  539. bool hp = false;
  540. bool ll = false;
  541. if (astrcmpi(preset, "hq") == 0) {
  542. nv_preset = NV_ENC_PRESET_HQ_GUID;
  543. } else if (astrcmpi(preset, "mq") == 0) {
  544. nv_preset = NV_ENC_PRESET_HQ_GUID;
  545. twopass = true;
  546. } else if (astrcmpi(preset, "hp") == 0) {
  547. nv_preset = NV_ENC_PRESET_HP_GUID;
  548. hp = true;
  549. } else if (astrcmpi(preset, "ll") == 0) {
  550. nv_preset = NV_ENC_PRESET_LOW_LATENCY_DEFAULT_GUID;
  551. ll = true;
  552. } else if (astrcmpi(preset, "llhq") == 0) {
  553. nv_preset = NV_ENC_PRESET_LOW_LATENCY_HQ_GUID;
  554. ll = true;
  555. } else if (astrcmpi(preset, "llhp") == 0) {
  556. nv_preset = NV_ENC_PRESET_LOW_LATENCY_HP_GUID;
  557. hp = true;
  558. ll = true;
  559. }
  560. const bool rc_lossless = astrcmpi(rc, "lossless") == 0;
  561. bool lossless = rc_lossless;
  562. if (rc_lossless) {
  563. lossless = nv_get_cap_hevc(enc,
  564. NV_ENC_CAPS_SUPPORT_LOSSLESS_ENCODE);
  565. if (lossless) {
  566. nv_preset = hp ? NV_ENC_PRESET_LOSSLESS_HP_GUID
  567. : NV_ENC_PRESET_LOSSLESS_DEFAULT_GUID;
  568. } else {
  569. warn("lossless encode is not supported, ignoring");
  570. }
  571. }
  572. /* -------------------------- */
  573. /* get preset default config */
  574. NV_ENC_PRESET_CONFIG preset_config = {NV_ENC_PRESET_CONFIG_VER,
  575. {NV_ENC_CONFIG_VER}};
  576. err = nv.nvEncGetEncodePresetConfig(enc->session,
  577. NV_ENC_CODEC_HEVC_GUID, nv_preset,
  578. &preset_config);
  579. if (nv_failed(enc->encoder, err, __FUNCTION__,
  580. "nvEncGetEncodePresetConfig")) {
  581. return false;
  582. }
  583. /* -------------------------- */
  584. /* main configuration */
  585. enc->config = preset_config.presetCfg;
  586. uint32_t gop_size =
  587. (keyint_sec) ? keyint_sec * voi->fps_num / voi->fps_den : 250;
  588. NV_ENC_INITIALIZE_PARAMS *params = &enc->params;
  589. NV_ENC_CONFIG *config = &enc->config;
  590. NV_ENC_CONFIG_HEVC *hevc_config = &config->encodeCodecConfig.hevcConfig;
  591. NV_ENC_CONFIG_H264_VUI_PARAMETERS *vui_params =
  592. &hevc_config->hevcVUIParameters;
  593. int darWidth, darHeight;
  594. av_reduce(&darWidth, &darHeight, voi->width, voi->height, 1024 * 1024);
  595. memset(params, 0, sizeof(*params));
  596. params->version = NV_ENC_INITIALIZE_PARAMS_VER;
  597. params->encodeGUID = NV_ENC_CODEC_HEVC_GUID;
  598. params->presetGUID = nv_preset;
  599. params->encodeWidth = voi->width;
  600. params->encodeHeight = voi->height;
  601. params->darWidth = darWidth;
  602. params->darHeight = darHeight;
  603. params->frameRateNum = voi->fps_num;
  604. params->frameRateDen = voi->fps_den;
  605. params->enableEncodeAsync = 1;
  606. params->enablePTD = 1;
  607. params->encodeConfig = &enc->config;
  608. config->gopLength = gop_size;
  609. config->frameIntervalP = 1 + bf;
  610. hevc_config->idrPeriod = gop_size;
  611. bool repeat_headers = obs_data_get_bool(settings, "repeat_headers");
  612. if (repeat_headers) {
  613. hevc_config->repeatSPSPPS = 1;
  614. hevc_config->disableSPSPPS = 0;
  615. hevc_config->outputAUD = 1;
  616. }
  617. hevc_config->sliceMode = 3;
  618. hevc_config->sliceModeData = 1;
  619. hevc_config->useBFramesAsRef = NV_ENC_BFRAME_REF_MODE_DISABLED;
  620. vui_params->videoSignalTypePresentFlag = 1;
  621. vui_params->videoFullRangeFlag = (voi->range == VIDEO_RANGE_FULL);
  622. vui_params->colourDescriptionPresentFlag = 1;
  623. hevc_config->pixelBitDepthMinus8 = obs_p010_tex_active() ? 2 : 0;
  624. switch (voi->colorspace) {
  625. case VIDEO_CS_601:
  626. vui_params->colourPrimaries = 6;
  627. vui_params->transferCharacteristics = 6;
  628. vui_params->colourMatrix = 6;
  629. break;
  630. case VIDEO_CS_DEFAULT:
  631. case VIDEO_CS_709:
  632. vui_params->colourPrimaries = 1;
  633. vui_params->transferCharacteristics = 1;
  634. vui_params->colourMatrix = 1;
  635. break;
  636. case VIDEO_CS_SRGB:
  637. vui_params->colourPrimaries = 1;
  638. vui_params->transferCharacteristics = 13;
  639. vui_params->colourMatrix = 1;
  640. break;
  641. case VIDEO_CS_2100_PQ:
  642. vui_params->colourPrimaries = 9;
  643. vui_params->transferCharacteristics = 16;
  644. vui_params->colourMatrix = 9;
  645. vui_params->chromaSampleLocationFlag = 1;
  646. vui_params->chromaSampleLocationTop = 2;
  647. vui_params->chromaSampleLocationBot = 2;
  648. break;
  649. case VIDEO_CS_2100_HLG:
  650. vui_params->colourPrimaries = 9;
  651. vui_params->transferCharacteristics = 18;
  652. vui_params->colourMatrix = 9;
  653. vui_params->chromaSampleLocationFlag = 1;
  654. vui_params->chromaSampleLocationTop = 2;
  655. vui_params->chromaSampleLocationBot = 2;
  656. }
  657. enc->bframes = bf;
  658. /* lookahead */
  659. const bool use_profile_lookahead = config->rcParams.enableLookahead;
  660. lookahead = nv_get_cap_hevc(enc, NV_ENC_CAPS_SUPPORT_LOOKAHEAD) &&
  661. (lookahead || use_profile_lookahead);
  662. if (lookahead) {
  663. enc->rc_lookahead = use_profile_lookahead
  664. ? config->rcParams.lookaheadDepth
  665. : 8;
  666. }
  667. int buf_count = max(4, config->frameIntervalP * 2 * 2);
  668. if (lookahead) {
  669. buf_count = max(buf_count, config->frameIntervalP +
  670. enc->rc_lookahead +
  671. EXTRA_BUFFERS);
  672. }
  673. buf_count = min(64, buf_count);
  674. enc->buf_count = buf_count;
  675. const int output_delay = buf_count - 1;
  676. enc->output_delay = output_delay;
  677. if (lookahead) {
  678. const int lkd_bound = output_delay - config->frameIntervalP - 4;
  679. if (lkd_bound >= 0) {
  680. config->rcParams.enableLookahead = 1;
  681. config->rcParams.lookaheadDepth =
  682. max(enc->rc_lookahead, lkd_bound);
  683. config->rcParams.disableIadapt = 0;
  684. config->rcParams.disableBadapt = 0;
  685. } else {
  686. lookahead = false;
  687. }
  688. }
  689. /* psycho aq */
  690. if (nv_get_cap_hevc(enc, NV_ENC_CAPS_SUPPORT_TEMPORAL_AQ)) {
  691. config->rcParams.enableAQ = psycho_aq;
  692. config->rcParams.aqStrength = 8;
  693. config->rcParams.enableTemporalAQ = psycho_aq;
  694. } else if (psycho_aq) {
  695. warn("Ignoring Psycho Visual Tuning request since GPU is not capable");
  696. }
  697. /* -------------------------- */
  698. /* rate control */
  699. enc->can_change_bitrate =
  700. nv_get_cap_hevc(enc, NV_ENC_CAPS_SUPPORT_DYN_BITRATE_CHANGE) &&
  701. !lookahead;
  702. config->rcParams.rateControlMode = twopass ? NV_ENC_PARAMS_RC_VBR_HQ
  703. : NV_ENC_PARAMS_RC_VBR;
  704. if (astrcmpi(rc, "cqp") == 0 || rc_lossless) {
  705. if (lossless)
  706. cqp = 0;
  707. config->rcParams.rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
  708. config->rcParams.constQP.qpInterP = cqp;
  709. config->rcParams.constQP.qpInterB = cqp;
  710. config->rcParams.constQP.qpIntra = cqp;
  711. enc->can_change_bitrate = false;
  712. bitrate = 0;
  713. max_bitrate = 0;
  714. } else if (astrcmpi(rc, "vbr") != 0) { /* CBR by default */
  715. hevc_config->outputBufferingPeriodSEI = 1;
  716. config->rcParams.rateControlMode =
  717. twopass ? NV_ENC_PARAMS_RC_2_PASS_QUALITY
  718. : NV_ENC_PARAMS_RC_CBR;
  719. }
  720. hevc_config->outputPictureTimingSEI = 1;
  721. config->rcParams.averageBitRate = bitrate * 1000;
  722. config->rcParams.maxBitRate = vbr ? max_bitrate * 1000 : bitrate * 1000;
  723. config->rcParams.vbvBufferSize = bitrate * 1000;
  724. /* -------------------------- */
  725. /* profile */
  726. if (astrcmpi(profile, "main10") == 0) {
  727. config->profileGUID = NV_ENC_HEVC_PROFILE_MAIN10_GUID;
  728. } else if (obs_p010_tex_active()) {
  729. blog(LOG_WARNING, "[jim-nvenc] Forcing main10 for P010");
  730. config->profileGUID = NV_ENC_HEVC_PROFILE_MAIN10_GUID;
  731. } else {
  732. config->profileGUID = NV_ENC_HEVC_PROFILE_MAIN_GUID;
  733. }
  734. /* -------------------------- */
  735. /* initialize */
  736. if (NV_FAILED(nv.nvEncInitializeEncoder(enc->session, params))) {
  737. return false;
  738. }
  739. info("settings:\n"
  740. "\trate_control: %s\n"
  741. "\tbitrate: %d\n"
  742. "\tcqp: %d\n"
  743. "\tkeyint: %d\n"
  744. "\tpreset: %s\n"
  745. "\tprofile: %s\n"
  746. "\twidth: %d\n"
  747. "\theight: %d\n"
  748. "\t2-pass: %s\n"
  749. "\tb-frames: %d\n"
  750. "\tlookahead: %s\n"
  751. "\tpsycho_aq: %s\n",
  752. rc, bitrate, cqp, gop_size, preset, profile, enc->cx, enc->cy,
  753. twopass ? "true" : "false", bf, lookahead ? "true" : "false",
  754. psycho_aq ? "true" : "false");
  755. return true;
  756. }
  757. #endif
  758. static bool init_bitstreams(struct nvenc_data *enc)
  759. {
  760. da_reserve(enc->bitstreams, enc->buf_count);
  761. for (int i = 0; i < enc->buf_count; i++) {
  762. struct nv_bitstream bitstream;
  763. if (!nv_bitstream_init(enc, &bitstream)) {
  764. return false;
  765. }
  766. da_push_back(enc->bitstreams, &bitstream);
  767. }
  768. return true;
  769. }
  770. static bool init_textures(struct nvenc_data *enc)
  771. {
  772. da_reserve(enc->textures, enc->buf_count);
  773. for (int i = 0; i < enc->buf_count; i++) {
  774. struct nv_texture texture;
  775. if (!nv_texture_init(enc, &texture)) {
  776. return false;
  777. }
  778. da_push_back(enc->textures, &texture);
  779. }
  780. return true;
  781. }
  782. static void nvenc_destroy(void *data);
  783. static bool init_specific_encoder(struct nvenc_data *enc, bool hevc,
  784. obs_data_t *settings, int bf, bool psycho_aq)
  785. {
  786. #ifdef ENABLE_HEVC
  787. if (hevc)
  788. return init_encoder_hevc(enc, settings, bf, psycho_aq);
  789. #endif
  790. return init_encoder_h264(enc, settings, bf, psycho_aq);
  791. }
  792. static bool init_encoder(struct nvenc_data *enc, bool hevc,
  793. obs_data_t *settings)
  794. {
  795. const int bf = (int)obs_data_get_int(settings, "bf");
  796. const bool psycho_aq = obs_data_get_bool(settings, "psycho_aq");
  797. #ifdef ENABLE_HEVC
  798. const bool support_10bit =
  799. hevc ? nv_get_cap_hevc(enc, NV_ENC_CAPS_SUPPORT_10BIT_ENCODE)
  800. : nv_get_cap_h264(enc, NV_ENC_CAPS_SUPPORT_10BIT_ENCODE);
  801. const int bf_max =
  802. hevc ? nv_get_cap_hevc(enc, NV_ENC_CAPS_NUM_MAX_BFRAMES)
  803. : nv_get_cap_h264(enc, NV_ENC_CAPS_NUM_MAX_BFRAMES);
  804. #else
  805. const bool support_10bit =
  806. nv_get_cap_h264(enc, NV_ENC_CAPS_SUPPORT_10BIT_ENCODE);
  807. const int bf_max = nv_get_cap_h264(enc, NV_ENC_CAPS_NUM_MAX_BFRAMES);
  808. #endif
  809. if (obs_p010_tex_active() && !support_10bit) {
  810. NV_FAIL(obs_module_text("NVENC.10bitUnsupported"));
  811. return false;
  812. }
  813. video_t *video = obs_encoder_video(enc->encoder);
  814. const struct video_output_info *voi = video_output_get_info(video);
  815. switch (voi->format) {
  816. case VIDEO_FORMAT_I010:
  817. case VIDEO_FORMAT_P010:
  818. break;
  819. default:
  820. switch (voi->colorspace) {
  821. case VIDEO_CS_2100_PQ:
  822. case VIDEO_CS_2100_HLG:
  823. NV_FAIL(obs_module_text("NVENC.8bitUnsupportedHdr"));
  824. return false;
  825. }
  826. }
  827. if (bf > bf_max) {
  828. NV_FAIL(obs_module_text("NVENC.TooManyBFrames"), bf, bf_max);
  829. return false;
  830. }
  831. if (!init_specific_encoder(enc, hevc, settings, bf, psycho_aq)) {
  832. if (!psycho_aq)
  833. return false;
  834. blog(LOG_WARNING, "[jim-nvenc] init_specific_encoder failed, "
  835. "trying again without Psycho Visual Tuning");
  836. nv.nvEncDestroyEncoder(enc->session);
  837. enc->session = NULL;
  838. if (!init_session(enc) ||
  839. !init_specific_encoder(enc, hevc, settings, bf, false)) {
  840. return false;
  841. }
  842. }
  843. return true;
  844. }
  845. static void *nvenc_create_internal(bool hevc, obs_data_t *settings,
  846. obs_encoder_t *encoder)
  847. {
  848. NV_ENCODE_API_FUNCTION_LIST init = {NV_ENCODE_API_FUNCTION_LIST_VER};
  849. struct nvenc_data *enc = bzalloc(sizeof(*enc));
  850. enc->encoder = encoder;
  851. enc->first_packet = true;
  852. if (!init_nvenc(encoder)) {
  853. goto fail;
  854. }
  855. if (NV_FAILED(nv_create_instance(&init))) {
  856. goto fail;
  857. }
  858. if (!init_d3d11(enc, settings)) {
  859. goto fail;
  860. }
  861. if (!init_session(enc)) {
  862. goto fail;
  863. }
  864. if (!init_encoder(enc, hevc, settings)) {
  865. goto fail;
  866. }
  867. if (!init_bitstreams(enc)) {
  868. goto fail;
  869. }
  870. if (!init_textures(enc)) {
  871. goto fail;
  872. }
  873. #ifdef ENABLE_HEVC
  874. enc->hevc = hevc;
  875. #endif
  876. return enc;
  877. fail:
  878. nvenc_destroy(enc);
  879. return NULL;
  880. }
  881. static void *nvenc_create_h264_hevc(bool hevc, obs_data_t *settings,
  882. obs_encoder_t *encoder)
  883. {
  884. /* this encoder requires shared textures, this cannot be used on a
  885. * gpu other than the one OBS is currently running on. */
  886. const int gpu = (int)obs_data_get_int(settings, "gpu");
  887. if (gpu != 0) {
  888. blog(LOG_INFO,
  889. "[jim-nvenc] different GPU selected by user, falling back to ffmpeg");
  890. goto reroute;
  891. }
  892. if (obs_encoder_scaling_enabled(encoder)) {
  893. blog(LOG_INFO,
  894. "[jim-nvenc] scaling enabled, falling back to ffmpeg");
  895. goto reroute;
  896. }
  897. if (!obs_p010_tex_active() && !obs_nv12_tex_active()) {
  898. blog(LOG_INFO,
  899. "[jim-nvenc] nv12/p010 not active, falling back to ffmpeg");
  900. goto reroute;
  901. }
  902. struct nvenc_data *enc = nvenc_create_internal(hevc, settings, encoder);
  903. if (enc) {
  904. return enc;
  905. }
  906. reroute:
  907. const char *fallback_name = "ffmpeg_nvenc";
  908. #ifdef ENABLE_HEVC
  909. if (hevc)
  910. fallback_name = "ffmpeg_hevc_nvenc";
  911. #endif
  912. return obs_encoder_create_rerouted(encoder, fallback_name);
  913. }
  914. static void *h264_nvenc_create(obs_data_t *settings, obs_encoder_t *encoder)
  915. {
  916. return nvenc_create_h264_hevc(false, settings, encoder);
  917. }
  918. #ifdef ENABLE_HEVC
  919. static void *hevc_nvenc_create(obs_data_t *settings, obs_encoder_t *encoder)
  920. {
  921. return nvenc_create_h264_hevc(true, settings, encoder);
  922. }
  923. #endif
  924. static bool get_encoded_packet(struct nvenc_data *enc, bool finalize);
  925. static void nvenc_destroy(void *data)
  926. {
  927. struct nvenc_data *enc = data;
  928. if (enc->encode_started) {
  929. size_t next_bitstream = enc->next_bitstream;
  930. HANDLE next_event = enc->bitstreams.array[next_bitstream].event;
  931. NV_ENC_PIC_PARAMS params = {NV_ENC_PIC_PARAMS_VER};
  932. params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
  933. params.completionEvent = next_event;
  934. nv.nvEncEncodePicture(enc->session, &params);
  935. get_encoded_packet(enc, true);
  936. }
  937. for (size_t i = 0; i < enc->textures.num; i++) {
  938. nv_texture_free(enc, &enc->textures.array[i]);
  939. }
  940. for (size_t i = 0; i < enc->bitstreams.num; i++) {
  941. nv_bitstream_free(enc, &enc->bitstreams.array[i]);
  942. }
  943. if (enc->session) {
  944. nv.nvEncDestroyEncoder(enc->session);
  945. }
  946. for (size_t i = 0; i < enc->input_textures.num; i++) {
  947. ID3D11Texture2D *tex = enc->input_textures.array[i].tex;
  948. IDXGIKeyedMutex *km = enc->input_textures.array[i].km;
  949. tex->lpVtbl->Release(tex);
  950. km->lpVtbl->Release(km);
  951. }
  952. if (enc->context) {
  953. enc->context->lpVtbl->Release(enc->context);
  954. }
  955. if (enc->device) {
  956. enc->device->lpVtbl->Release(enc->device);
  957. }
  958. bfree(enc->header);
  959. bfree(enc->sei);
  960. circlebuf_free(&enc->dts_list);
  961. da_free(enc->textures);
  962. da_free(enc->bitstreams);
  963. da_free(enc->input_textures);
  964. da_free(enc->packet_data);
  965. bfree(enc);
  966. }
  967. static ID3D11Texture2D *get_tex_from_handle(struct nvenc_data *enc,
  968. uint32_t handle,
  969. IDXGIKeyedMutex **km_out)
  970. {
  971. ID3D11Device *device = enc->device;
  972. IDXGIKeyedMutex *km;
  973. ID3D11Texture2D *input_tex;
  974. HRESULT hr;
  975. for (size_t i = 0; i < enc->input_textures.num; i++) {
  976. struct handle_tex *ht = &enc->input_textures.array[i];
  977. if (ht->handle == handle) {
  978. *km_out = ht->km;
  979. return ht->tex;
  980. }
  981. }
  982. hr = device->lpVtbl->OpenSharedResource(device,
  983. (HANDLE)(uintptr_t)handle,
  984. &IID_ID3D11Texture2D,
  985. &input_tex);
  986. if (FAILED(hr)) {
  987. error_hr("OpenSharedResource failed");
  988. return NULL;
  989. }
  990. hr = input_tex->lpVtbl->QueryInterface(input_tex, &IID_IDXGIKeyedMutex,
  991. &km);
  992. if (FAILED(hr)) {
  993. error_hr("QueryInterface(IDXGIKeyedMutex) failed");
  994. input_tex->lpVtbl->Release(input_tex);
  995. return NULL;
  996. }
  997. input_tex->lpVtbl->SetEvictionPriority(input_tex,
  998. DXGI_RESOURCE_PRIORITY_MAXIMUM);
  999. *km_out = km;
  1000. struct handle_tex new_ht = {handle, input_tex, km};
  1001. da_push_back(enc->input_textures, &new_ht);
  1002. return input_tex;
  1003. }
  1004. static bool get_encoded_packet(struct nvenc_data *enc, bool finalize)
  1005. {
  1006. void *s = enc->session;
  1007. da_resize(enc->packet_data, 0);
  1008. if (!enc->buffers_queued)
  1009. return true;
  1010. if (!finalize && enc->buffers_queued < enc->output_delay)
  1011. return true;
  1012. size_t count = finalize ? enc->buffers_queued : 1;
  1013. for (size_t i = 0; i < count; i++) {
  1014. size_t cur_bs_idx = enc->cur_bitstream;
  1015. struct nv_bitstream *bs = &enc->bitstreams.array[cur_bs_idx];
  1016. struct nv_texture *nvtex = &enc->textures.array[cur_bs_idx];
  1017. /* ---------------- */
  1018. NV_ENC_LOCK_BITSTREAM lock = {NV_ENC_LOCK_BITSTREAM_VER};
  1019. lock.outputBitstream = bs->ptr;
  1020. lock.doNotWait = false;
  1021. if (NV_FAILED(nv.nvEncLockBitstream(s, &lock))) {
  1022. return false;
  1023. }
  1024. if (enc->first_packet) {
  1025. uint8_t *new_packet;
  1026. size_t size;
  1027. enc->first_packet = false;
  1028. #ifdef ENABLE_HEVC
  1029. if (enc->hevc) {
  1030. obs_extract_hevc_headers(
  1031. lock.bitstreamBufferPtr,
  1032. lock.bitstreamSizeInBytes, &new_packet,
  1033. &size, &enc->header, &enc->header_size,
  1034. &enc->sei, &enc->sei_size);
  1035. } else
  1036. #endif
  1037. {
  1038. obs_extract_avc_headers(
  1039. lock.bitstreamBufferPtr,
  1040. lock.bitstreamSizeInBytes, &new_packet,
  1041. &size, &enc->header, &enc->header_size,
  1042. &enc->sei, &enc->sei_size);
  1043. }
  1044. da_copy_array(enc->packet_data, new_packet, size);
  1045. bfree(new_packet);
  1046. } else {
  1047. da_copy_array(enc->packet_data, lock.bitstreamBufferPtr,
  1048. lock.bitstreamSizeInBytes);
  1049. }
  1050. enc->packet_pts = (int64_t)lock.outputTimeStamp;
  1051. enc->packet_keyframe = lock.pictureType == NV_ENC_PIC_TYPE_IDR;
  1052. if (NV_FAILED(nv.nvEncUnlockBitstream(s, bs->ptr))) {
  1053. return false;
  1054. }
  1055. /* ---------------- */
  1056. if (nvtex->mapped_res) {
  1057. NVENCSTATUS err;
  1058. err = nv.nvEncUnmapInputResource(s, nvtex->mapped_res);
  1059. if (nv_failed(enc->encoder, err, __FUNCTION__,
  1060. "unmap")) {
  1061. return false;
  1062. }
  1063. nvtex->mapped_res = NULL;
  1064. }
  1065. /* ---------------- */
  1066. if (++enc->cur_bitstream == enc->buf_count)
  1067. enc->cur_bitstream = 0;
  1068. enc->buffers_queued--;
  1069. }
  1070. return true;
  1071. }
  1072. static bool nvenc_encode_tex(void *data, uint32_t handle, int64_t pts,
  1073. uint64_t lock_key, uint64_t *next_key,
  1074. struct encoder_packet *packet,
  1075. bool *received_packet)
  1076. {
  1077. struct nvenc_data *enc = data;
  1078. ID3D11Device *device = enc->device;
  1079. ID3D11DeviceContext *context = enc->context;
  1080. ID3D11Texture2D *input_tex;
  1081. ID3D11Texture2D *output_tex;
  1082. IDXGIKeyedMutex *km;
  1083. struct nv_texture *nvtex;
  1084. struct nv_bitstream *bs;
  1085. NVENCSTATUS err;
  1086. if (handle == GS_INVALID_HANDLE) {
  1087. error("Encode failed: bad texture handle");
  1088. *next_key = lock_key;
  1089. return false;
  1090. }
  1091. bs = &enc->bitstreams.array[enc->next_bitstream];
  1092. nvtex = &enc->textures.array[enc->next_bitstream];
  1093. input_tex = get_tex_from_handle(enc, handle, &km);
  1094. output_tex = nvtex->tex;
  1095. if (!input_tex) {
  1096. *next_key = lock_key;
  1097. return false;
  1098. }
  1099. circlebuf_push_back(&enc->dts_list, &pts, sizeof(pts));
  1100. /* ------------------------------------ */
  1101. /* wait for output bitstream/tex */
  1102. WaitForSingleObject(bs->event, INFINITE);
  1103. /* ------------------------------------ */
  1104. /* copy to output tex */
  1105. km->lpVtbl->AcquireSync(km, lock_key, INFINITE);
  1106. context->lpVtbl->CopyResource(context, (ID3D11Resource *)output_tex,
  1107. (ID3D11Resource *)input_tex);
  1108. km->lpVtbl->ReleaseSync(km, *next_key);
  1109. /* ------------------------------------ */
  1110. /* map output tex so nvenc can use it */
  1111. NV_ENC_MAP_INPUT_RESOURCE map = {NV_ENC_MAP_INPUT_RESOURCE_VER};
  1112. map.registeredResource = nvtex->res;
  1113. if (NV_FAILED(nv.nvEncMapInputResource(enc->session, &map))) {
  1114. return false;
  1115. }
  1116. nvtex->mapped_res = map.mappedResource;
  1117. /* ------------------------------------ */
  1118. /* do actual encode call */
  1119. NV_ENC_PIC_PARAMS params = {0};
  1120. params.version = NV_ENC_PIC_PARAMS_VER;
  1121. params.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
  1122. params.inputBuffer = nvtex->mapped_res;
  1123. params.bufferFmt = obs_p010_tex_active()
  1124. ? NV_ENC_BUFFER_FORMAT_YUV420_10BIT
  1125. : NV_ENC_BUFFER_FORMAT_NV12;
  1126. params.inputTimeStamp = (uint64_t)pts;
  1127. params.inputWidth = enc->cx;
  1128. params.inputHeight = enc->cy;
  1129. params.inputPitch = enc->cx;
  1130. params.outputBitstream = bs->ptr;
  1131. params.completionEvent = bs->event;
  1132. err = nv.nvEncEncodePicture(enc->session, &params);
  1133. if (err != NV_ENC_SUCCESS && err != NV_ENC_ERR_NEED_MORE_INPUT) {
  1134. nv_failed(enc->encoder, err, __FUNCTION__,
  1135. "nvEncEncodePicture");
  1136. return false;
  1137. }
  1138. enc->encode_started = true;
  1139. enc->buffers_queued++;
  1140. if (++enc->next_bitstream == enc->buf_count) {
  1141. enc->next_bitstream = 0;
  1142. }
  1143. /* ------------------------------------ */
  1144. /* check for encoded packet and parse */
  1145. if (!get_encoded_packet(enc, false)) {
  1146. return false;
  1147. }
  1148. /* ------------------------------------ */
  1149. /* output encoded packet */
  1150. if (enc->packet_data.num) {
  1151. int64_t dts;
  1152. circlebuf_pop_front(&enc->dts_list, &dts, sizeof(dts));
  1153. /* subtract bframe delay from dts */
  1154. dts -= (int64_t)enc->bframes * packet->timebase_num;
  1155. *received_packet = true;
  1156. packet->data = enc->packet_data.array;
  1157. packet->size = enc->packet_data.num;
  1158. packet->type = OBS_ENCODER_VIDEO;
  1159. packet->pts = enc->packet_pts;
  1160. packet->dts = dts;
  1161. packet->keyframe = enc->packet_keyframe;
  1162. } else {
  1163. *received_packet = false;
  1164. }
  1165. return true;
  1166. }
  1167. extern void h264_nvenc_defaults(obs_data_t *settings);
  1168. extern obs_properties_t *h264_nvenc_properties(void *unused);
  1169. #ifdef ENABLE_HEVC
  1170. extern void hevc_nvenc_defaults(obs_data_t *settings);
  1171. extern obs_properties_t *hevc_nvenc_properties(void *unused);
  1172. #endif
  1173. static bool nvenc_extra_data(void *data, uint8_t **header, size_t *size)
  1174. {
  1175. struct nvenc_data *enc = data;
  1176. if (!enc->header) {
  1177. return false;
  1178. }
  1179. *header = enc->header;
  1180. *size = enc->header_size;
  1181. return true;
  1182. }
  1183. static bool nvenc_sei_data(void *data, uint8_t **sei, size_t *size)
  1184. {
  1185. struct nvenc_data *enc = data;
  1186. if (!enc->sei) {
  1187. return false;
  1188. }
  1189. *sei = enc->sei;
  1190. *size = enc->sei_size;
  1191. return true;
  1192. }
  1193. struct obs_encoder_info h264_nvenc_info = {
  1194. .id = "jim_nvenc",
  1195. .codec = "h264",
  1196. .type = OBS_ENCODER_VIDEO,
  1197. .caps = OBS_ENCODER_CAP_PASS_TEXTURE | OBS_ENCODER_CAP_DYN_BITRATE,
  1198. .get_name = h264_nvenc_get_name,
  1199. .create = h264_nvenc_create,
  1200. .destroy = nvenc_destroy,
  1201. .update = nvenc_update,
  1202. .encode_texture = nvenc_encode_tex,
  1203. .get_defaults = h264_nvenc_defaults,
  1204. .get_properties = h264_nvenc_properties,
  1205. .get_extra_data = nvenc_extra_data,
  1206. .get_sei_data = nvenc_sei_data,
  1207. };
  1208. #ifdef ENABLE_HEVC
  1209. struct obs_encoder_info hevc_nvenc_info = {
  1210. .id = "jim_hevc_nvenc",
  1211. .codec = "hevc",
  1212. .type = OBS_ENCODER_VIDEO,
  1213. .caps = OBS_ENCODER_CAP_PASS_TEXTURE | OBS_ENCODER_CAP_DYN_BITRATE,
  1214. .get_name = hevc_nvenc_get_name,
  1215. .create = hevc_nvenc_create,
  1216. .destroy = nvenc_destroy,
  1217. .update = nvenc_update,
  1218. .encode_texture = nvenc_encode_tex,
  1219. .get_defaults = hevc_nvenc_defaults,
  1220. .get_properties = hevc_nvenc_properties,
  1221. .get_extra_data = nvenc_extra_data,
  1222. .get_sei_data = nvenc_sei_data,
  1223. };
  1224. #endif