jim-nvenc.c 38 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460
  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. break;
  646. case VIDEO_CS_2100_HLG:
  647. vui_params->colourPrimaries = 9;
  648. vui_params->transferCharacteristics = 18;
  649. vui_params->colourMatrix = 9;
  650. }
  651. enc->bframes = bf;
  652. /* lookahead */
  653. const bool use_profile_lookahead = config->rcParams.enableLookahead;
  654. lookahead = nv_get_cap_hevc(enc, NV_ENC_CAPS_SUPPORT_LOOKAHEAD) &&
  655. (lookahead || use_profile_lookahead);
  656. if (lookahead) {
  657. enc->rc_lookahead = use_profile_lookahead
  658. ? config->rcParams.lookaheadDepth
  659. : 8;
  660. }
  661. int buf_count = max(4, config->frameIntervalP * 2 * 2);
  662. if (lookahead) {
  663. buf_count = max(buf_count, config->frameIntervalP +
  664. enc->rc_lookahead +
  665. EXTRA_BUFFERS);
  666. }
  667. buf_count = min(64, buf_count);
  668. enc->buf_count = buf_count;
  669. const int output_delay = buf_count - 1;
  670. enc->output_delay = output_delay;
  671. if (lookahead) {
  672. const int lkd_bound = output_delay - config->frameIntervalP - 4;
  673. if (lkd_bound >= 0) {
  674. config->rcParams.enableLookahead = 1;
  675. config->rcParams.lookaheadDepth =
  676. max(enc->rc_lookahead, lkd_bound);
  677. config->rcParams.disableIadapt = 0;
  678. config->rcParams.disableBadapt = 0;
  679. } else {
  680. lookahead = false;
  681. }
  682. }
  683. /* psycho aq */
  684. if (nv_get_cap_hevc(enc, NV_ENC_CAPS_SUPPORT_TEMPORAL_AQ)) {
  685. config->rcParams.enableAQ = psycho_aq;
  686. config->rcParams.aqStrength = 8;
  687. config->rcParams.enableTemporalAQ = psycho_aq;
  688. } else if (psycho_aq) {
  689. warn("Ignoring Psycho Visual Tuning request since GPU is not capable");
  690. }
  691. /* -------------------------- */
  692. /* rate control */
  693. enc->can_change_bitrate =
  694. nv_get_cap_hevc(enc, NV_ENC_CAPS_SUPPORT_DYN_BITRATE_CHANGE) &&
  695. !lookahead;
  696. config->rcParams.rateControlMode = twopass ? NV_ENC_PARAMS_RC_VBR_HQ
  697. : NV_ENC_PARAMS_RC_VBR;
  698. if (astrcmpi(rc, "cqp") == 0 || rc_lossless) {
  699. if (lossless)
  700. cqp = 0;
  701. config->rcParams.rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
  702. config->rcParams.constQP.qpInterP = cqp;
  703. config->rcParams.constQP.qpInterB = cqp;
  704. config->rcParams.constQP.qpIntra = cqp;
  705. enc->can_change_bitrate = false;
  706. bitrate = 0;
  707. max_bitrate = 0;
  708. } else if (astrcmpi(rc, "vbr") != 0) { /* CBR by default */
  709. hevc_config->outputBufferingPeriodSEI = 1;
  710. config->rcParams.rateControlMode =
  711. twopass ? NV_ENC_PARAMS_RC_2_PASS_QUALITY
  712. : NV_ENC_PARAMS_RC_CBR;
  713. }
  714. hevc_config->outputPictureTimingSEI = 1;
  715. config->rcParams.averageBitRate = bitrate * 1000;
  716. config->rcParams.maxBitRate = vbr ? max_bitrate * 1000 : bitrate * 1000;
  717. config->rcParams.vbvBufferSize = bitrate * 1000;
  718. /* -------------------------- */
  719. /* profile */
  720. if (astrcmpi(profile, "main10") == 0) {
  721. config->profileGUID = NV_ENC_HEVC_PROFILE_MAIN10_GUID;
  722. } else if (obs_p010_tex_active()) {
  723. blog(LOG_WARNING, "[jim-nvenc] Forcing main10 for P010");
  724. config->profileGUID = NV_ENC_HEVC_PROFILE_MAIN10_GUID;
  725. } else {
  726. config->profileGUID = NV_ENC_HEVC_PROFILE_MAIN_GUID;
  727. }
  728. /* -------------------------- */
  729. /* initialize */
  730. if (NV_FAILED(nv.nvEncInitializeEncoder(enc->session, params))) {
  731. return false;
  732. }
  733. info("settings:\n"
  734. "\trate_control: %s\n"
  735. "\tbitrate: %d\n"
  736. "\tcqp: %d\n"
  737. "\tkeyint: %d\n"
  738. "\tpreset: %s\n"
  739. "\tprofile: %s\n"
  740. "\twidth: %d\n"
  741. "\theight: %d\n"
  742. "\t2-pass: %s\n"
  743. "\tb-frames: %d\n"
  744. "\tlookahead: %s\n"
  745. "\tpsycho_aq: %s\n",
  746. rc, bitrate, cqp, gop_size, preset, profile, enc->cx, enc->cy,
  747. twopass ? "true" : "false", bf, lookahead ? "true" : "false",
  748. psycho_aq ? "true" : "false");
  749. return true;
  750. }
  751. #endif
  752. static bool init_bitstreams(struct nvenc_data *enc)
  753. {
  754. da_reserve(enc->bitstreams, enc->buf_count);
  755. for (int i = 0; i < enc->buf_count; i++) {
  756. struct nv_bitstream bitstream;
  757. if (!nv_bitstream_init(enc, &bitstream)) {
  758. return false;
  759. }
  760. da_push_back(enc->bitstreams, &bitstream);
  761. }
  762. return true;
  763. }
  764. static bool init_textures(struct nvenc_data *enc)
  765. {
  766. da_reserve(enc->textures, enc->buf_count);
  767. for (int i = 0; i < enc->buf_count; i++) {
  768. struct nv_texture texture;
  769. if (!nv_texture_init(enc, &texture)) {
  770. return false;
  771. }
  772. da_push_back(enc->textures, &texture);
  773. }
  774. return true;
  775. }
  776. static void nvenc_destroy(void *data);
  777. static bool init_specific_encoder(struct nvenc_data *enc, bool hevc,
  778. obs_data_t *settings, obs_encoder_t *encoder,
  779. int bf, bool psycho_aq)
  780. {
  781. bool init = false;
  782. #ifdef ENABLE_HEVC
  783. if (hevc)
  784. return init_encoder_hevc(enc, settings, bf, psycho_aq);
  785. #endif
  786. return init_encoder_h264(enc, settings, bf, psycho_aq);
  787. }
  788. static bool init_encoder(struct nvenc_data *enc, bool hevc,
  789. obs_data_t *settings, obs_encoder_t *encoder)
  790. {
  791. const int bf = (int)obs_data_get_int(settings, "bf");
  792. const bool psycho_aq = obs_data_get_bool(settings, "psycho_aq");
  793. #ifdef ENABLE_HEVC
  794. const bool support_10bit =
  795. hevc ? nv_get_cap_hevc(enc, NV_ENC_CAPS_SUPPORT_10BIT_ENCODE)
  796. : nv_get_cap_h264(enc, NV_ENC_CAPS_SUPPORT_10BIT_ENCODE);
  797. const int bf_max =
  798. hevc ? nv_get_cap_hevc(enc, NV_ENC_CAPS_NUM_MAX_BFRAMES)
  799. : nv_get_cap_h264(enc, NV_ENC_CAPS_NUM_MAX_BFRAMES);
  800. #else
  801. const bool support_10bit =
  802. nv_get_cap_h264(enc, NV_ENC_CAPS_SUPPORT_10BIT_ENCODE);
  803. const int bf_max = nv_get_cap_h264(enc, NV_ENC_CAPS_NUM_MAX_BFRAMES);
  804. #endif
  805. if (obs_p010_tex_active() && !support_10bit) {
  806. NV_FAIL(obs_module_text("NVENC.10bitUnsupported"));
  807. return false;
  808. }
  809. if (bf > bf_max) {
  810. NV_FAIL(obs_module_text("NVENC.TooManyBFrames"), bf, bf_max);
  811. return false;
  812. }
  813. if (!init_specific_encoder(enc, hevc, settings, encoder, bf,
  814. psycho_aq)) {
  815. if (!psycho_aq)
  816. return false;
  817. blog(LOG_WARNING, "[jim-nvenc] init_specific_encoder failed, "
  818. "trying again without Psycho Visual Tuning");
  819. if (!init_specific_encoder(enc, hevc, settings, encoder, bf,
  820. psycho_aq)) {
  821. return false;
  822. }
  823. }
  824. return true;
  825. }
  826. static void *nvenc_create_internal(bool hevc, obs_data_t *settings,
  827. obs_encoder_t *encoder)
  828. {
  829. NV_ENCODE_API_FUNCTION_LIST init = {NV_ENCODE_API_FUNCTION_LIST_VER};
  830. struct nvenc_data *enc = bzalloc(sizeof(*enc));
  831. enc->encoder = encoder;
  832. enc->first_packet = true;
  833. if (!init_nvenc(encoder)) {
  834. goto fail;
  835. }
  836. if (NV_FAILED(nv_create_instance(&init))) {
  837. goto fail;
  838. }
  839. if (!init_d3d11(enc, settings)) {
  840. goto fail;
  841. }
  842. if (!init_session(enc)) {
  843. goto fail;
  844. }
  845. if (!init_encoder(enc, hevc, settings, encoder)) {
  846. goto fail;
  847. }
  848. if (!init_bitstreams(enc)) {
  849. goto fail;
  850. }
  851. if (!init_textures(enc)) {
  852. goto fail;
  853. }
  854. #ifdef ENABLE_HEVC
  855. enc->hevc = hevc;
  856. #endif
  857. return enc;
  858. fail:
  859. nvenc_destroy(enc);
  860. return NULL;
  861. }
  862. static void *nvenc_create_h264_hevc(bool hevc, obs_data_t *settings,
  863. obs_encoder_t *encoder)
  864. {
  865. /* this encoder requires shared textures, this cannot be used on a
  866. * gpu other than the one OBS is currently running on. */
  867. const int gpu = (int)obs_data_get_int(settings, "gpu");
  868. if (gpu != 0) {
  869. blog(LOG_INFO,
  870. "[jim-nvenc] different GPU selected by user, falling back to ffmpeg");
  871. goto reroute;
  872. }
  873. if (obs_encoder_scaling_enabled(encoder)) {
  874. blog(LOG_INFO,
  875. "[jim-nvenc] scaling enabled, falling back to ffmpeg");
  876. goto reroute;
  877. }
  878. if (!obs_p010_tex_active() && !obs_nv12_tex_active()) {
  879. blog(LOG_INFO,
  880. "[jim-nvenc] nv12/p010 not active, falling back to ffmpeg");
  881. goto reroute;
  882. }
  883. struct nvenc_data *enc = nvenc_create_internal(hevc, settings, encoder);
  884. if (enc) {
  885. return enc;
  886. }
  887. reroute:
  888. const char *fallback_name = "ffmpeg_nvenc";
  889. #ifdef ENABLE_HEVC
  890. if (hevc)
  891. fallback_name = "ffmpeg_hevc_nvenc";
  892. #endif
  893. return obs_encoder_create_rerouted(encoder, fallback_name);
  894. }
  895. static void *h264_nvenc_create(obs_data_t *settings, obs_encoder_t *encoder)
  896. {
  897. return nvenc_create_h264_hevc(false, settings, encoder);
  898. }
  899. #ifdef ENABLE_HEVC
  900. static void *hevc_nvenc_create(obs_data_t *settings, obs_encoder_t *encoder)
  901. {
  902. return nvenc_create_h264_hevc(true, settings, encoder);
  903. }
  904. #endif
  905. static bool get_encoded_packet(struct nvenc_data *enc, bool finalize);
  906. static void nvenc_destroy(void *data)
  907. {
  908. struct nvenc_data *enc = data;
  909. if (enc->encode_started) {
  910. size_t next_bitstream = enc->next_bitstream;
  911. HANDLE next_event = enc->bitstreams.array[next_bitstream].event;
  912. NV_ENC_PIC_PARAMS params = {NV_ENC_PIC_PARAMS_VER};
  913. params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
  914. params.completionEvent = next_event;
  915. nv.nvEncEncodePicture(enc->session, &params);
  916. get_encoded_packet(enc, true);
  917. }
  918. for (size_t i = 0; i < enc->textures.num; i++) {
  919. nv_texture_free(enc, &enc->textures.array[i]);
  920. }
  921. for (size_t i = 0; i < enc->bitstreams.num; i++) {
  922. nv_bitstream_free(enc, &enc->bitstreams.array[i]);
  923. }
  924. if (enc->session) {
  925. nv.nvEncDestroyEncoder(enc->session);
  926. }
  927. for (size_t i = 0; i < enc->input_textures.num; i++) {
  928. ID3D11Texture2D *tex = enc->input_textures.array[i].tex;
  929. IDXGIKeyedMutex *km = enc->input_textures.array[i].km;
  930. tex->lpVtbl->Release(tex);
  931. km->lpVtbl->Release(km);
  932. }
  933. if (enc->context) {
  934. enc->context->lpVtbl->Release(enc->context);
  935. }
  936. if (enc->device) {
  937. enc->device->lpVtbl->Release(enc->device);
  938. }
  939. bfree(enc->header);
  940. bfree(enc->sei);
  941. circlebuf_free(&enc->dts_list);
  942. da_free(enc->textures);
  943. da_free(enc->bitstreams);
  944. da_free(enc->input_textures);
  945. da_free(enc->packet_data);
  946. bfree(enc);
  947. }
  948. static ID3D11Texture2D *get_tex_from_handle(struct nvenc_data *enc,
  949. uint32_t handle,
  950. IDXGIKeyedMutex **km_out)
  951. {
  952. ID3D11Device *device = enc->device;
  953. IDXGIKeyedMutex *km;
  954. ID3D11Texture2D *input_tex;
  955. HRESULT hr;
  956. for (size_t i = 0; i < enc->input_textures.num; i++) {
  957. struct handle_tex *ht = &enc->input_textures.array[i];
  958. if (ht->handle == handle) {
  959. *km_out = ht->km;
  960. return ht->tex;
  961. }
  962. }
  963. hr = device->lpVtbl->OpenSharedResource(device,
  964. (HANDLE)(uintptr_t)handle,
  965. &IID_ID3D11Texture2D,
  966. &input_tex);
  967. if (FAILED(hr)) {
  968. error_hr("OpenSharedResource failed");
  969. return NULL;
  970. }
  971. hr = input_tex->lpVtbl->QueryInterface(input_tex, &IID_IDXGIKeyedMutex,
  972. &km);
  973. if (FAILED(hr)) {
  974. error_hr("QueryInterface(IDXGIKeyedMutex) failed");
  975. input_tex->lpVtbl->Release(input_tex);
  976. return NULL;
  977. }
  978. input_tex->lpVtbl->SetEvictionPriority(input_tex,
  979. DXGI_RESOURCE_PRIORITY_MAXIMUM);
  980. *km_out = km;
  981. struct handle_tex new_ht = {handle, input_tex, km};
  982. da_push_back(enc->input_textures, &new_ht);
  983. return input_tex;
  984. }
  985. static bool get_encoded_packet(struct nvenc_data *enc, bool finalize)
  986. {
  987. void *s = enc->session;
  988. da_resize(enc->packet_data, 0);
  989. if (!enc->buffers_queued)
  990. return true;
  991. if (!finalize && enc->buffers_queued < enc->output_delay)
  992. return true;
  993. size_t count = finalize ? enc->buffers_queued : 1;
  994. for (size_t i = 0; i < count; i++) {
  995. size_t cur_bs_idx = enc->cur_bitstream;
  996. struct nv_bitstream *bs = &enc->bitstreams.array[cur_bs_idx];
  997. struct nv_texture *nvtex = &enc->textures.array[cur_bs_idx];
  998. /* ---------------- */
  999. NV_ENC_LOCK_BITSTREAM lock = {NV_ENC_LOCK_BITSTREAM_VER};
  1000. lock.outputBitstream = bs->ptr;
  1001. lock.doNotWait = false;
  1002. if (NV_FAILED(nv.nvEncLockBitstream(s, &lock))) {
  1003. return false;
  1004. }
  1005. if (enc->first_packet) {
  1006. uint8_t *new_packet;
  1007. size_t size;
  1008. enc->first_packet = false;
  1009. #ifdef ENABLE_HEVC
  1010. if (enc->hevc) {
  1011. obs_extract_hevc_headers(
  1012. lock.bitstreamBufferPtr,
  1013. lock.bitstreamSizeInBytes, &new_packet,
  1014. &size, &enc->header, &enc->header_size,
  1015. &enc->sei, &enc->sei_size);
  1016. } else
  1017. #endif
  1018. {
  1019. obs_extract_avc_headers(
  1020. lock.bitstreamBufferPtr,
  1021. lock.bitstreamSizeInBytes, &new_packet,
  1022. &size, &enc->header, &enc->header_size,
  1023. &enc->sei, &enc->sei_size);
  1024. }
  1025. da_copy_array(enc->packet_data, new_packet, size);
  1026. bfree(new_packet);
  1027. } else {
  1028. da_copy_array(enc->packet_data, lock.bitstreamBufferPtr,
  1029. lock.bitstreamSizeInBytes);
  1030. }
  1031. enc->packet_pts = (int64_t)lock.outputTimeStamp;
  1032. enc->packet_keyframe = lock.pictureType == NV_ENC_PIC_TYPE_IDR;
  1033. if (NV_FAILED(nv.nvEncUnlockBitstream(s, bs->ptr))) {
  1034. return false;
  1035. }
  1036. /* ---------------- */
  1037. if (nvtex->mapped_res) {
  1038. NVENCSTATUS err;
  1039. err = nv.nvEncUnmapInputResource(s, nvtex->mapped_res);
  1040. if (nv_failed(enc->encoder, err, __FUNCTION__,
  1041. "unmap")) {
  1042. return false;
  1043. }
  1044. nvtex->mapped_res = NULL;
  1045. }
  1046. /* ---------------- */
  1047. if (++enc->cur_bitstream == enc->buf_count)
  1048. enc->cur_bitstream = 0;
  1049. enc->buffers_queued--;
  1050. }
  1051. return true;
  1052. }
  1053. static bool nvenc_encode_tex(void *data, uint32_t handle, int64_t pts,
  1054. uint64_t lock_key, uint64_t *next_key,
  1055. struct encoder_packet *packet,
  1056. bool *received_packet)
  1057. {
  1058. struct nvenc_data *enc = data;
  1059. ID3D11Device *device = enc->device;
  1060. ID3D11DeviceContext *context = enc->context;
  1061. ID3D11Texture2D *input_tex;
  1062. ID3D11Texture2D *output_tex;
  1063. IDXGIKeyedMutex *km;
  1064. struct nv_texture *nvtex;
  1065. struct nv_bitstream *bs;
  1066. NVENCSTATUS err;
  1067. if (handle == GS_INVALID_HANDLE) {
  1068. error("Encode failed: bad texture handle");
  1069. *next_key = lock_key;
  1070. return false;
  1071. }
  1072. bs = &enc->bitstreams.array[enc->next_bitstream];
  1073. nvtex = &enc->textures.array[enc->next_bitstream];
  1074. input_tex = get_tex_from_handle(enc, handle, &km);
  1075. output_tex = nvtex->tex;
  1076. if (!input_tex) {
  1077. *next_key = lock_key;
  1078. return false;
  1079. }
  1080. circlebuf_push_back(&enc->dts_list, &pts, sizeof(pts));
  1081. /* ------------------------------------ */
  1082. /* wait for output bitstream/tex */
  1083. WaitForSingleObject(bs->event, INFINITE);
  1084. /* ------------------------------------ */
  1085. /* copy to output tex */
  1086. km->lpVtbl->AcquireSync(km, lock_key, INFINITE);
  1087. context->lpVtbl->CopyResource(context, (ID3D11Resource *)output_tex,
  1088. (ID3D11Resource *)input_tex);
  1089. km->lpVtbl->ReleaseSync(km, *next_key);
  1090. /* ------------------------------------ */
  1091. /* map output tex so nvenc can use it */
  1092. NV_ENC_MAP_INPUT_RESOURCE map = {NV_ENC_MAP_INPUT_RESOURCE_VER};
  1093. map.registeredResource = nvtex->res;
  1094. if (NV_FAILED(nv.nvEncMapInputResource(enc->session, &map))) {
  1095. return false;
  1096. }
  1097. nvtex->mapped_res = map.mappedResource;
  1098. /* ------------------------------------ */
  1099. /* do actual encode call */
  1100. NV_ENC_PIC_PARAMS params = {0};
  1101. params.version = NV_ENC_PIC_PARAMS_VER;
  1102. params.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
  1103. params.inputBuffer = nvtex->mapped_res;
  1104. params.bufferFmt = obs_p010_tex_active()
  1105. ? NV_ENC_BUFFER_FORMAT_YUV420_10BIT
  1106. : NV_ENC_BUFFER_FORMAT_NV12;
  1107. params.inputTimeStamp = (uint64_t)pts;
  1108. params.inputWidth = enc->cx;
  1109. params.inputHeight = enc->cy;
  1110. params.inputPitch = enc->cx;
  1111. params.outputBitstream = bs->ptr;
  1112. params.completionEvent = bs->event;
  1113. err = nv.nvEncEncodePicture(enc->session, &params);
  1114. if (err != NV_ENC_SUCCESS && err != NV_ENC_ERR_NEED_MORE_INPUT) {
  1115. nv_failed(enc->encoder, err, __FUNCTION__,
  1116. "nvEncEncodePicture");
  1117. return false;
  1118. }
  1119. enc->encode_started = true;
  1120. enc->buffers_queued++;
  1121. if (++enc->next_bitstream == enc->buf_count) {
  1122. enc->next_bitstream = 0;
  1123. }
  1124. /* ------------------------------------ */
  1125. /* check for encoded packet and parse */
  1126. if (!get_encoded_packet(enc, false)) {
  1127. return false;
  1128. }
  1129. /* ------------------------------------ */
  1130. /* output encoded packet */
  1131. if (enc->packet_data.num) {
  1132. int64_t dts;
  1133. circlebuf_pop_front(&enc->dts_list, &dts, sizeof(dts));
  1134. /* subtract bframe delay from dts */
  1135. dts -= (int64_t)enc->bframes * packet->timebase_num;
  1136. *received_packet = true;
  1137. packet->data = enc->packet_data.array;
  1138. packet->size = enc->packet_data.num;
  1139. packet->type = OBS_ENCODER_VIDEO;
  1140. packet->pts = enc->packet_pts;
  1141. packet->dts = dts;
  1142. packet->keyframe = enc->packet_keyframe;
  1143. } else {
  1144. *received_packet = false;
  1145. }
  1146. return true;
  1147. }
  1148. extern void h264_nvenc_defaults(obs_data_t *settings);
  1149. extern obs_properties_t *h264_nvenc_properties(void *unused);
  1150. #ifdef ENABLE_HEVC
  1151. extern void hevc_nvenc_defaults(obs_data_t *settings);
  1152. extern obs_properties_t *hevc_nvenc_properties(void *unused);
  1153. #endif
  1154. static bool nvenc_extra_data(void *data, uint8_t **header, size_t *size)
  1155. {
  1156. struct nvenc_data *enc = data;
  1157. if (!enc->header) {
  1158. return false;
  1159. }
  1160. *header = enc->header;
  1161. *size = enc->header_size;
  1162. return true;
  1163. }
  1164. static bool nvenc_sei_data(void *data, uint8_t **sei, size_t *size)
  1165. {
  1166. struct nvenc_data *enc = data;
  1167. if (!enc->sei) {
  1168. return false;
  1169. }
  1170. *sei = enc->sei;
  1171. *size = enc->sei_size;
  1172. return true;
  1173. }
  1174. struct obs_encoder_info h264_nvenc_info = {
  1175. .id = "jim_nvenc",
  1176. .codec = "h264",
  1177. .type = OBS_ENCODER_VIDEO,
  1178. .caps = OBS_ENCODER_CAP_PASS_TEXTURE | OBS_ENCODER_CAP_DYN_BITRATE,
  1179. .get_name = h264_nvenc_get_name,
  1180. .create = h264_nvenc_create,
  1181. .destroy = nvenc_destroy,
  1182. .update = nvenc_update,
  1183. .encode_texture = nvenc_encode_tex,
  1184. .get_defaults = h264_nvenc_defaults,
  1185. .get_properties = h264_nvenc_properties,
  1186. .get_extra_data = nvenc_extra_data,
  1187. .get_sei_data = nvenc_sei_data,
  1188. };
  1189. #ifdef ENABLE_HEVC
  1190. struct obs_encoder_info hevc_nvenc_info = {
  1191. .id = "jim_hevc_nvenc",
  1192. .codec = "hevc",
  1193. .type = OBS_ENCODER_VIDEO,
  1194. .caps = OBS_ENCODER_CAP_PASS_TEXTURE | OBS_ENCODER_CAP_DYN_BITRATE,
  1195. .get_name = hevc_nvenc_get_name,
  1196. .create = hevc_nvenc_create,
  1197. .destroy = nvenc_destroy,
  1198. .update = nvenc_update,
  1199. .encode_texture = nvenc_encode_tex,
  1200. .get_defaults = hevc_nvenc_defaults,
  1201. .get_properties = hevc_nvenc_properties,
  1202. .get_extra_data = nvenc_extra_data,
  1203. .get_sei_data = nvenc_sei_data,
  1204. };
  1205. #endif