You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

1553 lines
52KB

  1. /*
  2. * H.264 hardware encoding using nvidia nvenc
  3. * Copyright (c) 2014 Timo Rothenpieler <timo@rothenpieler.org>
  4. *
  5. * This file is part of FFmpeg.
  6. *
  7. * FFmpeg is free software; you can redistribute it and/or
  8. * modify it under the terms of the GNU Lesser General Public
  9. * License as published by the Free Software Foundation; either
  10. * version 2.1 of the License, or (at your option) any later version.
  11. *
  12. * FFmpeg is distributed in the hope that it will be useful,
  13. * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15. * Lesser General Public License for more details.
  16. *
  17. * You should have received a copy of the GNU Lesser General Public
  18. * License along with FFmpeg; if not, write to the Free Software
  19. * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20. */
  21. #if defined(_WIN32)
  22. #include <windows.h>
  23. #else
  24. #include <dlfcn.h>
  25. #endif
  26. #include <nvEncodeAPI.h>
  27. #include "libavutil/internal.h"
  28. #include "libavutil/imgutils.h"
  29. #include "libavutil/avassert.h"
  30. #include "libavutil/opt.h"
  31. #include "libavutil/mem.h"
  32. #include "avcodec.h"
  33. #include "internal.h"
  34. #include "thread.h"
  35. #if defined(_WIN32)
  36. #define CUDAAPI __stdcall
  37. #else
  38. #define CUDAAPI
  39. #endif
  40. #if defined(_WIN32)
  41. #define LOAD_FUNC(l, s) GetProcAddress(l, s)
  42. #define DL_CLOSE_FUNC(l) FreeLibrary(l)
  43. #else
  44. #define LOAD_FUNC(l, s) dlsym(l, s)
  45. #define DL_CLOSE_FUNC(l) dlclose(l)
  46. #endif
  47. typedef enum cudaError_enum {
  48. CUDA_SUCCESS = 0
  49. } CUresult;
  50. typedef int CUdevice;
  51. typedef void* CUcontext;
  52. typedef CUresult(CUDAAPI *PCUINIT)(unsigned int Flags);
  53. typedef CUresult(CUDAAPI *PCUDEVICEGETCOUNT)(int *count);
  54. typedef CUresult(CUDAAPI *PCUDEVICEGET)(CUdevice *device, int ordinal);
  55. typedef CUresult(CUDAAPI *PCUDEVICEGETNAME)(char *name, int len, CUdevice dev);
  56. typedef CUresult(CUDAAPI *PCUDEVICECOMPUTECAPABILITY)(int *major, int *minor, CUdevice dev);
  57. typedef CUresult(CUDAAPI *PCUCTXCREATE)(CUcontext *pctx, unsigned int flags, CUdevice dev);
  58. typedef CUresult(CUDAAPI *PCUCTXPOPCURRENT)(CUcontext *pctx);
  59. typedef CUresult(CUDAAPI *PCUCTXDESTROY)(CUcontext ctx);
  60. typedef NVENCSTATUS (NVENCAPI* PNVENCODEAPICREATEINSTANCE)(NV_ENCODE_API_FUNCTION_LIST *functionList);
  61. typedef struct NvencInputSurface
  62. {
  63. NV_ENC_INPUT_PTR input_surface;
  64. int width;
  65. int height;
  66. int lockCount;
  67. NV_ENC_BUFFER_FORMAT format;
  68. } NvencInputSurface;
  69. typedef struct NvencOutputSurface
  70. {
  71. NV_ENC_OUTPUT_PTR output_surface;
  72. int size;
  73. NvencInputSurface* input_surface;
  74. int busy;
  75. } NvencOutputSurface;
  76. typedef struct NvencData
  77. {
  78. union {
  79. int64_t timestamp;
  80. NvencOutputSurface *surface;
  81. } u;
  82. } NvencData;
  83. typedef struct NvencDataList
  84. {
  85. NvencData* data;
  86. uint32_t pos;
  87. uint32_t count;
  88. uint32_t size;
  89. } NvencDataList;
  90. typedef struct NvencDynLoadFunctions
  91. {
  92. PCUINIT cu_init;
  93. PCUDEVICEGETCOUNT cu_device_get_count;
  94. PCUDEVICEGET cu_device_get;
  95. PCUDEVICEGETNAME cu_device_get_name;
  96. PCUDEVICECOMPUTECAPABILITY cu_device_compute_capability;
  97. PCUCTXCREATE cu_ctx_create;
  98. PCUCTXPOPCURRENT cu_ctx_pop_current;
  99. PCUCTXDESTROY cu_ctx_destroy;
  100. NV_ENCODE_API_FUNCTION_LIST nvenc_funcs;
  101. int nvenc_device_count;
  102. CUdevice nvenc_devices[16];
  103. #if defined(_WIN32)
  104. HMODULE cuda_lib;
  105. HMODULE nvenc_lib;
  106. #else
  107. void* cuda_lib;
  108. void* nvenc_lib;
  109. #endif
  110. } NvencDynLoadFunctions;
  111. typedef struct NvencValuePair
  112. {
  113. const char *str;
  114. uint32_t num;
  115. } NvencValuePair;
  116. typedef struct NvencContext
  117. {
  118. AVClass *avclass;
  119. NvencDynLoadFunctions nvenc_dload_funcs;
  120. NV_ENC_INITIALIZE_PARAMS init_encode_params;
  121. NV_ENC_CONFIG encode_config;
  122. CUcontext cu_context;
  123. int max_surface_count;
  124. NvencInputSurface *input_surfaces;
  125. NvencOutputSurface *output_surfaces;
  126. NvencDataList output_surface_queue;
  127. NvencDataList output_surface_ready_queue;
  128. NvencDataList timestamp_list;
  129. int64_t last_dts;
  130. void *nvencoder;
  131. char *preset;
  132. char *profile;
  133. char *level;
  134. char *tier;
  135. int cbr;
  136. int twopass;
  137. int gpu;
  138. int buffer_delay;
  139. } NvencContext;
  140. static const NvencValuePair nvenc_h264_level_pairs[] = {
  141. { "auto", NV_ENC_LEVEL_AUTOSELECT },
  142. { "1" , NV_ENC_LEVEL_H264_1 },
  143. { "1.0" , NV_ENC_LEVEL_H264_1 },
  144. { "1b" , NV_ENC_LEVEL_H264_1b },
  145. { "1.0b", NV_ENC_LEVEL_H264_1b },
  146. { "1.1" , NV_ENC_LEVEL_H264_11 },
  147. { "1.2" , NV_ENC_LEVEL_H264_12 },
  148. { "1.3" , NV_ENC_LEVEL_H264_13 },
  149. { "2" , NV_ENC_LEVEL_H264_2 },
  150. { "2.0" , NV_ENC_LEVEL_H264_2 },
  151. { "2.1" , NV_ENC_LEVEL_H264_21 },
  152. { "2.2" , NV_ENC_LEVEL_H264_22 },
  153. { "3" , NV_ENC_LEVEL_H264_3 },
  154. { "3.0" , NV_ENC_LEVEL_H264_3 },
  155. { "3.1" , NV_ENC_LEVEL_H264_31 },
  156. { "3.2" , NV_ENC_LEVEL_H264_32 },
  157. { "4" , NV_ENC_LEVEL_H264_4 },
  158. { "4.0" , NV_ENC_LEVEL_H264_4 },
  159. { "4.1" , NV_ENC_LEVEL_H264_41 },
  160. { "4.2" , NV_ENC_LEVEL_H264_42 },
  161. { "5" , NV_ENC_LEVEL_H264_5 },
  162. { "5.0" , NV_ENC_LEVEL_H264_5 },
  163. { "5.1" , NV_ENC_LEVEL_H264_51 },
  164. { NULL }
  165. };
  166. static const NvencValuePair nvenc_hevc_level_pairs[] = {
  167. { "auto", NV_ENC_LEVEL_AUTOSELECT },
  168. { "1" , NV_ENC_LEVEL_HEVC_1 },
  169. { "1.0" , NV_ENC_LEVEL_HEVC_1 },
  170. { "2" , NV_ENC_LEVEL_HEVC_2 },
  171. { "2.0" , NV_ENC_LEVEL_HEVC_2 },
  172. { "2.1" , NV_ENC_LEVEL_HEVC_21 },
  173. { "3" , NV_ENC_LEVEL_HEVC_3 },
  174. { "3.0" , NV_ENC_LEVEL_HEVC_3 },
  175. { "3.1" , NV_ENC_LEVEL_HEVC_31 },
  176. { "4" , NV_ENC_LEVEL_HEVC_4 },
  177. { "4.0" , NV_ENC_LEVEL_HEVC_4 },
  178. { "4.1" , NV_ENC_LEVEL_HEVC_41 },
  179. { "5" , NV_ENC_LEVEL_HEVC_5 },
  180. { "5.0" , NV_ENC_LEVEL_HEVC_5 },
  181. { "5.1" , NV_ENC_LEVEL_HEVC_51 },
  182. { "5.2" , NV_ENC_LEVEL_HEVC_52 },
  183. { "6" , NV_ENC_LEVEL_HEVC_6 },
  184. { "6.0" , NV_ENC_LEVEL_HEVC_6 },
  185. { "6.1" , NV_ENC_LEVEL_HEVC_61 },
  186. { "6.2" , NV_ENC_LEVEL_HEVC_62 },
  187. { NULL }
  188. };
  189. static int input_string_to_uint32(AVCodecContext *avctx, const NvencValuePair *pair, const char *input, uint32_t *output)
  190. {
  191. for (; pair->str; ++pair) {
  192. if (!strcmp(input, pair->str)) {
  193. *output = pair->num;
  194. return 0;
  195. }
  196. }
  197. return AVERROR(EINVAL);
  198. }
  199. static NvencData* data_queue_dequeue(NvencDataList* queue)
  200. {
  201. uint32_t mask;
  202. uint32_t read_pos;
  203. av_assert0(queue);
  204. av_assert0(queue->size);
  205. av_assert0(queue->data);
  206. if (!queue->count)
  207. return NULL;
  208. /* Size always is a multiple of two */
  209. mask = queue->size - 1;
  210. read_pos = (queue->pos - queue->count) & mask;
  211. queue->count--;
  212. return &queue->data[read_pos];
  213. }
  214. static int data_queue_enqueue(NvencDataList* queue, NvencData *data)
  215. {
  216. NvencDataList new_queue;
  217. NvencData* tmp_data;
  218. uint32_t mask;
  219. if (!queue->size) {
  220. /* size always has to be a multiple of two */
  221. queue->size = 4;
  222. queue->pos = 0;
  223. queue->count = 0;
  224. queue->data = av_malloc(queue->size * sizeof(*(queue->data)));
  225. if (!queue->data) {
  226. queue->size = 0;
  227. return AVERROR(ENOMEM);
  228. }
  229. }
  230. if (queue->count == queue->size) {
  231. new_queue.size = queue->size << 1;
  232. new_queue.pos = 0;
  233. new_queue.count = 0;
  234. new_queue.data = av_malloc(new_queue.size * sizeof(*(queue->data)));
  235. if (!new_queue.data)
  236. return AVERROR(ENOMEM);
  237. while (tmp_data = data_queue_dequeue(queue))
  238. data_queue_enqueue(&new_queue, tmp_data);
  239. av_free(queue->data);
  240. *queue = new_queue;
  241. }
  242. mask = queue->size - 1;
  243. queue->data[queue->pos] = *data;
  244. queue->pos = (queue->pos + 1) & mask;
  245. queue->count++;
  246. return 0;
  247. }
  248. static int out_surf_queue_enqueue(NvencDataList* queue, NvencOutputSurface* surface)
  249. {
  250. NvencData data;
  251. data.u.surface = surface;
  252. return data_queue_enqueue(queue, &data);
  253. }
  254. static NvencOutputSurface* out_surf_queue_dequeue(NvencDataList* queue)
  255. {
  256. NvencData* res = data_queue_dequeue(queue);
  257. if (!res)
  258. return NULL;
  259. return res->u.surface;
  260. }
  261. static int timestamp_queue_enqueue(NvencDataList* queue, int64_t timestamp)
  262. {
  263. NvencData data;
  264. data.u.timestamp = timestamp;
  265. return data_queue_enqueue(queue, &data);
  266. }
  267. static int64_t timestamp_queue_dequeue(NvencDataList* queue)
  268. {
  269. NvencData* res = data_queue_dequeue(queue);
  270. if (!res)
  271. return AV_NOPTS_VALUE;
  272. return res->u.timestamp;
  273. }
  274. #define CHECK_LOAD_FUNC(t, f, s) \
  275. do { \
  276. (f) = (t)LOAD_FUNC(dl_fn->cuda_lib, s); \
  277. if (!(f)) { \
  278. av_log(avctx, AV_LOG_FATAL, "Failed loading %s from CUDA library\n", s); \
  279. goto error; \
  280. } \
  281. } while (0)
  282. static av_cold int nvenc_dyload_cuda(AVCodecContext *avctx)
  283. {
  284. NvencContext *ctx = avctx->priv_data;
  285. NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
  286. if (dl_fn->cuda_lib)
  287. return 1;
  288. #if defined(_WIN32)
  289. dl_fn->cuda_lib = LoadLibrary(TEXT("nvcuda.dll"));
  290. #else
  291. dl_fn->cuda_lib = dlopen("libcuda.so", RTLD_LAZY);
  292. #endif
  293. if (!dl_fn->cuda_lib) {
  294. av_log(avctx, AV_LOG_FATAL, "Failed loading CUDA library\n");
  295. goto error;
  296. }
  297. CHECK_LOAD_FUNC(PCUINIT, dl_fn->cu_init, "cuInit");
  298. CHECK_LOAD_FUNC(PCUDEVICEGETCOUNT, dl_fn->cu_device_get_count, "cuDeviceGetCount");
  299. CHECK_LOAD_FUNC(PCUDEVICEGET, dl_fn->cu_device_get, "cuDeviceGet");
  300. CHECK_LOAD_FUNC(PCUDEVICEGETNAME, dl_fn->cu_device_get_name, "cuDeviceGetName");
  301. CHECK_LOAD_FUNC(PCUDEVICECOMPUTECAPABILITY, dl_fn->cu_device_compute_capability, "cuDeviceComputeCapability");
  302. CHECK_LOAD_FUNC(PCUCTXCREATE, dl_fn->cu_ctx_create, "cuCtxCreate_v2");
  303. CHECK_LOAD_FUNC(PCUCTXPOPCURRENT, dl_fn->cu_ctx_pop_current, "cuCtxPopCurrent_v2");
  304. CHECK_LOAD_FUNC(PCUCTXDESTROY, dl_fn->cu_ctx_destroy, "cuCtxDestroy_v2");
  305. return 1;
  306. error:
  307. if (dl_fn->cuda_lib)
  308. DL_CLOSE_FUNC(dl_fn->cuda_lib);
  309. dl_fn->cuda_lib = NULL;
  310. return 0;
  311. }
  312. static av_cold int check_cuda_errors(AVCodecContext *avctx, CUresult err, const char *func)
  313. {
  314. if (err != CUDA_SUCCESS) {
  315. av_log(avctx, AV_LOG_FATAL, ">> %s - failed with error code 0x%x\n", func, err);
  316. return 0;
  317. }
  318. return 1;
  319. }
  320. #define check_cuda_errors(f) if (!check_cuda_errors(avctx, f, #f)) goto error
  321. static av_cold int nvenc_check_cuda(AVCodecContext *avctx)
  322. {
  323. int device_count = 0;
  324. CUdevice cu_device = 0;
  325. char gpu_name[128];
  326. int smminor = 0, smmajor = 0;
  327. int i, smver, target_smver;
  328. NvencContext *ctx = avctx->priv_data;
  329. NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
  330. switch (avctx->codec->id) {
  331. case AV_CODEC_ID_H264:
  332. target_smver = avctx->pix_fmt == AV_PIX_FMT_YUV444P ? 0x52 : 0x30;
  333. break;
  334. case AV_CODEC_ID_H265:
  335. target_smver = 0x52;
  336. break;
  337. default:
  338. av_log(avctx, AV_LOG_FATAL, "Unknown codec name\n");
  339. goto error;
  340. }
  341. if (!nvenc_dyload_cuda(avctx))
  342. return 0;
  343. if (dl_fn->nvenc_device_count > 0)
  344. return 1;
  345. check_cuda_errors(dl_fn->cu_init(0));
  346. check_cuda_errors(dl_fn->cu_device_get_count(&device_count));
  347. if (!device_count) {
  348. av_log(avctx, AV_LOG_FATAL, "No CUDA capable devices found\n");
  349. goto error;
  350. }
  351. av_log(avctx, AV_LOG_VERBOSE, "%d CUDA capable devices found\n", device_count);
  352. dl_fn->nvenc_device_count = 0;
  353. for (i = 0; i < device_count; ++i) {
  354. check_cuda_errors(dl_fn->cu_device_get(&cu_device, i));
  355. check_cuda_errors(dl_fn->cu_device_get_name(gpu_name, sizeof(gpu_name), cu_device));
  356. check_cuda_errors(dl_fn->cu_device_compute_capability(&smmajor, &smminor, cu_device));
  357. smver = (smmajor << 4) | smminor;
  358. av_log(avctx, AV_LOG_VERBOSE, "[ GPU #%d - < %s > has Compute SM %d.%d, NVENC %s ]\n", i, gpu_name, smmajor, smminor, (smver >= target_smver) ? "Available" : "Not Available");
  359. if (smver >= target_smver)
  360. dl_fn->nvenc_devices[dl_fn->nvenc_device_count++] = cu_device;
  361. }
  362. if (!dl_fn->nvenc_device_count) {
  363. av_log(avctx, AV_LOG_FATAL, "No NVENC capable devices found\n");
  364. goto error;
  365. }
  366. return 1;
  367. error:
  368. dl_fn->nvenc_device_count = 0;
  369. return 0;
  370. }
  371. static av_cold int nvenc_dyload_nvenc(AVCodecContext *avctx)
  372. {
  373. PNVENCODEAPICREATEINSTANCE nvEncodeAPICreateInstance = 0;
  374. NVENCSTATUS nvstatus;
  375. NvencContext *ctx = avctx->priv_data;
  376. NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
  377. if (!nvenc_check_cuda(avctx))
  378. return 0;
  379. if (dl_fn->nvenc_lib)
  380. return 1;
  381. #if defined(_WIN32)
  382. if (sizeof(void*) == 8) {
  383. dl_fn->nvenc_lib = LoadLibrary(TEXT("nvEncodeAPI64.dll"));
  384. } else {
  385. dl_fn->nvenc_lib = LoadLibrary(TEXT("nvEncodeAPI.dll"));
  386. }
  387. #else
  388. dl_fn->nvenc_lib = dlopen("libnvidia-encode.so.1", RTLD_LAZY);
  389. #endif
  390. if (!dl_fn->nvenc_lib) {
  391. av_log(avctx, AV_LOG_FATAL, "Failed loading the nvenc library\n");
  392. goto error;
  393. }
  394. nvEncodeAPICreateInstance = (PNVENCODEAPICREATEINSTANCE)LOAD_FUNC(dl_fn->nvenc_lib, "NvEncodeAPICreateInstance");
  395. if (!nvEncodeAPICreateInstance) {
  396. av_log(avctx, AV_LOG_FATAL, "Failed to load nvenc entrypoint\n");
  397. goto error;
  398. }
  399. dl_fn->nvenc_funcs.version = NV_ENCODE_API_FUNCTION_LIST_VER;
  400. nvstatus = nvEncodeAPICreateInstance(&dl_fn->nvenc_funcs);
  401. if (nvstatus != NV_ENC_SUCCESS) {
  402. av_log(avctx, AV_LOG_FATAL, "Failed to create nvenc instance\n");
  403. goto error;
  404. }
  405. av_log(avctx, AV_LOG_VERBOSE, "Nvenc initialized successfully\n");
  406. return 1;
  407. error:
  408. if (dl_fn->nvenc_lib)
  409. DL_CLOSE_FUNC(dl_fn->nvenc_lib);
  410. dl_fn->nvenc_lib = NULL;
  411. return 0;
  412. }
  413. static av_cold void nvenc_unload_nvenc(AVCodecContext *avctx)
  414. {
  415. NvencContext *ctx = avctx->priv_data;
  416. NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
  417. DL_CLOSE_FUNC(dl_fn->nvenc_lib);
  418. dl_fn->nvenc_lib = NULL;
  419. dl_fn->nvenc_device_count = 0;
  420. DL_CLOSE_FUNC(dl_fn->cuda_lib);
  421. dl_fn->cuda_lib = NULL;
  422. dl_fn->cu_init = NULL;
  423. dl_fn->cu_device_get_count = NULL;
  424. dl_fn->cu_device_get = NULL;
  425. dl_fn->cu_device_get_name = NULL;
  426. dl_fn->cu_device_compute_capability = NULL;
  427. dl_fn->cu_ctx_create = NULL;
  428. dl_fn->cu_ctx_pop_current = NULL;
  429. dl_fn->cu_ctx_destroy = NULL;
  430. av_log(avctx, AV_LOG_VERBOSE, "Nvenc unloaded\n");
  431. }
  432. static av_cold int nvenc_encode_init(AVCodecContext *avctx)
  433. {
  434. NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS encode_session_params = { 0 };
  435. NV_ENC_PRESET_CONFIG preset_config = { 0 };
  436. CUcontext cu_context_curr;
  437. CUresult cu_res;
  438. GUID encoder_preset = NV_ENC_PRESET_HQ_GUID;
  439. GUID codec;
  440. NVENCSTATUS nv_status = NV_ENC_SUCCESS;
  441. int surfaceCount = 0;
  442. int i, num_mbs;
  443. int isLL = 0;
  444. int lossless = 0;
  445. int res = 0;
  446. int dw, dh;
  447. int qp_inter_p;
  448. NvencContext *ctx = avctx->priv_data;
  449. NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
  450. NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
  451. if (!nvenc_dyload_nvenc(avctx))
  452. return AVERROR_EXTERNAL;
  453. ctx->last_dts = AV_NOPTS_VALUE;
  454. ctx->encode_config.version = NV_ENC_CONFIG_VER;
  455. ctx->init_encode_params.version = NV_ENC_INITIALIZE_PARAMS_VER;
  456. preset_config.version = NV_ENC_PRESET_CONFIG_VER;
  457. preset_config.presetCfg.version = NV_ENC_CONFIG_VER;
  458. encode_session_params.version = NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER;
  459. encode_session_params.apiVersion = NVENCAPI_VERSION;
  460. if (ctx->gpu >= dl_fn->nvenc_device_count) {
  461. av_log(avctx, AV_LOG_FATAL, "Requested GPU %d, but only %d GPUs are available!\n", ctx->gpu, dl_fn->nvenc_device_count);
  462. res = AVERROR(EINVAL);
  463. goto error;
  464. }
  465. ctx->cu_context = NULL;
  466. cu_res = dl_fn->cu_ctx_create(&ctx->cu_context, 4, dl_fn->nvenc_devices[ctx->gpu]); // CU_CTX_SCHED_BLOCKING_SYNC=4, avoid CPU spins
  467. if (cu_res != CUDA_SUCCESS) {
  468. av_log(avctx, AV_LOG_FATAL, "Failed creating CUDA context for NVENC: 0x%x\n", (int)cu_res);
  469. res = AVERROR_EXTERNAL;
  470. goto error;
  471. }
  472. cu_res = dl_fn->cu_ctx_pop_current(&cu_context_curr);
  473. if (cu_res != CUDA_SUCCESS) {
  474. av_log(avctx, AV_LOG_FATAL, "Failed popping CUDA context: 0x%x\n", (int)cu_res);
  475. res = AVERROR_EXTERNAL;
  476. goto error;
  477. }
  478. encode_session_params.device = ctx->cu_context;
  479. encode_session_params.deviceType = NV_ENC_DEVICE_TYPE_CUDA;
  480. nv_status = p_nvenc->nvEncOpenEncodeSessionEx(&encode_session_params, &ctx->nvencoder);
  481. if (nv_status != NV_ENC_SUCCESS) {
  482. ctx->nvencoder = NULL;
  483. av_log(avctx, AV_LOG_FATAL, "OpenEncodeSessionEx failed: 0x%x\n", (int)nv_status);
  484. res = AVERROR_EXTERNAL;
  485. goto error;
  486. }
  487. if (ctx->preset) {
  488. if (!strcmp(ctx->preset, "slow")) {
  489. encoder_preset = NV_ENC_PRESET_HQ_GUID;
  490. ctx->twopass = 1;
  491. } else if (!strcmp(ctx->preset, "medium")) {
  492. encoder_preset = NV_ENC_PRESET_HQ_GUID;
  493. ctx->twopass = 0;
  494. } else if (!strcmp(ctx->preset, "fast")) {
  495. encoder_preset = NV_ENC_PRESET_HP_GUID;
  496. ctx->twopass = 0;
  497. } else if (!strcmp(ctx->preset, "hq")) {
  498. encoder_preset = NV_ENC_PRESET_HQ_GUID;
  499. } else if (!strcmp(ctx->preset, "hp")) {
  500. encoder_preset = NV_ENC_PRESET_HP_GUID;
  501. } else if (!strcmp(ctx->preset, "bd")) {
  502. encoder_preset = NV_ENC_PRESET_BD_GUID;
  503. } else if (!strcmp(ctx->preset, "ll")) {
  504. encoder_preset = NV_ENC_PRESET_LOW_LATENCY_DEFAULT_GUID;
  505. isLL = 1;
  506. } else if (!strcmp(ctx->preset, "llhp")) {
  507. encoder_preset = NV_ENC_PRESET_LOW_LATENCY_HP_GUID;
  508. isLL = 1;
  509. } else if (!strcmp(ctx->preset, "llhq")) {
  510. encoder_preset = NV_ENC_PRESET_LOW_LATENCY_HQ_GUID;
  511. isLL = 1;
  512. } else if (!strcmp(ctx->preset, "lossless")) {
  513. encoder_preset = NV_ENC_PRESET_LOSSLESS_DEFAULT_GUID;
  514. lossless = 1;
  515. } else if (!strcmp(ctx->preset, "losslesshp")) {
  516. encoder_preset = NV_ENC_PRESET_LOSSLESS_HP_GUID;
  517. lossless = 1;
  518. } else if (!strcmp(ctx->preset, "default")) {
  519. encoder_preset = NV_ENC_PRESET_DEFAULT_GUID;
  520. } else {
  521. av_log(avctx, AV_LOG_FATAL, "Preset \"%s\" is unknown! Supported presets: slow, medium, high, hp, hq, bd, ll, llhp, llhq, lossless, losslesshp, default\n", ctx->preset);
  522. res = AVERROR(EINVAL);
  523. goto error;
  524. }
  525. }
  526. if (ctx->twopass < 0) {
  527. ctx->twopass = isLL;
  528. }
  529. switch (avctx->codec->id) {
  530. case AV_CODEC_ID_H264:
  531. codec = NV_ENC_CODEC_H264_GUID;
  532. break;
  533. case AV_CODEC_ID_H265:
  534. codec = NV_ENC_CODEC_HEVC_GUID;
  535. break;
  536. default:
  537. av_log(avctx, AV_LOG_ERROR, "Unknown codec name\n");
  538. res = AVERROR(EINVAL);
  539. goto error;
  540. }
  541. nv_status = p_nvenc->nvEncGetEncodePresetConfig(ctx->nvencoder, codec, encoder_preset, &preset_config);
  542. if (nv_status != NV_ENC_SUCCESS) {
  543. av_log(avctx, AV_LOG_FATAL, "GetEncodePresetConfig failed: 0x%x\n", (int)nv_status);
  544. res = AVERROR_EXTERNAL;
  545. goto error;
  546. }
  547. ctx->init_encode_params.encodeGUID = codec;
  548. ctx->init_encode_params.encodeHeight = avctx->height;
  549. ctx->init_encode_params.encodeWidth = avctx->width;
  550. if (avctx->sample_aspect_ratio.num && avctx->sample_aspect_ratio.den &&
  551. (avctx->sample_aspect_ratio.num != 1 || avctx->sample_aspect_ratio.num != 1)) {
  552. av_reduce(&dw, &dh,
  553. avctx->width * avctx->sample_aspect_ratio.num,
  554. avctx->height * avctx->sample_aspect_ratio.den,
  555. 1024 * 1024);
  556. ctx->init_encode_params.darHeight = dh;
  557. ctx->init_encode_params.darWidth = dw;
  558. } else {
  559. ctx->init_encode_params.darHeight = avctx->height;
  560. ctx->init_encode_params.darWidth = avctx->width;
  561. }
  562. // De-compensate for hardware, dubiously, trying to compensate for
  563. // playback at 704 pixel width.
  564. if (avctx->width == 720 &&
  565. (avctx->height == 480 || avctx->height == 576)) {
  566. av_reduce(&dw, &dh,
  567. ctx->init_encode_params.darWidth * 44,
  568. ctx->init_encode_params.darHeight * 45,
  569. 1024 * 1024);
  570. ctx->init_encode_params.darHeight = dh;
  571. ctx->init_encode_params.darWidth = dw;
  572. }
  573. ctx->init_encode_params.frameRateNum = avctx->time_base.den;
  574. ctx->init_encode_params.frameRateDen = avctx->time_base.num * avctx->ticks_per_frame;
  575. num_mbs = ((avctx->width + 15) >> 4) * ((avctx->height + 15) >> 4);
  576. ctx->max_surface_count = (num_mbs >= 8160) ? 32 : 48;
  577. if (ctx->buffer_delay >= ctx->max_surface_count)
  578. ctx->buffer_delay = ctx->max_surface_count - 1;
  579. ctx->init_encode_params.enableEncodeAsync = 0;
  580. ctx->init_encode_params.enablePTD = 1;
  581. ctx->init_encode_params.presetGUID = encoder_preset;
  582. ctx->init_encode_params.encodeConfig = &ctx->encode_config;
  583. memcpy(&ctx->encode_config, &preset_config.presetCfg, sizeof(ctx->encode_config));
  584. ctx->encode_config.version = NV_ENC_CONFIG_VER;
  585. if (avctx->refs >= 0) {
  586. /* 0 means "let the hardware decide" */
  587. switch (avctx->codec->id) {
  588. case AV_CODEC_ID_H264:
  589. ctx->encode_config.encodeCodecConfig.h264Config.maxNumRefFrames = avctx->refs;
  590. break;
  591. case AV_CODEC_ID_H265:
  592. ctx->encode_config.encodeCodecConfig.hevcConfig.maxNumRefFramesInDPB = avctx->refs;
  593. break;
  594. /* Earlier switch/case will return if unknown codec is passed. */
  595. }
  596. }
  597. if (avctx->gop_size > 0) {
  598. if (avctx->max_b_frames >= 0) {
  599. /* 0 is intra-only, 1 is I/P only, 2 is one B Frame, 3 two B frames, and so on. */
  600. ctx->encode_config.frameIntervalP = avctx->max_b_frames + 1;
  601. }
  602. ctx->encode_config.gopLength = avctx->gop_size;
  603. switch (avctx->codec->id) {
  604. case AV_CODEC_ID_H264:
  605. ctx->encode_config.encodeCodecConfig.h264Config.idrPeriod = avctx->gop_size;
  606. break;
  607. case AV_CODEC_ID_H265:
  608. ctx->encode_config.encodeCodecConfig.hevcConfig.idrPeriod = avctx->gop_size;
  609. break;
  610. /* Earlier switch/case will return if unknown codec is passed. */
  611. }
  612. } else if (avctx->gop_size == 0) {
  613. ctx->encode_config.frameIntervalP = 0;
  614. ctx->encode_config.gopLength = 1;
  615. switch (avctx->codec->id) {
  616. case AV_CODEC_ID_H264:
  617. ctx->encode_config.encodeCodecConfig.h264Config.idrPeriod = 1;
  618. break;
  619. case AV_CODEC_ID_H265:
  620. ctx->encode_config.encodeCodecConfig.hevcConfig.idrPeriod = 1;
  621. break;
  622. /* Earlier switch/case will return if unknown codec is passed. */
  623. }
  624. }
  625. /* when there're b frames, set dts offset */
  626. if (ctx->encode_config.frameIntervalP >= 2)
  627. ctx->last_dts = -2;
  628. if (avctx->bit_rate > 0)
  629. ctx->encode_config.rcParams.averageBitRate = avctx->bit_rate;
  630. if (avctx->rc_max_rate > 0)
  631. ctx->encode_config.rcParams.maxBitRate = avctx->rc_max_rate;
  632. if (lossless) {
  633. if (avctx->codec->id == AV_CODEC_ID_H264)
  634. ctx->encode_config.encodeCodecConfig.h264Config.qpPrimeYZeroTransformBypassFlag = 1;
  635. ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
  636. ctx->encode_config.rcParams.constQP.qpInterB = 0;
  637. ctx->encode_config.rcParams.constQP.qpInterP = 0;
  638. ctx->encode_config.rcParams.constQP.qpIntra = 0;
  639. avctx->qmin = -1;
  640. avctx->qmax = -1;
  641. } else if (ctx->cbr) {
  642. if (!ctx->twopass) {
  643. ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CBR;
  644. } else {
  645. ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_2_PASS_QUALITY;
  646. if (avctx->codec->id == AV_CODEC_ID_H264) {
  647. ctx->encode_config.encodeCodecConfig.h264Config.adaptiveTransformMode = NV_ENC_H264_ADAPTIVE_TRANSFORM_ENABLE;
  648. ctx->encode_config.encodeCodecConfig.h264Config.fmoMode = NV_ENC_H264_FMO_DISABLE;
  649. }
  650. }
  651. } else if (avctx->global_quality > 0) {
  652. ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
  653. ctx->encode_config.rcParams.constQP.qpInterB = avctx->global_quality;
  654. ctx->encode_config.rcParams.constQP.qpInterP = avctx->global_quality;
  655. ctx->encode_config.rcParams.constQP.qpIntra = avctx->global_quality;
  656. avctx->qmin = -1;
  657. avctx->qmax = -1;
  658. } else {
  659. if (avctx->qmin >= 0 && avctx->qmax >= 0) {
  660. ctx->encode_config.rcParams.enableMinQP = 1;
  661. ctx->encode_config.rcParams.enableMaxQP = 1;
  662. ctx->encode_config.rcParams.minQP.qpInterB = avctx->qmin;
  663. ctx->encode_config.rcParams.minQP.qpInterP = avctx->qmin;
  664. ctx->encode_config.rcParams.minQP.qpIntra = avctx->qmin;
  665. ctx->encode_config.rcParams.maxQP.qpInterB = avctx->qmax;
  666. ctx->encode_config.rcParams.maxQP.qpInterP = avctx->qmax;
  667. ctx->encode_config.rcParams.maxQP.qpIntra = avctx->qmax;
  668. qp_inter_p = (avctx->qmax + 3 * avctx->qmin) / 4; // biased towards Qmin
  669. if (ctx->twopass) {
  670. ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_2_PASS_VBR;
  671. if (avctx->codec->id == AV_CODEC_ID_H264) {
  672. ctx->encode_config.encodeCodecConfig.h264Config.adaptiveTransformMode = NV_ENC_H264_ADAPTIVE_TRANSFORM_ENABLE;
  673. ctx->encode_config.encodeCodecConfig.h264Config.fmoMode = NV_ENC_H264_FMO_DISABLE;
  674. }
  675. } else {
  676. ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_VBR_MINQP;
  677. }
  678. } else {
  679. qp_inter_p = 26; // default to 26
  680. if (ctx->twopass) {
  681. ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_2_PASS_VBR;
  682. } else {
  683. ctx->encode_config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_VBR;
  684. }
  685. }
  686. ctx->encode_config.rcParams.enableInitialRCQP = 1;
  687. ctx->encode_config.rcParams.initialRCQP.qpInterP = qp_inter_p;
  688. if(avctx->i_quant_factor != 0.0 && avctx->b_quant_factor != 0.0) {
  689. ctx->encode_config.rcParams.initialRCQP.qpIntra = qp_inter_p * fabs(avctx->i_quant_factor);
  690. ctx->encode_config.rcParams.initialRCQP.qpIntra += avctx->i_quant_offset;
  691. ctx->encode_config.rcParams.initialRCQP.qpInterB = qp_inter_p * fabs(avctx->b_quant_factor);
  692. ctx->encode_config.rcParams.initialRCQP.qpInterB += avctx->b_quant_offset;
  693. } else {
  694. ctx->encode_config.rcParams.initialRCQP.qpIntra = qp_inter_p;
  695. ctx->encode_config.rcParams.initialRCQP.qpInterB = qp_inter_p;
  696. }
  697. }
  698. if (avctx->rc_buffer_size > 0)
  699. ctx->encode_config.rcParams.vbvBufferSize = avctx->rc_buffer_size;
  700. if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
  701. ctx->encode_config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FIELD;
  702. } else {
  703. ctx->encode_config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FRAME;
  704. }
  705. switch (avctx->codec->id) {
  706. case AV_CODEC_ID_H264:
  707. ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourDescriptionPresentFlag = 1;
  708. ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.videoSignalTypePresentFlag = 1;
  709. ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourMatrix = avctx->colorspace;
  710. ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.colourPrimaries = avctx->color_primaries;
  711. ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.transferCharacteristics = avctx->color_trc;
  712. ctx->encode_config.encodeCodecConfig.h264Config.h264VUIParameters.videoFullRangeFlag = avctx->color_range == AVCOL_RANGE_JPEG;
  713. ctx->encode_config.encodeCodecConfig.h264Config.disableSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
  714. ctx->encode_config.encodeCodecConfig.h264Config.repeatSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 0 : 1;
  715. if (!ctx->profile) {
  716. switch (avctx->profile) {
  717. case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
  718. ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_444_GUID;
  719. break;
  720. case FF_PROFILE_H264_BASELINE:
  721. ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_BASELINE_GUID;
  722. break;
  723. case FF_PROFILE_H264_MAIN:
  724. ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_MAIN_GUID;
  725. break;
  726. case FF_PROFILE_H264_HIGH:
  727. case FF_PROFILE_UNKNOWN:
  728. ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
  729. break;
  730. default:
  731. av_log(avctx, AV_LOG_WARNING, "Unsupported profile requested, falling back to high\n");
  732. ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
  733. break;
  734. }
  735. } else {
  736. if (!strcmp(ctx->profile, "high")) {
  737. ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_GUID;
  738. avctx->profile = FF_PROFILE_H264_HIGH;
  739. } else if (!strcmp(ctx->profile, "main")) {
  740. ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_MAIN_GUID;
  741. avctx->profile = FF_PROFILE_H264_MAIN;
  742. } else if (!strcmp(ctx->profile, "baseline")) {
  743. ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_BASELINE_GUID;
  744. avctx->profile = FF_PROFILE_H264_BASELINE;
  745. } else if (!strcmp(ctx->profile, "high444p")) {
  746. ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_444_GUID;
  747. avctx->profile = FF_PROFILE_H264_HIGH_444_PREDICTIVE;
  748. } else {
  749. av_log(avctx, AV_LOG_FATAL, "Profile \"%s\" is unknown! Supported profiles: high, main, baseline\n", ctx->profile);
  750. res = AVERROR(EINVAL);
  751. goto error;
  752. }
  753. }
  754. // force setting profile as high444p if input is AV_PIX_FMT_YUV444P
  755. if (avctx->pix_fmt == AV_PIX_FMT_YUV444P) {
  756. ctx->encode_config.profileGUID = NV_ENC_H264_PROFILE_HIGH_444_GUID;
  757. avctx->profile = FF_PROFILE_H264_HIGH_444_PREDICTIVE;
  758. }
  759. ctx->encode_config.encodeCodecConfig.h264Config.chromaFormatIDC = avctx->profile == FF_PROFILE_H264_HIGH_444_PREDICTIVE ? 3 : 1;
  760. if (ctx->level) {
  761. res = input_string_to_uint32(avctx, nvenc_h264_level_pairs, ctx->level, &ctx->encode_config.encodeCodecConfig.h264Config.level);
  762. if (res) {
  763. av_log(avctx, AV_LOG_FATAL, "Level \"%s\" is unknown! Supported levels: auto, 1, 1b, 1.1, 1.2, 1.3, 2, 2.1, 2.2, 3, 3.1, 3.2, 4, 4.1, 4.2, 5, 5.1\n", ctx->level);
  764. goto error;
  765. }
  766. } else {
  767. ctx->encode_config.encodeCodecConfig.h264Config.level = NV_ENC_LEVEL_AUTOSELECT;
  768. }
  769. break;
  770. case AV_CODEC_ID_H265:
  771. ctx->encode_config.encodeCodecConfig.hevcConfig.disableSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
  772. ctx->encode_config.encodeCodecConfig.hevcConfig.repeatSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 0 : 1;
  773. /* No other profile is supported in the current SDK version 5 */
  774. ctx->encode_config.profileGUID = NV_ENC_HEVC_PROFILE_MAIN_GUID;
  775. avctx->profile = FF_PROFILE_HEVC_MAIN;
  776. if (ctx->level) {
  777. res = input_string_to_uint32(avctx, nvenc_hevc_level_pairs, ctx->level, &ctx->encode_config.encodeCodecConfig.hevcConfig.level);
  778. if (res) {
  779. av_log(avctx, AV_LOG_FATAL, "Level \"%s\" is unknown! Supported levels: auto, 1, 2, 2.1, 3, 3.1, 4, 4.1, 5, 5.1, 5.2, 6, 6.1, 6.2\n", ctx->level);
  780. goto error;
  781. }
  782. } else {
  783. ctx->encode_config.encodeCodecConfig.hevcConfig.level = NV_ENC_LEVEL_AUTOSELECT;
  784. }
  785. if (ctx->tier) {
  786. if (!strcmp(ctx->tier, "main")) {
  787. ctx->encode_config.encodeCodecConfig.hevcConfig.tier = NV_ENC_TIER_HEVC_MAIN;
  788. } else if (!strcmp(ctx->tier, "high")) {
  789. ctx->encode_config.encodeCodecConfig.hevcConfig.tier = NV_ENC_TIER_HEVC_HIGH;
  790. } else {
  791. av_log(avctx, AV_LOG_FATAL, "Tier \"%s\" is unknown! Supported tiers: main, high\n", ctx->tier);
  792. res = AVERROR(EINVAL);
  793. goto error;
  794. }
  795. }
  796. break;
  797. /* Earlier switch/case will return if unknown codec is passed. */
  798. }
  799. nv_status = p_nvenc->nvEncInitializeEncoder(ctx->nvencoder, &ctx->init_encode_params);
  800. if (nv_status != NV_ENC_SUCCESS) {
  801. av_log(avctx, AV_LOG_FATAL, "InitializeEncoder failed: 0x%x\n", (int)nv_status);
  802. res = AVERROR_EXTERNAL;
  803. goto error;
  804. }
  805. ctx->input_surfaces = av_malloc(ctx->max_surface_count * sizeof(*ctx->input_surfaces));
  806. if (!ctx->input_surfaces) {
  807. res = AVERROR(ENOMEM);
  808. goto error;
  809. }
  810. ctx->output_surfaces = av_malloc(ctx->max_surface_count * sizeof(*ctx->output_surfaces));
  811. if (!ctx->output_surfaces) {
  812. res = AVERROR(ENOMEM);
  813. goto error;
  814. }
  815. for (surfaceCount = 0; surfaceCount < ctx->max_surface_count; ++surfaceCount) {
  816. NV_ENC_CREATE_INPUT_BUFFER allocSurf = { 0 };
  817. NV_ENC_CREATE_BITSTREAM_BUFFER allocOut = { 0 };
  818. allocSurf.version = NV_ENC_CREATE_INPUT_BUFFER_VER;
  819. allocOut.version = NV_ENC_CREATE_BITSTREAM_BUFFER_VER;
  820. allocSurf.width = (avctx->width + 31) & ~31;
  821. allocSurf.height = (avctx->height + 31) & ~31;
  822. allocSurf.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_CACHED;
  823. switch (avctx->pix_fmt) {
  824. case AV_PIX_FMT_YUV420P:
  825. allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_YV12_PL;
  826. break;
  827. case AV_PIX_FMT_NV12:
  828. allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_NV12_PL;
  829. break;
  830. case AV_PIX_FMT_YUV444P:
  831. allocSurf.bufferFmt = NV_ENC_BUFFER_FORMAT_YUV444_PL;
  832. break;
  833. default:
  834. av_log(avctx, AV_LOG_FATAL, "Invalid input pixel format\n");
  835. res = AVERROR(EINVAL);
  836. goto error;
  837. }
  838. nv_status = p_nvenc->nvEncCreateInputBuffer(ctx->nvencoder, &allocSurf);
  839. if (nv_status != NV_ENC_SUCCESS) {
  840. av_log(avctx, AV_LOG_FATAL, "CreateInputBuffer failed\n");
  841. res = AVERROR_EXTERNAL;
  842. goto error;
  843. }
  844. ctx->input_surfaces[surfaceCount].lockCount = 0;
  845. ctx->input_surfaces[surfaceCount].input_surface = allocSurf.inputBuffer;
  846. ctx->input_surfaces[surfaceCount].format = allocSurf.bufferFmt;
  847. ctx->input_surfaces[surfaceCount].width = allocSurf.width;
  848. ctx->input_surfaces[surfaceCount].height = allocSurf.height;
  849. /* 1MB is large enough to hold most output frames. NVENC increases this automaticaly if it's not enough. */
  850. allocOut.size = 1024 * 1024;
  851. allocOut.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_CACHED;
  852. nv_status = p_nvenc->nvEncCreateBitstreamBuffer(ctx->nvencoder, &allocOut);
  853. if (nv_status != NV_ENC_SUCCESS) {
  854. av_log(avctx, AV_LOG_FATAL, "CreateBitstreamBuffer failed\n");
  855. ctx->output_surfaces[surfaceCount++].output_surface = NULL;
  856. res = AVERROR_EXTERNAL;
  857. goto error;
  858. }
  859. ctx->output_surfaces[surfaceCount].output_surface = allocOut.bitstreamBuffer;
  860. ctx->output_surfaces[surfaceCount].size = allocOut.size;
  861. ctx->output_surfaces[surfaceCount].busy = 0;
  862. }
  863. if (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) {
  864. uint32_t outSize = 0;
  865. char tmpHeader[256];
  866. NV_ENC_SEQUENCE_PARAM_PAYLOAD payload = { 0 };
  867. payload.version = NV_ENC_SEQUENCE_PARAM_PAYLOAD_VER;
  868. payload.spsppsBuffer = tmpHeader;
  869. payload.inBufferSize = sizeof(tmpHeader);
  870. payload.outSPSPPSPayloadSize = &outSize;
  871. nv_status = p_nvenc->nvEncGetSequenceParams(ctx->nvencoder, &payload);
  872. if (nv_status != NV_ENC_SUCCESS) {
  873. av_log(avctx, AV_LOG_FATAL, "GetSequenceParams failed\n");
  874. goto error;
  875. }
  876. avctx->extradata_size = outSize;
  877. avctx->extradata = av_mallocz(outSize + AV_INPUT_BUFFER_PADDING_SIZE);
  878. if (!avctx->extradata) {
  879. res = AVERROR(ENOMEM);
  880. goto error;
  881. }
  882. memcpy(avctx->extradata, tmpHeader, outSize);
  883. }
  884. if (ctx->encode_config.frameIntervalP > 1)
  885. avctx->has_b_frames = 2;
  886. if (ctx->encode_config.rcParams.averageBitRate > 0)
  887. avctx->bit_rate = ctx->encode_config.rcParams.averageBitRate;
  888. return 0;
  889. error:
  890. for (i = 0; i < surfaceCount; ++i) {
  891. p_nvenc->nvEncDestroyInputBuffer(ctx->nvencoder, ctx->input_surfaces[i].input_surface);
  892. if (ctx->output_surfaces[i].output_surface)
  893. p_nvenc->nvEncDestroyBitstreamBuffer(ctx->nvencoder, ctx->output_surfaces[i].output_surface);
  894. }
  895. if (ctx->nvencoder)
  896. p_nvenc->nvEncDestroyEncoder(ctx->nvencoder);
  897. if (ctx->cu_context)
  898. dl_fn->cu_ctx_destroy(ctx->cu_context);
  899. nvenc_unload_nvenc(avctx);
  900. ctx->nvencoder = NULL;
  901. ctx->cu_context = NULL;
  902. return res;
  903. }
  904. static av_cold int nvenc_encode_close(AVCodecContext *avctx)
  905. {
  906. NvencContext *ctx = avctx->priv_data;
  907. NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
  908. NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
  909. int i;
  910. av_freep(&ctx->timestamp_list.data);
  911. av_freep(&ctx->output_surface_ready_queue.data);
  912. av_freep(&ctx->output_surface_queue.data);
  913. for (i = 0; i < ctx->max_surface_count; ++i) {
  914. p_nvenc->nvEncDestroyInputBuffer(ctx->nvencoder, ctx->input_surfaces[i].input_surface);
  915. p_nvenc->nvEncDestroyBitstreamBuffer(ctx->nvencoder, ctx->output_surfaces[i].output_surface);
  916. }
  917. ctx->max_surface_count = 0;
  918. p_nvenc->nvEncDestroyEncoder(ctx->nvencoder);
  919. ctx->nvencoder = NULL;
  920. dl_fn->cu_ctx_destroy(ctx->cu_context);
  921. ctx->cu_context = NULL;
  922. nvenc_unload_nvenc(avctx);
  923. return 0;
  924. }
  925. static int process_output_surface(AVCodecContext *avctx, AVPacket *pkt, NvencOutputSurface *tmpoutsurf)
  926. {
  927. NvencContext *ctx = avctx->priv_data;
  928. NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
  929. NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
  930. uint32_t slice_mode_data;
  931. uint32_t *slice_offsets;
  932. NV_ENC_LOCK_BITSTREAM lock_params = { 0 };
  933. NVENCSTATUS nv_status;
  934. int res = 0;
  935. switch (avctx->codec->id) {
  936. case AV_CODEC_ID_H264:
  937. slice_mode_data = ctx->encode_config.encodeCodecConfig.h264Config.sliceModeData;
  938. break;
  939. case AV_CODEC_ID_H265:
  940. slice_mode_data = ctx->encode_config.encodeCodecConfig.hevcConfig.sliceModeData;
  941. break;
  942. default:
  943. av_log(avctx, AV_LOG_ERROR, "Unknown codec name\n");
  944. res = AVERROR(EINVAL);
  945. goto error;
  946. }
  947. slice_offsets = av_mallocz(slice_mode_data * sizeof(*slice_offsets));
  948. if (!slice_offsets)
  949. return AVERROR(ENOMEM);
  950. lock_params.version = NV_ENC_LOCK_BITSTREAM_VER;
  951. lock_params.doNotWait = 0;
  952. lock_params.outputBitstream = tmpoutsurf->output_surface;
  953. lock_params.sliceOffsets = slice_offsets;
  954. nv_status = p_nvenc->nvEncLockBitstream(ctx->nvencoder, &lock_params);
  955. if (nv_status != NV_ENC_SUCCESS) {
  956. av_log(avctx, AV_LOG_ERROR, "Failed locking bitstream buffer\n");
  957. res = AVERROR_EXTERNAL;
  958. goto error;
  959. }
  960. if (res = ff_alloc_packet2(avctx, pkt, lock_params.bitstreamSizeInBytes,0)) {
  961. p_nvenc->nvEncUnlockBitstream(ctx->nvencoder, tmpoutsurf->output_surface);
  962. goto error;
  963. }
  964. memcpy(pkt->data, lock_params.bitstreamBufferPtr, lock_params.bitstreamSizeInBytes);
  965. nv_status = p_nvenc->nvEncUnlockBitstream(ctx->nvencoder, tmpoutsurf->output_surface);
  966. if (nv_status != NV_ENC_SUCCESS)
  967. av_log(avctx, AV_LOG_ERROR, "Failed unlocking bitstream buffer, expect the gates of mordor to open\n");
  968. switch (lock_params.pictureType) {
  969. case NV_ENC_PIC_TYPE_IDR:
  970. pkt->flags |= AV_PKT_FLAG_KEY;
  971. #if FF_API_CODED_FRAME
  972. FF_DISABLE_DEPRECATION_WARNINGS
  973. case NV_ENC_PIC_TYPE_I:
  974. avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
  975. break;
  976. case NV_ENC_PIC_TYPE_P:
  977. avctx->coded_frame->pict_type = AV_PICTURE_TYPE_P;
  978. break;
  979. case NV_ENC_PIC_TYPE_B:
  980. avctx->coded_frame->pict_type = AV_PICTURE_TYPE_B;
  981. break;
  982. case NV_ENC_PIC_TYPE_BI:
  983. avctx->coded_frame->pict_type = AV_PICTURE_TYPE_BI;
  984. break;
  985. default:
  986. av_log(avctx, AV_LOG_ERROR, "Unknown picture type encountered, expect the output to be broken.\n");
  987. av_log(avctx, AV_LOG_ERROR, "Please report this error and include as much information on how to reproduce it as possible.\n");
  988. res = AVERROR_EXTERNAL;
  989. goto error;
  990. FF_ENABLE_DEPRECATION_WARNINGS
  991. #endif
  992. }
  993. pkt->pts = lock_params.outputTimeStamp;
  994. pkt->dts = timestamp_queue_dequeue(&ctx->timestamp_list);
  995. /* when there're b frame(s), set dts offset */
  996. if (ctx->encode_config.frameIntervalP >= 2)
  997. pkt->dts -= 1;
  998. if (pkt->dts > pkt->pts)
  999. pkt->dts = pkt->pts;
  1000. if (ctx->last_dts != AV_NOPTS_VALUE && pkt->dts <= ctx->last_dts)
  1001. pkt->dts = ctx->last_dts + 1;
  1002. ctx->last_dts = pkt->dts;
  1003. av_free(slice_offsets);
  1004. return 0;
  1005. error:
  1006. av_free(slice_offsets);
  1007. timestamp_queue_dequeue(&ctx->timestamp_list);
  1008. return res;
  1009. }
  1010. static int nvenc_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
  1011. const AVFrame *frame, int *got_packet)
  1012. {
  1013. NVENCSTATUS nv_status;
  1014. NvencOutputSurface *tmpoutsurf;
  1015. int res, i = 0;
  1016. NvencContext *ctx = avctx->priv_data;
  1017. NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
  1018. NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
  1019. NV_ENC_PIC_PARAMS pic_params = { 0 };
  1020. pic_params.version = NV_ENC_PIC_PARAMS_VER;
  1021. if (frame) {
  1022. NV_ENC_LOCK_INPUT_BUFFER lockBufferParams = { 0 };
  1023. NvencInputSurface *inSurf = NULL;
  1024. for (i = 0; i < ctx->max_surface_count; ++i) {
  1025. if (!ctx->input_surfaces[i].lockCount) {
  1026. inSurf = &ctx->input_surfaces[i];
  1027. break;
  1028. }
  1029. }
  1030. av_assert0(inSurf);
  1031. inSurf->lockCount = 1;
  1032. lockBufferParams.version = NV_ENC_LOCK_INPUT_BUFFER_VER;
  1033. lockBufferParams.inputBuffer = inSurf->input_surface;
  1034. nv_status = p_nvenc->nvEncLockInputBuffer(ctx->nvencoder, &lockBufferParams);
  1035. if (nv_status != NV_ENC_SUCCESS) {
  1036. av_log(avctx, AV_LOG_ERROR, "Failed locking nvenc input buffer\n");
  1037. return 0;
  1038. }
  1039. if (avctx->pix_fmt == AV_PIX_FMT_YUV420P) {
  1040. uint8_t *buf = lockBufferParams.bufferDataPtr;
  1041. av_image_copy_plane(buf, lockBufferParams.pitch,
  1042. frame->data[0], frame->linesize[0],
  1043. avctx->width, avctx->height);
  1044. buf += inSurf->height * lockBufferParams.pitch;
  1045. av_image_copy_plane(buf, lockBufferParams.pitch >> 1,
  1046. frame->data[2], frame->linesize[2],
  1047. avctx->width >> 1, avctx->height >> 1);
  1048. buf += (inSurf->height * lockBufferParams.pitch) >> 2;
  1049. av_image_copy_plane(buf, lockBufferParams.pitch >> 1,
  1050. frame->data[1], frame->linesize[1],
  1051. avctx->width >> 1, avctx->height >> 1);
  1052. } else if (avctx->pix_fmt == AV_PIX_FMT_NV12) {
  1053. uint8_t *buf = lockBufferParams.bufferDataPtr;
  1054. av_image_copy_plane(buf, lockBufferParams.pitch,
  1055. frame->data[0], frame->linesize[0],
  1056. avctx->width, avctx->height);
  1057. buf += inSurf->height * lockBufferParams.pitch;
  1058. av_image_copy_plane(buf, lockBufferParams.pitch,
  1059. frame->data[1], frame->linesize[1],
  1060. avctx->width, avctx->height >> 1);
  1061. } else if (avctx->pix_fmt == AV_PIX_FMT_YUV444P) {
  1062. uint8_t *buf = lockBufferParams.bufferDataPtr;
  1063. av_image_copy_plane(buf, lockBufferParams.pitch,
  1064. frame->data[0], frame->linesize[0],
  1065. avctx->width, avctx->height);
  1066. buf += inSurf->height * lockBufferParams.pitch;
  1067. av_image_copy_plane(buf, lockBufferParams.pitch,
  1068. frame->data[1], frame->linesize[1],
  1069. avctx->width, avctx->height);
  1070. buf += inSurf->height * lockBufferParams.pitch;
  1071. av_image_copy_plane(buf, lockBufferParams.pitch,
  1072. frame->data[2], frame->linesize[2],
  1073. avctx->width, avctx->height);
  1074. } else {
  1075. av_log(avctx, AV_LOG_FATAL, "Invalid pixel format!\n");
  1076. return AVERROR(EINVAL);
  1077. }
  1078. nv_status = p_nvenc->nvEncUnlockInputBuffer(ctx->nvencoder, inSurf->input_surface);
  1079. if (nv_status != NV_ENC_SUCCESS) {
  1080. av_log(avctx, AV_LOG_FATAL, "Failed unlocking input buffer!\n");
  1081. return AVERROR_EXTERNAL;
  1082. }
  1083. for (i = 0; i < ctx->max_surface_count; ++i)
  1084. if (!ctx->output_surfaces[i].busy)
  1085. break;
  1086. if (i == ctx->max_surface_count) {
  1087. inSurf->lockCount = 0;
  1088. av_log(avctx, AV_LOG_FATAL, "No free output surface found!\n");
  1089. return AVERROR_EXTERNAL;
  1090. }
  1091. ctx->output_surfaces[i].input_surface = inSurf;
  1092. pic_params.inputBuffer = inSurf->input_surface;
  1093. pic_params.bufferFmt = inSurf->format;
  1094. pic_params.inputWidth = avctx->width;
  1095. pic_params.inputHeight = avctx->height;
  1096. pic_params.outputBitstream = ctx->output_surfaces[i].output_surface;
  1097. pic_params.completionEvent = 0;
  1098. if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
  1099. if (frame->top_field_first) {
  1100. pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_TOP_BOTTOM;
  1101. } else {
  1102. pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FIELD_BOTTOM_TOP;
  1103. }
  1104. } else {
  1105. pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
  1106. }
  1107. pic_params.encodePicFlags = 0;
  1108. pic_params.inputTimeStamp = frame->pts;
  1109. pic_params.inputDuration = 0;
  1110. switch (avctx->codec->id) {
  1111. case AV_CODEC_ID_H264:
  1112. pic_params.codecPicParams.h264PicParams.sliceMode = ctx->encode_config.encodeCodecConfig.h264Config.sliceMode;
  1113. pic_params.codecPicParams.h264PicParams.sliceModeData = ctx->encode_config.encodeCodecConfig.h264Config.sliceModeData;
  1114. break;
  1115. case AV_CODEC_ID_H265:
  1116. pic_params.codecPicParams.hevcPicParams.sliceMode = ctx->encode_config.encodeCodecConfig.hevcConfig.sliceMode;
  1117. pic_params.codecPicParams.hevcPicParams.sliceModeData = ctx->encode_config.encodeCodecConfig.hevcConfig.sliceModeData;
  1118. break;
  1119. default:
  1120. av_log(avctx, AV_LOG_ERROR, "Unknown codec name\n");
  1121. return AVERROR(EINVAL);
  1122. }
  1123. res = timestamp_queue_enqueue(&ctx->timestamp_list, frame->pts);
  1124. if (res)
  1125. return res;
  1126. } else {
  1127. pic_params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
  1128. }
  1129. nv_status = p_nvenc->nvEncEncodePicture(ctx->nvencoder, &pic_params);
  1130. if (frame && nv_status == NV_ENC_ERR_NEED_MORE_INPUT) {
  1131. res = out_surf_queue_enqueue(&ctx->output_surface_queue, &ctx->output_surfaces[i]);
  1132. if (res)
  1133. return res;
  1134. ctx->output_surfaces[i].busy = 1;
  1135. }
  1136. if (nv_status != NV_ENC_SUCCESS && nv_status != NV_ENC_ERR_NEED_MORE_INPUT) {
  1137. av_log(avctx, AV_LOG_ERROR, "EncodePicture failed!\n");
  1138. return AVERROR_EXTERNAL;
  1139. }
  1140. if (nv_status != NV_ENC_ERR_NEED_MORE_INPUT) {
  1141. while (ctx->output_surface_queue.count) {
  1142. tmpoutsurf = out_surf_queue_dequeue(&ctx->output_surface_queue);
  1143. res = out_surf_queue_enqueue(&ctx->output_surface_ready_queue, tmpoutsurf);
  1144. if (res)
  1145. return res;
  1146. }
  1147. if (frame) {
  1148. res = out_surf_queue_enqueue(&ctx->output_surface_ready_queue, &ctx->output_surfaces[i]);
  1149. if (res)
  1150. return res;
  1151. ctx->output_surfaces[i].busy = 1;
  1152. }
  1153. }
  1154. if (ctx->output_surface_ready_queue.count && (!frame || ctx->output_surface_ready_queue.count + ctx->output_surface_queue.count >= ctx->buffer_delay)) {
  1155. tmpoutsurf = out_surf_queue_dequeue(&ctx->output_surface_ready_queue);
  1156. res = process_output_surface(avctx, pkt, tmpoutsurf);
  1157. if (res)
  1158. return res;
  1159. tmpoutsurf->busy = 0;
  1160. av_assert0(tmpoutsurf->input_surface->lockCount);
  1161. tmpoutsurf->input_surface->lockCount--;
  1162. *got_packet = 1;
  1163. } else {
  1164. *got_packet = 0;
  1165. }
  1166. return 0;
  1167. }
  1168. static const enum AVPixelFormat pix_fmts_nvenc[] = {
  1169. AV_PIX_FMT_YUV420P,
  1170. AV_PIX_FMT_NV12,
  1171. AV_PIX_FMT_YUV444P,
  1172. AV_PIX_FMT_NONE
  1173. };
  1174. #define OFFSET(x) offsetof(NvencContext, x)
  1175. #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
  1176. static const AVOption options[] = {
  1177. { "preset", "Set the encoding preset (one of slow = hq 2pass, medium = hq, fast = hp, hq, hp, bd, ll, llhq, llhp, default)", OFFSET(preset), AV_OPT_TYPE_STRING, { .str = "hq" }, 0, 0, VE },
  1178. { "profile", "Set the encoding profile (high, main, baseline or high444p)", OFFSET(profile), AV_OPT_TYPE_STRING, { 0 }, 0, 0, VE },
  1179. { "level", "Set the encoding level restriction (auto, 1.0, 1.0b, 1.1, 1.2, ..., 4.2, 5.0, 5.1)", OFFSET(level), AV_OPT_TYPE_STRING, { 0 }, 0, 0, VE },
  1180. { "tier", "Set the encoding tier (main or high)", OFFSET(tier), AV_OPT_TYPE_STRING, { 0 }, 0, 0, VE },
  1181. { "cbr", "Use cbr encoding mode", OFFSET(cbr), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
  1182. { "2pass", "Use 2pass encoding mode", OFFSET(twopass), AV_OPT_TYPE_BOOL, { .i64 = -1 }, -1, 1, VE },
  1183. { "gpu", "Selects which NVENC capable GPU to use. First GPU is 0, second is 1, and so on.", OFFSET(gpu), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE },
  1184. { "delay", "Delays frame output by the given amount of frames.", OFFSET(buffer_delay), AV_OPT_TYPE_INT, { .i64 = INT_MAX }, 0, INT_MAX, VE },
  1185. { NULL }
  1186. };
  1187. static const AVCodecDefault nvenc_defaults[] = {
  1188. { "b", "0" },
  1189. { "qmin", "-1" },
  1190. { "qmax", "-1" },
  1191. { "qdiff", "-1" },
  1192. { "qblur", "-1" },
  1193. { "qcomp", "-1" },
  1194. { NULL },
  1195. };
  1196. #if CONFIG_NVENC_ENCODER
  1197. static const AVClass nvenc_class = {
  1198. .class_name = "nvenc",
  1199. .item_name = av_default_item_name,
  1200. .option = options,
  1201. .version = LIBAVUTIL_VERSION_INT,
  1202. };
  1203. AVCodec ff_nvenc_encoder = {
  1204. .name = "nvenc",
  1205. .long_name = NULL_IF_CONFIG_SMALL("NVIDIA NVENC h264 encoder"),
  1206. .type = AVMEDIA_TYPE_VIDEO,
  1207. .id = AV_CODEC_ID_H264,
  1208. .priv_data_size = sizeof(NvencContext),
  1209. .init = nvenc_encode_init,
  1210. .encode2 = nvenc_encode_frame,
  1211. .close = nvenc_encode_close,
  1212. .capabilities = AV_CODEC_CAP_DELAY,
  1213. .priv_class = &nvenc_class,
  1214. .defaults = nvenc_defaults,
  1215. .pix_fmts = pix_fmts_nvenc,
  1216. };
  1217. #endif
  1218. /* Add an alias for nvenc_h264 */
  1219. #if CONFIG_NVENC_H264_ENCODER
  1220. static const AVClass nvenc_h264_class = {
  1221. .class_name = "nvenc_h264",
  1222. .item_name = av_default_item_name,
  1223. .option = options,
  1224. .version = LIBAVUTIL_VERSION_INT,
  1225. };
  1226. AVCodec ff_nvenc_h264_encoder = {
  1227. .name = "nvenc_h264",
  1228. .long_name = NULL_IF_CONFIG_SMALL("NVIDIA NVENC h264 encoder"),
  1229. .type = AVMEDIA_TYPE_VIDEO,
  1230. .id = AV_CODEC_ID_H264,
  1231. .priv_data_size = sizeof(NvencContext),
  1232. .init = nvenc_encode_init,
  1233. .encode2 = nvenc_encode_frame,
  1234. .close = nvenc_encode_close,
  1235. .capabilities = AV_CODEC_CAP_DELAY,
  1236. .priv_class = &nvenc_h264_class,
  1237. .defaults = nvenc_defaults,
  1238. .pix_fmts = pix_fmts_nvenc,
  1239. };
  1240. #endif
  1241. #if CONFIG_NVENC_HEVC_ENCODER
  1242. static const AVClass nvenc_hevc_class = {
  1243. .class_name = "nvenc_hevc",
  1244. .item_name = av_default_item_name,
  1245. .option = options,
  1246. .version = LIBAVUTIL_VERSION_INT,
  1247. };
  1248. AVCodec ff_nvenc_hevc_encoder = {
  1249. .name = "nvenc_hevc",
  1250. .long_name = NULL_IF_CONFIG_SMALL("NVIDIA NVENC hevc encoder"),
  1251. .type = AVMEDIA_TYPE_VIDEO,
  1252. .id = AV_CODEC_ID_H265,
  1253. .priv_data_size = sizeof(NvencContext),
  1254. .init = nvenc_encode_init,
  1255. .encode2 = nvenc_encode_frame,
  1256. .close = nvenc_encode_close,
  1257. .capabilities = AV_CODEC_CAP_DELAY,
  1258. .priv_class = &nvenc_hevc_class,
  1259. .defaults = nvenc_defaults,
  1260. .pix_fmts = pix_fmts_nvenc,
  1261. };
  1262. #endif