libavutil/hwcontext_cuda.c

   1 /*
   2  * This file is part of FFmpeg.
   3  *
   4  * FFmpeg is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU Lesser General Public
   6  * License as published by the Free Software Foundation; either
   7  * version 2.1 of the License, or (at your option) any later version.
   8  *
   9  * FFmpeg is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  12  * Lesser General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU Lesser General Public
  15  * License along with FFmpeg; if not, write to the Free Software
  16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  17  */
  18
  19 #include "buffer.h"
  20 #include "common.h"
  21 #include "hwcontext.h"
  22 #include "hwcontext_internal.h"
  23 #include "hwcontext_cuda_internal.h"
  24 #if CONFIG_VULKAN
  25 #include "hwcontext_vulkan.h"
  26 #endif
  27 #include "cuda_check.h"
  28 #include "mem.h"
  29 #include "pixdesc.h"
  30 #include "pixfmt.h"
  31 #include "imgutils.h"
  32
  33 typedef struct CUDAFramesContext {
  34     int shift_width, shift_height;
  35     int tex_alignment;
  36 } CUDAFramesContext;
  37
  38 typedef struct CUDADeviceContext {
  39     AVCUDADeviceContext p;
  40     AVCUDADeviceContextInternal internal;
  41 } CUDADeviceContext;
  42
  43 static const enum AVPixelFormat supported_formats[] = {
  44     AV_PIX_FMT_NV12,
  45     AV_PIX_FMT_NV16,
  46     AV_PIX_FMT_YUV420P,
  47     AV_PIX_FMT_YUVA420P,
  48     AV_PIX_FMT_YUV444P,
  49     AV_PIX_FMT_P010,
  50     AV_PIX_FMT_P016,
  51     AV_PIX_FMT_P210,
  52     AV_PIX_FMT_P216,
  53     AV_PIX_FMT_YUV422P,
  54     AV_PIX_FMT_YUV420P10,
  55     AV_PIX_FMT_YUV422P10,
  56     AV_PIX_FMT_YUV444P10,
  57     AV_PIX_FMT_YUV444P10MSB,
  58     AV_PIX_FMT_YUV444P12MSB,
  59     AV_PIX_FMT_YUV444P16,
  60     AV_PIX_FMT_0RGB32,
  61     AV_PIX_FMT_0BGR32,
  62     AV_PIX_FMT_RGB32,
  63     AV_PIX_FMT_BGR32,
  64 #if CONFIG_VULKAN
  65     AV_PIX_FMT_VULKAN,
  66 #endif
  67 };
  68
  69 #define CHECK_CU(x) FF_CUDA_CHECK_DL(device_ctx, cu, x)
  70
  71 static int cuda_frames_get_constraints(AVHWDeviceContext *ctx,
  72                                        const void *hwconfig,
  73                                        AVHWFramesConstraints *constraints)
  74 {
  75     int i;
  76
  77     constraints->valid_sw_formats = av_malloc_array(FF_ARRAY_ELEMS(supported_formats) + 1,
  78                                                     sizeof(*constraints->valid_sw_formats));
  79     if (!constraints->valid_sw_formats)
  80         return AVERROR(ENOMEM);
  81
  82     for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++)
  83         constraints->valid_sw_formats[i] = supported_formats[i];
  84     constraints->valid_sw_formats[FF_ARRAY_ELEMS(supported_formats)] = AV_PIX_FMT_NONE;
  85
  86     constraints->valid_hw_formats = av_malloc_array(2, sizeof(*constraints->valid_hw_formats));
  87     if (!constraints->valid_hw_formats)
  88         return AVERROR(ENOMEM);
  89
  90     constraints->valid_hw_formats[0] = AV_PIX_FMT_CUDA;
  91     constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE;
  92
  93     return 0;
  94 }
  95
  96 static void cuda_buffer_free(void *opaque, uint8_t *data)
  97 {
  98     AVHWFramesContext        *ctx = opaque;
  99     AVHWDeviceContext *device_ctx = ctx->device_ctx;
 100     AVCUDADeviceContext    *hwctx = device_ctx->hwctx;
 101     CudaFunctions             *cu = hwctx->internal->cuda_dl;
 102
 103     CUcontext dummy;
 104
 105     CHECK_CU(cu->cuCtxPushCurrent(hwctx->cuda_ctx));
 106
 107     CHECK_CU(cu->cuMemFree((CUdeviceptr)data));
 108
 109     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
 110 }
 111
 112 static AVBufferRef *cuda_pool_alloc(void *opaque, size_t size)
 113 {
 114     AVHWFramesContext        *ctx = opaque;
 115     AVHWDeviceContext *device_ctx = ctx->device_ctx;
 116     AVCUDADeviceContext    *hwctx = device_ctx->hwctx;
 117     CudaFunctions             *cu = hwctx->internal->cuda_dl;
 118
 119     AVBufferRef *ret = NULL;
 120     CUcontext dummy = NULL;
 121     CUdeviceptr data;
 122     int err;
 123
 124     err = CHECK_CU(cu->cuCtxPushCurrent(hwctx->cuda_ctx));
 125     if (err < 0)
 126         return NULL;
 127
 128     err = CHECK_CU(cu->cuMemAlloc(&data, size));
 129     if (err < 0)
 130         goto fail;
 131
 132     ret = av_buffer_create((uint8_t*)data, size, cuda_buffer_free, ctx, 0);
 133     if (!ret) {
 134         CHECK_CU(cu->cuMemFree(data));
 135         goto fail;
 136     }
 137
 138 fail:
 139     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
 140     return ret;
 141 }
 142
 143 static int cuda_frames_init(AVHWFramesContext *ctx)
 144 {
 145     AVHWDeviceContext *device_ctx = ctx->device_ctx;
 146     AVCUDADeviceContext    *hwctx = device_ctx->hwctx;
 147     CUDAFramesContext       *priv = ctx->hwctx;
 148     CudaFunctions             *cu = hwctx->internal->cuda_dl;
 149     int err, i;
 150
 151     for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++) {
 152         if (ctx->sw_format == supported_formats[i])
 153             break;
 154     }
 155     if (i == FF_ARRAY_ELEMS(supported_formats)) {
 156         av_log(ctx, AV_LOG_ERROR, "Pixel format '%s' is not supported\n",
 157                av_get_pix_fmt_name(ctx->sw_format));
 158         return AVERROR(ENOSYS);
 159     }
 160
 161     err = CHECK_CU(cu->cuDeviceGetAttribute(&priv->tex_alignment,
 162                                             14 /* CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT */,
 163                                             hwctx->internal->cuda_device));
 164     if (err < 0)
 165         return err;
 166
 167     av_log(ctx, AV_LOG_DEBUG, "CUDA texture alignment: %d\n", priv->tex_alignment);
 168
 169     // YUV420P is a special case.
 170     // Since nvenc expects the U/V planes to have half the linesize of the Y plane
 171     // alignment has to be doubled to ensure the U/V planes still end up aligned.
 172     if (ctx->sw_format == AV_PIX_FMT_YUV420P)
 173         priv->tex_alignment *= 2;
 174
 175     av_pix_fmt_get_chroma_sub_sample(ctx->sw_format, &priv->shift_width, &priv->shift_height);
 176
 177     if (!ctx->pool) {
 178         int size = av_image_get_buffer_size(ctx->sw_format, ctx->width, ctx->height, priv->tex_alignment);
 179         if (size < 0)
 180             return size;
 181
 182         ffhwframesctx(ctx)->pool_internal =
 183             av_buffer_pool_init2(size, ctx, cuda_pool_alloc, NULL);
 184         if (!ffhwframesctx(ctx)->pool_internal)
 185             return AVERROR(ENOMEM);
 186     }
 187
 188     return 0;
 189 }
 190
 191 static int cuda_get_buffer(AVHWFramesContext *ctx, AVFrame *frame)
 192 {
 193     CUDAFramesContext *priv = ctx->hwctx;
 194     int res;
 195
 196     frame->buf[0] = av_buffer_pool_get(ctx->pool);
 197     if (!frame->buf[0])
 198         return AVERROR(ENOMEM);
 199
 200     res = av_image_fill_arrays(frame->data, frame->linesize, frame->buf[0]->data,
 201                                ctx->sw_format, ctx->width, ctx->height, priv->tex_alignment);
 202     if (res < 0)
 203         return res;
 204
 205     // YUV420P is a special case.
 206     // Nvenc expects the U/V planes in swapped order from how ffmpeg expects them, also chroma is half-aligned
 207     if (ctx->sw_format == AV_PIX_FMT_YUV420P) {
 208         frame->linesize[1] = frame->linesize[2] = frame->linesize[0] / 2;
 209         frame->data[2]     = frame->data[1];
 210         frame->data[1]     = frame->data[2] + frame->linesize[2] * (ctx->height / 2);
 211     }
 212
 213     frame->format = AV_PIX_FMT_CUDA;
 214     frame->width  = ctx->width;
 215     frame->height = ctx->height;
 216
 217     return 0;
 218 }
 219
 220 static int cuda_transfer_get_formats(AVHWFramesContext *ctx,
 221                                      enum AVHWFrameTransferDirection dir,
 222                                      enum AVPixelFormat **formats)
 223 {
 224     enum AVPixelFormat *fmts;
 225
 226     fmts = av_malloc_array(2, sizeof(*fmts));
 227     if (!fmts)
 228         return AVERROR(ENOMEM);
 229
 230     fmts[0] = ctx->sw_format;
 231     fmts[1] = AV_PIX_FMT_NONE;
 232
 233     *formats = fmts;
 234
 235     return 0;
 236 }
 237
 238 static int cuda_transfer_data(AVHWFramesContext *ctx, AVFrame *dst,
 239                                  const AVFrame *src)
 240 {
 241     CUDAFramesContext       *priv = ctx->hwctx;
 242     AVHWDeviceContext *device_ctx = ctx->device_ctx;
 243     AVCUDADeviceContext    *hwctx = device_ctx->hwctx;
 244     CudaFunctions             *cu = hwctx->internal->cuda_dl;
 245
 246     CUcontext dummy;
 247     int i, ret;
 248
 249     if ((src->hw_frames_ctx && ((AVHWFramesContext*)src->hw_frames_ctx->data)->format != AV_PIX_FMT_CUDA) ||
 250         (dst->hw_frames_ctx && ((AVHWFramesContext*)dst->hw_frames_ctx->data)->format != AV_PIX_FMT_CUDA))
 251         return AVERROR(ENOSYS);
 252
 253     ret = CHECK_CU(cu->cuCtxPushCurrent(hwctx->cuda_ctx));
 254     if (ret < 0)
 255         return ret;
 256
 257     for (i = 0; i < FF_ARRAY_ELEMS(src->data) && src->data[i]; i++) {
 258         CUDA_MEMCPY2D cpy = {
 259             .srcPitch      = src->linesize[i],
 260             .dstPitch      = dst->linesize[i],
 261             .WidthInBytes  = FFMIN(src->linesize[i], dst->linesize[i]),
 262             .Height        = src->height >> ((i == 0 || i == 3) ? 0 : priv->shift_height),
 263         };
 264
 265         if (src->hw_frames_ctx) {
 266             cpy.srcMemoryType = CU_MEMORYTYPE_DEVICE;
 267             cpy.srcDevice     = (CUdeviceptr)src->data[i];
 268         } else {
 269             cpy.srcMemoryType = CU_MEMORYTYPE_HOST;
 270             cpy.srcHost       = src->data[i];
 271         }
 272
 273         if (dst->hw_frames_ctx) {
 274             cpy.dstMemoryType = CU_MEMORYTYPE_DEVICE;
 275             cpy.dstDevice     = (CUdeviceptr)dst->data[i];
 276         } else {
 277             cpy.dstMemoryType = CU_MEMORYTYPE_HOST;
 278             cpy.dstHost       = dst->data[i];
 279         }
 280
 281         ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, hwctx->stream));
 282         if (ret < 0)
 283             goto exit;
 284     }
 285
 286     if (!dst->hw_frames_ctx) {
 287         ret = CHECK_CU(cu->cuStreamSynchronize(hwctx->stream));
 288         if (ret < 0)
 289             goto exit;
 290     }
 291
 292 exit:
 293     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
 294
 295     return 0;
 296 }
 297
 298 static void cuda_device_uninit(AVHWDeviceContext *device_ctx)
 299 {
 300     CUDADeviceContext *hwctx = device_ctx->hwctx;
 301
 302     if (hwctx->p.internal) {
 303         CudaFunctions *cu = hwctx->internal.cuda_dl;
 304
 305         if (hwctx->internal.is_allocated && hwctx->p.cuda_ctx) {
 306             if (hwctx->internal.flags & AV_CUDA_USE_PRIMARY_CONTEXT)
 307                 CHECK_CU(cu->cuDevicePrimaryCtxRelease(hwctx->internal.cuda_device));
 308             else if (!(hwctx->internal.flags & AV_CUDA_USE_CURRENT_CONTEXT))
 309                 CHECK_CU(cu->cuCtxDestroy(hwctx->p.cuda_ctx));
 310
 311             hwctx->p.cuda_ctx = NULL;
 312         }
 313
 314         cuda_free_functions(&hwctx->internal.cuda_dl);
 315         memset(&hwctx->internal, 0, sizeof(hwctx->internal));
 316         hwctx->p.internal = NULL;
 317     }
 318 }
 319
 320 static int cuda_device_init(AVHWDeviceContext *ctx)
 321 {
 322     CUDADeviceContext *hwctx = ctx->hwctx;
 323     int ret;
 324
 325     hwctx->p.internal = &hwctx->internal;
 326
 327     if (!hwctx->internal.cuda_dl) {
 328         ret = cuda_load_functions(&hwctx->internal.cuda_dl, ctx);
 329         if (ret < 0) {
 330             av_log(ctx, AV_LOG_ERROR, "Could not dynamically load CUDA\n");
 331             goto error;
 332         }
 333     }
 334
 335     return 0;
 336
 337 error:
 338     cuda_device_uninit(ctx);
 339     return ret;
 340 }
 341
 342 static int cuda_context_init(AVHWDeviceContext *device_ctx, int flags) {
 343     AVCUDADeviceContext *hwctx = device_ctx->hwctx;
 344     CudaFunctions *cu;
 345     CUcontext dummy;
 346     int ret, dev_active = 0;
 347     unsigned int dev_flags = 0;
 348
 349     const unsigned int desired_flags = CU_CTX_SCHED_BLOCKING_SYNC;
 350
 351     cu = hwctx->internal->cuda_dl;
 352
 353     hwctx->internal->flags = flags;
 354
 355     if (flags & AV_CUDA_USE_PRIMARY_CONTEXT) {
 356         ret = CHECK_CU(cu->cuDevicePrimaryCtxGetState(hwctx->internal->cuda_device,
 357                        &dev_flags, &dev_active));
 358         if (ret < 0)
 359             return ret;
 360
 361         if (dev_active && dev_flags != desired_flags) {
 362             av_log(device_ctx, AV_LOG_ERROR, "Primary context already active with incompatible flags.\n");
 363             return AVERROR(ENOTSUP);
 364         } else if (dev_flags != desired_flags) {
 365             ret = CHECK_CU(cu->cuDevicePrimaryCtxSetFlags(hwctx->internal->cuda_device,
 366                            desired_flags));
 367             if (ret < 0)
 368                 return ret;
 369         }
 370
 371         ret = CHECK_CU(cu->cuDevicePrimaryCtxRetain(&hwctx->cuda_ctx,
 372                                                     hwctx->internal->cuda_device));
 373         if (ret < 0)
 374             return ret;
 375     } else if (flags & AV_CUDA_USE_CURRENT_CONTEXT) {
 376         ret = CHECK_CU(cu->cuCtxGetCurrent(&hwctx->cuda_ctx));
 377         if (ret < 0)
 378             return ret;
 379         av_log(device_ctx, AV_LOG_INFO, "Using current CUDA context.\n");
 380     } else {
 381         ret = CHECK_CU(cu->cuCtxCreate(&hwctx->cuda_ctx, desired_flags,
 382                                        hwctx->internal->cuda_device));
 383         if (ret < 0)
 384             return ret;
 385
 386         CHECK_CU(cu->cuCtxPopCurrent(&dummy));
 387     }
 388
 389     hwctx->internal->is_allocated = 1;
 390
 391     // Setting stream to NULL will make functions automatically use the default CUstream
 392     hwctx->stream = NULL;
 393
 394     return 0;
 395 }
 396
 397 static int cuda_flags_from_opts(AVHWDeviceContext *device_ctx,
 398                                 AVDictionary *opts, int *flags)
 399 {
 400     AVDictionaryEntry *primary_ctx_opt = av_dict_get(opts, "primary_ctx", NULL, 0);
 401     AVDictionaryEntry *current_ctx_opt = av_dict_get(opts, "current_ctx", NULL, 0);
 402
 403     int use_primary_ctx = 0, use_current_ctx = 0;
 404     if (primary_ctx_opt)
 405         use_primary_ctx = strtol(primary_ctx_opt->value, NULL, 10);
 406
 407     if (current_ctx_opt)
 408         use_current_ctx = strtol(current_ctx_opt->value, NULL, 10);
 409
 410     if (use_primary_ctx && use_current_ctx) {
 411         av_log(device_ctx, AV_LOG_ERROR, "Requested both primary and current CUDA context simultaneously.\n");
 412         return AVERROR(EINVAL);
 413     }
 414
 415     if (primary_ctx_opt && use_primary_ctx) {
 416         av_log(device_ctx, AV_LOG_VERBOSE, "Using CUDA primary device context\n");
 417         *flags |= AV_CUDA_USE_PRIMARY_CONTEXT;
 418     } else if (primary_ctx_opt) {
 419         av_log(device_ctx, AV_LOG_VERBOSE, "Disabling use of CUDA primary device context\n");
 420         *flags &= ~AV_CUDA_USE_PRIMARY_CONTEXT;
 421     }
 422
 423     if (current_ctx_opt && use_current_ctx) {
 424         av_log(device_ctx, AV_LOG_VERBOSE, "Using CUDA current device context\n");
 425         *flags |= AV_CUDA_USE_CURRENT_CONTEXT;
 426     } else if (current_ctx_opt) {
 427         av_log(device_ctx, AV_LOG_VERBOSE, "Disabling use of CUDA current device context\n");
 428         *flags &= ~AV_CUDA_USE_CURRENT_CONTEXT;
 429     }
 430
 431     return 0;
 432 }
 433
 434 static int cuda_device_create(AVHWDeviceContext *device_ctx,
 435                               const char *device,
 436                               AVDictionary *opts, int flags)
 437 {
 438     AVCUDADeviceContext *hwctx = device_ctx->hwctx;
 439     CudaFunctions *cu;
 440     int ret, device_idx = 0;
 441
 442     ret = cuda_flags_from_opts(device_ctx, opts, &flags);
 443     if (ret < 0)
 444         goto error;
 445
 446     if (device)
 447         device_idx = strtol(device, NULL, 0);
 448
 449     ret = cuda_device_init(device_ctx);
 450     if (ret < 0)
 451         goto error;
 452
 453     cu = hwctx->internal->cuda_dl;
 454
 455     ret = CHECK_CU(cu->cuInit(0));
 456     if (ret < 0)
 457         goto error;
 458
 459     ret = CHECK_CU(cu->cuDeviceGet(&hwctx->internal->cuda_device, device_idx));
 460     if (ret < 0)
 461         goto error;
 462
 463     ret = cuda_context_init(device_ctx, flags);
 464     if (ret < 0)
 465         goto error;
 466
 467     return 0;
 468
 469 error:
 470     cuda_device_uninit(device_ctx);
 471     return ret;
 472 }
 473
 474 static int cuda_device_derive(AVHWDeviceContext *device_ctx,
 475                               AVHWDeviceContext *src_ctx, AVDictionary *opts,
 476                               int flags) {
 477     AVCUDADeviceContext *hwctx = device_ctx->hwctx;
 478     CudaFunctions *cu;
 479     const char *src_uuid = NULL;
 480 #if CONFIG_VULKAN
 481     VkPhysicalDeviceIDProperties vk_idp;
 482 #endif
 483     int ret, i, device_count;
 484
 485     ret = cuda_flags_from_opts(device_ctx, opts, &flags);
 486     if (ret < 0)
 487         goto error;
 488
 489 #if CONFIG_VULKAN
 490     vk_idp = (VkPhysicalDeviceIDProperties) {
 491         .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES,
 492     };
 493 #endif
 494
 495     switch (src_ctx->type) {
 496 #if CONFIG_VULKAN
 497 #define TYPE PFN_vkGetPhysicalDeviceProperties2
 498     case AV_HWDEVICE_TYPE_VULKAN: {
 499         AVVulkanDeviceContext *vkctx = src_ctx->hwctx;
 500         TYPE prop_fn = (TYPE)vkctx->get_proc_addr(vkctx->inst, "vkGetPhysicalDeviceProperties2");
 501         VkPhysicalDeviceProperties2 vk_dev_props = {
 502             .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2,
 503             .pNext = &vk_idp,
 504         };
 505         prop_fn(vkctx->phys_dev, &vk_dev_props);
 506         src_uuid = vk_idp.deviceUUID;
 507         break;
 508     }
 509 #undef TYPE
 510 #endif
 511     default:
 512         ret = AVERROR(ENOSYS);
 513         goto error;
 514     }
 515
 516     if (!src_uuid) {
 517         av_log(device_ctx, AV_LOG_ERROR,
 518                "Failed to get UUID of source device.\n");
 519         ret = AVERROR(EINVAL);
 520         goto error;
 521     }
 522
 523     ret = cuda_device_init(device_ctx);
 524     if (ret < 0)
 525         goto error;
 526
 527     cu = hwctx->internal->cuda_dl;
 528
 529     ret = CHECK_CU(cu->cuInit(0));
 530     if (ret < 0)
 531         goto error;
 532
 533     ret = CHECK_CU(cu->cuDeviceGetCount(&device_count));
 534     if (ret < 0)
 535         goto error;
 536
 537     hwctx->internal->cuda_device = -1;
 538     for (i = 0; i < device_count; i++) {
 539         CUdevice dev;
 540         CUuuid uuid;
 541
 542         ret = CHECK_CU(cu->cuDeviceGet(&dev, i));
 543         if (ret < 0)
 544             goto error;
 545
 546         ret = CHECK_CU(cu->cuDeviceGetUuid(&uuid, dev));
 547         if (ret < 0)
 548             goto error;
 549
 550         if (memcmp(src_uuid, uuid.bytes, sizeof (uuid.bytes)) == 0) {
 551             hwctx->internal->cuda_device = dev;
 552             break;
 553         }
 554     }
 555
 556     if (hwctx->internal->cuda_device == -1) {
 557         av_log(device_ctx, AV_LOG_ERROR, "Could not derive CUDA device.\n");
 558         goto error;
 559     }
 560
 561     ret = cuda_context_init(device_ctx, flags);
 562     if (ret < 0)
 563         goto error;
 564
 565     return 0;
 566
 567 error:
 568     cuda_device_uninit(device_ctx);
 569     return ret;
 570 }
 571
 572 const HWContextType ff_hwcontext_type_cuda = {
 573     .type                 = AV_HWDEVICE_TYPE_CUDA,
 574     .name                 = "CUDA",
 575
 576     .device_hwctx_size    = sizeof(CUDADeviceContext),
 577     .frames_hwctx_size    = sizeof(CUDAFramesContext),
 578
 579     .device_create        = cuda_device_create,
 580     .device_derive        = cuda_device_derive,
 581     .device_init          = cuda_device_init,
 582     .device_uninit        = cuda_device_uninit,
 583     .frames_get_constraints = cuda_frames_get_constraints,
 584     .frames_init          = cuda_frames_init,
 585     .frames_get_buffer    = cuda_get_buffer,
 586     .transfer_get_formats = cuda_transfer_get_formats,
 587     .transfer_data_to     = cuda_transfer_data,
 588     .transfer_data_from   = cuda_transfer_data,
 589
 590     .pix_fmts             = (const enum AVPixelFormat[]){ AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE },
 591 };