git.sesse.net Git - ffmpeg/blob - libavutil/hwcontext_vulkan.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 "config.h"
  20 #include "pixdesc.h"
  21 #include "avstring.h"
  22 #include "imgutils.h"
  23 #include "hwcontext.h"
  24 #include "hwcontext_internal.h"
  25 #include "hwcontext_vulkan.h"
  26
  27 #if CONFIG_LIBDRM
  28 #include <unistd.h>
  29 #include <xf86drm.h>
  30 #include <drm_fourcc.h>
  31 #include "hwcontext_drm.h"
  32 #if CONFIG_VAAPI
  33 #include <va/va_drmcommon.h>
  34 #include "hwcontext_vaapi.h"
  35 #endif
  36 #endif
  37
  38 #if CONFIG_CUDA
  39 #include "hwcontext_cuda_internal.h"
  40 #include "cuda_check.h"
  41 #define CHECK_CU(x) FF_CUDA_CHECK_DL(cuda_cu, cu, x)
  42 #endif
  43
  44 typedef struct VulkanExecCtx {
  45     VkCommandPool pool;
  46     VkCommandBuffer buf;
  47     VkQueue queue;
  48     VkFence fence;
  49 } VulkanExecCtx;
  50
  51 typedef struct VulkanDevicePriv {
  52     /* Properties */
  53     VkPhysicalDeviceProperties props;
  54     VkPhysicalDeviceMemoryProperties mprops;
  55
  56     /* Debug callback */
  57     VkDebugUtilsMessengerEXT debug_ctx;
  58
  59     /* Image uploading */
  60     VulkanExecCtx cmd;
  61
  62     /* Extensions */
  63     uint64_t extensions;
  64
  65     /* Settings */
  66     int use_linear_images;
  67
  68     /* Nvidia */
  69     int dev_is_nvidia;
  70 } VulkanDevicePriv;
  71
  72 typedef struct VulkanFramesPriv {
  73     VulkanExecCtx cmd;
  74 } VulkanFramesPriv;
  75
  76 typedef struct AVVkFrameInternal {
  77 #if CONFIG_CUDA
  78     /* Importing external memory into cuda is really expensive so we keep the
  79      * memory imported all the time */
  80     AVBufferRef *cuda_fc_ref; /* Need to keep it around for uninit */
  81     CUexternalMemory ext_mem[AV_NUM_DATA_POINTERS];
  82     CUmipmappedArray cu_mma[AV_NUM_DATA_POINTERS];
  83     CUarray cu_array[AV_NUM_DATA_POINTERS];
  84     CUexternalSemaphore cu_sem[AV_NUM_DATA_POINTERS];
  85 #endif
  86 } AVVkFrameInternal;
  87
  88 #define VK_LOAD_PFN(inst, name) PFN_##name pfn_##name = (PFN_##name)           \
  89                                               vkGetInstanceProcAddr(inst, #name)
  90
  91 #define DEFAULT_USAGE_FLAGS (VK_IMAGE_USAGE_SAMPLED_BIT      |                 \
  92                              VK_IMAGE_USAGE_STORAGE_BIT      |                 \
  93                              VK_IMAGE_USAGE_TRANSFER_SRC_BIT |                 \
  94                              VK_IMAGE_USAGE_TRANSFER_DST_BIT)
  95
  96 #define ADD_VAL_TO_LIST(list, count, val)                                      \
  97     do {                                                                       \
  98         list = av_realloc_array(list, sizeof(*list), ++count);                 \
  99         if (!list) {                                                           \
 100             err = AVERROR(ENOMEM);                                             \
 101             goto end;                                                          \
 102         }                                                                      \
 103         list[count - 1] = val;                                                 \
 104     } while(0)
 105
 106 static const struct {
 107     enum AVPixelFormat pixfmt;
 108     const VkFormat vkfmts[3];
 109 } vk_pixfmt_map[] = {
 110     { AV_PIX_FMT_GRAY8,   { VK_FORMAT_R8_UNORM } },
 111     { AV_PIX_FMT_GRAY16,  { VK_FORMAT_R16_UNORM } },
 112     { AV_PIX_FMT_GRAYF32, { VK_FORMAT_R32_SFLOAT } },
 113
 114     { AV_PIX_FMT_NV12, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } },
 115     { AV_PIX_FMT_P010, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
 116     { AV_PIX_FMT_P016, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
 117
 118     { AV_PIX_FMT_YUV420P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
 119     { AV_PIX_FMT_YUV422P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
 120     { AV_PIX_FMT_YUV444P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
 121
 122     { AV_PIX_FMT_YUV420P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
 123     { AV_PIX_FMT_YUV422P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
 124     { AV_PIX_FMT_YUV444P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
 125
 126     { AV_PIX_FMT_ABGR,   { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
 127     { AV_PIX_FMT_BGRA,   { VK_FORMAT_B8G8R8A8_UNORM } },
 128     { AV_PIX_FMT_RGBA,   { VK_FORMAT_R8G8B8A8_UNORM } },
 129     { AV_PIX_FMT_RGB24,  { VK_FORMAT_R8G8B8_UNORM } },
 130     { AV_PIX_FMT_BGR24,  { VK_FORMAT_B8G8R8_UNORM } },
 131     { AV_PIX_FMT_RGB48,  { VK_FORMAT_R16G16B16_UNORM } },
 132     { AV_PIX_FMT_RGBA64, { VK_FORMAT_R16G16B16A16_UNORM } },
 133     { AV_PIX_FMT_RGB565, { VK_FORMAT_R5G6B5_UNORM_PACK16 } },
 134     { AV_PIX_FMT_BGR565, { VK_FORMAT_B5G6R5_UNORM_PACK16 } },
 135     { AV_PIX_FMT_BGR0,   { VK_FORMAT_B8G8R8A8_UNORM } },
 136     { AV_PIX_FMT_0BGR,   { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
 137     { AV_PIX_FMT_RGB0,   { VK_FORMAT_R8G8B8A8_UNORM } },
 138
 139     { AV_PIX_FMT_GBRPF32, { VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT } },
 140 };
 141
 142 const VkFormat *av_vkfmt_from_pixfmt(enum AVPixelFormat p)
 143 {
 144     for (enum AVPixelFormat i = 0; i < FF_ARRAY_ELEMS(vk_pixfmt_map); i++)
 145         if (vk_pixfmt_map[i].pixfmt == p)
 146             return vk_pixfmt_map[i].vkfmts;
 147     return NULL;
 148 }
 149
 150 static int pixfmt_is_supported(AVVulkanDeviceContext *hwctx, enum AVPixelFormat p,
 151                                int linear)
 152 {
 153     const VkFormat *fmt = av_vkfmt_from_pixfmt(p);
 154     int planes = av_pix_fmt_count_planes(p);
 155
 156     if (!fmt)
 157         return 0;
 158
 159     for (int i = 0; i < planes; i++) {
 160         VkFormatFeatureFlags flags;
 161         VkFormatProperties2 prop = {
 162             .sType = VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2,
 163         };
 164         vkGetPhysicalDeviceFormatProperties2(hwctx->phys_dev, fmt[i], &prop);
 165         flags = linear ? prop.formatProperties.linearTilingFeatures :
 166                          prop.formatProperties.optimalTilingFeatures;
 167         if (!(flags & DEFAULT_USAGE_FLAGS))
 168             return 0;
 169     }
 170
 171     return 1;
 172 }
 173
 174 enum VulkanExtensions {
 175     EXT_EXTERNAL_DMABUF_MEMORY = 1ULL <<  0, /* VK_EXT_external_memory_dma_buf */
 176     EXT_DRM_MODIFIER_FLAGS     = 1ULL <<  1, /* VK_EXT_image_drm_format_modifier */
 177     EXT_EXTERNAL_FD_MEMORY     = 1ULL <<  2, /* VK_KHR_external_memory_fd */
 178     EXT_EXTERNAL_FD_SEM        = 1ULL <<  3, /* VK_KHR_external_semaphore_fd */
 179
 180     EXT_OPTIONAL               = 1ULL << 62,
 181     EXT_REQUIRED               = 1ULL << 63,
 182 };
 183
 184 typedef struct VulkanOptExtension {
 185     const char *name;
 186     uint64_t flag;
 187 } VulkanOptExtension;
 188
 189 static const VulkanOptExtension optional_instance_exts[] = {
 190     /* For future use */
 191 };
 192
 193 static const VulkanOptExtension optional_device_exts[] = {
 194     { VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME,               EXT_EXTERNAL_FD_MEMORY,     },
 195     { VK_EXT_EXTERNAL_MEMORY_DMA_BUF_EXTENSION_NAME,          EXT_EXTERNAL_DMABUF_MEMORY, },
 196     { VK_EXT_IMAGE_DRM_FORMAT_MODIFIER_EXTENSION_NAME,        EXT_DRM_MODIFIER_FLAGS,     },
 197     { VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME,            EXT_EXTERNAL_FD_SEM,        },
 198 };
 199
 200 /* Converts return values to strings */
 201 static const char *vk_ret2str(VkResult res)
 202 {
 203 #define CASE(VAL) case VAL: return #VAL
 204     switch (res) {
 205     CASE(VK_SUCCESS);
 206     CASE(VK_NOT_READY);
 207     CASE(VK_TIMEOUT);
 208     CASE(VK_EVENT_SET);
 209     CASE(VK_EVENT_RESET);
 210     CASE(VK_INCOMPLETE);
 211     CASE(VK_ERROR_OUT_OF_HOST_MEMORY);
 212     CASE(VK_ERROR_OUT_OF_DEVICE_MEMORY);
 213     CASE(VK_ERROR_INITIALIZATION_FAILED);
 214     CASE(VK_ERROR_DEVICE_LOST);
 215     CASE(VK_ERROR_MEMORY_MAP_FAILED);
 216     CASE(VK_ERROR_LAYER_NOT_PRESENT);
 217     CASE(VK_ERROR_EXTENSION_NOT_PRESENT);
 218     CASE(VK_ERROR_FEATURE_NOT_PRESENT);
 219     CASE(VK_ERROR_INCOMPATIBLE_DRIVER);
 220     CASE(VK_ERROR_TOO_MANY_OBJECTS);
 221     CASE(VK_ERROR_FORMAT_NOT_SUPPORTED);
 222     CASE(VK_ERROR_FRAGMENTED_POOL);
 223     CASE(VK_ERROR_SURFACE_LOST_KHR);
 224     CASE(VK_ERROR_NATIVE_WINDOW_IN_USE_KHR);
 225     CASE(VK_SUBOPTIMAL_KHR);
 226     CASE(VK_ERROR_OUT_OF_DATE_KHR);
 227     CASE(VK_ERROR_INCOMPATIBLE_DISPLAY_KHR);
 228     CASE(VK_ERROR_VALIDATION_FAILED_EXT);
 229     CASE(VK_ERROR_INVALID_SHADER_NV);
 230     CASE(VK_ERROR_OUT_OF_POOL_MEMORY);
 231     CASE(VK_ERROR_INVALID_EXTERNAL_HANDLE);
 232     CASE(VK_ERROR_NOT_PERMITTED_EXT);
 233     CASE(VK_ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT);
 234     CASE(VK_ERROR_INVALID_DEVICE_ADDRESS_EXT);
 235     CASE(VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT);
 236     default: return "Unknown error";
 237     }
 238 #undef CASE
 239 }
 240
 241 static VkBool32 vk_dbg_callback(VkDebugUtilsMessageSeverityFlagBitsEXT severity,
 242                                 VkDebugUtilsMessageTypeFlagsEXT messageType,
 243                                 const VkDebugUtilsMessengerCallbackDataEXT *data,
 244                                 void *priv)
 245 {
 246     int l;
 247     AVHWDeviceContext *ctx = priv;
 248
 249     switch (severity) {
 250     case VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT: l = AV_LOG_VERBOSE; break;
 251     case VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT:    l = AV_LOG_INFO;    break;
 252     case VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT: l = AV_LOG_WARNING; break;
 253     case VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT:   l = AV_LOG_ERROR;   break;
 254     default:                                              l = AV_LOG_DEBUG;   break;
 255     }
 256
 257     av_log(ctx, l, "%s\n", data->pMessage);
 258     for (int i = 0; i < data->cmdBufLabelCount; i++)
 259         av_log(ctx, l, "\t%i: %s\n", i, data->pCmdBufLabels[i].pLabelName);
 260
 261     return 0;
 262 }
 263
 264 static int check_extensions(AVHWDeviceContext *ctx, int dev,
 265                             const char * const **dst, uint32_t *num, int debug)
 266 {
 267     const char *tstr;
 268     const char **extension_names = NULL;
 269     VulkanDevicePriv *p = ctx->internal->priv;
 270     AVVulkanDeviceContext *hwctx = ctx->hwctx;
 271     int err = 0, found, extensions_found = 0;
 272
 273     const char *mod;
 274     int optional_exts_num;
 275     uint32_t sup_ext_count;
 276     VkExtensionProperties *sup_ext;
 277     const VulkanOptExtension *optional_exts;
 278
 279     if (!dev) {
 280         mod = "instance";
 281         optional_exts = optional_instance_exts;
 282         optional_exts_num = FF_ARRAY_ELEMS(optional_instance_exts);
 283         vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, NULL);
 284         sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
 285         if (!sup_ext)
 286             return AVERROR(ENOMEM);
 287         vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, sup_ext);
 288     } else {
 289         mod = "device";
 290         optional_exts = optional_device_exts;
 291         optional_exts_num = FF_ARRAY_ELEMS(optional_device_exts);
 292         vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
 293                                              &sup_ext_count, NULL);
 294         sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
 295         if (!sup_ext)
 296             return AVERROR(ENOMEM);
 297         vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
 298                                              &sup_ext_count, sup_ext);
 299     }
 300
 301     for (int i = 0; i < optional_exts_num; i++) {
 302         int req = optional_exts[i].flag & EXT_REQUIRED;
 303         tstr = optional_exts[i].name;
 304
 305         found = 0;
 306         for (int j = 0; j < sup_ext_count; j++) {
 307             if (!strcmp(tstr, sup_ext[j].extensionName)) {
 308                 found = 1;
 309                 break;
 310             }
 311         }
 312         if (!found) {
 313             int lvl = req ? AV_LOG_ERROR : AV_LOG_VERBOSE;
 314             av_log(ctx, lvl, "Extension \"%s\" not found!\n", tstr);
 315             if (req) {
 316                 err = AVERROR(EINVAL);
 317                 goto end;
 318             }
 319             continue;
 320         }
 321         if (!req)
 322             p->extensions |= optional_exts[i].flag;
 323
 324         av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
 325
 326         ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
 327     }
 328
 329     if (debug && !dev) {
 330         tstr = VK_EXT_DEBUG_UTILS_EXTENSION_NAME;
 331         found = 0;
 332         for (int j = 0; j < sup_ext_count; j++) {
 333             if (!strcmp(tstr, sup_ext[j].extensionName)) {
 334                 found = 1;
 335                 break;
 336             }
 337         }
 338         if (found) {
 339             ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
 340         } else {
 341             av_log(ctx, AV_LOG_ERROR, "Debug extension \"%s\" not found!\n",
 342                    tstr);
 343             err = AVERROR(EINVAL);
 344             goto end;
 345         }
 346     }
 347
 348     *dst = extension_names;
 349     *num = extensions_found;
 350
 351 end:
 352     av_free(sup_ext);
 353     return err;
 354 }
 355
 356 /* Creates a VkInstance */
 357 static int create_instance(AVHWDeviceContext *ctx, AVDictionary *opts)
 358 {
 359     int err = 0;
 360     VkResult ret;
 361     VulkanDevicePriv *p = ctx->internal->priv;
 362     AVVulkanDeviceContext *hwctx = ctx->hwctx;
 363     AVDictionaryEntry *debug_opt = av_dict_get(opts, "debug", NULL, 0);
 364     const int debug_mode = debug_opt && strtol(debug_opt->value, NULL, 10);
 365     VkApplicationInfo application_info = {
 366         .sType              = VK_STRUCTURE_TYPE_APPLICATION_INFO,
 367         .pEngineName        = "libavutil",
 368         .apiVersion         = VK_API_VERSION_1_1,
 369         .engineVersion      = VK_MAKE_VERSION(LIBAVUTIL_VERSION_MAJOR,
 370                                               LIBAVUTIL_VERSION_MINOR,
 371                                               LIBAVUTIL_VERSION_MICRO),
 372     };
 373     VkInstanceCreateInfo inst_props = {
 374         .sType            = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
 375         .pApplicationInfo = &application_info,
 376     };
 377
 378     /* Check for present/missing extensions */
 379     err = check_extensions(ctx, 0, &inst_props.ppEnabledExtensionNames,
 380                            &inst_props.enabledExtensionCount, debug_mode);
 381     if (err < 0)
 382         return err;
 383
 384     if (debug_mode) {
 385         static const char *layers[] = { "VK_LAYER_LUNARG_standard_validation" };
 386         inst_props.ppEnabledLayerNames = layers;
 387         inst_props.enabledLayerCount = FF_ARRAY_ELEMS(layers);
 388     }
 389
 390     /* Try to create the instance */
 391     ret = vkCreateInstance(&inst_props, hwctx->alloc, &hwctx->inst);
 392
 393     /* Free used memory */
 394     av_free((void *)inst_props.ppEnabledExtensionNames);
 395
 396     /* Check for errors */
 397     if (ret != VK_SUCCESS) {
 398         av_log(ctx, AV_LOG_ERROR, "Instance creation failure: %s\n",
 399                vk_ret2str(ret));
 400         return AVERROR_EXTERNAL;
 401     }
 402
 403     if (debug_mode) {
 404         VkDebugUtilsMessengerCreateInfoEXT dbg = {
 405             .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_MESSENGER_CREATE_INFO_EXT,
 406             .messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT |
 407                                VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT    |
 408                                VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT |
 409                                VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT,
 410             .messageType = VK_DEBUG_UTILS_MESSAGE_TYPE_GENERAL_BIT_EXT    |
 411                            VK_DEBUG_UTILS_MESSAGE_TYPE_VALIDATION_BIT_EXT |
 412                            VK_DEBUG_UTILS_MESSAGE_TYPE_PERFORMANCE_BIT_EXT,
 413             .pfnUserCallback = vk_dbg_callback,
 414             .pUserData = ctx,
 415         };
 416         VK_LOAD_PFN(hwctx->inst, vkCreateDebugUtilsMessengerEXT);
 417
 418         pfn_vkCreateDebugUtilsMessengerEXT(hwctx->inst, &dbg,
 419                                            hwctx->alloc, &p->debug_ctx);
 420     }
 421
 422     return 0;
 423 }
 424
 425 typedef struct VulkanDeviceSelection {
 426     uint8_t uuid[VK_UUID_SIZE]; /* Will use this first unless !has_uuid */
 427     int has_uuid;
 428     const char *name; /* Will use this second unless NULL */
 429     uint32_t pci_device; /* Will use this third unless 0x0 */
 430     uint32_t vendor_id; /* Last resort to find something deterministic */
 431     int index; /* Finally fall back to index */
 432 } VulkanDeviceSelection;
 433
 434 static const char *vk_dev_type(enum VkPhysicalDeviceType type)
 435 {
 436     switch (type) {
 437     case VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU: return "integrated";
 438     case VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU:   return "discrete";
 439     case VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU:    return "virtual";
 440     case VK_PHYSICAL_DEVICE_TYPE_CPU:            return "software";
 441     default:                                     return "unknown";
 442     }
 443 }
 444
 445 /* Finds a device */
 446 static int find_device(AVHWDeviceContext *ctx, VulkanDeviceSelection *select)
 447 {
 448     int err = 0, choice = -1;
 449     uint32_t num;
 450     VkResult ret;
 451     VkPhysicalDevice *devices = NULL;
 452     VkPhysicalDeviceIDProperties *idp = NULL;
 453     VkPhysicalDeviceProperties2 *prop = NULL;
 454     VulkanDevicePriv *p = ctx->internal->priv;
 455     AVVulkanDeviceContext *hwctx = ctx->hwctx;
 456
 457     ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, NULL);
 458     if (ret != VK_SUCCESS || !num) {
 459         av_log(ctx, AV_LOG_ERROR, "No devices found: %s!\n", vk_ret2str(ret));
 460         return AVERROR(ENODEV);
 461     }
 462
 463     devices = av_malloc_array(num, sizeof(VkPhysicalDevice));
 464     if (!devices)
 465         return AVERROR(ENOMEM);
 466
 467     ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, devices);
 468     if (ret != VK_SUCCESS) {
 469         av_log(ctx, AV_LOG_ERROR, "Failed enumerating devices: %s\n",
 470                vk_ret2str(ret));
 471         err = AVERROR(ENODEV);
 472         goto end;
 473     }
 474
 475     prop = av_mallocz_array(num, sizeof(*prop));
 476     if (!prop) {
 477         err = AVERROR(ENOMEM);
 478         goto end;
 479     }
 480
 481     idp = av_mallocz_array(num, sizeof(*idp));
 482     if (!idp) {
 483         err = AVERROR(ENOMEM);
 484         goto end;
 485     }
 486
 487     av_log(ctx, AV_LOG_VERBOSE, "GPU listing:\n");
 488     for (int i = 0; i < num; i++) {
 489         idp[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES;
 490         prop[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2;
 491         prop[i].pNext = &idp[i];
 492
 493         vkGetPhysicalDeviceProperties2(devices[i], &prop[i]);
 494         av_log(ctx, AV_LOG_VERBOSE, "    %d: %s (%s) (0x%x)\n", i,
 495                prop[i].properties.deviceName,
 496                vk_dev_type(prop[i].properties.deviceType),
 497                prop[i].properties.deviceID);
 498     }
 499
 500     if (select->has_uuid) {
 501         for (int i = 0; i < num; i++) {
 502             if (!strncmp(idp[i].deviceUUID, select->uuid, VK_UUID_SIZE)) {
 503                 choice = i;
 504                 goto end;
 505              }
 506         }
 507         av_log(ctx, AV_LOG_ERROR, "Unable to find device by given UUID!\n");
 508         err = AVERROR(ENODEV);
 509         goto end;
 510     } else if (select->name) {
 511         av_log(ctx, AV_LOG_VERBOSE, "Requested device: %s\n", select->name);
 512         for (int i = 0; i < num; i++) {
 513             if (strstr(prop[i].properties.deviceName, select->name)) {
 514                 choice = i;
 515                 goto end;
 516              }
 517         }
 518         av_log(ctx, AV_LOG_ERROR, "Unable to find device \"%s\"!\n",
 519                select->name);
 520         err = AVERROR(ENODEV);
 521         goto end;
 522     } else if (select->pci_device) {
 523         av_log(ctx, AV_LOG_VERBOSE, "Requested device: 0x%x\n", select->pci_device);
 524         for (int i = 0; i < num; i++) {
 525             if (select->pci_device == prop[i].properties.deviceID) {
 526                 choice = i;
 527                 goto end;
 528             }
 529         }
 530         av_log(ctx, AV_LOG_ERROR, "Unable to find device with PCI ID 0x%x!\n",
 531                select->pci_device);
 532         err = AVERROR(EINVAL);
 533         goto end;
 534     } else if (select->vendor_id) {
 535         av_log(ctx, AV_LOG_VERBOSE, "Requested vendor: 0x%x\n", select->vendor_id);
 536         for (int i = 0; i < num; i++) {
 537             if (select->vendor_id == prop[i].properties.vendorID) {
 538                 choice = i;
 539                 goto end;
 540             }
 541         }
 542         av_log(ctx, AV_LOG_ERROR, "Unable to find device with Vendor ID 0x%x!\n",
 543                select->vendor_id);
 544         err = AVERROR(ENODEV);
 545         goto end;
 546     } else {
 547         if (select->index < num) {
 548             choice = select->index;
 549             goto end;
 550         }
 551         av_log(ctx, AV_LOG_ERROR, "Unable to find device with index %i!\n",
 552                select->index);
 553         err = AVERROR(ENODEV);
 554         goto end;
 555     }
 556
 557 end:
 558     if (choice > -1) {
 559         p->dev_is_nvidia = (prop[choice].properties.vendorID == 0x10de);
 560         hwctx->phys_dev = devices[choice];
 561     }
 562     av_free(devices);
 563     av_free(prop);
 564     av_free(idp);
 565
 566     return err;
 567 }
 568
 569 static int search_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
 570 {
 571     uint32_t num;
 572     VkQueueFamilyProperties *qs = NULL;
 573     AVVulkanDeviceContext *hwctx = ctx->hwctx;
 574     int graph_index = -1, comp_index = -1, tx_index = -1;
 575     VkDeviceQueueCreateInfo *pc = (VkDeviceQueueCreateInfo *)cd->pQueueCreateInfos;
 576
 577     /* First get the number of queue families */
 578     vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, NULL);
 579     if (!num) {
 580         av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
 581         return AVERROR_EXTERNAL;
 582     }
 583
 584     /* Then allocate memory */
 585     qs = av_malloc_array(num, sizeof(VkQueueFamilyProperties));
 586     if (!qs)
 587         return AVERROR(ENOMEM);
 588
 589     /* Finally retrieve the queue families */
 590     vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, qs);
 591
 592 #define SEARCH_FLAGS(expr, out)                                                \
 593     for (int i = 0; i < num; i++) {                                            \
 594         const VkQueueFlagBits flags = qs[i].queueFlags;                        \
 595         if (expr) {                                                            \
 596             out = i;                                                           \
 597             break;                                                             \
 598         }                                                                      \
 599     }
 600
 601     SEARCH_FLAGS(flags & VK_QUEUE_GRAPHICS_BIT, graph_index)
 602
 603     SEARCH_FLAGS((flags &  VK_QUEUE_COMPUTE_BIT) && (i != graph_index),
 604                  comp_index)
 605
 606     SEARCH_FLAGS((flags & VK_QUEUE_TRANSFER_BIT) && (i != graph_index) &&
 607                  (i != comp_index), tx_index)
 608
 609 #undef SEARCH_FLAGS
 610 #define QF_FLAGS(flags)                                                        \
 611     ((flags) & VK_QUEUE_GRAPHICS_BIT      ) ? "(graphics) " : "",              \
 612     ((flags) & VK_QUEUE_COMPUTE_BIT       ) ? "(compute) "  : "",              \
 613     ((flags) & VK_QUEUE_TRANSFER_BIT      ) ? "(transfer) " : "",              \
 614     ((flags) & VK_QUEUE_SPARSE_BINDING_BIT) ? "(sparse) "   : ""
 615
 616     av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for graphics, "
 617            "flags: %s%s%s%s\n", graph_index, QF_FLAGS(qs[graph_index].queueFlags));
 618
 619     hwctx->queue_family_index      = graph_index;
 620     hwctx->queue_family_tx_index   = graph_index;
 621     hwctx->queue_family_comp_index = graph_index;
 622
 623     pc[cd->queueCreateInfoCount++].queueFamilyIndex = graph_index;
 624
 625     if (comp_index != -1) {
 626         av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for compute, "
 627                "flags: %s%s%s%s\n", comp_index, QF_FLAGS(qs[comp_index].queueFlags));
 628         hwctx->queue_family_tx_index                    = comp_index;
 629         hwctx->queue_family_comp_index                  = comp_index;
 630         pc[cd->queueCreateInfoCount++].queueFamilyIndex = comp_index;
 631     }
 632
 633     if (tx_index != -1) {
 634         av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for transfers, "
 635                "flags: %s%s%s%s\n", tx_index, QF_FLAGS(qs[tx_index].queueFlags));
 636         hwctx->queue_family_tx_index                    = tx_index;
 637         pc[cd->queueCreateInfoCount++].queueFamilyIndex = tx_index;
 638     }
 639
 640 #undef QF_FLAGS
 641
 642     av_free(qs);
 643
 644     return 0;
 645 }
 646
 647 static int create_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
 648                            int queue_family_index)
 649 {
 650     VkResult ret;
 651     AVVulkanDeviceContext *hwctx = ctx->hwctx;
 652
 653     VkCommandPoolCreateInfo cqueue_create = {
 654         .sType              = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
 655         .flags              = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
 656         .queueFamilyIndex   = queue_family_index,
 657     };
 658     VkCommandBufferAllocateInfo cbuf_create = {
 659         .sType              = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
 660         .level              = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
 661         .commandBufferCount = 1,
 662     };
 663
 664     VkFenceCreateInfo fence_spawn = {
 665         .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
 666     };
 667
 668     ret = vkCreateFence(hwctx->act_dev, &fence_spawn,
 669                         hwctx->alloc, &cmd->fence);
 670     if (ret != VK_SUCCESS) {
 671         av_log(ctx, AV_LOG_ERROR, "Failed to create frame fence: %s\n",
 672                vk_ret2str(ret));
 673         return AVERROR_EXTERNAL;
 674     }
 675
 676     ret = vkCreateCommandPool(hwctx->act_dev, &cqueue_create,
 677                               hwctx->alloc, &cmd->pool);
 678     if (ret != VK_SUCCESS) {
 679         av_log(ctx, AV_LOG_ERROR, "Command pool creation failure: %s\n",
 680                vk_ret2str(ret));
 681         return AVERROR_EXTERNAL;
 682     }
 683
 684     cbuf_create.commandPool = cmd->pool;
 685
 686     ret = vkAllocateCommandBuffers(hwctx->act_dev, &cbuf_create, &cmd->buf);
 687     if (ret != VK_SUCCESS) {
 688         av_log(ctx, AV_LOG_ERROR, "Command buffer alloc failure: %s\n",
 689                vk_ret2str(ret));
 690         return AVERROR_EXTERNAL;
 691     }
 692
 693     vkGetDeviceQueue(hwctx->act_dev, cqueue_create.queueFamilyIndex, 0,
 694                      &cmd->queue);
 695
 696     return 0;
 697 }
 698
 699 static void free_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd)
 700 {
 701     AVVulkanDeviceContext *hwctx = ctx->hwctx;
 702
 703     if (cmd->fence)
 704         vkDestroyFence(hwctx->act_dev, cmd->fence, hwctx->alloc);
 705     if (cmd->buf)
 706         vkFreeCommandBuffers(hwctx->act_dev, cmd->pool, 1, &cmd->buf);
 707     if (cmd->pool)
 708         vkDestroyCommandPool(hwctx->act_dev, cmd->pool, hwctx->alloc);
 709 }
 710
 711 static void vulkan_device_free(AVHWDeviceContext *ctx)
 712 {
 713     VulkanDevicePriv *p = ctx->internal->priv;
 714     AVVulkanDeviceContext *hwctx = ctx->hwctx;
 715
 716     free_exec_ctx(ctx, &p->cmd);
 717
 718     vkDestroyDevice(hwctx->act_dev, hwctx->alloc);
 719
 720     if (p->debug_ctx) {
 721         VK_LOAD_PFN(hwctx->inst, vkDestroyDebugUtilsMessengerEXT);
 722         pfn_vkDestroyDebugUtilsMessengerEXT(hwctx->inst, p->debug_ctx,
 723                                             hwctx->alloc);
 724     }
 725
 726     vkDestroyInstance(hwctx->inst, hwctx->alloc);
 727 }
 728
 729 static int vulkan_device_create_internal(AVHWDeviceContext *ctx,
 730                                          VulkanDeviceSelection *dev_select,
 731                                          AVDictionary *opts, int flags)
 732 {
 733     int err = 0;
 734     VkResult ret;
 735     AVDictionaryEntry *opt_d;
 736     VulkanDevicePriv *p = ctx->internal->priv;
 737     AVVulkanDeviceContext *hwctx = ctx->hwctx;
 738     VkDeviceQueueCreateInfo queue_create_info[3] = {
 739         {   .sType            = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
 740             .pQueuePriorities = (float []){ 1.0f },
 741             .queueCount       = 1, },
 742         {   .sType            = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
 743             .pQueuePriorities = (float []){ 1.0f },
 744             .queueCount       = 1, },
 745         {   .sType            = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
 746             .pQueuePriorities = (float []){ 1.0f },
 747             .queueCount       = 1, },
 748     };
 749
 750     VkDeviceCreateInfo dev_info = {
 751         .sType                = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
 752         .pQueueCreateInfos    = queue_create_info,
 753         .queueCreateInfoCount = 0,
 754     };
 755
 756     ctx->free = vulkan_device_free;
 757
 758     /* Create an instance if not given one */
 759     if ((err = create_instance(ctx, opts)))
 760         goto end;
 761
 762     /* Find a device (if not given one) */
 763     if ((err = find_device(ctx, dev_select)))
 764         goto end;
 765
 766     vkGetPhysicalDeviceProperties(hwctx->phys_dev, &p->props);
 767     av_log(ctx, AV_LOG_VERBOSE, "Using device: %s\n", p->props.deviceName);
 768     av_log(ctx, AV_LOG_VERBOSE, "Alignments:\n");
 769     av_log(ctx, AV_LOG_VERBOSE, "    optimalBufferCopyOffsetAlignment:   %li\n",
 770            p->props.limits.optimalBufferCopyOffsetAlignment);
 771     av_log(ctx, AV_LOG_VERBOSE, "    optimalBufferCopyRowPitchAlignment: %li\n",
 772            p->props.limits.optimalBufferCopyRowPitchAlignment);
 773     av_log(ctx, AV_LOG_VERBOSE, "    minMemoryMapAlignment:              %li\n",
 774            p->props.limits.minMemoryMapAlignment);
 775
 776     /* Search queue family */
 777     if ((err = search_queue_families(ctx, &dev_info)))
 778         goto end;
 779
 780     if ((err = check_extensions(ctx, 1, &dev_info.ppEnabledExtensionNames,
 781                                 &dev_info.enabledExtensionCount, 0)))
 782         goto end;
 783
 784     ret = vkCreateDevice(hwctx->phys_dev, &dev_info, hwctx->alloc,
 785                          &hwctx->act_dev);
 786
 787     av_free((void *)dev_info.ppEnabledExtensionNames);
 788
 789     if (ret != VK_SUCCESS) {
 790         av_log(ctx, AV_LOG_ERROR, "Device creation failure: %s\n",
 791                vk_ret2str(ret));
 792         err = AVERROR_EXTERNAL;
 793         goto end;
 794     }
 795
 796     /* Tiled images setting, use them by default */
 797     opt_d = av_dict_get(opts, "linear_images", NULL, 0);
 798     if (opt_d)
 799         p->use_linear_images = strtol(opt_d->value, NULL, 10);
 800
 801 end:
 802     return err;
 803 }
 804
 805 static int vulkan_device_init(AVHWDeviceContext *ctx)
 806 {
 807     int err;
 808     uint32_t queue_num;
 809     AVVulkanDeviceContext *hwctx = ctx->hwctx;
 810     VulkanDevicePriv *p = ctx->internal->priv;
 811
 812     vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &queue_num, NULL);
 813     if (!queue_num) {
 814         av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
 815         return AVERROR_EXTERNAL;
 816     }
 817
 818 #define CHECK_QUEUE(type, n)                                                         \
 819 if (n >= queue_num) {                                                                \
 820     av_log(ctx, AV_LOG_ERROR, "Invalid %s queue index %i (device has %i queues)!\n", \
 821            type, n, queue_num);                                                      \
 822     return AVERROR(EINVAL);                                                          \
 823 }
 824
 825     CHECK_QUEUE("graphics", hwctx->queue_family_index)
 826     CHECK_QUEUE("upload",   hwctx->queue_family_tx_index)
 827     CHECK_QUEUE("compute",  hwctx->queue_family_comp_index)
 828
 829 #undef CHECK_QUEUE
 830
 831     /* Create exec context - if there's something invalid this will error out */
 832     err = create_exec_ctx(ctx, &p->cmd, hwctx->queue_family_tx_index);
 833     if (err)
 834         return err;
 835
 836     /* Get device capabilities */
 837     vkGetPhysicalDeviceMemoryProperties(hwctx->phys_dev, &p->mprops);
 838
 839     return 0;
 840 }
 841
 842 static int vulkan_device_create(AVHWDeviceContext *ctx, const char *device,
 843                                 AVDictionary *opts, int flags)
 844 {
 845     VulkanDeviceSelection dev_select = { 0 };
 846     if (device && device[0]) {
 847         char *end = NULL;
 848         dev_select.index = strtol(device, &end, 10);
 849         if (end == device) {
 850             dev_select.index = 0;
 851             dev_select.name  = device;
 852         }
 853     }
 854
 855     return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
 856 }
 857
 858 static int vulkan_device_derive(AVHWDeviceContext *ctx,
 859                                 AVHWDeviceContext *src_ctx, int flags)
 860 {
 861     av_unused VulkanDeviceSelection dev_select = { 0 };
 862
 863     /* If there's only one device on the system, then even if its not covered
 864      * by the following checks (e.g. non-PCIe ARM GPU), having an empty
 865      * dev_select will mean it'll get picked. */
 866     switch(src_ctx->type) {
 867 #if CONFIG_LIBDRM
 868 #if CONFIG_VAAPI
 869     case AV_HWDEVICE_TYPE_VAAPI: {
 870         AVVAAPIDeviceContext *src_hwctx = src_ctx->hwctx;
 871
 872         const char *vendor = vaQueryVendorString(src_hwctx->display);
 873         if (!vendor) {
 874             av_log(ctx, AV_LOG_ERROR, "Unable to get device info from VAAPI!\n");
 875             return AVERROR_EXTERNAL;
 876         }
 877
 878         if (strstr(vendor, "Intel"))
 879             dev_select.vendor_id = 0x8086;
 880         if (strstr(vendor, "AMD"))
 881             dev_select.vendor_id = 0x1002;
 882
 883         return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
 884     }
 885 #endif
 886     case AV_HWDEVICE_TYPE_DRM: {
 887         AVDRMDeviceContext *src_hwctx = src_ctx->hwctx;
 888
 889         drmDevice *drm_dev_info;
 890         int err = drmGetDevice(src_hwctx->fd, &drm_dev_info);
 891         if (err) {
 892             av_log(ctx, AV_LOG_ERROR, "Unable to get device info from DRM fd!\n");
 893             return AVERROR_EXTERNAL;
 894         }
 895
 896         if (drm_dev_info->bustype == DRM_BUS_PCI)
 897             dev_select.pci_device = drm_dev_info->deviceinfo.pci->device_id;
 898
 899         drmFreeDevice(&drm_dev_info);
 900
 901         return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
 902     }
 903 #endif
 904 #if CONFIG_CUDA
 905     case AV_HWDEVICE_TYPE_CUDA: {
 906         AVHWDeviceContext *cuda_cu = src_ctx;
 907         AVCUDADeviceContext *src_hwctx = src_ctx->hwctx;
 908         AVCUDADeviceContextInternal *cu_internal = src_hwctx->internal;
 909         CudaFunctions *cu = cu_internal->cuda_dl;
 910
 911         int ret = CHECK_CU(cu->cuDeviceGetUuid((CUuuid *)&dev_select.uuid,
 912                                                cu_internal->cuda_device));
 913         if (ret < 0) {
 914             av_log(ctx, AV_LOG_ERROR, "Unable to get UUID from CUDA!\n");
 915             return AVERROR_EXTERNAL;
 916         }
 917
 918         dev_select.has_uuid = 1;
 919
 920         return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
 921     }
 922 #endif
 923     default:
 924         return AVERROR(ENOSYS);
 925     }
 926 }
 927
 928 static int vulkan_frames_get_constraints(AVHWDeviceContext *ctx,
 929                                          const void *hwconfig,
 930                                          AVHWFramesConstraints *constraints)
 931 {
 932     int count = 0;
 933     AVVulkanDeviceContext *hwctx = ctx->hwctx;
 934     VulkanDevicePriv *p = ctx->internal->priv;
 935
 936     for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
 937         count += pixfmt_is_supported(hwctx, i, p->use_linear_images);
 938
 939 #if CONFIG_CUDA
 940     if (p->dev_is_nvidia)
 941         count++;
 942 #endif
 943
 944     constraints->valid_sw_formats = av_malloc_array(count + 1,
 945                                                     sizeof(enum AVPixelFormat));
 946     if (!constraints->valid_sw_formats)
 947         return AVERROR(ENOMEM);
 948
 949     count = 0;
 950     for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
 951         if (pixfmt_is_supported(hwctx, i, p->use_linear_images))
 952             constraints->valid_sw_formats[count++] = i;
 953
 954 #if CONFIG_CUDA
 955     if (p->dev_is_nvidia)
 956         constraints->valid_sw_formats[count++] = AV_PIX_FMT_CUDA;
 957 #endif
 958     constraints->valid_sw_formats[count++] = AV_PIX_FMT_NONE;
 959
 960     constraints->min_width  = 0;
 961     constraints->min_height = 0;
 962     constraints->max_width  = p->props.limits.maxImageDimension2D;
 963     constraints->max_height = p->props.limits.maxImageDimension2D;
 964
 965     constraints->valid_hw_formats = av_malloc_array(2, sizeof(enum AVPixelFormat));
 966     if (!constraints->valid_hw_formats)
 967         return AVERROR(ENOMEM);
 968
 969     constraints->valid_hw_formats[0] = AV_PIX_FMT_VULKAN;
 970     constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE;
 971
 972     return 0;
 973 }
 974
 975 static int alloc_mem(AVHWDeviceContext *ctx, VkMemoryRequirements *req,
 976                      VkMemoryPropertyFlagBits req_flags, void *alloc_extension,
 977                      VkMemoryPropertyFlagBits *mem_flags, VkDeviceMemory *mem)
 978 {
 979     VkResult ret;
 980     int index = -1;
 981     VulkanDevicePriv *p = ctx->internal->priv;
 982     AVVulkanDeviceContext *dev_hwctx = ctx->hwctx;
 983     VkMemoryAllocateInfo alloc_info = {
 984         .sType           = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
 985         .pNext           = alloc_extension,
 986     };
 987
 988     /* Align if we need to */
 989     if (req_flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
 990         req->size = FFALIGN(req->size, p->props.limits.minMemoryMapAlignment);
 991
 992     alloc_info.allocationSize = req->size;
 993
 994     /* The vulkan spec requires memory types to be sorted in the "optimal"
 995      * order, so the first matching type we find will be the best/fastest one */
 996     for (int i = 0; i < p->mprops.memoryTypeCount; i++) {
 997         /* The memory type must be supported by the requirements (bitfield) */
 998         if (!(req->memoryTypeBits & (1 << i)))
 999             continue;
1000
1001         /* The memory type flags must include our properties */
1002         if ((p->mprops.memoryTypes[i].propertyFlags & req_flags) != req_flags)
1003             continue;
1004
1005         /* Found a suitable memory type */
1006         index = i;
1007         break;
1008     }
1009
1010     if (index < 0) {
1011         av_log(ctx, AV_LOG_ERROR, "No memory type found for flags 0x%x\n",
1012                req_flags);
1013         return AVERROR(EINVAL);
1014     }
1015
1016     alloc_info.memoryTypeIndex = index;
1017
1018     ret = vkAllocateMemory(dev_hwctx->act_dev, &alloc_info,
1019                            dev_hwctx->alloc, mem);
1020     if (ret != VK_SUCCESS) {
1021         av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory: %s\n",
1022                vk_ret2str(ret));
1023         return AVERROR(ENOMEM);
1024     }
1025
1026     *mem_flags |= p->mprops.memoryTypes[index].propertyFlags;
1027
1028     return 0;
1029 }
1030
1031 static void vulkan_free_internal(AVVkFrameInternal *internal)
1032 {
1033     if (!internal)
1034         return;
1035
1036 #if CONFIG_CUDA
1037     if (internal->cuda_fc_ref) {
1038         AVHWFramesContext *cuda_fc = (AVHWFramesContext *)internal->cuda_fc_ref->data;
1039         int planes = av_pix_fmt_count_planes(cuda_fc->sw_format);
1040         AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
1041         AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
1042         AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
1043         CudaFunctions *cu = cu_internal->cuda_dl;
1044
1045         for (int i = 0; i < planes; i++) {
1046             if (internal->cu_sem[i])
1047                 CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem[i]));
1048             if (internal->cu_mma[i])
1049                 CHECK_CU(cu->cuMipmappedArrayDestroy(internal->cu_mma[i]));
1050             if (internal->ext_mem[i])
1051                 CHECK_CU(cu->cuDestroyExternalMemory(internal->ext_mem[i]));
1052         }
1053
1054         av_buffer_unref(&internal->cuda_fc_ref);
1055     }
1056 #endif
1057
1058     av_free(internal);
1059 }
1060
1061 static void vulkan_frame_free(void *opaque, uint8_t *data)
1062 {
1063     AVVkFrame *f = (AVVkFrame *)data;
1064     AVHWFramesContext *hwfc = opaque;
1065     AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1066     int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1067
1068     vulkan_free_internal(f->internal);
1069
1070     for (int i = 0; i < planes; i++) {
1071         vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1072         vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1073         vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
1074     }
1075
1076     av_free(f);
1077 }
1078
1079 static int alloc_bind_mem(AVHWFramesContext *hwfc, AVVkFrame *f,
1080                           void *alloc_pnext, size_t alloc_pnext_stride)
1081 {
1082     int err;
1083     VkResult ret;
1084     AVHWDeviceContext *ctx = hwfc->device_ctx;
1085     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1086     VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { { 0 } };
1087
1088     AVVulkanDeviceContext *hwctx = ctx->hwctx;
1089
1090     for (int i = 0; i < planes; i++) {
1091         int use_ded_mem;
1092         VkImageMemoryRequirementsInfo2 req_desc = {
1093             .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
1094             .image = f->img[i],
1095         };
1096         VkMemoryDedicatedAllocateInfo ded_alloc = {
1097             .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
1098             .pNext = (void *)(((uint8_t *)alloc_pnext) + i*alloc_pnext_stride),
1099         };
1100         VkMemoryDedicatedRequirements ded_req = {
1101             .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
1102         };
1103         VkMemoryRequirements2 req = {
1104             .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
1105             .pNext = &ded_req,
1106         };
1107
1108         vkGetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req);
1109
1110         /* In case the implementation prefers/requires dedicated allocation */
1111         use_ded_mem = ded_req.prefersDedicatedAllocation |
1112                       ded_req.requiresDedicatedAllocation;
1113         if (use_ded_mem)
1114             ded_alloc.image = f->img[i];
1115
1116         /* Allocate memory */
1117         if ((err = alloc_mem(ctx, &req.memoryRequirements,
1118                              f->tiling == VK_IMAGE_TILING_LINEAR ?
1119                              VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT :
1120                              VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1121                              use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext,
1122                              &f->flags, &f->mem[i])))
1123             return err;
1124
1125         f->size[i] = req.memoryRequirements.size;
1126         bind_info[i].sType  = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1127         bind_info[i].image  = f->img[i];
1128         bind_info[i].memory = f->mem[i];
1129     }
1130
1131     /* Bind the allocated memory to the images */
1132     ret = vkBindImageMemory2(hwctx->act_dev, planes, bind_info);
1133     if (ret != VK_SUCCESS) {
1134         av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1135                vk_ret2str(ret));
1136         return AVERROR_EXTERNAL;
1137     }
1138
1139     return 0;
1140 }
1141
1142 enum PrepMode {
1143     PREP_MODE_WRITE,
1144     PREP_MODE_RO_SHADER,
1145 };
1146
1147 static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
1148                          AVVkFrame *frame, enum PrepMode pmode)
1149 {
1150     VkResult ret;
1151     VkImageLayout new_layout;
1152     VkAccessFlags new_access;
1153     AVHWDeviceContext *ctx = hwfc->device_ctx;
1154     AVVulkanDeviceContext *hwctx = ctx->hwctx;
1155     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1156
1157     VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
1158
1159     VkCommandBufferBeginInfo cmd_start = {
1160         .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
1161         .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
1162     };
1163
1164     VkSubmitInfo s_info = {
1165         .sType                = VK_STRUCTURE_TYPE_SUBMIT_INFO,
1166         .commandBufferCount   = 1,
1167         .pCommandBuffers      = &ectx->buf,
1168
1169         .pSignalSemaphores    = frame->sem,
1170         .signalSemaphoreCount = planes,
1171     };
1172
1173     switch (pmode) {
1174     case PREP_MODE_WRITE:
1175         new_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
1176         new_access = VK_ACCESS_TRANSFER_WRITE_BIT;
1177         break;
1178     case PREP_MODE_RO_SHADER:
1179         new_layout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
1180         new_access = VK_ACCESS_TRANSFER_READ_BIT;
1181         break;
1182     }
1183
1184     ret = vkBeginCommandBuffer(ectx->buf, &cmd_start);
1185     if (ret != VK_SUCCESS)
1186         return AVERROR_EXTERNAL;
1187
1188     /* Change the image layout to something more optimal for writes.
1189      * This also signals the newly created semaphore, making it usable
1190      * for synchronization */
1191     for (int i = 0; i < planes; i++) {
1192         img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
1193         img_bar[i].srcAccessMask = 0x0;
1194         img_bar[i].dstAccessMask = new_access;
1195         img_bar[i].oldLayout = frame->layout[i];
1196         img_bar[i].newLayout = new_layout;
1197         img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
1198         img_bar[i].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
1199         img_bar[i].image = frame->img[i];
1200         img_bar[i].subresourceRange.levelCount = 1;
1201         img_bar[i].subresourceRange.layerCount = 1;
1202         img_bar[i].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
1203
1204         frame->layout[i] = img_bar[i].newLayout;
1205         frame->access[i] = img_bar[i].dstAccessMask;
1206     }
1207
1208     vkCmdPipelineBarrier(ectx->buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
1209                          VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
1210                          0, NULL, 0, NULL, planes, img_bar);
1211
1212     ret = vkEndCommandBuffer(ectx->buf);
1213     if (ret != VK_SUCCESS)
1214         return AVERROR_EXTERNAL;
1215
1216     ret = vkQueueSubmit(ectx->queue, 1, &s_info, ectx->fence);
1217     if (ret != VK_SUCCESS) {
1218         return AVERROR_EXTERNAL;
1219     } else {
1220         vkWaitForFences(hwctx->act_dev, 1, &ectx->fence, VK_TRUE, UINT64_MAX);
1221         vkResetFences(hwctx->act_dev, 1, &ectx->fence);
1222     }
1223
1224     return 0;
1225 }
1226
1227 static int create_frame(AVHWFramesContext *hwfc, AVVkFrame **frame,
1228                         VkImageTiling tiling, VkImageUsageFlagBits usage,
1229                         void *create_pnext)
1230 {
1231     int err;
1232     VkResult ret;
1233     AVHWDeviceContext *ctx = hwfc->device_ctx;
1234     VulkanDevicePriv *p = ctx->internal->priv;
1235     AVVulkanDeviceContext *hwctx = ctx->hwctx;
1236     enum AVPixelFormat format = hwfc->sw_format;
1237     const VkFormat *img_fmts = av_vkfmt_from_pixfmt(format);
1238     const int planes = av_pix_fmt_count_planes(format);
1239
1240     VkExportSemaphoreCreateInfo ext_sem_info = {
1241         .sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO,
1242         .handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
1243     };
1244
1245     VkSemaphoreCreateInfo sem_spawn = {
1246         .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1247         .pNext = p->extensions & EXT_EXTERNAL_FD_SEM ? &ext_sem_info : NULL,
1248     };
1249
1250     AVVkFrame *f = av_vk_frame_alloc();
1251     if (!f) {
1252         av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1253         return AVERROR(ENOMEM);
1254     }
1255
1256     /* Create the images */
1257     for (int i = 0; i < planes; i++) {
1258         const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
1259         int w = hwfc->width;
1260         int h = hwfc->height;
1261         const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
1262         const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
1263
1264         VkImageCreateInfo image_create_info = {
1265             .sType         = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1266             .pNext         = create_pnext,
1267             .imageType     = VK_IMAGE_TYPE_2D,
1268             .format        = img_fmts[i],
1269             .extent.width  = p_w,
1270             .extent.height = p_h,
1271             .extent.depth  = 1,
1272             .mipLevels     = 1,
1273             .arrayLayers   = 1,
1274             .flags         = VK_IMAGE_CREATE_ALIAS_BIT,
1275             .tiling        = tiling,
1276             .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
1277             .usage         = usage,
1278             .sharingMode   = VK_SHARING_MODE_EXCLUSIVE,
1279             .samples       = VK_SAMPLE_COUNT_1_BIT,
1280         };
1281
1282         ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1283                             hwctx->alloc, &f->img[i]);
1284         if (ret != VK_SUCCESS) {
1285             av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1286                    vk_ret2str(ret));
1287             err = AVERROR(EINVAL);
1288             goto fail;
1289         }
1290
1291         /* Create semaphore */
1292         ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1293                                 hwctx->alloc, &f->sem[i]);
1294         if (ret != VK_SUCCESS) {
1295             av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1296                    vk_ret2str(ret));
1297             return AVERROR_EXTERNAL;
1298         }
1299
1300         f->layout[i] = image_create_info.initialLayout;
1301         f->access[i] = 0x0;
1302     }
1303
1304     f->flags     = 0x0;
1305     f->tiling    = tiling;
1306
1307     *frame = f;
1308     return 0;
1309
1310 fail:
1311     vulkan_frame_free(hwfc, (uint8_t *)f);
1312     return err;
1313 }
1314
1315 /* Checks if an export flag is enabled, and if it is ORs it with *iexp */
1316 static void try_export_flags(AVHWFramesContext *hwfc,
1317                              VkExternalMemoryHandleTypeFlags *comp_handle_types,
1318                              VkExternalMemoryHandleTypeFlagBits *iexp,
1319                              VkExternalMemoryHandleTypeFlagBits exp)
1320 {
1321     VkResult ret;
1322     AVVulkanFramesContext *hwctx = hwfc->hwctx;
1323     AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1324     VkExternalImageFormatProperties eprops = {
1325         .sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR,
1326     };
1327     VkImageFormatProperties2 props = {
1328         .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2,
1329         .pNext = &eprops,
1330     };
1331     VkPhysicalDeviceExternalImageFormatInfo enext = {
1332         .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO,
1333         .handleType = exp,
1334     };
1335     VkPhysicalDeviceImageFormatInfo2 pinfo = {
1336         .sType  = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2,
1337         .pNext  = !exp ? NULL : &enext,
1338         .format = av_vkfmt_from_pixfmt(hwfc->sw_format)[0],
1339         .type   = VK_IMAGE_TYPE_2D,
1340         .tiling = hwctx->tiling,
1341         .usage  = hwctx->usage,
1342         .flags  = VK_IMAGE_CREATE_ALIAS_BIT,
1343     };
1344
1345     ret = vkGetPhysicalDeviceImageFormatProperties2(dev_hwctx->phys_dev,
1346                                                     &pinfo, &props);
1347     if (ret == VK_SUCCESS) {
1348         *iexp |= exp;
1349         *comp_handle_types |= eprops.externalMemoryProperties.compatibleHandleTypes;
1350     }
1351 }
1352
1353 static AVBufferRef *vulkan_pool_alloc(void *opaque, int size)
1354 {
1355     int err;
1356     AVVkFrame *f;
1357     AVBufferRef *avbuf = NULL;
1358     AVHWFramesContext *hwfc = opaque;
1359     AVVulkanFramesContext *hwctx = hwfc->hwctx;
1360     VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1361     VkExportMemoryAllocateInfo eminfo[AV_NUM_DATA_POINTERS];
1362     VkExternalMemoryHandleTypeFlags e = 0x0;
1363
1364     VkExternalMemoryImageCreateInfo eiinfo = {
1365         .sType       = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1366         .pNext       = hwctx->create_pnext,
1367     };
1368
1369     if (p->extensions & EXT_EXTERNAL_FD_MEMORY)
1370         try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1371                          VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT);
1372
1373     if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
1374         try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1375                          VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1376
1377     for (int i = 0; i < av_pix_fmt_count_planes(hwfc->sw_format); i++) {
1378         eminfo[i].sType       = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO;
1379         eminfo[i].pNext       = hwctx->alloc_pnext[i];
1380         eminfo[i].handleTypes = e;
1381     }
1382
1383     err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1384                        eiinfo.handleTypes ? &eiinfo : NULL);
1385     if (err)
1386         return NULL;
1387
1388     err = alloc_bind_mem(hwfc, f, eminfo, sizeof(*eminfo));
1389     if (err)
1390         goto fail;
1391
1392     err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_WRITE);
1393     if (err)
1394         goto fail;
1395
1396     avbuf = av_buffer_create((uint8_t *)f, sizeof(AVVkFrame),
1397                              vulkan_frame_free, hwfc, 0);
1398     if (!avbuf)
1399         goto fail;
1400
1401     return avbuf;
1402
1403 fail:
1404     vulkan_frame_free(hwfc, (uint8_t *)f);
1405     return NULL;
1406 }
1407
1408 static void vulkan_frames_uninit(AVHWFramesContext *hwfc)
1409 {
1410     VulkanFramesPriv *fp = hwfc->internal->priv;
1411
1412     free_exec_ctx(hwfc->device_ctx, &fp->cmd);
1413 }
1414
1415 static int vulkan_frames_init(AVHWFramesContext *hwfc)
1416 {
1417     int err;
1418     AVVkFrame *f;
1419     AVVulkanFramesContext *hwctx = hwfc->hwctx;
1420     VulkanFramesPriv *fp = hwfc->internal->priv;
1421     AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1422     VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1423
1424     if (hwfc->pool)
1425         return 0;
1426
1427     /* Default pool flags */
1428     hwctx->tiling = hwctx->tiling ? hwctx->tiling : p->use_linear_images ?
1429                     VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1430
1431     hwctx->usage |= DEFAULT_USAGE_FLAGS;
1432
1433     err = create_exec_ctx(hwfc->device_ctx, &fp->cmd,
1434                           dev_hwctx->queue_family_tx_index);
1435     if (err)
1436         return err;
1437
1438     /* Test to see if allocation will fail */
1439     err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1440                        hwctx->create_pnext);
1441     if (err) {
1442         free_exec_ctx(hwfc->device_ctx, &p->cmd);
1443         return err;
1444     }
1445
1446     vulkan_frame_free(hwfc, (uint8_t *)f);
1447
1448     hwfc->internal->pool_internal = av_buffer_pool_init2(sizeof(AVVkFrame),
1449                                                          hwfc, vulkan_pool_alloc,
1450                                                          NULL);
1451     if (!hwfc->internal->pool_internal) {
1452         free_exec_ctx(hwfc->device_ctx, &p->cmd);
1453         return AVERROR(ENOMEM);
1454     }
1455
1456     return 0;
1457 }
1458
1459 static int vulkan_get_buffer(AVHWFramesContext *hwfc, AVFrame *frame)
1460 {
1461     frame->buf[0] = av_buffer_pool_get(hwfc->pool);
1462     if (!frame->buf[0])
1463         return AVERROR(ENOMEM);
1464
1465     frame->data[0] = frame->buf[0]->data;
1466     frame->format  = AV_PIX_FMT_VULKAN;
1467     frame->width   = hwfc->width;
1468     frame->height  = hwfc->height;
1469
1470     return 0;
1471 }
1472
1473 static int vulkan_transfer_get_formats(AVHWFramesContext *hwfc,
1474                                        enum AVHWFrameTransferDirection dir,
1475                                        enum AVPixelFormat **formats)
1476 {
1477     enum AVPixelFormat *fmts = av_malloc_array(2, sizeof(*fmts));
1478     if (!fmts)
1479         return AVERROR(ENOMEM);
1480
1481     fmts[0] = hwfc->sw_format;
1482     fmts[1] = AV_PIX_FMT_NONE;
1483
1484     *formats = fmts;
1485     return 0;
1486 }
1487
1488 typedef struct VulkanMapping {
1489     AVVkFrame *frame;
1490     int flags;
1491 } VulkanMapping;
1492
1493 static void vulkan_unmap_frame(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1494 {
1495     VulkanMapping *map = hwmap->priv;
1496     AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1497     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1498
1499     /* Check if buffer needs flushing */
1500     if ((map->flags & AV_HWFRAME_MAP_WRITE) &&
1501         !(map->frame->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1502         VkResult ret;
1503         VkMappedMemoryRange flush_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1504
1505         for (int i = 0; i < planes; i++) {
1506             flush_ranges[i].sType  = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1507             flush_ranges[i].memory = map->frame->mem[i];
1508             flush_ranges[i].size   = VK_WHOLE_SIZE;
1509         }
1510
1511         ret = vkFlushMappedMemoryRanges(hwctx->act_dev, planes,
1512                                         flush_ranges);
1513         if (ret != VK_SUCCESS) {
1514             av_log(hwfc, AV_LOG_ERROR, "Failed to flush memory: %s\n",
1515                    vk_ret2str(ret));
1516         }
1517     }
1518
1519     for (int i = 0; i < planes; i++)
1520         vkUnmapMemory(hwctx->act_dev, map->frame->mem[i]);
1521
1522     av_free(map);
1523 }
1524
1525 static int vulkan_map_frame_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
1526                                    const AVFrame *src, int flags)
1527 {
1528     VkResult ret;
1529     int err, mapped_mem_count = 0;
1530     AVVkFrame *f = (AVVkFrame *)src->data[0];
1531     AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1532     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1533
1534     VulkanMapping *map = av_mallocz(sizeof(VulkanMapping));
1535     if (!map)
1536         return AVERROR(EINVAL);
1537
1538     if (src->format != AV_PIX_FMT_VULKAN) {
1539         av_log(hwfc, AV_LOG_ERROR, "Cannot map from pixel format %s!\n",
1540                av_get_pix_fmt_name(src->format));
1541         err = AVERROR(EINVAL);
1542         goto fail;
1543     }
1544
1545     if (!(f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) ||
1546         !(f->tiling == VK_IMAGE_TILING_LINEAR)) {
1547         av_log(hwfc, AV_LOG_ERROR, "Unable to map frame, not host visible "
1548                "and linear!\n");
1549         err = AVERROR(EINVAL);
1550         goto fail;
1551     }
1552
1553     dst->width  = src->width;
1554     dst->height = src->height;
1555
1556     for (int i = 0; i < planes; i++) {
1557         ret = vkMapMemory(hwctx->act_dev, f->mem[i], 0,
1558                           VK_WHOLE_SIZE, 0, (void **)&dst->data[i]);
1559         if (ret != VK_SUCCESS) {
1560             av_log(hwfc, AV_LOG_ERROR, "Failed to map image memory: %s\n",
1561                 vk_ret2str(ret));
1562             err = AVERROR_EXTERNAL;
1563             goto fail;
1564         }
1565         mapped_mem_count++;
1566     }
1567
1568     /* Check if the memory contents matter */
1569     if (((flags & AV_HWFRAME_MAP_READ) || !(flags & AV_HWFRAME_MAP_OVERWRITE)) &&
1570         !(f->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1571         VkMappedMemoryRange map_mem_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1572         for (int i = 0; i < planes; i++) {
1573             map_mem_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1574             map_mem_ranges[i].size = VK_WHOLE_SIZE;
1575             map_mem_ranges[i].memory = f->mem[i];
1576         }
1577
1578         ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, planes,
1579                                              map_mem_ranges);
1580         if (ret != VK_SUCCESS) {
1581             av_log(hwfc, AV_LOG_ERROR, "Failed to invalidate memory: %s\n",
1582                    vk_ret2str(ret));
1583             err = AVERROR_EXTERNAL;
1584             goto fail;
1585         }
1586     }
1587
1588     for (int i = 0; i < planes; i++) {
1589         VkImageSubresource sub = {
1590             .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
1591         };
1592         VkSubresourceLayout layout;
1593         vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
1594         dst->linesize[i] = layout.rowPitch;
1595     }
1596
1597     map->frame = f;
1598     map->flags = flags;
1599
1600     err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src,
1601                                 &vulkan_unmap_frame, map);
1602     if (err < 0)
1603         goto fail;
1604
1605     return 0;
1606
1607 fail:
1608     for (int i = 0; i < mapped_mem_count; i++)
1609         vkUnmapMemory(hwctx->act_dev, f->mem[i]);
1610
1611     av_free(map);
1612     return err;
1613 }
1614
1615 #if CONFIG_LIBDRM
1616 static void vulkan_unmap_from(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1617 {
1618     VulkanMapping *map = hwmap->priv;
1619     AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1620     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1621
1622     for (int i = 0; i < planes; i++) {
1623         vkDestroyImage(hwctx->act_dev, map->frame->img[i], hwctx->alloc);
1624         vkFreeMemory(hwctx->act_dev, map->frame->mem[i], hwctx->alloc);
1625         vkDestroySemaphore(hwctx->act_dev, map->frame->sem[i], hwctx->alloc);
1626     }
1627
1628     av_freep(&map->frame);
1629 }
1630
1631 static const struct {
1632     uint32_t drm_fourcc;
1633     VkFormat vk_format;
1634 } vulkan_drm_format_map[] = {
1635     { DRM_FORMAT_R8,       VK_FORMAT_R8_UNORM       },
1636     { DRM_FORMAT_R16,      VK_FORMAT_R16_UNORM      },
1637     { DRM_FORMAT_GR88,     VK_FORMAT_R8G8_UNORM     },
1638     { DRM_FORMAT_RG88,     VK_FORMAT_R8G8_UNORM     },
1639     { DRM_FORMAT_GR1616,   VK_FORMAT_R16G16_UNORM   },
1640     { DRM_FORMAT_RG1616,   VK_FORMAT_R16G16_UNORM   },
1641     { DRM_FORMAT_ARGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1642     { DRM_FORMAT_XRGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1643     { DRM_FORMAT_ABGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1644     { DRM_FORMAT_XBGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1645 };
1646
1647 static inline VkFormat drm_to_vulkan_fmt(uint32_t drm_fourcc)
1648 {
1649     for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
1650         if (vulkan_drm_format_map[i].drm_fourcc == drm_fourcc)
1651             return vulkan_drm_format_map[i].vk_format;
1652     return VK_FORMAT_UNDEFINED;
1653 }
1654
1655 static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **frame,
1656                                           AVDRMFrameDescriptor *desc)
1657 {
1658     int err = 0;
1659     VkResult ret;
1660     AVVkFrame *f;
1661     AVHWDeviceContext *ctx = hwfc->device_ctx;
1662     AVVulkanDeviceContext *hwctx = ctx->hwctx;
1663     VulkanDevicePriv *p = ctx->internal->priv;
1664     const AVPixFmtDescriptor *fmt_desc = av_pix_fmt_desc_get(hwfc->sw_format);
1665     const int has_modifiers = p->extensions & EXT_DRM_MODIFIER_FLAGS;
1666     VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS];
1667     VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS];
1668     VkExternalMemoryHandleTypeFlagBits htype = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT;
1669
1670     VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdPropertiesKHR);
1671
1672     for (int i = 0; i < desc->nb_layers; i++) {
1673         if (desc->layers[i].nb_planes > 1) {
1674             av_log(ctx, AV_LOG_ERROR, "Cannot import DMABUFS with more than 1 "
1675                                       "plane per layer!\n");
1676             return AVERROR(EINVAL);
1677         }
1678
1679         if (drm_to_vulkan_fmt(desc->layers[i].format) == VK_FORMAT_UNDEFINED) {
1680             av_log(ctx, AV_LOG_ERROR, "Unsupported DMABUF layer format %#08x!\n",
1681                    desc->layers[i].format);
1682             return AVERROR(EINVAL);
1683         }
1684     }
1685
1686     if (!(f = av_vk_frame_alloc())) {
1687         av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1688         err = AVERROR(ENOMEM);
1689         goto fail;
1690     }
1691
1692     for (int i = 0; i < desc->nb_objects; i++) {
1693         VkMemoryFdPropertiesKHR fdmp = {
1694             .sType = VK_STRUCTURE_TYPE_MEMORY_FD_PROPERTIES_KHR,
1695         };
1696         VkMemoryRequirements req = {
1697             .size = desc->objects[i].size,
1698         };
1699         VkImportMemoryFdInfoKHR idesc = {
1700             .sType      = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR,
1701             .handleType = htype,
1702             .fd         = dup(desc->objects[i].fd),
1703         };
1704
1705         ret = pfn_vkGetMemoryFdPropertiesKHR(hwctx->act_dev, htype,
1706                                              idesc.fd, &fdmp);
1707         if (ret != VK_SUCCESS) {
1708             av_log(hwfc, AV_LOG_ERROR, "Failed to get FD properties: %s\n",
1709                    vk_ret2str(ret));
1710             err = AVERROR_EXTERNAL;
1711             close(idesc.fd);
1712             goto fail;
1713         }
1714
1715         req.memoryTypeBits = fdmp.memoryTypeBits;
1716
1717         err = alloc_mem(ctx, &req, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1718                         &idesc, &f->flags, &f->mem[i]);
1719         if (err) {
1720             close(idesc.fd);
1721             return err;
1722         }
1723
1724         f->size[i] = desc->objects[i].size;
1725     }
1726
1727     f->tiling = has_modifiers ? VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT :
1728                 desc->objects[0].format_modifier == DRM_FORMAT_MOD_LINEAR ?
1729                 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1730
1731     for (int i = 0; i < desc->nb_layers; i++) {
1732         VkImageDrmFormatModifierExplicitCreateInfoEXT drm_info = {
1733             .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT,
1734             .drmFormatModifier = desc->objects[0].format_modifier,
1735             .drmFormatModifierPlaneCount = desc->layers[i].nb_planes,
1736             .pPlaneLayouts = (const VkSubresourceLayout *)&plane_data,
1737         };
1738
1739         VkExternalMemoryImageCreateInfo einfo = {
1740             .sType       = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1741             .pNext       = has_modifiers ? &drm_info : NULL,
1742             .handleTypes = htype,
1743         };
1744
1745         VkSemaphoreCreateInfo sem_spawn = {
1746             .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1747         };
1748
1749         const int p_w = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, fmt_desc->log2_chroma_w) : hwfc->width;
1750         const int p_h = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, fmt_desc->log2_chroma_h) : hwfc->height;
1751
1752         VkImageCreateInfo image_create_info = {
1753             .sType         = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1754             .pNext         = &einfo,
1755             .imageType     = VK_IMAGE_TYPE_2D,
1756             .format        = drm_to_vulkan_fmt(desc->layers[i].format),
1757             .extent.width  = p_w,
1758             .extent.height = p_h,
1759             .extent.depth  = 1,
1760             .mipLevels     = 1,
1761             .arrayLayers   = 1,
1762             .flags         = VK_IMAGE_CREATE_ALIAS_BIT,
1763             .tiling        = f->tiling,
1764             .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */
1765             .usage         = DEFAULT_USAGE_FLAGS,
1766             .sharingMode   = VK_SHARING_MODE_EXCLUSIVE,
1767             .samples       = VK_SAMPLE_COUNT_1_BIT,
1768         };
1769
1770         for (int j = 0; j < desc->layers[i].nb_planes; j++) {
1771             plane_data[j].offset     = desc->layers[i].planes[j].offset;
1772             plane_data[j].rowPitch   = desc->layers[i].planes[j].pitch;
1773             plane_data[j].size       = 0; /* The specs say so for all 3 */
1774             plane_data[j].arrayPitch = 0;
1775             plane_data[j].depthPitch = 0;
1776         }
1777
1778         /* Create image */
1779         ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1780                             hwctx->alloc, &f->img[i]);
1781         if (ret != VK_SUCCESS) {
1782             av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1783                    vk_ret2str(ret));
1784             err = AVERROR(EINVAL);
1785             goto fail;
1786         }
1787
1788         ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1789                                 hwctx->alloc, &f->sem[i]);
1790         if (ret != VK_SUCCESS) {
1791             av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1792                    vk_ret2str(ret));
1793             return AVERROR_EXTERNAL;
1794         }
1795
1796         /* We'd import a semaphore onto the one we created using
1797          * vkImportSemaphoreFdKHR but unfortunately neither DRM nor VAAPI
1798          * offer us anything we could import and sync with, so instead
1799          * just signal the semaphore we created. */
1800
1801         f->layout[i] = image_create_info.initialLayout;
1802         f->access[i] = 0x0;
1803
1804         /* TODO: Fix to support more than 1 plane per layer */
1805         bind_info[i].sType  = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1806         bind_info[i].pNext  = NULL;
1807         bind_info[i].image  = f->img[i];
1808         bind_info[i].memory = f->mem[desc->layers[i].planes[0].object_index];
1809         bind_info[i].memoryOffset = desc->layers[i].planes[0].offset;
1810     }
1811
1812     /* Bind the allocated memory to the images */
1813     ret = vkBindImageMemory2(hwctx->act_dev, desc->nb_layers, bind_info);
1814     if (ret != VK_SUCCESS) {
1815         av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1816                vk_ret2str(ret));
1817         return AVERROR_EXTERNAL;
1818     }
1819
1820     /* NOTE: This is completely uneccesary and unneeded once we can import
1821      * semaphores from DRM. Otherwise we have to activate the semaphores.
1822      * We're reusing the exec context that's also used for uploads/downloads. */
1823     err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_RO_SHADER);
1824     if (err)
1825         goto fail;
1826
1827     *frame = f;
1828
1829     return 0;
1830
1831 fail:
1832     for (int i = 0; i < desc->nb_layers; i++) {
1833         vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1834         vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
1835     }
1836     for (int i = 0; i < desc->nb_objects; i++)
1837         vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1838
1839     av_free(f);
1840
1841     return err;
1842 }
1843
1844 static int vulkan_map_from_drm(AVHWFramesContext *hwfc, AVFrame *dst,
1845                                const AVFrame *src, int flags)
1846 {
1847     int err = 0;
1848     AVVkFrame *f;
1849     VulkanMapping *map = NULL;
1850
1851     err = vulkan_map_from_drm_frame_desc(hwfc, &f,
1852                                          (AVDRMFrameDescriptor *)src->data[0]);
1853     if (err)
1854         return err;
1855
1856     /* The unmapping function will free this */
1857     dst->data[0] = (uint8_t *)f;
1858     dst->width   = src->width;
1859     dst->height  = src->height;
1860
1861     map = av_mallocz(sizeof(VulkanMapping));
1862     if (!map)
1863         goto fail;
1864
1865     map->frame = f;
1866     map->flags = flags;
1867
1868     err = ff_hwframe_map_create(dst->hw_frames_ctx, dst, src,
1869                                 &vulkan_unmap_from, map);
1870     if (err < 0)
1871         goto fail;
1872
1873     av_log(hwfc, AV_LOG_DEBUG, "Mapped DRM object to Vulkan!\n");
1874
1875     return 0;
1876
1877 fail:
1878     vulkan_frame_free(hwfc->device_ctx->hwctx, (uint8_t *)f);
1879     av_free(map);
1880     return err;
1881 }
1882
1883 #if CONFIG_VAAPI
1884 static int vulkan_map_from_vaapi(AVHWFramesContext *dst_fc,
1885                                  AVFrame *dst, const AVFrame *src,
1886                                  int flags)
1887 {
1888     int err;
1889     AVFrame *tmp = av_frame_alloc();
1890     AVHWFramesContext *vaapi_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
1891     AVVAAPIDeviceContext *vaapi_ctx = vaapi_fc->device_ctx->hwctx;
1892     VASurfaceID surface_id = (VASurfaceID)(uintptr_t)src->data[3];
1893
1894     if (!tmp)
1895         return AVERROR(ENOMEM);
1896
1897     /* We have to sync since like the previous comment said, no semaphores */
1898     vaSyncSurface(vaapi_ctx->display, surface_id);
1899
1900     tmp->format = AV_PIX_FMT_DRM_PRIME;
1901
1902     err = av_hwframe_map(tmp, src, flags);
1903     if (err < 0)
1904         goto fail;
1905
1906     err = vulkan_map_from_drm(dst_fc, dst, tmp, flags);
1907     if (err < 0)
1908         goto fail;
1909
1910     err = ff_hwframe_map_replace(dst, src);
1911
1912 fail:
1913     av_frame_free(&tmp);
1914     return err;
1915 }
1916 #endif
1917 #endif
1918
1919 #if CONFIG_CUDA
1920 static int vulkan_export_to_cuda(AVHWFramesContext *hwfc,
1921                                  AVBufferRef *cuda_hwfc,
1922                                  const AVFrame *frame)
1923 {
1924     int err;
1925     VkResult ret;
1926     AVVkFrame *dst_f;
1927     AVVkFrameInternal *dst_int;
1928     AVHWDeviceContext *ctx = hwfc->device_ctx;
1929     AVVulkanDeviceContext *hwctx = ctx->hwctx;
1930     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1931     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
1932     VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
1933     VK_LOAD_PFN(hwctx->inst, vkGetSemaphoreFdKHR);
1934
1935     AVHWFramesContext *cuda_fc = (AVHWFramesContext*)cuda_hwfc->data;
1936     AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
1937     AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
1938     AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
1939     CudaFunctions *cu = cu_internal->cuda_dl;
1940     CUarray_format cufmt = desc->comp[0].depth > 8 ? CU_AD_FORMAT_UNSIGNED_INT16 :
1941                                                      CU_AD_FORMAT_UNSIGNED_INT8;
1942
1943     dst_f = (AVVkFrame *)frame->data[0];
1944
1945     dst_int = dst_f->internal;
1946     if (!dst_int || !dst_int->cuda_fc_ref) {
1947         if (!dst_f->internal)
1948             dst_f->internal = dst_int = av_mallocz(sizeof(*dst_f->internal));
1949
1950         if (!dst_int) {
1951             err = AVERROR(ENOMEM);
1952             goto fail;
1953         }
1954
1955         dst_int->cuda_fc_ref = av_buffer_ref(cuda_hwfc);
1956         if (!dst_int->cuda_fc_ref) {
1957             err = AVERROR(ENOMEM);
1958             goto fail;
1959         }
1960
1961         for (int i = 0; i < planes; i++) {
1962             CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC tex_desc = {
1963                 .offset = 0,
1964                 .arrayDesc = {
1965                     .Width  = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
1966                                     : hwfc->width,
1967                     .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
1968                                     : hwfc->height,
1969                     .Depth = 0,
1970                     .Format = cufmt,
1971                     .NumChannels = 1 + ((planes == 2) && i),
1972                     .Flags = 0,
1973                 },
1974                 .numLevels = 1,
1975             };
1976             CUDA_EXTERNAL_MEMORY_HANDLE_DESC ext_desc = {
1977                 .type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD,
1978                 .size = dst_f->size[i],
1979             };
1980             VkMemoryGetFdInfoKHR export_info = {
1981                 .sType      = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
1982                 .memory     = dst_f->mem[i],
1983                 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR,
1984             };
1985             VkSemaphoreGetFdInfoKHR sem_export = {
1986                 .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
1987                 .semaphore = dst_f->sem[i],
1988                 .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
1989             };
1990             CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = {
1991                 .type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
1992             };
1993
1994             ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
1995                                        &ext_desc.handle.fd);
1996             if (ret != VK_SUCCESS) {
1997                 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
1998                 err = AVERROR_EXTERNAL;
1999                 goto fail;
2000             }
2001
2002             ret = CHECK_CU(cu->cuImportExternalMemory(&dst_int->ext_mem[i], &ext_desc));
2003             if (ret < 0) {
2004                 err = AVERROR_EXTERNAL;
2005                 goto fail;
2006             }
2007
2008             ret = CHECK_CU(cu->cuExternalMemoryGetMappedMipmappedArray(&dst_int->cu_mma[i],
2009                                                                        dst_int->ext_mem[i],
2010                                                                        &tex_desc));
2011             if (ret < 0) {
2012                 err = AVERROR_EXTERNAL;
2013                 goto fail;
2014             }
2015
2016             ret = CHECK_CU(cu->cuMipmappedArrayGetLevel(&dst_int->cu_array[i],
2017                                                         dst_int->cu_mma[i], 0));
2018             if (ret < 0) {
2019                 err = AVERROR_EXTERNAL;
2020                 goto fail;
2021             }
2022
2023             ret = pfn_vkGetSemaphoreFdKHR(hwctx->act_dev, &sem_export,
2024                                           &ext_sem_desc.handle.fd);
2025             if (ret != VK_SUCCESS) {
2026                 av_log(ctx, AV_LOG_ERROR, "Failed to export semaphore: %s\n",
2027                        vk_ret2str(ret));
2028                 err = AVERROR_EXTERNAL;
2029                 goto fail;
2030             }
2031
2032             ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem[i],
2033                                                          &ext_sem_desc));
2034             if (ret < 0) {
2035                 err = AVERROR_EXTERNAL;
2036                 goto fail;
2037             }
2038         }
2039     }
2040
2041     return 0;
2042
2043 fail:
2044     return err;
2045 }
2046
2047 static int vulkan_transfer_data_from_cuda(AVHWFramesContext *hwfc,
2048                                           AVFrame *dst, const AVFrame *src)
2049 {
2050     int err;
2051     VkResult ret;
2052     CUcontext dummy;
2053     AVVkFrame *dst_f;
2054     AVVkFrameInternal *dst_int;
2055     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2056     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2057
2058     AVHWFramesContext *cuda_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2059     AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2060     AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2061     AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2062     CudaFunctions *cu = cu_internal->cuda_dl;
2063     CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS] = { 0 };
2064     CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS] = { 0 };
2065
2066     ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2067     if (ret < 0) {
2068         err = AVERROR_EXTERNAL;
2069         goto fail;
2070     }
2071
2072     dst_f = (AVVkFrame *)dst->data[0];
2073
2074     ret = vulkan_export_to_cuda(hwfc, src->hw_frames_ctx, dst);
2075     if (ret < 0) {
2076         goto fail;
2077     }
2078     dst_int = dst_f->internal;
2079
2080     ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par,
2081                                                      planes, cuda_dev->stream));
2082     if (ret < 0) {
2083         err = AVERROR_EXTERNAL;
2084         goto fail;
2085     }
2086
2087     for (int i = 0; i < planes; i++) {
2088         CUDA_MEMCPY2D cpy = {
2089             .srcMemoryType = CU_MEMORYTYPE_DEVICE,
2090             .srcDevice     = (CUdeviceptr)src->data[i],
2091             .srcPitch      = src->linesize[i],
2092             .srcY          = 0,
2093
2094             .dstMemoryType = CU_MEMORYTYPE_ARRAY,
2095             .dstArray      = dst_int->cu_array[i],
2096             .WidthInBytes  = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2097                                     : hwfc->width) * desc->comp[i].step,
2098             .Height        = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2099                                    : hwfc->height,
2100         };
2101
2102         ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2103         if (ret < 0) {
2104             err = AVERROR_EXTERNAL;
2105             goto fail;
2106         }
2107     }
2108
2109     ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par,
2110                                                        planes, cuda_dev->stream));
2111     if (ret < 0) {
2112         err = AVERROR_EXTERNAL;
2113         goto fail;
2114     }
2115
2116     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2117
2118     av_log(hwfc, AV_LOG_VERBOSE, "Transfered CUDA image to Vulkan!\n");
2119
2120     return 0;
2121
2122 fail:
2123     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2124     vulkan_free_internal(dst_int);
2125     dst_f->internal = NULL;
2126     av_buffer_unref(&dst->buf[0]);
2127     return err;
2128 }
2129 #endif
2130
2131 static int vulkan_map_to(AVHWFramesContext *hwfc, AVFrame *dst,
2132                          const AVFrame *src, int flags)
2133 {
2134     av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2135
2136     switch (src->format) {
2137 #if CONFIG_LIBDRM
2138 #if CONFIG_VAAPI
2139     case AV_PIX_FMT_VAAPI:
2140         if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2141             return vulkan_map_from_vaapi(hwfc, dst, src, flags);
2142 #endif
2143     case AV_PIX_FMT_DRM_PRIME:
2144         if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2145             return vulkan_map_from_drm(hwfc, dst, src, flags);
2146 #endif
2147     default:
2148         return AVERROR(ENOSYS);
2149     }
2150 }
2151
2152 #if CONFIG_LIBDRM
2153 typedef struct VulkanDRMMapping {
2154     AVDRMFrameDescriptor drm_desc;
2155     AVVkFrame *source;
2156 } VulkanDRMMapping;
2157
2158 static void vulkan_unmap_to_drm(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
2159 {
2160     AVDRMFrameDescriptor *drm_desc = hwmap->priv;
2161
2162     for (int i = 0; i < drm_desc->nb_objects; i++)
2163         close(drm_desc->objects[i].fd);
2164
2165     av_free(drm_desc);
2166 }
2167
2168 static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt)
2169 {
2170     for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
2171         if (vulkan_drm_format_map[i].vk_format == vkfmt)
2172             return vulkan_drm_format_map[i].drm_fourcc;
2173     return DRM_FORMAT_INVALID;
2174 }
2175
2176 static int vulkan_map_to_drm(AVHWFramesContext *hwfc, AVFrame *dst,
2177                              const AVFrame *src, int flags)
2178 {
2179     int err = 0;
2180     VkResult ret;
2181     AVVkFrame *f = (AVVkFrame *)src->data[0];
2182     VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2183     AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
2184     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2185     VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2186     VkImageDrmFormatModifierPropertiesEXT drm_mod = {
2187         .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_PROPERTIES_EXT,
2188     };
2189
2190     AVDRMFrameDescriptor *drm_desc = av_mallocz(sizeof(*drm_desc));
2191     if (!drm_desc)
2192         return AVERROR(ENOMEM);
2193
2194     err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src, &vulkan_unmap_to_drm, drm_desc);
2195     if (err < 0)
2196         goto end;
2197
2198     if (p->extensions & EXT_DRM_MODIFIER_FLAGS) {
2199         VK_LOAD_PFN(hwctx->inst, vkGetImageDrmFormatModifierPropertiesEXT);
2200         ret = pfn_vkGetImageDrmFormatModifierPropertiesEXT(hwctx->act_dev, f->img[0],
2201                                                            &drm_mod);
2202         if (ret != VK_SUCCESS) {
2203             av_log(hwfc, AV_LOG_ERROR, "Failed to retrieve DRM format modifier!\n");
2204             err = AVERROR_EXTERNAL;
2205             goto end;
2206         }
2207     }
2208
2209     for (int i = 0; (i < planes) && (f->mem[i]); i++) {
2210         VkMemoryGetFdInfoKHR export_info = {
2211             .sType      = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2212             .memory     = f->mem[i],
2213             .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT,
2214         };
2215
2216         ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2217                                    &drm_desc->objects[i].fd);
2218         if (ret != VK_SUCCESS) {
2219             av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2220             err = AVERROR_EXTERNAL;
2221             goto end;
2222         }
2223
2224         drm_desc->nb_objects++;
2225         drm_desc->objects[i].size = f->size[i];
2226         drm_desc->objects[i].format_modifier = drm_mod.drmFormatModifier;
2227     }
2228
2229     drm_desc->nb_layers = planes;
2230     for (int i = 0; i < drm_desc->nb_layers; i++) {
2231         VkSubresourceLayout layout;
2232         VkImageSubresource sub = {
2233             .aspectMask = p->extensions & EXT_DRM_MODIFIER_FLAGS ?
2234                           VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
2235                           VK_IMAGE_ASPECT_COLOR_BIT,
2236         };
2237         VkFormat plane_vkfmt = av_vkfmt_from_pixfmt(hwfc->sw_format)[i];
2238
2239         drm_desc->layers[i].format    = vulkan_fmt_to_drm(plane_vkfmt);
2240         drm_desc->layers[i].nb_planes = 1;
2241
2242         if (drm_desc->layers[i].format == DRM_FORMAT_INVALID) {
2243             av_log(hwfc, AV_LOG_ERROR, "Cannot map to DRM layer, unsupported!\n");
2244             err = AVERROR_PATCHWELCOME;
2245             goto end;
2246         }
2247
2248         drm_desc->layers[i].planes[0].object_index = FFMIN(i, drm_desc->nb_objects - 1);
2249
2250         if (f->tiling != VK_IMAGE_TILING_OPTIMAL)
2251             continue;
2252
2253         vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
2254         drm_desc->layers[i].planes[0].offset       = layout.offset;
2255         drm_desc->layers[i].planes[0].pitch        = layout.rowPitch;
2256     }
2257
2258     dst->width   = src->width;
2259     dst->height  = src->height;
2260     dst->data[0] = (uint8_t *)drm_desc;
2261
2262     av_log(hwfc, AV_LOG_VERBOSE, "Mapped AVVkFrame to a DRM object!\n");
2263
2264     return 0;
2265
2266 end:
2267     av_free(drm_desc);
2268     return err;
2269 }
2270
2271 #if CONFIG_VAAPI
2272 static int vulkan_map_to_vaapi(AVHWFramesContext *hwfc, AVFrame *dst,
2273                                const AVFrame *src, int flags)
2274 {
2275     int err;
2276     AVFrame *tmp = av_frame_alloc();
2277     if (!tmp)
2278         return AVERROR(ENOMEM);
2279
2280     tmp->format = AV_PIX_FMT_DRM_PRIME;
2281
2282     err = vulkan_map_to_drm(hwfc, tmp, src, flags);
2283     if (err < 0)
2284         goto fail;
2285
2286     err = av_hwframe_map(dst, tmp, flags);
2287     if (err < 0)
2288         goto fail;
2289
2290     err = ff_hwframe_map_replace(dst, src);
2291
2292 fail:
2293     av_frame_free(&tmp);
2294     return err;
2295 }
2296 #endif
2297 #endif
2298
2299 static int vulkan_map_from(AVHWFramesContext *hwfc, AVFrame *dst,
2300                            const AVFrame *src, int flags)
2301 {
2302     av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2303
2304     switch (dst->format) {
2305 #if CONFIG_LIBDRM
2306     case AV_PIX_FMT_DRM_PRIME:
2307         if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2308             return vulkan_map_to_drm(hwfc, dst, src, flags);
2309 #if CONFIG_VAAPI
2310     case AV_PIX_FMT_VAAPI:
2311         if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2312             return vulkan_map_to_vaapi(hwfc, dst, src, flags);
2313 #endif
2314 #endif
2315     default:
2316         return vulkan_map_frame_to_mem(hwfc, dst, src, flags);
2317     }
2318 }
2319
2320 typedef struct ImageBuffer {
2321     VkBuffer buf;
2322     VkDeviceMemory mem;
2323     VkMemoryPropertyFlagBits flags;
2324 } ImageBuffer;
2325
2326 static void free_buf(AVHWDeviceContext *ctx, ImageBuffer *buf)
2327 {
2328     AVVulkanDeviceContext *hwctx = ctx->hwctx;
2329     if (!buf)
2330         return;
2331
2332     vkDestroyBuffer(hwctx->act_dev, buf->buf, hwctx->alloc);
2333     vkFreeMemory(hwctx->act_dev, buf->mem, hwctx->alloc);
2334 }
2335
2336 static int create_buf(AVHWDeviceContext *ctx, ImageBuffer *buf, int height,
2337                       int *stride, VkBufferUsageFlags usage,
2338                       VkMemoryPropertyFlagBits flags, void *create_pnext,
2339                       void *alloc_pnext)
2340 {
2341     int err;
2342     VkResult ret;
2343     VkMemoryRequirements req;
2344     AVVulkanDeviceContext *hwctx = ctx->hwctx;
2345     VulkanDevicePriv *p = ctx->internal->priv;
2346
2347     VkBufferCreateInfo buf_spawn = {
2348         .sType       = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
2349         .pNext       = create_pnext,
2350         .usage       = usage,
2351         .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
2352     };
2353
2354     *stride = FFALIGN(*stride, p->props.limits.optimalBufferCopyRowPitchAlignment);
2355     buf_spawn.size = height*(*stride);
2356
2357     ret = vkCreateBuffer(hwctx->act_dev, &buf_spawn, NULL, &buf->buf);
2358     if (ret != VK_SUCCESS) {
2359         av_log(ctx, AV_LOG_ERROR, "Failed to create buffer: %s\n",
2360                vk_ret2str(ret));
2361         return AVERROR_EXTERNAL;
2362     }
2363
2364     vkGetBufferMemoryRequirements(hwctx->act_dev, buf->buf, &req);
2365
2366     err = alloc_mem(ctx, &req, flags, alloc_pnext, &buf->flags, &buf->mem);
2367     if (err)
2368         return err;
2369
2370     ret = vkBindBufferMemory(hwctx->act_dev, buf->buf, buf->mem, 0);
2371     if (ret != VK_SUCCESS) {
2372         av_log(ctx, AV_LOG_ERROR, "Failed to bind memory to buffer: %s\n",
2373                vk_ret2str(ret));
2374         free_buf(ctx, buf);
2375         return AVERROR_EXTERNAL;
2376     }
2377
2378     return 0;
2379 }
2380
2381 static int map_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf, uint8_t *mem[],
2382                        int nb_buffers, int invalidate)
2383 {
2384     VkResult ret;
2385     AVVulkanDeviceContext *hwctx = ctx->hwctx;
2386     VkMappedMemoryRange invalidate_ctx[AV_NUM_DATA_POINTERS];
2387     int invalidate_count = 0;
2388
2389     for (int i = 0; i < nb_buffers; i++) {
2390         ret = vkMapMemory(hwctx->act_dev, buf[i].mem, 0,
2391                           VK_WHOLE_SIZE, 0, (void **)&mem[i]);
2392         if (ret != VK_SUCCESS) {
2393             av_log(ctx, AV_LOG_ERROR, "Failed to map buffer memory: %s\n",
2394                    vk_ret2str(ret));
2395             return AVERROR_EXTERNAL;
2396         }
2397     }
2398
2399     if (!invalidate)
2400         return 0;
2401
2402     for (int i = 0; i < nb_buffers; i++) {
2403         const VkMappedMemoryRange ival_buf = {
2404             .sType  = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2405             .memory = buf[i].mem,
2406             .size   = VK_WHOLE_SIZE,
2407         };
2408         if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2409             continue;
2410         invalidate_ctx[invalidate_count++] = ival_buf;
2411     }
2412
2413     if (invalidate_count) {
2414         ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, invalidate_count,
2415                                              invalidate_ctx);
2416         if (ret != VK_SUCCESS)
2417             av_log(ctx, AV_LOG_WARNING, "Failed to invalidate memory: %s\n",
2418                    vk_ret2str(ret));
2419     }
2420
2421     return 0;
2422 }
2423
2424 static int unmap_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf,
2425                          int nb_buffers, int flush)
2426 {
2427     int err = 0;
2428     VkResult ret;
2429     AVVulkanDeviceContext *hwctx = ctx->hwctx;
2430     VkMappedMemoryRange flush_ctx[AV_NUM_DATA_POINTERS];
2431     int flush_count = 0;
2432
2433     if (flush) {
2434         for (int i = 0; i < nb_buffers; i++) {
2435             const VkMappedMemoryRange flush_buf = {
2436                 .sType  = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2437                 .memory = buf[i].mem,
2438                 .size   = VK_WHOLE_SIZE,
2439             };
2440             if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2441                 continue;
2442             flush_ctx[flush_count++] = flush_buf;
2443         }
2444     }
2445
2446     if (flush_count) {
2447         ret = vkFlushMappedMemoryRanges(hwctx->act_dev, flush_count, flush_ctx);
2448         if (ret != VK_SUCCESS) {
2449             av_log(ctx, AV_LOG_ERROR, "Failed to flush memory: %s\n",
2450                     vk_ret2str(ret));
2451             err = AVERROR_EXTERNAL; /* We still want to try to unmap them */
2452         }
2453     }
2454
2455     for (int i = 0; i < nb_buffers; i++)
2456         vkUnmapMemory(hwctx->act_dev, buf[i].mem);
2457
2458     return err;
2459 }
2460
2461 static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
2462                               ImageBuffer *buffer, const int *buf_stride, int w,
2463                               int h, enum AVPixelFormat pix_fmt, int to_buf)
2464 {
2465     VkResult ret;
2466     AVVulkanDeviceContext *hwctx = ctx->hwctx;
2467     VulkanDevicePriv *s = ctx->internal->priv;
2468
2469     int bar_num = 0;
2470     VkPipelineStageFlagBits sem_wait_dst[AV_NUM_DATA_POINTERS];
2471
2472     const int planes = av_pix_fmt_count_planes(pix_fmt);
2473     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
2474
2475     VkCommandBufferBeginInfo cmd_start = {
2476         .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
2477         .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
2478     };
2479
2480     VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
2481
2482     VkSubmitInfo s_info = {
2483         .sType                = VK_STRUCTURE_TYPE_SUBMIT_INFO,
2484         .commandBufferCount   = 1,
2485         .pCommandBuffers      = &s->cmd.buf,
2486         .pSignalSemaphores    = frame->sem,
2487         .pWaitSemaphores      = frame->sem,
2488         .pWaitDstStageMask    = sem_wait_dst,
2489         .signalSemaphoreCount = planes,
2490         .waitSemaphoreCount   = planes,
2491     };
2492
2493     ret = vkBeginCommandBuffer(s->cmd.buf, &cmd_start);
2494     if (ret != VK_SUCCESS) {
2495         av_log(ctx, AV_LOG_ERROR, "Unable to init command buffer: %s\n",
2496                vk_ret2str(ret));
2497         return AVERROR_EXTERNAL;
2498     }
2499
2500     /* Change the image layout to something more optimal for transfers */
2501     for (int i = 0; i < planes; i++) {
2502         VkImageLayout new_layout = to_buf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL :
2503                                             VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
2504         VkAccessFlags new_access = to_buf ? VK_ACCESS_TRANSFER_READ_BIT :
2505                                             VK_ACCESS_TRANSFER_WRITE_BIT;
2506
2507         sem_wait_dst[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
2508
2509         /* If the layout matches and we have read access skip the barrier */
2510         if ((frame->layout[i] == new_layout) && (frame->access[i] & new_access))
2511             continue;
2512
2513         img_bar[bar_num].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
2514         img_bar[bar_num].srcAccessMask = 0x0;
2515         img_bar[bar_num].dstAccessMask = new_access;
2516         img_bar[bar_num].oldLayout = frame->layout[i];
2517         img_bar[bar_num].newLayout = new_layout;
2518         img_bar[bar_num].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2519         img_bar[bar_num].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2520         img_bar[bar_num].image = frame->img[i];
2521         img_bar[bar_num].subresourceRange.levelCount = 1;
2522         img_bar[bar_num].subresourceRange.layerCount = 1;
2523         img_bar[bar_num].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
2524
2525         frame->layout[i] = img_bar[bar_num].newLayout;
2526         frame->access[i] = img_bar[bar_num].dstAccessMask;
2527
2528         bar_num++;
2529     }
2530
2531     if (bar_num)
2532         vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
2533                              VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
2534                              0, NULL, 0, NULL, bar_num, img_bar);
2535
2536     /* Schedule a copy for each plane */
2537     for (int i = 0; i < planes; i++) {
2538         const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
2539         const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
2540         VkBufferImageCopy buf_reg = {
2541             .bufferOffset = 0,
2542             /* Buffer stride isn't in bytes, it's in samples, the implementation
2543              * uses the image's VkFormat to know how many bytes per sample
2544              * the buffer has. So we have to convert by dividing. Stupid.
2545              * Won't work with YUVA or other planar formats with alpha. */
2546             .bufferRowLength = buf_stride[i] / desc->comp[i].step,
2547             .bufferImageHeight = p_h,
2548             .imageSubresource.layerCount = 1,
2549             .imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
2550             .imageOffset = { 0, 0, 0, },
2551             .imageExtent = { p_w, p_h, 1, },
2552         };
2553
2554         if (to_buf)
2555             vkCmdCopyImageToBuffer(s->cmd.buf, frame->img[i], frame->layout[i],
2556                                    buffer[i].buf, 1, &buf_reg);
2557         else
2558             vkCmdCopyBufferToImage(s->cmd.buf, buffer[i].buf, frame->img[i],
2559                                    frame->layout[i], 1, &buf_reg);
2560     }
2561
2562     ret = vkEndCommandBuffer(s->cmd.buf);
2563     if (ret != VK_SUCCESS) {
2564         av_log(ctx, AV_LOG_ERROR, "Unable to finish command buffer: %s\n",
2565                vk_ret2str(ret));
2566         return AVERROR_EXTERNAL;
2567     }
2568
2569     /* Wait for the download/upload to finish if uploading, otherwise the
2570      * semaphore will take care of synchronization when uploading */
2571     ret = vkQueueSubmit(s->cmd.queue, 1, &s_info, s->cmd.fence);
2572     if (ret != VK_SUCCESS) {
2573         av_log(ctx, AV_LOG_ERROR, "Unable to submit command buffer: %s\n",
2574                vk_ret2str(ret));
2575         return AVERROR_EXTERNAL;
2576     } else {
2577         vkWaitForFences(hwctx->act_dev, 1, &s->cmd.fence, VK_TRUE, UINT64_MAX);
2578         vkResetFences(hwctx->act_dev, 1, &s->cmd.fence);
2579     }
2580
2581     return 0;
2582 }
2583
2584 /* Technically we can use VK_EXT_external_memory_host to upload and download,
2585  * however the alignment requirements make this unfeasible as both the pointer
2586  * and the size of each plane need to be aligned to the minimum alignment
2587  * requirement, which on all current implementations (anv, radv) is 4096.
2588  * If the requirement gets relaxed (unlikely) this can easily be implemented. */
2589 static int vulkan_transfer_data_from_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2590                                          const AVFrame *src)
2591 {
2592     int err = 0;
2593     AVFrame tmp;
2594     AVVkFrame *f = (AVVkFrame *)dst->data[0];
2595     AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2596     ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
2597     const int planes = av_pix_fmt_count_planes(src->format);
2598     int log2_chroma = av_pix_fmt_desc_get(src->format)->log2_chroma_h;
2599
2600     if ((src->format != AV_PIX_FMT_NONE && !av_vkfmt_from_pixfmt(src->format))) {
2601         av_log(hwfc, AV_LOG_ERROR, "Unsupported source pixel format!\n");
2602         return AVERROR(EINVAL);
2603     }
2604
2605     if (src->width > hwfc->width || src->height > hwfc->height)
2606         return AVERROR(EINVAL);
2607
2608     /* For linear, host visiable images */
2609     if (f->tiling == VK_IMAGE_TILING_LINEAR &&
2610         f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
2611         AVFrame *map = av_frame_alloc();
2612         if (!map)
2613             return AVERROR(ENOMEM);
2614         map->format = src->format;
2615
2616         err = vulkan_map_frame_to_mem(hwfc, map, dst, AV_HWFRAME_MAP_WRITE);
2617         if (err)
2618             goto end;
2619
2620         err = av_frame_copy(map, src);
2621         av_frame_free(&map);
2622         goto end;
2623     }
2624
2625     /* Create buffers */
2626     for (int i = 0; i < planes; i++) {
2627         int h = src->height;
2628         int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
2629
2630         tmp.linesize[i] = src->linesize[i];
2631         err = create_buf(dev_ctx, &buf[i], p_height,
2632                          &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
2633                          VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
2634         if (err)
2635             goto end;
2636     }
2637
2638     /* Map, copy image to buffer, unmap */
2639     if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 0)))
2640         goto end;
2641
2642     av_image_copy(tmp.data, tmp.linesize, (const uint8_t **)src->data,
2643                   src->linesize, src->format, src->width, src->height);
2644
2645     if ((err = unmap_buffers(dev_ctx, buf, planes, 1)))
2646         goto end;
2647
2648     /* Copy buffers to image */
2649     err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
2650                              src->width, src->height, src->format, 0);
2651
2652 end:
2653     for (int i = 0; i < planes; i++)
2654         free_buf(dev_ctx, &buf[i]);
2655
2656     return err;
2657 }
2658
2659 static int vulkan_transfer_data_to(AVHWFramesContext *hwfc, AVFrame *dst,
2660                                         const AVFrame *src)
2661 {
2662     av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2663
2664     switch (src->format) {
2665 #if CONFIG_CUDA
2666     case AV_PIX_FMT_CUDA:
2667         if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
2668             (p->extensions & EXT_EXTERNAL_FD_SEM))
2669             return vulkan_transfer_data_from_cuda(hwfc, dst, src);
2670 #endif
2671     default:
2672         if (src->hw_frames_ctx)
2673             return AVERROR(ENOSYS);
2674         else
2675             return vulkan_transfer_data_from_mem(hwfc, dst, src);
2676     }
2677 }
2678
2679 #if CONFIG_CUDA
2680 static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
2681                                       const AVFrame *src)
2682 {
2683     int err;
2684     VkResult ret;
2685     CUcontext dummy;
2686     AVVkFrame *dst_f;
2687     AVVkFrameInternal *dst_int;
2688     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2689     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2690
2691     AVHWFramesContext *cuda_fc = (AVHWFramesContext*)dst->hw_frames_ctx->data;
2692     AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2693     AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2694     AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2695     CudaFunctions *cu = cu_internal->cuda_dl;
2696
2697     ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2698     if (ret < 0) {
2699         err = AVERROR_EXTERNAL;
2700         goto fail;
2701     }
2702
2703     dst_f = (AVVkFrame *)src->data[0];
2704
2705     err = vulkan_export_to_cuda(hwfc, dst->hw_frames_ctx, src);
2706     if (err < 0) {
2707         goto fail;
2708     }
2709
2710     dst_int = dst_f->internal;
2711
2712     for (int i = 0; i < planes; i++) {
2713         CUDA_MEMCPY2D cpy = {
2714             .dstMemoryType = CU_MEMORYTYPE_DEVICE,
2715             .dstDevice     = (CUdeviceptr)dst->data[i],
2716             .dstPitch      = dst->linesize[i],
2717             .dstY          = 0,
2718
2719             .srcMemoryType = CU_MEMORYTYPE_ARRAY,
2720             .srcArray      = dst_int->cu_array[i],
2721             .WidthInBytes  = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2722                                     : hwfc->width) * desc->comp[i].step,
2723             .Height        = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2724                                    : hwfc->height,
2725         };
2726
2727         ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2728         if (ret < 0) {
2729             err = AVERROR_EXTERNAL;
2730             goto fail;
2731         }
2732     }
2733
2734     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2735
2736     av_log(hwfc, AV_LOG_VERBOSE, "Transfered Vulkan image to CUDA!\n");
2737
2738     return 0;
2739
2740 fail:
2741     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2742     vulkan_free_internal(dst_int);
2743     dst_f->internal = NULL;
2744     av_buffer_unref(&dst->buf[0]);
2745     return err;
2746 }
2747 #endif
2748
2749 static int vulkan_transfer_data_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2750                                        const AVFrame *src)
2751 {
2752     int err = 0;
2753     AVFrame tmp;
2754     AVVkFrame *f = (AVVkFrame *)src->data[0];
2755     AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2756     ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
2757     const int planes = av_pix_fmt_count_planes(dst->format);
2758     int log2_chroma = av_pix_fmt_desc_get(dst->format)->log2_chroma_h;
2759
2760     if (dst->width > hwfc->width || dst->height > hwfc->height)
2761         return AVERROR(EINVAL);
2762
2763     /* For linear, host visiable images */
2764     if (f->tiling == VK_IMAGE_TILING_LINEAR &&
2765         f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
2766         AVFrame *map = av_frame_alloc();
2767         if (!map)
2768             return AVERROR(ENOMEM);
2769         map->format = dst->format;
2770
2771         err = vulkan_map_frame_to_mem(hwfc, map, src, AV_HWFRAME_MAP_READ);
2772         if (err)
2773             return err;
2774
2775         err = av_frame_copy(dst, map);
2776         av_frame_free(&map);
2777         return err;
2778     }
2779
2780     /* Create buffers */
2781     for (int i = 0; i < planes; i++) {
2782         int h = dst->height;
2783         int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
2784
2785         tmp.linesize[i] = dst->linesize[i];
2786         err = create_buf(dev_ctx, &buf[i], p_height,
2787                          &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_DST_BIT,
2788                          VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
2789     }
2790
2791     /* Copy image to buffer */
2792     if ((err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
2793                                   dst->width, dst->height, dst->format, 1)))
2794         goto end;
2795
2796     /* Map, copy buffer to frame, unmap */
2797     if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 1)))
2798         goto end;
2799
2800     av_image_copy(dst->data, dst->linesize, (const uint8_t **)tmp.data,
2801                   tmp.linesize, dst->format, dst->width, dst->height);
2802
2803     err = unmap_buffers(dev_ctx, buf, planes, 0);
2804
2805 end:
2806     for (int i = 0; i < planes; i++)
2807         free_buf(dev_ctx, &buf[i]);
2808
2809     return err;
2810 }
2811
2812 static int vulkan_transfer_data_from(AVHWFramesContext *hwfc, AVFrame *dst,
2813                                      const AVFrame *src)
2814 {
2815     av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2816
2817     switch (dst->format) {
2818 #if CONFIG_CUDA
2819     case AV_PIX_FMT_CUDA:
2820         if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
2821             (p->extensions & EXT_EXTERNAL_FD_SEM))
2822             return vulkan_transfer_data_to_cuda(hwfc, dst, src);
2823 #endif
2824     default:
2825         if (dst->hw_frames_ctx)
2826             return AVERROR(ENOSYS);
2827         else
2828             return vulkan_transfer_data_to_mem(hwfc, dst, src);
2829     }
2830 }
2831
2832 AVVkFrame *av_vk_frame_alloc(void)
2833 {
2834     return av_mallocz(sizeof(AVVkFrame));
2835 }
2836
2837 const HWContextType ff_hwcontext_type_vulkan = {
2838     .type                   = AV_HWDEVICE_TYPE_VULKAN,
2839     .name                   = "Vulkan",
2840
2841     .device_hwctx_size      = sizeof(AVVulkanDeviceContext),
2842     .device_priv_size       = sizeof(VulkanDevicePriv),
2843     .frames_hwctx_size      = sizeof(AVVulkanFramesContext),
2844     .frames_priv_size       = sizeof(VulkanFramesPriv),
2845
2846     .device_init            = &vulkan_device_init,
2847     .device_create          = &vulkan_device_create,
2848     .device_derive          = &vulkan_device_derive,
2849
2850     .frames_get_constraints = &vulkan_frames_get_constraints,
2851     .frames_init            = vulkan_frames_init,
2852     .frames_get_buffer      = vulkan_get_buffer,
2853     .frames_uninit          = vulkan_frames_uninit,
2854
2855     .transfer_get_formats   = vulkan_transfer_get_formats,
2856     .transfer_data_to       = vulkan_transfer_data_to,
2857     .transfer_data_from     = vulkan_transfer_data_from,
2858
2859     .map_to                 = vulkan_map_to,
2860     .map_from               = vulkan_map_from,
2861
2862     .pix_fmts = (const enum AVPixelFormat []) {
2863         AV_PIX_FMT_VULKAN,
2864         AV_PIX_FMT_NONE
2865     },
2866 };