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;
  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         if (internal->cu_sem)
1046             CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem));
1047
1048         for (int i = 0; i < planes; i++) {
1049             if (internal->cu_mma[i])
1050                 CHECK_CU(cu->cuMipmappedArrayDestroy(internal->cu_mma[i]));
1051             if (internal->ext_mem[i])
1052                 CHECK_CU(cu->cuDestroyExternalMemory(internal->ext_mem[i]));
1053         }
1054
1055         av_buffer_unref(&internal->cuda_fc_ref);
1056     }
1057 #endif
1058
1059     av_free(internal);
1060 }
1061
1062 static void vulkan_frame_free(void *opaque, uint8_t *data)
1063 {
1064     AVVkFrame *f = (AVVkFrame *)data;
1065     AVHWFramesContext *hwfc = opaque;
1066     AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1067     int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1068
1069     vulkan_free_internal(f->internal);
1070
1071     for (int i = 0; i < planes; i++) {
1072         vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1073         vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1074     }
1075
1076     vkDestroySemaphore(hwctx->act_dev, f->sem, hwctx->alloc);
1077
1078     av_free(f);
1079 }
1080
1081 static int alloc_bind_mem(AVHWFramesContext *hwfc, AVVkFrame *f,
1082                           void *alloc_pnext, size_t alloc_pnext_stride)
1083 {
1084     int err;
1085     VkResult ret;
1086     AVHWDeviceContext *ctx = hwfc->device_ctx;
1087     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1088     VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { { 0 } };
1089
1090     AVVulkanDeviceContext *hwctx = ctx->hwctx;
1091
1092     for (int i = 0; i < planes; i++) {
1093         int use_ded_mem;
1094         VkImageMemoryRequirementsInfo2 req_desc = {
1095             .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
1096             .image = f->img[i],
1097         };
1098         VkMemoryDedicatedAllocateInfo ded_alloc = {
1099             .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
1100             .pNext = (void *)(((uint8_t *)alloc_pnext) + i*alloc_pnext_stride),
1101         };
1102         VkMemoryDedicatedRequirements ded_req = {
1103             .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
1104         };
1105         VkMemoryRequirements2 req = {
1106             .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
1107             .pNext = &ded_req,
1108         };
1109
1110         vkGetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req);
1111
1112         /* In case the implementation prefers/requires dedicated allocation */
1113         use_ded_mem = ded_req.prefersDedicatedAllocation |
1114                       ded_req.requiresDedicatedAllocation;
1115         if (use_ded_mem)
1116             ded_alloc.image = f->img[i];
1117
1118         /* Allocate memory */
1119         if ((err = alloc_mem(ctx, &req.memoryRequirements,
1120                              f->tiling == VK_IMAGE_TILING_LINEAR ?
1121                              VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT :
1122                              VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1123                              use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext,
1124                              &f->flags, &f->mem[i])))
1125             return err;
1126
1127         f->size[i] = req.memoryRequirements.size;
1128         bind_info[i].sType  = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1129         bind_info[i].image  = f->img[i];
1130         bind_info[i].memory = f->mem[i];
1131     }
1132
1133     /* Bind the allocated memory to the images */
1134     ret = vkBindImageMemory2(hwctx->act_dev, planes, bind_info);
1135     if (ret != VK_SUCCESS) {
1136         av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1137                vk_ret2str(ret));
1138         return AVERROR_EXTERNAL;
1139     }
1140
1141     return 0;
1142 }
1143
1144 enum PrepMode {
1145     PREP_MODE_WRITE,
1146     PREP_MODE_RO_SHADER,
1147 };
1148
1149 static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
1150                          AVVkFrame *frame, enum PrepMode pmode)
1151 {
1152     VkResult ret;
1153     VkImageLayout new_layout;
1154     VkAccessFlags new_access;
1155     AVHWDeviceContext *ctx = hwfc->device_ctx;
1156     AVVulkanDeviceContext *hwctx = ctx->hwctx;
1157     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1158
1159     VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
1160
1161     VkCommandBufferBeginInfo cmd_start = {
1162         .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
1163         .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
1164     };
1165
1166     VkSubmitInfo s_info = {
1167         .sType                = VK_STRUCTURE_TYPE_SUBMIT_INFO,
1168         .commandBufferCount   = 1,
1169         .pCommandBuffers      = &ectx->buf,
1170
1171         .pSignalSemaphores    = &frame->sem,
1172         .signalSemaphoreCount = 1,
1173     };
1174
1175     switch (pmode) {
1176     case PREP_MODE_WRITE:
1177         new_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
1178         new_access = VK_ACCESS_TRANSFER_WRITE_BIT;
1179         break;
1180     case PREP_MODE_RO_SHADER:
1181         new_layout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
1182         new_access = VK_ACCESS_TRANSFER_READ_BIT;
1183         break;
1184     }
1185
1186     ret = vkBeginCommandBuffer(ectx->buf, &cmd_start);
1187     if (ret != VK_SUCCESS)
1188         return AVERROR_EXTERNAL;
1189
1190     /* Change the image layout to something more optimal for writes.
1191      * This also signals the newly created semaphore, making it usable
1192      * for synchronization */
1193     for (int i = 0; i < planes; i++) {
1194         img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
1195         img_bar[i].srcAccessMask = 0x0;
1196         img_bar[i].dstAccessMask = new_access;
1197         img_bar[i].oldLayout = frame->layout[i];
1198         img_bar[i].newLayout = new_layout;
1199         img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
1200         img_bar[i].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
1201         img_bar[i].image = frame->img[i];
1202         img_bar[i].subresourceRange.levelCount = 1;
1203         img_bar[i].subresourceRange.layerCount = 1;
1204         img_bar[i].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
1205
1206         frame->layout[i] = img_bar[i].newLayout;
1207         frame->access[i] = img_bar[i].dstAccessMask;
1208     }
1209
1210     vkCmdPipelineBarrier(ectx->buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
1211                          VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
1212                          0, NULL, 0, NULL, planes, img_bar);
1213
1214     ret = vkEndCommandBuffer(ectx->buf);
1215     if (ret != VK_SUCCESS)
1216         return AVERROR_EXTERNAL;
1217
1218     ret = vkQueueSubmit(ectx->queue, 1, &s_info, ectx->fence);
1219     if (ret != VK_SUCCESS) {
1220         return AVERROR_EXTERNAL;
1221     } else {
1222         vkWaitForFences(hwctx->act_dev, 1, &ectx->fence, VK_TRUE, UINT64_MAX);
1223         vkResetFences(hwctx->act_dev, 1, &ectx->fence);
1224     }
1225
1226     return 0;
1227 }
1228
1229 static int create_frame(AVHWFramesContext *hwfc, AVVkFrame **frame,
1230                         VkImageTiling tiling, VkImageUsageFlagBits usage,
1231                         void *create_pnext)
1232 {
1233     int err;
1234     VkResult ret;
1235     AVHWDeviceContext *ctx = hwfc->device_ctx;
1236     VulkanDevicePriv *p = ctx->internal->priv;
1237     AVVulkanDeviceContext *hwctx = ctx->hwctx;
1238     enum AVPixelFormat format = hwfc->sw_format;
1239     const VkFormat *img_fmts = av_vkfmt_from_pixfmt(format);
1240     const int planes = av_pix_fmt_count_planes(format);
1241
1242     VkExportSemaphoreCreateInfo ext_sem_info = {
1243         .sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO,
1244         .handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
1245     };
1246
1247     VkSemaphoreCreateInfo sem_spawn = {
1248         .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1249         .pNext = p->extensions & EXT_EXTERNAL_FD_SEM ? &ext_sem_info : NULL,
1250     };
1251
1252     AVVkFrame *f = av_vk_frame_alloc();
1253     if (!f) {
1254         av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1255         return AVERROR(ENOMEM);
1256     }
1257
1258     /* Create the images */
1259     for (int i = 0; i < planes; i++) {
1260         const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
1261         int w = hwfc->width;
1262         int h = hwfc->height;
1263         const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
1264         const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
1265
1266         VkImageCreateInfo image_create_info = {
1267             .sType         = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1268             .pNext         = create_pnext,
1269             .imageType     = VK_IMAGE_TYPE_2D,
1270             .format        = img_fmts[i],
1271             .extent.width  = p_w,
1272             .extent.height = p_h,
1273             .extent.depth  = 1,
1274             .mipLevels     = 1,
1275             .arrayLayers   = 1,
1276             .flags         = VK_IMAGE_CREATE_ALIAS_BIT,
1277             .tiling        = tiling,
1278             .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
1279             .usage         = usage,
1280             .sharingMode   = VK_SHARING_MODE_EXCLUSIVE,
1281             .samples       = VK_SAMPLE_COUNT_1_BIT,
1282         };
1283
1284         ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1285                             hwctx->alloc, &f->img[i]);
1286         if (ret != VK_SUCCESS) {
1287             av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1288                    vk_ret2str(ret));
1289             err = AVERROR(EINVAL);
1290             goto fail;
1291         }
1292
1293         f->layout[i] = image_create_info.initialLayout;
1294         f->access[i] = 0x0;
1295     }
1296
1297     /* Create semaphore */
1298     ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1299                             hwctx->alloc, &f->sem);
1300     if (ret != VK_SUCCESS) {
1301         av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1302                vk_ret2str(ret));
1303         return AVERROR_EXTERNAL;
1304     }
1305
1306     f->flags     = 0x0;
1307     f->tiling    = tiling;
1308
1309     *frame = f;
1310     return 0;
1311
1312 fail:
1313     vulkan_frame_free(hwfc, (uint8_t *)f);
1314     return err;
1315 }
1316
1317 /* Checks if an export flag is enabled, and if it is ORs it with *iexp */
1318 static void try_export_flags(AVHWFramesContext *hwfc,
1319                              VkExternalMemoryHandleTypeFlags *comp_handle_types,
1320                              VkExternalMemoryHandleTypeFlagBits *iexp,
1321                              VkExternalMemoryHandleTypeFlagBits exp)
1322 {
1323     VkResult ret;
1324     AVVulkanFramesContext *hwctx = hwfc->hwctx;
1325     AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1326     VkExternalImageFormatProperties eprops = {
1327         .sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR,
1328     };
1329     VkImageFormatProperties2 props = {
1330         .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2,
1331         .pNext = &eprops,
1332     };
1333     VkPhysicalDeviceExternalImageFormatInfo enext = {
1334         .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO,
1335         .handleType = exp,
1336     };
1337     VkPhysicalDeviceImageFormatInfo2 pinfo = {
1338         .sType  = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2,
1339         .pNext  = !exp ? NULL : &enext,
1340         .format = av_vkfmt_from_pixfmt(hwfc->sw_format)[0],
1341         .type   = VK_IMAGE_TYPE_2D,
1342         .tiling = hwctx->tiling,
1343         .usage  = hwctx->usage,
1344         .flags  = VK_IMAGE_CREATE_ALIAS_BIT,
1345     };
1346
1347     ret = vkGetPhysicalDeviceImageFormatProperties2(dev_hwctx->phys_dev,
1348                                                     &pinfo, &props);
1349     if (ret == VK_SUCCESS) {
1350         *iexp |= exp;
1351         *comp_handle_types |= eprops.externalMemoryProperties.compatibleHandleTypes;
1352     }
1353 }
1354
1355 static AVBufferRef *vulkan_pool_alloc(void *opaque, int size)
1356 {
1357     int err;
1358     AVVkFrame *f;
1359     AVBufferRef *avbuf = NULL;
1360     AVHWFramesContext *hwfc = opaque;
1361     AVVulkanFramesContext *hwctx = hwfc->hwctx;
1362     VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1363     VkExportMemoryAllocateInfo eminfo[AV_NUM_DATA_POINTERS];
1364     VkExternalMemoryHandleTypeFlags e = 0x0;
1365
1366     VkExternalMemoryImageCreateInfo eiinfo = {
1367         .sType       = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1368         .pNext       = hwctx->create_pnext,
1369     };
1370
1371     if (p->extensions & EXT_EXTERNAL_FD_MEMORY)
1372         try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1373                          VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT);
1374
1375     if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
1376         try_export_flags(hwfc, &eiinfo.handleTypes, &e,
1377                          VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
1378
1379     for (int i = 0; i < av_pix_fmt_count_planes(hwfc->sw_format); i++) {
1380         eminfo[i].sType       = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO;
1381         eminfo[i].pNext       = hwctx->alloc_pnext[i];
1382         eminfo[i].handleTypes = e;
1383     }
1384
1385     err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1386                        eiinfo.handleTypes ? &eiinfo : NULL);
1387     if (err)
1388         return NULL;
1389
1390     err = alloc_bind_mem(hwfc, f, eminfo, sizeof(*eminfo));
1391     if (err)
1392         goto fail;
1393
1394     err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_WRITE);
1395     if (err)
1396         goto fail;
1397
1398     avbuf = av_buffer_create((uint8_t *)f, sizeof(AVVkFrame),
1399                              vulkan_frame_free, hwfc, 0);
1400     if (!avbuf)
1401         goto fail;
1402
1403     return avbuf;
1404
1405 fail:
1406     vulkan_frame_free(hwfc, (uint8_t *)f);
1407     return NULL;
1408 }
1409
1410 static void vulkan_frames_uninit(AVHWFramesContext *hwfc)
1411 {
1412     VulkanFramesPriv *fp = hwfc->internal->priv;
1413
1414     free_exec_ctx(hwfc->device_ctx, &fp->cmd);
1415 }
1416
1417 static int vulkan_frames_init(AVHWFramesContext *hwfc)
1418 {
1419     int err;
1420     AVVkFrame *f;
1421     AVVulkanFramesContext *hwctx = hwfc->hwctx;
1422     VulkanFramesPriv *fp = hwfc->internal->priv;
1423     AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
1424     VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
1425
1426     if (hwfc->pool)
1427         return 0;
1428
1429     /* Default pool flags */
1430     hwctx->tiling = hwctx->tiling ? hwctx->tiling : p->use_linear_images ?
1431                     VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1432
1433     hwctx->usage |= DEFAULT_USAGE_FLAGS;
1434
1435     err = create_exec_ctx(hwfc->device_ctx, &fp->cmd,
1436                           dev_hwctx->queue_family_tx_index);
1437     if (err)
1438         return err;
1439
1440     /* Test to see if allocation will fail */
1441     err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
1442                        hwctx->create_pnext);
1443     if (err) {
1444         free_exec_ctx(hwfc->device_ctx, &p->cmd);
1445         return err;
1446     }
1447
1448     vulkan_frame_free(hwfc, (uint8_t *)f);
1449
1450     hwfc->internal->pool_internal = av_buffer_pool_init2(sizeof(AVVkFrame),
1451                                                          hwfc, vulkan_pool_alloc,
1452                                                          NULL);
1453     if (!hwfc->internal->pool_internal) {
1454         free_exec_ctx(hwfc->device_ctx, &p->cmd);
1455         return AVERROR(ENOMEM);
1456     }
1457
1458     return 0;
1459 }
1460
1461 static int vulkan_get_buffer(AVHWFramesContext *hwfc, AVFrame *frame)
1462 {
1463     frame->buf[0] = av_buffer_pool_get(hwfc->pool);
1464     if (!frame->buf[0])
1465         return AVERROR(ENOMEM);
1466
1467     frame->data[0] = frame->buf[0]->data;
1468     frame->format  = AV_PIX_FMT_VULKAN;
1469     frame->width   = hwfc->width;
1470     frame->height  = hwfc->height;
1471
1472     return 0;
1473 }
1474
1475 static int vulkan_transfer_get_formats(AVHWFramesContext *hwfc,
1476                                        enum AVHWFrameTransferDirection dir,
1477                                        enum AVPixelFormat **formats)
1478 {
1479     enum AVPixelFormat *fmts = av_malloc_array(2, sizeof(*fmts));
1480     if (!fmts)
1481         return AVERROR(ENOMEM);
1482
1483     fmts[0] = hwfc->sw_format;
1484     fmts[1] = AV_PIX_FMT_NONE;
1485
1486     *formats = fmts;
1487     return 0;
1488 }
1489
1490 typedef struct VulkanMapping {
1491     AVVkFrame *frame;
1492     int flags;
1493 } VulkanMapping;
1494
1495 static void vulkan_unmap_frame(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1496 {
1497     VulkanMapping *map = hwmap->priv;
1498     AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1499     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1500
1501     /* Check if buffer needs flushing */
1502     if ((map->flags & AV_HWFRAME_MAP_WRITE) &&
1503         !(map->frame->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1504         VkResult ret;
1505         VkMappedMemoryRange flush_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1506
1507         for (int i = 0; i < planes; i++) {
1508             flush_ranges[i].sType  = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1509             flush_ranges[i].memory = map->frame->mem[i];
1510             flush_ranges[i].size   = VK_WHOLE_SIZE;
1511         }
1512
1513         ret = vkFlushMappedMemoryRanges(hwctx->act_dev, planes,
1514                                         flush_ranges);
1515         if (ret != VK_SUCCESS) {
1516             av_log(hwfc, AV_LOG_ERROR, "Failed to flush memory: %s\n",
1517                    vk_ret2str(ret));
1518         }
1519     }
1520
1521     for (int i = 0; i < planes; i++)
1522         vkUnmapMemory(hwctx->act_dev, map->frame->mem[i]);
1523
1524     av_free(map);
1525 }
1526
1527 static int vulkan_map_frame_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
1528                                    const AVFrame *src, int flags)
1529 {
1530     VkResult ret;
1531     int err, mapped_mem_count = 0;
1532     AVVkFrame *f = (AVVkFrame *)src->data[0];
1533     AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1534     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1535
1536     VulkanMapping *map = av_mallocz(sizeof(VulkanMapping));
1537     if (!map)
1538         return AVERROR(EINVAL);
1539
1540     if (src->format != AV_PIX_FMT_VULKAN) {
1541         av_log(hwfc, AV_LOG_ERROR, "Cannot map from pixel format %s!\n",
1542                av_get_pix_fmt_name(src->format));
1543         err = AVERROR(EINVAL);
1544         goto fail;
1545     }
1546
1547     if (!(f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) ||
1548         !(f->tiling == VK_IMAGE_TILING_LINEAR)) {
1549         av_log(hwfc, AV_LOG_ERROR, "Unable to map frame, not host visible "
1550                "and linear!\n");
1551         err = AVERROR(EINVAL);
1552         goto fail;
1553     }
1554
1555     dst->width  = src->width;
1556     dst->height = src->height;
1557
1558     for (int i = 0; i < planes; i++) {
1559         ret = vkMapMemory(hwctx->act_dev, f->mem[i], 0,
1560                           VK_WHOLE_SIZE, 0, (void **)&dst->data[i]);
1561         if (ret != VK_SUCCESS) {
1562             av_log(hwfc, AV_LOG_ERROR, "Failed to map image memory: %s\n",
1563                 vk_ret2str(ret));
1564             err = AVERROR_EXTERNAL;
1565             goto fail;
1566         }
1567         mapped_mem_count++;
1568     }
1569
1570     /* Check if the memory contents matter */
1571     if (((flags & AV_HWFRAME_MAP_READ) || !(flags & AV_HWFRAME_MAP_OVERWRITE)) &&
1572         !(f->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
1573         VkMappedMemoryRange map_mem_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
1574         for (int i = 0; i < planes; i++) {
1575             map_mem_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1576             map_mem_ranges[i].size = VK_WHOLE_SIZE;
1577             map_mem_ranges[i].memory = f->mem[i];
1578         }
1579
1580         ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, planes,
1581                                              map_mem_ranges);
1582         if (ret != VK_SUCCESS) {
1583             av_log(hwfc, AV_LOG_ERROR, "Failed to invalidate memory: %s\n",
1584                    vk_ret2str(ret));
1585             err = AVERROR_EXTERNAL;
1586             goto fail;
1587         }
1588     }
1589
1590     for (int i = 0; i < planes; i++) {
1591         VkImageSubresource sub = {
1592             .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
1593         };
1594         VkSubresourceLayout layout;
1595         vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
1596         dst->linesize[i] = layout.rowPitch;
1597     }
1598
1599     map->frame = f;
1600     map->flags = flags;
1601
1602     err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src,
1603                                 &vulkan_unmap_frame, map);
1604     if (err < 0)
1605         goto fail;
1606
1607     return 0;
1608
1609 fail:
1610     for (int i = 0; i < mapped_mem_count; i++)
1611         vkUnmapMemory(hwctx->act_dev, f->mem[i]);
1612
1613     av_free(map);
1614     return err;
1615 }
1616
1617 #if CONFIG_LIBDRM
1618 static void vulkan_unmap_from(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
1619 {
1620     VulkanMapping *map = hwmap->priv;
1621     AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
1622     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1623
1624     for (int i = 0; i < planes; i++) {
1625         vkDestroyImage(hwctx->act_dev, map->frame->img[i], hwctx->alloc);
1626         vkFreeMemory(hwctx->act_dev, map->frame->mem[i], hwctx->alloc);
1627     }
1628
1629     vkDestroySemaphore(hwctx->act_dev, map->frame->sem, hwctx->alloc);
1630
1631     av_freep(&map->frame);
1632 }
1633
1634 static const struct {
1635     uint32_t drm_fourcc;
1636     VkFormat vk_format;
1637 } vulkan_drm_format_map[] = {
1638     { DRM_FORMAT_R8,       VK_FORMAT_R8_UNORM       },
1639     { DRM_FORMAT_R16,      VK_FORMAT_R16_UNORM      },
1640     { DRM_FORMAT_GR88,     VK_FORMAT_R8G8_UNORM     },
1641     { DRM_FORMAT_RG88,     VK_FORMAT_R8G8_UNORM     },
1642     { DRM_FORMAT_GR1616,   VK_FORMAT_R16G16_UNORM   },
1643     { DRM_FORMAT_RG1616,   VK_FORMAT_R16G16_UNORM   },
1644     { DRM_FORMAT_ARGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1645     { DRM_FORMAT_XRGB8888, VK_FORMAT_B8G8R8A8_UNORM },
1646     { DRM_FORMAT_ABGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1647     { DRM_FORMAT_XBGR8888, VK_FORMAT_R8G8B8A8_UNORM },
1648 };
1649
1650 static inline VkFormat drm_to_vulkan_fmt(uint32_t drm_fourcc)
1651 {
1652     for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
1653         if (vulkan_drm_format_map[i].drm_fourcc == drm_fourcc)
1654             return vulkan_drm_format_map[i].vk_format;
1655     return VK_FORMAT_UNDEFINED;
1656 }
1657
1658 static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **frame,
1659                                           AVDRMFrameDescriptor *desc)
1660 {
1661     int err = 0;
1662     VkResult ret;
1663     AVVkFrame *f;
1664     int bind_counts = 0;
1665     AVHWDeviceContext *ctx = hwfc->device_ctx;
1666     AVVulkanDeviceContext *hwctx = ctx->hwctx;
1667     VulkanDevicePriv *p = ctx->internal->priv;
1668     const AVPixFmtDescriptor *fmt_desc = av_pix_fmt_desc_get(hwfc->sw_format);
1669     const int has_modifiers = p->extensions & EXT_DRM_MODIFIER_FLAGS;
1670     VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS] = { 0 };
1671     VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { 0 };
1672     VkBindImagePlaneMemoryInfo plane_info[AV_NUM_DATA_POINTERS] = { 0 };
1673     VkExternalMemoryHandleTypeFlagBits htype = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT;
1674     VkSemaphoreCreateInfo sem_spawn = {
1675         .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
1676     };
1677
1678     VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdPropertiesKHR);
1679
1680     for (int i = 0; i < desc->nb_layers; i++) {
1681         if (drm_to_vulkan_fmt(desc->layers[i].format) == VK_FORMAT_UNDEFINED) {
1682             av_log(ctx, AV_LOG_ERROR, "Unsupported DMABUF layer format %#08x!\n",
1683                    desc->layers[i].format);
1684             return AVERROR(EINVAL);
1685         }
1686     }
1687
1688     if (!(f = av_vk_frame_alloc())) {
1689         av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
1690         err = AVERROR(ENOMEM);
1691         goto fail;
1692     }
1693
1694     for (int i = 0; i < desc->nb_objects; i++) {
1695         VkMemoryFdPropertiesKHR fdmp = {
1696             .sType = VK_STRUCTURE_TYPE_MEMORY_FD_PROPERTIES_KHR,
1697         };
1698         VkMemoryRequirements req = {
1699             .size = desc->objects[i].size,
1700         };
1701         VkImportMemoryFdInfoKHR idesc = {
1702             .sType      = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR,
1703             .handleType = htype,
1704             .fd         = dup(desc->objects[i].fd),
1705         };
1706
1707         ret = pfn_vkGetMemoryFdPropertiesKHR(hwctx->act_dev, htype,
1708                                              idesc.fd, &fdmp);
1709         if (ret != VK_SUCCESS) {
1710             av_log(hwfc, AV_LOG_ERROR, "Failed to get FD properties: %s\n",
1711                    vk_ret2str(ret));
1712             err = AVERROR_EXTERNAL;
1713             close(idesc.fd);
1714             goto fail;
1715         }
1716
1717         req.memoryTypeBits = fdmp.memoryTypeBits;
1718
1719         err = alloc_mem(ctx, &req, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1720                         &idesc, &f->flags, &f->mem[i]);
1721         if (err) {
1722             close(idesc.fd);
1723             return err;
1724         }
1725
1726         f->size[i] = desc->objects[i].size;
1727     }
1728
1729     f->tiling = has_modifiers ? VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT :
1730                 desc->objects[0].format_modifier == DRM_FORMAT_MOD_LINEAR ?
1731                 VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
1732
1733     for (int i = 0; i < desc->nb_layers; i++) {
1734         const int planes = desc->layers[i].nb_planes;
1735         const int signal_p = has_modifiers && (planes > 1);
1736
1737         VkImageDrmFormatModifierExplicitCreateInfoEXT drm_info = {
1738             .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT,
1739             .drmFormatModifier = desc->objects[0].format_modifier,
1740             .drmFormatModifierPlaneCount = planes,
1741             .pPlaneLayouts = (const VkSubresourceLayout *)&plane_data,
1742         };
1743
1744         VkExternalMemoryImageCreateInfo einfo = {
1745             .sType       = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
1746             .pNext       = has_modifiers ? &drm_info : NULL,
1747             .handleTypes = htype,
1748         };
1749
1750         const int p_w = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, fmt_desc->log2_chroma_w) : hwfc->width;
1751         const int p_h = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, fmt_desc->log2_chroma_h) : hwfc->height;
1752
1753         VkImageCreateInfo image_create_info = {
1754             .sType         = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1755             .pNext         = &einfo,
1756             .imageType     = VK_IMAGE_TYPE_2D,
1757             .format        = drm_to_vulkan_fmt(desc->layers[i].format),
1758             .extent.width  = p_w,
1759             .extent.height = p_h,
1760             .extent.depth  = 1,
1761             .mipLevels     = 1,
1762             .arrayLayers   = 1,
1763             .flags         = VK_IMAGE_CREATE_ALIAS_BIT |
1764                              (signal_p ? VK_IMAGE_CREATE_DISJOINT_BIT : 0x0),
1765             .tiling        = f->tiling,
1766             .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */
1767             .usage         = DEFAULT_USAGE_FLAGS,
1768             .sharingMode   = VK_SHARING_MODE_EXCLUSIVE,
1769             .samples       = VK_SAMPLE_COUNT_1_BIT,
1770         };
1771
1772         for (int j = 0; j < planes; j++) {
1773             plane_data[j].offset     = desc->layers[i].planes[j].offset;
1774             plane_data[j].rowPitch   = desc->layers[i].planes[j].pitch;
1775             plane_data[j].size       = 0; /* The specs say so for all 3 */
1776             plane_data[j].arrayPitch = 0;
1777             plane_data[j].depthPitch = 0;
1778         }
1779
1780         /* Create image */
1781         ret = vkCreateImage(hwctx->act_dev, &image_create_info,
1782                             hwctx->alloc, &f->img[i]);
1783         if (ret != VK_SUCCESS) {
1784             av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
1785                    vk_ret2str(ret));
1786             err = AVERROR(EINVAL);
1787             goto fail;
1788         }
1789
1790         f->layout[i] = image_create_info.initialLayout;
1791         f->access[i] = 0x0;
1792
1793         for (int j = 0; j < planes; j++) {
1794             VkImageAspectFlagBits aspect = j == 0 ? VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
1795                                            j == 1 ? VK_IMAGE_ASPECT_MEMORY_PLANE_1_BIT_EXT :
1796                                                     VK_IMAGE_ASPECT_MEMORY_PLANE_2_BIT_EXT;
1797
1798             plane_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO;
1799             plane_info[bind_counts].planeAspect = aspect;
1800
1801             bind_info[bind_counts].sType  = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
1802             bind_info[bind_counts].pNext  = signal_p ? &plane_info[bind_counts] : NULL;
1803             bind_info[bind_counts].image  = f->img[i];
1804             bind_info[bind_counts].memory = f->mem[desc->layers[i].planes[j].object_index];
1805             bind_info[bind_counts].memoryOffset = desc->layers[i].planes[j].offset;
1806             bind_counts++;
1807         }
1808     }
1809
1810     ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
1811                             hwctx->alloc, &f->sem);
1812     if (ret != VK_SUCCESS) {
1813         av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
1814                vk_ret2str(ret));
1815         return AVERROR_EXTERNAL;
1816     }
1817
1818     /* We'd import a semaphore onto the one we created using
1819      * vkImportSemaphoreFdKHR but unfortunately neither DRM nor VAAPI
1820      * offer us anything we could import and sync with, so instead
1821      * just signal the semaphore we created. */
1822
1823     /* Bind the allocated memory to the images */
1824     ret = vkBindImageMemory2(hwctx->act_dev, bind_counts, bind_info);
1825     if (ret != VK_SUCCESS) {
1826         av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
1827                vk_ret2str(ret));
1828         return AVERROR_EXTERNAL;
1829     }
1830
1831     /* NOTE: This is completely uneccesary and unneeded once we can import
1832      * semaphores from DRM. Otherwise we have to activate the semaphores.
1833      * We're reusing the exec context that's also used for uploads/downloads. */
1834     err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_RO_SHADER);
1835     if (err)
1836         goto fail;
1837
1838     *frame = f;
1839
1840     return 0;
1841
1842 fail:
1843     for (int i = 0; i < desc->nb_layers; i++)
1844         vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
1845     for (int i = 0; i < desc->nb_objects; i++)
1846         vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
1847     vkDestroySemaphore(hwctx->act_dev, f->sem, hwctx->alloc);
1848
1849     av_free(f);
1850
1851     return err;
1852 }
1853
1854 static int vulkan_map_from_drm(AVHWFramesContext *hwfc, AVFrame *dst,
1855                                const AVFrame *src, int flags)
1856 {
1857     int err = 0;
1858     AVVkFrame *f;
1859     VulkanMapping *map = NULL;
1860
1861     err = vulkan_map_from_drm_frame_desc(hwfc, &f,
1862                                          (AVDRMFrameDescriptor *)src->data[0]);
1863     if (err)
1864         return err;
1865
1866     /* The unmapping function will free this */
1867     dst->data[0] = (uint8_t *)f;
1868     dst->width   = src->width;
1869     dst->height  = src->height;
1870
1871     map = av_mallocz(sizeof(VulkanMapping));
1872     if (!map)
1873         goto fail;
1874
1875     map->frame = f;
1876     map->flags = flags;
1877
1878     err = ff_hwframe_map_create(dst->hw_frames_ctx, dst, src,
1879                                 &vulkan_unmap_from, map);
1880     if (err < 0)
1881         goto fail;
1882
1883     av_log(hwfc, AV_LOG_DEBUG, "Mapped DRM object to Vulkan!\n");
1884
1885     return 0;
1886
1887 fail:
1888     vulkan_frame_free(hwfc->device_ctx->hwctx, (uint8_t *)f);
1889     av_free(map);
1890     return err;
1891 }
1892
1893 #if CONFIG_VAAPI
1894 static int vulkan_map_from_vaapi(AVHWFramesContext *dst_fc,
1895                                  AVFrame *dst, const AVFrame *src,
1896                                  int flags)
1897 {
1898     int err;
1899     AVFrame *tmp = av_frame_alloc();
1900     AVHWFramesContext *vaapi_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
1901     AVVAAPIDeviceContext *vaapi_ctx = vaapi_fc->device_ctx->hwctx;
1902     VASurfaceID surface_id = (VASurfaceID)(uintptr_t)src->data[3];
1903
1904     if (!tmp)
1905         return AVERROR(ENOMEM);
1906
1907     /* We have to sync since like the previous comment said, no semaphores */
1908     vaSyncSurface(vaapi_ctx->display, surface_id);
1909
1910     tmp->format = AV_PIX_FMT_DRM_PRIME;
1911
1912     err = av_hwframe_map(tmp, src, flags);
1913     if (err < 0)
1914         goto fail;
1915
1916     err = vulkan_map_from_drm(dst_fc, dst, tmp, flags);
1917     if (err < 0)
1918         goto fail;
1919
1920     err = ff_hwframe_map_replace(dst, src);
1921
1922 fail:
1923     av_frame_free(&tmp);
1924     return err;
1925 }
1926 #endif
1927 #endif
1928
1929 #if CONFIG_CUDA
1930 static int vulkan_export_to_cuda(AVHWFramesContext *hwfc,
1931                                  AVBufferRef *cuda_hwfc,
1932                                  const AVFrame *frame)
1933 {
1934     int err;
1935     VkResult ret;
1936     AVVkFrame *dst_f;
1937     AVVkFrameInternal *dst_int;
1938     AVHWDeviceContext *ctx = hwfc->device_ctx;
1939     AVVulkanDeviceContext *hwctx = ctx->hwctx;
1940     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
1941     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
1942     VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
1943     VK_LOAD_PFN(hwctx->inst, vkGetSemaphoreFdKHR);
1944
1945     AVHWFramesContext *cuda_fc = (AVHWFramesContext*)cuda_hwfc->data;
1946     AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
1947     AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
1948     AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
1949     CudaFunctions *cu = cu_internal->cuda_dl;
1950     CUarray_format cufmt = desc->comp[0].depth > 8 ? CU_AD_FORMAT_UNSIGNED_INT16 :
1951                                                      CU_AD_FORMAT_UNSIGNED_INT8;
1952
1953     dst_f = (AVVkFrame *)frame->data[0];
1954
1955     dst_int = dst_f->internal;
1956     if (!dst_int || !dst_int->cuda_fc_ref) {
1957         VkSemaphoreGetFdInfoKHR sem_export = {
1958             .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
1959             .semaphore = dst_f->sem,
1960             .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
1961         };
1962         CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = {
1963             .type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
1964         };
1965
1966         if (!dst_f->internal)
1967             dst_f->internal = dst_int = av_mallocz(sizeof(*dst_f->internal));
1968
1969         if (!dst_int) {
1970             err = AVERROR(ENOMEM);
1971             goto fail;
1972         }
1973
1974         dst_int->cuda_fc_ref = av_buffer_ref(cuda_hwfc);
1975         if (!dst_int->cuda_fc_ref) {
1976             err = AVERROR(ENOMEM);
1977             goto fail;
1978         }
1979
1980         for (int i = 0; i < planes; i++) {
1981             CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC tex_desc = {
1982                 .offset = 0,
1983                 .arrayDesc = {
1984                     .Width  = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
1985                                     : hwfc->width,
1986                     .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
1987                                     : hwfc->height,
1988                     .Depth = 0,
1989                     .Format = cufmt,
1990                     .NumChannels = 1 + ((planes == 2) && i),
1991                     .Flags = 0,
1992                 },
1993                 .numLevels = 1,
1994             };
1995             CUDA_EXTERNAL_MEMORY_HANDLE_DESC ext_desc = {
1996                 .type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD,
1997                 .size = dst_f->size[i],
1998             };
1999             VkMemoryGetFdInfoKHR export_info = {
2000                 .sType      = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2001                 .memory     = dst_f->mem[i],
2002                 .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR,
2003             };
2004
2005             ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2006                                        &ext_desc.handle.fd);
2007             if (ret != VK_SUCCESS) {
2008                 av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2009                 err = AVERROR_EXTERNAL;
2010                 goto fail;
2011             }
2012
2013             ret = CHECK_CU(cu->cuImportExternalMemory(&dst_int->ext_mem[i], &ext_desc));
2014             if (ret < 0) {
2015                 err = AVERROR_EXTERNAL;
2016                 goto fail;
2017             }
2018
2019             ret = CHECK_CU(cu->cuExternalMemoryGetMappedMipmappedArray(&dst_int->cu_mma[i],
2020                                                                        dst_int->ext_mem[i],
2021                                                                        &tex_desc));
2022             if (ret < 0) {
2023                 err = AVERROR_EXTERNAL;
2024                 goto fail;
2025             }
2026
2027             ret = CHECK_CU(cu->cuMipmappedArrayGetLevel(&dst_int->cu_array[i],
2028                                                         dst_int->cu_mma[i], 0));
2029             if (ret < 0) {
2030                 err = AVERROR_EXTERNAL;
2031                 goto fail;
2032             }
2033         }
2034
2035         ret = pfn_vkGetSemaphoreFdKHR(hwctx->act_dev, &sem_export,
2036                                       &ext_sem_desc.handle.fd);
2037         if (ret != VK_SUCCESS) {
2038             av_log(ctx, AV_LOG_ERROR, "Failed to export semaphore: %s\n",
2039                    vk_ret2str(ret));
2040             err = AVERROR_EXTERNAL;
2041             goto fail;
2042         }
2043
2044         ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem,
2045                                                      &ext_sem_desc));
2046         if (ret < 0) {
2047             err = AVERROR_EXTERNAL;
2048             goto fail;
2049         }
2050     }
2051
2052     return 0;
2053
2054 fail:
2055     return err;
2056 }
2057
2058 static int vulkan_transfer_data_from_cuda(AVHWFramesContext *hwfc,
2059                                           AVFrame *dst, const AVFrame *src)
2060 {
2061     int err;
2062     VkResult ret;
2063     CUcontext dummy;
2064     AVVkFrame *dst_f;
2065     AVVkFrameInternal *dst_int;
2066     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2067     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2068
2069     AVHWFramesContext *cuda_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
2070     AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2071     AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2072     AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2073     CudaFunctions *cu = cu_internal->cuda_dl;
2074     CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par = { 0 };
2075     CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par = { 0 };
2076
2077     ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2078     if (ret < 0) {
2079         err = AVERROR_EXTERNAL;
2080         goto fail;
2081     }
2082
2083     dst_f = (AVVkFrame *)dst->data[0];
2084
2085     ret = vulkan_export_to_cuda(hwfc, src->hw_frames_ctx, dst);
2086     if (ret < 0) {
2087         goto fail;
2088     }
2089     dst_int = dst_f->internal;
2090
2091     ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(&dst_int->cu_sem, &s_w_par,
2092                                                      1, cuda_dev->stream));
2093     if (ret < 0) {
2094         err = AVERROR_EXTERNAL;
2095         goto fail;
2096     }
2097
2098     for (int i = 0; i < planes; i++) {
2099         CUDA_MEMCPY2D cpy = {
2100             .srcMemoryType = CU_MEMORYTYPE_DEVICE,
2101             .srcDevice     = (CUdeviceptr)src->data[i],
2102             .srcPitch      = src->linesize[i],
2103             .srcY          = 0,
2104
2105             .dstMemoryType = CU_MEMORYTYPE_ARRAY,
2106             .dstArray      = dst_int->cu_array[i],
2107             .WidthInBytes  = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2108                                     : hwfc->width) * desc->comp[i].step,
2109             .Height        = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2110                                    : hwfc->height,
2111         };
2112
2113         ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2114         if (ret < 0) {
2115             err = AVERROR_EXTERNAL;
2116             goto fail;
2117         }
2118     }
2119
2120     ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(&dst_int->cu_sem, &s_s_par,
2121                                                        1, cuda_dev->stream));
2122     if (ret < 0) {
2123         err = AVERROR_EXTERNAL;
2124         goto fail;
2125     }
2126
2127     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2128
2129     av_log(hwfc, AV_LOG_VERBOSE, "Transfered CUDA image to Vulkan!\n");
2130
2131     return 0;
2132
2133 fail:
2134     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2135     vulkan_free_internal(dst_int);
2136     dst_f->internal = NULL;
2137     av_buffer_unref(&dst->buf[0]);
2138     return err;
2139 }
2140 #endif
2141
2142 static int vulkan_map_to(AVHWFramesContext *hwfc, AVFrame *dst,
2143                          const AVFrame *src, int flags)
2144 {
2145     av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2146
2147     switch (src->format) {
2148 #if CONFIG_LIBDRM
2149 #if CONFIG_VAAPI
2150     case AV_PIX_FMT_VAAPI:
2151         if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2152             return vulkan_map_from_vaapi(hwfc, dst, src, flags);
2153 #endif
2154     case AV_PIX_FMT_DRM_PRIME:
2155         if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2156             return vulkan_map_from_drm(hwfc, dst, src, flags);
2157 #endif
2158     default:
2159         return AVERROR(ENOSYS);
2160     }
2161 }
2162
2163 #if CONFIG_LIBDRM
2164 typedef struct VulkanDRMMapping {
2165     AVDRMFrameDescriptor drm_desc;
2166     AVVkFrame *source;
2167 } VulkanDRMMapping;
2168
2169 static void vulkan_unmap_to_drm(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
2170 {
2171     AVDRMFrameDescriptor *drm_desc = hwmap->priv;
2172
2173     for (int i = 0; i < drm_desc->nb_objects; i++)
2174         close(drm_desc->objects[i].fd);
2175
2176     av_free(drm_desc);
2177 }
2178
2179 static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt)
2180 {
2181     for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
2182         if (vulkan_drm_format_map[i].vk_format == vkfmt)
2183             return vulkan_drm_format_map[i].drm_fourcc;
2184     return DRM_FORMAT_INVALID;
2185 }
2186
2187 static int vulkan_map_to_drm(AVHWFramesContext *hwfc, AVFrame *dst,
2188                              const AVFrame *src, int flags)
2189 {
2190     int err = 0;
2191     VkResult ret;
2192     AVVkFrame *f = (AVVkFrame *)src->data[0];
2193     VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2194     AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
2195     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2196     VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
2197     VkImageDrmFormatModifierPropertiesEXT drm_mod = {
2198         .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_PROPERTIES_EXT,
2199     };
2200
2201     AVDRMFrameDescriptor *drm_desc = av_mallocz(sizeof(*drm_desc));
2202     if (!drm_desc)
2203         return AVERROR(ENOMEM);
2204
2205     err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src, &vulkan_unmap_to_drm, drm_desc);
2206     if (err < 0)
2207         goto end;
2208
2209     if (p->extensions & EXT_DRM_MODIFIER_FLAGS) {
2210         VK_LOAD_PFN(hwctx->inst, vkGetImageDrmFormatModifierPropertiesEXT);
2211         ret = pfn_vkGetImageDrmFormatModifierPropertiesEXT(hwctx->act_dev, f->img[0],
2212                                                            &drm_mod);
2213         if (ret != VK_SUCCESS) {
2214             av_log(hwfc, AV_LOG_ERROR, "Failed to retrieve DRM format modifier!\n");
2215             err = AVERROR_EXTERNAL;
2216             goto end;
2217         }
2218     }
2219
2220     for (int i = 0; (i < planes) && (f->mem[i]); i++) {
2221         VkMemoryGetFdInfoKHR export_info = {
2222             .sType      = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
2223             .memory     = f->mem[i],
2224             .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT,
2225         };
2226
2227         ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
2228                                    &drm_desc->objects[i].fd);
2229         if (ret != VK_SUCCESS) {
2230             av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
2231             err = AVERROR_EXTERNAL;
2232             goto end;
2233         }
2234
2235         drm_desc->nb_objects++;
2236         drm_desc->objects[i].size = f->size[i];
2237         drm_desc->objects[i].format_modifier = drm_mod.drmFormatModifier;
2238     }
2239
2240     drm_desc->nb_layers = planes;
2241     for (int i = 0; i < drm_desc->nb_layers; i++) {
2242         VkSubresourceLayout layout;
2243         VkImageSubresource sub = {
2244             .aspectMask = p->extensions & EXT_DRM_MODIFIER_FLAGS ?
2245                           VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
2246                           VK_IMAGE_ASPECT_COLOR_BIT,
2247         };
2248         VkFormat plane_vkfmt = av_vkfmt_from_pixfmt(hwfc->sw_format)[i];
2249
2250         drm_desc->layers[i].format    = vulkan_fmt_to_drm(plane_vkfmt);
2251         drm_desc->layers[i].nb_planes = 1;
2252
2253         if (drm_desc->layers[i].format == DRM_FORMAT_INVALID) {
2254             av_log(hwfc, AV_LOG_ERROR, "Cannot map to DRM layer, unsupported!\n");
2255             err = AVERROR_PATCHWELCOME;
2256             goto end;
2257         }
2258
2259         drm_desc->layers[i].planes[0].object_index = FFMIN(i, drm_desc->nb_objects - 1);
2260
2261         if (f->tiling != VK_IMAGE_TILING_OPTIMAL)
2262             continue;
2263
2264         vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
2265         drm_desc->layers[i].planes[0].offset       = layout.offset;
2266         drm_desc->layers[i].planes[0].pitch        = layout.rowPitch;
2267     }
2268
2269     dst->width   = src->width;
2270     dst->height  = src->height;
2271     dst->data[0] = (uint8_t *)drm_desc;
2272
2273     av_log(hwfc, AV_LOG_VERBOSE, "Mapped AVVkFrame to a DRM object!\n");
2274
2275     return 0;
2276
2277 end:
2278     av_free(drm_desc);
2279     return err;
2280 }
2281
2282 #if CONFIG_VAAPI
2283 static int vulkan_map_to_vaapi(AVHWFramesContext *hwfc, AVFrame *dst,
2284                                const AVFrame *src, int flags)
2285 {
2286     int err;
2287     AVFrame *tmp = av_frame_alloc();
2288     if (!tmp)
2289         return AVERROR(ENOMEM);
2290
2291     tmp->format = AV_PIX_FMT_DRM_PRIME;
2292
2293     err = vulkan_map_to_drm(hwfc, tmp, src, flags);
2294     if (err < 0)
2295         goto fail;
2296
2297     err = av_hwframe_map(dst, tmp, flags);
2298     if (err < 0)
2299         goto fail;
2300
2301     err = ff_hwframe_map_replace(dst, src);
2302
2303 fail:
2304     av_frame_free(&tmp);
2305     return err;
2306 }
2307 #endif
2308 #endif
2309
2310 static int vulkan_map_from(AVHWFramesContext *hwfc, AVFrame *dst,
2311                            const AVFrame *src, int flags)
2312 {
2313     av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2314
2315     switch (dst->format) {
2316 #if CONFIG_LIBDRM
2317     case AV_PIX_FMT_DRM_PRIME:
2318         if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2319             return vulkan_map_to_drm(hwfc, dst, src, flags);
2320 #if CONFIG_VAAPI
2321     case AV_PIX_FMT_VAAPI:
2322         if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
2323             return vulkan_map_to_vaapi(hwfc, dst, src, flags);
2324 #endif
2325 #endif
2326     default:
2327         return vulkan_map_frame_to_mem(hwfc, dst, src, flags);
2328     }
2329 }
2330
2331 typedef struct ImageBuffer {
2332     VkBuffer buf;
2333     VkDeviceMemory mem;
2334     VkMemoryPropertyFlagBits flags;
2335 } ImageBuffer;
2336
2337 static void free_buf(AVHWDeviceContext *ctx, ImageBuffer *buf)
2338 {
2339     AVVulkanDeviceContext *hwctx = ctx->hwctx;
2340     if (!buf)
2341         return;
2342
2343     vkDestroyBuffer(hwctx->act_dev, buf->buf, hwctx->alloc);
2344     vkFreeMemory(hwctx->act_dev, buf->mem, hwctx->alloc);
2345 }
2346
2347 static int create_buf(AVHWDeviceContext *ctx, ImageBuffer *buf, int height,
2348                       int *stride, VkBufferUsageFlags usage,
2349                       VkMemoryPropertyFlagBits flags, void *create_pnext,
2350                       void *alloc_pnext)
2351 {
2352     int err;
2353     VkResult ret;
2354     VkMemoryRequirements req;
2355     AVVulkanDeviceContext *hwctx = ctx->hwctx;
2356     VulkanDevicePriv *p = ctx->internal->priv;
2357
2358     VkBufferCreateInfo buf_spawn = {
2359         .sType       = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
2360         .pNext       = create_pnext,
2361         .usage       = usage,
2362         .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
2363     };
2364
2365     *stride = FFALIGN(*stride, p->props.limits.optimalBufferCopyRowPitchAlignment);
2366     buf_spawn.size = height*(*stride);
2367
2368     ret = vkCreateBuffer(hwctx->act_dev, &buf_spawn, NULL, &buf->buf);
2369     if (ret != VK_SUCCESS) {
2370         av_log(ctx, AV_LOG_ERROR, "Failed to create buffer: %s\n",
2371                vk_ret2str(ret));
2372         return AVERROR_EXTERNAL;
2373     }
2374
2375     vkGetBufferMemoryRequirements(hwctx->act_dev, buf->buf, &req);
2376
2377     err = alloc_mem(ctx, &req, flags, alloc_pnext, &buf->flags, &buf->mem);
2378     if (err)
2379         return err;
2380
2381     ret = vkBindBufferMemory(hwctx->act_dev, buf->buf, buf->mem, 0);
2382     if (ret != VK_SUCCESS) {
2383         av_log(ctx, AV_LOG_ERROR, "Failed to bind memory to buffer: %s\n",
2384                vk_ret2str(ret));
2385         free_buf(ctx, buf);
2386         return AVERROR_EXTERNAL;
2387     }
2388
2389     return 0;
2390 }
2391
2392 static int map_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf, uint8_t *mem[],
2393                        int nb_buffers, int invalidate)
2394 {
2395     VkResult ret;
2396     AVVulkanDeviceContext *hwctx = ctx->hwctx;
2397     VkMappedMemoryRange invalidate_ctx[AV_NUM_DATA_POINTERS];
2398     int invalidate_count = 0;
2399
2400     for (int i = 0; i < nb_buffers; i++) {
2401         ret = vkMapMemory(hwctx->act_dev, buf[i].mem, 0,
2402                           VK_WHOLE_SIZE, 0, (void **)&mem[i]);
2403         if (ret != VK_SUCCESS) {
2404             av_log(ctx, AV_LOG_ERROR, "Failed to map buffer memory: %s\n",
2405                    vk_ret2str(ret));
2406             return AVERROR_EXTERNAL;
2407         }
2408     }
2409
2410     if (!invalidate)
2411         return 0;
2412
2413     for (int i = 0; i < nb_buffers; i++) {
2414         const VkMappedMemoryRange ival_buf = {
2415             .sType  = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2416             .memory = buf[i].mem,
2417             .size   = VK_WHOLE_SIZE,
2418         };
2419         if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2420             continue;
2421         invalidate_ctx[invalidate_count++] = ival_buf;
2422     }
2423
2424     if (invalidate_count) {
2425         ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, invalidate_count,
2426                                              invalidate_ctx);
2427         if (ret != VK_SUCCESS)
2428             av_log(ctx, AV_LOG_WARNING, "Failed to invalidate memory: %s\n",
2429                    vk_ret2str(ret));
2430     }
2431
2432     return 0;
2433 }
2434
2435 static int unmap_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf,
2436                          int nb_buffers, int flush)
2437 {
2438     int err = 0;
2439     VkResult ret;
2440     AVVulkanDeviceContext *hwctx = ctx->hwctx;
2441     VkMappedMemoryRange flush_ctx[AV_NUM_DATA_POINTERS];
2442     int flush_count = 0;
2443
2444     if (flush) {
2445         for (int i = 0; i < nb_buffers; i++) {
2446             const VkMappedMemoryRange flush_buf = {
2447                 .sType  = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
2448                 .memory = buf[i].mem,
2449                 .size   = VK_WHOLE_SIZE,
2450             };
2451             if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
2452                 continue;
2453             flush_ctx[flush_count++] = flush_buf;
2454         }
2455     }
2456
2457     if (flush_count) {
2458         ret = vkFlushMappedMemoryRanges(hwctx->act_dev, flush_count, flush_ctx);
2459         if (ret != VK_SUCCESS) {
2460             av_log(ctx, AV_LOG_ERROR, "Failed to flush memory: %s\n",
2461                     vk_ret2str(ret));
2462             err = AVERROR_EXTERNAL; /* We still want to try to unmap them */
2463         }
2464     }
2465
2466     for (int i = 0; i < nb_buffers; i++)
2467         vkUnmapMemory(hwctx->act_dev, buf[i].mem);
2468
2469     return err;
2470 }
2471
2472 static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
2473                               ImageBuffer *buffer, const int *buf_stride, int w,
2474                               int h, enum AVPixelFormat pix_fmt, int to_buf)
2475 {
2476     VkResult ret;
2477     AVVulkanDeviceContext *hwctx = ctx->hwctx;
2478     VulkanDevicePriv *s = ctx->internal->priv;
2479
2480     int bar_num = 0;
2481     VkPipelineStageFlagBits sem_wait_dst[AV_NUM_DATA_POINTERS];
2482
2483     const int planes = av_pix_fmt_count_planes(pix_fmt);
2484     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
2485
2486     VkCommandBufferBeginInfo cmd_start = {
2487         .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
2488         .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
2489     };
2490
2491     VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
2492
2493     VkSubmitInfo s_info = {
2494         .sType                = VK_STRUCTURE_TYPE_SUBMIT_INFO,
2495         .commandBufferCount   = 1,
2496         .pCommandBuffers      = &s->cmd.buf,
2497         .pSignalSemaphores    = &frame->sem,
2498         .pWaitSemaphores      = &frame->sem,
2499         .pWaitDstStageMask    = sem_wait_dst,
2500         .signalSemaphoreCount = 1,
2501         .waitSemaphoreCount   = 1,
2502     };
2503
2504     ret = vkBeginCommandBuffer(s->cmd.buf, &cmd_start);
2505     if (ret != VK_SUCCESS) {
2506         av_log(ctx, AV_LOG_ERROR, "Unable to init command buffer: %s\n",
2507                vk_ret2str(ret));
2508         return AVERROR_EXTERNAL;
2509     }
2510
2511     /* Change the image layout to something more optimal for transfers */
2512     for (int i = 0; i < planes; i++) {
2513         VkImageLayout new_layout = to_buf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL :
2514                                             VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
2515         VkAccessFlags new_access = to_buf ? VK_ACCESS_TRANSFER_READ_BIT :
2516                                             VK_ACCESS_TRANSFER_WRITE_BIT;
2517
2518         sem_wait_dst[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
2519
2520         /* If the layout matches and we have read access skip the barrier */
2521         if ((frame->layout[i] == new_layout) && (frame->access[i] & new_access))
2522             continue;
2523
2524         img_bar[bar_num].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
2525         img_bar[bar_num].srcAccessMask = 0x0;
2526         img_bar[bar_num].dstAccessMask = new_access;
2527         img_bar[bar_num].oldLayout = frame->layout[i];
2528         img_bar[bar_num].newLayout = new_layout;
2529         img_bar[bar_num].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2530         img_bar[bar_num].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
2531         img_bar[bar_num].image = frame->img[i];
2532         img_bar[bar_num].subresourceRange.levelCount = 1;
2533         img_bar[bar_num].subresourceRange.layerCount = 1;
2534         img_bar[bar_num].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
2535
2536         frame->layout[i] = img_bar[bar_num].newLayout;
2537         frame->access[i] = img_bar[bar_num].dstAccessMask;
2538
2539         bar_num++;
2540     }
2541
2542     if (bar_num)
2543         vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
2544                              VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
2545                              0, NULL, 0, NULL, bar_num, img_bar);
2546
2547     /* Schedule a copy for each plane */
2548     for (int i = 0; i < planes; i++) {
2549         const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
2550         const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
2551         VkBufferImageCopy buf_reg = {
2552             .bufferOffset = 0,
2553             /* Buffer stride isn't in bytes, it's in samples, the implementation
2554              * uses the image's VkFormat to know how many bytes per sample
2555              * the buffer has. So we have to convert by dividing. Stupid.
2556              * Won't work with YUVA or other planar formats with alpha. */
2557             .bufferRowLength = buf_stride[i] / desc->comp[i].step,
2558             .bufferImageHeight = p_h,
2559             .imageSubresource.layerCount = 1,
2560             .imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
2561             .imageOffset = { 0, 0, 0, },
2562             .imageExtent = { p_w, p_h, 1, },
2563         };
2564
2565         if (to_buf)
2566             vkCmdCopyImageToBuffer(s->cmd.buf, frame->img[i], frame->layout[i],
2567                                    buffer[i].buf, 1, &buf_reg);
2568         else
2569             vkCmdCopyBufferToImage(s->cmd.buf, buffer[i].buf, frame->img[i],
2570                                    frame->layout[i], 1, &buf_reg);
2571     }
2572
2573     ret = vkEndCommandBuffer(s->cmd.buf);
2574     if (ret != VK_SUCCESS) {
2575         av_log(ctx, AV_LOG_ERROR, "Unable to finish command buffer: %s\n",
2576                vk_ret2str(ret));
2577         return AVERROR_EXTERNAL;
2578     }
2579
2580     /* Wait for the download/upload to finish if uploading, otherwise the
2581      * semaphore will take care of synchronization when uploading */
2582     ret = vkQueueSubmit(s->cmd.queue, 1, &s_info, s->cmd.fence);
2583     if (ret != VK_SUCCESS) {
2584         av_log(ctx, AV_LOG_ERROR, "Unable to submit command buffer: %s\n",
2585                vk_ret2str(ret));
2586         return AVERROR_EXTERNAL;
2587     } else {
2588         vkWaitForFences(hwctx->act_dev, 1, &s->cmd.fence, VK_TRUE, UINT64_MAX);
2589         vkResetFences(hwctx->act_dev, 1, &s->cmd.fence);
2590     }
2591
2592     return 0;
2593 }
2594
2595 /* Technically we can use VK_EXT_external_memory_host to upload and download,
2596  * however the alignment requirements make this unfeasible as both the pointer
2597  * and the size of each plane need to be aligned to the minimum alignment
2598  * requirement, which on all current implementations (anv, radv) is 4096.
2599  * If the requirement gets relaxed (unlikely) this can easily be implemented. */
2600 static int vulkan_transfer_data_from_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2601                                          const AVFrame *src)
2602 {
2603     int err = 0;
2604     AVFrame tmp;
2605     AVVkFrame *f = (AVVkFrame *)dst->data[0];
2606     AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2607     ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
2608     const int planes = av_pix_fmt_count_planes(src->format);
2609     int log2_chroma = av_pix_fmt_desc_get(src->format)->log2_chroma_h;
2610
2611     if ((src->format != AV_PIX_FMT_NONE && !av_vkfmt_from_pixfmt(src->format))) {
2612         av_log(hwfc, AV_LOG_ERROR, "Unsupported source pixel format!\n");
2613         return AVERROR(EINVAL);
2614     }
2615
2616     if (src->width > hwfc->width || src->height > hwfc->height)
2617         return AVERROR(EINVAL);
2618
2619     /* For linear, host visiable images */
2620     if (f->tiling == VK_IMAGE_TILING_LINEAR &&
2621         f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
2622         AVFrame *map = av_frame_alloc();
2623         if (!map)
2624             return AVERROR(ENOMEM);
2625         map->format = src->format;
2626
2627         err = vulkan_map_frame_to_mem(hwfc, map, dst, AV_HWFRAME_MAP_WRITE);
2628         if (err)
2629             goto end;
2630
2631         err = av_frame_copy(map, src);
2632         av_frame_free(&map);
2633         goto end;
2634     }
2635
2636     /* Create buffers */
2637     for (int i = 0; i < planes; i++) {
2638         int h = src->height;
2639         int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
2640
2641         tmp.linesize[i] = FFABS(src->linesize[i]);
2642         err = create_buf(dev_ctx, &buf[i], p_height,
2643                          &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
2644                          VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
2645         if (err)
2646             goto end;
2647     }
2648
2649     /* Map, copy image to buffer, unmap */
2650     if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 0)))
2651         goto end;
2652
2653     av_image_copy(tmp.data, tmp.linesize, (const uint8_t **)src->data,
2654                   src->linesize, src->format, src->width, src->height);
2655
2656     if ((err = unmap_buffers(dev_ctx, buf, planes, 1)))
2657         goto end;
2658
2659     /* Copy buffers to image */
2660     err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
2661                              src->width, src->height, src->format, 0);
2662
2663 end:
2664     for (int i = 0; i < planes; i++)
2665         free_buf(dev_ctx, &buf[i]);
2666
2667     return err;
2668 }
2669
2670 static int vulkan_transfer_data_to(AVHWFramesContext *hwfc, AVFrame *dst,
2671                                         const AVFrame *src)
2672 {
2673     av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2674
2675     switch (src->format) {
2676 #if CONFIG_CUDA
2677     case AV_PIX_FMT_CUDA:
2678         if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
2679             (p->extensions & EXT_EXTERNAL_FD_SEM))
2680             return vulkan_transfer_data_from_cuda(hwfc, dst, src);
2681 #endif
2682     default:
2683         if (src->hw_frames_ctx)
2684             return AVERROR(ENOSYS);
2685         else
2686             return vulkan_transfer_data_from_mem(hwfc, dst, src);
2687     }
2688 }
2689
2690 #if CONFIG_CUDA
2691 static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
2692                                       const AVFrame *src)
2693 {
2694     int err;
2695     VkResult ret;
2696     CUcontext dummy;
2697     AVVkFrame *dst_f;
2698     AVVkFrameInternal *dst_int;
2699     const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
2700     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
2701
2702     AVHWFramesContext *cuda_fc = (AVHWFramesContext*)dst->hw_frames_ctx->data;
2703     AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
2704     AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
2705     AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
2706     CudaFunctions *cu = cu_internal->cuda_dl;
2707
2708     ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
2709     if (ret < 0) {
2710         err = AVERROR_EXTERNAL;
2711         goto fail;
2712     }
2713
2714     dst_f = (AVVkFrame *)src->data[0];
2715
2716     err = vulkan_export_to_cuda(hwfc, dst->hw_frames_ctx, src);
2717     if (err < 0) {
2718         goto fail;
2719     }
2720
2721     dst_int = dst_f->internal;
2722
2723     for (int i = 0; i < planes; i++) {
2724         CUDA_MEMCPY2D cpy = {
2725             .dstMemoryType = CU_MEMORYTYPE_DEVICE,
2726             .dstDevice     = (CUdeviceptr)dst->data[i],
2727             .dstPitch      = dst->linesize[i],
2728             .dstY          = 0,
2729
2730             .srcMemoryType = CU_MEMORYTYPE_ARRAY,
2731             .srcArray      = dst_int->cu_array[i],
2732             .WidthInBytes  = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
2733                                     : hwfc->width) * desc->comp[i].step,
2734             .Height        = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
2735                                    : hwfc->height,
2736         };
2737
2738         ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
2739         if (ret < 0) {
2740             err = AVERROR_EXTERNAL;
2741             goto fail;
2742         }
2743     }
2744
2745     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2746
2747     av_log(hwfc, AV_LOG_VERBOSE, "Transfered Vulkan image to CUDA!\n");
2748
2749     return 0;
2750
2751 fail:
2752     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
2753     vulkan_free_internal(dst_int);
2754     dst_f->internal = NULL;
2755     av_buffer_unref(&dst->buf[0]);
2756     return err;
2757 }
2758 #endif
2759
2760 static int vulkan_transfer_data_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
2761                                        const AVFrame *src)
2762 {
2763     int err = 0;
2764     AVFrame tmp;
2765     AVVkFrame *f = (AVVkFrame *)src->data[0];
2766     AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
2767     ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
2768     const int planes = av_pix_fmt_count_planes(dst->format);
2769     int log2_chroma = av_pix_fmt_desc_get(dst->format)->log2_chroma_h;
2770
2771     if (dst->width > hwfc->width || dst->height > hwfc->height)
2772         return AVERROR(EINVAL);
2773
2774     /* For linear, host visiable images */
2775     if (f->tiling == VK_IMAGE_TILING_LINEAR &&
2776         f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
2777         AVFrame *map = av_frame_alloc();
2778         if (!map)
2779             return AVERROR(ENOMEM);
2780         map->format = dst->format;
2781
2782         err = vulkan_map_frame_to_mem(hwfc, map, src, AV_HWFRAME_MAP_READ);
2783         if (err)
2784             return err;
2785
2786         err = av_frame_copy(dst, map);
2787         av_frame_free(&map);
2788         return err;
2789     }
2790
2791     /* Create buffers */
2792     for (int i = 0; i < planes; i++) {
2793         int h = dst->height;
2794         int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
2795
2796         tmp.linesize[i] = FFABS(dst->linesize[i]);
2797         err = create_buf(dev_ctx, &buf[i], p_height,
2798                          &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_DST_BIT,
2799                          VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
2800     }
2801
2802     /* Copy image to buffer */
2803     if ((err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
2804                                   dst->width, dst->height, dst->format, 1)))
2805         goto end;
2806
2807     /* Map, copy buffer to frame, unmap */
2808     if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 1)))
2809         goto end;
2810
2811     av_image_copy(dst->data, dst->linesize, (const uint8_t **)tmp.data,
2812                   tmp.linesize, dst->format, dst->width, dst->height);
2813
2814     err = unmap_buffers(dev_ctx, buf, planes, 0);
2815
2816 end:
2817     for (int i = 0; i < planes; i++)
2818         free_buf(dev_ctx, &buf[i]);
2819
2820     return err;
2821 }
2822
2823 static int vulkan_transfer_data_from(AVHWFramesContext *hwfc, AVFrame *dst,
2824                                      const AVFrame *src)
2825 {
2826     av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
2827
2828     switch (dst->format) {
2829 #if CONFIG_CUDA
2830     case AV_PIX_FMT_CUDA:
2831         if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
2832             (p->extensions & EXT_EXTERNAL_FD_SEM))
2833             return vulkan_transfer_data_to_cuda(hwfc, dst, src);
2834 #endif
2835     default:
2836         if (dst->hw_frames_ctx)
2837             return AVERROR(ENOSYS);
2838         else
2839             return vulkan_transfer_data_to_mem(hwfc, dst, src);
2840     }
2841 }
2842
2843 AVVkFrame *av_vk_frame_alloc(void)
2844 {
2845     return av_mallocz(sizeof(AVVkFrame));
2846 }
2847
2848 const HWContextType ff_hwcontext_type_vulkan = {
2849     .type                   = AV_HWDEVICE_TYPE_VULKAN,
2850     .name                   = "Vulkan",
2851
2852     .device_hwctx_size      = sizeof(AVVulkanDeviceContext),
2853     .device_priv_size       = sizeof(VulkanDevicePriv),
2854     .frames_hwctx_size      = sizeof(AVVulkanFramesContext),
2855     .frames_priv_size       = sizeof(VulkanFramesPriv),
2856
2857     .device_init            = &vulkan_device_init,
2858     .device_create          = &vulkan_device_create,
2859     .device_derive          = &vulkan_device_derive,
2860
2861     .frames_get_constraints = &vulkan_frames_get_constraints,
2862     .frames_init            = vulkan_frames_init,
2863     .frames_get_buffer      = vulkan_get_buffer,
2864     .frames_uninit          = vulkan_frames_uninit,
2865
2866     .transfer_get_formats   = vulkan_transfer_get_formats,
2867     .transfer_data_to       = vulkan_transfer_data_to,
2868     .transfer_data_from     = vulkan_transfer_data_from,
2869
2870     .map_to                 = vulkan_map_to,
2871     .map_from               = vulkan_map_from,
2872
2873     .pix_fmts = (const enum AVPixelFormat []) {
2874         AV_PIX_FMT_VULKAN,
2875         AV_PIX_FMT_NONE
2876     },
2877 };