| File: | root/firefox-clang/media/ffvpx/libavutil/hwcontext_vulkan.c |
| Warning: | line 1215, column 17 Value stored to 'token' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* |
| 2 | * Copyright (c) Lynne |
| 3 | * |
| 4 | * This file is part of FFmpeg. |
| 5 | * |
| 6 | * FFmpeg is free software; you can redistribute it and/or |
| 7 | * modify it under the terms of the GNU Lesser General Public |
| 8 | * License as published by the Free Software Foundation; either |
| 9 | * version 2.1 of the License, or (at your option) any later version. |
| 10 | * |
| 11 | * FFmpeg is distributed in the hope that it will be useful, |
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 14 | * Lesser General Public License for more details. |
| 15 | * |
| 16 | * You should have received a copy of the GNU Lesser General Public |
| 17 | * License along with FFmpeg; if not, write to the Free Software |
| 18 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 19 | */ |
| 20 | |
| 21 | #define VK_NO_PROTOTYPES |
| 22 | #define VK_ENABLE_BETA_EXTENSIONS |
| 23 | |
| 24 | #ifdef _WIN32 |
| 25 | #include <windows.h> /* Included to prevent conflicts with CreateSemaphore */ |
| 26 | #include <versionhelpers.h> |
| 27 | #include "compat/w32dlfcn.h" |
| 28 | #else |
| 29 | #include <dlfcn.h> |
| 30 | #include <unistd.h> |
| 31 | #endif |
| 32 | |
| 33 | #include "thread.h" |
| 34 | |
| 35 | #include "config.h" |
| 36 | #include "pixdesc.h" |
| 37 | #include "avstring.h" |
| 38 | #include "imgutils.h" |
| 39 | #include "hwcontext.h" |
| 40 | #include "hwcontext_internal.h" |
| 41 | #include "hwcontext_vulkan.h" |
| 42 | #include "mem.h" |
| 43 | #include "uuid.h" |
| 44 | |
| 45 | #include "vulkan.h" |
| 46 | #include "vulkan_loader.h" |
| 47 | |
| 48 | #if CONFIG_VAAPI1 |
| 49 | #include "hwcontext_vaapi.h" |
| 50 | #endif |
| 51 | |
| 52 | #if CONFIG_LIBDRM1 |
| 53 | #if CONFIG_VAAPI1 |
| 54 | #include <va/va_drmcommon.h> |
| 55 | #endif |
| 56 | #ifdef __linux__1 |
| 57 | #include <sys/sysmacros.h> |
| 58 | #endif |
| 59 | #include <sys/stat.h> |
| 60 | #include <xf86drm.h> |
| 61 | #include <drm_fourcc.h> |
| 62 | #include "hwcontext_drm.h" |
| 63 | #endif |
| 64 | |
| 65 | #if HAVE_LINUX_DMA_BUF_H0 |
| 66 | #include <sys/ioctl.h> |
| 67 | #include <linux1/dma-buf.h> |
| 68 | #elif defined(__linux__1) |
| 69 | #include <stdint.h> |
| 70 | #include <sys/ioctl.h> |
| 71 | struct dma_buf_export_sync_file { |
| 72 | uint32_t flags; |
| 73 | int32_t fd; |
| 74 | }; |
| 75 | |
| 76 | struct dma_buf_import_sync_file { |
| 77 | uint32_t flags; |
| 78 | int32_t fd; |
| 79 | }; |
| 80 | |
| 81 | #define DMA_BUF_SYNC_READ(1 << 0) (1 << 0) |
| 82 | #define DMA_BUF_SYNC_WRITE(2 << 0) (2 << 0) |
| 83 | #define DMA_BUF_BASE'b' 'b' |
| 84 | #define DMA_BUF_IOCTL_EXPORT_SYNC_FILE(((2U|1U) << (((0 +8)+8)+14)) | ((('b')) << (0 +8 )) | (((2)) << 0) | ((((sizeof(struct dma_buf_export_sync_file )))) << ((0 +8)+8))) \ |
| 85 | _IOWR(DMA_BUF_BASE, 2, struct dma_buf_export_sync_file)(((2U|1U) << (((0 +8)+8)+14)) | ((('b')) << (0 +8 )) | (((2)) << 0) | ((((sizeof(struct dma_buf_export_sync_file )))) << ((0 +8)+8))) |
| 86 | #define DMA_BUF_IOCTL_IMPORT_SYNC_FILE(((1U) << (((0 +8)+8)+14)) | ((('b')) << (0 +8)) | (((3)) << 0) | ((((sizeof(struct dma_buf_import_sync_file )))) << ((0 +8)+8))) \ |
| 87 | _IOW(DMA_BUF_BASE, 3, struct dma_buf_import_sync_file)(((1U) << (((0 +8)+8)+14)) | ((('b')) << (0 +8)) | (((3)) << 0) | ((((sizeof(struct dma_buf_import_sync_file )))) << ((0 +8)+8))) |
| 88 | #endif |
| 89 | |
| 90 | #if CONFIG_CUDA0 |
| 91 | #include "hwcontext_cuda_internal.h" |
| 92 | #include "cuda_check.h" |
| 93 | #define CHECK_CU(x) FF_CUDA_CHECK_DL(cuda_cu, cu, x) |
| 94 | #endif |
| 95 | |
| 96 | typedef struct VulkanDeviceFeatures { |
| 97 | VkPhysicalDeviceFeatures2 device; |
| 98 | |
| 99 | VkPhysicalDeviceVulkan11Features vulkan_1_1; |
| 100 | VkPhysicalDeviceVulkan12Features vulkan_1_2; |
| 101 | VkPhysicalDeviceVulkan13Features vulkan_1_3; |
| 102 | VkPhysicalDeviceTimelineSemaphoreFeatures timeline_semaphore; |
| 103 | VkPhysicalDeviceShaderSubgroupRotateFeaturesKHR subgroup_rotate; |
| 104 | VkPhysicalDeviceHostImageCopyFeaturesEXT host_image_copy; |
| 105 | VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR explicit_mem_layout; |
| 106 | |
| 107 | #ifdef VK_EXT_shader_long_vector1 |
| 108 | VkPhysicalDeviceShaderLongVectorFeaturesEXT long_vector; |
| 109 | #endif |
| 110 | |
| 111 | #ifdef VK_EXT_shader_replicated_composites1 |
| 112 | VkPhysicalDeviceShaderReplicatedCompositesFeaturesEXT replicated_composites; |
| 113 | #endif |
| 114 | |
| 115 | #ifdef VK_EXT_zero_initialize_device_memory1 |
| 116 | VkPhysicalDeviceZeroInitializeDeviceMemoryFeaturesEXT zero_initialize; |
| 117 | #endif |
| 118 | |
| 119 | #ifdef VK_KHR_shader_expect_assume1 |
| 120 | VkPhysicalDeviceShaderExpectAssumeFeaturesKHR expect_assume; |
| 121 | #endif |
| 122 | |
| 123 | #ifdef VK_KHR_shader_maximal_reconvergence1 |
| 124 | VkPhysicalDeviceShaderMaximalReconvergenceFeaturesKHR maximal_reconvergence; |
| 125 | #endif |
| 126 | |
| 127 | #ifdef VK_KHR_maintenance91 |
| 128 | VkPhysicalDeviceMaintenance9FeaturesKHR maintenance_9; |
| 129 | #endif |
| 130 | #ifdef VK_KHR_unified_image_layouts1 |
| 131 | VkPhysicalDeviceUnifiedImageLayoutsFeaturesKHR unified_layouts; |
| 132 | #endif |
| 133 | |
| 134 | VkPhysicalDeviceVideoMaintenance1FeaturesKHR video_maintenance_1; |
| 135 | #ifdef VK_KHR_video_maintenance21 |
| 136 | VkPhysicalDeviceVideoMaintenance2FeaturesKHR video_maintenance_2; |
| 137 | #endif |
| 138 | #ifdef VK_KHR_video_decode_vp91 |
| 139 | VkPhysicalDeviceVideoDecodeVP9FeaturesKHR vp9_decode; |
| 140 | #endif |
| 141 | #ifdef VK_KHR_video_encode_av11 |
| 142 | VkPhysicalDeviceVideoEncodeAV1FeaturesKHR av1_encode; |
| 143 | #endif |
| 144 | |
| 145 | VkPhysicalDeviceShaderObjectFeaturesEXT shader_object; |
| 146 | VkPhysicalDeviceCooperativeMatrixFeaturesKHR cooperative_matrix; |
| 147 | VkPhysicalDeviceShaderAtomicFloatFeaturesEXT atomic_float; |
| 148 | |
| 149 | #ifdef VK_KHR_shader_relaxed_extended_instruction1 |
| 150 | VkPhysicalDeviceShaderRelaxedExtendedInstructionFeaturesKHR relaxed_extended_instruction; |
| 151 | #endif |
| 152 | |
| 153 | #ifdef VK_KHR_internally_synchronized_queues1 |
| 154 | VkPhysicalDeviceInternallySynchronizedQueuesFeaturesKHR internal_queue_sync; |
| 155 | #endif |
| 156 | VkPhysicalDeviceOpticalFlowFeaturesNV optical_flow; |
| 157 | } VulkanDeviceFeatures; |
| 158 | |
| 159 | typedef struct VulkanDevicePriv { |
| 160 | /** |
| 161 | * The public AVVulkanDeviceContext. See hwcontext_vulkan.h for it. |
| 162 | */ |
| 163 | AVVulkanDeviceContext p; |
| 164 | |
| 165 | /* Vulkan library and loader functions */ |
| 166 | void *libvulkan; |
| 167 | |
| 168 | FFVulkanContext vkctx; |
| 169 | AVVulkanDeviceQueueFamily *compute_qf; |
| 170 | AVVulkanDeviceQueueFamily *transfer_qf; |
| 171 | |
| 172 | /* Properties */ |
| 173 | VkPhysicalDeviceProperties2 props; |
| 174 | VkPhysicalDeviceMemoryProperties mprops; |
| 175 | VkPhysicalDeviceExternalMemoryHostPropertiesEXT hprops; |
| 176 | VkPhysicalDeviceDriverProperties dprops; |
| 177 | |
| 178 | /* Opaque FD external semaphore properties */ |
| 179 | VkExternalSemaphoreProperties ext_sem_props_opaque; |
| 180 | |
| 181 | /* Enabled features */ |
| 182 | VulkanDeviceFeatures feats; |
| 183 | |
| 184 | /* Queues */ |
| 185 | pthread_mutex_t **qf_mutex; |
| 186 | uint32_t nb_tot_qfs; |
| 187 | uint32_t img_qfs[64]; |
| 188 | uint32_t nb_img_qfs; |
| 189 | |
| 190 | /* Debug callback */ |
| 191 | VkDebugUtilsMessengerEXT debug_ctx; |
| 192 | |
| 193 | /* Settings */ |
| 194 | int use_linear_images; |
| 195 | |
| 196 | /* Option to allocate all image planes in a single allocation */ |
| 197 | int contiguous_planes; |
| 198 | |
| 199 | /* Disable multiplane images */ |
| 200 | int disable_multiplane; |
| 201 | |
| 202 | /* Prefer memcpy over dynamic host pointer imports */ |
| 203 | int avoid_host_import; |
| 204 | |
| 205 | /* Maximum queues */ |
| 206 | int limit_queues; |
| 207 | } VulkanDevicePriv; |
| 208 | |
| 209 | typedef struct VulkanFramesPriv { |
| 210 | /** |
| 211 | * The public AVVulkanFramesContext. See hwcontext_vulkan.h for it. |
| 212 | */ |
| 213 | AVVulkanFramesContext p; |
| 214 | |
| 215 | /* Image conversions */ |
| 216 | FFVkExecPool compute_exec; |
| 217 | |
| 218 | /* Image transfers */ |
| 219 | FFVkExecPool upload_exec; |
| 220 | FFVkExecPool download_exec; |
| 221 | |
| 222 | /* Temporary buffer pools */ |
| 223 | AVRefStructPool *tmp; |
| 224 | |
| 225 | /* Modifier info list to free at uninit */ |
| 226 | VkImageDrmFormatModifierListCreateInfoEXT *modifier_info; |
| 227 | |
| 228 | /* Properties for DRM modifier for each plane in the image */ |
| 229 | VkDrmFormatModifierPropertiesEXT drm_format_modifier_properties[5]; |
| 230 | |
| 231 | /* Set when physical device reports DEDICATED_ONLY for DMA-BUF export (try_export_flags) */ |
| 232 | int export_requires_dedicated; |
| 233 | } VulkanFramesPriv; |
| 234 | |
| 235 | typedef struct AVVkFrameInternal { |
| 236 | pthread_mutex_t update_mutex; |
| 237 | |
| 238 | /* Binary semaphore for SYNC_FD export at DRM map time. Created once lazily, |
| 239 | * re-signaled each time via a submit in vulkan_map_to_drm. */ |
| 240 | VkSemaphore drm_sync_sem; |
| 241 | |
| 242 | #if CONFIG_CUDA0 |
| 243 | /* Importing external memory into cuda is really expensive so we keep the |
| 244 | * memory imported all the time */ |
| 245 | AVBufferRef *cuda_fc_ref; /* Need to keep it around for uninit */ |
| 246 | CUexternalMemory ext_mem[AV_NUM_DATA_POINTERS8]; |
| 247 | CUmipmappedArray cu_mma[AV_NUM_DATA_POINTERS8]; |
| 248 | CUarray cu_array[AV_NUM_DATA_POINTERS8]; |
| 249 | CUexternalSemaphore cu_sem[AV_NUM_DATA_POINTERS8]; |
| 250 | #ifdef _WIN32 |
| 251 | HANDLE ext_mem_handle[AV_NUM_DATA_POINTERS8]; |
| 252 | HANDLE ext_sem_handle[AV_NUM_DATA_POINTERS8]; |
| 253 | #endif |
| 254 | #endif |
| 255 | } AVVkFrameInternal; |
| 256 | |
| 257 | /* Initialize all structs in VulkanDeviceFeatures */ |
| 258 | static void device_features_init(AVHWDeviceContext *ctx, VulkanDeviceFeatures *feats) |
| 259 | { |
| 260 | VulkanDevicePriv *p = ctx->hwctx; |
| 261 | FFVulkanContext *s = &p->vkctx; |
| 262 | |
| 263 | feats->device = (VkPhysicalDeviceFeatures2) { |
| 264 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2, |
| 265 | }; |
| 266 | |
| 267 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->vulkan_1_1, FF_VK_EXT_NO_FLAG,do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&feats->vulkan_1_1 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES ; ff_vk_link_struct(&feats->device, &feats->vulkan_1_1 ); } } while (0) |
| 268 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES)do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&feats->vulkan_1_1 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES ; ff_vk_link_struct(&feats->device, &feats->vulkan_1_1 ); } } while (0); |
| 269 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->vulkan_1_2, FF_VK_EXT_NO_FLAG,do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&feats->vulkan_1_2 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES ; ff_vk_link_struct(&feats->device, &feats->vulkan_1_2 ); } } while (0) |
| 270 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES)do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&feats->vulkan_1_2 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES ; ff_vk_link_struct(&feats->device, &feats->vulkan_1_2 ); } } while (0); |
| 271 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->vulkan_1_3, FF_VK_EXT_NO_FLAG,do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&feats->vulkan_1_3 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_FEATURES ; ff_vk_link_struct(&feats->device, &feats->vulkan_1_3 ); } } while (0) |
| 272 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_FEATURES)do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&feats->vulkan_1_3 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_FEATURES ; ff_vk_link_struct(&feats->device, &feats->vulkan_1_3 ); } } while (0); |
| 273 | |
| 274 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->timeline_semaphore, FF_VK_EXT_PORTABILITY_SUBSET,do { if (((1ULL << 62) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 62))) { (&feats->timeline_semaphore )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES ; ff_vk_link_struct(&feats->device, &feats->timeline_semaphore ); } } while (0) |
| 275 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES)do { if (((1ULL << 62) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 62))) { (&feats->timeline_semaphore )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES ; ff_vk_link_struct(&feats->device, &feats->timeline_semaphore ); } } while (0); |
| 276 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->subgroup_rotate, FF_VK_EXT_SUBGROUP_ROTATE,do { if (((1ULL << 17) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 17))) { (&feats->subgroup_rotate )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_ROTATE_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->subgroup_rotate ); } } while (0) |
| 277 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_ROTATE_FEATURES_KHR)do { if (((1ULL << 17) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 17))) { (&feats->subgroup_rotate )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_ROTATE_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->subgroup_rotate ); } } while (0); |
| 278 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->host_image_copy, FF_VK_EXT_HOST_IMAGE_COPY,do { if (((1ULL << 18) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 18))) { (&feats->host_image_copy )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_IMAGE_COPY_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->host_image_copy ); } } while (0) |
| 279 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_IMAGE_COPY_FEATURES_EXT)do { if (((1ULL << 18) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 18))) { (&feats->host_image_copy )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_IMAGE_COPY_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->host_image_copy ); } } while (0); |
| 280 | |
| 281 | #ifdef VK_EXT_shader_long_vector1 |
| 282 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->long_vector, FF_VK_EXT_LONG_VECTOR,do { if (((1ULL << 22) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 22))) { (&feats->long_vector )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_LONG_VECTOR_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->long_vector ); } } while (0) |
| 283 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_LONG_VECTOR_FEATURES_EXT)do { if (((1ULL << 22) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 22))) { (&feats->long_vector )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_LONG_VECTOR_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->long_vector ); } } while (0); |
| 284 | #endif |
| 285 | |
| 286 | #ifdef VK_EXT_shader_replicated_composites1 |
| 287 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->replicated_composites, FF_VK_EXT_REPLICATED_COMPOSITES,do { if (((1ULL << 21) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 21))) { (&feats->replicated_composites )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_REPLICATED_COMPOSITES_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->replicated_composites ); } } while (0) |
| 288 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_REPLICATED_COMPOSITES_FEATURES_EXT)do { if (((1ULL << 21) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 21))) { (&feats->replicated_composites )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_REPLICATED_COMPOSITES_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->replicated_composites ); } } while (0); |
| 289 | #endif |
| 290 | |
| 291 | #ifdef VK_EXT_zero_initialize_device_memory1 |
| 292 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->zero_initialize, FF_VK_EXT_ZERO_INITIALIZE,do { if (((1ULL << 19) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 19))) { (&feats->zero_initialize )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ZERO_INITIALIZE_DEVICE_MEMORY_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->zero_initialize ); } } while (0) |
| 293 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ZERO_INITIALIZE_DEVICE_MEMORY_FEATURES_EXT)do { if (((1ULL << 19) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 19))) { (&feats->zero_initialize )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ZERO_INITIALIZE_DEVICE_MEMORY_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->zero_initialize ); } } while (0); |
| 294 | #endif |
| 295 | |
| 296 | #ifdef VK_KHR_shader_expect_assume1 |
| 297 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->expect_assume, FF_VK_EXT_EXPECT_ASSUME,do { if (((1ULL << 16) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 16))) { (&feats->expect_assume )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_EXPECT_ASSUME_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->expect_assume ); } } while (0) |
| 298 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_EXPECT_ASSUME_FEATURES_KHR)do { if (((1ULL << 16) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 16))) { (&feats->expect_assume )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_EXPECT_ASSUME_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->expect_assume ); } } while (0); |
| 299 | #endif |
| 300 | |
| 301 | #ifdef VK_KHR_shader_maximal_reconvergence1 |
| 302 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->maximal_reconvergence, FF_VK_EXT_MAXIMAL_RECONVERGENCE,do { if (((1ULL << 24) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 24))) { (&feats->maximal_reconvergence )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_MAXIMAL_RECONVERGENCE_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->maximal_reconvergence ); } } while (0) |
| 303 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_MAXIMAL_RECONVERGENCE_FEATURES_KHR)do { if (((1ULL << 24) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 24))) { (&feats->maximal_reconvergence )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_MAXIMAL_RECONVERGENCE_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->maximal_reconvergence ); } } while (0); |
| 304 | #endif |
| 305 | |
| 306 | #ifdef VK_KHR_maintenance91 |
| 307 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->maintenance_9, FF_VK_EXT_MAINTENANCE_9,do { if (((1ULL << 25) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 25))) { (&feats->maintenance_9 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_9_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->maintenance_9 ); } } while (0) |
| 308 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_9_FEATURES_KHR)do { if (((1ULL << 25) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 25))) { (&feats->maintenance_9 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_9_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->maintenance_9 ); } } while (0); |
| 309 | #endif |
| 310 | #ifdef VK_KHR_unified_image_layouts1 |
| 311 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->unified_layouts, FF_VK_EXT_UNIFIED_IMAGE_LAYOUTS,do { if (((1ULL << 26) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 26))) { (&feats->unified_layouts )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFIED_IMAGE_LAYOUTS_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->unified_layouts ); } } while (0) |
| 312 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFIED_IMAGE_LAYOUTS_FEATURES_KHR)do { if (((1ULL << 26) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 26))) { (&feats->unified_layouts )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFIED_IMAGE_LAYOUTS_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->unified_layouts ); } } while (0); |
| 313 | #endif |
| 314 | |
| 315 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->video_maintenance_1, FF_VK_EXT_VIDEO_MAINTENANCE_1,do { if (((1ULL << 37) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 37))) { (&feats->video_maintenance_1 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_MAINTENANCE_1_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->video_maintenance_1 ); } } while (0) |
| 316 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_MAINTENANCE_1_FEATURES_KHR)do { if (((1ULL << 37) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 37))) { (&feats->video_maintenance_1 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_MAINTENANCE_1_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->video_maintenance_1 ); } } while (0); |
| 317 | #ifdef VK_KHR_video_maintenance21 |
| 318 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->video_maintenance_2, FF_VK_EXT_VIDEO_MAINTENANCE_2,do { if (((1ULL << 38) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 38))) { (&feats->video_maintenance_2 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_MAINTENANCE_2_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->video_maintenance_2 ); } } while (0) |
| 319 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_MAINTENANCE_2_FEATURES_KHR)do { if (((1ULL << 38) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 38))) { (&feats->video_maintenance_2 )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_MAINTENANCE_2_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->video_maintenance_2 ); } } while (0); |
| 320 | #endif |
| 321 | #ifdef VK_KHR_video_decode_vp91 |
| 322 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->vp9_decode, FF_VK_EXT_VIDEO_DECODE_VP9,do { if (((1ULL << 43) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 43))) { (&feats->vp9_decode )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_DECODE_VP9_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->vp9_decode ); } } while (0) |
| 323 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_DECODE_VP9_FEATURES_KHR)do { if (((1ULL << 43) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 43))) { (&feats->vp9_decode )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_DECODE_VP9_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->vp9_decode ); } } while (0); |
| 324 | #endif |
| 325 | #ifdef VK_KHR_video_encode_av11 |
| 326 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->av1_encode, FF_VK_EXT_VIDEO_ENCODE_AV1,do { if (((1ULL << 53) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 53))) { (&feats->av1_encode )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_ENCODE_AV1_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->av1_encode ); } } while (0) |
| 327 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_ENCODE_AV1_FEATURES_KHR)do { if (((1ULL << 53) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 53))) { (&feats->av1_encode )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VIDEO_ENCODE_AV1_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->av1_encode ); } } while (0); |
| 328 | #endif |
| 329 | |
| 330 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->shader_object, FF_VK_EXT_SHADER_OBJECT,do { if (((1ULL << 13) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 13))) { (&feats->shader_object )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_OBJECT_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->shader_object ); } } while (0) |
| 331 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_OBJECT_FEATURES_EXT)do { if (((1ULL << 13) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 13))) { (&feats->shader_object )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_OBJECT_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->shader_object ); } } while (0); |
| 332 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->cooperative_matrix, FF_VK_EXT_COOP_MATRIX,do { if (((1ULL << 11) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 11))) { (&feats->cooperative_matrix )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->cooperative_matrix ); } } while (0) |
| 333 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR)do { if (((1ULL << 11) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 11))) { (&feats->cooperative_matrix )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->cooperative_matrix ); } } while (0); |
| 334 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->atomic_float, FF_VK_EXT_ATOMIC_FLOAT,do { if (((1ULL << 10) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 10))) { (&feats->atomic_float )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->atomic_float ); } } while (0) |
| 335 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT)do { if (((1ULL << 10) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 10))) { (&feats->atomic_float )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT ; ff_vk_link_struct(&feats->device, &feats->atomic_float ); } } while (0); |
| 336 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->explicit_mem_layout, FF_VK_EXT_EXPLICIT_MEM_LAYOUT,do { if (((1ULL << 20) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 20))) { (&feats->explicit_mem_layout )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->explicit_mem_layout ); } } while (0) |
| 337 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR)do { if (((1ULL << 20) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 20))) { (&feats->explicit_mem_layout )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->explicit_mem_layout ); } } while (0); |
| 338 | |
| 339 | #ifdef VK_KHR_shader_relaxed_extended_instruction1 |
| 340 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->relaxed_extended_instruction, FF_VK_EXT_RELAXED_EXTENDED_INSTR,do { if (((1ULL << 15) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 15))) { (&feats->relaxed_extended_instruction )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_RELAXED_EXTENDED_INSTRUCTION_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->relaxed_extended_instruction ); } } while (0) |
| 341 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_RELAXED_EXTENDED_INSTRUCTION_FEATURES_KHR)do { if (((1ULL << 15) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 15))) { (&feats->relaxed_extended_instruction )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_RELAXED_EXTENDED_INSTRUCTION_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->relaxed_extended_instruction ); } } while (0); |
| 342 | #endif |
| 343 | |
| 344 | #ifdef VK_KHR_internally_synchronized_queues1 |
| 345 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->internal_queue_sync, FF_VK_EXT_INTERNAL_QUEUE_SYNC,do { if (((1ULL << 23) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 23))) { (&feats->internal_queue_sync )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INTERNALLY_SYNCHRONIZED_QUEUES_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->internal_queue_sync ); } } while (0) |
| 346 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INTERNALLY_SYNCHRONIZED_QUEUES_FEATURES_KHR)do { if (((1ULL << 23) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 23))) { (&feats->internal_queue_sync )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INTERNALLY_SYNCHRONIZED_QUEUES_FEATURES_KHR ; ff_vk_link_struct(&feats->device, &feats->internal_queue_sync ); } } while (0); |
| 347 | #endif |
| 348 | |
| 349 | FF_VK_STRUCT_EXT(s, &feats->device, &feats->optical_flow, FF_VK_EXT_OPTICAL_FLOW,do { if (((1ULL << 12) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 12))) { (&feats->optical_flow )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_OPTICAL_FLOW_FEATURES_NV ; ff_vk_link_struct(&feats->device, &feats->optical_flow ); } } while (0) |
| 350 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_OPTICAL_FLOW_FEATURES_NV)do { if (((1ULL << 12) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 12))) { (&feats->optical_flow )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_OPTICAL_FLOW_FEATURES_NV ; ff_vk_link_struct(&feats->device, &feats->optical_flow ); } } while (0); |
| 351 | } |
| 352 | |
| 353 | /* Copy all needed device features */ |
| 354 | static void device_features_copy_needed(VulkanDeviceFeatures *dst, VulkanDeviceFeatures *src) |
| 355 | { |
| 356 | #define COPY_VAL(VAL) \ |
| 357 | do { \ |
| 358 | dst->VAL = src->VAL; \ |
| 359 | } while (0) \ |
| 360 | |
| 361 | COPY_VAL(device.features.shaderImageGatherExtended); |
| 362 | COPY_VAL(device.features.shaderStorageImageReadWithoutFormat); |
| 363 | COPY_VAL(device.features.shaderStorageImageWriteWithoutFormat); |
| 364 | COPY_VAL(device.features.fragmentStoresAndAtomics); |
| 365 | COPY_VAL(device.features.vertexPipelineStoresAndAtomics); |
| 366 | COPY_VAL(device.features.shaderInt64); |
| 367 | COPY_VAL(device.features.shaderInt16); |
| 368 | COPY_VAL(device.features.shaderFloat64); |
| 369 | COPY_VAL(device.features.shaderStorageImageReadWithoutFormat); |
| 370 | COPY_VAL(device.features.shaderStorageImageWriteWithoutFormat); |
| 371 | |
| 372 | COPY_VAL(vulkan_1_1.samplerYcbcrConversion); |
| 373 | COPY_VAL(vulkan_1_1.storagePushConstant16); |
| 374 | COPY_VAL(vulkan_1_1.storageBuffer16BitAccess); |
| 375 | COPY_VAL(vulkan_1_1.uniformAndStorageBuffer16BitAccess); |
| 376 | |
| 377 | COPY_VAL(vulkan_1_2.timelineSemaphore); |
| 378 | COPY_VAL(vulkan_1_2.scalarBlockLayout); |
| 379 | COPY_VAL(vulkan_1_2.bufferDeviceAddress); |
| 380 | COPY_VAL(vulkan_1_2.hostQueryReset); |
| 381 | COPY_VAL(vulkan_1_2.storagePushConstant8); |
| 382 | COPY_VAL(vulkan_1_2.shaderInt8); |
| 383 | COPY_VAL(vulkan_1_2.storageBuffer8BitAccess); |
| 384 | COPY_VAL(vulkan_1_2.uniformAndStorageBuffer8BitAccess); |
| 385 | COPY_VAL(vulkan_1_2.shaderFloat16); |
| 386 | COPY_VAL(vulkan_1_2.shaderBufferInt64Atomics); |
| 387 | COPY_VAL(vulkan_1_2.shaderSharedInt64Atomics); |
| 388 | COPY_VAL(vulkan_1_2.vulkanMemoryModel); |
| 389 | COPY_VAL(vulkan_1_2.vulkanMemoryModelDeviceScope); |
| 390 | COPY_VAL(vulkan_1_2.vulkanMemoryModelAvailabilityVisibilityChains); |
| 391 | COPY_VAL(vulkan_1_2.uniformBufferStandardLayout); |
| 392 | COPY_VAL(vulkan_1_2.runtimeDescriptorArray); |
| 393 | COPY_VAL(vulkan_1_2.shaderSubgroupExtendedTypes); |
| 394 | COPY_VAL(vulkan_1_2.shaderUniformBufferArrayNonUniformIndexing); |
| 395 | COPY_VAL(vulkan_1_2.shaderSampledImageArrayNonUniformIndexing); |
| 396 | COPY_VAL(vulkan_1_2.shaderStorageBufferArrayNonUniformIndexing); |
| 397 | COPY_VAL(vulkan_1_2.shaderStorageImageArrayNonUniformIndexing); |
| 398 | |
| 399 | COPY_VAL(vulkan_1_3.dynamicRendering); |
| 400 | COPY_VAL(vulkan_1_3.maintenance4); |
| 401 | COPY_VAL(vulkan_1_3.synchronization2); |
| 402 | COPY_VAL(vulkan_1_3.computeFullSubgroups); |
| 403 | COPY_VAL(vulkan_1_3.subgroupSizeControl); |
| 404 | COPY_VAL(vulkan_1_3.shaderZeroInitializeWorkgroupMemory); |
| 405 | COPY_VAL(vulkan_1_3.dynamicRendering); |
| 406 | |
| 407 | COPY_VAL(timeline_semaphore.timelineSemaphore); |
| 408 | COPY_VAL(subgroup_rotate.shaderSubgroupRotate); |
| 409 | COPY_VAL(host_image_copy.hostImageCopy); |
| 410 | |
| 411 | #ifdef VK_EXT_shader_long_vector1 |
| 412 | COPY_VAL(long_vector.longVector); |
| 413 | #endif |
| 414 | |
| 415 | #ifdef VK_EXT_shader_replicated_composites1 |
| 416 | COPY_VAL(replicated_composites.shaderReplicatedComposites); |
| 417 | #endif |
| 418 | |
| 419 | #ifdef VK_EXT_zero_initialize_device_memory1 |
| 420 | COPY_VAL(zero_initialize.zeroInitializeDeviceMemory); |
| 421 | #endif |
| 422 | |
| 423 | COPY_VAL(video_maintenance_1.videoMaintenance1); |
| 424 | #ifdef VK_KHR_video_maintenance21 |
| 425 | COPY_VAL(video_maintenance_2.videoMaintenance2); |
| 426 | #endif |
| 427 | |
| 428 | #ifdef VK_KHR_video_decode_vp91 |
| 429 | COPY_VAL(vp9_decode.videoDecodeVP9); |
| 430 | #endif |
| 431 | |
| 432 | #ifdef VK_KHR_video_encode_av11 |
| 433 | COPY_VAL(av1_encode.videoEncodeAV1); |
| 434 | #endif |
| 435 | |
| 436 | COPY_VAL(shader_object.shaderObject); |
| 437 | |
| 438 | COPY_VAL(cooperative_matrix.cooperativeMatrix); |
| 439 | |
| 440 | COPY_VAL(atomic_float.shaderBufferFloat32Atomics); |
| 441 | COPY_VAL(atomic_float.shaderBufferFloat32AtomicAdd); |
| 442 | |
| 443 | COPY_VAL(explicit_mem_layout.workgroupMemoryExplicitLayout); |
| 444 | COPY_VAL(explicit_mem_layout.workgroupMemoryExplicitLayoutScalarBlockLayout); |
| 445 | COPY_VAL(explicit_mem_layout.workgroupMemoryExplicitLayout8BitAccess); |
| 446 | COPY_VAL(explicit_mem_layout.workgroupMemoryExplicitLayout16BitAccess); |
| 447 | |
| 448 | #ifdef VK_KHR_shader_relaxed_extended_instruction1 |
| 449 | COPY_VAL(relaxed_extended_instruction.shaderRelaxedExtendedInstruction); |
| 450 | #endif |
| 451 | |
| 452 | #ifdef VK_KHR_shader_expect_assume1 |
| 453 | COPY_VAL(expect_assume.shaderExpectAssume); |
| 454 | #endif |
| 455 | |
| 456 | #ifdef VK_KHR_shader_maximal_reconvergence1 |
| 457 | COPY_VAL(maximal_reconvergence.shaderMaximalReconvergence); |
| 458 | #endif |
| 459 | |
| 460 | #ifdef VK_KHR_maintenance91 |
| 461 | COPY_VAL(maintenance_9.maintenance9); |
| 462 | #endif |
| 463 | #ifdef VK_KHR_unified_image_layouts1 |
| 464 | COPY_VAL(unified_layouts.unifiedImageLayouts); |
| 465 | COPY_VAL(unified_layouts.unifiedImageLayoutsVideo); |
| 466 | #endif |
| 467 | |
| 468 | #ifdef VK_KHR_internally_synchronized_queues1 |
| 469 | COPY_VAL(internal_queue_sync.internallySynchronizedQueues); |
| 470 | #endif |
| 471 | |
| 472 | COPY_VAL(optical_flow.opticalFlow); |
| 473 | |
| 474 | #undef COPY_VAL |
| 475 | } |
| 476 | |
| 477 | #define ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT) (VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT) |
| 478 | #define ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ) (VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT) |
| 479 | |
| 480 | static const struct FFVkFormatEntry { |
| 481 | VkFormat vkf; |
| 482 | enum AVPixelFormat pixfmt; |
| 483 | VkImageAspectFlags aspect; |
| 484 | int vk_planes; |
| 485 | int nb_images; |
| 486 | int nb_images_fallback; |
| 487 | const VkFormat fallback[5]; |
| 488 | } vk_formats_list[] = { |
| 489 | /* Gray formats */ |
| 490 | { VK_FORMAT_R8_UNORM, AV_PIX_FMT_GRAY8, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R8_UNORM } }, |
| 491 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GRAY10AV_PIX_FMT_GRAY10LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16_UNORM } }, |
| 492 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GRAY12AV_PIX_FMT_GRAY12LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16_UNORM } }, |
| 493 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GRAY14AV_PIX_FMT_GRAY14LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16_UNORM } }, |
| 494 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GRAY16AV_PIX_FMT_GRAY16LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16_UNORM } }, |
| 495 | { VK_FORMAT_R32_UINT, AV_PIX_FMT_GRAY32AV_PIX_FMT_GRAY32LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R32_UINT } }, |
| 496 | { VK_FORMAT_R32_SFLOAT, AV_PIX_FMT_GRAYF32AV_PIX_FMT_GRAYF32LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R32_SFLOAT } }, |
| 497 | |
| 498 | /* RGB formats */ |
| 499 | { VK_FORMAT_R8G8B8A8_UNORM, AV_PIX_FMT_BGRA, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R8G8B8A8_UNORM } }, |
| 500 | { VK_FORMAT_R8G8B8A8_UNORM, AV_PIX_FMT_RGBA, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R8G8B8A8_UNORM } }, |
| 501 | { VK_FORMAT_R8G8B8_UNORM, AV_PIX_FMT_RGB24, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R8G8B8_UNORM } }, |
| 502 | { VK_FORMAT_B8G8R8_UNORM, AV_PIX_FMT_BGR24, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_B8G8R8_UNORM } }, |
| 503 | { VK_FORMAT_R16G16B16_UNORM, AV_PIX_FMT_RGB48AV_PIX_FMT_RGB48LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16G16B16_UNORM } }, |
| 504 | { VK_FORMAT_R16G16B16A16_UNORM, AV_PIX_FMT_RGBA64AV_PIX_FMT_RGBA64LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16G16B16A16_UNORM } }, |
| 505 | { VK_FORMAT_R8G8B8A8_UNORM, AV_PIX_FMT_BGR0, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R8G8B8A8_UNORM } }, |
| 506 | { VK_FORMAT_R8G8B8A8_UNORM, AV_PIX_FMT_RGB0, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R8G8B8A8_UNORM } }, |
| 507 | { VK_FORMAT_A2R10G10B10_UNORM_PACK32, AV_PIX_FMT_X2RGB10AV_PIX_FMT_X2RGB10LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_A2R10G10B10_UNORM_PACK32 } }, |
| 508 | { VK_FORMAT_A2B10G10R10_UNORM_PACK32, AV_PIX_FMT_X2BGR10AV_PIX_FMT_X2BGR10LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_A2B10G10R10_UNORM_PACK32 } }, |
| 509 | { VK_FORMAT_R32G32B32_SFLOAT, AV_PIX_FMT_RGBF32AV_PIX_FMT_RGBF32LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R32G32B32_SFLOAT } }, |
| 510 | { VK_FORMAT_R16G16B16A16_SFLOAT, AV_PIX_FMT_RGBAF16AV_PIX_FMT_RGBAF16LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16G16B16A16_SFLOAT } }, |
| 511 | { VK_FORMAT_R32G32B32A32_SFLOAT, AV_PIX_FMT_RGBAF32AV_PIX_FMT_RGBAF32LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R32G32B32A32_SFLOAT } }, |
| 512 | { VK_FORMAT_R32G32B32_UINT, AV_PIX_FMT_RGB96AV_PIX_FMT_RGB96LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R32G32B32_UINT } }, |
| 513 | { VK_FORMAT_R32G32B32A32_UINT, AV_PIX_FMT_RGBA128AV_PIX_FMT_RGBA128LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R32G32B32A32_UINT } }, |
| 514 | |
| 515 | /* Planar RGB */ |
| 516 | { VK_FORMAT_R8_UNORM, AV_PIX_FMT_GBRP, VK_IMAGE_ASPECT_COLOR_BIT, 3, 3, 3, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } }, |
| 517 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GBRP10AV_PIX_FMT_GBRP10LE, VK_IMAGE_ASPECT_COLOR_BIT, 3, 3, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 518 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GBRP12AV_PIX_FMT_GBRP12LE, VK_IMAGE_ASPECT_COLOR_BIT, 3, 3, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 519 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GBRP14AV_PIX_FMT_GBRP14LE, VK_IMAGE_ASPECT_COLOR_BIT, 3, 3, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 520 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GBRP16AV_PIX_FMT_GBRP16LE, VK_IMAGE_ASPECT_COLOR_BIT, 3, 3, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 521 | { VK_FORMAT_R16_SFLOAT, AV_PIX_FMT_GBRPF16AV_PIX_FMT_GBRPF16LE, VK_IMAGE_ASPECT_COLOR_BIT, 3, 3, 3, { VK_FORMAT_R16_SFLOAT, VK_FORMAT_R16_SFLOAT, VK_FORMAT_R16_SFLOAT } }, |
| 522 | { VK_FORMAT_R32_SFLOAT, AV_PIX_FMT_GBRPF32AV_PIX_FMT_GBRPF32LE, VK_IMAGE_ASPECT_COLOR_BIT, 3, 3, 3, { VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT } }, |
| 523 | |
| 524 | /* Planar RGB + Alpha */ |
| 525 | { VK_FORMAT_R8_UNORM, AV_PIX_FMT_GBRAP, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } }, |
| 526 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GBRAP10AV_PIX_FMT_GBRAP10LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 527 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GBRAP12AV_PIX_FMT_GBRAP12LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 528 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GBRAP14AV_PIX_FMT_GBRAP14LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 529 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_GBRAP16AV_PIX_FMT_GBRAP16LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 530 | { VK_FORMAT_R16_SFLOAT, AV_PIX_FMT_GBRAPF16AV_PIX_FMT_GBRAPF16LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_SFLOAT, VK_FORMAT_R16_SFLOAT, VK_FORMAT_R16_SFLOAT, VK_FORMAT_R16_SFLOAT } }, |
| 531 | { VK_FORMAT_R32_UINT, AV_PIX_FMT_GBRAP32AV_PIX_FMT_GBRAP32LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R32_UINT, VK_FORMAT_R32_UINT, VK_FORMAT_R32_UINT, VK_FORMAT_R32_UINT } }, |
| 532 | { VK_FORMAT_R32_SFLOAT, AV_PIX_FMT_GBRAPF32AV_PIX_FMT_GBRAPF32LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT } }, |
| 533 | |
| 534 | /* Bayer */ |
| 535 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_BAYER_RGGB16AV_PIX_FMT_BAYER_RGGB16LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16_UNORM } }, |
| 536 | |
| 537 | /* Two-plane 420 YUV at 8, 10, 12 and 16 bits */ |
| 538 | { VK_FORMAT_G8_B8R8_2PLANE_420_UNORM, AV_PIX_FMT_NV12, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } }, |
| 539 | { VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16, AV_PIX_FMT_P010AV_PIX_FMT_P010LE, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } }, |
| 540 | { VK_FORMAT_G12X4_B12X4R12X4_2PLANE_420_UNORM_3PACK16, AV_PIX_FMT_P012AV_PIX_FMT_P012LE, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } }, |
| 541 | { VK_FORMAT_G16_B16R16_2PLANE_420_UNORM, AV_PIX_FMT_P016AV_PIX_FMT_P016LE, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } }, |
| 542 | |
| 543 | /* Two-plane 422 YUV at 8, 10 and 16 bits */ |
| 544 | { VK_FORMAT_G8_B8R8_2PLANE_422_UNORM, AV_PIX_FMT_NV16, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } }, |
| 545 | { VK_FORMAT_G10X6_B10X6R10X6_2PLANE_422_UNORM_3PACK16, AV_PIX_FMT_P210AV_PIX_FMT_P210LE, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } }, |
| 546 | { VK_FORMAT_G12X4_B12X4R12X4_2PLANE_422_UNORM_3PACK16, AV_PIX_FMT_P212AV_PIX_FMT_P212LE, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } }, |
| 547 | { VK_FORMAT_G16_B16R16_2PLANE_422_UNORM, AV_PIX_FMT_P216AV_PIX_FMT_P216LE, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } }, |
| 548 | |
| 549 | /* Two-plane 444 YUV at 8, 10 and 16 bits */ |
| 550 | { VK_FORMAT_G8_B8R8_2PLANE_444_UNORM, AV_PIX_FMT_NV24, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } }, |
| 551 | { VK_FORMAT_G10X6_B10X6R10X6_2PLANE_444_UNORM_3PACK16, AV_PIX_FMT_P410AV_PIX_FMT_P410LE, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } }, |
| 552 | { VK_FORMAT_G12X4_B12X4R12X4_2PLANE_444_UNORM_3PACK16, AV_PIX_FMT_P412AV_PIX_FMT_P412LE, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } }, |
| 553 | { VK_FORMAT_G16_B16R16_2PLANE_444_UNORM, AV_PIX_FMT_P416AV_PIX_FMT_P416LE, ASPECT_2PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT), 2, 1, 2, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } }, |
| 554 | |
| 555 | /* Three-plane 420, 422, 444 at 8, 10, 12 and 16 bits */ |
| 556 | { VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM, AV_PIX_FMT_YUV420P, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } }, |
| 557 | { VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM, AV_PIX_FMT_YUV420P10AV_PIX_FMT_YUV420P10LE, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 558 | { VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM, AV_PIX_FMT_YUV420P12AV_PIX_FMT_YUV420P12LE, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 559 | { VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM, AV_PIX_FMT_YUV420P16AV_PIX_FMT_YUV420P16LE, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 560 | { VK_FORMAT_G8_B8_R8_3PLANE_422_UNORM, AV_PIX_FMT_YUV422P, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } }, |
| 561 | { VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM, AV_PIX_FMT_YUV422P10AV_PIX_FMT_YUV422P10LE, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 562 | { VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM, AV_PIX_FMT_YUV422P12AV_PIX_FMT_YUV422P12LE, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 563 | { VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM, AV_PIX_FMT_YUV422P16AV_PIX_FMT_YUV422P16LE, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 564 | { VK_FORMAT_G8_B8_R8_3PLANE_444_UNORM, AV_PIX_FMT_YUV444P, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } }, |
| 565 | { VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM, AV_PIX_FMT_YUV444P10AV_PIX_FMT_YUV444P10LE, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 566 | { VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM, AV_PIX_FMT_YUV444P12AV_PIX_FMT_YUV444P12LE, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 567 | { VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM, AV_PIX_FMT_YUV444P16AV_PIX_FMT_YUV444P16LE, ASPECT_3PLANE(VK_IMAGE_ASPECT_PLANE_0_BIT | VK_IMAGE_ASPECT_PLANE_1_BIT | VK_IMAGE_ASPECT_PLANE_2_BIT ), 3, 1, 3, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 568 | |
| 569 | /* Single plane 422 at 8, 10, 12 and 16 bits */ |
| 570 | { VK_FORMAT_G8B8G8R8_422_UNORM, AV_PIX_FMT_YUYV422, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R8G8B8A8_UNORM } }, |
| 571 | { VK_FORMAT_B8G8R8G8_422_UNORM, AV_PIX_FMT_UYVY422, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R8G8B8A8_UNORM } }, |
| 572 | { VK_FORMAT_G10X6B10X6G10X6R10X6_422_UNORM_4PACK16, AV_PIX_FMT_Y210AV_PIX_FMT_Y210LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16G16B16A16_UNORM } }, |
| 573 | { VK_FORMAT_G12X4B12X4G12X4R12X4_422_UNORM_4PACK16, AV_PIX_FMT_Y212AV_PIX_FMT_Y212LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16G16B16A16_UNORM } }, |
| 574 | { VK_FORMAT_G16B16G16R16_422_UNORM, AV_PIX_FMT_Y216AV_PIX_FMT_Y216LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16G16B16A16_UNORM } }, |
| 575 | |
| 576 | /* Planar YUVA 420 at 8, 10 and 16 bits */ |
| 577 | { VK_FORMAT_R8_UNORM, AV_PIX_FMT_YUVA420P, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } }, |
| 578 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_YUVA420P10AV_PIX_FMT_YUVA420P10LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 579 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_YUVA420P16AV_PIX_FMT_YUVA420P16LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 580 | |
| 581 | /* Planar YUVA 422 at 8, 10, 12 and 16 bits */ |
| 582 | { VK_FORMAT_R8_UNORM, AV_PIX_FMT_YUVA422P, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } }, |
| 583 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_YUVA422P10AV_PIX_FMT_YUVA422P10LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 584 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_YUVA422P12AV_PIX_FMT_YUVA422P12LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 585 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_YUVA422P16AV_PIX_FMT_YUVA422P16LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 586 | |
| 587 | /* Planar YUVA 444 at 8, 10, 12 and 16 bits */ |
| 588 | { VK_FORMAT_R8_UNORM, AV_PIX_FMT_YUVA444P, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } }, |
| 589 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_YUVA444P10AV_PIX_FMT_YUVA444P10LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 590 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_YUVA444P12AV_PIX_FMT_YUVA444P12LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 591 | { VK_FORMAT_R16_UNORM, AV_PIX_FMT_YUVA444P16AV_PIX_FMT_YUVA444P16LE, VK_IMAGE_ASPECT_COLOR_BIT, 4, 4, 4, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } }, |
| 592 | |
| 593 | /* Single plane 444 at 8, 10, 12 and 16 bits */ |
| 594 | { VK_FORMAT_A2R10G10B10_UNORM_PACK32, AV_PIX_FMT_XV30AV_PIX_FMT_XV30LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_A2R10G10B10_UNORM_PACK32 } }, |
| 595 | { VK_FORMAT_R12X4G12X4B12X4A12X4_UNORM_4PACK16, AV_PIX_FMT_XV36AV_PIX_FMT_XV36LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16G16B16A16_UNORM } }, |
| 596 | { VK_FORMAT_R16G16B16A16_UNORM, AV_PIX_FMT_XV48AV_PIX_FMT_XV48LE, VK_IMAGE_ASPECT_COLOR_BIT, 1, 1, 1, { VK_FORMAT_R16G16B16A16_UNORM } }, |
| 597 | }; |
| 598 | static const int nb_vk_formats_list = FF_ARRAY_ELEMS(vk_formats_list)(sizeof(vk_formats_list) / sizeof((vk_formats_list)[0])); |
| 599 | |
| 600 | const VkFormat *av_vkfmt_from_pixfmt(enum AVPixelFormat p) |
| 601 | { |
| 602 | for (int i = 0; i < nb_vk_formats_list; i++) |
| 603 | if (vk_formats_list[i].pixfmt == p) |
| 604 | return vk_formats_list[i].fallback; |
| 605 | return NULL((void*)0); |
| 606 | } |
| 607 | |
| 608 | static const struct FFVkFormatEntry *vk_find_format_entry(enum AVPixelFormat p) |
| 609 | { |
| 610 | for (int i = 0; i < nb_vk_formats_list; i++) |
| 611 | if (vk_formats_list[i].pixfmt == p) |
| 612 | return &vk_formats_list[i]; |
| 613 | return NULL((void*)0); |
| 614 | } |
| 615 | |
| 616 | static int vkfmt_from_pixfmt2(AVHWDeviceContext *dev_ctx, enum AVPixelFormat p, |
| 617 | VkImageTiling tiling, |
| 618 | VkFormat fmts[AV_NUM_DATA_POINTERS8], /* Output format list */ |
| 619 | int *nb_images, /* Output number of images */ |
| 620 | VkImageAspectFlags *aspect, /* Output aspect */ |
| 621 | VkImageUsageFlags *supported_usage, /* Output supported usage */ |
| 622 | int disable_multiplane, int need_storage) |
| 623 | { |
| 624 | VulkanDevicePriv *priv = dev_ctx->hwctx; |
| 625 | AVVulkanDeviceContext *hwctx = &priv->p; |
| 626 | FFVulkanFunctions *vk = &priv->vkctx.vkfn; |
| 627 | |
| 628 | const VkFormatFeatureFlagBits2 basic_flags = VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_BIT | |
| 629 | VK_FORMAT_FEATURE_2_TRANSFER_SRC_BIT | |
| 630 | VK_FORMAT_FEATURE_2_TRANSFER_DST_BIT; |
| 631 | |
| 632 | for (int i = 0; i < nb_vk_formats_list; i++) { |
| 633 | if (vk_formats_list[i].pixfmt == p) { |
| 634 | VkFormatProperties3 fprops = { |
| 635 | .sType = VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_3, |
| 636 | }; |
| 637 | VkFormatProperties2 prop = { |
| 638 | .sType = VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2, |
| 639 | .pNext = &fprops, |
| 640 | }; |
| 641 | VkFormatFeatureFlagBits2 feats_primary, feats_secondary; |
| 642 | int basics_primary = 0, basics_secondary = 0; |
| 643 | int storage_primary = 0, storage_secondary = 0; |
| 644 | |
| 645 | vk->GetPhysicalDeviceFormatProperties2(hwctx->phys_dev, |
| 646 | vk_formats_list[i].vkf, |
| 647 | &prop); |
| 648 | |
| 649 | feats_primary = tiling == VK_IMAGE_TILING_LINEAR ? |
| 650 | fprops.linearTilingFeatures : fprops.optimalTilingFeatures; |
| 651 | basics_primary = (feats_primary & basic_flags) == basic_flags; |
| 652 | storage_primary = !!(feats_primary & VK_FORMAT_FEATURE_2_STORAGE_IMAGE_BIT); |
| 653 | |
| 654 | if (vk_formats_list[i].vkf != vk_formats_list[i].fallback[0]) { |
| 655 | vk->GetPhysicalDeviceFormatProperties2(hwctx->phys_dev, |
| 656 | vk_formats_list[i].fallback[0], |
| 657 | &prop); |
| 658 | feats_secondary = tiling == VK_IMAGE_TILING_LINEAR ? |
| 659 | fprops.linearTilingFeatures : fprops.optimalTilingFeatures; |
| 660 | basics_secondary = (feats_secondary & basic_flags) == basic_flags; |
| 661 | storage_secondary = !!(feats_secondary & VK_FORMAT_FEATURE_2_STORAGE_IMAGE_BIT); |
| 662 | } else { |
| 663 | basics_secondary = basics_primary; |
| 664 | storage_secondary = storage_primary; |
| 665 | } |
| 666 | |
| 667 | if (basics_primary && |
| 668 | !(disable_multiplane && vk_formats_list[i].vk_planes > 1) && |
| 669 | (!need_storage || (need_storage && (storage_primary | storage_secondary)))) { |
| 670 | if (fmts) { |
| 671 | if (vk_formats_list[i].nb_images > 1) { |
| 672 | for (int j = 0; j < vk_formats_list[i].nb_images_fallback; j++) |
| 673 | fmts[j] = vk_formats_list[i].fallback[j]; |
| 674 | } else { |
| 675 | fmts[0] = vk_formats_list[i].vkf; |
| 676 | } |
| 677 | } |
| 678 | if (nb_images) |
| 679 | *nb_images = 1; |
| 680 | if (aspect) |
| 681 | *aspect = vk_formats_list[i].aspect; |
| 682 | if (supported_usage) |
| 683 | *supported_usage = ff_vk_map_feats_to_usage(feats_primary) | |
| 684 | ((need_storage && (storage_primary | storage_secondary)) ? |
| 685 | VK_IMAGE_USAGE_STORAGE_BIT : 0); |
| 686 | return 0; |
| 687 | } else if (basics_secondary && |
| 688 | (!need_storage || (need_storage && storage_secondary))) { |
| 689 | if (fmts) { |
| 690 | for (int j = 0; j < vk_formats_list[i].nb_images_fallback; j++) |
| 691 | fmts[j] = vk_formats_list[i].fallback[j]; |
| 692 | } |
| 693 | if (nb_images) |
| 694 | *nb_images = vk_formats_list[i].nb_images_fallback; |
| 695 | if (aspect) |
| 696 | *aspect = vk_formats_list[i].aspect; |
| 697 | if (supported_usage) |
| 698 | *supported_usage = ff_vk_map_feats_to_usage(feats_secondary); |
| 699 | return 0; |
| 700 | } else { |
| 701 | return AVERROR(ENOTSUP)(-(95)); |
| 702 | } |
| 703 | } |
| 704 | } |
| 705 | |
| 706 | return AVERROR(EINVAL)(-(22)); |
| 707 | } |
| 708 | |
| 709 | #if CONFIG_VULKAN_STATIC0 |
| 710 | VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vkGetInstanceProcAddr(VkInstance instance, |
| 711 | const char *pName); |
| 712 | #endif |
| 713 | |
| 714 | static int load_libvulkan(AVHWDeviceContext *ctx) |
| 715 | { |
| 716 | VulkanDevicePriv *p = ctx->hwctx; |
| 717 | AVVulkanDeviceContext *hwctx = &p->p; |
| 718 | |
| 719 | #if CONFIG_VULKAN_STATIC0 |
| 720 | hwctx->get_proc_addr = vkGetInstanceProcAddr; |
| 721 | #else |
| 722 | static const char *lib_names[] = { |
| 723 | #if defined(_WIN32) |
| 724 | "vulkan-1.dll", |
| 725 | #elif defined(__APPLE__) |
| 726 | "libvulkan.dylib", |
| 727 | "libvulkan.1.dylib", |
| 728 | "libMoltenVK.dylib", |
| 729 | #else |
| 730 | "libvulkan.so.1", |
| 731 | "libvulkan.so", |
| 732 | #endif |
| 733 | }; |
| 734 | |
| 735 | for (int i = 0; i < FF_ARRAY_ELEMS(lib_names)(sizeof(lib_names) / sizeof((lib_names)[0])); i++) { |
| 736 | p->libvulkan = dlopen(lib_names[i], RTLD_NOW0x00002 | RTLD_LOCAL0); |
| 737 | if (p->libvulkan) |
| 738 | break; |
| 739 | } |
| 740 | |
| 741 | if (!p->libvulkan) { |
| 742 | av_log(ctx, AV_LOG_ERROR16, "Unable to open the libvulkan library!\n"); |
| 743 | return AVERROR_UNKNOWN(-(int)(('U') | (('N') << 8) | (('K') << 16) | (( unsigned)('N') << 24))); |
| 744 | } |
| 745 | |
| 746 | hwctx->get_proc_addr = (PFN_vkGetInstanceProcAddr)dlsym(p->libvulkan, "vkGetInstanceProcAddr"); |
| 747 | #endif /* CONFIG_VULKAN_STATIC */ |
| 748 | |
| 749 | return 0; |
| 750 | } |
| 751 | |
| 752 | typedef struct VulkanOptExtension { |
| 753 | const char *name; |
| 754 | FFVulkanExtensions flag; |
| 755 | } VulkanOptExtension; |
| 756 | |
| 757 | static const VulkanOptExtension optional_instance_exts[] = { |
| 758 | #ifdef __APPLE__ |
| 759 | { VK_KHR_PORTABILITY_ENUMERATION_EXTENSION_NAME"VK_KHR_portability_enumeration", FF_VK_EXT_NO_FLAG(1ULL << 63) }, |
| 760 | #endif |
| 761 | { 0 }, |
| 762 | }; |
| 763 | |
| 764 | static const VulkanOptExtension optional_device_exts[] = { |
| 765 | /* Misc or required by other extensions */ |
| 766 | { VK_KHR_PORTABILITY_SUBSET_EXTENSION_NAME"VK_KHR_portability_subset", FF_VK_EXT_PORTABILITY_SUBSET(1ULL << 62) }, |
| 767 | { VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME"VK_KHR_push_descriptor", FF_VK_EXT_PUSH_DESCRIPTOR(1ULL << 14) }, |
| 768 | { VK_EXT_PHYSICAL_DEVICE_DRM_EXTENSION_NAME"VK_EXT_physical_device_drm", FF_VK_EXT_DEVICE_DRM(1ULL << 9) }, |
| 769 | { VK_EXT_SHADER_ATOMIC_FLOAT_EXTENSION_NAME"VK_EXT_shader_atomic_float", FF_VK_EXT_ATOMIC_FLOAT(1ULL << 10) }, |
| 770 | { VK_KHR_COOPERATIVE_MATRIX_EXTENSION_NAME"VK_KHR_cooperative_matrix", FF_VK_EXT_COOP_MATRIX(1ULL << 11) }, |
| 771 | { VK_EXT_SHADER_OBJECT_EXTENSION_NAME"VK_EXT_shader_object", FF_VK_EXT_SHADER_OBJECT(1ULL << 13) }, |
| 772 | { VK_KHR_SHADER_SUBGROUP_ROTATE_EXTENSION_NAME"VK_KHR_shader_subgroup_rotate", FF_VK_EXT_SUBGROUP_ROTATE(1ULL << 17) }, |
| 773 | { VK_EXT_HOST_IMAGE_COPY_EXTENSION_NAME"VK_EXT_host_image_copy", FF_VK_EXT_HOST_IMAGE_COPY(1ULL << 18) }, |
| 774 | { VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME"VK_KHR_workgroup_memory_explicit_layout", FF_VK_EXT_EXPLICIT_MEM_LAYOUT(1ULL << 20) }, |
| 775 | #ifdef VK_KHR_shader_relaxed_extended_instruction1 |
| 776 | { VK_KHR_SHADER_RELAXED_EXTENDED_INSTRUCTION_EXTENSION_NAME"VK_KHR_shader_relaxed_extended_instruction", FF_VK_EXT_RELAXED_EXTENDED_INSTR(1ULL << 15) }, |
| 777 | #endif |
| 778 | #ifdef VK_EXT_shader_long_vector1 |
| 779 | { VK_EXT_SHADER_LONG_VECTOR_EXTENSION_NAME"VK_EXT_shader_long_vector", FF_VK_EXT_LONG_VECTOR(1ULL << 22) }, |
| 780 | #endif |
| 781 | #ifdef VK_EXT_shader_replicated_composites1 |
| 782 | { VK_EXT_SHADER_REPLICATED_COMPOSITES_EXTENSION_NAME"VK_EXT_shader_replicated_composites", FF_VK_EXT_REPLICATED_COMPOSITES(1ULL << 21) }, |
| 783 | #endif |
| 784 | #ifdef VK_EXT_zero_initialize_device_memory1 |
| 785 | { VK_EXT_ZERO_INITIALIZE_DEVICE_MEMORY_EXTENSION_NAME"VK_EXT_zero_initialize_device_memory", FF_VK_EXT_ZERO_INITIALIZE(1ULL << 19) }, |
| 786 | #endif |
| 787 | #ifdef VK_KHR_shader_expect_assume1 |
| 788 | { VK_KHR_SHADER_EXPECT_ASSUME_EXTENSION_NAME"VK_KHR_shader_expect_assume", FF_VK_EXT_EXPECT_ASSUME(1ULL << 16) }, |
| 789 | #endif |
| 790 | #ifdef VK_KHR_shader_maximal_reconvergence1 |
| 791 | { VK_KHR_SHADER_MAXIMAL_RECONVERGENCE_EXTENSION_NAME"VK_KHR_shader_maximal_reconvergence", FF_VK_EXT_MAXIMAL_RECONVERGENCE(1ULL << 24) }, |
| 792 | #endif |
| 793 | #ifdef VK_KHR_maintenance91 |
| 794 | { VK_KHR_MAINTENANCE_9_EXTENSION_NAME"VK_KHR_maintenance9", FF_VK_EXT_MAINTENANCE_9(1ULL << 25) }, |
| 795 | #endif |
| 796 | #ifdef VK_KHR_unified_image_layouts1 |
| 797 | { VK_KHR_UNIFIED_IMAGE_LAYOUTS_EXTENSION_NAME"VK_KHR_unified_image_layouts", FF_VK_EXT_UNIFIED_IMAGE_LAYOUTS(1ULL << 26) }, |
| 798 | #endif |
| 799 | { VK_KHR_VIDEO_MAINTENANCE_1_EXTENSION_NAME"VK_KHR_video_maintenance1", FF_VK_EXT_VIDEO_MAINTENANCE_1(1ULL << 37) }, |
| 800 | #ifdef VK_KHR_video_maintenance21 |
| 801 | { VK_KHR_VIDEO_MAINTENANCE_2_EXTENSION_NAME"VK_KHR_video_maintenance2", FF_VK_EXT_VIDEO_MAINTENANCE_2(1ULL << 38) }, |
| 802 | #endif |
| 803 | #ifdef VK_KHR_internally_synchronized_queues1 |
| 804 | { VK_KHR_INTERNALLY_SYNCHRONIZED_QUEUES_EXTENSION_NAME"VK_KHR_internally_synchronized_queues", FF_VK_EXT_INTERNAL_QUEUE_SYNC(1ULL << 23) }, |
| 805 | #endif |
| 806 | |
| 807 | /* Imports/exports */ |
| 808 | { VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME"VK_KHR_external_memory_fd", FF_VK_EXT_EXTERNAL_FD_MEMORY(1ULL << 2) }, |
| 809 | { VK_EXT_EXTERNAL_MEMORY_DMA_BUF_EXTENSION_NAME"VK_EXT_external_memory_dma_buf", FF_VK_EXT_EXTERNAL_DMABUF_MEMORY(1ULL << 0) }, |
| 810 | { VK_EXT_IMAGE_DRM_FORMAT_MODIFIER_EXTENSION_NAME"VK_EXT_image_drm_format_modifier", FF_VK_EXT_DRM_MODIFIER_FLAGS(1ULL << 1) }, |
| 811 | { VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME"VK_KHR_external_semaphore_fd", FF_VK_EXT_EXTERNAL_FD_SEM(1ULL << 3) }, |
| 812 | { VK_EXT_EXTERNAL_MEMORY_HOST_EXTENSION_NAME"VK_EXT_external_memory_host", FF_VK_EXT_EXTERNAL_HOST_MEMORY(1ULL << 4) }, |
| 813 | #ifdef _WIN32 |
| 814 | { VK_KHR_EXTERNAL_MEMORY_WIN32_EXTENSION_NAME, FF_VK_EXT_EXTERNAL_WIN32_MEMORY(1ULL << 6) }, |
| 815 | { VK_KHR_EXTERNAL_SEMAPHORE_WIN32_EXTENSION_NAME, FF_VK_EXT_EXTERNAL_WIN32_SEM(1ULL << 7) }, |
| 816 | #endif |
| 817 | |
| 818 | /* Video encoding/decoding */ |
| 819 | { VK_KHR_VIDEO_QUEUE_EXTENSION_NAME"VK_KHR_video_queue", FF_VK_EXT_VIDEO_QUEUE(1ULL << 36) }, |
| 820 | { VK_KHR_VIDEO_ENCODE_QUEUE_EXTENSION_NAME"VK_KHR_video_encode_queue", FF_VK_EXT_VIDEO_ENCODE_QUEUE(1ULL << 50) }, |
| 821 | { VK_KHR_VIDEO_DECODE_QUEUE_EXTENSION_NAME"VK_KHR_video_decode_queue", FF_VK_EXT_VIDEO_DECODE_QUEUE(1ULL << 40) }, |
| 822 | { VK_KHR_VIDEO_ENCODE_H264_EXTENSION_NAME"VK_KHR_video_encode_h264", FF_VK_EXT_VIDEO_ENCODE_H264(1ULL << 51) }, |
| 823 | { VK_KHR_VIDEO_DECODE_H264_EXTENSION_NAME"VK_KHR_video_decode_h264", FF_VK_EXT_VIDEO_DECODE_H264(1ULL << 41) }, |
| 824 | { VK_KHR_VIDEO_ENCODE_H265_EXTENSION_NAME"VK_KHR_video_encode_h265", FF_VK_EXT_VIDEO_ENCODE_H265(1ULL << 52) }, |
| 825 | { VK_KHR_VIDEO_DECODE_H265_EXTENSION_NAME"VK_KHR_video_decode_h265", FF_VK_EXT_VIDEO_DECODE_H265(1ULL << 42) }, |
| 826 | #ifdef VK_KHR_video_decode_vp91 |
| 827 | { VK_KHR_VIDEO_DECODE_VP9_EXTENSION_NAME"VK_KHR_video_decode_vp9", FF_VK_EXT_VIDEO_DECODE_VP9(1ULL << 43) }, |
| 828 | #endif |
| 829 | #ifdef VK_KHR_video_encode_av11 |
| 830 | { VK_KHR_VIDEO_ENCODE_AV1_EXTENSION_NAME"VK_KHR_video_encode_av1", FF_VK_EXT_VIDEO_ENCODE_AV1(1ULL << 53) }, |
| 831 | #endif |
| 832 | { VK_KHR_VIDEO_DECODE_AV1_EXTENSION_NAME"VK_KHR_video_decode_av1", FF_VK_EXT_VIDEO_DECODE_AV1(1ULL << 44) }, |
| 833 | }; |
| 834 | |
| 835 | const char **av_vk_get_optional_instance_extensions(int *count) |
| 836 | { |
| 837 | const char **exts = av_malloc_array(sizeof(*exts), |
| 838 | FF_ARRAY_ELEMS(optional_instance_exts)(sizeof(optional_instance_exts) / sizeof((optional_instance_exts )[0])) - 1); |
| 839 | if (!exts) |
| 840 | return NULL((void*)0); |
| 841 | |
| 842 | for (int i = 0; i < (int)FF_ARRAY_ELEMS(optional_instance_exts)(sizeof(optional_instance_exts) / sizeof((optional_instance_exts )[0])) - 1; i++) |
| 843 | exts[i] = optional_instance_exts[i].name; |
| 844 | |
| 845 | *count = FF_ARRAY_ELEMS(optional_instance_exts)(sizeof(optional_instance_exts) / sizeof((optional_instance_exts )[0])) - 1; |
| 846 | return exts; |
| 847 | } |
| 848 | |
| 849 | const char **av_vk_get_optional_device_extensions(int *count) |
| 850 | { |
| 851 | const char **exts = av_malloc_array(sizeof(*exts), |
| 852 | FF_ARRAY_ELEMS(optional_device_exts)(sizeof(optional_device_exts) / sizeof((optional_device_exts) [0]))); |
| 853 | if (!exts) |
| 854 | return NULL((void*)0); |
| 855 | |
| 856 | for (int i = 0; i < FF_ARRAY_ELEMS(optional_device_exts)(sizeof(optional_device_exts) / sizeof((optional_device_exts) [0])); i++) |
| 857 | exts[i] = optional_device_exts[i].name; |
| 858 | |
| 859 | *count = FF_ARRAY_ELEMS(optional_device_exts)(sizeof(optional_device_exts) / sizeof((optional_device_exts) [0])); |
| 860 | return exts; |
| 861 | } |
| 862 | |
| 863 | static VKAPI_ATTR |
| 864 | VkBool32 VKAPI_CALL vk_dbg_callback(VkDebugUtilsMessageSeverityFlagBitsEXT severity, |
| 865 | VkDebugUtilsMessageTypeFlagsEXT messageType, |
| 866 | const VkDebugUtilsMessengerCallbackDataEXT *data, |
| 867 | void *priv) |
| 868 | { |
| 869 | int l; |
| 870 | AVHWDeviceContext *ctx = priv; |
| 871 | |
| 872 | /* Ignore false positives */ |
| 873 | switch (data->messageIdNumber) { |
| 874 | case 0x086974c1: /* BestPractices-vkCreateCommandPool-command-buffer-reset */ |
| 875 | case 0xfd92477a: /* BestPractices-vkAllocateMemory-small-allocation */ |
| 876 | return VK_FALSE0U; |
| 877 | default: |
| 878 | break; |
| 879 | } |
| 880 | |
| 881 | switch (severity) { |
| 882 | case VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT: l = AV_LOG_VERBOSE40; break; |
| 883 | case VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT: l = AV_LOG_INFO32; break; |
| 884 | case VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT: l = AV_LOG_WARNING24; break; |
| 885 | case VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT: l = AV_LOG_ERROR16; break; |
| 886 | default: l = AV_LOG_DEBUG48; break; |
| 887 | } |
| 888 | |
| 889 | av_log(ctx, l, "%s\n", data->pMessage); |
| 890 | for (int i = 0; i < data->cmdBufLabelCount; i++) |
| 891 | av_log(ctx, l, "\t%i: %s\n", i, data->pCmdBufLabels[i].pLabelName); |
| 892 | |
| 893 | return VK_FALSE0U; |
| 894 | } |
| 895 | |
| 896 | #define ADD_VAL_TO_LIST(list, count, val)do { list = av_realloc_array(list, ++count, sizeof(*list)); if (!list) { err = (-(12)); goto fail; } list[count - 1] = av_strdup (val); if (!list[count - 1]) { err = (-(12)); goto fail; } } while (0) \ |
| 897 | do { \ |
| 898 | list = av_realloc_array(list, ++count, sizeof(*list)); \ |
| 899 | if (!list) { \ |
| 900 | err = AVERROR(ENOMEM)(-(12)); \ |
| 901 | goto fail; \ |
| 902 | } \ |
| 903 | list[count - 1] = av_strdup(val); \ |
| 904 | if (!list[count - 1]) { \ |
| 905 | err = AVERROR(ENOMEM)(-(12)); \ |
| 906 | goto fail; \ |
| 907 | } \ |
| 908 | } while(0) |
| 909 | |
| 910 | #define RELEASE_PROPS(props, count)if (props) { for (int i = 0; i < count; i++) av_free((void *)((props)[i])); av_free((void *)props); } \ |
| 911 | if (props) { \ |
| 912 | for (int i = 0; i < count; i++) \ |
| 913 | av_free((void *)((props)[i])); \ |
| 914 | av_free((void *)props); \ |
| 915 | } |
| 916 | |
| 917 | static int vulkan_device_has_rebar(AVHWDeviceContext *ctx) |
| 918 | { |
| 919 | VulkanDevicePriv *p = ctx->hwctx; |
| 920 | VkDeviceSize max_vram = 0, max_visible_vram = 0; |
| 921 | |
| 922 | /* Get device memory properties */ |
| 923 | av_assert0(p->mprops.memoryTypeCount)do { if (!(p->mprops.memoryTypeCount)) { av_log(((void*)0) , 0, "Assertion %s failed at %s:%d\n", "p->mprops.memoryTypeCount" , "/root/firefox-clang/media/ffvpx/libavutil/hwcontext_vulkan.c" , 923); abort(); } } while (0); |
| 924 | for (int i = 0; i < p->mprops.memoryTypeCount; i++) { |
| 925 | const VkMemoryType type = p->mprops.memoryTypes[i]; |
| 926 | const VkMemoryHeap heap = p->mprops.memoryHeaps[type.heapIndex]; |
| 927 | if (!(type.propertyFlags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT)) |
| 928 | continue; |
| 929 | max_vram = FFMAX(max_vram, heap.size)((max_vram) > (heap.size) ? (max_vram) : (heap.size)); |
| 930 | if (type.propertyFlags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) |
| 931 | max_visible_vram = FFMAX(max_visible_vram, heap.size)((max_visible_vram) > (heap.size) ? (max_visible_vram) : ( heap.size)); |
| 932 | } |
| 933 | |
| 934 | return max_vram - max_visible_vram < 1024; /* 1 kB tolerance */ |
| 935 | } |
| 936 | |
| 937 | enum FFVulkanDebugMode { |
| 938 | FF_VULKAN_DEBUG_NONE = 0, |
| 939 | /* Standard GPU-assisted validation */ |
| 940 | FF_VULKAN_DEBUG_VALIDATE = 1, |
| 941 | /* Passes printfs in shaders to the debug callback */ |
| 942 | FF_VULKAN_DEBUG_PRINTF = 2, |
| 943 | /* Enables extra printouts */ |
| 944 | FF_VULKAN_DEBUG_PRACTICES = 3, |
| 945 | |
| 946 | FF_VULKAN_DEBUG_NB, |
| 947 | }; |
| 948 | |
| 949 | static int check_extensions(AVHWDeviceContext *ctx, int dev, AVDictionary *opts, |
| 950 | const char * const **dst, uint32_t *num, |
| 951 | enum FFVulkanDebugMode debug_mode) |
| 952 | { |
| 953 | const char *tstr; |
| 954 | const char **extension_names = NULL((void*)0); |
| 955 | VulkanDevicePriv *p = ctx->hwctx; |
| 956 | AVVulkanDeviceContext *hwctx = &p->p; |
| 957 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 958 | int err = 0, found, extensions_found = 0; |
| 959 | |
| 960 | const char *mod; |
| 961 | int optional_exts_num; |
| 962 | uint32_t sup_ext_count; |
| 963 | char *user_exts_str = NULL((void*)0); |
| 964 | AVDictionaryEntry *user_exts; |
| 965 | VkExtensionProperties *sup_ext; |
| 966 | const VulkanOptExtension *optional_exts; |
| 967 | |
| 968 | if (!dev) { |
| 969 | mod = "instance"; |
| 970 | optional_exts = optional_instance_exts; |
| 971 | optional_exts_num = FF_ARRAY_ELEMS(optional_instance_exts)(sizeof(optional_instance_exts) / sizeof((optional_instance_exts )[0])) - 1; |
| 972 | user_exts = av_dict_get(opts, "instance_extensions", NULL((void*)0), 0); |
| 973 | if (user_exts) { |
| 974 | user_exts_str = av_strdup(user_exts->value); |
| 975 | if (!user_exts_str) { |
| 976 | err = AVERROR(ENOMEM)(-(12)); |
| 977 | goto fail; |
| 978 | } |
| 979 | } |
| 980 | vk->EnumerateInstanceExtensionProperties(NULL((void*)0), &sup_ext_count, NULL((void*)0)); |
| 981 | sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties)); |
| 982 | if (!sup_ext) |
| 983 | return AVERROR(ENOMEM)(-(12)); |
| 984 | vk->EnumerateInstanceExtensionProperties(NULL((void*)0), &sup_ext_count, sup_ext); |
| 985 | } else { |
| 986 | mod = "device"; |
| 987 | optional_exts = optional_device_exts; |
| 988 | optional_exts_num = FF_ARRAY_ELEMS(optional_device_exts)(sizeof(optional_device_exts) / sizeof((optional_device_exts) [0])); |
| 989 | user_exts = av_dict_get(opts, "device_extensions", NULL((void*)0), 0); |
| 990 | if (user_exts) { |
| 991 | user_exts_str = av_strdup(user_exts->value); |
| 992 | if (!user_exts_str) { |
| 993 | err = AVERROR(ENOMEM)(-(12)); |
| 994 | goto fail; |
| 995 | } |
| 996 | } |
| 997 | vk->EnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL((void*)0), |
| 998 | &sup_ext_count, NULL((void*)0)); |
| 999 | sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties)); |
| 1000 | if (!sup_ext) |
| 1001 | return AVERROR(ENOMEM)(-(12)); |
| 1002 | vk->EnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL((void*)0), |
| 1003 | &sup_ext_count, sup_ext); |
| 1004 | } |
| 1005 | |
| 1006 | for (int i = 0; i < optional_exts_num; i++) { |
| 1007 | tstr = optional_exts[i].name; |
| 1008 | found = 0; |
| 1009 | |
| 1010 | /* Check if the device has ReBAR for host image copies */ |
| 1011 | if (!strcmp(tstr, VK_EXT_HOST_IMAGE_COPY_EXTENSION_NAME"VK_EXT_host_image_copy") && |
| 1012 | !vulkan_device_has_rebar(ctx)) |
| 1013 | continue; |
| 1014 | |
| 1015 | if (dev && |
| 1016 | ((debug_mode == FF_VULKAN_DEBUG_VALIDATE) || |
| 1017 | (debug_mode == FF_VULKAN_DEBUG_PRINTF) || |
| 1018 | (debug_mode == FF_VULKAN_DEBUG_PRACTICES)) && |
| 1019 | (!strcmp(tstr, VK_EXT_SHADER_OBJECT_EXTENSION_NAME"VK_EXT_shader_object"))) { |
| 1020 | continue; |
| 1021 | } |
| 1022 | |
| 1023 | for (int j = 0; j < sup_ext_count; j++) { |
| 1024 | if (!strcmp(tstr, sup_ext[j].extensionName)) { |
| 1025 | found = 1; |
| 1026 | break; |
| 1027 | } |
| 1028 | } |
| 1029 | if (!found) |
| 1030 | continue; |
| 1031 | |
| 1032 | av_log(ctx, AV_LOG_VERBOSE40, "Using %s extension %s\n", mod, tstr); |
| 1033 | p->vkctx.extensions |= optional_exts[i].flag; |
| 1034 | ADD_VAL_TO_LIST(extension_names, extensions_found, tstr)do { extension_names = av_realloc_array(extension_names, ++extensions_found , sizeof(*extension_names)); if (!extension_names) { err = (- (12)); goto fail; } extension_names[extensions_found - 1] = av_strdup (tstr); if (!extension_names[extensions_found - 1]) { err = ( -(12)); goto fail; } } while(0); |
| 1035 | } |
| 1036 | |
| 1037 | if (!dev && |
| 1038 | ((debug_mode == FF_VULKAN_DEBUG_VALIDATE) || |
| 1039 | (debug_mode == FF_VULKAN_DEBUG_PRINTF) || |
| 1040 | (debug_mode == FF_VULKAN_DEBUG_PRACTICES))) { |
| 1041 | tstr = VK_EXT_DEBUG_UTILS_EXTENSION_NAME"VK_EXT_debug_utils"; |
| 1042 | found = 0; |
| 1043 | for (int j = 0; j < sup_ext_count; j++) { |
| 1044 | if (!strcmp(tstr, sup_ext[j].extensionName)) { |
| 1045 | found = 1; |
| 1046 | break; |
| 1047 | } |
| 1048 | } |
| 1049 | if (found) { |
| 1050 | av_log(ctx, AV_LOG_VERBOSE40, "Using %s extension %s\n", mod, tstr); |
| 1051 | ADD_VAL_TO_LIST(extension_names, extensions_found, tstr)do { extension_names = av_realloc_array(extension_names, ++extensions_found , sizeof(*extension_names)); if (!extension_names) { err = (- (12)); goto fail; } extension_names[extensions_found - 1] = av_strdup (tstr); if (!extension_names[extensions_found - 1]) { err = ( -(12)); goto fail; } } while(0); |
| 1052 | } else { |
| 1053 | av_log(ctx, AV_LOG_ERROR16, "Debug extension \"%s\" not found!\n", |
| 1054 | tstr); |
| 1055 | err = AVERROR(EINVAL)(-(22)); |
| 1056 | goto fail; |
| 1057 | } |
| 1058 | } |
| 1059 | |
| 1060 | #ifdef VK_KHR_shader_relaxed_extended_instruction1 |
| 1061 | if ((debug_mode == FF_VULKAN_DEBUG_PRINTF) && dev) { |
| 1062 | tstr = VK_KHR_SHADER_RELAXED_EXTENDED_INSTRUCTION_EXTENSION_NAME"VK_KHR_shader_relaxed_extended_instruction"; |
| 1063 | found = 0; |
| 1064 | for (int j = 0; j < sup_ext_count; j++) { |
| 1065 | if (!strcmp(tstr, sup_ext[j].extensionName)) { |
| 1066 | found = 1; |
| 1067 | break; |
| 1068 | } |
| 1069 | } |
| 1070 | if (!found) { |
| 1071 | av_log(ctx, AV_LOG_ERROR16, "Debug_printf/profile enabled, but extension \"%s\" not found!\n", |
| 1072 | tstr); |
| 1073 | err = AVERROR(EINVAL)(-(22)); |
| 1074 | goto fail; |
| 1075 | } |
| 1076 | } |
| 1077 | #endif |
| 1078 | |
| 1079 | if (user_exts_str) { |
| 1080 | char *save, *token = av_strtok(user_exts_str, "+", &save); |
| 1081 | while (token) { |
| 1082 | found = 0; |
| 1083 | for (int j = 0; j < sup_ext_count; j++) { |
| 1084 | if (!strcmp(token, sup_ext[j].extensionName)) { |
| 1085 | found = 1; |
| 1086 | break; |
| 1087 | } |
| 1088 | } |
| 1089 | if (found) { |
| 1090 | av_log(ctx, AV_LOG_VERBOSE40, "Using %s extension \"%s\"\n", mod, token); |
| 1091 | ADD_VAL_TO_LIST(extension_names, extensions_found, token)do { extension_names = av_realloc_array(extension_names, ++extensions_found , sizeof(*extension_names)); if (!extension_names) { err = (- (12)); goto fail; } extension_names[extensions_found - 1] = av_strdup (token); if (!extension_names[extensions_found - 1]) { err = ( -(12)); goto fail; } } while(0); |
| 1092 | } else { |
| 1093 | av_log(ctx, AV_LOG_WARNING24, "%s extension \"%s\" not found, excluding.\n", |
| 1094 | mod, token); |
| 1095 | } |
| 1096 | token = av_strtok(NULL((void*)0), "+", &save); |
| 1097 | } |
| 1098 | } |
| 1099 | |
| 1100 | *dst = extension_names; |
| 1101 | *num = extensions_found; |
| 1102 | |
| 1103 | av_free(user_exts_str); |
| 1104 | av_free(sup_ext); |
| 1105 | return 0; |
| 1106 | |
| 1107 | fail: |
| 1108 | RELEASE_PROPS(extension_names, extensions_found)if (extension_names) { for (int i = 0; i < extensions_found ; i++) av_free((void *)((extension_names)[i])); av_free((void *)extension_names); }; |
| 1109 | av_free(user_exts_str); |
| 1110 | av_free(sup_ext); |
| 1111 | return err; |
| 1112 | } |
| 1113 | |
| 1114 | static int check_layers(AVHWDeviceContext *ctx, AVDictionary *opts, |
| 1115 | const char * const **dst, uint32_t *num, |
| 1116 | enum FFVulkanDebugMode *debug_mode) |
| 1117 | { |
| 1118 | int err = 0; |
| 1119 | VulkanDevicePriv *priv = ctx->hwctx; |
| 1120 | FFVulkanFunctions *vk = &priv->vkctx.vkfn; |
| 1121 | |
| 1122 | static const char layer_standard_validation[] = { "VK_LAYER_KHRONOS_validation" }; |
| 1123 | int layer_standard_validation_found = 0; |
| 1124 | |
| 1125 | uint32_t sup_layer_count; |
| 1126 | VkLayerProperties *sup_layers; |
| 1127 | |
| 1128 | AVDictionaryEntry *user_layers = av_dict_get(opts, "layers", NULL((void*)0), 0); |
| 1129 | char *user_layers_str = NULL((void*)0); |
| 1130 | char *save, *token; |
| 1131 | |
| 1132 | const char **enabled_layers = NULL((void*)0); |
| 1133 | uint32_t enabled_layers_count = 0; |
| 1134 | |
| 1135 | AVDictionaryEntry *debug_opt = av_dict_get(opts, "debug", NULL((void*)0), 0); |
| 1136 | enum FFVulkanDebugMode mode; |
| 1137 | |
| 1138 | *debug_mode = mode = FF_VULKAN_DEBUG_NONE; |
| 1139 | |
| 1140 | /* Get a list of all layers */ |
| 1141 | vk->EnumerateInstanceLayerProperties(&sup_layer_count, NULL((void*)0)); |
| 1142 | sup_layers = av_malloc_array(sup_layer_count, sizeof(VkLayerProperties)); |
| 1143 | if (!sup_layers) |
| 1144 | return AVERROR(ENOMEM)(-(12)); |
| 1145 | vk->EnumerateInstanceLayerProperties(&sup_layer_count, sup_layers); |
| 1146 | |
| 1147 | av_log(ctx, AV_LOG_VERBOSE40, "Supported layers:\n"); |
| 1148 | for (int i = 0; i < sup_layer_count; i++) |
| 1149 | av_log(ctx, AV_LOG_VERBOSE40, "\t%s\n", sup_layers[i].layerName); |
| 1150 | |
| 1151 | /* If no user layers or debug layers are given, return */ |
| 1152 | if (!debug_opt && !user_layers) |
| 1153 | goto end; |
| 1154 | |
| 1155 | /* Check for any properly supported validation layer */ |
| 1156 | if (debug_opt) { |
| 1157 | if (!strcmp(debug_opt->value, "printf")) { |
| 1158 | mode = FF_VULKAN_DEBUG_PRINTF; |
| 1159 | } else if (!strcmp(debug_opt->value, "validate")) { |
| 1160 | mode = FF_VULKAN_DEBUG_VALIDATE; |
| 1161 | } else if (!strcmp(debug_opt->value, "practices")) { |
| 1162 | mode = FF_VULKAN_DEBUG_PRACTICES; |
| 1163 | } else { |
| 1164 | char *end_ptr = NULL((void*)0); |
| 1165 | int idx = strtol(debug_opt->value, &end_ptr, 10); |
| 1166 | if (end_ptr == debug_opt->value || end_ptr[0] != '\0' || |
| 1167 | idx < 0 || idx >= FF_VULKAN_DEBUG_NB) { |
| 1168 | av_log(ctx, AV_LOG_ERROR16, "Invalid debugging mode \"%s\"\n", |
| 1169 | debug_opt->value); |
| 1170 | err = AVERROR(EINVAL)(-(22)); |
| 1171 | goto end; |
| 1172 | } |
| 1173 | mode = idx; |
| 1174 | } |
| 1175 | } |
| 1176 | |
| 1177 | /* If mode is VALIDATE or PRINTF, try to find the standard validation layer extension */ |
| 1178 | if ((mode == FF_VULKAN_DEBUG_VALIDATE) || |
| 1179 | (mode == FF_VULKAN_DEBUG_PRINTF) || |
| 1180 | (mode == FF_VULKAN_DEBUG_PRACTICES)) { |
| 1181 | for (int i = 0; i < sup_layer_count; i++) { |
| 1182 | if (!strcmp(layer_standard_validation, sup_layers[i].layerName)) { |
| 1183 | av_log(ctx, AV_LOG_VERBOSE40, "Standard validation layer %s is enabled\n", |
| 1184 | layer_standard_validation); |
| 1185 | ADD_VAL_TO_LIST(enabled_layers, enabled_layers_count, layer_standard_validation)do { enabled_layers = av_realloc_array(enabled_layers, ++enabled_layers_count , sizeof(*enabled_layers)); if (!enabled_layers) { err = (-(12 )); goto fail; } enabled_layers[enabled_layers_count - 1] = av_strdup (layer_standard_validation); if (!enabled_layers[enabled_layers_count - 1]) { err = (-(12)); goto fail; } } while(0); |
| 1186 | *debug_mode = mode; |
| 1187 | layer_standard_validation_found = 1; |
| 1188 | break; |
| 1189 | } |
| 1190 | } |
| 1191 | if (!layer_standard_validation_found) { |
| 1192 | av_log(ctx, AV_LOG_ERROR16, |
| 1193 | "Validation Layer \"%s\" not supported\n", layer_standard_validation); |
| 1194 | err = AVERROR(ENOTSUP)(-(95)); |
| 1195 | goto end; |
| 1196 | } |
| 1197 | } |
| 1198 | |
| 1199 | /* Process any custom layers enabled */ |
| 1200 | if (user_layers) { |
| 1201 | int found; |
| 1202 | |
| 1203 | user_layers_str = av_strdup(user_layers->value); |
| 1204 | if (!user_layers_str) { |
| 1205 | err = AVERROR(ENOMEM)(-(12)); |
| 1206 | goto fail; |
| 1207 | } |
| 1208 | |
| 1209 | token = av_strtok(user_layers_str, "+", &save); |
| 1210 | while (token) { |
| 1211 | found = 0; |
| 1212 | |
| 1213 | /* If debug=1/2 was specified as an option, skip this layer */ |
| 1214 | if (!strcmp(layer_standard_validation, token) && layer_standard_validation_found) { |
| 1215 | token = av_strtok(NULL((void*)0), "+", &save); |
Value stored to 'token' is never read | |
| 1216 | break; |
| 1217 | } |
| 1218 | |
| 1219 | /* Try to find the layer in the list of supported layers */ |
| 1220 | for (int j = 0; j < sup_layer_count; j++) { |
| 1221 | if (!strcmp(token, sup_layers[j].layerName)) { |
| 1222 | found = 1; |
| 1223 | break; |
| 1224 | } |
| 1225 | } |
| 1226 | |
| 1227 | if (found) { |
| 1228 | av_log(ctx, AV_LOG_VERBOSE40, "Using layer: %s\n", token); |
| 1229 | ADD_VAL_TO_LIST(enabled_layers, enabled_layers_count, token)do { enabled_layers = av_realloc_array(enabled_layers, ++enabled_layers_count , sizeof(*enabled_layers)); if (!enabled_layers) { err = (-(12 )); goto fail; } enabled_layers[enabled_layers_count - 1] = av_strdup (token); if (!enabled_layers[enabled_layers_count - 1]) { err = (-(12)); goto fail; } } while(0); |
| 1230 | |
| 1231 | /* If debug was not set as an option, force it */ |
| 1232 | if (!strcmp(layer_standard_validation, token)) |
| 1233 | *debug_mode = FF_VULKAN_DEBUG_VALIDATE; |
| 1234 | } else { |
| 1235 | av_log(ctx, AV_LOG_ERROR16, |
| 1236 | "Layer \"%s\" not supported\n", token); |
| 1237 | err = AVERROR(EINVAL)(-(22)); |
| 1238 | goto end; |
| 1239 | } |
| 1240 | |
| 1241 | token = av_strtok(NULL((void*)0), "+", &save); |
| 1242 | } |
| 1243 | } |
| 1244 | |
| 1245 | fail: |
| 1246 | end: |
| 1247 | av_free(sup_layers); |
| 1248 | av_free(user_layers_str); |
| 1249 | |
| 1250 | if (err < 0) { |
| 1251 | RELEASE_PROPS(enabled_layers, enabled_layers_count)if (enabled_layers) { for (int i = 0; i < enabled_layers_count ; i++) av_free((void *)((enabled_layers)[i])); av_free((void * )enabled_layers); }; |
| 1252 | } else { |
| 1253 | *dst = enabled_layers; |
| 1254 | *num = enabled_layers_count; |
| 1255 | } |
| 1256 | |
| 1257 | return err; |
| 1258 | } |
| 1259 | |
| 1260 | /* Creates a VkInstance */ |
| 1261 | static int create_instance(AVHWDeviceContext *ctx, AVDictionary *opts, |
| 1262 | enum FFVulkanDebugMode *debug_mode) |
| 1263 | { |
| 1264 | int err = 0; |
| 1265 | VkResult ret; |
| 1266 | VulkanDevicePriv *p = ctx->hwctx; |
| 1267 | AVVulkanDeviceContext *hwctx = &p->p; |
| 1268 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 1269 | VkApplicationInfo application_info = { |
| 1270 | .sType = VK_STRUCTURE_TYPE_APPLICATION_INFO, |
| 1271 | .pApplicationName = "ffmpeg", |
| 1272 | .applicationVersion = VK_MAKE_VERSION(LIBAVUTIL_VERSION_MAJOR,((((uint32_t)(61)) << 22U) | (((uint32_t)(1)) << 12U ) | ((uint32_t)(101))) |
| 1273 | LIBAVUTIL_VERSION_MINOR,((((uint32_t)(61)) << 22U) | (((uint32_t)(1)) << 12U ) | ((uint32_t)(101))) |
| 1274 | LIBAVUTIL_VERSION_MICRO)((((uint32_t)(61)) << 22U) | (((uint32_t)(1)) << 12U ) | ((uint32_t)(101))), |
| 1275 | .pEngineName = "libavutil", |
| 1276 | .apiVersion = VK_API_VERSION_1_3((((uint32_t)(0)) << 29U) | (((uint32_t)(1)) << 22U ) | (((uint32_t)(3)) << 12U) | ((uint32_t)(0))), |
| 1277 | .engineVersion = VK_MAKE_VERSION(LIBAVUTIL_VERSION_MAJOR,((((uint32_t)(61)) << 22U) | (((uint32_t)(1)) << 12U ) | ((uint32_t)(101))) |
| 1278 | LIBAVUTIL_VERSION_MINOR,((((uint32_t)(61)) << 22U) | (((uint32_t)(1)) << 12U ) | ((uint32_t)(101))) |
| 1279 | LIBAVUTIL_VERSION_MICRO)((((uint32_t)(61)) << 22U) | (((uint32_t)(1)) << 12U ) | ((uint32_t)(101))), |
| 1280 | }; |
| 1281 | VkValidationFeaturesEXT validation_features = { |
| 1282 | .sType = VK_STRUCTURE_TYPE_VALIDATION_FEATURES_EXT, |
| 1283 | }; |
| 1284 | VkInstanceCreateInfo inst_props = { |
| 1285 | .sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO, |
| 1286 | .pApplicationInfo = &application_info, |
| 1287 | }; |
| 1288 | |
| 1289 | if (!hwctx->get_proc_addr) { |
| 1290 | err = load_libvulkan(ctx); |
| 1291 | if (err < 0) |
| 1292 | return err; |
| 1293 | } |
| 1294 | |
| 1295 | err = ff_vk_load_functions(ctx, vk, p->vkctx.extensions, 0, 0); |
| 1296 | if (err < 0) { |
| 1297 | av_log(ctx, AV_LOG_ERROR16, "Unable to load instance enumeration functions!\n"); |
| 1298 | return err; |
| 1299 | } |
| 1300 | |
| 1301 | err = check_layers(ctx, opts, &inst_props.ppEnabledLayerNames, |
| 1302 | &inst_props.enabledLayerCount, debug_mode); |
| 1303 | if (err) |
| 1304 | goto fail; |
| 1305 | |
| 1306 | /* Check for present/missing extensions */ |
| 1307 | err = check_extensions(ctx, 0, opts, &inst_props.ppEnabledExtensionNames, |
| 1308 | &inst_props.enabledExtensionCount, *debug_mode); |
| 1309 | hwctx->enabled_inst_extensions = inst_props.ppEnabledExtensionNames; |
| 1310 | hwctx->nb_enabled_inst_extensions = inst_props.enabledExtensionCount; |
| 1311 | if (err < 0) |
| 1312 | goto fail; |
| 1313 | |
| 1314 | /* Enable debug features if needed */ |
| 1315 | if (*debug_mode == FF_VULKAN_DEBUG_VALIDATE) { |
| 1316 | static const VkValidationFeatureEnableEXT feat_list_validate[] = { |
| 1317 | VK_VALIDATION_FEATURE_ENABLE_SYNCHRONIZATION_VALIDATION_EXT, |
| 1318 | VK_VALIDATION_FEATURE_ENABLE_GPU_ASSISTED_RESERVE_BINDING_SLOT_EXT, |
| 1319 | VK_VALIDATION_FEATURE_ENABLE_GPU_ASSISTED_EXT, |
| 1320 | }; |
| 1321 | validation_features.pEnabledValidationFeatures = feat_list_validate; |
| 1322 | validation_features.enabledValidationFeatureCount = FF_ARRAY_ELEMS(feat_list_validate)(sizeof(feat_list_validate) / sizeof((feat_list_validate)[0]) ); |
| 1323 | inst_props.pNext = &validation_features; |
| 1324 | } else if (*debug_mode == FF_VULKAN_DEBUG_PRINTF) { |
| 1325 | static const VkValidationFeatureEnableEXT feat_list_debug[] = { |
| 1326 | VK_VALIDATION_FEATURE_ENABLE_SYNCHRONIZATION_VALIDATION_EXT, |
| 1327 | VK_VALIDATION_FEATURE_ENABLE_GPU_ASSISTED_RESERVE_BINDING_SLOT_EXT, |
| 1328 | VK_VALIDATION_FEATURE_ENABLE_DEBUG_PRINTF_EXT, |
| 1329 | }; |
| 1330 | validation_features.pEnabledValidationFeatures = feat_list_debug; |
| 1331 | validation_features.enabledValidationFeatureCount = FF_ARRAY_ELEMS(feat_list_debug)(sizeof(feat_list_debug) / sizeof((feat_list_debug)[0])); |
| 1332 | inst_props.pNext = &validation_features; |
| 1333 | } else if (*debug_mode == FF_VULKAN_DEBUG_PRACTICES) { |
| 1334 | static const VkValidationFeatureEnableEXT feat_list_practices[] = { |
| 1335 | VK_VALIDATION_FEATURE_ENABLE_SYNCHRONIZATION_VALIDATION_EXT, |
| 1336 | VK_VALIDATION_FEATURE_ENABLE_BEST_PRACTICES_EXT, |
| 1337 | }; |
| 1338 | validation_features.pEnabledValidationFeatures = feat_list_practices; |
| 1339 | validation_features.enabledValidationFeatureCount = FF_ARRAY_ELEMS(feat_list_practices)(sizeof(feat_list_practices) / sizeof((feat_list_practices)[0 ])); |
| 1340 | inst_props.pNext = &validation_features; |
| 1341 | } |
| 1342 | |
| 1343 | #ifdef __APPLE__ |
| 1344 | for (int i = 0; i < inst_props.enabledExtensionCount; i++) { |
| 1345 | if (!strcmp(VK_KHR_PORTABILITY_ENUMERATION_EXTENSION_NAME"VK_KHR_portability_enumeration", |
| 1346 | inst_props.ppEnabledExtensionNames[i])) { |
| 1347 | inst_props.flags |= VK_INSTANCE_CREATE_ENUMERATE_PORTABILITY_BIT_KHR; |
| 1348 | break; |
| 1349 | } |
| 1350 | } |
| 1351 | #endif |
| 1352 | |
| 1353 | /* Try to create the instance */ |
| 1354 | ret = vk->CreateInstance(&inst_props, hwctx->alloc, &hwctx->inst); |
| 1355 | |
| 1356 | /* Check for errors */ |
| 1357 | if (ret != VK_SUCCESS) { |
| 1358 | av_log(ctx, AV_LOG_ERROR16, "Instance creation failure: %s\n", |
| 1359 | ff_vk_ret2str(ret)); |
| 1360 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 1361 | goto fail; |
| 1362 | } |
| 1363 | |
| 1364 | err = ff_vk_load_functions(ctx, vk, p->vkctx.extensions, 1, 0); |
| 1365 | if (err < 0) { |
| 1366 | av_log(ctx, AV_LOG_ERROR16, "Unable to load instance functions!\n"); |
| 1367 | goto fail; |
| 1368 | } |
| 1369 | |
| 1370 | /* Setup debugging callback if needed */ |
| 1371 | if ((*debug_mode == FF_VULKAN_DEBUG_VALIDATE) || |
| 1372 | (*debug_mode == FF_VULKAN_DEBUG_PRINTF) || |
| 1373 | (*debug_mode == FF_VULKAN_DEBUG_PRACTICES)) { |
| 1374 | VkDebugUtilsMessengerCreateInfoEXT dbg = { |
| 1375 | .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_MESSENGER_CREATE_INFO_EXT, |
| 1376 | .messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT | |
| 1377 | VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT | |
| 1378 | VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT | |
| 1379 | VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT, |
| 1380 | .messageType = VK_DEBUG_UTILS_MESSAGE_TYPE_GENERAL_BIT_EXT | |
| 1381 | VK_DEBUG_UTILS_MESSAGE_TYPE_VALIDATION_BIT_EXT | |
| 1382 | VK_DEBUG_UTILS_MESSAGE_TYPE_PERFORMANCE_BIT_EXT, |
| 1383 | .pfnUserCallback = vk_dbg_callback, |
| 1384 | .pUserData = ctx, |
| 1385 | }; |
| 1386 | |
| 1387 | vk->CreateDebugUtilsMessengerEXT(hwctx->inst, &dbg, |
| 1388 | hwctx->alloc, &p->debug_ctx); |
| 1389 | } |
| 1390 | |
| 1391 | err = 0; |
| 1392 | |
| 1393 | fail: |
| 1394 | RELEASE_PROPS(inst_props.ppEnabledLayerNames, inst_props.enabledLayerCount)if (inst_props.ppEnabledLayerNames) { for (int i = 0; i < inst_props .enabledLayerCount; i++) av_free((void *)((inst_props.ppEnabledLayerNames )[i])); av_free((void *)inst_props.ppEnabledLayerNames); }; |
| 1395 | return err; |
| 1396 | } |
| 1397 | |
| 1398 | typedef struct VulkanDeviceSelection { |
| 1399 | uint8_t uuid[VK_UUID_SIZE16U]; /* Will use this first unless !has_uuid */ |
| 1400 | int has_uuid; |
| 1401 | uint32_t drm_major; /* Will use this second unless !has_drm */ |
| 1402 | uint32_t drm_minor; /* Will use this second unless !has_drm */ |
| 1403 | uint32_t has_drm; /* has drm node info */ |
| 1404 | const char *name; /* Will use this third unless NULL */ |
| 1405 | uint32_t pci_device; /* Will use this fourth unless 0x0 */ |
| 1406 | uint32_t vendor_id; /* Last resort to find something deterministic */ |
| 1407 | int index; /* Finally fall back to index */ |
| 1408 | } VulkanDeviceSelection; |
| 1409 | |
| 1410 | static const char *vk_dev_type(enum VkPhysicalDeviceType type) |
| 1411 | { |
| 1412 | switch (type) { |
| 1413 | case VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU: return "integrated"; |
| 1414 | case VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU: return "discrete"; |
| 1415 | case VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU: return "virtual"; |
| 1416 | case VK_PHYSICAL_DEVICE_TYPE_CPU: return "software"; |
| 1417 | default: return "unknown"; |
| 1418 | } |
| 1419 | } |
| 1420 | |
| 1421 | /* Finds a device */ |
| 1422 | static int find_device(AVHWDeviceContext *ctx, VulkanDeviceSelection *select) |
| 1423 | { |
| 1424 | int err = 0, choice = -1; |
| 1425 | uint32_t num; |
| 1426 | VkResult ret; |
| 1427 | VulkanDevicePriv *p = ctx->hwctx; |
| 1428 | AVVulkanDeviceContext *hwctx = &p->p; |
| 1429 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 1430 | VkPhysicalDevice *devices = NULL((void*)0); |
| 1431 | VkPhysicalDeviceIDProperties *idp = NULL((void*)0); |
| 1432 | VkPhysicalDeviceProperties2 *prop = NULL((void*)0); |
| 1433 | VkPhysicalDeviceDriverProperties *driver_prop = NULL((void*)0); |
| 1434 | VkPhysicalDeviceDrmPropertiesEXT *drm_prop = NULL((void*)0); |
| 1435 | |
| 1436 | ret = vk->EnumeratePhysicalDevices(hwctx->inst, &num, NULL((void*)0)); |
| 1437 | if (ret != VK_SUCCESS || !num) { |
| 1438 | av_log(ctx, AV_LOG_ERROR16, "No devices found: %s!\n", ff_vk_ret2str(ret)); |
| 1439 | return AVERROR(ENODEV)(-(19)); |
| 1440 | } |
| 1441 | |
| 1442 | devices = av_malloc_array(num, sizeof(VkPhysicalDevice)); |
| 1443 | if (!devices) |
| 1444 | return AVERROR(ENOMEM)(-(12)); |
| 1445 | |
| 1446 | ret = vk->EnumeratePhysicalDevices(hwctx->inst, &num, devices); |
| 1447 | if (ret != VK_SUCCESS) { |
| 1448 | av_log(ctx, AV_LOG_ERROR16, "Failed enumerating devices: %s\n", |
| 1449 | ff_vk_ret2str(ret)); |
| 1450 | err = AVERROR(ENODEV)(-(19)); |
| 1451 | goto end; |
| 1452 | } |
| 1453 | |
| 1454 | prop = av_calloc(num, sizeof(*prop)); |
| 1455 | if (!prop) { |
| 1456 | err = AVERROR(ENOMEM)(-(12)); |
| 1457 | goto end; |
| 1458 | } |
| 1459 | |
| 1460 | idp = av_calloc(num, sizeof(*idp)); |
| 1461 | if (!idp) { |
| 1462 | err = AVERROR(ENOMEM)(-(12)); |
| 1463 | goto end; |
| 1464 | } |
| 1465 | |
| 1466 | driver_prop = av_calloc(num, sizeof(*driver_prop)); |
| 1467 | if (!driver_prop) { |
| 1468 | err = AVERROR(ENOMEM)(-(12)); |
| 1469 | goto end; |
| 1470 | } |
| 1471 | |
| 1472 | if (p->vkctx.extensions & FF_VK_EXT_DEVICE_DRM(1ULL << 9)) { |
| 1473 | drm_prop = av_calloc(num, sizeof(*drm_prop)); |
| 1474 | if (!drm_prop) { |
| 1475 | err = AVERROR(ENOMEM)(-(12)); |
| 1476 | goto end; |
| 1477 | } |
| 1478 | } |
| 1479 | |
| 1480 | av_log(ctx, AV_LOG_VERBOSE40, "GPU listing:\n"); |
| 1481 | for (int i = 0; i < num; i++) { |
| 1482 | if (p->vkctx.extensions & FF_VK_EXT_DEVICE_DRM(1ULL << 9)) { |
| 1483 | drm_prop[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRM_PROPERTIES_EXT; |
| 1484 | driver_prop[i].pNext = &drm_prop[i]; |
| 1485 | } |
| 1486 | driver_prop[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES; |
| 1487 | idp[i].pNext = &driver_prop[i]; |
| 1488 | idp[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES; |
| 1489 | prop[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2; |
| 1490 | prop[i].pNext = &idp[i]; |
| 1491 | |
| 1492 | vk->GetPhysicalDeviceProperties2(devices[i], &prop[i]); |
| 1493 | av_log(ctx, AV_LOG_VERBOSE40, " %d: %s (%s) (0x%x)\n", i, |
| 1494 | prop[i].properties.deviceName, |
| 1495 | vk_dev_type(prop[i].properties.deviceType), |
| 1496 | prop[i].properties.deviceID); |
| 1497 | } |
| 1498 | |
| 1499 | if (select->has_uuid) { |
| 1500 | for (int i = 0; i < num; i++) { |
| 1501 | if (!memcmp(idp[i].deviceUUID, select->uuid, VK_UUID_SIZE16U)) { |
| 1502 | choice = i; |
| 1503 | goto end; |
| 1504 | } |
| 1505 | } |
| 1506 | av_log(ctx, AV_LOG_ERROR16, "Unable to find device by given UUID!\n"); |
| 1507 | err = AVERROR(ENODEV)(-(19)); |
| 1508 | goto end; |
| 1509 | } else if ((p->vkctx.extensions & FF_VK_EXT_DEVICE_DRM(1ULL << 9)) && select->has_drm) { |
| 1510 | for (int i = 0; i < num; i++) { |
| 1511 | if ((select->drm_major == drm_prop[i].primaryMajor && |
| 1512 | select->drm_minor == drm_prop[i].primaryMinor) || |
| 1513 | (select->drm_major == drm_prop[i].renderMajor && |
| 1514 | select->drm_minor == drm_prop[i].renderMinor)) { |
| 1515 | choice = i; |
| 1516 | goto end; |
| 1517 | } |
| 1518 | } |
| 1519 | av_log(ctx, AV_LOG_ERROR16, "Unable to find device by given DRM node numbers %i:%i!\n", |
| 1520 | select->drm_major, select->drm_minor); |
| 1521 | err = AVERROR(ENODEV)(-(19)); |
| 1522 | goto end; |
| 1523 | } else if (select->name) { |
| 1524 | av_log(ctx, AV_LOG_VERBOSE40, "Requested device: %s\n", select->name); |
| 1525 | for (int i = 0; i < num; i++) { |
| 1526 | if (strstr(prop[i].properties.deviceName, select->name)) { |
| 1527 | choice = i; |
| 1528 | goto end; |
| 1529 | } |
| 1530 | } |
| 1531 | av_log(ctx, AV_LOG_ERROR16, "Unable to find device \"%s\"!\n", |
| 1532 | select->name); |
| 1533 | err = AVERROR(ENODEV)(-(19)); |
| 1534 | goto end; |
| 1535 | } else if (select->pci_device) { |
| 1536 | av_log(ctx, AV_LOG_VERBOSE40, "Requested device: 0x%x\n", select->pci_device); |
| 1537 | for (int i = 0; i < num; i++) { |
| 1538 | if (select->pci_device == prop[i].properties.deviceID) { |
| 1539 | choice = i; |
| 1540 | goto end; |
| 1541 | } |
| 1542 | } |
| 1543 | av_log(ctx, AV_LOG_ERROR16, "Unable to find device with PCI ID 0x%x!\n", |
| 1544 | select->pci_device); |
| 1545 | err = AVERROR(EINVAL)(-(22)); |
| 1546 | goto end; |
| 1547 | } else if (select->vendor_id) { |
| 1548 | av_log(ctx, AV_LOG_VERBOSE40, "Requested vendor: 0x%x\n", select->vendor_id); |
| 1549 | for (int i = 0; i < num; i++) { |
| 1550 | if (select->vendor_id == prop[i].properties.vendorID) { |
| 1551 | choice = i; |
| 1552 | goto end; |
| 1553 | } |
| 1554 | } |
| 1555 | av_log(ctx, AV_LOG_ERROR16, "Unable to find device with Vendor ID 0x%x!\n", |
| 1556 | select->vendor_id); |
| 1557 | err = AVERROR(ENODEV)(-(19)); |
| 1558 | goto end; |
| 1559 | } else { |
| 1560 | if (select->index < num) { |
| 1561 | choice = select->index; |
| 1562 | goto end; |
| 1563 | } |
| 1564 | av_log(ctx, AV_LOG_ERROR16, "Unable to find device with index %i!\n", |
| 1565 | select->index); |
| 1566 | err = AVERROR(ENODEV)(-(19)); |
| 1567 | goto end; |
| 1568 | } |
| 1569 | |
| 1570 | end: |
| 1571 | if (choice > -1) { |
| 1572 | av_log(ctx, AV_LOG_VERBOSE40, "Device %d selected: %s (%s) (0x%x)\n", |
| 1573 | choice, prop[choice].properties.deviceName, |
| 1574 | vk_dev_type(prop[choice].properties.deviceType), |
| 1575 | prop[choice].properties.deviceID); |
| 1576 | hwctx->phys_dev = devices[choice]; |
| 1577 | p->props = prop[choice]; |
| 1578 | p->props.pNext = NULL((void*)0); |
| 1579 | p->dprops = driver_prop[choice]; |
| 1580 | p->dprops.pNext = NULL((void*)0); |
| 1581 | } |
| 1582 | |
| 1583 | av_free(devices); |
| 1584 | av_free(prop); |
| 1585 | av_free(idp); |
| 1586 | av_free(drm_prop); |
| 1587 | av_free(driver_prop); |
| 1588 | |
| 1589 | return err; |
| 1590 | } |
| 1591 | |
| 1592 | /* Picks the least used qf with the fewest unneeded flags, or -1 if none found */ |
| 1593 | static inline int pick_queue_family(VkQueueFamilyProperties2 *qf, uint32_t num_qf, |
| 1594 | VkQueueFlagBits flags) |
| 1595 | { |
| 1596 | int index = -1; |
| 1597 | uint32_t min_score = UINT32_MAX(4294967295U); |
| 1598 | |
| 1599 | for (int i = 0; i < num_qf; i++) { |
| 1600 | VkQueueFlagBits qflags = qf[i].queueFamilyProperties.queueFlags; |
| 1601 | |
| 1602 | /* Per the spec, reporting transfer caps is optional for these 2 types */ |
| 1603 | if ((flags & VK_QUEUE_TRANSFER_BIT) && |
| 1604 | (qflags & (VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT))) |
| 1605 | qflags |= VK_QUEUE_TRANSFER_BIT; |
| 1606 | |
| 1607 | if (qflags & flags) { |
| 1608 | uint32_t score = av_popcountav_popcount_c(qflags) + qf[i].queueFamilyProperties.timestampValidBits; |
| 1609 | if (score < min_score) { |
| 1610 | index = i; |
| 1611 | min_score = score; |
| 1612 | } |
| 1613 | } |
| 1614 | } |
| 1615 | |
| 1616 | if (index > -1) |
| 1617 | qf[index].queueFamilyProperties.timestampValidBits++; |
| 1618 | |
| 1619 | return index; |
| 1620 | } |
| 1621 | |
| 1622 | static inline int pick_video_queue_family(VkQueueFamilyProperties2 *qf, |
| 1623 | VkQueueFamilyVideoPropertiesKHR *qf_vid, uint32_t num_qf, |
| 1624 | VkVideoCodecOperationFlagsKHR flags) |
| 1625 | { |
| 1626 | int index = -1; |
| 1627 | uint32_t min_score = UINT32_MAX(4294967295U); |
| 1628 | |
| 1629 | for (int i = 0; i < num_qf; i++) { |
| 1630 | const VkQueueFlags qflags = qf[i].queueFamilyProperties.queueFlags; |
| 1631 | const VkVideoCodecOperationFlagsKHR vflags = qf_vid[i].videoCodecOperations; |
| 1632 | |
| 1633 | if (!(qflags & (VK_QUEUE_VIDEO_ENCODE_BIT_KHR | VK_QUEUE_VIDEO_DECODE_BIT_KHR))) |
| 1634 | continue; |
| 1635 | |
| 1636 | if (vflags & flags) { |
| 1637 | uint32_t score = av_popcountav_popcount_c(vflags) + qf[i].queueFamilyProperties.timestampValidBits; |
| 1638 | if (score < min_score) { |
| 1639 | index = i; |
| 1640 | min_score = score; |
| 1641 | } |
| 1642 | } |
| 1643 | } |
| 1644 | |
| 1645 | if (index > -1) |
| 1646 | qf[index].queueFamilyProperties.timestampValidBits++; |
| 1647 | |
| 1648 | return index; |
| 1649 | } |
| 1650 | |
| 1651 | static int setup_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd) |
| 1652 | { |
| 1653 | uint32_t num; |
| 1654 | VulkanDevicePriv *p = ctx->hwctx; |
| 1655 | AVVulkanDeviceContext *hwctx = &p->p; |
| 1656 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 1657 | |
| 1658 | VkQueueFamilyProperties2 *qf = NULL((void*)0); |
| 1659 | VkQueueFamilyVideoPropertiesKHR *qf_vid = NULL((void*)0); |
| 1660 | |
| 1661 | /* First get the number of queue families */ |
| 1662 | vk->GetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, NULL((void*)0)); |
| 1663 | if (!num) { |
| 1664 | av_log(ctx, AV_LOG_ERROR16, "Failed to get queues!\n"); |
| 1665 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 1666 | } |
| 1667 | |
| 1668 | /* Then allocate memory */ |
| 1669 | qf = av_malloc_array(num, sizeof(VkQueueFamilyProperties2)); |
| 1670 | if (!qf) |
| 1671 | return AVERROR(ENOMEM)(-(12)); |
| 1672 | |
| 1673 | qf_vid = av_malloc_array(num, sizeof(VkQueueFamilyVideoPropertiesKHR)); |
| 1674 | if (!qf_vid) |
| 1675 | return AVERROR(ENOMEM)(-(12)); |
| 1676 | |
| 1677 | for (uint32_t i = 0; i < num; i++) { |
| 1678 | qf_vid[i] = (VkQueueFamilyVideoPropertiesKHR) { |
| 1679 | .sType = VK_STRUCTURE_TYPE_QUEUE_FAMILY_VIDEO_PROPERTIES_KHR, |
| 1680 | }; |
| 1681 | qf[i] = (VkQueueFamilyProperties2) { |
| 1682 | .sType = VK_STRUCTURE_TYPE_QUEUE_FAMILY_PROPERTIES_2, |
| 1683 | .pNext = p->vkctx.extensions & FF_VK_EXT_VIDEO_QUEUE(1ULL << 36) ? &qf_vid[i] : NULL((void*)0), |
| 1684 | }; |
| 1685 | } |
| 1686 | |
| 1687 | /* Finally retrieve the queue families */ |
| 1688 | vk->GetPhysicalDeviceQueueFamilyProperties2(hwctx->phys_dev, &num, qf); |
| 1689 | |
| 1690 | av_log(ctx, AV_LOG_VERBOSE40, "Queue families:\n"); |
| 1691 | for (int i = 0; i < num; i++) { |
| 1692 | av_log(ctx, AV_LOG_VERBOSE40, " %i:%s%s%s%s%s%s%s%s (queues: %i)\n", i, |
| 1693 | ((qf[i].queueFamilyProperties.queueFlags) & VK_QUEUE_GRAPHICS_BIT) ? " graphics" : "", |
| 1694 | ((qf[i].queueFamilyProperties.queueFlags) & VK_QUEUE_COMPUTE_BIT) ? " compute" : "", |
| 1695 | ((qf[i].queueFamilyProperties.queueFlags) & VK_QUEUE_TRANSFER_BIT) ? " transfer" : "", |
| 1696 | ((qf[i].queueFamilyProperties.queueFlags) & VK_QUEUE_VIDEO_ENCODE_BIT_KHR) ? " encode" : "", |
| 1697 | ((qf[i].queueFamilyProperties.queueFlags) & VK_QUEUE_VIDEO_DECODE_BIT_KHR) ? " decode" : "", |
| 1698 | ((qf[i].queueFamilyProperties.queueFlags) & VK_QUEUE_SPARSE_BINDING_BIT) ? " sparse" : "", |
| 1699 | ((qf[i].queueFamilyProperties.queueFlags) & VK_QUEUE_OPTICAL_FLOW_BIT_NV) ? " optical_flow" : "", |
| 1700 | ((qf[i].queueFamilyProperties.queueFlags) & VK_QUEUE_PROTECTED_BIT) ? " protected" : "", |
| 1701 | qf[i].queueFamilyProperties.queueCount); |
| 1702 | |
| 1703 | /* We use this field to keep a score of how many times we've used that |
| 1704 | * queue family in order to make better choices. */ |
| 1705 | qf[i].queueFamilyProperties.timestampValidBits = 0; |
| 1706 | } |
| 1707 | |
| 1708 | hwctx->nb_qf = 0; |
| 1709 | hwctx->queue_flags = 0; |
| 1710 | #ifdef VK_KHR_internally_synchronized_queues1 |
| 1711 | if (p->vkctx.extensions & FF_VK_EXT_INTERNAL_QUEUE_SYNC(1ULL << 23)) |
| 1712 | hwctx->queue_flags |= VK_DEVICE_QUEUE_CREATE_INTERNALLY_SYNCHRONIZED_BIT_KHR; |
| 1713 | #endif |
| 1714 | |
| 1715 | /* Pick each queue family to use. */ |
| 1716 | #define PICK_QF(type, vid_op) \ |
| 1717 | do { \ |
| 1718 | uint32_t i; \ |
| 1719 | uint32_t idx; \ |
| 1720 | \ |
| 1721 | if (vid_op) \ |
| 1722 | idx = pick_video_queue_family(qf, qf_vid, num, vid_op); \ |
| 1723 | else \ |
| 1724 | idx = pick_queue_family(qf, num, type); \ |
| 1725 | \ |
| 1726 | if (idx == -1) \ |
| 1727 | continue; \ |
| 1728 | \ |
| 1729 | for (i = 0; i < hwctx->nb_qf; i++) { \ |
| 1730 | if (hwctx->qf[i].idx == idx) { \ |
| 1731 | hwctx->qf[i].flags |= type; \ |
| 1732 | hwctx->qf[i].video_caps |= vid_op; \ |
| 1733 | break; \ |
| 1734 | } \ |
| 1735 | } \ |
| 1736 | if (i == hwctx->nb_qf) { \ |
| 1737 | hwctx->qf[i].idx = idx; \ |
| 1738 | hwctx->qf[i].num = qf[idx].queueFamilyProperties.queueCount; \ |
| 1739 | if (p->limit_queues || \ |
| 1740 | p->dprops.driverID == VK_DRIVER_ID_NVIDIA_PROPRIETARY) { \ |
| 1741 | int max = p->limit_queues; \ |
| 1742 | if (type == VK_QUEUE_GRAPHICS_BIT || \ |
| 1743 | type == VK_QUEUE_COMPUTE_BIT) \ |
| 1744 | hwctx->qf[i].num = FFMIN(hwctx->qf[i].num, \((hwctx->qf[i].num) > (max ? max : 1) ? (max ? max : 1) : (hwctx->qf[i].num)) |
| 1745 | max ? max : 1)((hwctx->qf[i].num) > (max ? max : 1) ? (max ? max : 1) : (hwctx->qf[i].num)); \ |
| 1746 | else if (max) \ |
| 1747 | hwctx->qf[i].num = FFMIN(hwctx->qf[i].num, max)((hwctx->qf[i].num) > (max) ? (max) : (hwctx->qf[i]. num)); \ |
| 1748 | } \ |
| 1749 | hwctx->qf[i].flags = type; \ |
| 1750 | hwctx->qf[i].video_caps = vid_op; \ |
| 1751 | hwctx->nb_qf++; \ |
| 1752 | } \ |
| 1753 | } while (0) |
| 1754 | |
| 1755 | PICK_QF(VK_QUEUE_GRAPHICS_BIT, VK_VIDEO_CODEC_OPERATION_NONE_KHR); |
| 1756 | PICK_QF(VK_QUEUE_COMPUTE_BIT, VK_VIDEO_CODEC_OPERATION_NONE_KHR); |
| 1757 | PICK_QF(VK_QUEUE_TRANSFER_BIT, VK_VIDEO_CODEC_OPERATION_NONE_KHR); |
| 1758 | |
| 1759 | PICK_QF(VK_QUEUE_VIDEO_ENCODE_BIT_KHR, VK_VIDEO_CODEC_OPERATION_ENCODE_H264_BIT_KHR); |
| 1760 | PICK_QF(VK_QUEUE_VIDEO_DECODE_BIT_KHR, VK_VIDEO_CODEC_OPERATION_DECODE_H264_BIT_KHR); |
| 1761 | |
| 1762 | PICK_QF(VK_QUEUE_VIDEO_ENCODE_BIT_KHR, VK_VIDEO_CODEC_OPERATION_ENCODE_H265_BIT_KHR); |
| 1763 | PICK_QF(VK_QUEUE_VIDEO_DECODE_BIT_KHR, VK_VIDEO_CODEC_OPERATION_DECODE_H265_BIT_KHR); |
| 1764 | |
| 1765 | #ifdef VK_KHR_video_decode_vp91 |
| 1766 | PICK_QF(VK_QUEUE_VIDEO_DECODE_BIT_KHR, VK_VIDEO_CODEC_OPERATION_DECODE_VP9_BIT_KHR); |
| 1767 | #endif |
| 1768 | |
| 1769 | #ifdef VK_KHR_video_encode_av11 |
| 1770 | PICK_QF(VK_QUEUE_VIDEO_ENCODE_BIT_KHR, VK_VIDEO_CODEC_OPERATION_ENCODE_AV1_BIT_KHR); |
| 1771 | #endif |
| 1772 | PICK_QF(VK_QUEUE_VIDEO_DECODE_BIT_KHR, VK_VIDEO_CODEC_OPERATION_DECODE_AV1_BIT_KHR); |
| 1773 | |
| 1774 | if (p->vkctx.extensions & FF_VK_EXT_OPTICAL_FLOW(1ULL << 12)) |
| 1775 | PICK_QF(VK_QUEUE_OPTICAL_FLOW_BIT_NV, VK_VIDEO_CODEC_OPERATION_NONE_KHR); |
| 1776 | |
| 1777 | av_free(qf); |
| 1778 | av_free(qf_vid); |
| 1779 | |
| 1780 | #undef PICK_QF |
| 1781 | |
| 1782 | cd->pQueueCreateInfos = av_malloc_array(hwctx->nb_qf, |
| 1783 | sizeof(VkDeviceQueueCreateInfo)); |
| 1784 | if (!cd->pQueueCreateInfos) |
| 1785 | return AVERROR(ENOMEM)(-(12)); |
| 1786 | |
| 1787 | for (uint32_t i = 0; i < hwctx->nb_qf; i++) { |
| 1788 | int dup = 0; |
| 1789 | float *weights = NULL((void*)0); |
| 1790 | VkDeviceQueueCreateInfo *pc; |
| 1791 | for (uint32_t j = 0; j < cd->queueCreateInfoCount; j++) { |
| 1792 | if (hwctx->qf[i].idx == cd->pQueueCreateInfos[j].queueFamilyIndex) { |
| 1793 | dup = 1; |
| 1794 | break; |
| 1795 | } |
| 1796 | } |
| 1797 | if (dup) |
| 1798 | continue; |
| 1799 | |
| 1800 | weights = av_malloc_array(hwctx->qf[i].num, sizeof(float)); |
| 1801 | if (!weights) { |
| 1802 | for (uint32_t j = 0; j < cd->queueCreateInfoCount; j++) |
| 1803 | av_free((void *)cd->pQueueCreateInfos[i].pQueuePriorities); |
| 1804 | av_free((void *)cd->pQueueCreateInfos); |
| 1805 | return AVERROR(ENOMEM)(-(12)); |
| 1806 | } |
| 1807 | |
| 1808 | for (uint32_t j = 0; j < hwctx->qf[i].num; j++) |
| 1809 | weights[j] = 1.0; |
| 1810 | |
| 1811 | pc = (VkDeviceQueueCreateInfo *)cd->pQueueCreateInfos; |
| 1812 | pc[cd->queueCreateInfoCount++] = (VkDeviceQueueCreateInfo) { |
| 1813 | .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, |
| 1814 | .flags = hwctx->queue_flags, |
| 1815 | .queueFamilyIndex = hwctx->qf[i].idx, |
| 1816 | .queueCount = hwctx->qf[i].num, |
| 1817 | .pQueuePriorities = weights, |
| 1818 | }; |
| 1819 | } |
| 1820 | |
| 1821 | return 0; |
| 1822 | } |
| 1823 | |
| 1824 | /* Only resources created by vulkan_device_create should be released here, |
| 1825 | * resources created by vulkan_device_init should be released by |
| 1826 | * vulkan_device_uninit, to make sure we don't free user provided resources, |
| 1827 | * and there is no leak. |
| 1828 | */ |
| 1829 | static void vulkan_device_free(AVHWDeviceContext *ctx) |
| 1830 | { |
| 1831 | VulkanDevicePriv *p = ctx->hwctx; |
| 1832 | AVVulkanDeviceContext *hwctx = &p->p; |
| 1833 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 1834 | |
| 1835 | if (hwctx->act_dev) |
| 1836 | vk->DestroyDevice(hwctx->act_dev, hwctx->alloc); |
| 1837 | |
| 1838 | if (p->debug_ctx) |
| 1839 | vk->DestroyDebugUtilsMessengerEXT(hwctx->inst, p->debug_ctx, |
| 1840 | hwctx->alloc); |
| 1841 | |
| 1842 | if (hwctx->inst) |
| 1843 | vk->DestroyInstance(hwctx->inst, hwctx->alloc); |
| 1844 | |
| 1845 | if (p->libvulkan) |
| 1846 | dlclose(p->libvulkan); |
| 1847 | |
| 1848 | RELEASE_PROPS(hwctx->enabled_inst_extensions, hwctx->nb_enabled_inst_extensions)if (hwctx->enabled_inst_extensions) { for (int i = 0; i < hwctx->nb_enabled_inst_extensions; i++) av_free((void *)( (hwctx->enabled_inst_extensions)[i])); av_free((void *)hwctx ->enabled_inst_extensions); }; |
| 1849 | RELEASE_PROPS(hwctx->enabled_dev_extensions, hwctx->nb_enabled_dev_extensions)if (hwctx->enabled_dev_extensions) { for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++) av_free((void *)(( hwctx->enabled_dev_extensions)[i])); av_free((void *)hwctx ->enabled_dev_extensions); }; |
| 1850 | } |
| 1851 | |
| 1852 | static void vulkan_device_uninit(AVHWDeviceContext *ctx) |
| 1853 | { |
| 1854 | VulkanDevicePriv *p = ctx->hwctx; |
| 1855 | |
| 1856 | if (p->qf_mutex) { |
| 1857 | for (uint32_t i = 0; i < p->nb_tot_qfs; i++) { |
| 1858 | pthread_mutex_destroystrict_pthread_mutex_destroy(p->qf_mutex[i]); |
| 1859 | av_freep(&p->qf_mutex[i]); |
| 1860 | } |
| 1861 | av_freep(&p->qf_mutex); |
| 1862 | } |
| 1863 | |
| 1864 | ff_vk_uninit(&p->vkctx); |
| 1865 | } |
| 1866 | |
| 1867 | static int vulkan_device_create_internal(AVHWDeviceContext *ctx, |
| 1868 | VulkanDeviceSelection *dev_select, |
| 1869 | int disable_multiplane, |
| 1870 | AVDictionary *opts, int flags) |
| 1871 | { |
| 1872 | int err = 0; |
| 1873 | VkResult ret; |
| 1874 | AVDictionaryEntry *opt_d; |
| 1875 | VulkanDevicePriv *p = ctx->hwctx; |
| 1876 | AVVulkanDeviceContext *hwctx = &p->p; |
| 1877 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 1878 | enum FFVulkanDebugMode debug_mode = FF_VULKAN_DEBUG_NONE; |
| 1879 | VulkanDeviceFeatures supported_feats = { 0 }; |
| 1880 | VkDeviceCreateInfo dev_info = { |
| 1881 | .sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO, |
| 1882 | }; |
| 1883 | |
| 1884 | /* Create an instance if not given one */ |
| 1885 | if ((err = create_instance(ctx, opts, &debug_mode))) |
| 1886 | goto end; |
| 1887 | |
| 1888 | /* Find a physical device (if not given one) */ |
| 1889 | if ((err = find_device(ctx, dev_select))) |
| 1890 | goto end; |
| 1891 | |
| 1892 | /* Get supported memory types */ |
| 1893 | vk->GetPhysicalDeviceMemoryProperties(hwctx->phys_dev, &p->mprops); |
| 1894 | |
| 1895 | /* Find and enable extensions for the physical device */ |
| 1896 | if ((err = check_extensions(ctx, 1, opts, &dev_info.ppEnabledExtensionNames, |
| 1897 | &dev_info.enabledExtensionCount, debug_mode))) { |
| 1898 | for (int i = 0; i < dev_info.queueCreateInfoCount; i++) |
| 1899 | av_free((void *)dev_info.pQueueCreateInfos[i].pQueuePriorities); |
| 1900 | av_free((void *)dev_info.pQueueCreateInfos); |
| 1901 | goto end; |
| 1902 | } |
| 1903 | |
| 1904 | /* Get all supported features for the physical device */ |
| 1905 | device_features_init(ctx, &supported_feats); |
| 1906 | vk->GetPhysicalDeviceFeatures2(hwctx->phys_dev, &supported_feats.device); |
| 1907 | |
| 1908 | /* Copy all needed features from those supported and activate them */ |
| 1909 | device_features_init(ctx, &p->feats); |
| 1910 | device_features_copy_needed(&p->feats, &supported_feats); |
| 1911 | dev_info.pNext = p->feats.device.pNext; |
| 1912 | dev_info.pEnabledFeatures = &p->feats.device.features; |
| 1913 | |
| 1914 | /* Limit queues to a given number if needed */ |
| 1915 | opt_d = av_dict_get(opts, "limit_queues", NULL((void*)0), 0); |
| 1916 | if (opt_d) |
| 1917 | p->limit_queues = strtol(opt_d->value, NULL((void*)0), 10); |
| 1918 | |
| 1919 | /* Setup enabled queue families */ |
| 1920 | if ((err = setup_queue_families(ctx, &dev_info))) |
| 1921 | goto end; |
| 1922 | |
| 1923 | /* Finally create the device */ |
| 1924 | ret = vk->CreateDevice(hwctx->phys_dev, &dev_info, hwctx->alloc, |
| 1925 | &hwctx->act_dev); |
| 1926 | |
| 1927 | for (int i = 0; i < dev_info.queueCreateInfoCount; i++) |
| 1928 | av_free((void *)dev_info.pQueueCreateInfos[i].pQueuePriorities); |
| 1929 | av_free((void *)dev_info.pQueueCreateInfos); |
| 1930 | |
| 1931 | if (ret != VK_SUCCESS) { |
| 1932 | av_log(ctx, AV_LOG_ERROR16, "Device creation failure: %s\n", |
| 1933 | ff_vk_ret2str(ret)); |
| 1934 | for (int i = 0; i < dev_info.enabledExtensionCount; i++) |
| 1935 | av_free((void *)dev_info.ppEnabledExtensionNames[i]); |
| 1936 | av_free((void *)dev_info.ppEnabledExtensionNames); |
| 1937 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 1938 | goto end; |
| 1939 | } |
| 1940 | |
| 1941 | /* Tiled images setting, use them by default */ |
| 1942 | opt_d = av_dict_get(opts, "linear_images", NULL((void*)0), 0); |
| 1943 | if (opt_d) |
| 1944 | p->use_linear_images = strtol(opt_d->value, NULL((void*)0), 10); |
| 1945 | |
| 1946 | /* The disable_multiplane argument takes precedent over the option */ |
| 1947 | p->disable_multiplane = disable_multiplane; |
| 1948 | if (!p->disable_multiplane) { |
| 1949 | opt_d = av_dict_get(opts, "disable_multiplane", NULL((void*)0), 0); |
| 1950 | if (opt_d) |
| 1951 | p->disable_multiplane = strtol(opt_d->value, NULL((void*)0), 10); |
| 1952 | } |
| 1953 | |
| 1954 | /* Disable host pointer imports (by default on nvidia) */ |
| 1955 | p->avoid_host_import = p->dprops.driverID == VK_DRIVER_ID_NVIDIA_PROPRIETARY; |
| 1956 | opt_d = av_dict_get(opts, "avoid_host_import", NULL((void*)0), 0); |
| 1957 | if (opt_d) |
| 1958 | p->avoid_host_import = strtol(opt_d->value, NULL((void*)0), 10); |
| 1959 | |
| 1960 | /* Set the public device feature struct and its pNext chain */ |
| 1961 | hwctx->device_features = p->feats.device; |
| 1962 | |
| 1963 | /* Set the list of all active extensions */ |
| 1964 | hwctx->enabled_dev_extensions = dev_info.ppEnabledExtensionNames; |
| 1965 | hwctx->nb_enabled_dev_extensions = dev_info.enabledExtensionCount; |
| 1966 | |
| 1967 | /* The extension lists need to be freed */ |
| 1968 | ctx->free = vulkan_device_free; |
| 1969 | |
| 1970 | end: |
| 1971 | return err; |
| 1972 | } |
| 1973 | |
| 1974 | static void lock_queue(AVHWDeviceContext *ctx, uint32_t queue_family, uint32_t index) |
| 1975 | { |
| 1976 | VulkanDevicePriv *p = ctx->hwctx; |
| 1977 | if (p->qf_mutex) |
| 1978 | pthread_mutex_lockstrict_pthread_mutex_lock(&p->qf_mutex[queue_family][index]); |
| 1979 | } |
| 1980 | |
| 1981 | static void unlock_queue(AVHWDeviceContext *ctx, uint32_t queue_family, uint32_t index) |
| 1982 | { |
| 1983 | VulkanDevicePriv *p = ctx->hwctx; |
| 1984 | if (p->qf_mutex) |
| 1985 | pthread_mutex_unlockstrict_pthread_mutex_unlock(&p->qf_mutex[queue_family][index]); |
| 1986 | } |
| 1987 | |
| 1988 | static int vulkan_device_init(AVHWDeviceContext *ctx) |
| 1989 | { |
| 1990 | int err = 0; |
| 1991 | uint32_t qf_num; |
| 1992 | VulkanDevicePriv *p = ctx->hwctx; |
| 1993 | AVVulkanDeviceContext *hwctx = &p->p; |
| 1994 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 1995 | VkQueueFamilyProperties2 *qf; |
| 1996 | VkQueueFamilyVideoPropertiesKHR *qf_vid; |
| 1997 | VkPhysicalDeviceExternalSemaphoreInfo ext_sem_props_info; |
| 1998 | |
| 1999 | /* Set device extension flags */ |
| 2000 | for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++) { |
| 2001 | for (int j = 0; j < FF_ARRAY_ELEMS(optional_device_exts)(sizeof(optional_device_exts) / sizeof((optional_device_exts) [0])); j++) { |
| 2002 | if (!strcmp(hwctx->enabled_dev_extensions[i], |
| 2003 | optional_device_exts[j].name)) { |
| 2004 | p->vkctx.extensions |= optional_device_exts[j].flag; |
| 2005 | break; |
| 2006 | } |
| 2007 | } |
| 2008 | } |
| 2009 | |
| 2010 | err = ff_vk_load_functions(ctx, vk, p->vkctx.extensions, 1, 1); |
| 2011 | if (err < 0) { |
| 2012 | av_log(ctx, AV_LOG_ERROR16, "Unable to load functions!\n"); |
| 2013 | return err; |
| 2014 | } |
| 2015 | |
| 2016 | p->props.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2; |
| 2017 | p->props.pNext = &p->hprops; |
| 2018 | p->hprops.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT; |
| 2019 | p->hprops.pNext = &p->dprops; |
| 2020 | p->dprops.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES; |
| 2021 | |
| 2022 | vk->GetPhysicalDeviceProperties2(hwctx->phys_dev, &p->props); |
| 2023 | av_log(ctx, AV_LOG_VERBOSE40, "Using device: %s\n", |
| 2024 | p->props.properties.deviceName); |
| 2025 | av_log(ctx, AV_LOG_VERBOSE40, "Alignments:\n"); |
| 2026 | av_log(ctx, AV_LOG_VERBOSE40, " optimalBufferCopyRowPitchAlignment: %"PRIu64"l" "u""\n", |
| 2027 | p->props.properties.limits.optimalBufferCopyRowPitchAlignment); |
| 2028 | av_log(ctx, AV_LOG_VERBOSE40, " minMemoryMapAlignment: %zu\n", |
| 2029 | p->props.properties.limits.minMemoryMapAlignment); |
| 2030 | av_log(ctx, AV_LOG_VERBOSE40, " nonCoherentAtomSize: %"PRIu64"l" "u""\n", |
| 2031 | p->props.properties.limits.nonCoherentAtomSize); |
| 2032 | if (p->vkctx.extensions & FF_VK_EXT_EXTERNAL_HOST_MEMORY(1ULL << 4)) |
| 2033 | av_log(ctx, AV_LOG_VERBOSE40, " minImportedHostPointerAlignment: %"PRIu64"l" "u""\n", |
| 2034 | p->hprops.minImportedHostPointerAlignment); |
| 2035 | |
| 2036 | vk->GetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &qf_num, NULL((void*)0)); |
| 2037 | if (!qf_num) { |
| 2038 | av_log(ctx, AV_LOG_ERROR16, "Failed to get queues!\n"); |
| 2039 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 2040 | } |
| 2041 | |
| 2042 | ext_sem_props_info = (VkPhysicalDeviceExternalSemaphoreInfo) { |
| 2043 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_SEMAPHORE_INFO, |
| 2044 | }; |
| 2045 | |
| 2046 | /* Opaque FD semaphore properties */ |
| 2047 | ext_sem_props_info.handleType = |
| 2048 | #ifdef _WIN32 |
| 2049 | IsWindows8OrGreater() |
| 2050 | ? VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_BIT |
| 2051 | : VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT; |
| 2052 | #else |
| 2053 | VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT; |
| 2054 | #endif |
| 2055 | p->ext_sem_props_opaque.sType = VK_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_PROPERTIES; |
| 2056 | vk->GetPhysicalDeviceExternalSemaphoreProperties(hwctx->phys_dev, |
| 2057 | &ext_sem_props_info, |
| 2058 | &p->ext_sem_props_opaque); |
| 2059 | |
| 2060 | qf = av_malloc_array(qf_num, sizeof(VkQueueFamilyProperties2)); |
| 2061 | if (!qf) |
| 2062 | return AVERROR(ENOMEM)(-(12)); |
| 2063 | |
| 2064 | qf_vid = av_malloc_array(qf_num, sizeof(VkQueueFamilyVideoPropertiesKHR)); |
| 2065 | if (!qf_vid) { |
| 2066 | av_free(qf); |
| 2067 | return AVERROR(ENOMEM)(-(12)); |
| 2068 | } |
| 2069 | |
| 2070 | for (uint32_t i = 0; i < qf_num; i++) { |
| 2071 | qf_vid[i] = (VkQueueFamilyVideoPropertiesKHR) { |
| 2072 | .sType = VK_STRUCTURE_TYPE_QUEUE_FAMILY_VIDEO_PROPERTIES_KHR, |
| 2073 | }; |
| 2074 | qf[i] = (VkQueueFamilyProperties2) { |
| 2075 | .sType = VK_STRUCTURE_TYPE_QUEUE_FAMILY_PROPERTIES_2, |
| 2076 | .pNext = p->vkctx.extensions & FF_VK_EXT_VIDEO_QUEUE(1ULL << 36) ? &qf_vid[i] : NULL((void*)0), |
| 2077 | }; |
| 2078 | } |
| 2079 | |
| 2080 | vk->GetPhysicalDeviceQueueFamilyProperties2(hwctx->phys_dev, &qf_num, qf); |
| 2081 | |
| 2082 | p->nb_tot_qfs = qf_num; |
| 2083 | |
| 2084 | if (!(p->vkctx.extensions & FF_VK_EXT_INTERNAL_QUEUE_SYNC(1ULL << 23))) { |
| 2085 | p->qf_mutex = av_calloc(qf_num, sizeof(*p->qf_mutex)); |
| 2086 | if (!p->qf_mutex) { |
| 2087 | err = AVERROR(ENOMEM)(-(12)); |
| 2088 | goto end; |
| 2089 | } |
| 2090 | |
| 2091 | for (uint32_t i = 0; i < qf_num; i++) { |
| 2092 | p->qf_mutex[i] = av_calloc(qf[i].queueFamilyProperties.queueCount, |
| 2093 | sizeof(**p->qf_mutex)); |
| 2094 | if (!p->qf_mutex[i]) { |
| 2095 | err = AVERROR(ENOMEM)(-(12)); |
| 2096 | goto end; |
| 2097 | } |
| 2098 | for (uint32_t j = 0; j < qf[i].queueFamilyProperties.queueCount; j++) { |
| 2099 | err = pthread_mutex_initstrict_pthread_mutex_init(&p->qf_mutex[i][j], NULL((void*)0)); |
| 2100 | if (err != 0) { |
| 2101 | av_log(ctx, AV_LOG_ERROR16, "pthread_mutex_init failed : %s\n", |
| 2102 | av_err2str(err)av_make_error_string((char[64]){0}, 64, err)); |
| 2103 | err = AVERROR(err)(-(err)); |
| 2104 | goto end; |
| 2105 | } |
| 2106 | } |
| 2107 | } |
| 2108 | } |
| 2109 | |
| 2110 | for (int i = 0; i < hwctx->nb_qf; i++) { |
| 2111 | if (!hwctx->qf[i].video_caps && |
| 2112 | hwctx->qf[i].flags & (VK_QUEUE_VIDEO_DECODE_BIT_KHR | |
| 2113 | VK_QUEUE_VIDEO_ENCODE_BIT_KHR)) { |
| 2114 | hwctx->qf[i].video_caps = qf_vid[hwctx->qf[i].idx].videoCodecOperations; |
| 2115 | } |
| 2116 | } |
| 2117 | |
| 2118 | /* Setup array for pQueueFamilyIndices with used queue families */ |
| 2119 | p->nb_img_qfs = 0; |
| 2120 | for (int i = 0; i < hwctx->nb_qf; i++) { |
| 2121 | int seen = 0; |
| 2122 | /* Make sure each entry is unique |
| 2123 | * (VUID-VkBufferCreateInfo-sharingMode-01419) */ |
| 2124 | for (int j = (i - 1); j >= 0; j--) { |
| 2125 | if (hwctx->qf[i].idx == hwctx->qf[j].idx) { |
| 2126 | seen = 1; |
| 2127 | break; |
| 2128 | } |
| 2129 | } |
| 2130 | if (!seen) |
| 2131 | p->img_qfs[p->nb_img_qfs++] = hwctx->qf[i].idx; |
| 2132 | } |
| 2133 | |
| 2134 | #if FF_API_VULKAN_SYNC_QUEUES(61 < 62) |
| 2135 | FF_DISABLE_DEPRECATION_WARNINGSGCC diagnostic push
GCC diagnostic ignored "-Wdeprecated-declarations" |
| 2136 | if (!hwctx->lock_queue) |
| 2137 | hwctx->lock_queue = lock_queue; |
| 2138 | if (!hwctx->unlock_queue) |
| 2139 | hwctx->unlock_queue = unlock_queue; |
| 2140 | FF_ENABLE_DEPRECATION_WARNINGSGCC diagnostic pop |
| 2141 | #endif |
| 2142 | |
| 2143 | /* Re-query device capabilities, in case the device was created externally */ |
| 2144 | vk->GetPhysicalDeviceMemoryProperties(hwctx->phys_dev, &p->mprops); |
| 2145 | |
| 2146 | p->vkctx.device = ctx; |
| 2147 | p->vkctx.hwctx = hwctx; |
| 2148 | |
| 2149 | ff_vk_load_props(&p->vkctx); |
| 2150 | p->compute_qf = ff_vk_qf_find(&p->vkctx, VK_QUEUE_COMPUTE_BIT, 0); |
| 2151 | p->transfer_qf = ff_vk_qf_find(&p->vkctx, VK_QUEUE_TRANSFER_BIT, 0); |
| 2152 | |
| 2153 | /* Re-query device capabilities, in case the device was created externally */ |
| 2154 | vk->GetPhysicalDeviceMemoryProperties(hwctx->phys_dev, &p->mprops); |
| 2155 | |
| 2156 | end: |
| 2157 | av_free(qf_vid); |
| 2158 | av_free(qf); |
| 2159 | return err; |
| 2160 | } |
| 2161 | |
| 2162 | static int vulkan_device_create(AVHWDeviceContext *ctx, const char *device, |
| 2163 | AVDictionary *opts, int flags) |
| 2164 | { |
| 2165 | VulkanDeviceSelection dev_select = { 0 }; |
| 2166 | if (device && device[0]) { |
| 2167 | char *end = NULL((void*)0); |
| 2168 | |
| 2169 | if (!av_uuid_parse(device, dev_select.uuid)) { |
| 2170 | dev_select.has_uuid = 1; |
| 2171 | } else { |
| 2172 | dev_select.index = strtol(device, &end, 10); |
| 2173 | if (end == device || *end) { |
| 2174 | dev_select.index = 0; |
| 2175 | dev_select.name = device; |
| 2176 | } |
| 2177 | } |
| 2178 | } |
| 2179 | |
| 2180 | return vulkan_device_create_internal(ctx, &dev_select, 0, opts, flags); |
| 2181 | } |
| 2182 | |
| 2183 | static int vulkan_device_derive(AVHWDeviceContext *ctx, |
| 2184 | AVHWDeviceContext *src_ctx, |
| 2185 | AVDictionary *opts, int flags) |
| 2186 | { |
| 2187 | av_unused__attribute__((unused)) VulkanDeviceSelection dev_select = { 0 }; |
| 2188 | |
| 2189 | /* If there's only one device on the system, then even if its not covered |
| 2190 | * by the following checks (e.g. non-PCIe ARM GPU), having an empty |
| 2191 | * dev_select will mean it'll get picked. */ |
| 2192 | switch(src_ctx->type) { |
| 2193 | #if CONFIG_VAAPI1 |
| 2194 | case AV_HWDEVICE_TYPE_VAAPI: { |
| 2195 | AVVAAPIDeviceContext *src_hwctx = src_ctx->hwctx; |
| 2196 | VADisplay dpy = src_hwctx->display; |
| 2197 | #if VA_CHECK_VERSION(1, 15, 0)(1 > (1) || (1 == (1) && 13 > (15)) || (1 == (1 ) && 13 == (15) && 0 >= (0))) |
| 2198 | VAStatus vas; |
| 2199 | VADisplayAttribute attr = { |
| 2200 | .type = VADisplayPCIID, |
| 2201 | }; |
| 2202 | #endif |
| 2203 | const char *vendor; |
| 2204 | |
| 2205 | #if VA_CHECK_VERSION(1, 15, 0)(1 > (1) || (1 == (1) && 13 > (15)) || (1 == (1 ) && 13 == (15) && 0 >= (0))) |
| 2206 | vas = vaGetDisplayAttributes(dpy, &attr, 1); |
| 2207 | if (vas == VA_STATUS_SUCCESS0x00000000 && attr.flags != VA_DISPLAY_ATTRIB_NOT_SUPPORTED0x0000) |
| 2208 | dev_select.pci_device = (attr.value & 0xFFFF); |
| 2209 | #endif |
| 2210 | |
| 2211 | if (!dev_select.pci_device) { |
| 2212 | vendor = vaQueryVendorString(dpy); |
| 2213 | if (!vendor) { |
| 2214 | av_log(ctx, AV_LOG_ERROR16, "Unable to get device info from VAAPI!\n"); |
| 2215 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 2216 | } |
| 2217 | |
| 2218 | if (strstr(vendor, "AMD")) |
| 2219 | dev_select.vendor_id = 0x1002; |
| 2220 | } |
| 2221 | |
| 2222 | return vulkan_device_create_internal(ctx, &dev_select, 0, opts, flags); |
| 2223 | } |
| 2224 | #endif |
| 2225 | #if CONFIG_LIBDRM1 |
| 2226 | case AV_HWDEVICE_TYPE_DRM: { |
| 2227 | int err; |
| 2228 | struct stat drm_node_info; |
| 2229 | drmDevice *drm_dev_info; |
| 2230 | AVDRMDeviceContext *src_hwctx = src_ctx->hwctx; |
| 2231 | |
| 2232 | err = fstat(src_hwctx->fd, &drm_node_info); |
| 2233 | if (err) { |
| 2234 | av_log(ctx, AV_LOG_ERROR16, "Unable to get node info from DRM fd: %s!\n", |
| 2235 | av_err2str(AVERROR(errno))av_make_error_string((char[64]){0}, 64, (-((*__errno_location ()))))); |
| 2236 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 2237 | } |
| 2238 | |
| 2239 | dev_select.drm_major = major(drm_node_info.st_dev)gnu_dev_major (drm_node_info.st_dev); |
| 2240 | dev_select.drm_minor = minor(drm_node_info.st_dev)gnu_dev_minor (drm_node_info.st_dev); |
| 2241 | dev_select.has_drm = 1; |
| 2242 | |
| 2243 | err = drmGetDevice(src_hwctx->fd, &drm_dev_info); |
| 2244 | if (err) { |
| 2245 | av_log(ctx, AV_LOG_ERROR16, "Unable to get device info from DRM fd: %s!\n", |
| 2246 | av_err2str(AVERROR(errno))av_make_error_string((char[64]){0}, 64, (-((*__errno_location ()))))); |
| 2247 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 2248 | } |
| 2249 | |
| 2250 | if (drm_dev_info->bustype == DRM_BUS_PCI0) |
| 2251 | dev_select.pci_device = drm_dev_info->deviceinfo.pci->device_id; |
| 2252 | |
| 2253 | drmFreeDevice(&drm_dev_info); |
| 2254 | |
| 2255 | return vulkan_device_create_internal(ctx, &dev_select, 0, opts, flags); |
| 2256 | } |
| 2257 | #endif |
| 2258 | #if CONFIG_CUDA0 |
| 2259 | case AV_HWDEVICE_TYPE_CUDA: { |
| 2260 | AVHWDeviceContext *cuda_cu = src_ctx; |
| 2261 | AVCUDADeviceContext *src_hwctx = src_ctx->hwctx; |
| 2262 | AVCUDADeviceContextInternal *cu_internal = src_hwctx->internal; |
| 2263 | CudaFunctions *cu = cu_internal->cuda_dl; |
| 2264 | |
| 2265 | int ret = CHECK_CU(cu->cuDeviceGetUuid((CUuuid *)&dev_select.uuid, |
| 2266 | cu_internal->cuda_device)); |
| 2267 | if (ret < 0) { |
| 2268 | av_log(ctx, AV_LOG_ERROR16, "Unable to get UUID from CUDA!\n"); |
| 2269 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 2270 | } |
| 2271 | |
| 2272 | dev_select.has_uuid = 1; |
| 2273 | |
| 2274 | /* |
| 2275 | * CUDA is not able to import multiplane images, so always derive a |
| 2276 | * Vulkan device with multiplane disabled. |
| 2277 | */ |
| 2278 | return vulkan_device_create_internal(ctx, &dev_select, 1, opts, flags); |
| 2279 | } |
| 2280 | #endif |
| 2281 | default: |
| 2282 | return AVERROR(ENOSYS)(-(38)); |
| 2283 | } |
| 2284 | } |
| 2285 | |
| 2286 | static int vulkan_frames_get_constraints(AVHWDeviceContext *ctx, |
| 2287 | const void *hwconfig, |
| 2288 | AVHWFramesConstraints *constraints) |
| 2289 | { |
| 2290 | int count = 0; |
| 2291 | VulkanDevicePriv *p = ctx->hwctx; |
| 2292 | |
| 2293 | for (enum AVPixelFormat i = 0; i < nb_vk_formats_list; i++) { |
| 2294 | count += vkfmt_from_pixfmt2(ctx, vk_formats_list[i].pixfmt, |
| 2295 | p->use_linear_images ? VK_IMAGE_TILING_LINEAR : |
| 2296 | VK_IMAGE_TILING_OPTIMAL, |
| 2297 | NULL((void*)0), NULL((void*)0), NULL((void*)0), NULL((void*)0), p->disable_multiplane, 1) >= 0; |
| 2298 | } |
| 2299 | |
| 2300 | constraints->valid_sw_formats = av_malloc_array(count + 1 + CONFIG_CUDA0, |
| 2301 | sizeof(enum AVPixelFormat)); |
| 2302 | if (!constraints->valid_sw_formats) |
| 2303 | return AVERROR(ENOMEM)(-(12)); |
| 2304 | |
| 2305 | count = 0; |
| 2306 | for (enum AVPixelFormat i = 0; i < nb_vk_formats_list; i++) { |
| 2307 | if (vkfmt_from_pixfmt2(ctx, vk_formats_list[i].pixfmt, |
| 2308 | p->use_linear_images ? VK_IMAGE_TILING_LINEAR : |
| 2309 | VK_IMAGE_TILING_OPTIMAL, |
| 2310 | NULL((void*)0), NULL((void*)0), NULL((void*)0), NULL((void*)0), p->disable_multiplane, 1) >= 0) { |
| 2311 | constraints->valid_sw_formats[count++] = vk_formats_list[i].pixfmt; |
| 2312 | } |
| 2313 | } |
| 2314 | |
| 2315 | #if CONFIG_CUDA0 |
| 2316 | constraints->valid_sw_formats[count++] = AV_PIX_FMT_CUDA; |
| 2317 | #endif |
| 2318 | |
| 2319 | constraints->valid_sw_formats[count++] = AV_PIX_FMT_NONE; |
| 2320 | |
| 2321 | constraints->min_width = 1; |
| 2322 | constraints->min_height = 1; |
| 2323 | constraints->max_width = p->props.properties.limits.maxImageDimension2D; |
| 2324 | constraints->max_height = p->props.properties.limits.maxImageDimension2D; |
| 2325 | |
| 2326 | constraints->valid_hw_formats = av_malloc_array(2, sizeof(enum AVPixelFormat)); |
| 2327 | if (!constraints->valid_hw_formats) |
| 2328 | return AVERROR(ENOMEM)(-(12)); |
| 2329 | |
| 2330 | constraints->valid_hw_formats[0] = AV_PIX_FMT_VULKAN; |
| 2331 | constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE; |
| 2332 | |
| 2333 | return 0; |
| 2334 | } |
| 2335 | |
| 2336 | static int alloc_mem(AVHWDeviceContext *ctx, VkMemoryRequirements *req, |
| 2337 | VkMemoryPropertyFlagBits req_flags, const void *alloc_extension, |
| 2338 | VkMemoryPropertyFlagBits *mem_flags, VkDeviceMemory *mem) |
| 2339 | { |
| 2340 | VkResult ret; |
| 2341 | int index = -1; |
| 2342 | VulkanDevicePriv *p = ctx->hwctx; |
| 2343 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 2344 | AVVulkanDeviceContext *dev_hwctx = &p->p; |
| 2345 | VkMemoryAllocateInfo alloc_info = { |
| 2346 | .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO, |
| 2347 | .pNext = alloc_extension, |
| 2348 | .allocationSize = req->size, |
| 2349 | }; |
| 2350 | |
| 2351 | /* The vulkan spec requires memory types to be sorted in the "optimal" |
| 2352 | * order, so the first matching type we find will be the best/fastest one */ |
| 2353 | for (int i = 0; i < p->mprops.memoryTypeCount; i++) { |
| 2354 | const VkMemoryType *type = &p->mprops.memoryTypes[i]; |
| 2355 | |
| 2356 | /* The memory type must be supported by the requirements (bitfield) */ |
| 2357 | if (!(req->memoryTypeBits & (1 << i))) |
| 2358 | continue; |
| 2359 | |
| 2360 | /* The memory type flags must include our properties */ |
| 2361 | if ((type->propertyFlags & req_flags) != req_flags) |
| 2362 | continue; |
| 2363 | |
| 2364 | /* The memory type must be large enough */ |
| 2365 | if (req->size > p->mprops.memoryHeaps[type->heapIndex].size) |
| 2366 | continue; |
| 2367 | |
| 2368 | /* Found a suitable memory type */ |
| 2369 | index = i; |
| 2370 | break; |
| 2371 | } |
| 2372 | |
| 2373 | if (index < 0) { |
| 2374 | av_log(ctx, AV_LOG_ERROR16, "No memory type found for flags 0x%x\n", |
| 2375 | req_flags); |
| 2376 | return AVERROR(EINVAL)(-(22)); |
| 2377 | } |
| 2378 | |
| 2379 | alloc_info.memoryTypeIndex = index; |
| 2380 | |
| 2381 | ret = vk->AllocateMemory(dev_hwctx->act_dev, &alloc_info, |
| 2382 | dev_hwctx->alloc, mem); |
| 2383 | if (ret != VK_SUCCESS) { |
| 2384 | av_log(ctx, AV_LOG_ERROR16, "Failed to allocate memory: %s\n", |
| 2385 | ff_vk_ret2str(ret)); |
| 2386 | return AVERROR(ENOMEM)(-(12)); |
| 2387 | } |
| 2388 | |
| 2389 | *mem_flags |= p->mprops.memoryTypes[index].propertyFlags; |
| 2390 | |
| 2391 | return 0; |
| 2392 | } |
| 2393 | |
| 2394 | static void vulkan_free_internal(VulkanDevicePriv *p, AVVkFrame *f) |
| 2395 | { |
| 2396 | av_unused__attribute__((unused)) AVVkFrameInternal *internal = f->internal; |
| 2397 | |
| 2398 | // Make this function safe to call repeatedly |
| 2399 | if (!internal) |
| 2400 | return; |
| 2401 | |
| 2402 | #if CONFIG_CUDA0 |
| 2403 | if (internal->cuda_fc_ref) { |
| 2404 | AVHWFramesContext *cuda_fc = (AVHWFramesContext *)internal->cuda_fc_ref->data; |
| 2405 | int planes = av_pix_fmt_count_planes(cuda_fc->sw_format); |
| 2406 | AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx; |
| 2407 | AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx; |
| 2408 | AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal; |
| 2409 | CudaFunctions *cu = cu_internal->cuda_dl; |
| 2410 | |
| 2411 | for (int i = 0; i < planes; i++) { |
| 2412 | if (internal->cu_sem[i]) |
| 2413 | CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem[i])); |
| 2414 | if (internal->cu_mma[i]) |
| 2415 | CHECK_CU(cu->cuMipmappedArrayDestroy(internal->cu_mma[i])); |
| 2416 | if (internal->ext_mem[i]) |
| 2417 | CHECK_CU(cu->cuDestroyExternalMemory(internal->ext_mem[i])); |
| 2418 | #ifdef _WIN32 |
| 2419 | if (internal->ext_sem_handle[i]) |
| 2420 | CloseHandle(internal->ext_sem_handle[i]); |
| 2421 | if (internal->ext_mem_handle[i]) |
| 2422 | CloseHandle(internal->ext_mem_handle[i]); |
| 2423 | #endif |
| 2424 | } |
| 2425 | |
| 2426 | av_buffer_unref(&internal->cuda_fc_ref); |
| 2427 | } |
| 2428 | #endif |
| 2429 | |
| 2430 | if (internal->drm_sync_sem != VK_NULL_HANDLE((void*)0)) |
| 2431 | p->vkctx.vkfn.DestroySemaphore(p->p.act_dev, internal->drm_sync_sem, |
| 2432 | p->p.alloc); |
| 2433 | |
| 2434 | pthread_mutex_destroystrict_pthread_mutex_destroy(&internal->update_mutex); |
| 2435 | av_freep(&f->internal); |
| 2436 | } |
| 2437 | |
| 2438 | static void vulkan_frame_free(AVHWFramesContext *hwfc, AVVkFrame *f) |
| 2439 | { |
| 2440 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 2441 | AVVulkanDeviceContext *hwctx = &p->p; |
| 2442 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 2443 | int nb_images = ff_vk_count_images(f); |
| 2444 | int nb_sems = 0; |
| 2445 | |
| 2446 | while (nb_sems < FF_ARRAY_ELEMS(f->sem)(sizeof(f->sem) / sizeof((f->sem)[0])) && f->sem[nb_sems]) |
| 2447 | nb_sems++; |
| 2448 | |
| 2449 | if (nb_sems) { |
| 2450 | VkSemaphoreWaitInfo sem_wait = { |
| 2451 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_WAIT_INFO, |
| 2452 | .flags = 0x0, |
| 2453 | .pSemaphores = f->sem, |
| 2454 | .pValues = f->sem_value, |
| 2455 | .semaphoreCount = nb_sems, |
| 2456 | }; |
| 2457 | |
| 2458 | vk->WaitSemaphores(hwctx->act_dev, &sem_wait, UINT64_MAX(18446744073709551615UL)); |
| 2459 | } |
| 2460 | |
| 2461 | vulkan_free_internal(p, f); |
| 2462 | |
| 2463 | for (int i = 0; i < nb_images; i++) { |
| 2464 | vk->DestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc); |
| 2465 | vk->FreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc); |
| 2466 | vk->DestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc); |
| 2467 | } |
| 2468 | |
| 2469 | av_free(f); |
| 2470 | } |
| 2471 | |
| 2472 | static void vulkan_frame_free_cb(void *opaque, uint8_t *data) |
| 2473 | { |
| 2474 | vulkan_frame_free(opaque, (AVVkFrame*)data); |
| 2475 | } |
| 2476 | |
| 2477 | static int alloc_bind_mem(AVHWFramesContext *hwfc, AVVkFrame *f, |
| 2478 | void *alloc_pnext, size_t alloc_pnext_stride) |
| 2479 | { |
| 2480 | int img_cnt = 0, err; |
| 2481 | VkResult ret; |
| 2482 | AVHWDeviceContext *ctx = hwfc->device_ctx; |
| 2483 | VulkanDevicePriv *p = ctx->hwctx; |
| 2484 | AVVulkanDeviceContext *hwctx = &p->p; |
| 2485 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 2486 | VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS8] = { { 0 } }; |
| 2487 | |
| 2488 | while (f->img[img_cnt]) { |
| 2489 | int use_ded_mem; |
| 2490 | VkImageMemoryRequirementsInfo2 req_desc = { |
| 2491 | .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2, |
| 2492 | .image = f->img[img_cnt], |
| 2493 | }; |
| 2494 | VkMemoryDedicatedAllocateInfo ded_alloc = { |
| 2495 | .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO, |
| 2496 | .pNext = (void *)(((uint8_t *)alloc_pnext) + img_cnt*alloc_pnext_stride), |
| 2497 | }; |
| 2498 | VkMemoryDedicatedRequirements ded_req = { |
| 2499 | .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS, |
| 2500 | }; |
| 2501 | VkMemoryRequirements2 req = { |
| 2502 | .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2, |
| 2503 | .pNext = &ded_req, |
| 2504 | }; |
| 2505 | |
| 2506 | vk->GetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req); |
| 2507 | |
| 2508 | av_log(hwfc, AV_LOG_TRACE56, |
| 2509 | "plane %d: driver reports prefersDedicatedAllocation=%i requiresDedicatedAllocation=%i\n", |
| 2510 | img_cnt, ded_req.prefersDedicatedAllocation, ded_req.requiresDedicatedAllocation); |
| 2511 | |
| 2512 | if (f->tiling == VK_IMAGE_TILING_LINEAR) |
| 2513 | req.memoryRequirements.size = FFALIGN(req.memoryRequirements.size,(((req.memoryRequirements.size)+(p->props.properties.limits .minMemoryMapAlignment)-1)&~((p->props.properties.limits .minMemoryMapAlignment)-1)) |
| 2514 | p->props.properties.limits.minMemoryMapAlignment)(((req.memoryRequirements.size)+(p->props.properties.limits .minMemoryMapAlignment)-1)&~((p->props.properties.limits .minMemoryMapAlignment)-1)); |
| 2515 | |
| 2516 | /* In case the implementation prefers/requires dedicated allocation */ |
| 2517 | use_ded_mem = ded_req.prefersDedicatedAllocation | |
| 2518 | ded_req.requiresDedicatedAllocation; |
| 2519 | if (((VulkanFramesPriv *)hwfc->hwctx)->export_requires_dedicated) |
| 2520 | use_ded_mem = 1; |
| 2521 | if (use_ded_mem) |
| 2522 | ded_alloc.image = f->img[img_cnt]; |
| 2523 | |
| 2524 | /* Allocate memory */ |
| 2525 | if ((err = alloc_mem(ctx, &req.memoryRequirements, |
| 2526 | f->tiling == VK_IMAGE_TILING_LINEAR ? |
| 2527 | VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT : |
| 2528 | VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, |
| 2529 | use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext, |
| 2530 | &f->flags, &f->mem[img_cnt]))) |
| 2531 | return err; |
| 2532 | |
| 2533 | f->size[img_cnt] = req.memoryRequirements.size; |
| 2534 | bind_info[img_cnt].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO; |
| 2535 | bind_info[img_cnt].image = f->img[img_cnt]; |
| 2536 | bind_info[img_cnt].memory = f->mem[img_cnt]; |
| 2537 | |
| 2538 | img_cnt++; |
| 2539 | } |
| 2540 | |
| 2541 | /* Bind the allocated memory to the images */ |
| 2542 | ret = vk->BindImageMemory2(hwctx->act_dev, img_cnt, bind_info); |
| 2543 | if (ret != VK_SUCCESS) { |
| 2544 | av_log(ctx, AV_LOG_ERROR16, "Failed to bind memory: %s\n", |
| 2545 | ff_vk_ret2str(ret)); |
| 2546 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 2547 | } |
| 2548 | |
| 2549 | return 0; |
| 2550 | } |
| 2551 | |
| 2552 | enum PrepMode { |
| 2553 | PREP_MODE_GENERAL, |
| 2554 | PREP_MODE_WRITE, |
| 2555 | PREP_MODE_EXTERNAL_EXPORT, |
| 2556 | PREP_MODE_EXTERNAL_IMPORT, |
| 2557 | PREP_MODE_DECODING_DST, |
| 2558 | PREP_MODE_DECODING_DPB, |
| 2559 | PREP_MODE_ENCODING_DPB, |
| 2560 | }; |
| 2561 | |
| 2562 | static void switch_new_props(enum PrepMode pmode, VkImageLayout *new_layout, |
| 2563 | VkAccessFlags2 *new_access) |
| 2564 | { |
| 2565 | switch (pmode) { |
| 2566 | case PREP_MODE_GENERAL: |
| 2567 | *new_layout = VK_IMAGE_LAYOUT_GENERAL; |
| 2568 | *new_access = VK_ACCESS_TRANSFER_WRITE_BIT; |
| 2569 | break; |
| 2570 | case PREP_MODE_WRITE: |
| 2571 | *new_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL; |
| 2572 | *new_access = VK_ACCESS_TRANSFER_WRITE_BIT; |
| 2573 | break; |
| 2574 | case PREP_MODE_EXTERNAL_IMPORT: |
| 2575 | *new_layout = VK_IMAGE_LAYOUT_GENERAL; |
| 2576 | *new_access = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT; |
| 2577 | break; |
| 2578 | case PREP_MODE_EXTERNAL_EXPORT: |
| 2579 | *new_layout = VK_IMAGE_LAYOUT_GENERAL; |
| 2580 | *new_access = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT; |
| 2581 | break; |
| 2582 | case PREP_MODE_DECODING_DST: |
| 2583 | *new_layout = VK_IMAGE_LAYOUT_VIDEO_DECODE_DST_KHR; |
| 2584 | *new_access = VK_ACCESS_TRANSFER_WRITE_BIT; |
| 2585 | break; |
| 2586 | case PREP_MODE_DECODING_DPB: |
| 2587 | *new_layout = VK_IMAGE_LAYOUT_VIDEO_DECODE_DPB_KHR; |
| 2588 | *new_access = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT; |
| 2589 | break; |
| 2590 | case PREP_MODE_ENCODING_DPB: |
| 2591 | *new_layout = VK_IMAGE_LAYOUT_VIDEO_ENCODE_DPB_KHR; |
| 2592 | *new_access = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT; |
| 2593 | break; |
| 2594 | } |
| 2595 | } |
| 2596 | |
| 2597 | static int switch_layout(AVHWFramesContext *hwfc, FFVkExecPool *ectx, |
| 2598 | AVVkFrame *frame, enum PrepMode pmode) |
| 2599 | { |
| 2600 | int err; |
| 2601 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 2602 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 2603 | VkImageMemoryBarrier2 img_bar[AV_NUM_DATA_POINTERS8]; |
| 2604 | int nb_img_bar = 0; |
| 2605 | |
| 2606 | VkImageLayout new_layout; |
| 2607 | VkAccessFlags2 new_access; |
| 2608 | switch_new_props(pmode, &new_layout, &new_access); |
| 2609 | |
| 2610 | uint32_t dst_qf = p->nb_img_qfs > 1 ? VK_QUEUE_FAMILY_IGNORED(~0U) : p->img_qfs[0]; |
| 2611 | VkPipelineStageFlagBits2 src_stage = VK_PIPELINE_STAGE_2_NONE; |
| 2612 | if (pmode == PREP_MODE_EXTERNAL_EXPORT) { |
| 2613 | dst_qf = VK_QUEUE_FAMILY_EXTERNAL_KHR(~1U); |
| 2614 | src_stage = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT; |
| 2615 | } |
| 2616 | |
| 2617 | /* This is dirty - but it works. The vulkan.c dependency system doesn't |
| 2618 | * free non-refcounted frames, and non-refcounted hardware frames cannot |
| 2619 | * happen anywhere outside of here. */ |
| 2620 | AVBufferRef tmp_ref = { |
| 2621 | .data = (uint8_t *)hwfc, |
| 2622 | }; |
| 2623 | AVFrame tmp_frame = { |
| 2624 | .data[0] = (uint8_t *)frame, |
| 2625 | .hw_frames_ctx = &tmp_ref, |
| 2626 | }; |
| 2627 | |
| 2628 | VkCommandBuffer cmd_buf; |
| 2629 | FFVkExecContext *exec = ff_vk_exec_get(&p->vkctx, ectx); |
| 2630 | cmd_buf = exec->buf; |
| 2631 | err = ff_vk_exec_start(&p->vkctx, exec); |
| 2632 | if (err < 0) |
| 2633 | return err; |
| 2634 | |
| 2635 | err = ff_vk_exec_add_dep_frame(&p->vkctx, exec, &tmp_frame, |
| 2636 | VK_PIPELINE_STAGE_2_NONE, |
| 2637 | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT); |
| 2638 | if (err < 0) |
| 2639 | return err; |
| 2640 | |
| 2641 | ff_vk_frame_barrier(&p->vkctx, exec, &tmp_frame, img_bar, &nb_img_bar, |
| 2642 | src_stage, |
| 2643 | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, |
| 2644 | new_access, new_layout, dst_qf); |
| 2645 | |
| 2646 | vk->CmdPipelineBarrier2(cmd_buf, &(VkDependencyInfo) { |
| 2647 | .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO, |
| 2648 | .pImageMemoryBarriers = img_bar, |
| 2649 | .imageMemoryBarrierCount = nb_img_bar, |
| 2650 | }); |
| 2651 | |
| 2652 | err = ff_vk_exec_submit(&p->vkctx, exec); |
| 2653 | if (err < 0) |
| 2654 | return err; |
| 2655 | |
| 2656 | /* Drops the stack-based frame above from the dependency list */ |
| 2657 | ff_vk_exec_wait(&p->vkctx, exec); |
| 2658 | |
| 2659 | return 0; |
| 2660 | } |
| 2661 | |
| 2662 | static int switch_layout_host(AVHWFramesContext *hwfc, FFVkExecPool *ectx, |
| 2663 | AVVkFrame *frame, enum PrepMode pmode) |
| 2664 | { |
| 2665 | VkResult ret; |
| 2666 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 2667 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 2668 | VkHostImageLayoutTransitionInfoEXT layout_change[AV_NUM_DATA_POINTERS8]; |
| 2669 | int nb_images = ff_vk_count_images(frame); |
| 2670 | |
| 2671 | VkImageLayout new_layout; |
| 2672 | VkAccessFlags2 new_access; |
| 2673 | switch_new_props(pmode, &new_layout, &new_access); |
| 2674 | |
| 2675 | int i; |
| 2676 | for (i = 0; i < p->vkctx.host_image_props.copyDstLayoutCount; i++) { |
| 2677 | if (p->vkctx.host_image_props.pCopyDstLayouts[i] == new_layout) |
| 2678 | break; |
| 2679 | } |
| 2680 | if (i == p->vkctx.host_image_props.copyDstLayoutCount) |
| 2681 | return AVERROR(ENOTSUP)(-(95)); |
| 2682 | |
| 2683 | for (i = 0; i < nb_images; i++) { |
| 2684 | layout_change[i] = (VkHostImageLayoutTransitionInfoEXT) { |
| 2685 | .sType = VK_STRUCTURE_TYPE_HOST_IMAGE_LAYOUT_TRANSITION_INFO_EXT, |
| 2686 | .image = frame->img[i], |
| 2687 | .oldLayout = frame->layout[i], |
| 2688 | .newLayout = new_layout, |
| 2689 | .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, |
| 2690 | .subresourceRange.layerCount = 1, |
| 2691 | .subresourceRange.levelCount = 1, |
| 2692 | }; |
| 2693 | frame->layout[i] = new_layout; |
| 2694 | } |
| 2695 | |
| 2696 | ret = vk->TransitionImageLayoutEXT(p->vkctx.hwctx->act_dev, |
| 2697 | nb_images, layout_change); |
| 2698 | if (ret != VK_SUCCESS) { |
| 2699 | av_log(hwfc, AV_LOG_ERROR16, "Unable to prepare frame: %s\n", |
| 2700 | ff_vk_ret2str(ret)); |
| 2701 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 2702 | } |
| 2703 | |
| 2704 | return 0; |
| 2705 | } |
| 2706 | |
| 2707 | static int prepare_frame(AVHWFramesContext *hwfc, FFVkExecPool *ectx, |
| 2708 | AVVkFrame *frame, enum PrepMode pmode) |
| 2709 | { |
| 2710 | int err = 0; |
| 2711 | AVVulkanFramesContext *hwfc_vk = hwfc->hwctx; |
| 2712 | if (hwfc_vk->usage & VK_IMAGE_USAGE_HOST_TRANSFER_BIT_EXT && |
| 2713 | (pmode != PREP_MODE_EXTERNAL_EXPORT) && |
| 2714 | (pmode != PREP_MODE_EXTERNAL_IMPORT)) { |
| 2715 | err = switch_layout_host(hwfc, ectx, frame, pmode); |
| 2716 | if (err != AVERROR(ENOTSUP)(-(95))) |
| 2717 | return err; |
| 2718 | } |
| 2719 | |
| 2720 | return switch_layout(hwfc, ectx, frame, pmode); |
| 2721 | } |
| 2722 | |
| 2723 | static inline void get_plane_wh(uint32_t *w, uint32_t *h, enum AVPixelFormat format, |
| 2724 | int frame_w, int frame_h, int plane) |
| 2725 | { |
| 2726 | const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format); |
| 2727 | |
| 2728 | /* Currently always true unless gray + alpha support is added */ |
| 2729 | if (!plane || (plane == 3) || desc->flags & AV_PIX_FMT_FLAG_RGB(1 << 5) || |
| 2730 | !(desc->flags & AV_PIX_FMT_FLAG_PLANAR(1 << 4))) { |
| 2731 | *w = frame_w; |
| 2732 | *h = frame_h; |
| 2733 | return; |
| 2734 | } |
| 2735 | |
| 2736 | *w = AV_CEIL_RSHIFT(frame_w, desc->log2_chroma_w)(!__builtin_constant_p(desc->log2_chroma_w) ? -((-(frame_w )) >> (desc->log2_chroma_w)) : ((frame_w) + (1<< (desc->log2_chroma_w)) - 1) >> (desc->log2_chroma_w )); |
| 2737 | *h = AV_CEIL_RSHIFT(frame_h, desc->log2_chroma_h)(!__builtin_constant_p(desc->log2_chroma_h) ? -((-(frame_h )) >> (desc->log2_chroma_h)) : ((frame_h) + (1<< (desc->log2_chroma_h)) - 1) >> (desc->log2_chroma_h )); |
| 2738 | } |
| 2739 | |
| 2740 | static int create_frame(AVHWFramesContext *hwfc, AVVkFrame **frame, |
| 2741 | VkImageTiling tiling, VkImageUsageFlagBits usage, |
| 2742 | VkImageCreateFlags flags, int nb_layers, |
| 2743 | void *create_pnext) |
| 2744 | { |
| 2745 | int err; |
| 2746 | VkResult ret; |
| 2747 | AVVulkanFramesContext *hwfc_vk = hwfc->hwctx; |
| 2748 | AVHWDeviceContext *ctx = hwfc->device_ctx; |
| 2749 | VulkanDevicePriv *p = ctx->hwctx; |
| 2750 | AVVulkanDeviceContext *hwctx = &p->p; |
| 2751 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 2752 | AVVkFrame *f; |
| 2753 | |
| 2754 | VkSemaphoreTypeCreateInfo sem_type_info = { |
| 2755 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_TYPE_CREATE_INFO, |
| 2756 | .semaphoreType = VK_SEMAPHORE_TYPE_TIMELINE, |
| 2757 | .initialValue = 0, |
| 2758 | }; |
| 2759 | VkSemaphoreCreateInfo sem_spawn = { |
| 2760 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO, |
| 2761 | .pNext = &sem_type_info, |
| 2762 | }; |
| 2763 | |
| 2764 | VkExportSemaphoreCreateInfo ext_sem_info_opaque = { |
| 2765 | .sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO, |
| 2766 | #ifdef _WIN32 |
| 2767 | .handleTypes = IsWindows8OrGreater() |
| 2768 | ? VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_BIT |
| 2769 | : VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT, |
| 2770 | #else |
| 2771 | .handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT, |
| 2772 | #endif |
| 2773 | }; |
| 2774 | |
| 2775 | /* Check if exporting is supported before chaining any structs */ |
| 2776 | if (p->ext_sem_props_opaque.externalSemaphoreFeatures & VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT) { |
| 2777 | if (p->vkctx.extensions & (FF_VK_EXT_EXTERNAL_WIN32_SEM(1ULL << 7) | FF_VK_EXT_EXTERNAL_FD_SEM(1ULL << 3))) |
| 2778 | ff_vk_link_struct(&sem_type_info, &ext_sem_info_opaque); |
| 2779 | } |
| 2780 | |
| 2781 | f = av_vk_frame_alloc(); |
| 2782 | if (!f) { |
| 2783 | av_log(ctx, AV_LOG_ERROR16, "Unable to allocate memory for AVVkFrame!\n"); |
| 2784 | return AVERROR(ENOMEM)(-(12)); |
| 2785 | } |
| 2786 | |
| 2787 | // TODO: check width and height for alignment in case of multiplanar (must be mod-2 if subsampled) |
| 2788 | |
| 2789 | /* Create the images */ |
| 2790 | for (int i = 0; (hwfc_vk->format[i] != VK_FORMAT_UNDEFINED); i++) { |
| 2791 | VkImageCreateInfo create_info = { |
| 2792 | .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO, |
| 2793 | .pNext = create_pnext, |
| 2794 | .imageType = VK_IMAGE_TYPE_2D, |
| 2795 | .format = hwfc_vk->format[i], |
| 2796 | .extent.depth = 1, |
| 2797 | .mipLevels = 1, |
| 2798 | .arrayLayers = nb_layers, |
| 2799 | .flags = flags, |
| 2800 | .tiling = tiling, |
| 2801 | .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, |
| 2802 | .usage = usage, |
| 2803 | .samples = VK_SAMPLE_COUNT_1_BIT, |
| 2804 | .pQueueFamilyIndices = p->img_qfs, |
| 2805 | .queueFamilyIndexCount = p->nb_img_qfs, |
| 2806 | .sharingMode = p->nb_img_qfs > 1 ? VK_SHARING_MODE_CONCURRENT : |
| 2807 | VK_SHARING_MODE_EXCLUSIVE, |
| 2808 | }; |
| 2809 | |
| 2810 | get_plane_wh(&create_info.extent.width, &create_info.extent.height, |
| 2811 | hwfc->sw_format, hwfc->width, hwfc->height, i); |
| 2812 | |
| 2813 | ret = vk->CreateImage(hwctx->act_dev, &create_info, |
| 2814 | hwctx->alloc, &f->img[i]); |
| 2815 | if (ret != VK_SUCCESS) { |
| 2816 | av_log(ctx, AV_LOG_ERROR16, "Image creation failure: %s\n", |
| 2817 | ff_vk_ret2str(ret)); |
| 2818 | err = AVERROR(EINVAL)(-(22)); |
| 2819 | goto fail; |
| 2820 | } |
| 2821 | |
| 2822 | /* Create semaphore */ |
| 2823 | ret = vk->CreateSemaphore(hwctx->act_dev, &sem_spawn, |
| 2824 | hwctx->alloc, &f->sem[i]); |
| 2825 | if (ret != VK_SUCCESS) { |
| 2826 | av_log(hwctx, AV_LOG_ERROR16, "Failed to create semaphore: %s\n", |
| 2827 | ff_vk_ret2str(ret)); |
| 2828 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 2829 | goto fail; |
| 2830 | } |
| 2831 | |
| 2832 | f->queue_family[i] = p->nb_img_qfs > 1 ? VK_QUEUE_FAMILY_IGNORED(~0U) : p->img_qfs[0]; |
| 2833 | f->layout[i] = create_info.initialLayout; |
| 2834 | f->access[i] = 0x0; |
| 2835 | f->sem_value[i] = 0; |
| 2836 | } |
| 2837 | |
| 2838 | f->flags = 0x0; |
| 2839 | f->tiling = tiling; |
| 2840 | |
| 2841 | *frame = f; |
| 2842 | return 0; |
| 2843 | |
| 2844 | fail: |
| 2845 | vulkan_frame_free(hwfc, f); |
| 2846 | return err; |
| 2847 | } |
| 2848 | |
| 2849 | /* Checks if an export flag is enabled, and if it is ORs it with *iexp */ |
| 2850 | static void try_export_flags(AVHWFramesContext *hwfc, |
| 2851 | VkExternalMemoryHandleTypeFlags *comp_handle_types, |
| 2852 | VkExternalMemoryHandleTypeFlags *iexp, |
| 2853 | VkExternalMemoryHandleTypeFlagBits exp) |
| 2854 | { |
| 2855 | VkResult ret; |
| 2856 | AVVulkanFramesContext *hwctx = hwfc->hwctx; |
| 2857 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 2858 | AVVulkanDeviceContext *dev_hwctx = &p->p; |
| 2859 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 2860 | |
| 2861 | const VkImageDrmFormatModifierListCreateInfoEXT *drm_mod_info = |
| 2862 | ff_vk_find_struct(hwctx->create_pnext, |
| 2863 | VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_LIST_CREATE_INFO_EXT); |
| 2864 | int has_mods = hwctx->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT && drm_mod_info; |
| 2865 | int nb_mods; |
| 2866 | |
| 2867 | VkExternalImageFormatProperties eprops = { |
| 2868 | .sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR, |
| 2869 | }; |
| 2870 | VkImageFormatProperties2 props = { |
| 2871 | .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2, |
| 2872 | .pNext = &eprops, |
| 2873 | }; |
| 2874 | VkPhysicalDeviceImageDrmFormatModifierInfoEXT phy_dev_mod_info = { |
| 2875 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_DRM_FORMAT_MODIFIER_INFO_EXT, |
| 2876 | .pNext = NULL((void*)0), |
| 2877 | .pQueueFamilyIndices = p->img_qfs, |
| 2878 | .queueFamilyIndexCount = p->nb_img_qfs, |
| 2879 | .sharingMode = p->nb_img_qfs > 1 ? VK_SHARING_MODE_CONCURRENT : |
| 2880 | VK_SHARING_MODE_EXCLUSIVE, |
| 2881 | }; |
| 2882 | VkPhysicalDeviceExternalImageFormatInfo enext = { |
| 2883 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO, |
| 2884 | .handleType = exp, |
| 2885 | .pNext = has_mods ? &phy_dev_mod_info : NULL((void*)0), |
| 2886 | }; |
| 2887 | VkPhysicalDeviceImageFormatInfo2 pinfo = { |
| 2888 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2, |
| 2889 | .pNext = !exp ? NULL((void*)0) : &enext, |
| 2890 | .format = hwctx->format[0], |
| 2891 | .type = VK_IMAGE_TYPE_2D, |
| 2892 | .tiling = hwctx->tiling, |
| 2893 | .usage = hwctx->usage, |
| 2894 | .flags = (hwctx->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT && has_mods) ? |
| 2895 | (hwctx->img_flags) : hwctx->img_flags ? hwctx->img_flags : (VkImageCreateFlags)(VK_IMAGE_CREATE_ALIAS_BIT), |
| 2896 | }; |
| 2897 | |
| 2898 | nb_mods = has_mods ? drm_mod_info->drmFormatModifierCount : 1; |
| 2899 | for (int i = 0; i < nb_mods; i++) { |
| 2900 | if (has_mods) |
| 2901 | phy_dev_mod_info.drmFormatModifier = drm_mod_info->pDrmFormatModifiers[i]; |
| 2902 | |
| 2903 | ret = vk->GetPhysicalDeviceImageFormatProperties2(dev_hwctx->phys_dev, |
| 2904 | &pinfo, &props); |
| 2905 | |
| 2906 | if (has_mods) |
| 2907 | av_log(hwfc, AV_LOG_VERBOSE40, "GetPhysicalDeviceImageFormatProperties2: mod[%d]=0x%llx -> %s\n", |
| 2908 | i, (unsigned long long)phy_dev_mod_info.drmFormatModifier, |
| 2909 | ret == VK_SUCCESS ? "OK" : "FAIL"); |
| 2910 | if (ret == VK_SUCCESS) { |
| 2911 | *iexp |= exp; |
| 2912 | *comp_handle_types |= eprops.externalMemoryProperties.compatibleHandleTypes; |
| 2913 | if (exp == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT) { |
| 2914 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 2915 | fp->export_requires_dedicated = !!(eprops.externalMemoryProperties.externalMemoryFeatures & |
| 2916 | VK_EXTERNAL_MEMORY_FEATURE_DEDICATED_ONLY_BIT); |
| 2917 | } |
| 2918 | } |
| 2919 | } |
| 2920 | } |
| 2921 | |
| 2922 | /* Computes the external memory handle types for the frame context. */ |
| 2923 | static void get_export_handle_types(AVHWFramesContext *hwfc, |
| 2924 | VkExternalMemoryHandleTypeFlags *comp_handle_types, |
| 2925 | VkExternalMemoryHandleTypeFlags *export_types) |
| 2926 | { |
| 2927 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 2928 | av_unused__attribute__((unused)) AVVulkanFramesContext *hwctx = &((VulkanFramesPriv *)hwfc->hwctx)->p; |
| 2929 | |
| 2930 | *comp_handle_types = 0x0; |
| 2931 | *export_types = 0x0; |
| 2932 | |
| 2933 | #ifdef _WIN32 |
| 2934 | if (p->vkctx.extensions & FF_VK_EXT_EXTERNAL_WIN32_MEMORY(1ULL << 6)) |
| 2935 | try_export_flags(hwfc, comp_handle_types, export_types, IsWindows8OrGreater() |
| 2936 | ? VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT |
| 2937 | : VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT); |
| 2938 | #else |
| 2939 | if ((p->vkctx.extensions & FF_VK_EXT_EXTERNAL_FD_MEMORY(1ULL << 2)) && |
| 2940 | (hwctx->tiling != VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT)) |
| 2941 | try_export_flags(hwfc, comp_handle_types, export_types, |
| 2942 | VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT); |
| 2943 | |
| 2944 | if (p->vkctx.extensions & FF_VK_EXT_EXTERNAL_DMABUF_MEMORY(1ULL << 0) && |
| 2945 | hwctx->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT) |
| 2946 | try_export_flags(hwfc, comp_handle_types, export_types, |
| 2947 | VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT); |
| 2948 | #endif |
| 2949 | } |
| 2950 | |
| 2951 | static AVBufferRef *vulkan_pool_alloc(void *opaque, size_t size) |
| 2952 | { |
| 2953 | int err; |
| 2954 | AVVkFrame *f; |
| 2955 | AVBufferRef *avbuf = NULL((void*)0); |
| 2956 | AVHWFramesContext *hwfc = opaque; |
| 2957 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 2958 | AVVulkanFramesContext *hwctx = &fp->p; |
| 2959 | VkExternalMemoryHandleTypeFlags e; |
| 2960 | VkExportMemoryAllocateInfo eminfo[AV_NUM_DATA_POINTERS8]; |
| 2961 | |
| 2962 | VkExternalMemoryImageCreateInfo eiinfo = { |
| 2963 | .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO, |
| 2964 | .pNext = hwctx->create_pnext, |
| 2965 | }; |
| 2966 | |
| 2967 | get_export_handle_types(hwfc, &eiinfo.handleTypes, &e); |
| 2968 | |
| 2969 | for (int i = 0; i < av_pix_fmt_count_planes(hwfc->sw_format); i++) { |
| 2970 | eminfo[i].sType = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO; |
| 2971 | eminfo[i].pNext = hwctx->alloc_pnext[i]; |
| 2972 | eminfo[i].handleTypes = e; |
| 2973 | } |
| 2974 | |
| 2975 | err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage, hwctx->img_flags, |
| 2976 | hwctx->nb_layers, |
| 2977 | eiinfo.handleTypes ? &eiinfo : hwctx->create_pnext); |
| 2978 | if (err) { |
| 2979 | av_log(hwfc, AV_LOG_ERROR16, "vulkan_pool_alloc failed: create_frame failed: %d\n", err); |
| 2980 | return NULL((void*)0); |
| 2981 | } |
| 2982 | |
| 2983 | err = alloc_bind_mem(hwfc, f, eminfo, sizeof(*eminfo)); |
| 2984 | if (err) |
| 2985 | goto fail; |
| 2986 | |
| 2987 | if ( (hwctx->usage & VK_IMAGE_USAGE_VIDEO_DECODE_DPB_BIT_KHR) && |
| 2988 | !(hwctx->usage & VK_IMAGE_USAGE_VIDEO_DECODE_DST_BIT_KHR)) |
| 2989 | err = prepare_frame(hwfc, &fp->compute_exec, f, PREP_MODE_DECODING_DPB); |
| 2990 | else if (hwctx->usage & VK_IMAGE_USAGE_VIDEO_DECODE_DST_BIT_KHR) |
| 2991 | err = prepare_frame(hwfc, &fp->compute_exec, f, PREP_MODE_DECODING_DST); |
| 2992 | else if (hwctx->usage & VK_IMAGE_USAGE_VIDEO_ENCODE_DPB_BIT_KHR) |
| 2993 | err = prepare_frame(hwfc, &fp->compute_exec, f, PREP_MODE_ENCODING_DPB); |
| 2994 | else if (hwctx->usage & VK_IMAGE_USAGE_TRANSFER_DST_BIT) |
| 2995 | err = prepare_frame(hwfc, &fp->compute_exec, f, PREP_MODE_WRITE); |
| 2996 | else |
| 2997 | err = prepare_frame(hwfc, &fp->compute_exec, f, PREP_MODE_GENERAL); |
| 2998 | if (err) |
| 2999 | goto fail; |
| 3000 | |
| 3001 | avbuf = av_buffer_create((uint8_t *)f, sizeof(AVVkFrame), |
| 3002 | vulkan_frame_free_cb, hwfc, 0); |
| 3003 | if (!avbuf) |
| 3004 | goto fail; |
| 3005 | |
| 3006 | return avbuf; |
| 3007 | |
| 3008 | fail: |
| 3009 | av_log(hwfc, AV_LOG_ERROR16, "vulkan_pool_alloc failed with error %d\n", err); |
| 3010 | vulkan_frame_free(hwfc, f); |
| 3011 | return NULL((void*)0); |
| 3012 | } |
| 3013 | |
| 3014 | static void lock_frame(AVHWFramesContext *fc, AVVkFrame *vkf) |
| 3015 | { |
| 3016 | pthread_mutex_lockstrict_pthread_mutex_lock(&vkf->internal->update_mutex); |
| 3017 | } |
| 3018 | |
| 3019 | static void unlock_frame(AVHWFramesContext *fc, AVVkFrame *vkf) |
| 3020 | { |
| 3021 | pthread_mutex_unlockstrict_pthread_mutex_unlock(&vkf->internal->update_mutex); |
| 3022 | } |
| 3023 | |
| 3024 | static void vulkan_frames_uninit(AVHWFramesContext *hwfc) |
| 3025 | { |
| 3026 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 3027 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 3028 | |
| 3029 | if (fp->modifier_info) { |
| 3030 | if (fp->modifier_info->pDrmFormatModifiers) |
| 3031 | av_freep(&fp->modifier_info->pDrmFormatModifiers); |
| 3032 | av_freep(&fp->modifier_info); |
| 3033 | } |
| 3034 | |
| 3035 | ff_vk_exec_pool_free(&p->vkctx, &fp->compute_exec); |
| 3036 | ff_vk_exec_pool_free(&p->vkctx, &fp->upload_exec); |
| 3037 | ff_vk_exec_pool_free(&p->vkctx, &fp->download_exec); |
| 3038 | |
| 3039 | av_refstruct_pool_uninit(&fp->tmp); |
| 3040 | } |
| 3041 | |
| 3042 | /* Probes whether the host transfer usage bit is actually usable in combination |
| 3043 | * with the rest of the image creation parameters. Drivers may expose the |
| 3044 | * format feature, yet reject the final image, or provide no memory type from |
| 3045 | * which such an image can be allocated. Notably video decode/encode images |
| 3046 | * are often not possible to allocate from the host visible memory type. */ |
| 3047 | static int vulkan_host_transfer_usable(AVHWFramesContext *hwfc) |
| 3048 | { |
| 3049 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 3050 | AVVulkanFramesContext *hwctx = &fp->p; |
| 3051 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 3052 | AVVulkanDeviceContext *dev_hwctx = &p->p; |
| 3053 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 3054 | VkResult ret; |
| 3055 | |
| 3056 | /* NVIDIA's host image copy resolves every plane's parameters as if it were |
| 3057 | * plane 0, so every plane but the first is corrupted on both upload and |
| 3058 | * download. Frames with one image per plane are unaffected. |
| 3059 | */ |
| 3060 | const struct FFVkFormatEntry *fmt = vk_find_format_entry(hwfc->sw_format); |
| 3061 | if (p->dprops.driverID == VK_DRIVER_ID_NVIDIA_PROPRIETARY && |
| 3062 | fmt && fmt->vk_planes > 1 && hwctx->format[0] == fmt->vkf) { |
| 3063 | av_log(hwfc, AV_LOG_VERBOSE40, "Disabling host image transfers: " |
| 3064 | "NVIDIA drivers mishandle multi-plane images\n"); |
| 3065 | return 0; |
| 3066 | } |
| 3067 | |
| 3068 | const VkImageDrmFormatModifierListCreateInfoEXT *mod_list = |
| 3069 | ff_vk_find_struct(hwctx->create_pnext, |
| 3070 | VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_LIST_CREATE_INFO_EXT); |
| 3071 | int has_mods = hwctx->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT && |
| 3072 | mod_list && mod_list->drmFormatModifierCount; |
| 3073 | int nb_mods = has_mods ? mod_list->drmFormatModifierCount : 1; |
| 3074 | |
| 3075 | /* Without a modifier list the final modifier is not knowable here. */ |
| 3076 | if (hwctx->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT && !has_mods) |
| 3077 | return 0; |
| 3078 | |
| 3079 | VkPhysicalDeviceImageDrmFormatModifierInfoEXT mod_info = { |
| 3080 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_DRM_FORMAT_MODIFIER_INFO_EXT, |
| 3081 | .pQueueFamilyIndices = p->img_qfs, |
| 3082 | .queueFamilyIndexCount = p->nb_img_qfs, |
| 3083 | .sharingMode = p->nb_img_qfs > 1 ? VK_SHARING_MODE_CONCURRENT : |
| 3084 | VK_SHARING_MODE_EXCLUSIVE, |
| 3085 | }; |
| 3086 | VkPhysicalDeviceImageFormatInfo2 pinfo = { |
| 3087 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2, |
| 3088 | .type = VK_IMAGE_TYPE_2D, |
| 3089 | .tiling = hwctx->tiling, |
| 3090 | .usage = hwctx->usage, |
| 3091 | .flags = hwctx->img_flags, |
| 3092 | }; |
| 3093 | |
| 3094 | if (has_mods) |
| 3095 | ff_vk_link_struct(&pinfo, &mod_info); |
| 3096 | |
| 3097 | VkVideoProfileListInfoKHR profile_list; |
| 3098 | const VkVideoProfileListInfoKHR *pl = |
| 3099 | ff_vk_find_struct(hwctx->create_pnext, |
| 3100 | VK_STRUCTURE_TYPE_VIDEO_PROFILE_LIST_INFO_KHR); |
| 3101 | if (pl) { |
| 3102 | profile_list = *pl; |
| 3103 | profile_list.pNext = NULL((void*)0); |
| 3104 | ff_vk_link_struct(&pinfo, &profile_list); |
| 3105 | } |
| 3106 | |
| 3107 | VkImageFormatListCreateInfo format_list; |
| 3108 | const VkImageFormatListCreateInfo *fl = |
| 3109 | ff_vk_find_struct(hwctx->create_pnext, |
| 3110 | VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO); |
| 3111 | if (fl) { |
| 3112 | format_list = *fl; |
| 3113 | format_list.pNext = NULL((void*)0); |
| 3114 | ff_vk_link_struct(&pinfo, &format_list); |
| 3115 | } |
| 3116 | |
| 3117 | /* The same usage is applied to the images of every plane, so each plane |
| 3118 | * format has to support host transfers for the usage to be usable. */ |
| 3119 | for (int i = 0; i < AV_NUM_DATA_POINTERS8 && |
| 3120 | hwctx->format[i] != VK_FORMAT_UNDEFINED; i++) { |
| 3121 | pinfo.format = hwctx->format[i]; |
| 3122 | |
| 3123 | /* The driver is free to pick any modifier from the list, so all of |
| 3124 | * them have to be compatible. */ |
| 3125 | for (int j = 0; j < nb_mods; j++) { |
| 3126 | VkHostImageCopyDevicePerformanceQueryEXT perf = { |
| 3127 | .sType = VK_STRUCTURE_TYPE_HOST_IMAGE_COPY_DEVICE_PERFORMANCE_QUERY_EXT, |
| 3128 | }; |
| 3129 | VkImageFormatProperties2 props = { |
| 3130 | .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2, |
| 3131 | .pNext = &perf, |
| 3132 | }; |
| 3133 | |
| 3134 | if (has_mods) |
| 3135 | mod_info.drmFormatModifier = mod_list->pDrmFormatModifiers[j]; |
| 3136 | |
| 3137 | ret = vk->GetPhysicalDeviceImageFormatProperties2(dev_hwctx->phys_dev, |
| 3138 | &pinfo, &props); |
| 3139 | if (ret != VK_SUCCESS) { |
| 3140 | av_log(hwfc, AV_LOG_VERBOSE40, "Disabling host image transfers: " |
| 3141 | "format %i is not supported: %s\n", |
| 3142 | pinfo.format, ff_vk_ret2str(ret)); |
| 3143 | return 0; |
| 3144 | } |
| 3145 | |
| 3146 | if (!perf.optimalDeviceAccess) { |
| 3147 | av_log(hwfc, AV_LOG_VERBOSE40, "Disabling host image transfers: " |
| 3148 | "format %i has no optimal device access (identical " |
| 3149 | "memory layout: %i)\n", |
| 3150 | pinfo.format, perf.identicalMemoryLayout); |
| 3151 | return 0; |
| 3152 | } |
| 3153 | } |
| 3154 | } |
| 3155 | |
| 3156 | if (!p->vkctx.host_image_props.identicalMemoryTypeRequirements) { |
| 3157 | int index; |
| 3158 | VkExternalMemoryHandleTypeFlags e; |
| 3159 | VkExternalMemoryImageCreateInfo eiinfo = { |
| 3160 | .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO, |
| 3161 | .pNext = hwctx->create_pnext, |
| 3162 | }; |
| 3163 | get_export_handle_types(hwfc, &eiinfo.handleTypes, &e); |
| 3164 | VkImageCreateInfo create_info = { |
| 3165 | .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO, |
| 3166 | .pNext = eiinfo.handleTypes ? &eiinfo : hwctx->create_pnext, |
| 3167 | .imageType = VK_IMAGE_TYPE_2D, |
| 3168 | .format = hwctx->format[0], |
| 3169 | .extent = { hwfc->width, hwfc->height, 1 }, |
| 3170 | .mipLevels = 1, |
| 3171 | .arrayLayers = hwctx->nb_layers, |
| 3172 | .flags = hwctx->img_flags, |
| 3173 | .tiling = hwctx->tiling, |
| 3174 | .usage = hwctx->usage, |
| 3175 | .samples = VK_SAMPLE_COUNT_1_BIT, |
| 3176 | .pQueueFamilyIndices = p->img_qfs, |
| 3177 | .queueFamilyIndexCount = p->nb_img_qfs, |
| 3178 | .sharingMode = p->nb_img_qfs > 1 ? VK_SHARING_MODE_CONCURRENT : |
| 3179 | VK_SHARING_MODE_EXCLUSIVE, |
| 3180 | }; |
| 3181 | VkDeviceImageMemoryRequirements req_info = { |
| 3182 | .sType = VK_STRUCTURE_TYPE_DEVICE_IMAGE_MEMORY_REQUIREMENTS, |
| 3183 | .pCreateInfo = &create_info, |
| 3184 | .planeAspect = hwctx->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT ? |
| 3185 | VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT : 0, |
| 3186 | }; |
| 3187 | VkMemoryRequirements2 req = { |
| 3188 | .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2, |
| 3189 | }; |
| 3190 | |
| 3191 | vk->GetDeviceImageMemoryRequirements(dev_hwctx->act_dev, &req_info, &req); |
| 3192 | /* Check that alloc_bind_mem() will find a compatible memory type. */ |
| 3193 | VkMemoryPropertyFlags req_flags = hwctx->tiling == VK_IMAGE_TILING_LINEAR ? |
| 3194 | VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT : |
| 3195 | VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT; |
| 3196 | for (index = 0; index < p->mprops.memoryTypeCount; index++) { |
| 3197 | if (!(req.memoryRequirements.memoryTypeBits & (1U << index))) |
| 3198 | continue; |
| 3199 | if ((p->mprops.memoryTypes[index].propertyFlags & req_flags) == req_flags) |
| 3200 | break; |
| 3201 | } |
| 3202 | if (index == p->mprops.memoryTypeCount) { |
| 3203 | av_log(hwfc, AV_LOG_VERBOSE40, "Disabling host image transfers: " |
| 3204 | "no compatible memory type\n"); |
| 3205 | return 0; |
| 3206 | } |
| 3207 | } |
| 3208 | |
| 3209 | return 1; |
| 3210 | } |
| 3211 | |
| 3212 | static int vulkan_frames_init(AVHWFramesContext *hwfc) |
| 3213 | { |
| 3214 | int err; |
| 3215 | AVVkFrame *f; |
| 3216 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 3217 | AVVulkanFramesContext *hwctx = &fp->p; |
| 3218 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 3219 | AVVulkanDeviceContext *dev_hwctx = &p->p; |
| 3220 | VkImageUsageFlags supported_usage; |
| 3221 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 3222 | const struct FFVkFormatEntry *fmt; |
| 3223 | int disable_multiplane = p->disable_multiplane || |
| 3224 | (hwctx->flags & AV_VK_FRAME_FLAG_DISABLE_MULTIPLANE); |
| 3225 | int is_lone_dpb = ((hwctx->usage & VK_IMAGE_USAGE_VIDEO_ENCODE_DPB_BIT_KHR) || |
| 3226 | ((hwctx->usage & VK_IMAGE_USAGE_VIDEO_DECODE_DPB_BIT_KHR) && |
| 3227 | !(hwctx->usage & VK_IMAGE_USAGE_VIDEO_DECODE_DST_BIT_KHR))); |
| 3228 | |
| 3229 | /* Defaults */ |
| 3230 | if (!hwctx->nb_layers) |
| 3231 | hwctx->nb_layers = 1; |
| 3232 | |
| 3233 | /* VK_IMAGE_TILING_OPTIMAL == 0, can't check for it really */ |
| 3234 | if (p->use_linear_images && |
| 3235 | (hwctx->tiling != VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT)) |
| 3236 | hwctx->tiling = VK_IMAGE_TILING_LINEAR; |
| 3237 | |
| 3238 | |
| 3239 | fmt = vk_find_format_entry(hwfc->sw_format); |
| 3240 | if (!fmt) { |
| 3241 | av_log(hwfc, AV_LOG_ERROR16, "Unsupported pixel format: %s!\n", |
| 3242 | av_get_pix_fmt_name(hwfc->sw_format)); |
| 3243 | return AVERROR(ENOTSUP)(-(95)); |
| 3244 | } |
| 3245 | |
| 3246 | if (hwctx->format[0] != VK_FORMAT_UNDEFINED) { |
| 3247 | if (hwctx->format[0] != fmt->vkf) { |
| 3248 | for (int i = 0; i < fmt->nb_images_fallback; i++) { |
| 3249 | if (hwctx->format[i] != fmt->fallback[i]) { |
| 3250 | av_log(hwfc, AV_LOG_ERROR16, "Incompatible Vulkan format given " |
| 3251 | "for the current sw_format %s!\n", |
| 3252 | av_get_pix_fmt_name(hwfc->sw_format)); |
| 3253 | return AVERROR(EINVAL)(-(22)); |
| 3254 | } |
| 3255 | } |
| 3256 | } |
| 3257 | |
| 3258 | /* Check if the sw_format itself is supported */ |
| 3259 | err = vkfmt_from_pixfmt2(hwfc->device_ctx, hwfc->sw_format, |
| 3260 | hwctx->tiling, NULL((void*)0), |
| 3261 | NULL((void*)0), NULL((void*)0), &supported_usage, 0, |
| 3262 | !hwctx->usage || |
| 3263 | (hwctx->usage & VK_IMAGE_USAGE_STORAGE_BIT)); |
| 3264 | if (err < 0) { |
| 3265 | av_log(hwfc, AV_LOG_ERROR16, "Unsupported sw format: %s!\n", |
| 3266 | av_get_pix_fmt_name(hwfc->sw_format)); |
| 3267 | return AVERROR(EINVAL)(-(22)); |
| 3268 | } |
| 3269 | } else { |
| 3270 | err = vkfmt_from_pixfmt2(hwfc->device_ctx, hwfc->sw_format, |
| 3271 | hwctx->tiling, hwctx->format, NULL((void*)0), |
| 3272 | NULL((void*)0), &supported_usage, |
| 3273 | disable_multiplane, |
| 3274 | !hwctx->usage || |
| 3275 | (hwctx->usage & VK_IMAGE_USAGE_STORAGE_BIT)); |
| 3276 | if (err < 0) |
| 3277 | return err; |
| 3278 | } |
| 3279 | |
| 3280 | /* Lone DPB images do not need additional flags. */ |
| 3281 | /* With DRM modifier + video profile the caller has already chosen a valid |
| 3282 | * usage/img_flags/chain; do not add usage or img_flags (supported_usage does |
| 3283 | * not consider the actual modifier or video profile). */ |
| 3284 | int drm_mod_with_video = (hwctx->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT && |
| 3285 | ff_vk_find_struct(hwctx->create_pnext, |
| 3286 | VK_STRUCTURE_TYPE_VIDEO_PROFILE_LIST_INFO_KHR)); |
| 3287 | |
| 3288 | if (!is_lone_dpb && !drm_mod_with_video) { |
| 3289 | /* Image usage flags */ |
| 3290 | hwctx->usage |= supported_usage & (VK_IMAGE_USAGE_TRANSFER_DST_BIT | |
| 3291 | VK_IMAGE_USAGE_TRANSFER_SRC_BIT | |
| 3292 | VK_IMAGE_USAGE_STORAGE_BIT | |
| 3293 | VK_IMAGE_USAGE_SAMPLED_BIT); |
| 3294 | |
| 3295 | /* Frames which may double as active references (coinciding decode |
| 3296 | * output) cannot support host transfers: decode submissions bake the |
| 3297 | * image layout into their barriers at record time, so a host-side |
| 3298 | * layout transition cannot be synchronized against them, not even |
| 3299 | * by the frame lock and timeline semaphore. */ |
| 3300 | if (p->vkctx.extensions & FF_VK_EXT_HOST_IMAGE_COPY(1ULL << 18) && |
| 3301 | !(hwctx->usage & VK_IMAGE_USAGE_VIDEO_DECODE_DPB_BIT_KHR)) |
| 3302 | hwctx->usage |= supported_usage & VK_IMAGE_USAGE_HOST_TRANSFER_BIT_EXT; |
| 3303 | |
| 3304 | /* Enables encoding of images, if supported by format and extensions */ |
| 3305 | if ((supported_usage & VK_IMAGE_USAGE_VIDEO_ENCODE_SRC_BIT_KHR) && |
| 3306 | (p->vkctx.extensions & FF_VK_EXT_VIDEO_ENCODE_QUEUE(1ULL << 50)) && |
| 3307 | (p->vkctx.extensions & FF_VK_EXT_VIDEO_MAINTENANCE_1(1ULL << 37))) |
| 3308 | hwctx->usage |= VK_IMAGE_USAGE_VIDEO_ENCODE_SRC_BIT_KHR; |
| 3309 | |
| 3310 | /* Image creation flags. |
| 3311 | * Only fill them in automatically if the image is not going to be used as |
| 3312 | * a DPB-only image, and we have SAMPLED/STORAGE bits set. */ |
| 3313 | if (!hwctx->img_flags) { |
| 3314 | int sampleable = hwctx->usage & (VK_IMAGE_USAGE_SAMPLED_BIT | |
| 3315 | VK_IMAGE_USAGE_STORAGE_BIT); |
| 3316 | hwctx->img_flags = VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT; |
| 3317 | if (sampleable) { |
| 3318 | hwctx->img_flags |= VK_IMAGE_CREATE_ALIAS_BIT; |
| 3319 | if ((fmt->vk_planes > 1) && (hwctx->format[0] == fmt->vkf)) |
| 3320 | hwctx->img_flags |= VK_IMAGE_CREATE_EXTENDED_USAGE_BIT; |
| 3321 | } |
| 3322 | } |
| 3323 | } |
| 3324 | |
| 3325 | /* If the image has an ENCODE_SRC usage, and the maintenance1 |
| 3326 | * extension is supported, check if it has a profile list. |
| 3327 | * If there's no profile list, or it has no encode operations, |
| 3328 | * then allow creating the image with no specific profile. */ |
| 3329 | if ((hwctx->usage & VK_IMAGE_USAGE_VIDEO_ENCODE_SRC_BIT_KHR) && |
| 3330 | (p->vkctx.extensions & FF_VK_EXT_VIDEO_ENCODE_QUEUE(1ULL << 50)) && |
| 3331 | (p->vkctx.extensions & FF_VK_EXT_VIDEO_MAINTENANCE_1(1ULL << 37))) { |
| 3332 | const VkVideoProfileListInfoKHR *pl; |
| 3333 | pl = ff_vk_find_struct(hwctx->create_pnext, VK_STRUCTURE_TYPE_VIDEO_PROFILE_LIST_INFO_KHR); |
| 3334 | if (!pl) { |
| 3335 | hwctx->img_flags |= VK_IMAGE_CREATE_VIDEO_PROFILE_INDEPENDENT_BIT_KHR; |
| 3336 | } else { |
| 3337 | uint32_t i; |
| 3338 | for (i = 0; i < pl->profileCount; i++) { |
| 3339 | /* Video ops start at exactly 0x00010000 */ |
| 3340 | if (pl->pProfiles[i].videoCodecOperation & 0xFFFF0000) |
| 3341 | break; |
| 3342 | } |
| 3343 | if (i == pl->profileCount) |
| 3344 | hwctx->img_flags |= VK_IMAGE_CREATE_VIDEO_PROFILE_INDEPENDENT_BIT_KHR; |
| 3345 | } |
| 3346 | } |
| 3347 | |
| 3348 | /* Drop the host transfer if it isn't usable for this image configuration. */ |
| 3349 | if ((hwctx->usage & VK_IMAGE_USAGE_HOST_TRANSFER_BIT_EXT) && |
| 3350 | !vulkan_host_transfer_usable(hwfc)) |
| 3351 | hwctx->usage &= ~VK_IMAGE_USAGE_HOST_TRANSFER_BIT_EXT; |
| 3352 | |
| 3353 | if (!hwctx->lock_frame) |
| 3354 | hwctx->lock_frame = lock_frame; |
| 3355 | |
| 3356 | if (!hwctx->unlock_frame) |
| 3357 | hwctx->unlock_frame = unlock_frame; |
| 3358 | |
| 3359 | err = ff_vk_exec_pool_init(&p->vkctx, p->compute_qf, &fp->compute_exec, |
| 3360 | 2, 0, 0, 0, NULL((void*)0)); |
| 3361 | if (err) |
| 3362 | return err; |
| 3363 | |
| 3364 | err = ff_vk_exec_pool_init(&p->vkctx, p->transfer_qf, &fp->upload_exec, |
| 3365 | FF_VK_DEFAULT_EXEC_CONTEXTS4, 0, 0, 0, NULL((void*)0)); |
| 3366 | if (err) |
| 3367 | return err; |
| 3368 | |
| 3369 | err = ff_vk_exec_pool_init(&p->vkctx, p->transfer_qf, &fp->download_exec, |
| 3370 | 2, 0, 0, 0, NULL((void*)0)); |
| 3371 | if (err) |
| 3372 | return err; |
| 3373 | |
| 3374 | /* Test to see if allocation will fail */ |
| 3375 | err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage, hwctx->img_flags, |
| 3376 | hwctx->nb_layers, hwctx->create_pnext); |
| 3377 | if (err) |
| 3378 | return err; |
| 3379 | |
| 3380 | /* Collect `VkDrmFormatModifierPropertiesEXT` for each plane. Required for DRM export. */ |
| 3381 | if (p->vkctx.extensions & FF_VK_EXT_DRM_MODIFIER_FLAGS(1ULL << 1) && hwctx->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT) { |
| 3382 | VkImageDrmFormatModifierPropertiesEXT drm_mod = { |
| 3383 | .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_PROPERTIES_EXT, |
| 3384 | }; |
| 3385 | err = vk->GetImageDrmFormatModifierPropertiesEXT(dev_hwctx->act_dev, f->img[0], |
| 3386 | &drm_mod); |
| 3387 | if (err != VK_SUCCESS) { |
| 3388 | av_log(hwfc, AV_LOG_ERROR16, "Failed to get image DRM format modifier properties"); |
| 3389 | vulkan_frame_free(hwfc, f); |
| 3390 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 3391 | } |
| 3392 | for (int i = 0; i < fmt->vk_planes; ++i) { |
| 3393 | VkDrmFormatModifierPropertiesListEXT modp; |
| 3394 | VkFormatProperties2 fmtp; |
| 3395 | VkDrmFormatModifierPropertiesEXT *mod_props = NULL((void*)0); |
| 3396 | |
| 3397 | modp = (VkDrmFormatModifierPropertiesListEXT) { |
| 3398 | .sType = VK_STRUCTURE_TYPE_DRM_FORMAT_MODIFIER_PROPERTIES_LIST_EXT, |
| 3399 | }; |
| 3400 | fmtp = (VkFormatProperties2) { |
| 3401 | .sType = VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2, |
| 3402 | .pNext = &modp, |
| 3403 | }; |
| 3404 | |
| 3405 | /* query drmFormatModifierCount by keeping pDrmFormatModifierProperties NULL */ |
| 3406 | vk->GetPhysicalDeviceFormatProperties2(dev_hwctx->phys_dev, fmt->fallback[i], &fmtp); |
| 3407 | |
| 3408 | modp.pDrmFormatModifierProperties = |
| 3409 | av_calloc(modp.drmFormatModifierCount, sizeof(*modp.pDrmFormatModifierProperties)); |
| 3410 | if (!modp.pDrmFormatModifierProperties) { |
| 3411 | vulkan_frame_free(hwfc, f); |
| 3412 | return AVERROR(ENOMEM)(-(12)); |
| 3413 | } |
| 3414 | vk->GetPhysicalDeviceFormatProperties2(dev_hwctx->phys_dev, fmt->fallback[i], &fmtp); |
| 3415 | |
| 3416 | for (uint32_t j = 0; j < modp.drmFormatModifierCount; ++j) { |
| 3417 | VkDrmFormatModifierPropertiesEXT *m = &modp.pDrmFormatModifierProperties[j]; |
| 3418 | if (m->drmFormatModifier == drm_mod.drmFormatModifier) { |
| 3419 | mod_props = m; |
| 3420 | break; |
| 3421 | } |
| 3422 | } |
| 3423 | |
| 3424 | if (mod_props == NULL((void*)0)) { |
| 3425 | av_log(hwfc, AV_LOG_ERROR16, "No DRM format modifier properties found for modifier 0x%016"PRIx64"l" "x""\n", |
| 3426 | drm_mod.drmFormatModifier); |
| 3427 | av_free(modp.pDrmFormatModifierProperties); |
| 3428 | vulkan_frame_free(hwfc, f); |
| 3429 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 3430 | } |
| 3431 | |
| 3432 | fp->drm_format_modifier_properties[i] = *mod_props; |
| 3433 | av_free(modp.pDrmFormatModifierProperties); |
| 3434 | } |
| 3435 | } |
| 3436 | |
| 3437 | vulkan_frame_free(hwfc, f); |
| 3438 | |
| 3439 | /* If user did not specify a pool, hwfc->pool will be set to the internal one |
| 3440 | * in hwcontext.c just after this gets called */ |
| 3441 | if (!hwfc->pool) { |
| 3442 | ffhwframesctx(hwfc)->pool_internal = av_buffer_pool_init2(sizeof(AVVkFrame), |
| 3443 | hwfc, vulkan_pool_alloc, |
| 3444 | NULL((void*)0)); |
| 3445 | if (!ffhwframesctx(hwfc)->pool_internal) |
| 3446 | return AVERROR(ENOMEM)(-(12)); |
| 3447 | } |
| 3448 | |
| 3449 | return 0; |
| 3450 | } |
| 3451 | |
| 3452 | static int vulkan_get_buffer(AVHWFramesContext *hwfc, AVFrame *frame) |
| 3453 | { |
| 3454 | frame->buf[0] = av_buffer_pool_get(hwfc->pool); |
| 3455 | if (!frame->buf[0]) |
| 3456 | return AVERROR(ENOMEM)(-(12)); |
| 3457 | |
| 3458 | frame->data[0] = frame->buf[0]->data; |
| 3459 | frame->format = AV_PIX_FMT_VULKAN; |
| 3460 | frame->width = hwfc->width; |
| 3461 | frame->height = hwfc->height; |
| 3462 | |
| 3463 | return 0; |
| 3464 | } |
| 3465 | |
| 3466 | static int vulkan_transfer_get_formats(AVHWFramesContext *hwfc, |
| 3467 | enum AVHWFrameTransferDirection dir, |
| 3468 | enum AVPixelFormat **formats) |
| 3469 | { |
| 3470 | enum AVPixelFormat *fmts; |
| 3471 | int n = 2; |
| 3472 | |
| 3473 | #if CONFIG_CUDA0 |
| 3474 | n++; |
| 3475 | #endif |
| 3476 | fmts = av_malloc_array(n, sizeof(*fmts)); |
| 3477 | if (!fmts) |
| 3478 | return AVERROR(ENOMEM)(-(12)); |
| 3479 | |
| 3480 | n = 0; |
| 3481 | fmts[n++] = hwfc->sw_format; |
| 3482 | #if CONFIG_CUDA0 |
| 3483 | fmts[n++] = AV_PIX_FMT_CUDA; |
| 3484 | #endif |
| 3485 | fmts[n++] = AV_PIX_FMT_NONE; |
| 3486 | |
| 3487 | *formats = fmts; |
| 3488 | return 0; |
| 3489 | } |
| 3490 | |
| 3491 | #if CONFIG_LIBDRM1 |
| 3492 | static void vulkan_unmap_from_drm(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap) |
| 3493 | { |
| 3494 | vulkan_frame_free(hwfc, hwmap->priv); |
| 3495 | } |
| 3496 | |
| 3497 | static const struct { |
| 3498 | uint32_t drm_fourcc; |
| 3499 | VkFormat vk_format; |
| 3500 | } vulkan_drm_format_map[] = { |
| 3501 | { DRM_FORMAT_R8((__u32)('R') | ((__u32)('8') << 8) | ((__u32)(' ') << 16) | ((__u32)(' ') << 24)), VK_FORMAT_R8_UNORM }, |
| 3502 | { DRM_FORMAT_R16((__u32)('R') | ((__u32)('1') << 8) | ((__u32)('6') << 16) | ((__u32)(' ') << 24)), VK_FORMAT_R16_UNORM }, |
| 3503 | { DRM_FORMAT_GR88((__u32)('G') | ((__u32)('R') << 8) | ((__u32)('8') << 16) | ((__u32)('8') << 24)), VK_FORMAT_R8G8_UNORM }, |
| 3504 | { DRM_FORMAT_RG88((__u32)('R') | ((__u32)('G') << 8) | ((__u32)('8') << 16) | ((__u32)('8') << 24)), VK_FORMAT_R8G8_UNORM }, |
| 3505 | { DRM_FORMAT_GR1616((__u32)('G') | ((__u32)('R') << 8) | ((__u32)('3') << 16) | ((__u32)('2') << 24)), VK_FORMAT_R16G16_UNORM }, |
| 3506 | { DRM_FORMAT_RG1616((__u32)('R') | ((__u32)('G') << 8) | ((__u32)('3') << 16) | ((__u32)('2') << 24)), VK_FORMAT_R16G16_UNORM }, |
| 3507 | { DRM_FORMAT_ARGB8888((__u32)('A') | ((__u32)('R') << 8) | ((__u32)('2') << 16) | ((__u32)('4') << 24)), VK_FORMAT_B8G8R8A8_UNORM }, |
| 3508 | { DRM_FORMAT_XRGB8888((__u32)('X') | ((__u32)('R') << 8) | ((__u32)('2') << 16) | ((__u32)('4') << 24)), VK_FORMAT_B8G8R8A8_UNORM }, |
| 3509 | { DRM_FORMAT_ABGR8888((__u32)('A') | ((__u32)('B') << 8) | ((__u32)('2') << 16) | ((__u32)('4') << 24)), VK_FORMAT_R8G8B8A8_UNORM }, |
| 3510 | { DRM_FORMAT_XBGR8888((__u32)('X') | ((__u32)('B') << 8) | ((__u32)('2') << 16) | ((__u32)('4') << 24)), VK_FORMAT_R8G8B8A8_UNORM }, |
| 3511 | { DRM_FORMAT_ARGB2101010((__u32)('A') | ((__u32)('R') << 8) | ((__u32)('3') << 16) | ((__u32)('0') << 24)), VK_FORMAT_A2B10G10R10_UNORM_PACK32 }, |
| 3512 | { DRM_FORMAT_ABGR2101010((__u32)('A') | ((__u32)('B') << 8) | ((__u32)('3') << 16) | ((__u32)('0') << 24)), VK_FORMAT_A2R10G10B10_UNORM_PACK32 }, |
| 3513 | { DRM_FORMAT_XRGB2101010((__u32)('X') | ((__u32)('R') << 8) | ((__u32)('3') << 16) | ((__u32)('0') << 24)), VK_FORMAT_A2B10G10R10_UNORM_PACK32 }, |
| 3514 | { DRM_FORMAT_XBGR2101010((__u32)('X') | ((__u32)('B') << 8) | ((__u32)('3') << 16) | ((__u32)('0') << 24)), VK_FORMAT_A2R10G10B10_UNORM_PACK32 }, |
| 3515 | |
| 3516 | // All these DRM_FORMATs were added in the same libdrm commit. |
| 3517 | #ifdef DRM_FORMAT_XYUV8888((__u32)('X') | ((__u32)('Y') << 8) | ((__u32)('U') << 16) | ((__u32)('V') << 24)) |
| 3518 | { DRM_FORMAT_XYUV8888((__u32)('X') | ((__u32)('Y') << 8) | ((__u32)('U') << 16) | ((__u32)('V') << 24)), VK_FORMAT_R8G8B8A8_UNORM }, |
| 3519 | { DRM_FORMAT_XVYU2101010((__u32)('X') | ((__u32)('V') << 8) | ((__u32)('3') << 16) | ((__u32)('0') << 24)), VK_FORMAT_A2R10G10B10_UNORM_PACK32 } , |
| 3520 | { DRM_FORMAT_XVYU12_16161616((__u32)('X') | ((__u32)('V') << 8) | ((__u32)('3') << 16) | ((__u32)('6') << 24)), VK_FORMAT_R12X4G12X4B12X4A12X4_UNORM_4PACK16 } , |
| 3521 | { DRM_FORMAT_XVYU16161616((__u32)('X') | ((__u32)('V') << 8) | ((__u32)('4') << 16) | ((__u32)('8') << 24)), VK_FORMAT_R16G16B16A16_UNORM } , |
| 3522 | #endif |
| 3523 | }; |
| 3524 | |
| 3525 | static inline VkFormat drm_to_vulkan_fmt(uint32_t drm_fourcc) |
| 3526 | { |
| 3527 | for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map)(sizeof(vulkan_drm_format_map) / sizeof((vulkan_drm_format_map )[0])); i++) |
| 3528 | if (vulkan_drm_format_map[i].drm_fourcc == drm_fourcc) |
| 3529 | return vulkan_drm_format_map[i].vk_format; |
| 3530 | return VK_FORMAT_UNDEFINED; |
| 3531 | } |
| 3532 | |
| 3533 | static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **frame, |
| 3534 | const AVFrame *src, int flags) |
| 3535 | { |
| 3536 | int err = 0; |
| 3537 | VkResult ret; |
| 3538 | AVVkFrame *f; |
| 3539 | int bind_counts = 0; |
| 3540 | AVHWDeviceContext *ctx = hwfc->device_ctx; |
| 3541 | VulkanDevicePriv *p = ctx->hwctx; |
| 3542 | AVVulkanDeviceContext *hwctx = &p->p; |
| 3543 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 3544 | const AVDRMFrameDescriptor *desc = (AVDRMFrameDescriptor *)src->data[0]; |
| 3545 | VkBindImageMemoryInfo bind_info[AV_DRM_MAX_PLANES]; |
| 3546 | VkBindImagePlaneMemoryInfo plane_info[AV_DRM_MAX_PLANES]; |
| 3547 | |
| 3548 | for (int i = 0; i < desc->nb_layers; i++) { |
| 3549 | if (drm_to_vulkan_fmt(desc->layers[i].format) == VK_FORMAT_UNDEFINED) { |
| 3550 | av_log(ctx, AV_LOG_ERROR16, "Unsupported DMABUF layer format %#08x!\n", |
| 3551 | desc->layers[i].format); |
| 3552 | return AVERROR(EINVAL)(-(22)); |
| 3553 | } |
| 3554 | } |
| 3555 | |
| 3556 | if (!(f = av_vk_frame_alloc())) { |
| 3557 | av_log(ctx, AV_LOG_ERROR16, "Unable to allocate memory for AVVkFrame!\n"); |
| 3558 | err = AVERROR(ENOMEM)(-(12)); |
| 3559 | goto fail; |
| 3560 | } |
| 3561 | |
| 3562 | f->tiling = VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT; |
| 3563 | |
| 3564 | for (int i = 0; i < desc->nb_layers; i++) { |
| 3565 | const int planes = desc->layers[i].nb_planes; |
| 3566 | |
| 3567 | /* Semaphore */ |
| 3568 | VkSemaphoreTypeCreateInfo sem_type_info = { |
| 3569 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_TYPE_CREATE_INFO, |
| 3570 | .semaphoreType = VK_SEMAPHORE_TYPE_TIMELINE, |
| 3571 | .initialValue = 0, |
| 3572 | }; |
| 3573 | VkSemaphoreCreateInfo sem_spawn = { |
| 3574 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO, |
| 3575 | .pNext = &sem_type_info, |
| 3576 | }; |
| 3577 | |
| 3578 | /* Image creation */ |
| 3579 | VkSubresourceLayout ext_img_layouts[AV_DRM_MAX_PLANES]; |
| 3580 | VkImageDrmFormatModifierExplicitCreateInfoEXT ext_img_mod_spec = { |
| 3581 | .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT, |
| 3582 | .drmFormatModifier = desc->objects[0].format_modifier, |
| 3583 | .drmFormatModifierPlaneCount = planes, |
| 3584 | .pPlaneLayouts = (const VkSubresourceLayout *)&ext_img_layouts, |
| 3585 | }; |
| 3586 | VkExternalMemoryImageCreateInfo ext_img_spec = { |
| 3587 | .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO, |
| 3588 | .pNext = &ext_img_mod_spec, |
| 3589 | .handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT, |
| 3590 | }; |
| 3591 | VkImageCreateInfo create_info = { |
| 3592 | .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO, |
| 3593 | .pNext = &ext_img_spec, |
| 3594 | .imageType = VK_IMAGE_TYPE_2D, |
| 3595 | .format = drm_to_vulkan_fmt(desc->layers[i].format), |
| 3596 | .extent.depth = 1, |
| 3597 | .mipLevels = 1, |
| 3598 | .arrayLayers = 1, |
| 3599 | .flags = 0x0, |
| 3600 | .tiling = VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT, |
| 3601 | .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */ |
| 3602 | .usage = 0x0, /* filled in below */ |
| 3603 | .samples = VK_SAMPLE_COUNT_1_BIT, |
| 3604 | .pQueueFamilyIndices = p->img_qfs, |
| 3605 | .queueFamilyIndexCount = p->nb_img_qfs, |
| 3606 | .sharingMode = p->nb_img_qfs > 1 ? VK_SHARING_MODE_CONCURRENT : |
| 3607 | VK_SHARING_MODE_EXCLUSIVE, |
| 3608 | }; |
| 3609 | |
| 3610 | /* The DRM fourcc fixes a specific channel order (f.ex ARGB8888 maps to |
| 3611 | * B8G8R8A8), but image views are always created from the destination |
| 3612 | * frames context sw_format (like how bgra maps to R8G8B8A8). For a |
| 3613 | * single plane layer, create the image with the sw_format's compatible |
| 3614 | * VkFormat so the image and its views agree without a mutable format |
| 3615 | * list, the format query below validates this */ |
| 3616 | if (planes == 1) { |
| 3617 | const VkFormat *sw_vkfmts = av_vkfmt_from_pixfmt(hwfc->sw_format); |
| 3618 | if (sw_vkfmts && sw_vkfmts[i] != VK_FORMAT_UNDEFINED) |
| 3619 | create_info.format = sw_vkfmts[i]; |
| 3620 | } |
| 3621 | |
| 3622 | /* Image format verification */ |
| 3623 | VkExternalImageFormatProperties ext_props = { |
| 3624 | .sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR, |
| 3625 | }; |
| 3626 | VkImageFormatProperties2 props_ret = { |
| 3627 | .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2, |
| 3628 | .pNext = &ext_props, |
| 3629 | }; |
| 3630 | VkPhysicalDeviceImageDrmFormatModifierInfoEXT props_drm_mod = { |
| 3631 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_DRM_FORMAT_MODIFIER_INFO_EXT, |
| 3632 | .drmFormatModifier = ext_img_mod_spec.drmFormatModifier, |
| 3633 | .pQueueFamilyIndices = create_info.pQueueFamilyIndices, |
| 3634 | .queueFamilyIndexCount = create_info.queueFamilyIndexCount, |
| 3635 | .sharingMode = create_info.sharingMode, |
| 3636 | }; |
| 3637 | VkPhysicalDeviceExternalImageFormatInfo props_ext = { |
| 3638 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO, |
| 3639 | .pNext = &props_drm_mod, |
| 3640 | .handleType = ext_img_spec.handleTypes, |
| 3641 | }; |
| 3642 | VkPhysicalDeviceImageFormatInfo2 fmt_props; |
| 3643 | |
| 3644 | if (flags & AV_HWFRAME_MAP_READ) |
| 3645 | create_info.usage |= VK_IMAGE_USAGE_SAMPLED_BIT | |
| 3646 | VK_IMAGE_USAGE_TRANSFER_SRC_BIT; |
| 3647 | if (flags & AV_HWFRAME_MAP_WRITE) |
| 3648 | create_info.usage |= VK_IMAGE_USAGE_STORAGE_BIT | |
| 3649 | VK_IMAGE_USAGE_TRANSFER_DST_BIT; |
| 3650 | |
| 3651 | fmt_props = (VkPhysicalDeviceImageFormatInfo2) { |
| 3652 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2, |
| 3653 | .pNext = &props_ext, |
| 3654 | .format = create_info.format, |
| 3655 | .type = create_info.imageType, |
| 3656 | .tiling = create_info.tiling, |
| 3657 | .usage = create_info.usage, |
| 3658 | .flags = create_info.flags, |
| 3659 | }; |
| 3660 | |
| 3661 | /* Check if importing is possible for this combination of parameters */ |
| 3662 | ret = vk->GetPhysicalDeviceImageFormatProperties2(hwctx->phys_dev, |
| 3663 | &fmt_props, &props_ret); |
| 3664 | if (ret != VK_SUCCESS) { |
| 3665 | av_log(ctx, AV_LOG_ERROR16, "Cannot map DRM frame to Vulkan: %s\n", |
| 3666 | ff_vk_ret2str(ret)); |
| 3667 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 3668 | goto fail; |
| 3669 | } |
| 3670 | |
| 3671 | /* Set the image width/height */ |
| 3672 | get_plane_wh(&create_info.extent.width, &create_info.extent.height, |
| 3673 | hwfc->sw_format, src->width, src->height, i); |
| 3674 | |
| 3675 | /* Set the subresource layout based on the layer properties */ |
| 3676 | for (int j = 0; j < planes; j++) { |
| 3677 | ext_img_layouts[j].offset = desc->layers[i].planes[j].offset; |
| 3678 | ext_img_layouts[j].rowPitch = desc->layers[i].planes[j].pitch; |
| 3679 | ext_img_layouts[j].size = 0; /* The specs say so for all 3 */ |
| 3680 | ext_img_layouts[j].arrayPitch = 0; |
| 3681 | ext_img_layouts[j].depthPitch = 0; |
| 3682 | } |
| 3683 | |
| 3684 | /* Create image */ |
| 3685 | ret = vk->CreateImage(hwctx->act_dev, &create_info, |
| 3686 | hwctx->alloc, &f->img[i]); |
| 3687 | if (ret != VK_SUCCESS) { |
| 3688 | av_log(ctx, AV_LOG_ERROR16, "Image creation failure: %s\n", |
| 3689 | ff_vk_ret2str(ret)); |
| 3690 | err = AVERROR(EINVAL)(-(22)); |
| 3691 | goto fail; |
| 3692 | } |
| 3693 | |
| 3694 | ret = vk->CreateSemaphore(hwctx->act_dev, &sem_spawn, |
| 3695 | hwctx->alloc, &f->sem[i]); |
| 3696 | if (ret != VK_SUCCESS) { |
| 3697 | av_log(hwctx, AV_LOG_ERROR16, "Failed to create semaphore: %s\n", |
| 3698 | ff_vk_ret2str(ret)); |
| 3699 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 3700 | goto fail; |
| 3701 | } |
| 3702 | |
| 3703 | f->queue_family[i] = VK_QUEUE_FAMILY_EXTERNAL(~1U); |
| 3704 | f->layout[i] = create_info.initialLayout; |
| 3705 | f->access[i] = 0x0; |
| 3706 | f->sem_value[i] = 0; |
| 3707 | } |
| 3708 | |
| 3709 | for (int i = 0; i < desc->nb_layers; i++) { |
| 3710 | /* Memory requirements */ |
| 3711 | VkImageMemoryRequirementsInfo2 req_desc = { |
| 3712 | .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2, |
| 3713 | .image = f->img[i], |
| 3714 | }; |
| 3715 | VkMemoryDedicatedRequirements ded_req = { |
| 3716 | .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS, |
| 3717 | }; |
| 3718 | VkMemoryRequirements2 req2 = { |
| 3719 | .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2, |
| 3720 | .pNext = &ded_req, |
| 3721 | }; |
| 3722 | |
| 3723 | /* Allocation/importing */ |
| 3724 | VkMemoryFdPropertiesKHR fdmp = { |
| 3725 | .sType = VK_STRUCTURE_TYPE_MEMORY_FD_PROPERTIES_KHR, |
| 3726 | }; |
| 3727 | /* This assumes that a layer will never be constructed from multiple |
| 3728 | * objects. If that was to happen in the real world, this code would |
| 3729 | * need to import each plane separately. |
| 3730 | */ |
| 3731 | VkImportMemoryFdInfoKHR idesc = { |
| 3732 | .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR, |
| 3733 | .fd = dup(desc->objects[desc->layers[i].planes[0].object_index].fd), |
| 3734 | .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT, |
| 3735 | }; |
| 3736 | VkMemoryDedicatedAllocateInfo ded_alloc = { |
| 3737 | .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO, |
| 3738 | .pNext = &idesc, |
| 3739 | .image = req_desc.image, |
| 3740 | }; |
| 3741 | |
| 3742 | /* Get object properties */ |
| 3743 | ret = vk->GetMemoryFdPropertiesKHR(hwctx->act_dev, |
| 3744 | VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT, |
| 3745 | idesc.fd, &fdmp); |
| 3746 | if (ret != VK_SUCCESS) { |
| 3747 | av_log(hwfc, AV_LOG_ERROR16, "Failed to get FD properties: %s\n", |
| 3748 | ff_vk_ret2str(ret)); |
| 3749 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 3750 | close(idesc.fd); |
| 3751 | goto fail; |
| 3752 | } |
| 3753 | |
| 3754 | vk->GetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req2); |
| 3755 | |
| 3756 | /* Only a single bit must be set, not a range, and it must match */ |
| 3757 | req2.memoryRequirements.memoryTypeBits = fdmp.memoryTypeBits; |
| 3758 | |
| 3759 | err = alloc_mem(ctx, &req2.memoryRequirements, |
| 3760 | VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, |
| 3761 | (ded_req.prefersDedicatedAllocation || |
| 3762 | ded_req.requiresDedicatedAllocation) ? |
| 3763 | &ded_alloc : ded_alloc.pNext, |
| 3764 | &f->flags, &f->mem[i]); |
| 3765 | if (err) { |
| 3766 | close(idesc.fd); |
| 3767 | goto fail; |
| 3768 | } |
| 3769 | |
| 3770 | f->size[i] = req2.memoryRequirements.size; |
| 3771 | } |
| 3772 | |
| 3773 | for (int i = 0; i < desc->nb_layers; i++) { |
| 3774 | const int planes = desc->layers[i].nb_planes; |
| 3775 | for (int j = 0; j < planes; j++) { |
| 3776 | VkImageAspectFlagBits aspect = j == 0 ? VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT : |
| 3777 | j == 1 ? VK_IMAGE_ASPECT_MEMORY_PLANE_1_BIT_EXT : |
| 3778 | VK_IMAGE_ASPECT_MEMORY_PLANE_2_BIT_EXT; |
| 3779 | |
| 3780 | plane_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO; |
| 3781 | plane_info[bind_counts].pNext = NULL((void*)0); |
| 3782 | plane_info[bind_counts].planeAspect = aspect; |
| 3783 | |
| 3784 | bind_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO; |
| 3785 | bind_info[bind_counts].pNext = planes > 1 ? &plane_info[bind_counts] : NULL((void*)0); |
| 3786 | bind_info[bind_counts].image = f->img[i]; |
| 3787 | bind_info[bind_counts].memory = f->mem[i]; |
| 3788 | |
| 3789 | /* Offset is already signalled via pPlaneLayouts above */ |
| 3790 | bind_info[bind_counts].memoryOffset = 0; |
| 3791 | |
| 3792 | bind_counts++; |
| 3793 | } |
| 3794 | } |
| 3795 | |
| 3796 | /* Bind the allocated memory to the images */ |
| 3797 | ret = vk->BindImageMemory2(hwctx->act_dev, bind_counts, bind_info); |
| 3798 | if (ret != VK_SUCCESS) { |
| 3799 | av_log(ctx, AV_LOG_ERROR16, "Failed to bind memory: %s\n", |
| 3800 | ff_vk_ret2str(ret)); |
| 3801 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 3802 | goto fail; |
| 3803 | } |
| 3804 | |
| 3805 | *frame = f; |
| 3806 | |
| 3807 | return 0; |
| 3808 | |
| 3809 | fail: |
| 3810 | vulkan_frame_free(hwfc, f); |
| 3811 | |
| 3812 | return err; |
| 3813 | } |
| 3814 | |
| 3815 | static int vulkan_map_from_drm_frame_sync(AVHWFramesContext *hwfc, AVFrame *dst, |
| 3816 | const AVDRMFrameDescriptor *desc, int flags) |
| 3817 | { |
| 3818 | int err; |
| 3819 | VkResult ret; |
| 3820 | AVHWDeviceContext *ctx = hwfc->device_ctx; |
| 3821 | VulkanDevicePriv *p = ctx->hwctx; |
| 3822 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 3823 | AVVulkanDeviceContext *hwctx = &p->p; |
| 3824 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 3825 | |
| 3826 | #ifdef DMA_BUF_IOCTL_EXPORT_SYNC_FILE(((2U|1U) << (((0 +8)+8)+14)) | ((('b')) << (0 +8 )) | (((2)) << 0) | ((((sizeof(struct dma_buf_export_sync_file )))) << ((0 +8)+8))) |
| 3827 | if (p->vkctx.extensions & FF_VK_EXT_EXTERNAL_FD_SEM(1ULL << 3)) { |
| 3828 | VkCommandBuffer cmd_buf; |
| 3829 | FFVkExecContext *exec; |
| 3830 | VkImageMemoryBarrier2 img_bar[AV_NUM_DATA_POINTERS8]; |
| 3831 | VkSemaphore drm_sync_sem[AV_DRM_MAX_PLANES] = { 0 }; |
| 3832 | int nb_img_bar = 0; |
| 3833 | |
| 3834 | for (int i = 0; i < desc->nb_objects; i++) { |
| 3835 | VkSemaphoreTypeCreateInfo sem_type_info = { |
| 3836 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_TYPE_CREATE_INFO, |
| 3837 | .semaphoreType = VK_SEMAPHORE_TYPE_BINARY, |
| 3838 | }; |
| 3839 | VkSemaphoreCreateInfo sem_spawn = { |
| 3840 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO, |
| 3841 | .pNext = &sem_type_info, |
| 3842 | }; |
| 3843 | VkImportSemaphoreFdInfoKHR import_info; |
| 3844 | struct dma_buf_export_sync_file implicit_fd_info = { |
| 3845 | .flags = DMA_BUF_SYNC_READ(1 << 0), |
| 3846 | .fd = -1, |
| 3847 | }; |
| 3848 | |
| 3849 | if (ioctl(desc->objects[i].fd, DMA_BUF_IOCTL_EXPORT_SYNC_FILE(((2U|1U) << (((0 +8)+8)+14)) | ((('b')) << (0 +8 )) | (((2)) << 0) | ((((sizeof(struct dma_buf_export_sync_file )))) << ((0 +8)+8))), |
| 3850 | &implicit_fd_info)) { |
| 3851 | err = AVERROR(errno)(-((*__errno_location ()))); |
| 3852 | av_log(hwctx, AV_LOG_ERROR16, "Failed to retrieve implicit DRM sync file: %s\n", |
| 3853 | av_err2str(err)av_make_error_string((char[64]){0}, 64, err)); |
| 3854 | for (; i >= 0; i--) |
| 3855 | vk->DestroySemaphore(hwctx->act_dev, drm_sync_sem[i], hwctx->alloc); |
| 3856 | return err; |
| 3857 | } |
| 3858 | |
| 3859 | ret = vk->CreateSemaphore(hwctx->act_dev, &sem_spawn, |
| 3860 | hwctx->alloc, &drm_sync_sem[i]); |
| 3861 | if (ret != VK_SUCCESS) { |
| 3862 | av_log(hwctx, AV_LOG_ERROR16, "Failed to create semaphore: %s\n", |
| 3863 | ff_vk_ret2str(ret)); |
| 3864 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 3865 | for (; i >= 0; i--) |
| 3866 | vk->DestroySemaphore(hwctx->act_dev, drm_sync_sem[i], hwctx->alloc); |
| 3867 | return err; |
| 3868 | } |
| 3869 | |
| 3870 | import_info = (VkImportSemaphoreFdInfoKHR) { |
| 3871 | .sType = VK_STRUCTURE_TYPE_IMPORT_SEMAPHORE_FD_INFO_KHR, |
| 3872 | .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT, |
| 3873 | .flags = VK_SEMAPHORE_IMPORT_TEMPORARY_BIT, |
| 3874 | .semaphore = drm_sync_sem[i], |
| 3875 | .fd = implicit_fd_info.fd, |
| 3876 | }; |
| 3877 | |
| 3878 | ret = vk->ImportSemaphoreFdKHR(hwctx->act_dev, &import_info); |
| 3879 | if (ret != VK_SUCCESS) { |
| 3880 | av_log(hwctx, AV_LOG_ERROR16, "Failed to import semaphore: %s\n", |
| 3881 | ff_vk_ret2str(ret)); |
| 3882 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 3883 | for (; i >= 0; i--) |
| 3884 | vk->DestroySemaphore(hwctx->act_dev, drm_sync_sem[i], hwctx->alloc); |
| 3885 | return err; |
| 3886 | } |
| 3887 | } |
| 3888 | |
| 3889 | exec = ff_vk_exec_get(&p->vkctx, &fp->compute_exec); |
| 3890 | cmd_buf = exec->buf; |
| 3891 | |
| 3892 | err = ff_vk_exec_start(&p->vkctx, exec); |
| 3893 | if (err < 0) { |
| 3894 | for (int i = 0; i < desc->nb_objects; i++) |
| 3895 | vk->DestroySemaphore(hwctx->act_dev, drm_sync_sem[i], hwctx->alloc); |
| 3896 | return err; |
| 3897 | } |
| 3898 | |
| 3899 | /* Ownership of semaphores is passed */ |
| 3900 | ff_vk_exec_add_dep_bool_sem(&p->vkctx, exec, |
| 3901 | drm_sync_sem, desc->nb_objects, |
| 3902 | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, 1); |
| 3903 | |
| 3904 | err = ff_vk_exec_add_dep_frame(&p->vkctx, exec, dst, |
| 3905 | VK_PIPELINE_STAGE_2_NONE, |
| 3906 | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT); |
| 3907 | if (err < 0) { |
| 3908 | ff_vk_exec_discard(&p->vkctx, exec); |
| 3909 | return err; |
| 3910 | } |
| 3911 | |
| 3912 | ff_vk_frame_barrier(&p->vkctx, exec, dst, img_bar, &nb_img_bar, |
| 3913 | VK_PIPELINE_STAGE_2_NONE, |
| 3914 | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, |
| 3915 | ((flags & AV_HWFRAME_MAP_READ) ? |
| 3916 | VK_ACCESS_2_SHADER_SAMPLED_READ_BIT : 0x0) | |
| 3917 | ((flags & AV_HWFRAME_MAP_WRITE) ? |
| 3918 | VK_ACCESS_2_SHADER_STORAGE_WRITE_BIT : 0x0), |
| 3919 | VK_IMAGE_LAYOUT_GENERAL, |
| 3920 | p->nb_img_qfs > 1 ? VK_QUEUE_FAMILY_IGNORED(~0U) : p->img_qfs[0]); |
| 3921 | |
| 3922 | vk->CmdPipelineBarrier2(cmd_buf, &(VkDependencyInfo) { |
| 3923 | .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO, |
| 3924 | .pImageMemoryBarriers = img_bar, |
| 3925 | .imageMemoryBarrierCount = nb_img_bar, |
| 3926 | }); |
| 3927 | |
| 3928 | err = ff_vk_exec_submit(&p->vkctx, exec); |
| 3929 | if (err < 0) |
| 3930 | return err; |
| 3931 | } else |
| 3932 | #endif |
| 3933 | { |
| 3934 | AVVkFrame *f = (AVVkFrame *)dst->data[0]; |
| 3935 | av_log(hwctx, AV_LOG_WARNING24, "No support for synchronization when importing DMA-BUFs, " |
| 3936 | "image may be corrupted.\n"); |
| 3937 | err = prepare_frame(hwfc, &fp->compute_exec, f, PREP_MODE_EXTERNAL_IMPORT); |
| 3938 | if (err) |
| 3939 | return err; |
| 3940 | } |
| 3941 | |
| 3942 | return 0; |
| 3943 | } |
| 3944 | |
| 3945 | static int vulkan_map_from_drm(AVHWFramesContext *hwfc, AVFrame *dst, |
| 3946 | const AVFrame *src, int flags) |
| 3947 | { |
| 3948 | int err = 0; |
| 3949 | AVVkFrame *f; |
| 3950 | const AVDRMFrameDescriptor *desc = (AVDRMFrameDescriptor *)src->data[0]; |
| 3951 | |
| 3952 | if ((err = vulkan_map_from_drm_frame_desc(hwfc, &f, src, flags))) |
| 3953 | return err; |
| 3954 | |
| 3955 | /* The unmapping function will free this */ |
| 3956 | dst->data[0] = (uint8_t *)f; |
| 3957 | dst->width = src->width; |
| 3958 | dst->height = src->height; |
| 3959 | |
| 3960 | err = ff_hwframe_map_create(dst->hw_frames_ctx, dst, src, |
| 3961 | &vulkan_unmap_from_drm, f); |
| 3962 | if (err < 0) |
| 3963 | goto fail; |
| 3964 | |
| 3965 | err = vulkan_map_from_drm_frame_sync(hwfc, dst, desc, flags); |
| 3966 | if (err < 0) |
| 3967 | return err; |
| 3968 | |
| 3969 | av_log(hwfc, AV_LOG_DEBUG48, "Mapped DRM object to Vulkan!\n"); |
| 3970 | |
| 3971 | return 0; |
| 3972 | |
| 3973 | fail: |
| 3974 | vulkan_frame_free(hwfc->device_ctx->hwctx, f); |
| 3975 | dst->data[0] = NULL((void*)0); |
| 3976 | return err; |
| 3977 | } |
| 3978 | |
| 3979 | #if CONFIG_VAAPI1 |
| 3980 | static int vulkan_map_from_vaapi(AVHWFramesContext *dst_fc, |
| 3981 | AVFrame *dst, const AVFrame *src, |
| 3982 | int flags) |
| 3983 | { |
| 3984 | int err; |
| 3985 | AVFrame *tmp = av_frame_alloc(); |
| 3986 | AVHWFramesContext *vaapi_fc = (AVHWFramesContext*)src->hw_frames_ctx->data; |
| 3987 | AVVAAPIDeviceContext *vaapi_ctx = vaapi_fc->device_ctx->hwctx; |
| 3988 | VASurfaceID surface_id = (VASurfaceID)(uintptr_t)src->data[3]; |
| 3989 | |
| 3990 | if (!tmp) |
| 3991 | return AVERROR(ENOMEM)(-(12)); |
| 3992 | |
| 3993 | /* We have to sync since like the previous comment said, no semaphores */ |
| 3994 | vaSyncSurface(vaapi_ctx->display, surface_id); |
| 3995 | |
| 3996 | tmp->format = AV_PIX_FMT_DRM_PRIME; |
| 3997 | |
| 3998 | err = av_hwframe_map(tmp, src, flags); |
| 3999 | if (err < 0) |
| 4000 | goto fail; |
| 4001 | |
| 4002 | err = vulkan_map_from_drm(dst_fc, dst, tmp, flags); |
| 4003 | if (err < 0) |
| 4004 | goto fail; |
| 4005 | |
| 4006 | err = ff_hwframe_map_replace(dst, src); |
| 4007 | |
| 4008 | fail: |
| 4009 | av_frame_free(&tmp); |
| 4010 | return err; |
| 4011 | } |
| 4012 | #endif |
| 4013 | #endif |
| 4014 | |
| 4015 | #if CONFIG_CUDA0 |
| 4016 | static int export_mem_to_cuda(AVHWDeviceContext *ctx, |
| 4017 | AVHWDeviceContext *cuda_cu, CudaFunctions *cu, |
| 4018 | AVVkFrameInternal *dst_int, int idx, |
| 4019 | VkDeviceMemory mem, size_t size) |
| 4020 | { |
| 4021 | VkResult ret; |
| 4022 | VulkanDevicePriv *p = ctx->hwctx; |
| 4023 | AVVulkanDeviceContext *hwctx = &p->p; |
| 4024 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 4025 | |
| 4026 | #ifdef _WIN32 |
| 4027 | CUDA_EXTERNAL_MEMORY_HANDLE_DESC ext_desc = { |
| 4028 | .type = IsWindows8OrGreater() |
| 4029 | ? CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32 |
| 4030 | : CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT, |
| 4031 | .size = size, |
| 4032 | }; |
| 4033 | VkMemoryGetWin32HandleInfoKHR export_info = { |
| 4034 | .sType = VK_STRUCTURE_TYPE_MEMORY_GET_WIN32_HANDLE_INFO_KHR, |
| 4035 | .memory = mem, |
| 4036 | .handleType = IsWindows8OrGreater() |
| 4037 | ? VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT |
| 4038 | : VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT, |
| 4039 | }; |
| 4040 | |
| 4041 | ret = vk->GetMemoryWin32HandleKHR(hwctx->act_dev, &export_info, |
| 4042 | &ext_desc.handle.win32.handle); |
| 4043 | if (ret != VK_SUCCESS) { |
| 4044 | av_log(ctx, AV_LOG_ERROR16, "Unable to export the image as a Win32 Handle: %s!\n", |
| 4045 | ff_vk_ret2str(ret)); |
| 4046 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4047 | } |
| 4048 | dst_int->ext_mem_handle[idx] = ext_desc.handle.win32.handle; |
| 4049 | #else |
| 4050 | CUDA_EXTERNAL_MEMORY_HANDLE_DESC ext_desc = { |
| 4051 | .type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD, |
| 4052 | .size = size, |
| 4053 | }; |
| 4054 | VkMemoryGetFdInfoKHR export_info = { |
| 4055 | .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR, |
| 4056 | .memory = mem, |
| 4057 | .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR, |
| 4058 | }; |
| 4059 | |
| 4060 | ret = vk->GetMemoryFdKHR(hwctx->act_dev, &export_info, |
| 4061 | &ext_desc.handle.fd); |
| 4062 | if (ret != VK_SUCCESS) { |
| 4063 | av_log(ctx, AV_LOG_ERROR16, "Unable to export the image as a FD: %s!\n", |
| 4064 | ff_vk_ret2str(ret)); |
| 4065 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4066 | } |
| 4067 | #endif |
| 4068 | |
| 4069 | ret = CHECK_CU(cu->cuImportExternalMemory(&dst_int->ext_mem[idx], &ext_desc)); |
| 4070 | if (ret < 0) { |
| 4071 | #ifndef _WIN32 |
| 4072 | close(ext_desc.handle.fd); |
| 4073 | #endif |
| 4074 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4075 | } |
| 4076 | |
| 4077 | return 0; |
| 4078 | } |
| 4079 | |
| 4080 | static int export_sem_to_cuda(AVHWDeviceContext *ctx, |
| 4081 | AVHWDeviceContext *cuda_cu, CudaFunctions *cu, |
| 4082 | AVVkFrameInternal *dst_int, int idx, |
| 4083 | VkSemaphore sem) |
| 4084 | { |
| 4085 | VkResult ret; |
| 4086 | VulkanDevicePriv *p = ctx->hwctx; |
| 4087 | AVVulkanDeviceContext *hwctx = &p->p; |
| 4088 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 4089 | |
| 4090 | #ifdef _WIN32 |
| 4091 | VkSemaphoreGetWin32HandleInfoKHR sem_export = { |
| 4092 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_WIN32_HANDLE_INFO_KHR, |
| 4093 | .semaphore = sem, |
| 4094 | .handleType = IsWindows8OrGreater() |
| 4095 | ? VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_BIT |
| 4096 | : VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT, |
| 4097 | }; |
| 4098 | CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = { |
| 4099 | .type = 10 /* TODO: CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_WIN32 */, |
| 4100 | }; |
| 4101 | #else |
| 4102 | VkSemaphoreGetFdInfoKHR sem_export = { |
| 4103 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR, |
| 4104 | .semaphore = sem, |
| 4105 | .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT, |
| 4106 | }; |
| 4107 | CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = { |
| 4108 | .type = 9 /* TODO: CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_FD */, |
| 4109 | }; |
| 4110 | #endif |
| 4111 | |
| 4112 | #ifdef _WIN32 |
| 4113 | ret = vk->GetSemaphoreWin32HandleKHR(hwctx->act_dev, &sem_export, |
| 4114 | &ext_sem_desc.handle.win32.handle); |
| 4115 | #else |
| 4116 | ret = vk->GetSemaphoreFdKHR(hwctx->act_dev, &sem_export, |
| 4117 | &ext_sem_desc.handle.fd); |
| 4118 | #endif |
| 4119 | if (ret != VK_SUCCESS) { |
| 4120 | av_log(ctx, AV_LOG_ERROR16, "Failed to export semaphore: %s\n", |
| 4121 | ff_vk_ret2str(ret)); |
| 4122 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4123 | } |
| 4124 | #ifdef _WIN32 |
| 4125 | dst_int->ext_sem_handle[idx] = ext_sem_desc.handle.win32.handle; |
| 4126 | #endif |
| 4127 | |
| 4128 | ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem[idx], |
| 4129 | &ext_sem_desc)); |
| 4130 | if (ret < 0) { |
| 4131 | #ifndef _WIN32 |
| 4132 | close(ext_sem_desc.handle.fd); |
| 4133 | #endif |
| 4134 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4135 | } |
| 4136 | |
| 4137 | return 0; |
| 4138 | } |
| 4139 | |
| 4140 | static int vulkan_export_to_cuda(AVHWFramesContext *hwfc, |
| 4141 | AVBufferRef *cuda_hwfc, |
| 4142 | const AVFrame *frame) |
| 4143 | { |
| 4144 | int err; |
| 4145 | VkResult ret; |
| 4146 | AVVkFrame *dst_f; |
| 4147 | AVVkFrameInternal *dst_int; |
| 4148 | AVHWDeviceContext *ctx = hwfc->device_ctx; |
| 4149 | const int planes = av_pix_fmt_count_planes(hwfc->sw_format); |
| 4150 | const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format); |
| 4151 | VulkanDevicePriv *p = ctx->hwctx; |
| 4152 | AVVulkanDeviceContext *hwctx = &p->p; |
| 4153 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 4154 | int nb_images; |
| 4155 | |
| 4156 | AVHWFramesContext *cuda_fc = (AVHWFramesContext*)cuda_hwfc->data; |
| 4157 | AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx; |
| 4158 | AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx; |
| 4159 | AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal; |
| 4160 | CudaFunctions *cu = cu_internal->cuda_dl; |
| 4161 | CUarray_format cufmt = desc->comp[0].depth > 8 ? CU_AD_FORMAT_UNSIGNED_INT16 : |
| 4162 | CU_AD_FORMAT_UNSIGNED_INT8; |
| 4163 | const int elem_size = 1 + (desc->comp[0].depth > 8); |
| 4164 | |
| 4165 | dst_f = (AVVkFrame *)frame->data[0]; |
| 4166 | dst_int = dst_f->internal; |
| 4167 | |
| 4168 | if (!dst_int->cuda_fc_ref) { |
| 4169 | size_t offsets[3] = { 0 }; |
| 4170 | |
| 4171 | dst_int->cuda_fc_ref = av_buffer_ref(cuda_hwfc); |
| 4172 | if (!dst_int->cuda_fc_ref) |
| 4173 | return AVERROR(ENOMEM)(-(12)); |
| 4174 | |
| 4175 | nb_images = ff_vk_count_images(dst_f); |
| 4176 | for (int i = 0; i < nb_images; i++) { |
| 4177 | err = export_mem_to_cuda(ctx, cuda_cu, cu, dst_int, i, |
| 4178 | dst_f->mem[i], dst_f->size[i]); |
| 4179 | if (err < 0) |
| 4180 | goto fail; |
| 4181 | |
| 4182 | err = export_sem_to_cuda(ctx, cuda_cu, cu, dst_int, i, |
| 4183 | dst_f->sem[i]); |
| 4184 | if (err < 0) |
| 4185 | goto fail; |
| 4186 | } |
| 4187 | |
| 4188 | if (nb_images != planes) { |
| 4189 | for (int i = 0; i < planes; i++) { |
| 4190 | /* Cuda now defines array formats for semi-planar, but these are |
| 4191 | * not currently supported for imported Vulkan images. */ |
| 4192 | if (desc->comp[i].step / elem_size > 1) { |
| 4193 | av_log(ctx, AV_LOG_ERROR16, |
| 4194 | "Cannot map a multiplane Vulkan image (%d image(s) " |
| 4195 | "for %d plane(s)) to CUDA; create the Vulkan device " |
| 4196 | "with the disable_multiplane=1 option (one image per " |
| 4197 | "plane) for CUDA interop.\n", nb_images, planes); |
| 4198 | err = AVERROR(ENOSYS)(-(38)); |
| 4199 | goto fail; |
| 4200 | } |
| 4201 | |
| 4202 | VkImageSubresource subres = { |
| 4203 | .aspectMask = i == 2 ? VK_IMAGE_ASPECT_MEMORY_PLANE_2_BIT_EXT : |
| 4204 | i == 1 ? VK_IMAGE_ASPECT_MEMORY_PLANE_1_BIT_EXT : |
| 4205 | VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT |
| 4206 | }; |
| 4207 | VkSubresourceLayout layout = { 0 }; |
| 4208 | vk->GetImageSubresourceLayout(hwctx->act_dev, dst_f->img[FFMIN(i, nb_images - 1)((i) > (nb_images - 1) ? (nb_images - 1) : (i))], |
| 4209 | &subres, &layout); |
| 4210 | offsets[i] = layout.offset; |
| 4211 | } |
| 4212 | } |
| 4213 | |
| 4214 | for (int i = 0; i < planes; i++) { |
| 4215 | CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC tex_desc = { |
| 4216 | .offset = offsets[i], |
| 4217 | .arrayDesc = { |
| 4218 | .Depth = 0, |
| 4219 | .Format = cufmt, |
| 4220 | .NumChannels = desc->comp[i].step / elem_size, |
| 4221 | .Flags = 0, |
| 4222 | }, |
| 4223 | .numLevels = 1, |
| 4224 | }; |
| 4225 | int p_w, p_h; |
| 4226 | |
| 4227 | get_plane_wh(&p_w, &p_h, hwfc->sw_format, hwfc->width, hwfc->height, i); |
| 4228 | tex_desc.arrayDesc.Width = p_w; |
| 4229 | tex_desc.arrayDesc.Height = p_h; |
| 4230 | |
| 4231 | ret = CHECK_CU(cu->cuExternalMemoryGetMappedMipmappedArray(&dst_int->cu_mma[i], |
| 4232 | dst_int->ext_mem[FFMIN(i, nb_images - 1)((i) > (nb_images - 1) ? (nb_images - 1) : (i))], |
| 4233 | &tex_desc)); |
| 4234 | if (ret < 0) { |
| 4235 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4236 | goto fail; |
| 4237 | } |
| 4238 | |
| 4239 | ret = CHECK_CU(cu->cuMipmappedArrayGetLevel(&dst_int->cu_array[i], |
| 4240 | dst_int->cu_mma[i], 0)); |
| 4241 | if (ret < 0) { |
| 4242 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4243 | goto fail; |
| 4244 | } |
| 4245 | |
| 4246 | } |
| 4247 | } |
| 4248 | |
| 4249 | return 0; |
| 4250 | |
| 4251 | fail: |
| 4252 | vulkan_free_internal(p, dst_f); |
| 4253 | return err; |
| 4254 | } |
| 4255 | |
| 4256 | static int vulkan_transfer_data_from_cuda(AVHWFramesContext *hwfc, |
| 4257 | AVFrame *dst, const AVFrame *src) |
| 4258 | { |
| 4259 | int err; |
| 4260 | CUcontext dummy; |
| 4261 | AVVkFrame *dst_f; |
| 4262 | AVVkFrameInternal *dst_int; |
| 4263 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 4264 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 4265 | const int planes = av_pix_fmt_count_planes(hwfc->sw_format); |
| 4266 | const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format); |
| 4267 | int nb_images; |
| 4268 | |
| 4269 | AVHWFramesContext *cuda_fc = (AVHWFramesContext*)src->hw_frames_ctx->data; |
| 4270 | AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx; |
| 4271 | AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx; |
| 4272 | AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal; |
| 4273 | CudaFunctions *cu = cu_internal->cuda_dl; |
| 4274 | CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS8] = { 0 }; |
| 4275 | CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS8] = { 0 }; |
| 4276 | |
| 4277 | dst_f = (AVVkFrame *)dst->data[0]; |
| 4278 | nb_images = ff_vk_count_images(dst_f); |
| 4279 | |
| 4280 | err = prepare_frame(hwfc, &fp->upload_exec, dst_f, PREP_MODE_EXTERNAL_EXPORT); |
| 4281 | if (err < 0) |
| 4282 | return err; |
| 4283 | |
| 4284 | err = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx)); |
| 4285 | if (err < 0) |
| 4286 | return err; |
| 4287 | |
| 4288 | err = vulkan_export_to_cuda(hwfc, src->hw_frames_ctx, dst); |
| 4289 | if (err < 0) { |
| 4290 | CHECK_CU(cu->cuCtxPopCurrent(&dummy)); |
| 4291 | return err; |
| 4292 | } |
| 4293 | |
| 4294 | dst_int = dst_f->internal; |
| 4295 | |
| 4296 | for (int i = 0; i < nb_images; i++) { |
| 4297 | s_w_par[i].params.fence.value = dst_f->sem_value[i] + 0; |
| 4298 | s_s_par[i].params.fence.value = dst_f->sem_value[i] + 1; |
| 4299 | } |
| 4300 | |
| 4301 | err = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par, |
| 4302 | nb_images, cuda_dev->stream)); |
| 4303 | if (err < 0) |
| 4304 | goto fail; |
| 4305 | |
| 4306 | for (int i = 0; i < planes; i++) { |
| 4307 | CUDA_MEMCPY2D cpy = { |
| 4308 | .srcMemoryType = CU_MEMORYTYPE_DEVICE, |
| 4309 | .srcDevice = (CUdeviceptr)src->data[i], |
| 4310 | .srcPitch = src->linesize[i], |
| 4311 | .srcY = 0, |
| 4312 | |
| 4313 | .dstMemoryType = CU_MEMORYTYPE_ARRAY, |
| 4314 | .dstArray = dst_int->cu_array[i], |
| 4315 | }; |
| 4316 | |
| 4317 | int p_w, p_h; |
| 4318 | get_plane_wh(&p_w, &p_h, hwfc->sw_format, hwfc->width, hwfc->height, i); |
| 4319 | |
| 4320 | cpy.WidthInBytes = p_w * desc->comp[i].step; |
| 4321 | cpy.Height = p_h; |
| 4322 | |
| 4323 | err = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream)); |
| 4324 | if (err < 0) |
| 4325 | goto fail; |
| 4326 | } |
| 4327 | |
| 4328 | err = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par, |
| 4329 | nb_images, cuda_dev->stream)); |
| 4330 | if (err < 0) |
| 4331 | goto fail; |
| 4332 | |
| 4333 | for (int i = 0; i < nb_images; i++) |
| 4334 | dst_f->sem_value[i]++; |
| 4335 | |
| 4336 | CHECK_CU(cu->cuCtxPopCurrent(&dummy)); |
| 4337 | |
| 4338 | av_log(hwfc, AV_LOG_VERBOSE40, "Transferred CUDA image to Vulkan!\n"); |
| 4339 | |
| 4340 | return err = prepare_frame(hwfc, &fp->upload_exec, dst_f, PREP_MODE_EXTERNAL_IMPORT); |
| 4341 | |
| 4342 | fail: |
| 4343 | CHECK_CU(cu->cuCtxPopCurrent(&dummy)); |
| 4344 | vulkan_free_internal(p, dst_f); |
| 4345 | av_buffer_unref(&dst->buf[0]); |
| 4346 | return err; |
| 4347 | } |
| 4348 | #endif |
| 4349 | |
| 4350 | static int vulkan_map_to(AVHWFramesContext *hwfc, AVFrame *dst, |
| 4351 | const AVFrame *src, int flags) |
| 4352 | { |
| 4353 | av_unused__attribute__((unused)) VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 4354 | |
| 4355 | switch (src->format) { |
| 4356 | #if CONFIG_LIBDRM1 |
| 4357 | #if CONFIG_VAAPI1 |
| 4358 | case AV_PIX_FMT_VAAPI: |
| 4359 | if (p->vkctx.extensions & FF_VK_EXT_DRM_MODIFIER_FLAGS(1ULL << 1)) |
| 4360 | return vulkan_map_from_vaapi(hwfc, dst, src, flags); |
| 4361 | else |
| 4362 | return AVERROR(ENOSYS)(-(38)); |
| 4363 | #endif |
| 4364 | case AV_PIX_FMT_DRM_PRIME: |
| 4365 | if (p->vkctx.extensions & FF_VK_EXT_DRM_MODIFIER_FLAGS(1ULL << 1)) |
| 4366 | return vulkan_map_from_drm(hwfc, dst, src, flags); |
| 4367 | else |
| 4368 | return AVERROR(ENOSYS)(-(38)); |
| 4369 | #endif |
| 4370 | default: |
| 4371 | return AVERROR(ENOSYS)(-(38)); |
| 4372 | } |
| 4373 | } |
| 4374 | |
| 4375 | #if CONFIG_LIBDRM1 |
| 4376 | typedef struct VulkanDRMMapping { |
| 4377 | AVDRMFrameDescriptor drm_desc; |
| 4378 | AVVkFrame *source; |
| 4379 | } VulkanDRMMapping; |
| 4380 | |
| 4381 | static void vulkan_unmap_to_drm(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap) |
| 4382 | { |
| 4383 | AVDRMFrameDescriptor *drm_desc = hwmap->priv; |
| 4384 | |
| 4385 | /* on unmap from DRM, make sure to import sync objects so that we are sync'd with any work that was |
| 4386 | * done on the buffer while exported. We don't know if who used the dmabuf did reads or writes, so protect against both */ |
| 4387 | vulkan_map_from_drm_frame_sync(hwfc, hwmap->source, drm_desc, AV_HWFRAME_MAP_READ | AV_HWFRAME_MAP_WRITE); |
| 4388 | |
| 4389 | for (int i = 0; i < drm_desc->nb_objects; i++) |
| 4390 | close(drm_desc->objects[i].fd); |
| 4391 | |
| 4392 | av_free(drm_desc); |
| 4393 | } |
| 4394 | |
| 4395 | static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt) |
| 4396 | { |
| 4397 | for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map)(sizeof(vulkan_drm_format_map) / sizeof((vulkan_drm_format_map )[0])); i++) |
| 4398 | if (vulkan_drm_format_map[i].vk_format == vkfmt) |
| 4399 | return vulkan_drm_format_map[i].drm_fourcc; |
| 4400 | return DRM_FORMAT_INVALID0; |
| 4401 | } |
| 4402 | |
| 4403 | #define MAX_MEMORY_PLANES4 4 |
| 4404 | static VkImageAspectFlags plane_index_to_aspect(int plane) { |
| 4405 | if (plane == 0) return VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT; |
| 4406 | if (plane == 1) return VK_IMAGE_ASPECT_MEMORY_PLANE_1_BIT_EXT; |
| 4407 | if (plane == 2) return VK_IMAGE_ASPECT_MEMORY_PLANE_2_BIT_EXT; |
| 4408 | if (plane == 3) return VK_IMAGE_ASPECT_MEMORY_PLANE_3_BIT_EXT; |
| 4409 | |
| 4410 | av_assert2 (0 && "Invalid plane index")do { if (!(0 && "Invalid plane index")) { av_log(((void *)0), 0, "Assertion %s failed at %s:%d\n", "0 && \"Invalid plane index\"" , "/root/firefox-clang/media/ffvpx/libavutil/hwcontext_vulkan.c" , 4410); abort(); } } while (0); |
| 4411 | return VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT; |
| 4412 | } |
| 4413 | |
| 4414 | #ifdef DMA_BUF_IOCTL_EXPORT_SYNC_FILE(((2U|1U) << (((0 +8)+8)+14)) | ((('b')) << (0 +8 )) | (((2)) << 0) | ((((sizeof(struct dma_buf_export_sync_file )))) << ((0 +8)+8))) |
| 4415 | static int vulkan_drm_export_sync_fd(AVHWFramesContext *hwfc, AVVkFrame *f, |
| 4416 | VulkanFramesPriv *fp, int nb_sems) |
| 4417 | { |
| 4418 | int sync_fd = -1; |
| 4419 | VkResult ret; |
| 4420 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 4421 | AVVulkanDeviceContext *hwctx = &p->p; |
| 4422 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 4423 | |
| 4424 | if (f->internal->drm_sync_sem == VK_NULL_HANDLE((void*)0)) { |
| 4425 | VkExportSemaphoreCreateInfo exp_info = { |
| 4426 | .sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO, |
| 4427 | .handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT, |
| 4428 | }; |
| 4429 | VkSemaphoreTypeCreateInfo type_info = { |
| 4430 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_TYPE_CREATE_INFO, |
| 4431 | .pNext = &exp_info, |
| 4432 | .semaphoreType = VK_SEMAPHORE_TYPE_BINARY, |
| 4433 | }; |
| 4434 | VkSemaphoreCreateInfo sem_create = { |
| 4435 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO, |
| 4436 | .pNext = &type_info, |
| 4437 | }; |
| 4438 | ret = vk->CreateSemaphore(hwctx->act_dev, &sem_create, hwctx->alloc, |
| 4439 | &f->internal->drm_sync_sem); |
| 4440 | if (ret != VK_SUCCESS) { |
| 4441 | av_log(hwctx, AV_LOG_ERROR16, "Failed to create DRM export semaphore: %s\n", |
| 4442 | ff_vk_ret2str(ret)); |
| 4443 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4444 | } |
| 4445 | } |
| 4446 | |
| 4447 | /* Submit a lightweight exec that waits on the timeline semaphore |
| 4448 | * (true last operation on the frame) and signals the binary semaphore, |
| 4449 | * so any Vulkan frame can get a SYNC_FD regardless of origin. */ |
| 4450 | FFVkExecContext *exec = ff_vk_exec_get(&p->vkctx, &fp->compute_exec); |
| 4451 | if (ff_vk_exec_start(&p->vkctx, exec) >= 0) { |
| 4452 | for (int i = 0; i < nb_sems; i++) |
| 4453 | ff_vk_exec_add_dep_wait_sem(&p->vkctx, exec, f->sem[i], |
| 4454 | f->sem_value[i], |
| 4455 | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT); |
| 4456 | ff_vk_exec_add_dep_bool_sem(&p->vkctx, exec, &f->internal->drm_sync_sem, 1, |
| 4457 | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, 0); |
| 4458 | if (ff_vk_exec_submit(&p->vkctx, exec) >= 0) { |
| 4459 | VkSemaphoreGetFdInfoKHR get_fd_info = { |
| 4460 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR, |
| 4461 | .semaphore = f->internal->drm_sync_sem, |
| 4462 | .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT, |
| 4463 | }; |
| 4464 | ret = vk->GetSemaphoreFdKHR(hwctx->act_dev, &get_fd_info, &sync_fd); |
| 4465 | if (ret != VK_SUCCESS) { |
| 4466 | av_log(hwctx, AV_LOG_WARNING24, |
| 4467 | "Failed to get sync fd from DRM map export semaphore: %s\n", |
| 4468 | ff_vk_ret2str(ret)); |
| 4469 | sync_fd = -1; |
| 4470 | } |
| 4471 | } |
| 4472 | } |
| 4473 | |
| 4474 | return sync_fd; |
| 4475 | } |
| 4476 | #endif |
| 4477 | |
| 4478 | static int vulkan_map_to_drm(AVHWFramesContext *hwfc, AVFrame *dst, |
| 4479 | const AVFrame *src, int flags) |
| 4480 | { |
| 4481 | int err = 0; |
| 4482 | VkResult ret; |
| 4483 | AVVkFrame *f = (AVVkFrame *)src->data[0]; |
| 4484 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 4485 | AVVulkanDeviceContext *hwctx = &p->p; |
| 4486 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 4487 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 4488 | const int planes = av_pix_fmt_count_planes(hwfc->sw_format); |
| 4489 | const int nb_images = ff_vk_count_images(f); |
| 4490 | VkImageDrmFormatModifierPropertiesEXT drm_mod = { |
| 4491 | .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_PROPERTIES_EXT, |
| 4492 | }; |
| 4493 | const int nb_sems = nb_images; |
| 4494 | int free_drm_desc_on_err = 1; |
| 4495 | int sync_fd = -1; |
| 4496 | |
| 4497 | AVDRMFrameDescriptor *drm_desc = av_mallocz(sizeof(*drm_desc)); |
| 4498 | if (!drm_desc) |
| 4499 | return AVERROR(ENOMEM)(-(12)); |
| 4500 | |
| 4501 | err = prepare_frame(hwfc, &fp->compute_exec, f, PREP_MODE_EXTERNAL_EXPORT); |
| 4502 | if (err < 0) |
| 4503 | goto end; |
| 4504 | |
| 4505 | #ifdef DMA_BUF_IOCTL_EXPORT_SYNC_FILE(((2U|1U) << (((0 +8)+8)+14)) | ((('b')) << (0 +8 )) | (((2)) << 0) | ((((sizeof(struct dma_buf_export_sync_file )))) << ((0 +8)+8))) |
| 4506 | if ((p->vkctx.extensions & FF_VK_EXT_EXTERNAL_FD_SEM(1ULL << 3)) && |
| 4507 | f->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT && |
| 4508 | vk->GetSemaphoreFdKHR && vk->CreateSemaphore) { |
| 4509 | err = vulkan_drm_export_sync_fd(hwfc, f, fp, nb_sems); |
| 4510 | if (err < 0) |
| 4511 | goto end; |
| 4512 | sync_fd = err; |
| 4513 | err = 0; |
| 4514 | } |
| 4515 | #endif |
| 4516 | |
| 4517 | if (sync_fd < 0) { |
| 4518 | VkSemaphoreWaitInfo wait_info = { |
| 4519 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_WAIT_INFO, |
| 4520 | .flags = 0x0, |
| 4521 | .semaphoreCount = nb_sems, |
| 4522 | .pSemaphores = f->sem, |
| 4523 | .pValues = f->sem_value, |
| 4524 | }; |
| 4525 | vk->WaitSemaphores(hwctx->act_dev, &wait_info, UINT64_MAX(18446744073709551615UL)); |
| 4526 | } |
| 4527 | |
| 4528 | err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src, &vulkan_unmap_to_drm, drm_desc); |
| 4529 | if (err < 0) |
| 4530 | goto end; |
| 4531 | |
| 4532 | /* It will be freed in ff_hwframe_map_create callback */ |
| 4533 | free_drm_desc_on_err = 0; |
| 4534 | |
| 4535 | ret = vk->GetImageDrmFormatModifierPropertiesEXT(hwctx->act_dev, f->img[0], |
| 4536 | &drm_mod); |
| 4537 | if (ret != VK_SUCCESS) { |
| 4538 | av_log(hwfc, AV_LOG_ERROR16, "Failed to retrieve DRM format modifier!\n"); |
| 4539 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4540 | goto end; |
| 4541 | } |
| 4542 | |
| 4543 | for (int i = 0; (i < planes) && (f->mem[i]); i++) { |
| 4544 | VkMemoryGetFdInfoKHR export_info = { |
| 4545 | .sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR, |
| 4546 | .memory = f->mem[i], |
| 4547 | .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT, |
| 4548 | }; |
| 4549 | |
| 4550 | ret = vk->GetMemoryFdKHR(hwctx->act_dev, &export_info, |
| 4551 | &drm_desc->objects[i].fd); |
| 4552 | if (ret != VK_SUCCESS) { |
| 4553 | av_log(hwfc, AV_LOG_ERROR16, "Unable to export the image as a FD!\n"); |
| 4554 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4555 | goto end; |
| 4556 | } |
| 4557 | |
| 4558 | #ifdef DMA_BUF_IOCTL_IMPORT_SYNC_FILE(((1U) << (((0 +8)+8)+14)) | ((('b')) << (0 +8)) | (((3)) << 0) | ((((sizeof(struct dma_buf_import_sync_file )))) << ((0 +8)+8))) |
| 4559 | if (sync_fd >= 0) { |
| 4560 | int dup_fd = dup(sync_fd); |
| 4561 | if (dup_fd >= 0) { |
| 4562 | struct dma_buf_import_sync_file import_info = { |
| 4563 | .flags = DMA_BUF_SYNC_WRITE(2 << 0), |
| 4564 | .fd = dup_fd, |
| 4565 | }; |
| 4566 | if (ioctl(drm_desc->objects[i].fd, DMA_BUF_IOCTL_IMPORT_SYNC_FILE(((1U) << (((0 +8)+8)+14)) | ((('b')) << (0 +8)) | (((3)) << 0) | ((((sizeof(struct dma_buf_import_sync_file )))) << ((0 +8)+8))), &import_info) < 0) |
| 4567 | av_log(hwfc, AV_LOG_WARNING24, "DMA_BUF_IOCTL_IMPORT_SYNC_FILE failed: %s\n", av_err2str(AVERROR(errno))av_make_error_string((char[64]){0}, 64, (-((*__errno_location ()))))); |
| 4568 | close(dup_fd); |
| 4569 | } else { |
| 4570 | av_log(hwfc, AV_LOG_WARNING24, "dup(sync_fd) failed: %s\n", av_err2str(AVERROR(errno))av_make_error_string((char[64]){0}, 64, (-((*__errno_location ()))))); |
| 4571 | } |
| 4572 | } |
| 4573 | #endif |
| 4574 | |
| 4575 | drm_desc->nb_objects++; |
| 4576 | drm_desc->objects[i].size = f->size[i]; |
| 4577 | drm_desc->objects[i].format_modifier = drm_mod.drmFormatModifier; |
| 4578 | } |
| 4579 | |
| 4580 | /* NV12 has 2 planes but 1 image/semaphore */ |
| 4581 | drm_desc->nb_layers = FFMAX(planes, nb_images)((planes) > (nb_images) ? (planes) : (nb_images)); |
| 4582 | for (int i = 0; i < drm_desc->nb_layers; i++) { |
| 4583 | VkFormat plane_vkfmt = av_vkfmt_from_pixfmt(hwfc->sw_format)[i]; |
| 4584 | |
| 4585 | drm_desc->layers[i].format = vulkan_fmt_to_drm(plane_vkfmt); |
| 4586 | drm_desc->layers[i].nb_planes = fp->drm_format_modifier_properties[i].drmFormatModifierPlaneCount; |
| 4587 | |
| 4588 | if (drm_desc->layers[i].nb_planes > MAX_MEMORY_PLANES4) { |
| 4589 | av_log(hwfc, AV_LOG_ERROR16, "Too many memory planes for DRM format!\n"); |
| 4590 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4591 | goto end; |
| 4592 | } |
| 4593 | |
| 4594 | for (int j = 0; j < drm_desc->layers[i].nb_planes; j++) { |
| 4595 | VkSubresourceLayout layout; |
| 4596 | int aspect_plane = (nb_images == 1) ? i : j; |
| 4597 | VkImageSubresource sub = { |
| 4598 | .aspectMask = plane_index_to_aspect(aspect_plane), |
| 4599 | }; |
| 4600 | |
| 4601 | drm_desc->layers[i].planes[j].object_index = FFMIN(i, drm_desc->nb_objects - 1)((i) > (drm_desc->nb_objects - 1) ? (drm_desc->nb_objects - 1) : (i)); |
| 4602 | |
| 4603 | vk->GetImageSubresourceLayout(hwctx->act_dev, f->img[FFMIN(i, nb_images - 1)((i) > (nb_images - 1) ? (nb_images - 1) : (i))], &sub, &layout); |
| 4604 | drm_desc->layers[i].planes[j].offset = layout.offset; |
| 4605 | drm_desc->layers[i].planes[j].pitch = layout.rowPitch; |
| 4606 | } |
| 4607 | |
| 4608 | if (drm_desc->layers[i].format == DRM_FORMAT_INVALID0) { |
| 4609 | av_log(hwfc, AV_LOG_ERROR16, "Cannot map to DRM layer, unsupported!\n"); |
| 4610 | err = AVERROR_PATCHWELCOME(-(int)(('P') | (('A') << 8) | (('W') << 16) | (( unsigned)('E') << 24))); |
| 4611 | goto end; |
| 4612 | } |
| 4613 | |
| 4614 | |
| 4615 | if (f->tiling == VK_IMAGE_TILING_OPTIMAL) |
| 4616 | continue; |
| 4617 | |
| 4618 | } |
| 4619 | |
| 4620 | dst->width = src->width; |
| 4621 | dst->height = src->height; |
| 4622 | dst->data[0] = (uint8_t *)drm_desc; |
| 4623 | dst->hw_frames_ctx = av_buffer_ref(src->hw_frames_ctx); |
| 4624 | |
| 4625 | if (sync_fd >= 0) |
| 4626 | close(sync_fd); |
| 4627 | |
| 4628 | av_log(hwfc, AV_LOG_VERBOSE40, "Mapped AVVkFrame to a DRM object!\n"); |
| 4629 | |
| 4630 | return 0; |
| 4631 | |
| 4632 | end: |
| 4633 | for (int i = 0; i < drm_desc->nb_objects; i++) |
| 4634 | close(drm_desc->objects[i].fd); |
| 4635 | if (free_drm_desc_on_err) |
| 4636 | av_free(drm_desc); |
| 4637 | if (sync_fd >= 0) |
| 4638 | close(sync_fd); |
| 4639 | return err; |
| 4640 | } |
| 4641 | |
| 4642 | #if CONFIG_VAAPI1 |
| 4643 | static int vulkan_map_to_vaapi(AVHWFramesContext *hwfc, AVFrame *dst, |
| 4644 | const AVFrame *src, int flags) |
| 4645 | { |
| 4646 | int err; |
| 4647 | AVFrame *tmp = av_frame_alloc(); |
| 4648 | if (!tmp) |
| 4649 | return AVERROR(ENOMEM)(-(12)); |
| 4650 | |
| 4651 | tmp->format = AV_PIX_FMT_DRM_PRIME; |
| 4652 | |
| 4653 | err = vulkan_map_to_drm(hwfc, tmp, src, flags); |
| 4654 | if (err < 0) |
| 4655 | goto fail; |
| 4656 | |
| 4657 | err = av_hwframe_map(dst, tmp, flags); |
| 4658 | if (err < 0) |
| 4659 | goto fail; |
| 4660 | |
| 4661 | err = ff_hwframe_map_replace(dst, src); |
| 4662 | |
| 4663 | fail: |
| 4664 | av_frame_free(&tmp); |
| 4665 | return err; |
| 4666 | } |
| 4667 | #endif |
| 4668 | #endif |
| 4669 | |
| 4670 | static int vulkan_map_from(AVHWFramesContext *hwfc, AVFrame *dst, |
| 4671 | const AVFrame *src, int flags) |
| 4672 | { |
| 4673 | av_unused__attribute__((unused)) VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 4674 | |
| 4675 | switch (dst->format) { |
| 4676 | #if CONFIG_LIBDRM1 |
| 4677 | case AV_PIX_FMT_DRM_PRIME: |
| 4678 | if (p->vkctx.extensions & FF_VK_EXT_DRM_MODIFIER_FLAGS(1ULL << 1)) |
| 4679 | return vulkan_map_to_drm(hwfc, dst, src, flags); |
| 4680 | else |
| 4681 | return AVERROR(ENOSYS)(-(38)); |
| 4682 | #if CONFIG_VAAPI1 |
| 4683 | case AV_PIX_FMT_VAAPI: |
| 4684 | if (p->vkctx.extensions & FF_VK_EXT_DRM_MODIFIER_FLAGS(1ULL << 1)) |
| 4685 | return vulkan_map_to_vaapi(hwfc, dst, src, flags); |
| 4686 | else |
| 4687 | return AVERROR(ENOSYS)(-(38)); |
| 4688 | #endif |
| 4689 | #endif |
| 4690 | default: |
| 4691 | break; |
| 4692 | } |
| 4693 | return AVERROR(ENOSYS)(-(38)); |
| 4694 | } |
| 4695 | |
| 4696 | static int copy_buffer_data(AVHWFramesContext *hwfc, FFVkBuffer *vkbuf, |
| 4697 | AVFrame *swf, VkBufferImageCopy *region, |
| 4698 | int planes, int upload) |
| 4699 | { |
| 4700 | int err; |
| 4701 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 4702 | |
| 4703 | if (upload) { |
| 4704 | for (int i = 0; i < planes; i++) |
| 4705 | av_image_copy_plane(vkbuf->mapped_mem + region[i].bufferOffset, |
| 4706 | region[i].bufferRowLength, |
| 4707 | swf->data[i], |
| 4708 | swf->linesize[i], |
| 4709 | swf->linesize[i], |
| 4710 | region[i].imageExtent.height); |
| 4711 | |
| 4712 | err = ff_vk_flush_buffer(&p->vkctx, vkbuf, 0, VK_WHOLE_SIZE(~0ULL), 1); |
| 4713 | if (err != VK_SUCCESS) { |
| 4714 | av_log(hwfc, AV_LOG_ERROR16, "Failed to flush buffer data: %s\n", |
| 4715 | av_err2str(err)av_make_error_string((char[64]){0}, 64, err)); |
| 4716 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4717 | } |
| 4718 | } else { |
| 4719 | err = ff_vk_flush_buffer(&p->vkctx, vkbuf, 0, VK_WHOLE_SIZE(~0ULL), 0); |
| 4720 | if (err != VK_SUCCESS) { |
| 4721 | av_log(hwfc, AV_LOG_ERROR16, "Failed to invalidate buffer data: %s\n", |
| 4722 | av_err2str(err)av_make_error_string((char[64]){0}, 64, err)); |
| 4723 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); |
| 4724 | } |
| 4725 | |
| 4726 | for (int i = 0; i < planes; i++) |
| 4727 | av_image_copy_plane(swf->data[i], |
| 4728 | swf->linesize[i], |
| 4729 | vkbuf->mapped_mem + region[i].bufferOffset, |
| 4730 | region[i].bufferRowLength, |
| 4731 | swf->linesize[i], |
| 4732 | region[i].imageExtent.height); |
| 4733 | } |
| 4734 | |
| 4735 | return 0; |
| 4736 | } |
| 4737 | |
| 4738 | static int get_plane_buf(AVHWFramesContext *hwfc, FFVkBuffer **dst, |
| 4739 | AVFrame *swf, VkBufferImageCopy *region, int upload) |
| 4740 | { |
| 4741 | int err; |
| 4742 | uint32_t p_w, p_h; |
| 4743 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 4744 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 4745 | const int planes = av_pix_fmt_count_planes(swf->format); |
| 4746 | VkBufferUsageFlags buf_usage = upload ? VK_BUFFER_USAGE_TRANSFER_SRC_BIT : |
| 4747 | VK_BUFFER_USAGE_TRANSFER_DST_BIT; |
| 4748 | |
| 4749 | size_t buf_offset = 0; |
| 4750 | for (int i = 0; i < planes; i++) { |
| 4751 | get_plane_wh(&p_w, &p_h, swf->format, swf->width, swf->height, i); |
| 4752 | |
| 4753 | region[i] = (VkBufferImageCopy) { |
| 4754 | .bufferOffset = buf_offset, |
| 4755 | .bufferRowLength = FFALIGN(swf->linesize[i],(((swf->linesize[i])+(p->props.properties.limits.optimalBufferCopyRowPitchAlignment )-1)&~((p->props.properties.limits.optimalBufferCopyRowPitchAlignment )-1)) |
| 4756 | p->props.properties.limits.optimalBufferCopyRowPitchAlignment)(((swf->linesize[i])+(p->props.properties.limits.optimalBufferCopyRowPitchAlignment )-1)&~((p->props.properties.limits.optimalBufferCopyRowPitchAlignment )-1)), |
| 4757 | .bufferImageHeight = p_h, |
| 4758 | .imageSubresource.layerCount = 1, |
| 4759 | .imageExtent = (VkExtent3D){ p_w, p_h, 1 }, |
| 4760 | /* Rest of the fields adjusted/filled in later */ |
| 4761 | }; |
| 4762 | |
| 4763 | buf_offset += FFALIGN(p_h*region[i].bufferRowLength,(((p_h*region[i].bufferRowLength)+(p->props.properties.limits .optimalBufferCopyOffsetAlignment)-1)&~((p->props.properties .limits.optimalBufferCopyOffsetAlignment)-1)) |
| 4764 | p->props.properties.limits.optimalBufferCopyOffsetAlignment)(((p_h*region[i].bufferRowLength)+(p->props.properties.limits .optimalBufferCopyOffsetAlignment)-1)&~((p->props.properties .limits.optimalBufferCopyOffsetAlignment)-1)); |
| 4765 | } |
| 4766 | |
| 4767 | err = ff_vk_get_pooled_buffer(&p->vkctx, &fp->tmp, dst, buf_usage, |
| 4768 | NULL((void*)0), buf_offset, |
| 4769 | VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | |
| 4770 | p->vkctx.host_cached_flag); |
| 4771 | if (err < 0) |
| 4772 | return err; |
| 4773 | |
| 4774 | return 0; |
| 4775 | } |
| 4776 | |
| 4777 | static int host_map_frame(AVHWFramesContext *hwfc, FFVkBuffer **dst, int *nb_bufs, |
| 4778 | AVFrame *swf, VkBufferImageCopy *region, int upload) |
| 4779 | { |
| 4780 | int err; |
| 4781 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 4782 | |
| 4783 | int nb_src_bufs; |
| 4784 | const int planes = av_pix_fmt_count_planes(swf->format); |
| 4785 | VkBufferUsageFlags buf_usage = upload ? VK_BUFFER_USAGE_TRANSFER_SRC_BIT : |
| 4786 | VK_BUFFER_USAGE_TRANSFER_DST_BIT; |
| 4787 | |
| 4788 | /* We can't host map images with negative strides */ |
| 4789 | for (int i = 0; i < planes; i++) |
| 4790 | if (swf->linesize[i] < 0) |
| 4791 | return AVERROR(EINVAL)(-(22)); |
| 4792 | |
| 4793 | /* Count the number of buffers in the software frame */ |
| 4794 | nb_src_bufs = 0; |
| 4795 | while (swf->buf[nb_src_bufs]) |
| 4796 | nb_src_bufs++; |
| 4797 | |
| 4798 | /* Single buffer contains all planes */ |
| 4799 | if (nb_src_bufs == 1) { |
| 4800 | err = ff_vk_host_map_buffer(&p->vkctx, &dst[0], |
| 4801 | swf->data[0], VK_WHOLE_SIZE(~0ULL), swf->buf[0], |
| 4802 | buf_usage); |
| 4803 | if (err < 0) |
| 4804 | return err; |
| 4805 | (*nb_bufs)++; |
| 4806 | |
| 4807 | for (int i = 0; i < planes; i++) |
| 4808 | region[i].bufferOffset = dst[0]->virtual_offset + |
| 4809 | swf->data[i] - swf->data[0]; |
| 4810 | } else if (nb_src_bufs == planes) { /* One buffer per plane */ |
| 4811 | for (int i = 0; i < planes; i++) { |
| 4812 | err = ff_vk_host_map_buffer(&p->vkctx, &dst[i], |
| 4813 | swf->data[i], VK_WHOLE_SIZE(~0ULL), |
| 4814 | swf->buf[i], buf_usage); |
| 4815 | if (err < 0) |
| 4816 | goto fail; |
| 4817 | (*nb_bufs)++; |
| 4818 | |
| 4819 | region[i].bufferOffset = dst[i]->virtual_offset; |
| 4820 | } |
| 4821 | } else { |
| 4822 | /* Weird layout (3 planes, 2 buffers), patch welcome, fallback to copy */ |
| 4823 | return AVERROR_PATCHWELCOME(-(int)(('P') | (('A') << 8) | (('W') << 16) | (( unsigned)('E') << 24))); |
| 4824 | } |
| 4825 | |
| 4826 | return 0; |
| 4827 | |
| 4828 | fail: |
| 4829 | for (int i = 0; i < (*nb_bufs); i++) |
| 4830 | av_refstruct_unref(&dst[i]); |
| 4831 | return err; |
| 4832 | } |
| 4833 | |
| 4834 | static int vulkan_transfer_host(AVHWFramesContext *hwfc, AVFrame *hwf, |
| 4835 | AVFrame *swf, int upload) |
| 4836 | { |
| 4837 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 4838 | AVVulkanFramesContext *hwfc_vk = &fp->p; |
| 4839 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 4840 | AVVulkanDeviceContext *hwctx = &p->p; |
| 4841 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 4842 | |
| 4843 | AVVkFrame *hwf_vk = (AVVkFrame *)hwf->data[0]; |
| 4844 | const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(swf->format); |
| 4845 | const int planes = av_pix_fmt_count_planes(swf->format); |
| 4846 | const int nb_images = ff_vk_count_images(hwf_vk); |
| 4847 | |
| 4848 | VkSemaphoreWaitInfo sem_wait; |
| 4849 | VkHostImageLayoutTransitionInfoEXT layout_ch_info[AV_NUM_DATA_POINTERS8]; |
| 4850 | int nb_layout_ch = 0; |
| 4851 | |
| 4852 | hwfc_vk->lock_frame(hwfc, hwf_vk); |
| 4853 | |
| 4854 | for (int i = 0; i < nb_images; i++) { |
| 4855 | int compat = 0; |
| 4856 | for (int j = 0; j < p->vkctx.host_image_props.copySrcLayoutCount; j++) { |
| 4857 | if (hwf_vk->layout[i] == p->vkctx.host_image_props.pCopySrcLayouts[j]) { |
| 4858 | compat = 1; |
| 4859 | break; |
| 4860 | } |
| 4861 | } |
| 4862 | if (compat) |
| 4863 | continue; |
| 4864 | |
| 4865 | /* This should only ever happen on uploads, so using UNDEFINED is safe */ |
| 4866 | av_assert1(upload)do { if (!(upload)) { av_log(((void*)0), 0, "Assertion %s failed at %s:%d\n" , "upload", "/root/firefox-clang/media/ffvpx/libavutil/hwcontext_vulkan.c" , 4866); abort(); } } while (0); |
| 4867 | layout_ch_info[nb_layout_ch] = (VkHostImageLayoutTransitionInfoEXT) { |
| 4868 | .sType = VK_STRUCTURE_TYPE_HOST_IMAGE_LAYOUT_TRANSITION_INFO_EXT, |
| 4869 | .image = hwf_vk->img[i], |
| 4870 | .oldLayout = VK_IMAGE_LAYOUT_UNDEFINED, |
| 4871 | .newLayout = VK_IMAGE_LAYOUT_GENERAL, |
| 4872 | .subresourceRange = { |
| 4873 | .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, |
| 4874 | .levelCount = 1, |
| 4875 | .layerCount = 1, |
| 4876 | }, |
| 4877 | }; |
| 4878 | |
| 4879 | hwf_vk->layout[i] = layout_ch_info[nb_layout_ch].newLayout; |
| 4880 | nb_layout_ch++; |
| 4881 | } |
| 4882 | |
| 4883 | sem_wait = (VkSemaphoreWaitInfo) { |
| 4884 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_WAIT_INFO, |
| 4885 | .pSemaphores = hwf_vk->sem, |
| 4886 | .pValues = hwf_vk->sem_value, |
| 4887 | .semaphoreCount = nb_images, |
| 4888 | }; |
| 4889 | |
| 4890 | vk->WaitSemaphores(hwctx->act_dev, &sem_wait, UINT64_MAX(18446744073709551615UL)); |
| 4891 | |
| 4892 | if (nb_layout_ch) |
| 4893 | vk->TransitionImageLayoutEXT(hwctx->act_dev, |
| 4894 | nb_layout_ch, layout_ch_info); |
| 4895 | |
| 4896 | if (upload) { |
| 4897 | VkMemoryToImageCopyEXT region_info = { |
| 4898 | .sType = VK_STRUCTURE_TYPE_MEMORY_TO_IMAGE_COPY_EXT, |
| 4899 | .imageSubresource = { |
| 4900 | .layerCount = 1, |
| 4901 | }, |
| 4902 | }; |
| 4903 | VkCopyMemoryToImageInfoEXT copy_info = { |
| 4904 | .sType = VK_STRUCTURE_TYPE_COPY_MEMORY_TO_IMAGE_INFO_EXT, |
| 4905 | .regionCount = 1, |
| 4906 | .pRegions = ®ion_info, |
| 4907 | }; |
| 4908 | for (int i = 0; i < planes; i++) { |
| 4909 | int img_idx = FFMIN(i, (nb_images - 1))((i) > ((nb_images - 1)) ? ((nb_images - 1)) : (i)); |
| 4910 | uint32_t p_w, p_h; |
| 4911 | get_plane_wh(&p_w, &p_h, swf->format, swf->width, swf->height, i); |
| 4912 | |
| 4913 | region_info.pHostPointer = swf->data[i]; |
| 4914 | region_info.memoryRowLength = swf->linesize[i] / desc->comp[i].step; |
| 4915 | region_info.imageSubresource.aspectMask = ff_vk_aspect_flag(hwf, i); |
| 4916 | region_info.imageExtent = (VkExtent3D){ p_w, p_h, 1 }; |
| 4917 | copy_info.dstImage = hwf_vk->img[img_idx]; |
| 4918 | copy_info.dstImageLayout = hwf_vk->layout[img_idx]; |
| 4919 | |
| 4920 | vk->CopyMemoryToImageEXT(hwctx->act_dev, ©_info); |
| 4921 | } |
| 4922 | } else { |
| 4923 | VkImageToMemoryCopyEXT region_info = { |
| 4924 | .sType = VK_STRUCTURE_TYPE_IMAGE_TO_MEMORY_COPY_EXT, |
| 4925 | .imageSubresource = { |
| 4926 | .layerCount = 1, |
| 4927 | }, |
| 4928 | }; |
| 4929 | VkCopyImageToMemoryInfoEXT copy_info = { |
| 4930 | .sType = VK_STRUCTURE_TYPE_COPY_IMAGE_TO_MEMORY_INFO_EXT, |
| 4931 | .regionCount = 1, |
| 4932 | .pRegions = ®ion_info, |
| 4933 | }; |
| 4934 | for (int i = 0; i < planes; i++) { |
| 4935 | int img_idx = FFMIN(i, (nb_images - 1))((i) > ((nb_images - 1)) ? ((nb_images - 1)) : (i)); |
| 4936 | uint32_t p_w, p_h; |
| 4937 | get_plane_wh(&p_w, &p_h, swf->format, swf->width, swf->height, i); |
| 4938 | |
| 4939 | region_info.pHostPointer = swf->data[i]; |
| 4940 | region_info.memoryRowLength = swf->linesize[i] / desc->comp[i].step; |
| 4941 | region_info.imageSubresource.aspectMask = ff_vk_aspect_flag(hwf, i); |
| 4942 | region_info.imageExtent = (VkExtent3D){ p_w, p_h, 1 }; |
| 4943 | copy_info.srcImage = hwf_vk->img[img_idx]; |
| 4944 | copy_info.srcImageLayout = hwf_vk->layout[img_idx]; |
| 4945 | |
| 4946 | vk->CopyImageToMemoryEXT(hwctx->act_dev, ©_info); |
| 4947 | } |
| 4948 | } |
| 4949 | |
| 4950 | hwfc_vk->unlock_frame(hwfc, hwf_vk); |
| 4951 | |
| 4952 | return 0; |
| 4953 | } |
| 4954 | |
| 4955 | static int vulkan_transfer_frame(AVHWFramesContext *hwfc, |
| 4956 | AVFrame *swf, AVFrame *hwf, |
| 4957 | int upload) |
| 4958 | { |
| 4959 | int err; |
| 4960 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 4961 | AVVulkanFramesContext *hwctx = &fp->p; |
| 4962 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 4963 | FFVulkanFunctions *vk = &p->vkctx.vkfn; |
| 4964 | |
| 4965 | int host_mapped = 0; |
| 4966 | |
| 4967 | AVVkFrame *hwf_vk = (AVVkFrame *)hwf->data[0]; |
| 4968 | VkBufferImageCopy region[AV_NUM_DATA_POINTERS8]; // always one per plane |
| 4969 | |
| 4970 | const int planes = av_pix_fmt_count_planes(swf->format); |
| 4971 | const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(swf->format); |
| 4972 | const int nb_images = ff_vk_count_images(hwf_vk); |
| 4973 | |
| 4974 | VkImageMemoryBarrier2 img_bar[AV_NUM_DATA_POINTERS8]; |
| 4975 | int nb_img_bar = 0; |
| 4976 | |
| 4977 | FFVkBuffer *bufs[AV_NUM_DATA_POINTERS8]; |
| 4978 | int nb_bufs = 0; |
| 4979 | |
| 4980 | VkCommandBuffer cmd_buf; |
| 4981 | FFVkExecContext *exec; |
| 4982 | |
| 4983 | /* Sanity checking */ |
| 4984 | if ((swf->format != AV_PIX_FMT_NONE && !av_vkfmt_from_pixfmt(swf->format))) { |
| 4985 | av_log(hwfc, AV_LOG_ERROR16, "Unsupported software frame pixel format!\n"); |
| 4986 | return AVERROR(EINVAL)(-(22)); |
| 4987 | } |
| 4988 | |
| 4989 | if (swf->width > hwfc->width || swf->height > hwfc->height) |
| 4990 | return AVERROR(EINVAL)(-(22)); |
| 4991 | |
| 4992 | int host_copy = hwctx->usage & VK_IMAGE_USAGE_HOST_TRANSFER_BIT_EXT; |
| 4993 | |
| 4994 | /* Host layout transitions may only originate from a host-copyable layout */ |
| 4995 | if (!upload && host_copy) { |
| 4996 | for (int i = 0; i < nb_images; i++) { |
| 4997 | int compat = 0; |
| 4998 | for (int j = 0; j < p->vkctx.host_image_props.copySrcLayoutCount; j++) { |
| 4999 | if (hwf_vk->layout[i] == p->vkctx.host_image_props.pCopySrcLayouts[j]) { |
| 5000 | compat = 1; |
| 5001 | break; |
| 5002 | } |
| 5003 | } |
| 5004 | if (!compat) { |
| 5005 | host_copy = 0; |
| 5006 | break; |
| 5007 | } |
| 5008 | } |
| 5009 | } |
| 5010 | |
| 5011 | if (host_copy) |
| 5012 | return vulkan_transfer_host(hwfc, hwf, swf, upload); |
| 5013 | |
| 5014 | for (int i = 0; i < av_pix_fmt_count_planes(swf->format); i++) { |
| 5015 | uint32_t p_w, p_h; |
| 5016 | get_plane_wh(&p_w, &p_h, swf->format, swf->width, swf->height, i); |
| 5017 | |
| 5018 | /* Buffer region for this plane */ |
| 5019 | region[i] = (VkBufferImageCopy) { |
| 5020 | .bufferOffset = 0, |
| 5021 | .bufferRowLength = swf->linesize[i], |
| 5022 | .bufferImageHeight = p_h, |
| 5023 | .imageSubresource.layerCount = 1, |
| 5024 | .imageExtent = (VkExtent3D){ p_w, p_h, 1 }, |
| 5025 | /* Rest of the fields adjusted/filled in later */ |
| 5026 | }; |
| 5027 | } |
| 5028 | |
| 5029 | /* Setup buffers first */ |
| 5030 | if (p->vkctx.extensions & FF_VK_EXT_EXTERNAL_HOST_MEMORY(1ULL << 4) && !p->avoid_host_import) { |
| 5031 | err = host_map_frame(hwfc, bufs, &nb_bufs, swf, region, upload); |
| 5032 | if (err >= 0) |
| 5033 | host_mapped = 1; |
| 5034 | } |
| 5035 | |
| 5036 | if (!host_mapped) { |
| 5037 | err = get_plane_buf(hwfc, &bufs[0], swf, region, upload); |
| 5038 | if (err < 0) |
| 5039 | goto end; |
| 5040 | nb_bufs = 1; |
| 5041 | |
| 5042 | if (upload) { |
| 5043 | err = copy_buffer_data(hwfc, bufs[0], swf, region, planes, 1); |
| 5044 | if (err < 0) |
| 5045 | goto end; |
| 5046 | } |
| 5047 | } |
| 5048 | |
| 5049 | exec = ff_vk_exec_get(&p->vkctx, &fp->upload_exec); |
| 5050 | cmd_buf = exec->buf; |
| 5051 | |
| 5052 | err = ff_vk_exec_start(&p->vkctx, exec); |
| 5053 | if (err < 0) |
| 5054 | goto end; |
| 5055 | |
| 5056 | /* Prep destination Vulkan frame */ |
| 5057 | err = ff_vk_exec_add_dep_frame(&p->vkctx, exec, hwf, |
| 5058 | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, |
| 5059 | VK_PIPELINE_STAGE_2_TRANSFER_BIT); |
| 5060 | if (err < 0) { |
| 5061 | ff_vk_exec_discard(&p->vkctx, exec); |
| 5062 | goto end; |
| 5063 | } |
| 5064 | |
| 5065 | /* No need to declare buf deps for synchronous transfers (downloads) */ |
| 5066 | if (upload) { |
| 5067 | /* Add the software frame backing the buffers if we're host mapping */ |
| 5068 | if (host_mapped) { |
| 5069 | err = ff_vk_exec_add_dep_sw_frame(&p->vkctx, exec, swf); |
| 5070 | if (err < 0) { |
| 5071 | ff_vk_exec_discard(&p->vkctx, exec); |
| 5072 | goto end; |
| 5073 | } |
| 5074 | } |
| 5075 | |
| 5076 | /* Add the buffers as a dependency */ |
| 5077 | for (int i = 0; i < nb_bufs; i++) |
| 5078 | ff_vk_exec_add_dep_refstruct(&p->vkctx, exec, bufs[i]); |
| 5079 | } |
| 5080 | |
| 5081 | ff_vk_frame_barrier(&p->vkctx, exec, hwf, img_bar, &nb_img_bar, |
| 5082 | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, |
| 5083 | VK_PIPELINE_STAGE_2_TRANSFER_BIT_KHR, |
| 5084 | upload ? VK_ACCESS_TRANSFER_WRITE_BIT : |
| 5085 | VK_ACCESS_TRANSFER_READ_BIT, |
| 5086 | upload ? VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL : |
| 5087 | VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, |
| 5088 | p->nb_img_qfs > 1 ? VK_QUEUE_FAMILY_IGNORED(~0U) : p->img_qfs[0]); |
| 5089 | |
| 5090 | vk->CmdPipelineBarrier2(cmd_buf, &(VkDependencyInfo) { |
| 5091 | .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO, |
| 5092 | .pImageMemoryBarriers = img_bar, |
| 5093 | .imageMemoryBarrierCount = nb_img_bar, |
| 5094 | }); |
| 5095 | |
| 5096 | for (int i = 0; i < planes; i++) { |
| 5097 | int buf_idx = FFMIN(i, (nb_bufs - 1))((i) > ((nb_bufs - 1)) ? ((nb_bufs - 1)) : (i)); |
| 5098 | int img_idx = FFMIN(i, (nb_images - 1))((i) > ((nb_images - 1)) ? ((nb_images - 1)) : (i)); |
| 5099 | FFVkBuffer *vkbuf = bufs[buf_idx]; |
| 5100 | |
| 5101 | uint32_t orig_stride = region[i].bufferRowLength; |
| 5102 | region[i].bufferRowLength /= desc->comp[i].step; |
| 5103 | region[i].imageSubresource.aspectMask = ff_vk_aspect_flag(hwf, i); |
| 5104 | |
| 5105 | if (upload) |
| 5106 | vk->CmdCopyBufferToImage(cmd_buf, vkbuf->buf, |
| 5107 | hwf_vk->img[img_idx], |
| 5108 | img_bar[img_idx].newLayout, |
| 5109 | 1, ®ion[i]); |
| 5110 | else |
| 5111 | vk->CmdCopyImageToBuffer(cmd_buf, hwf_vk->img[img_idx], |
| 5112 | img_bar[img_idx].newLayout, |
| 5113 | vkbuf->buf, |
| 5114 | 1, ®ion[i]); |
| 5115 | |
| 5116 | region[i].bufferRowLength = orig_stride; |
| 5117 | } |
| 5118 | |
| 5119 | err = ff_vk_exec_submit(&p->vkctx, exec); |
| 5120 | if (err >= 0 && !upload) { |
| 5121 | ff_vk_exec_wait(&p->vkctx, exec); |
| 5122 | if (!host_mapped) |
| 5123 | err = copy_buffer_data(hwfc, bufs[0], swf, region, planes, 0); |
| 5124 | } |
| 5125 | |
| 5126 | end: |
| 5127 | for (int i = 0; i < nb_bufs; i++) |
| 5128 | av_refstruct_unref(&bufs[i]); |
| 5129 | |
| 5130 | return err; |
| 5131 | } |
| 5132 | |
| 5133 | static int vulkan_transfer_data_to(AVHWFramesContext *hwfc, AVFrame *dst, |
| 5134 | const AVFrame *src) |
| 5135 | { |
| 5136 | av_unused__attribute__((unused)) VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 5137 | |
| 5138 | switch (src->format) { |
| 5139 | #if CONFIG_CUDA0 |
| 5140 | case AV_PIX_FMT_CUDA: |
| 5141 | #ifdef _WIN32 |
| 5142 | if ((p->vkctx.extensions & FF_VK_EXT_EXTERNAL_WIN32_MEMORY(1ULL << 6)) && |
| 5143 | (p->vkctx.extensions & FF_VK_EXT_EXTERNAL_WIN32_SEM(1ULL << 7))) |
| 5144 | #else |
| 5145 | if ((p->vkctx.extensions & FF_VK_EXT_EXTERNAL_FD_MEMORY(1ULL << 2)) && |
| 5146 | (p->vkctx.extensions & FF_VK_EXT_EXTERNAL_FD_SEM(1ULL << 3))) |
| 5147 | #endif |
| 5148 | return vulkan_transfer_data_from_cuda(hwfc, dst, src); |
| 5149 | av_fallthrough__attribute__((fallthrough)); |
| 5150 | #endif |
| 5151 | default: |
| 5152 | if (src->hw_frames_ctx) |
| 5153 | return AVERROR(ENOSYS)(-(38)); |
| 5154 | else |
| 5155 | return vulkan_transfer_frame(hwfc, (AVFrame *)src, dst, 1); |
| 5156 | } |
| 5157 | } |
| 5158 | |
| 5159 | #if CONFIG_CUDA0 |
| 5160 | static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst, |
| 5161 | const AVFrame *src) |
| 5162 | { |
| 5163 | int err; |
| 5164 | CUcontext dummy; |
| 5165 | AVVkFrame *dst_f; |
| 5166 | AVVkFrameInternal *dst_int; |
| 5167 | VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 5168 | VulkanFramesPriv *fp = hwfc->hwctx; |
| 5169 | const int planes = av_pix_fmt_count_planes(hwfc->sw_format); |
| 5170 | const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format); |
| 5171 | int nb_images; |
| 5172 | |
| 5173 | AVHWFramesContext *cuda_fc = (AVHWFramesContext*)dst->hw_frames_ctx->data; |
| 5174 | AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx; |
| 5175 | AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx; |
| 5176 | AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal; |
| 5177 | CudaFunctions *cu = cu_internal->cuda_dl; |
| 5178 | CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS8] = { 0 }; |
| 5179 | CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS8] = { 0 }; |
| 5180 | |
| 5181 | dst_f = (AVVkFrame *)src->data[0]; |
| 5182 | nb_images = ff_vk_count_images(dst_f); |
| 5183 | |
| 5184 | err = prepare_frame(hwfc, &fp->upload_exec, dst_f, PREP_MODE_EXTERNAL_EXPORT); |
| 5185 | if (err < 0) |
| 5186 | return err; |
| 5187 | |
| 5188 | err = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx)); |
| 5189 | if (err < 0) |
| 5190 | return err; |
| 5191 | |
| 5192 | err = vulkan_export_to_cuda(hwfc, dst->hw_frames_ctx, src); |
| 5193 | if (err < 0) { |
| 5194 | CHECK_CU(cu->cuCtxPopCurrent(&dummy)); |
| 5195 | return err; |
| 5196 | } |
| 5197 | |
| 5198 | dst_int = dst_f->internal; |
| 5199 | |
| 5200 | for (int i = 0; i < nb_images; i++) { |
| 5201 | s_w_par[i].params.fence.value = dst_f->sem_value[i] + 0; |
| 5202 | s_s_par[i].params.fence.value = dst_f->sem_value[i] + 1; |
| 5203 | } |
| 5204 | |
| 5205 | err = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par, |
| 5206 | nb_images, cuda_dev->stream)); |
| 5207 | if (err < 0) |
| 5208 | goto fail; |
| 5209 | |
| 5210 | for (int i = 0; i < planes; i++) { |
| 5211 | CUDA_MEMCPY2D cpy = { |
| 5212 | .dstMemoryType = CU_MEMORYTYPE_DEVICE, |
| 5213 | .dstDevice = (CUdeviceptr)dst->data[i], |
| 5214 | .dstPitch = dst->linesize[i], |
| 5215 | .dstY = 0, |
| 5216 | |
| 5217 | .srcMemoryType = CU_MEMORYTYPE_ARRAY, |
| 5218 | .srcArray = dst_int->cu_array[i], |
| 5219 | }; |
| 5220 | |
| 5221 | int w, h; |
| 5222 | get_plane_wh(&w, &h, hwfc->sw_format, hwfc->width, hwfc->height, i); |
| 5223 | |
| 5224 | cpy.WidthInBytes = w * desc->comp[i].step; |
| 5225 | cpy.Height = h; |
| 5226 | |
| 5227 | err = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream)); |
| 5228 | if (err < 0) |
| 5229 | goto fail; |
| 5230 | } |
| 5231 | |
| 5232 | err = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par, |
| 5233 | nb_images, cuda_dev->stream)); |
| 5234 | if (err < 0) |
| 5235 | goto fail; |
| 5236 | |
| 5237 | for (int i = 0; i < nb_images; i++) |
| 5238 | dst_f->sem_value[i]++; |
| 5239 | |
| 5240 | CHECK_CU(cu->cuCtxPopCurrent(&dummy)); |
| 5241 | |
| 5242 | av_log(hwfc, AV_LOG_VERBOSE40, "Transferred Vulkan image to CUDA!\n"); |
| 5243 | |
| 5244 | return prepare_frame(hwfc, &fp->upload_exec, dst_f, PREP_MODE_EXTERNAL_IMPORT); |
| 5245 | |
| 5246 | fail: |
| 5247 | CHECK_CU(cu->cuCtxPopCurrent(&dummy)); |
| 5248 | vulkan_free_internal(p, dst_f); |
| 5249 | av_buffer_unref(&dst->buf[0]); |
| 5250 | return err; |
| 5251 | } |
| 5252 | #endif |
| 5253 | |
| 5254 | static int vulkan_transfer_data_from(AVHWFramesContext *hwfc, AVFrame *dst, |
| 5255 | const AVFrame *src) |
| 5256 | { |
| 5257 | av_unused__attribute__((unused)) VulkanDevicePriv *p = hwfc->device_ctx->hwctx; |
| 5258 | |
| 5259 | switch (dst->format) { |
| 5260 | #if CONFIG_CUDA0 |
| 5261 | case AV_PIX_FMT_CUDA: |
| 5262 | #ifdef _WIN32 |
| 5263 | if ((p->vkctx.extensions & FF_VK_EXT_EXTERNAL_WIN32_MEMORY(1ULL << 6)) && |
| 5264 | (p->vkctx.extensions & FF_VK_EXT_EXTERNAL_WIN32_SEM(1ULL << 7))) |
| 5265 | #else |
| 5266 | if ((p->vkctx.extensions & FF_VK_EXT_EXTERNAL_FD_MEMORY(1ULL << 2)) && |
| 5267 | (p->vkctx.extensions & FF_VK_EXT_EXTERNAL_FD_SEM(1ULL << 3))) |
| 5268 | #endif |
| 5269 | return vulkan_transfer_data_to_cuda(hwfc, dst, src); |
| 5270 | av_fallthrough__attribute__((fallthrough)); |
| 5271 | #endif |
| 5272 | default: |
| 5273 | if (dst->hw_frames_ctx) |
| 5274 | return AVERROR(ENOSYS)(-(38)); |
| 5275 | else |
| 5276 | return vulkan_transfer_frame(hwfc, dst, (AVFrame *)src, 0); |
| 5277 | } |
| 5278 | } |
| 5279 | |
| 5280 | static int vulkan_frames_derive_to(AVHWFramesContext *dst_fc, |
| 5281 | AVHWFramesContext *src_fc, int flags) |
| 5282 | { |
| 5283 | return vulkan_frames_init(dst_fc); |
| 5284 | } |
| 5285 | |
| 5286 | AVVkFrame *av_vk_frame_alloc(void) |
| 5287 | { |
| 5288 | int err; |
| 5289 | AVVkFrame *f = av_mallocz(sizeof(AVVkFrame)); |
| 5290 | if (!f) |
| 5291 | return NULL((void*)0); |
| 5292 | |
| 5293 | f->internal = av_mallocz(sizeof(*f->internal)); |
| 5294 | if (!f->internal) { |
| 5295 | av_free(f); |
| 5296 | return NULL((void*)0); |
| 5297 | } |
| 5298 | |
| 5299 | err = pthread_mutex_initstrict_pthread_mutex_init(&f->internal->update_mutex, NULL((void*)0)); |
| 5300 | if (err != 0) { |
| 5301 | av_free(f->internal); |
| 5302 | av_free(f); |
| 5303 | return NULL((void*)0); |
| 5304 | } |
| 5305 | |
| 5306 | return f; |
| 5307 | } |
| 5308 | |
| 5309 | const HWContextType ff_hwcontext_type_vulkan = { |
| 5310 | .type = AV_HWDEVICE_TYPE_VULKAN, |
| 5311 | .name = "Vulkan", |
| 5312 | |
| 5313 | .device_hwctx_size = sizeof(VulkanDevicePriv), |
| 5314 | .frames_hwctx_size = sizeof(VulkanFramesPriv), |
| 5315 | |
| 5316 | .device_init = &vulkan_device_init, |
| 5317 | .device_uninit = &vulkan_device_uninit, |
| 5318 | .device_create = &vulkan_device_create, |
| 5319 | .device_derive = &vulkan_device_derive, |
| 5320 | |
| 5321 | .frames_get_constraints = &vulkan_frames_get_constraints, |
| 5322 | .frames_init = vulkan_frames_init, |
| 5323 | .frames_get_buffer = vulkan_get_buffer, |
| 5324 | .frames_uninit = vulkan_frames_uninit, |
| 5325 | |
| 5326 | .transfer_get_formats = vulkan_transfer_get_formats, |
| 5327 | .transfer_data_to = vulkan_transfer_data_to, |
| 5328 | .transfer_data_from = vulkan_transfer_data_from, |
| 5329 | |
| 5330 | .map_to = vulkan_map_to, |
| 5331 | .map_from = vulkan_map_from, |
| 5332 | .frames_derive_to = &vulkan_frames_derive_to, |
| 5333 | |
| 5334 | .pix_fmts = (const enum AVPixelFormat []) { |
| 5335 | AV_PIX_FMT_VULKAN, |
| 5336 | AV_PIX_FMT_NONE |
| 5337 | }, |
| 5338 | }; |