]> git.sesse.net Git - ffmpeg/blobdiff - libavutil/hwcontext_vulkan.c
avutil: use the buffer_size_t typedef where required
[ffmpeg] / libavutil / hwcontext_vulkan.c
index a8e61a15f6bda98de6e66141ec764a61df081d0b..2e3faaaa1ae1cc2b946f966e4dc60ce5b47772f2 100644 (file)
@@ -184,7 +184,6 @@ static const struct {
     { AV_PIX_FMT_YUVA444P12, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
     { AV_PIX_FMT_YUVA444P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
 
-    { AV_PIX_FMT_BGR32,  { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
     { AV_PIX_FMT_BGRA,   { VK_FORMAT_B8G8R8A8_UNORM } },
     { AV_PIX_FMT_RGBA,   { VK_FORMAT_R8G8B8A8_UNORM } },
     { AV_PIX_FMT_RGB24,  { VK_FORMAT_R8G8B8_UNORM } },
@@ -195,9 +194,13 @@ static const struct {
     { AV_PIX_FMT_RGB565, { VK_FORMAT_R5G6B5_UNORM_PACK16 } },
     { AV_PIX_FMT_BGR565, { VK_FORMAT_B5G6R5_UNORM_PACK16 } },
     { AV_PIX_FMT_BGR0,   { VK_FORMAT_B8G8R8A8_UNORM } },
-    { AV_PIX_FMT_0BGR32, { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
     { AV_PIX_FMT_RGB0,   { VK_FORMAT_R8G8B8A8_UNORM } },
 
+    /* Lower priority as there's an endianess-dependent overlap between these
+     * and rgba/bgr0, and PACK32 formats are more limited */
+    { AV_PIX_FMT_BGR32,  { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
+    { AV_PIX_FMT_0BGR32, { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
+
     { AV_PIX_FMT_X2RGB10, { VK_FORMAT_A2R10G10B10_UNORM_PACK32 } },
 
     { AV_PIX_FMT_GBRAP, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
@@ -244,6 +247,8 @@ enum VulkanExtensions {
     EXT_EXTERNAL_FD_MEMORY     = 1ULL <<  2, /* VK_KHR_external_memory_fd */
     EXT_EXTERNAL_FD_SEM        = 1ULL <<  3, /* VK_KHR_external_semaphore_fd */
     EXT_EXTERNAL_HOST_MEMORY   = 1ULL <<  4, /* VK_EXT_external_memory_host */
+    EXT_PUSH_DESCRIPTORS       = 1ULL <<  5, /* VK_KHR_push_descriptor */
+    EXT_HOST_QUERY_RESET       = 1ULL <<  6, /* VK_EXT_host_query_reset */
 
     EXT_NO_FLAG                = 1ULL << 63,
 };
@@ -263,6 +268,8 @@ static const VulkanOptExtension optional_device_exts[] = {
     { VK_EXT_IMAGE_DRM_FORMAT_MODIFIER_EXTENSION_NAME,        EXT_DRM_MODIFIER_FLAGS,     },
     { VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME,            EXT_EXTERNAL_FD_SEM,        },
     { VK_EXT_EXTERNAL_MEMORY_HOST_EXTENSION_NAME,             EXT_EXTERNAL_HOST_MEMORY,   },
+    { VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME,                  EXT_PUSH_DESCRIPTORS,       },
+    { VK_EXT_HOST_QUERY_RESET_EXTENSION_NAME,                 EXT_HOST_QUERY_RESET,       },
 };
 
 /* Converts return values to strings */
@@ -1041,6 +1048,8 @@ static int vulkan_device_create_internal(AVHWDeviceContext *ctx,
     vkGetPhysicalDeviceFeatures(hwctx->phys_dev, &dev_features);
 #define COPY_FEATURE(DST, NAME) (DST).features.NAME = dev_features.NAME;
     COPY_FEATURE(hwctx->device_features, shaderImageGatherExtended)
+    COPY_FEATURE(hwctx->device_features, shaderStorageImageReadWithoutFormat)
+    COPY_FEATURE(hwctx->device_features, shaderStorageImageWriteWithoutFormat)
     COPY_FEATURE(hwctx->device_features, fragmentStoresAndAtomics)
     COPY_FEATURE(hwctx->device_features, vertexPipelineStoresAndAtomics)
     COPY_FEATURE(hwctx->device_features, shaderInt64)
@@ -1684,7 +1693,7 @@ static void try_export_flags(AVHWFramesContext *hwfc,
     }
 }
 
-static AVBufferRef *vulkan_pool_alloc(void *opaque, int size)
+static AVBufferRef *vulkan_pool_alloc(void *opaque, buffer_size_t size)
 {
     int err;
     AVVkFrame *f;
@@ -2000,7 +2009,7 @@ static inline VkFormat drm_to_vulkan_fmt(uint32_t drm_fourcc)
 }
 
 static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **frame,
-                                          AVDRMFrameDescriptor *desc)
+                                          const AVFrame *src)
 {
     int err = 0;
     VkResult ret;
@@ -2011,6 +2020,7 @@ static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **f
     VulkanDevicePriv *p = ctx->internal->priv;
     VulkanFramesPriv *fp = hwfc->internal->priv;
     AVVulkanFramesContext *frames_hwctx = hwfc->hwctx;
+    const AVDRMFrameDescriptor *desc = (AVDRMFrameDescriptor *)src->data[0];
     const int has_modifiers = !!(p->extensions & EXT_DRM_MODIFIER_FLAGS);
     VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS] = { 0 };
     VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { 0 };
@@ -2076,7 +2086,7 @@ static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **f
         };
 
         get_plane_wh(&create_info.extent.width, &create_info.extent.height,
-                     hwfc->sw_format, hwfc->width, hwfc->height, i);
+                     hwfc->sw_format, src->width, src->height, i);
 
         for (int j = 0; j < planes; j++) {
             plane_data[j].offset     = desc->layers[i].planes[j].offset;
@@ -2237,9 +2247,7 @@ static int vulkan_map_from_drm(AVHWFramesContext *hwfc, AVFrame *dst,
     AVVkFrame *f;
     VulkanMapping *map = NULL;
 
-    err = vulkan_map_from_drm_frame_desc(hwfc, &f,
-                                         (AVDRMFrameDescriptor *)src->data[0]);
-    if (err)
+    if ((err = vulkan_map_from_drm_frame_desc(hwfc, &f, src)))
         return err;
 
     /* The unmapping function will free this */
@@ -3244,6 +3252,8 @@ static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
     AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
     AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
     CudaFunctions *cu = cu_internal->cuda_dl;
+    CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS] = { 0 };
+    CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS] = { 0 };
 
     ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
     if (ret < 0)
@@ -3259,6 +3269,13 @@ static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
 
     dst_int = dst_f->internal;
 
+    ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par,
+                                                     planes, cuda_dev->stream));
+    if (ret < 0) {
+        err = AVERROR_EXTERNAL;
+        goto fail;
+    }
+
     for (int i = 0; i < planes; i++) {
         CUDA_MEMCPY2D cpy = {
             .dstMemoryType = CU_MEMORYTYPE_DEVICE,
@@ -3283,6 +3300,13 @@ static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
         }
     }
 
+    ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par,
+                                                       planes, cuda_dev->stream));
+    if (ret < 0) {
+        err = AVERROR_EXTERNAL;
+        goto fail;
+    }
+
     CHECK_CU(cu->cuCtxPopCurrent(&dummy));
 
     av_log(hwfc, AV_LOG_VERBOSE, "Transfered Vulkan image to CUDA!\n");