X86静态库编译 CPU加载正常 预测正常 GPU崩溃

[1733617157]

正常报错：Loading CPU model...
CPU model index: 0
CPU predict running...
CPU predict count: 3
CPU box[0] label=2 score=0.994 x0=88.1 y0=125.7 x1=113.7 y1=152.6
CPU box[1] label=1 score=0.562 x0=48.5 y0=64.1 x1=70.2 y1=96.7
CPU box[2] label=3 score=0.271 x0=206.8 y0=96.6 x1=237.3 y1=128.0
GPU model load failed, skip

下方详细错误：
D:\Desktop\v8_ncnn [0:0] $ & "D:\Desktop\v8_ncnn\build\Release\Release\test_xd_dll.exe"

Environment configured for Vulkan loader debug
Loading CPU model...
CPU model index: 0
INFO: Using Vulkan Loader C:\WINDOWS\SYSTEM32\vulkan-1.dll
INFO: Vulkan Loader Version 1.4.304
INFO: No valid vk_loader_settings.json file found, no loader settings will be active
INFO: windows_get_app_package_manifest_path: Failed to find mapping layers packages by family name
LAYER: Checking for Layer Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json" from PnP registry: E
INFO | LAYER: Located json file "D:\Steam\SteamOverlayVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO | LAYER: Located json file "D:\Steam\SteamFossilizeVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO: Found manifest file C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json (file version 1.0.1)
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamOverlayVulkanLayer.json
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamFossilizeVulkanLayer.json
INFO: windows_get_app_package_manifest_path: Failed to find mapping layers packages by family name
DRIVER: Checking for Driver Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\Drivers
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\iigd_dch.inf_amd64_749d16baad25e716\igvk32.json" from PnP registry: E
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json" from PnP registry: E
DRIVER: Found no registry files in HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\Drivers
DRIVER: Found ICD manifest file C:\WINDOWS\System32\DriverStore\FileRepository\iigd_dch.inf_amd64_749d16baad25e716\igvk32.json, version 1.0.0
DEBUG | DRIVER: Searching for ICD drivers named .\igvk32.dll
DRIVER: Found ICD manifest file C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json, version 1.0.1
DEBUG | DRIVER: Searching for ICD drivers named .\nvoglv32.dll
INFO: windows_get_app_package_manifest_path: Failed to find mapping layers packages by family name
DRIVER: Checking for Driver Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\Drivers
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\iigd_dch.inf_amd64_749d16baad25e716\igvk32.json" from PnP registry: E
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json" from PnP registry: E
DRIVER: Found no registry files in HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\Drivers
DRIVER: Found ICD manifest file C:\WINDOWS\System32\DriverStore\FileRepository\iigd_dch.inf_amd64_749d16baad25e716\igvk32.json, version 1.0.0
DEBUG | DRIVER: Searching for ICD drivers named .\igvk32.dll
DRIVER: Found ICD manifest file C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json, version 1.0.1
DEBUG | DRIVER: Searching for ICD drivers named .\nvoglv32.dll
INFO: windows_get_app_package_manifest_path: Failed to find mapping layers packages by family name
LAYER: Checking for Layer Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json" from PnP registry: E
INFO | LAYER: Located json file "D:\Steam\SteamOverlayVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO | LAYER: Located json file "D:\Steam\SteamFossilizeVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO: Found manifest file C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json (file version 1.0.1)
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamOverlayVulkanLayer.json
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamFossilizeVulkanLayer.json
INFO: No valid vk_loader_settings.json file found, no loader settings will be active
INFO: windows_get_app_package_manifest_path: Failed to find mapping layers packages by family name
LAYER: Checking for Layer Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json" from PnP registry: E
INFO | LAYER: Located json file "D:\Steam\SteamOverlayVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO | LAYER: Located json file "D:\Steam\SteamFossilizeVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO: Found manifest file C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json (file version 1.0.1)
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamOverlayVulkanLayer.json
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamFossilizeVulkanLayer.json
INFO: windows_get_app_package_manifest_path: Failed to find mapping layers packages by family name
DRIVER: Checking for Driver Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\Drivers
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\iigd_dch.inf_amd64_749d16baad25e716\igvk32.json" from PnP registry: E
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json" from PnP registry: E
DRIVER: Found no registry files in HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\Drivers
DRIVER: Found ICD manifest file C:\WINDOWS\System32\DriverStore\FileRepository\iigd_dch.inf_amd64_749d16baad25e716\igvk32.json, version 1.0.0
DEBUG | DRIVER: Searching for ICD drivers named .\igvk32.dll
DRIVER: Found ICD manifest file C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json, version 1.0.1
DEBUG | DRIVER: Searching for ICD drivers named .\nvoglv32.dll
INFO: windows_get_app_package_manifest_path: Failed to find mapping layers packages by family name
LAYER: Checking for Layer Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json" from PnP registry: E
INFO | LAYER: Located json file "D:\Steam\SteamOverlayVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO | LAYER: Located json file "D:\Steam\SteamFossilizeVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO: Found manifest file C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json (file version 1.0.1)
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamOverlayVulkanLayer.json
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamFossilizeVulkanLayer.json
INFO: No valid vk_loader_settings.json file found, no loader settings will be active
INFO: windows_get_app_package_manifest_path: Failed to find mapping layers packages by family name
LAYER: Checking for Layer Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json" from PnP registry: E
INFO | LAYER: Located json file "D:\Steam\SteamOverlayVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO | LAYER: Located json file "D:\Steam\SteamFossilizeVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO: Found manifest file C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json (file version 1.0.1)
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamOverlayVulkanLayer.json
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamFossilizeVulkanLayer.json
INFO: No valid vk_loader_settings.json file found, no loader settings will be active
INFO: Portability enumeration bit was set, enumerating portability drivers.
INFO: windows_get_app_package_manifest_path: Failed to find mapping layers packages by family name
LAYER: Checking for Layer Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json" from PnP registry: E
INFO | LAYER: Located json file "D:\Steam\SteamOverlayVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO | LAYER: Located json file "D:\Steam\SteamFossilizeVulkanLayer.json" from registry "HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ImplicitLayers"
INFO: Found manifest file C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json (file version 1.0.1)
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamOverlayVulkanLayer.json
ERROR: loader_get_json: Failed to open JSON file D:\Steam\SteamFossilizeVulkanLayer.json
LAYER: Checking for Layer Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\ExplicitLayers
INFO | LAYER: windows_get_device_registry_files: opening device ROOT\DISPLAY\0000
INFO | DRIVER: windows_get_device_registry_entry: Device ID(4) Does not contain a value for "VulkanExplicitLayersWow"
INFO | LAYER: windows_get_device_registry_files: unable to open child-device error:13
INFO | LAYER: windows_get_device_registry_files: opening device ROOT\DISPLAY\0001
INFO | DRIVER: windows_get_device_registry_entry: Device ID(2) Does not contain a value for "VulkanExplicitLayersWow"
INFO | LAYER: windows_get_device_registry_files: unable to open child-device error:13
INFO | LAYER: windows_get_device_registry_files: opening device PCI\VEN_10DE&DEV_28E0&SUBSYS_3E4517AA&REV_A1\4&1eca561b&0&0008
INFO | DRIVER: windows_get_device_registry_entry: Device ID(3) Does not contain a value for "VulkanExplicitLayersWow"
INFO | LAYER: windows_get_device_registry_files: unable to open child-device error:13
INFO | LAYER: windows_get_device_registry_files: opening device PCI\VEN_8086&DEV_A788&SUBSYS_3E4517AA&REV_04\3&11583659&1&10
INFO | DRIVER: windows_get_device_registry_entry: Device ID(1) Does not contain a value for "VulkanExplicitLayersWow"
INFO | LAYER: windows_get_device_registry_files: Opening child device 6 - DISPLAY\BOE0B38\4&375EB5CA&0&UID8388688
DEBUG | LAYER: windows_get_device_registry_files: GUID for 6 is not SoftwareComponent skipping
LAYER: windows_get_device_registry_files: found no registry files
LAYER: Found no registry files in HKEY_CURRENT_USER\SOFTWARE\Khronos\Vulkan\ExplicitLayers
WARNING | LAYER: windows_read_data_files_in_registry: Registry lookup failed to get layer manifest files.
INFO: windows_get_app_package_manifest_path: Failed to find mapping layers packages by family name
DRIVER: Checking for Driver Manifest files in Registry at HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\Drivers
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\iigd_dch.inf_amd64_749d16baad25e716\igvk32.json" from PnP registry: E
INFO: windows_add_json_entry: Located json file "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json" from PnP registry: E
DRIVER: Found no registry files in HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\Vulkan\Drivers
DRIVER: Found ICD manifest file C:\WINDOWS\System32\DriverStore\FileRepository\iigd_dch.inf_amd64_749d16baad25e716\igvk32.json, version 1.0.0
DEBUG | DRIVER: Searching for ICD drivers named .\igvk32.dll
DRIVER: Found ICD manifest file C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json, version 1.0.1
DEBUG | DRIVER: Searching for ICD drivers named .\nvoglv32.dll
WARNING | LAYER: env var 'VK_INSTANCE_LAYERS' defined and adding layers "VK_LAYER_KHRONOS_validation"
ERROR | LAYER: Layer "VK_LAYER_KHRONOS_validation" was not found but was requested by env var VK_INSTANCE_LAYERS!
WARNING | LAYER: env var 'VK_INSTANCE_LAYERS' defined and adding layers "VK_LAYER_KHRONOS_validation"
ERROR | LAYER: Layer "VK_LAYER_KHRONOS_validation" was not found but was requested by env var VK_INSTANCE_LAYERS!
WARNING | LAYER: Layer VK_LAYER_NV_optimus uses API version 1.3 which is older than the application specified API version of 1.4. May cause issues.
DEBUG | LAYER: Loading layer library C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94.\nvoglv32.dll
INFO | LAYER: Insert instance layer "VK_LAYER_NV_optimus" (C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94.\nvoglv32.dll)
LAYER: vkCreateInstance layer callstack setup to:
LAYER:
LAYER: ||
LAYER:
LAYER: ||
LAYER: VK_LAYER_NV_optimus
LAYER: Type: Implicit
LAYER: Enabled By: Implicit Layer
LAYER: Disable Env Var: DISABLE_LAYER_NV_OPTIMUS_1
LAYER: Manifest: C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94\nv-vk32.json
LAYER: Library: C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94.\nvoglv32.dll
LAYER: ||
LAYER:
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] queueC=0[1] queueT=0[1]
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] subgroup=32(832) ops=1/1/1/1/1/1/1/1/1/1
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-cm=0 int8-cm=0 bf16-cm=0 fp8-cm=0
[1 NVIDIA GeForce RTX 4060 Laptop GPU] queueC=2[8] queueT=1[2]
[1 NVIDIA GeForce RTX 4060 Laptop GPU] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[1 NVIDIA GeForce RTX 4060 Laptop GPU] subgroup=32(3232) ops=1/1/1/1/1/1/1/1/1/1
[1 NVIDIA GeForce RTX 4060 Laptop GPU] fp16-cm=16x16x16/16x8x16/16x8x8 int8-cm=16x16x32/16x8x32 bf16-cm=0 fp8-cm=0
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] queueC=0[1] queueT=0[1]
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] subgroup=32(832) ops=1/1/1/1/1/1/1/1/1/1
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-cm=0 int8-cm=0 bf16-cm=0 fp8-cm=0
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] queueC=0[1] queueT=0[1]
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] subgroup=32(832) ops=1/1/1/1/1/1/1/1/1/1
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-cm=0 int8-cm=0 bf16-cm=0 fp8-cm=0
INFO | LAYER: Inserted device layer "VK_LAYER_NV_optimus" (C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94.\nvoglv32.dll)
DRIVER | LAYER: vkCreateDevice layer callstack setup to:
DRIVER | LAYER:
DRIVER | LAYER: ||
DRIVER | LAYER:
DRIVER | LAYER: ||
LAYER: VK_LAYER_NV_optimus
LAYER: Type: Implicit
LAYER: Enabled By: Implicit Layer
LAYER: Disable Env Var: DISABLE_LAYER_NV_OPTIMUS_1
LAYER: Manifest: C:\WIN|
DRIVER | LAYER:
DRIVER | LAYER: Using "NVIDIA GeForce RTX 4060 Laptop GPU" with driver: "C:\WINDOWS\System32\DriverStore\FileRepository\nvlt.inf_amd64_9312b88fbd16aa94.\nvoglv32.dll"

[dosubot]

Hi @1733617157! I'm Dosu and I’m helping the ncnn team.

你的CPU推理正常，但GPU模型加载失败，主要原因通常是Vulkan驱动或环境问题。根据ncnn官方FAQ和Vulkan加载机制，建议按以下步骤排查和修复：

1. 升级NVIDIA显卡驱动
   很多Vulkan相关的加载失败、崩溃、vkCreateInstance/vkEnumeratePhysicalDevices等错误，官方首推升级GPU驱动到最新版，尤其是笔记本独显经常因为驱动不全导致Vulkan不可用。驱动下载：NVIDIA官网
   参考：ncnn FAQ、Vulkan FAQ

2. 确认Vulkan环境和驱动状态
   用nvidia-smi确认GPU和驱动是否正常识别。如果nvidia-smi报错或找不到GPU，说明驱动有问题，需要重装驱动。
   参考：Vulkan FAQ

3. 确认ncnn代码设置
   必须在加载模型前设置net.opt.use_vulkan_compute = 1，否则不会启用GPU推理。
   参考：Vulkan driver loader

4. Vulkan层和ICD警告可忽略
   VK_LAYER_KHRONOS_validation缺失、SteamOverlay/Fossilize等json加载失败一般不会影响推理，只是警告。
   但如果VK_LAYER_NV_optimus版本过低，可能导致兼容性问题，升级驱动通常能解决。

5. 32位程序注意
   你是x86静态库，ncnn会自动在DriverStore下找nvoglv32.dll等32位驱动文件，驱动不全或版本过低也会导致加载失败。
   参考：Vulkan driver loader

6. 高级调试
   可以通过设置环境变量VK_ICD_FILENAMES或NCNN_VULKAN_DRIVER指定Vulkan驱动路径，排查驱动加载问题。
   参考：Vulkan driver loader

如果升级驱动后仍然GPU加载失败，建议用vulkaninfo工具检查Vulkan环境，或尝试在其他Vulkan程序（如Vulkan SDK自带demo）下验证Vulkan是否可用。

相关官方文档和FAQ：

• ncnn Vulkan FAQ
• ncnn FAQ
• Vulkan driver loader说明

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[1733617157]

#if NCNN_VULKAN
int gpu_count = ncnn::get_gpu_count();
if (gpu_index >= 0 && gpu_index < gpu_count)
{
const char* drv = getenv("NCNN_VULKAN_DRIVER");
int ret = 0;
if (drv && drv[0]) ret = ncnn::create_gpu_instance(drv);
else ret = ncnn::create_gpu_instance();
if (ret == 0)
{
m->net.opt.use_vulkan_compute = 1;
m->net.set_vulkan_device(gpu_index);
m->use_gpu = true;
g_gpu_ref.fetch_add(1);
}
} 我更新了最新的显卡驱动 按照要求 对代码进行更改 目前仍然是无法使用GPU的状态 Loading CPU model...
CPU model index: 0
CPU predict running...
CPU predict count: 3
CPU box[0] label=2 score=0.994 x0=88.1 y0=125.7 x1=113.7 y1=152.6
CPU box[1] label=1 score=0.562 x0=48.5 y0=64.1 x1=70.2 y1=96.7
CPU box[2] label=3 score=0.271 x0=206.8 y0=96.6 x1=237.3 y1=128.0
GPU model load failed, skip
(TraeAI-6) D:\Desktop\v8_ncnn [1:-1073741819] $
(TraeAI-6) D:\Desktop\v8_ncnn [1:-1073741819] $

[dosubot]

你的代码和驱动设置基本没问题，但“GPU model load failed, skip”且进程退出码 -1073741819（Windows 下的 0xC0000005）说明有内存访问冲突，通常是 GPU 初始化或资源清理阶段崩溃。

结合 ncnn 官方文档和源码，建议重点排查：

1. 确保 ncnn 是用 -DNCNN_VULKAN=ON 编译的
   如果不是，GPU 相关 API 可能会异常或崩溃（比如 create_gpu_instance/get_gpu_count 这些函数）参考。

2. Vulkan 驱动和环境
   你的日志显示 NVIDIA 驱动已被识别，Vulkan 层缺失（如 VK_LAYER_KHRONOS_validation）只是警告，不影响推理。
   建议用 vulkaninfo 工具确认 Vulkan 环境完全可用，能否正常列出 RTX 4060 相关信息。官方FAQ

3. 模型兼容性
   某些模型层如果不支持 Vulkan，会自动回退到 CPU，不会导致加载失败。但如果模型文件损坏或格式不对，ncnn 会直接 fail-fast 并清理资源，避免崩溃。PR #6383

4. 资源清理与生命周期
   ncnn 近期修复了 destroy_gpu_instance 崩溃问题（PR Fix destroy_gpu_instance crash #5353），但如果 create_gpu_instance 失败后没有正确清理，依然可能导致进程崩溃。建议每次用完 GPU 后显式调用 ncnn::destroy_gpu_instance()。PR #5353

5. 分配器设置
   如果你自定义了分配器（blob_vkallocator、workspace_vkallocator、staging_vkallocator），确保它们都来自同一个 VulkanDevice，否则会有潜在崩溃风险。Vulkan Notes

6. 调试建议

   • 用最新版 ncnn，确保包含 destroy_gpu_instance 崩溃修复。
   • 打开 ENABLE_VALIDATION_LAYER（src/gpu.cpp 里 #define ENABLE_VALIDATION_LAYER 1），可获得更详细的 Vulkan 报错。
   • 用 Debug/Release 两种编译模式都试试，部分 release 优化可能掩盖 bug。
   • 检查模型文件是否损坏或格式不对。

7. 最小复现代码建议
   参考官方 FAQ 推荐的最小 GPU 推理流程，逐步排查是哪一步出错：示例

如还有问题，建议贴出 ncnn 的详细日志（可用 NCNN_LOGE 打印），或用调试器定位崩溃堆栈。这样能更快定位是 Vulkan 初始化、模型加载还是资源清理阶段出错。

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[1733617157]

(base) PS D:\Desktop\v8_ncnn> ./build/Release/Release/test_xd_dll.exe
Loading CPU model...
CPU model index: 0
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] queueC=0[1] queueG=0[1] queueT=0[1]
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] subgroup=32(832) ops=1/1/1/1/1/1/1/1/1/1
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0
[1 NVIDIA GeForce RTX 4060 Laptop GPU] queueC=2[8] queueG=0[16] queueT=1[2]
[1 NVIDIA GeForce RTX 4060 Laptop GPU] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[1 NVIDIA GeForce RTX 4060 Laptop GPU] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[1 NVIDIA GeForce RTX 4060 Laptop GPU] subgroup=32(3232) ops=1/1/1/1/1/1/1/1/1/1
[1 NVIDIA GeForce RTX 4060 Laptop GPU] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/1/1/1
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] queueC=0[1] queueG=0[1] queueT=0[1]
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] subgroup=32(832) ops=1/1/1/1/1/1/1/1/1/1
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] queueC=0[1] queueG=0[1] queueT=0[1]
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] subgroup=32(832) ops=1/1/1/1/1/1/1/1/1/1
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0 依然会崩溃 cpu加载正常 GPU加载依然崩溃

#pragma once
#include <stdint.h>
#include <windows.h>

typedef struct XD_Box {
float x0;
float y0;
float x1;
float y1;
float score;
int32_t label;
} XD_Box;

#define XD_MAX_BOXES 1024

extern "C" __declspec(dllexport) int WINAPI XD_LoadModel(const unsigned char* param_data,
int param_size,
const unsigned char* bin_data,
int bin_size,
int shape,
int gpu_index);

extern "C" __declspec(dllexport) int WINAPI XD_Predict(int model_index,
const unsigned char* image_data,
int image_size,
XD_Box* boxes,
float conf_threshold,
float iou_threshold,
BOOL show);

extern "C" __declspec(dllexport) void WINAPI XD_FreeModel(int model_index);

#include <windows.h>
#include
#include
#include
#include
#include
#include "xd_yolov8_ncnn.h"
#include <opencv2/core.hpp>
#include <opencv2/imgcodecs.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
#include <ncnn/net.h>
#include <ncnn/mat.h>
#include <ncnn/datareader.h>

struct XD_Model {
ncnn::Net net;
int input_size;
int max_stride;
bool use_gpu;
};

static std::mutex g_mutex;
static std::vector<std::unique_ptr<XD_Model>> g_models;
static std::atomic g_gpu_ref{0};

static inline void softmax(const float* x, int n, std::vector& out)
{
out.resize(n);
float m = x[0];
for (int i = 1; i < n; ++i) if (x[i] > m) m = x[i];
float s = 0.f;
for (int i = 0; i < n; ++i) { out[i] = std::exp(x[i] - m); s += out[i]; }
float invs = s > 0 ? 1.f / s : 0.f;
for (int i = 0; i < n; ++i) out[i] *= invs;
}

static inline void letterbox(const cv::Mat& bgr, int target, int max_stride, cv::Mat& padded, float& scale)
{
int ih = bgr.rows, iw = bgr.cols;
if (iw >= ih) { scale = (float)target / (float)iw; } else { scale = (float)target / (float)ih; }
int nw = (int)std::round(iw * scale);
int nh = (int)std::round(ih * scale);
cv::Mat resized; cv::resize(bgr, resized, cv::Size(nw, nh), 0, 0, cv::INTER_LINEAR);
int wpad = ((nw + max_stride - 1) / max_stride) * max_stride - nw;
int hpad = ((nh + max_stride - 1) / max_stride) * max_stride - nh;
cv::copyMakeBorder(resized, padded, 0, hpad, 0, wpad, cv::BORDER_CONSTANT, cv::Scalar(114,114,114));
}

static inline float iou_xyxy(const XD_Box& a, const XD_Box& b)
{
float x0 = std::max(a.x0, b.x0);
float y0 = std::max(a.y0, b.y0);
float x1 = std::min(a.x1, b.x1);
float y1 = std::min(a.y1, b.y1);
float iw = std::max(0.f, x1 - x0);
float ih = std::max(0.f, y1 - y0);
float inter = iw * ih;
float area_a = (a.x1 - a.x0) * (a.y1 - a.y0);
float area_b = (b.x1 - b.x0) * (b.y1 - b.y0);
return inter / (area_a + area_b - inter + 1e-6f);
}

static int nms(std::vector<XD_Box>& boxes, float iou_thr)
{
std::sort(boxes.begin(), boxes.end(), [](const XD_Box& a, const XD_Box& b){ return a.score > b.score; });
std::vector<XD_Box> picked;
for (size_t i = 0; i < boxes.size(); ++i)
{
bool keep = true;
for (size_t j = 0; j < picked.size(); ++j)
{
if (iou_xyxy(boxes[i], picked[j]) > iou_thr) { keep = false; break; }
}
if (keep) picked.push_back(boxes[i]);
}
boxes.swap(picked);
return (int)boxes.size();
}

extern "C" __declspec(dllexport) int WINAPI XD_LoadModel(const unsigned char* param_data,
int param_size,
const unsigned char* bin_data,
int bin_size,
int shape,
int gpu_index)
{
try {
std::unique_ptr<XD_Model> m(new XD_Model());
m->input_size = shape > 0 ? shape : 640;
m->max_stride = 32;
m->use_gpu = false;
if (gpu_index >= 0)
{
#if NCNN_VULKAN
int gpu_count = ncnn::get_gpu_count();
if (gpu_index >= 0 && gpu_index < gpu_count)
{
const char* drv = getenv("NCNN_VULKAN_DRIVER");
int ret = 0;
if (drv && drv[0]) ret = ncnn::create_gpu_instance(drv);
else ret = ncnn::create_gpu_instance();
if (ret == 0)
{
m->net.opt.use_vulkan_compute = 1;
m->net.set_vulkan_device(gpu_index);
m->use_gpu = true;
g_gpu_ref.fetch_add(1);
}
}
#endif
}
m->net.opt.num_threads = (int)std::max(1u, std::thread::hardware_concurrency());
m->net.opt.use_packing_layout = true;
if (!param_data || param_size <= 0 || !bin_data || bin_size <= 0) return -1;
std::vector param_copy;
param_copy.reserve(param_size + 1);
param_copy.insert(param_copy.end(), param_data, param_data + param_size);
param_copy.push_back('\0');
const unsigned char* param_ptr = param_copy.data();
ncnn::DataReaderFromMemory dr_param(param_ptr);
int rp = m->net.load_param(dr_param);
if (rp != 0) return -1;
const unsigned char* bin_ptr = bin_data;
ncnn::DataReaderFromMemory dr_model(bin_ptr);
int rm = m->net.load_model(dr_model);
if (rm != 0) return -1;
std::lock_guardstd::mutex lk(g_mutex);
g_models.push_back(std::move(m));
return (int)g_models.size() - 1;
} catch (...) { return -1; }
}

extern "C" __declspec(dllexport) int WINAPI XD_Predict(int model_index,
const unsigned char* image_data,
int image_size,
XD_Box* boxes,
float conf_threshold,
float iou_threshold,
BOOL show)
{
std::unique_ptr<XD_Model>* pm = nullptr;
{
std::lock_guardstd::mutex lk(g_mutex);
if (model_index < 0 || model_index >= (int)g_models.size() || !g_models[model_index]) return -1;
pm = &g_models[model_index];
}
XD_Model* m = pm->get();
std::vector buf(image_data, image_data + image_size);
cv::Mat bgr = cv::imdecode(buf, cv::IMREAD_COLOR);
if (bgr.empty()) return -1;
cv::Mat padded; float scale = 1.f; letterbox(bgr, m->input_size, m->max_stride, padded, scale);
int pw = padded.cols, ph = padded.rows;

ncnn::Mat in = ncnn::Mat::from_pixels(padded.data, ncnn::Mat::PIXEL_BGR, pw, ph);
const float mean_vals[3] = {0.f,0.f,0.f};
const float norm_vals[3] = {1.f/255.f,1.f/255.f,1.f/255.f};
in.substract_mean_normalize(mean_vals, norm_vals);

ncnn::Extractor ex = m->net.create_extractor();
ex.input("in0", in);
ncnn::Mat out0, out1, out2;
if (ex.extract("out0", out0) != 0) return -1;
if (ex.extract("out1", out1) != 0) return -1;
if (ex.extract("out2", out2) != 0) return -1;


int reg_max = 16;
std::vector<XD_Box> all;
auto decode = [&](const ncnn::Mat& pred, int stride){
    int num_points = (pw / stride) * (ph / stride);
    int total = pred.w * pred.h * pred.c;
    if (num_points <= 0 || total <= 0) return;
    if (total % num_points != 0) return;
    int ch = total / num_points;
    int cls_channels = ch - reg_max * 4;
    if (cls_channels <= 0) return;
    std::vector<float> arr(total);
    int idx = 0;
    for (int c = 0; c < pred.c; ++c)
    {
        const ncnn::Mat plane = pred.channel(c);
        for (int y = 0; y < pred.h; ++y)
        {
            const float* row = plane.row(y);
            for (int x = 0; x < pred.w; ++x) arr[idx++] = row[x];
        }
    }
    for (int i = 0; i < num_points; ++i)
    {
        float dist[4];
        for (int k = 0; k < 4; ++k)
        {
            const float* p = &arr[i*ch + k*reg_max];
            float m = p[0]; for (int b = 1; b < reg_max; ++b) if (p[b] > m) m = p[b];
            float s = 0.f, e = 0.f; for (int b = 0; b < reg_max; ++b) { float v = std::exp(p[b]-m); s += v; e += v * (float)b; }
            dist[k] = (s > 0.f ? (e/s) : 0.f) * stride;
        }
        int label = 0; float score = arr[i*ch + reg_max*4 + 0];
        for (int c = 1; c < cls_channels; ++c)
        {
            float v = arr[i*ch + reg_max*4 + c];
            if (v > score) { score = v; label = c; }
        }
        if (score < conf_threshold) continue;
        int gw = pw / stride; int gx = i % gw; int gy = i / gw;
        float cx = (gx + 0.5f) * stride;
        float cy = (gy + 0.5f) * stride;
        XD_Box b;
        b.x0 = cx - dist[0]; b.y0 = cy - dist[1]; b.x1 = cx + dist[2]; b.y1 = cy + dist[3];
        b.score = score; b.label = label;
        all.push_back(b);
    }
};
decode(out0, 8);
decode(out1, 16);
decode(out2, 32);
if (all.empty()) return 0;
nms(all, iou_threshold);

int iw = bgr.cols, ih = bgr.rows;
int count = std::min((int)all.size(), XD_MAX_BOXES);
for (int i = 0; i < count; ++i)
{
    float x0 = all[i].x0 / scale;
    float y0 = all[i].y0 / scale;
    float x1 = all[i].x1 / scale;
    float y1 = all[i].y1 / scale;
    x0 = std::max(0.f, std::min((float)(iw-1), x0));
    y0 = std::max(0.f, std::min((float)(ih-1), y0));
    x1 = std::max(0.f, std::min((float)(iw-1), x1));
    y1 = std::max(0.f, std::min((float)(ih-1), y1));
    boxes[i].x0 = x0; boxes[i].y0 = y0; boxes[i].x1 = x1; boxes[i].y1 = y1;
    boxes[i].score = all[i].score; boxes[i].label = all[i].label;
    if (show)
    {
        cv::rectangle(bgr, cv::Rect(cv::Point((int)x0,(int)y0), cv::Point((int)x1,(int)y1)), cv::Scalar(0,255,0), 2);
    }
}
if (show)
{
    cv::imshow("xd_yolov8_ncnn", bgr);
    cv::waitKey(0);
    cv::destroyAllWindows();
}
return count;

}

extern "C" __declspec(dllexport) void WINAPI XD_FreeModel(int model_index)
{
std::unique_ptr<XD_Model> old;
{
std::lock_guardstd::mutex lk(g_mutex);
if (model_index < 0 || model_index >= (int)g_models.size()) return;
old = std::move(g_models[model_index]);
g_models[model_index].reset();
}
#if NCNN_VULKAN
if (old && old->use_gpu)
{
if (g_gpu_ref.fetch_sub(1) == 1) ncnn::destroy_gpu_instance();
}
#endif
}

[dosubot]

你的DLL接口和Vulkan初始化流程基本没问题，但ncnn在Windows下GPU加载崩溃（-1073741819）通常和Vulkan资源管理顺序、分配器、或模型兼容性有关。根据ncnn源码和类似案例，建议重点排查：

1. Vulkan资源管理顺序
   GPU相关的初始化和销毁（create_gpu_instance/destroy_gpu_instance）必须严格配对，且全局只初始化一次，销毁时机要早于DLL卸载（NVIDIA驱动尤其敏感）。建议在DLL加载时初始化，卸载前销毁，避免多次调用或顺序错乱。参考

2. 分配器必须匹配VulkanDevice
   如果你自定义了blob_vkallocator、workspace_vkallocator、staging_vkallocator，必须全部来自同一个VulkanDevice，不能跨device或在device销毁后继续用，否则极易崩溃。参考

3. 模型兼容性
   某些模型层（如PriorBox）在Vulkan下可能不支持，导致load_model失败。可以尝试只用CPU跑有问题的层，其余用GPU。案例

4. 调试建议

   • 用最新版ncnn，确保destroy_gpu_instance相关修复已合入。
   • 打开NCNN_LOGE和ENABLE_VALIDATION_LAYER，获取详细Vulkan报错。
   • 检查模型文件是否损坏或格式不对。
   • 用vulkaninfo确认Vulkan环境完全可用。

5. 最小复现流程
   推荐流程：create_gpu_instance → net.opt.use_vulkan_compute=1 → set_vulkan_device → load_param/load_model → 推理 → destroy_gpu_instance。所有Vulkan相关资源必须在destroy_gpu_instance前释放。官方FAQ

如果还有崩溃，建议用调试器定位具体崩溃堆栈，或贴出NCNN_LOGE详细日志，进一步定位是模型加载、推理还是资源清理阶段出错。

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[1733617157]

#include <windows.h>
#include <ncnn/gpu.h>
#include
#include
#include
#include
#include
#include "xd_yolov8_ncnn.h"
#include <opencv2/core.hpp>
#include <opencv2/imgcodecs.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
#include <ncnn/net.h>
#include <ncnn/mat.h>
#include <ncnn/datareader.h>

struct XD_Model {
ncnn::Net net;
int input_size;
int max_stride;
bool use_gpu;
ncnn::VkAllocator* blob_vkallocator;
ncnn::VkAllocator* workspace_vkallocator;
ncnn::VkAllocator* staging_vkallocator;
};

static std::mutex g_mutex;
static std::vector<std::unique_ptr<XD_Model>> g_models;
static std::atomic g_gpu_ref{0};

static inline void softmax(const float* x, int n, std::vector& out)
{
out.resize(n);
float m = x[0];
for (int i = 1; i < n; ++i) if (x[i] > m) m = x[i];
float s = 0.f;
for (int i = 0; i < n; ++i) { out[i] = std::exp(x[i] - m); s += out[i]; }
float invs = s > 0 ? 1.f / s : 0.f;
for (int i = 0; i < n; ++i) out[i] *= invs;
}

static inline void letterbox(const cv::Mat& bgr, int target, int max_stride, cv::Mat& padded, float& scale)
{
int ih = bgr.rows, iw = bgr.cols;
if (iw >= ih) { scale = (float)target / (float)iw; } else { scale = (float)target / (float)ih; }
int nw = (int)std::round(iw * scale);
int nh = (int)std::round(ih * scale);
cv::Mat resized; cv::resize(bgr, resized, cv::Size(nw, nh), 0, 0, cv::INTER_LINEAR);
int wpad = ((nw + max_stride - 1) / max_stride) * max_stride - nw;
int hpad = ((nh + max_stride - 1) / max_stride) * max_stride - nh;
cv::copyMakeBorder(resized, padded, 0, hpad, 0, wpad, cv::BORDER_CONSTANT, cv::Scalar(114,114,114));
}

static inline float iou_xyxy(const XD_Box& a, const XD_Box& b)
{
float x0 = std::max(a.x0, b.x0);
float y0 = std::max(a.y0, b.y0);
float x1 = std::min(a.x1, b.x1);
float y1 = std::min(a.y1, b.y1);
float iw = std::max(0.f, x1 - x0);
float ih = std::max(0.f, y1 - y0);
float inter = iw * ih;
float area_a = (a.x1 - a.x0) * (a.y1 - a.y0);
float area_b = (b.x1 - b.x0) * (b.y1 - b.y0);
return inter / (area_a + area_b - inter + 1e-6f);
}

static int nms(std::vector<XD_Box>& boxes, float iou_thr)
{
std::sort(boxes.begin(), boxes.end(), [](const XD_Box& a, const XD_Box& b){ return a.score > b.score; });
std::vector<XD_Box> picked;
for (size_t i = 0; i < boxes.size(); ++i)
{
bool keep = true;
for (size_t j = 0; j < picked.size(); ++j)
{
if (iou_xyxy(boxes[i], picked[j]) > iou_thr) { keep = false; break; }
}
if (keep) picked.push_back(boxes[i]);
}
boxes.swap(picked);
return (int)boxes.size();
}

extern "C" __declspec(dllexport) int WINAPI XD_LoadModel(const unsigned char* param_data,
int param_size,
const unsigned char* bin_data,
int bin_size,
int shape,
int gpu_index)
{
try {
std::unique_ptr<XD_Model> m(new XD_Model());
m->input_size = shape > 0 ? shape : 640;
m->max_stride = 32;
m->use_gpu = false;
m->blob_vkallocator = nullptr;
m->workspace_vkallocator = nullptr;
m->staging_vkallocator = nullptr;
if (gpu_index >= 0)
{
#if NCNN_VULKAN
int gpu_count = ncnn::get_gpu_count();
if (gpu_index >= 0 && gpu_index < gpu_count)
{
int ret = 0;
if (g_gpu_ref.load() == 0)
{
const char* drv = getenv("NCNN_VULKAN_DRIVER");
if (drv && drv[0]) ret = ncnn::create_gpu_instance(drv);
else ret = ncnn::create_gpu_instance();
if (ret != 0) ret = -1;
}
if (ret == 0 || g_gpu_ref.load() > 0)
{
ncnn::VulkanDevice* vkdev = ncnn::get_gpu_device(gpu_index);
m->blob_vkallocator = new ncnn::VkBlobAllocator(vkdev);
m->workspace_vkallocator = m->blob_vkallocator;
m->staging_vkallocator = new ncnn::VkStagingAllocator(vkdev);
m->net.opt.blob_vkallocator = m->blob_vkallocator;
m->net.opt.workspace_vkallocator = m->workspace_vkallocator;
m->net.opt.staging_vkallocator = m->staging_vkallocator;
m->net.opt.use_vulkan_compute = 1;
m->net.set_vulkan_device(gpu_index);
m->use_gpu = true;
g_gpu_ref.fetch_add(1);
}
}
#endif
}
m->net.opt.num_threads = (int)std::max(1u, std::thread::hardware_concurrency());
m->net.opt.use_packing_layout = true;
if (!param_data || param_size <= 0 || !bin_data || bin_size <= 0) return -1;
std::vector param_copy;
param_copy.reserve(param_size + 1);
param_copy.insert(param_copy.end(), param_data, param_data + param_size);
param_copy.push_back('\0');
const unsigned char* param_ptr = param_copy.data();
ncnn::DataReaderFromMemory dr_param(param_ptr);
int rp = m->net.load_param(dr_param);
if (rp != 0 && m->use_gpu)
{
m->net.clear();
m->net.opt.use_vulkan_compute = 0;
m->net.opt.blob_vkallocator = nullptr;
m->net.opt.workspace_vkallocator = nullptr;
m->net.opt.staging_vkallocator = nullptr;
m->use_gpu = false;
rp = m->net.load_param(dr_param);
}
if (rp != 0) return -1;
const unsigned char* bin_ptr = bin_data;
ncnn::DataReaderFromMemory dr_model(bin_ptr);
int rm = m->net.load_model(dr_model);
if (rm != 0 && m->use_gpu)
{
m->net.clear();
m->net.opt.use_vulkan_compute = 0;
m->net.opt.blob_vkallocator = nullptr;
m->net.opt.workspace_vkallocator = nullptr;
m->net.opt.staging_vkallocator = nullptr;
m->use_gpu = false;
rm = m->net.load_model(dr_model);
}
if (rm != 0) return -1;
std::lock_guardstd::mutex lk(g_mutex);
g_models.push_back(std::move(m));
return (int)g_models.size() - 1;
} catch (...) { return -1; }
}

extern "C" __declspec(dllexport) int WINAPI XD_Predict(int model_index,
const unsigned char* image_data,
int image_size,
XD_Box* boxes,
float conf_threshold,
float iou_threshold,
BOOL show)
{
std::unique_ptr<XD_Model>* pm = nullptr;
{
std::lock_guardstd::mutex lk(g_mutex);
if (model_index < 0 || model_index >= (int)g_models.size() || !g_models[model_index]) return -1;
pm = &g_models[model_index];
}
XD_Model* m = pm->get();
std::vector buf(image_data, image_data + image_size);
cv::Mat bgr = cv::imdecode(buf, cv::IMREAD_COLOR);
if (bgr.empty()) return -1;
cv::Mat padded; float scale = 1.f; letterbox(bgr, m->input_size, m->max_stride, padded, scale);
int pw = padded.cols, ph = padded.rows;

ncnn::Mat in = ncnn::Mat::from_pixels(padded.data, ncnn::Mat::PIXEL_BGR, pw, ph);
const float mean_vals[3] = {0.f,0.f,0.f};
const float norm_vals[3] = {1.f/255.f,1.f/255.f,1.f/255.f};
in.substract_mean_normalize(mean_vals, norm_vals);

ncnn::Extractor ex = m->net.create_extractor();
ex.input("in0", in);
ncnn::Mat out0, out1, out2;
if (ex.extract("out0", out0) != 0) return -1;
if (ex.extract("out1", out1) != 0) return -1;
if (ex.extract("out2", out2) != 0) return -1;


int reg_max = 16;
std::vector<XD_Box> all;
auto decode = [&](const ncnn::Mat& pred, int stride){
    int num_points = (pw / stride) * (ph / stride);
    int total = pred.w * pred.h * pred.c;
    if (num_points <= 0 || total <= 0) return;
    if (total % num_points != 0) return;
    int ch = total / num_points;
    int cls_channels = ch - reg_max * 4;
    if (cls_channels <= 0) return;
    std::vector<float> arr(total);
    int idx = 0;
    for (int c = 0; c < pred.c; ++c)
    {
        const ncnn::Mat plane = pred.channel(c);
        for (int y = 0; y < pred.h; ++y)
        {
            const float* row = plane.row(y);
            for (int x = 0; x < pred.w; ++x) arr[idx++] = row[x];
        }
    }
    for (int i = 0; i < num_points; ++i)
    {
        float dist[4];
        for (int k = 0; k < 4; ++k)
        {
            const float* p = &arr[i*ch + k*reg_max];
            float m = p[0]; for (int b = 1; b < reg_max; ++b) if (p[b] > m) m = p[b];
            float s = 0.f, e = 0.f; for (int b = 0; b < reg_max; ++b) { float v = std::exp(p[b]-m); s += v; e += v * (float)b; }
            dist[k] = (s > 0.f ? (e/s) : 0.f) * stride;
        }
        int label = 0; float score = arr[i*ch + reg_max*4 + 0];
        for (int c = 1; c < cls_channels; ++c)
        {
            float v = arr[i*ch + reg_max*4 + c];
            if (v > score) { score = v; label = c; }
        }
        if (score < conf_threshold) continue;
        int gw = pw / stride; int gx = i % gw; int gy = i / gw;
        float cx = (gx + 0.5f) * stride;
        float cy = (gy + 0.5f) * stride;
        XD_Box b;
        b.x0 = cx - dist[0]; b.y0 = cy - dist[1]; b.x1 = cx + dist[2]; b.y1 = cy + dist[3];
        b.score = score; b.label = label;
        all.push_back(b);
    }
};
decode(out0, 8);
decode(out1, 16);
decode(out2, 32);
if (all.empty()) return 0;
nms(all, iou_threshold);

int iw = bgr.cols, ih = bgr.rows;
int count = std::min((int)all.size(), XD_MAX_BOXES);
for (int i = 0; i < count; ++i)
{
    float x0 = all[i].x0 / scale;
    float y0 = all[i].y0 / scale;
    float x1 = all[i].x1 / scale;
    float y1 = all[i].y1 / scale;
    x0 = std::max(0.f, std::min((float)(iw-1), x0));
    y0 = std::max(0.f, std::min((float)(ih-1), y0));
    x1 = std::max(0.f, std::min((float)(iw-1), x1));
    y1 = std::max(0.f, std::min((float)(ih-1), y1));
    boxes[i].x0 = x0; boxes[i].y0 = y0; boxes[i].x1 = x1; boxes[i].y1 = y1;
    boxes[i].score = all[i].score; boxes[i].label = all[i].label;
    if (show)
    {
        cv::rectangle(bgr, cv::Rect(cv::Point((int)x0,(int)y0), cv::Point((int)x1,(int)y1)), cv::Scalar(0,255,0), 2);
    }
}
if (show)
{
    cv::imshow("xd_yolov8_ncnn", bgr);
    cv::waitKey(0);
    cv::destroyAllWindows();
}
return count;

}

extern "C" __declspec(dllexport) void WINAPI XD_FreeModel(int model_index)
{
std::unique_ptr<XD_Model> old;
{
std::lock_guardstd::mutex lk(g_mutex);
if (model_index < 0 || model_index >= (int)g_models.size()) return;
old = std::move(g_models[model_index]);
g_models[model_index].reset();
}
#if NCNN_VULKAN
if (old && old->use_gpu)
{
if (old->staging_vkallocator) { delete old->staging_vkallocator; old->staging_vkallocator = nullptr; }
if (old->workspace_vkallocator && old->workspace_vkallocator != old->blob_vkallocator) { delete old->workspace_vkallocator; old->workspace_vkallocator = nullptr; }
if (old->blob_vkallocator) { delete old->blob_vkallocator; old->blob_vkallocator = nullptr; }
if (g_gpu_ref.fetch_sub(1) == 1) ncnn::destroy_gpu_instance();
}
#endif
}已经按照你说的方法修改了代码 还是会出现 (base) PS D:\Desktop\v8_ncnn> ./build/Release/Release/test_xd_dll.exe
Loading CPU model...
CPU model index: 0
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] queueC=0[1] queueG=0[1] queueT=0[1]
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] subgroup=32(832) ops=1/1/1/1/1/1/1/1/1/1
[0 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0
[1 NVIDIA GeForce RTX 4060 Laptop GPU] queueC=2[8] queueG=0[16] queueT=1[2]
[1 NVIDIA GeForce RTX 4060 Laptop GPU] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[1 NVIDIA GeForce RTX 4060 Laptop GPU] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[1 NVIDIA GeForce RTX 4060 Laptop GPU] subgroup=32(3232) ops=1/1/1/1/1/1/1/1/1/1
[1 NVIDIA GeForce RTX 4060 Laptop GPU] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/1/1/1
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] queueC=0[1] queueG=0[1] queueT=0[1]
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] subgroup=32(832) ops=1/1/1/1/1/1/1/1/1/1
[2 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] queueC=0[1] queueG=0[1] queueT=0[1]
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] subgroup=32(832) ops=1/1/1/1/1/1/1/1/1/1
[3 Intel(R) RaptorLake-S Mobile Graphics Controller] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0
(base) PS D:\Desktop\v8_ncnn>
(base) PS D:\Desktop\v8_ncnn>

[dosubot]

你的代码已经非常接近 ncnn 官方推荐的 Vulkan 资源管理模式，但你遇到的 GPU 崩溃（-1073741819）很可能和 Windows 下 DLL 卸载时机、Vulkan 资源释放顺序、或编译环境有关。结合 ncnn 源码和大量类似案例，建议你重点排查以下几个方向：

1. DLL 卸载与资源释放顺序
   ncnn 官方明确要求：所有 ncnn::Net 实例必须在调用 ncnn::destroy_gpu_instance() 之前析构，否则极易崩溃（尤其是 Windows/NVIDIA 驱动）参考。如果 destroy_gpu_instance() 被 DLL 卸载时的全局析构器触发，往往已经太晚，驱动内部资源已被系统清理，导致访问冲突。建议：确保所有模型和分配器都在 DLL 卸载前手动释放，并且 destroy_gpu_instance() 必须在所有 Net 析构后主动调用。

2. ODR 问题导致的崩溃
   Windows 下如果 _WIN32_WINNT 宏定义不一致，可能导致 ncnn::Mutex 等类出现 One Definition Rule (ODR) 冲突，进而引发运行时崩溃参考。建议：确保所有编译单元都用统一的 _WIN32_WINNT 宏，并且 windows.h/platform.h 的包含顺序一致。

3. 建议的调试方法

   • 打开 ncnn 的 NCNN_LOGE 日志和 ENABLE_VALIDATION_LAYER（gpu.cpp 里 #define ENABLE_VALIDATION_LAYER 1），能获得详细的 Vulkan 报错和崩溃堆栈，有助于定位具体是哪一步出错。
   • 在 XD_LoadModel、XD_Predict、XD_FreeModel 每一步前后加日志，确认崩溃发生的精确位置（模型加载、推理、还是资源释放）。
   • 用 ncnn 官方 minimal GPU 推理 demo（如 examples/squeezenet_vulkan.cpp）在你的环境下单独测试，排除模型或环境问题。

4. 模型兼容性与分配器
   你的分配器创建和释放方式基本正确，但建议用 vkdev->acquire_blob_allocator()/acquire_staging_allocator()，并用 vkdev->reclaim_blob_allocator() 归还，避免野指针或重复释放。

5. 其他建议

   • 检查模型文件是否损坏，或是否包含 Vulkan 不支持的层（如 PriorBox）。
   • 确认 ncnn 版本为最新，避免踩到已修复的崩溃 bug。
   • 如果用 OpenMP，建议设置环境变量 OMP_WAIT_POLICY=passive，防止 DLL 卸载时线程池未退出导致崩溃参考。

核心建议：
请务必保证 destroy_gpu_instance() 在所有 Net/分配器析构后、DLL 卸载前手动调用，避免依赖全局析构器或 DLL 卸载自动清理。否则 Windows 下 NVIDIA 驱动极易崩溃。

如仍无法定位，建议贴出开启 NCNN_LOGE 和 ENABLE_VALIDATION_LAYER 后的详细日志，或用调试器定位崩溃堆栈。这样能更快锁定是 Vulkan 初始化、模型加载还是资源清理阶段出错。

相关案例和官方说明可参考：

• ncnn::destroy_gpu_instance 必须在 Net 析构后调用
• ODR 问题导致的崩溃
• Vulkan 资源管理官方说明

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}