Hey Andrey2007AMD,
Here's the response from Omar Alam, one of our developers:
The app has the following code to select a depth format:
const VkFormat depthFormats[] = {
#ifdef EMBEDDED_SYSTEM
VK_FORMAT_D16_UNORM,
VK_FORMAT_D16_UNORM_S8_UINT,
#endif
VK_FORMAT_D24_UNORM_S8_UINT,
VK_FORMAT_D32_SFLOAT,
VK_FORMAT_D32_SFLOAT_S8_UINT,
};
const char* strFormats[] = {
#ifdef EMBEDDED_SYSTEM
"VK_FORMAT_D16_UNORM",
"VK_FORMAT_D16_UNORM_S8_UINT",
#endif
"VK_FORMAT_D24_UNORM_S8_UINT",
"VK_FORMAT_D32_SFLOAT",
"VK_FORMAT_D32_SFLOAT_S8_UINT",
};
for (size_t i = 0; i < countof(depthFormats); ++i) {
VkFormatProperties formatProps;
VkFormat format = depthFormats[i];
vkPhysicalDevice.GetPhysicalDeviceFormatProperties(format, formatProps);
// Format must support depth stencil attachment for optimal tiling
if (formatProps.optimalTilingFeatures & VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT) {
vkDepthSurfaceFormat = format;
LogMsg("Depth format %s\n", strFormats[i]);
break;
}
}
return vkDepthSurfaceFormat;
}
in the InitImageFormatProperties (VulkanDevice.cpp of provided code).
The format that gets selected based on the conditional in the for loop that iterates over all formats to select a suitable one is: VK_FORMAT_D32_SFLOAT which is a depth only format.
The problem occurs when the Stencil aspect of the image is being resolved. Basically, when this depth format gets selected it is a depth only format and attempting a stencil resolve causes a segmentation fault since no memory and addresses get allocated for the stencil component and invalid pointers are accessed leading to a crash. An easy fix for the code on the app side is to double check which image formats are selected to make sure a stencil component is also present if stencil resolve is needed.
At present, the Vulkan specification seems to be unclear on the expected behavior and legality of this operation (stencil resolve on a depth only image). I have verified that the crash can be avoided with other hardware configurations such as my current development machine (Ryzen 7 5700G + RX 460 dGPU) using the amd vulkan driver but that is purely down to the underlying implementation and how the resolve is set up. An issue has been filed with Khronos to determine the correct behavior in this scenario, but it is unclear for now.
