vkGetBufferMemoryRequirements2KHR crash

[f32by]

I'm using Vulkan Memory Allocator and the program crashes when calling vkGetBufferMemoryRequirements2KHR(https://github.com/GPUOpen-LibrariesAndSDKs/VulkanMemoryAllocator/blob/master/src/vk_mem_alloc.h#L16489).

Following code is taken from VMA with some modifications:

VkBuffer hBuffer;

VkBufferCreateInfo creator = {
      .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
      .size = 256,
      .usage = VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT,
      .sharingMode = VK_SHARING_MODE_EXCLUSIVE
  };

// success
vkCreateBuffer(m_hDevice, &creator, nullptr, &buffer);

VkBufferMemoryRequirementsInfo2KHR memReqInfo = { VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2_KHR };

memReqInfo.buffer = hBuffer;

VkMemoryDedicatedRequirementsKHR memDedicatedReq = { VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR };

VkMemoryRequirements2KHR memReq2 = { VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2_KHR };
VmaPnextChainPushFront(&memReq2, &memDedicatedReq);

// crash
vkGetBufferMemoryRequirements2KHR(m_hDevice, &memReqInfo, &memReq2);

I'm using a MacBook Pro 16-inch 2019 and the program is linked against MoltenVK statically.

[billhollings]

After modifying your code to the code below, to replace VmaPnextChainPushFront, and then adding it to cube.c in the Cube Demo in MoltenVK, I do not encounter a crash.

What is the nature of the crash? Can you provide a stack call trace?

	VkBuffer hBuffer;

	VkBufferCreateInfo creator = {
		.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
		.size = 256,
		.usage = VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT,
		.sharingMode = VK_SHARING_MODE_EXCLUSIVE
	};

	// success
	vkCreateBuffer(demo->device, &creator, NULL, &hBuffer);

	VkBufferMemoryRequirementsInfo2KHR memReqInfo = { VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2_KHR };
	memReqInfo.buffer = hBuffer;

	VkMemoryDedicatedRequirementsKHR memDedicatedReq = { VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR };

	VkMemoryRequirements2KHR memReq2 = { VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2_KHR };

//	VmaPnextChainPushFront(&memReq2, &memDedicatedReq);
	memReq2.pNext = &memDedicatedReq;

	// crash
	vkGetBufferMemoryRequirements2KHR(demo->device, &memReqInfo, &memReq2);

[f32by]

On my Mac the created hBuffer is 0x0000290000000029 which is an invalid address and the program crashes when casting pInfo->buffer to MVKBuffer.

[billhollings]

Can you provide some further context?

Is that crash report a result of running my code above? Or your code above, which doesn't set hBuffer before using it in VkBufferMemoryRequirementsInfo2KHR. Or VMA?

Specifically, is the pInfo->buffer value assigned directly from what is returned from vkCreateBuffer()? In other words...is the value returned from vkCreateBuffer() that same bad 0x0000290000000029 pointer?

[billhollings]

And just to be certain, when you say you link against MoltenVK statically, are you linking directly to MoltenVK, or through the Vulkan SDK?

[f32by]

It's linked through the Vulkan SDK.
vkCreateBuffer() returns 0x0000290000000029.

I tried the latest Vulkan SDK(MoltenVK version 1.1.2. Vulkan version 1.1.162) and it still crashes.

[billhollings]

Okay. It's likely an SDK issue then.

Each layer in the SDK has different Vulkan handle types, so that 0x0000290000000029 handle is probably arising in one of the layers. It needs to be submitted back to that same layer for interpretation. It's possible that somewhere in the mix of VMA and SDK, the object is being submitted to the wrong layer and ending up coming into MoltenVK incorrectly.

Can you post a link to this issue in the Vulkan SDK issue list, please, so they can have a look?

BTW...Vulkan SDK 1.2.162 should be using MoltenVK 1.1.1, not 1.1.2.

[f32by]

Hi, finally I found this is 100% my mistake. I linked against both MoltenVK and the official Vulkan SDK. Now the problem is solved. Thank you for your help :)