Link

Before we go to implement textures, there is something we need to prepare first. Right now, we are implementing vertex buffers by storing them in CPU_TO_GPU memory type. This is doable, but it’s not the recommended way of dealing with mesh data. For vertex buffers, you want to have them accessible in the fastest memory type possible, which is GPU_ONLY memory type. The problem is that you cant write to it directly from the CPU.

To send data into GPU only memory, you need to first copy your data into a CPU writeable buffer, encode a copy command in a VkCommandBuffer, and then submit that command buffer to a queue. This will enqueue the transfer of the memory from your CPU writeable buffer into another buffer, which can be a GPU allocated buffer.

Because this is necessary when dealing with textures, we will implement this copy logic in our upload_mesh function. This will likely cause an immediate rendering speed-up if you have been trying to load heavy meshes.

Upload Context

As copying meshes from CPU buffer to GPU buffer won’t be the only thing we are going to do, we are going to create a small abstraction for this sort of short-lived commands.

Let’s begin by adding a new struct to VulkanEngine, and a function for immediate command execution.

struct UploadContext {
	VkFence _uploadFence;
	VkCommandPool _commandPool;
	VkCommandBuffer _commandBuffer;
};
class VulkanEngine {
public:
	UploadContext _uploadContext;

	void immediate_submit(std::function<void(VkCommandBuffer cmd)>&& function);
}

We are going to store the upload related structures in the struct, to keep the amount of objects in VulkanEngine class better organized. The immediate_submit() function uses an std::function in a very similar way as we do in the deletion queue.

Eventually we will add more instant-submit functions, but this one will be the default one.

We need to initialize that command pool and fence in the upload context. In init_sync_structures(), we will initialize the fence alongside the rendering fences that we have.


	VkFenceCreateInfo uploadFenceCreateInfo = vkinit::fence_create_info();

	VK_CHECK(vkCreateFence(_device, &uploadFenceCreateInfo, nullptr, &_uploadContext._uploadFence));
	_mainDeletionQueue.push_function([=]() {
		vkDestroyFence(_device, _uploadContext._uploadFence, nullptr);
	});

On this fence we won’t set the VK_FENCE_CREATE_SIGNALED_BIT flag, as we will not try to wait on it before sending commands like we do in the render loop.

Before we go further, lets define helpful initializers for beginning a command buffer and filling out VkSubmitInfo

// header
	VkCommandBufferBeginInfo command_buffer_begin_info(VkCommandBufferUsageFlags flags = 0);
	VkSubmitInfo submit_info(VkCommandBuffer* cmd);

// implementation
VkCommandBufferBeginInfo vkinit::command_buffer_begin_info(VkCommandBufferUsageFlags flags /*= 0*/)
{
	VkCommandBufferBeginInfo info = {};
	info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
	info.pNext = nullptr;

	info.pInheritanceInfo = nullptr;
	info.flags = flags;
	return info;
}

VkSubmitInfo vkinit::submit_info(VkCommandBuffer* cmd)
{
	VkSubmitInfo info = {};
	info.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
	info.pNext = nullptr;

	info.waitSemaphoreCount = 0;
	info.pWaitSemaphores = nullptr;
	info.pWaitDstStageMask = nullptr;
	info.commandBufferCount = 1;
	info.pCommandBuffers = cmd;
	info.signalSemaphoreCount = 0;
	info.pSignalSemaphores = nullptr;

	return info;
}

Now we create the command pool and allocate a command buffer from it, which we do in init_commands

	VkCommandPoolCreateInfo uploadCommandPoolInfo = vkinit::command_pool_create_info(_graphicsQueueFamily);
	//create pool for upload context
	VK_CHECK(vkCreateCommandPool(_device, &uploadCommandPoolInfo, nullptr, &_uploadContext._commandPool));

	_mainDeletionQueue.push_function([=]() {
		vkDestroyCommandPool(_device, _uploadContext._commandPool, nullptr);
	});

	//allocate the default command buffer that we will use for the instant commands
	VkCommandBufferAllocateInfo cmdAllocInfo = vkinit::command_buffer_allocate_info(_uploadContext._commandPool, 1);
	VK_CHECK(vkAllocateCommandBuffers(_device, &cmdAllocInfo, &_uploadContext._commandBuffer));

Right now we are creating the pool using the graphics queue family. This is because we will submit the commands to the same queue as the graphics one.

With that the command pool and fence initialized, let’s write the code for the immediate_submit function


void VulkanEngine::immediate_submit(std::function<void(VkCommandBuffer cmd)>&& function)
{
	VkCommandBuffer cmd = _uploadContext._commandBuffer;

	//begin the command buffer recording. We will use this command buffer exactly once before resetting, so we tell vulkan that
	VkCommandBufferBeginInfo cmdBeginInfo = vkinit::command_buffer_begin_info(VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT);

	VK_CHECK(vkBeginCommandBuffer(cmd, &cmdBeginInfo));

	//execute the function
	function(cmd);

	VK_CHECK(vkEndCommandBuffer(cmd));

	VkSubmitInfo submit = vkinit::submit_info(&cmd);


	//submit command buffer to the queue and execute it.
	// _uploadFence will now block until the graphic commands finish execution
	VK_CHECK(vkQueueSubmit(_graphicsQueue, 1, &submit, _uploadContext._uploadFence));

	vkWaitForFences(_device, 1, &_uploadContext._uploadFence, true, 9999999999);
	vkResetFences(_device, 1, &_uploadContext._uploadFence);

	// reset the command buffers inside the command pool
	vkResetCommandPool(_device, _uploadContext._commandPool, 0);
}

This is very similar to the logic that we do in the render loop. The most important one is that we are reusing the same command buffer from frame to frame. If we wanted to submit multiple command buffers, it would simply require allocating as many command buffers as we want ahead of time and reusing them.

We first allocate command buffer, we then call the function between begin/end command buffer, and then we submit it. Then we wait for the submit to be finished, and reset the command pool.

With this, we now have a way of instantly executing some commands to the GPU, without dealing with the render loop and other synchronization. This is great for compute calculations, and, if it submitted into a different queue, you could use this from a background thread, separated from the render loop.

Transferring memory.

Now that we have the system for instant commands, we will rewrite the function for upload_mesh() so that it uploads the mesh to a GPU local buffer, for best speed.

First, we need to allocate a CPU side buffer to hold the mesh data before uploading it to the GPU buffer.

void VulkanEngine::upload_mesh(Mesh& mesh)
{
	const size_t bufferSize= mesh._vertices.size() * sizeof(Vertex);
	//allocate staging buffer
	VkBufferCreateInfo stagingBufferInfo = {};
	stagingBufferInfo.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
	stagingBufferInfo.pNext = nullptr;

	stagingBufferInfo.size = bufferSize;
	stagingBufferInfo.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;

	//let the VMA library know that this data should be on CPU RAM
	VmaAllocationCreateInfo vmaallocInfo = {};
	vmaallocInfo.usage = VMA_MEMORY_USAGE_CPU_ONLY;

	AllocatedBuffer stagingBuffer;

	//allocate the buffer
	VK_CHECK(vmaCreateBuffer(_allocator, &stagingBufferInfo, &vmaallocInfo,
		&stagingBuffer._buffer,
		&stagingBuffer._allocation,
		nullptr));
}

We are creating the stagingBuffer with enough size to hold the mesh data, and giving it the VK_BUFFER_USAGE_TRANSFER_SRC_BIT usage flag. This flag tells Vulkan that this buffer will only be used as source for transfer commands. We won’t be using the staging buffer for rendering.

We can now copy the mesh data to this buffer

//copy vertex data
	void* data;
	vmaMapMemory(_allocator, stagingBuffer._allocation, &data);

	memcpy(data, mesh._vertices.data(), mesh._vertices.size() * sizeof(Vertex));

	vmaUnmapMemory(_allocator, stagingBuffer._allocation);

Similar map/unmap buffer logic as always. This is unchanged from the last version of upload_mesh().

With the vertex buffer now in a Vulkan CPU side buffer, we need to create the actual GPU-side buffer.

	//allocate vertex buffer
	VkBufferCreateInfo vertexBufferInfo = {};
	vertexBufferInfo.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
	vertexBufferInfo.pNext = nullptr;
	//this is the total size, in bytes, of the buffer we are allocating
	vertexBufferInfo.size = bufferSize;
	//this buffer is going to be used as a Vertex Buffer
	vertexBufferInfo.usage = VK_BUFFER_USAGE_VERTEX_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT;

	//let the VMA library know that this data should be GPU native
	vmaallocInfo.usage = VMA_MEMORY_USAGE_GPU_ONLY;

	//allocate the buffer
	VK_CHECK(vmaCreateBuffer(_allocator, &vertexBufferInfo, &vmaallocInfo,
		&mesh._vertexBuffer._buffer,
		&mesh._vertexBuffer._allocation,
		nullptr));

We give the buffer the VK_BUFFER_USAGE_VERTEX_BUFFER_BIT and VK_BUFFER_USAGE_TRANSFER_DST_BIT usage flags, so that the drivers knows that we will use this to render meshes, and to copy data into. We also make sure that VMA allocates it on GPU VRAM by using the VMA_MEMORY_USAGE_GPU_ONLY memory type.

With the buffers created, we can now execute the copy command.

immediate_submit([=](VkCommandBuffer cmd) {
		VkBufferCopy copy;
		copy.dstOffset = 0;
		copy.srcOffset = 0;
		copy.size = bufferSize;
		vkCmdCopyBuffer(cmd, stagingBuffer._buffer, mesh._vertexBuffer._buffer, 1, &copy);
	});

We use the immediate_submit function to enqueue a vkCmdCopyBuffer() command. this command will copy regions of one buffer into another buffer, using VkBufferCopy for each region detail. In here, we are copying the entire staging buffer into the vertex buffer.

With the memory now uploaded, we can clean up.

//add the destruction of mesh buffer to the deletion queue
_mainDeletionQueue.push_function([=]() {
	vmaDestroyBuffer(_allocator, mesh._vertexBuffer._buffer, mesh._vertexBuffer._allocation);
	});

vmaDestroyBuffer(_allocator, stagingBuffer._buffer, stagingBuffer._allocation);

For the GPU vertex buffer, we will add it to the deletion queue, but for the staging buffer, we will instantly delete it as we are done with it.

Try to run the engine now, everything should work completely fine.

With this new mesh loading code, you will have great performance even uploading meshes that are tens of millions of triangles.

All the logic for copying data is done for meshes, so it’s time to continue with textures.

Next: Vulkan Images