Hi !
Yes, I know, I lied, I said that my next article will be about buffers or images, but, finally, I’d prefer to talk about barriers first. However, barriers are, IMHO, a really difficult thing to well understand, so, this article might countain some mistakes.
In that case, please, let me know it by mail, or by one comment.
By the way, this article could remind you in some parts the article on GPU Open : Performance Tweets series: Barriers, fences, synchronization and Vulkan barriers explained
What memory barriers are for?
Memory barriers are source of bugs.
More seriously, barriers are used for three (actually four) things.
- Execution Barrier (synchronization) : To ensure that prior commands has finished
- Memory Barrier (memory visibility / availability): To ensure that prior writes are visible
- Layout Transitioning (useful for image) : To Optimize the usage of the resource
- Reformatting
I am not going to talk about reformating because (it is a shame) I am not very confident with it.
What exactly is an execution barrier ?
An execution barrier could remind you mutex on CPU thread. You write something in one resource. When you want to read what you write in, you must wait the write is finished.
What exactly is a memory barrier ?
When you write something from one thread, it could write it on some caches and you must flush them to ensure the visibility where you want to read that data. That is what memory barriers are for.
They ensure as well layout transition for image to get the best performance your graphic card can.
How it is done in Vulkan
Now that we understand why barriers are so important, we are going to see how can we use them in Vulkan.
Vulkan’s Pipeline
To be simple, the command enters in the top_of_pipe stage and end at bottom_of_pipe stage.
It exists an extra stage that refers to the host.
Barriers between stages
We are going to see two examples (that are inspired from GPU Open).
We will begin with the worse case : your first command writes at each stage everywhere it is possible, your second command reads at each stage everywhere it is possible.
It simply means that you want to wait for the first command totally finish before the second one begin.
To be simple, with a scheme it means that :
- In gray : All the stages that need to be executed before or after the barrier (or the ones that are never reached)
- In red : Above the barrier, it means where the data are produced. Below the barrier, it means where the data are consumed.
- In green : They are unblocked stages. You should try to have the maximum green stages as possible.
As you can see, here, you don’t have any green stages, so it is not good at all for performances.
In Vulkan C++, you should have something like that:
cmd.pipelineBarrier( vk::PipelineStageFlagBits::eAllCommands, vk::PipelineStageFlagBits::eAllCommands, ...);
Some people use BOTTOM_OF_PIPE as source and TOP_OF_PIPE as the destination. It is not false, but it is useful only for execution barrier. These stages do not access memory, so they can’t make memory access visible or even available!!!! You should not (must not?) issue a memory barrier on these stages, but we are going to see that later.
Now, we are going to see a better case
Imagine your first command fills an image or one buffer (SSBO or imageStore) through the VERTEX_SHADER. Now imagine you want to use these data in EVALUATION_SHADER.
The prior scheme, after modification, is :
As you can see, there is a lot of green stages and it is very good!
The Vulkan C++ code should be:
cmd.pipelineBarrier( vk::PipelineStageFlagBits::eVertexShader, vk::PipelineStageFlagBits::eTessellationEvaluationShader,...);
By Region or not?
This part may contain errors, so please, let me know if you disagree with me
To begin, what does by region means?
A region is a little part of your framebuffer. If you specify to use by region dependency, it means that (in fragment buffer space) operations need to be finished only in the region (that is specific to the implementation) and not in the whole image.
Well, it is not clear what is a fragment buffer space. In my opinion, and after reading the documentation, it could be from the EARLY_TEST (or at least FRAGMENT_SHADER if early depth is not enabled) to the COLOR_ATTACHMENT.
Actually, to me this flag lets the driver to optimize a bit. However, it must be used only (and should not be useful elsewhere IMHO) between subpasses for subpasses input attachments).
But I may be wrong !
Everything above about is wrong, if you want a plain explanation, see the comment from devsh. To make it simple, it means that the barrier will operate only on “one pixel” of the image. It could be used for input attachment or pre depth pass for example
Memory Barriers
Okay, now that we have seen how make a pure execution barrier (that means without memory barriers).
Memory barriers ensure the availability for the first half memory dependency and the visibility for the second one. We can see them as a “flushing” and “invalidation”. Make information available does not mean that it is visible.
In each kind of memory barrier you will have a srcAccessMask and a dstAccessMask.
How do they work?
Access and stage are somewhat coupled. For each stage of srcStage, all memory accesses using the set of access types defined in srcAccessMask will be made available. It can be seen as a flush of caches defined by srcAccessMask in all stages.
For dstStage / dstAccess, it is the same thing, but instead to make information available, the information is made visible for these stages and these accesses.
That’s why using BOTTOM/TOP_OF_PIPELINE is meaningless for memory barrier.
For buffer and image barriers, you could as well perform a “releasing of ownership” from a queue to another of the resource you are using.
An example, you transfer the image in your queue that is only used for transfers. At the end, you must perform a releasing from the transfer queue to the compute (or graphic) queue.
Global Memory Barriers
These kind of memory barriers applies to all memory objects that exist at the time of its execution.
I do not have any example of when to use this kind of memory barrier. Maybe if you have a lot of barriers to do, it is better to use global memory barriers.
An example:
vk::MemoryBarrier( vk::AccessFlagBits::eMemoryWrite, vk::AccessFlagBits::eMemoryRead);
Buffer Memory Barriers
Here, accessesMask are valid only for the buffer we are working on through the barrier.
Here is the example :
vk::BufferMemoryBarrier( vk::AccessFlagBits::eTransferWrite, vk::AccessFlagBits::eShaderRead, transferFamillyIndex, queueFamillyIndex, 0, VK_WHOLE_SIZE);
Image Memory Barriers
Image memory barriers have another kind of utility. They can perform layout transitions.
Example:
I want to create mipmaps associated to one image (we will see the complete function in another article) through vkCmdBlitImage.
After a vkCmdBlitImage, I want use the mipmap I just wrote as a source for the next mipmap level.
oldLayout must be DST_TRANSFER and newLayout must be SRC_TRANSFER.
Which kind of access I made and which kind of access I will do?
That is easy, I performed a TRANSFER_WRITE and I want to perform a TRANSFER_READ.
At each stage my last command “finish” and at each stage my new command “begin”? Both in TRANSFER_STAGE.
In C++ it is done by something like that:
cmd.blitImage(); vk::ImageMemoryBarrier imageBarrier( vk::AccessFlagBits::eTransferWrite, vk::AccessFlagBits::eTransferRead, vk::ImageLayout::eTransferDstOptimal, vk::ImageLayout::eTransferSrcOptimal, 0, 0, image, subResourceRange); cmd.pipelineBarrier( vk::PipelineStageFlagBits::eTransfer, vk::PipelineStageFlagBits::eTransfer, vk::DependencyFlags(), nullptr, nullptr, imageBarrier);
I hope that you enjoyed that article and that you have learned some things. Synchronization through Vulkan is not as easy to handle and all I wrote may (surely?) contains some errors.