Workaround for memory types and linear images on Nvidia

[K0bin]

WSI Changes

• Respect creating presentable image after GRVK has presented once
• Record the copy command buffer on the fly instead of prerecording it
• Respect Surface Capabilities min & max extent and use presentable extent size if possible. This fixes an issue where GetClientRect would return a 0x0 rectangle.

General fixes

• Do not use STORE_OP_STORE when the pipeline has a 0 write mask. This breaks rendering of the BF4 options menu on Nvidia.
• Ensure that the buffer barrier range in grPrepareMemoryRegion is within the bounds of the buffer. This fixes a validation error in BF4.
• Added more formats to GetTexelSize and GetTileSize.
• Ensure a minimum size of 1x1 for textures and the viewport. I don't know if that's correct but it at least fixes a validation error in BF4.

Nvidia workarounds

Running BF4 with GRVK ran into the following limitations:

• Nvidia has 11 memory types while Mantle only supports 8. Most of those memory types are non-mappable system memory which supports a specific type of (optimal tiled) image.
• Nvidia only supports linear 2D images with 1 array layer and 1 mip map.

So in order to make it all work, the PR does the following:

• If the game tries to create a linear image that is not supported, we use optimal tiling for the image and also create a accompanying buffer.
• Track all usages of images with the workaround applied and copy from the buffer to the image at submission time and from the image to the buffer at the end of a command buffer.
• Because the optimal image and a host visible buffer can not go into the same memory type / mantle heap, we need to take control of memory allocations: To do this, this pulls in VMA and uses that for all memory code when enabled.

We expose the following heaps:

• Device local
• Host visible
• Host visible + cached
• Host visible + device local (when supported by the Vulkan driver)

The host visible heaps basically work the same as before. grAllocateMemory will allocate a block of memory (using VMA in this case) that will get bound to an object later.

Device local heaps on the other hand will not allocate in grAllocateMemory. Instead we allocate in grBindObjectMemory when we know what kind of object it is in order to use the right memory type. It also brings back the old behavior of lazily creating Vulkan buffers for GrGpuMemory objects because otherwise memory usage would effectively double. It will still create a buffer in grAllocMemory for host visible heaps and when running on AMD GPUs.

When binding an image with the workaround activated (read: it has both an image and a buffer), we do one of the following things depending on the requested heap:
If the heap is host visible, we bind the buffer to the memory, allocate a new chunk of memory specifically for the image (can even be device local) and bind that. If the heap is not host visible, we just allocate memory for the image and destroy the buffer that was created for that image (effectively turning it into a regular device local optimal image).

[K0bin]

Given that Mantle apparently also doesn't have uniform buffers, I'm not sure if this is worth it at all. Even if everything worked correctly, it would be way slower than DXVK on Nvidia hardware.

We'd have to patch shaders to turn buffers into uniform buffers when binding small, correctly aligned ranges of memory (no idea if that would even work) and I don't think anyone wants to implement that.

[sehraf]

Any chance to get this revisited?
My only (half recent) AMD card doesn't support the newer required Vulkan extensions.

I see your point regarding performance and raise your mine: preservation 📚

[K0bin]

It's a very invasive and ugly change and I don't think @libcg is terribly excited about it. (which I can't blame him for.)

[libcg]

@sehraf You might want to look into NimeZ drivers if you're on Windows. Or use RADV on Linux, it's the only driver currently recommended for GRVK.

@K0bin It's a whole lot of code that I can't test or maintain.. Would be good to revisit when GRVK is mature. I appreciate the effort btw.