vulkan backend implementation 6: some optimization

[rinthel]

Hello! This is the 6th PR of my Vulkan backend implementation (#274)

Changes

Refactoring & optimizing descriptor set binding

• Uniform buffer objects which are set by setUniform() or predefined are now using dynamic uniform buffer.
• Thus descriptor set can be reused if renderBind information or pipeline layout is not changed.
• Allocating and writing descriptor sets is refactored into a function.
• I also tried to cache descriptor set like d3d12 implementation (using something like bindLru in the d3d12 backend), but failed. Maybe descriptor sets are already in a descriptor pool, so allocation and free is fast enough to be in a frame.
• After this optimization, example-17 now start to work, and active scissor rect drawing in example-21 also works. Perhaps the main reason that these examples don't work is the lack of descriptor sets (1024 for each frame)

Some minor change & bug fix

• Supports to indirect compute call (for enabling example-24 indirect call option)
• Image layout is reinitialized when texture is released. This fixes example-16 crash when switching a light source.
• Get / set device features when creating logical device, and check independent blending feature. This fixes example-19 crash when switching MRT-independent option.
• Create depth view for depth/stencil texture.

Thank you!

[cforfang]

Passing the output of vkGetPhysicalDeviceFeatures directly like this will unnecessary enable e.g. robustBufferAccess if available, which has a performance impact. I'd consider turning that off explicitly here (or only enable the features you actually want). :)

[rinthel]

Thanks for pointing!

[rinthel]

@cforfang Then how about this on line 1160?

VkPhysicalDeviceFeatures deviceFeatures;
vkGetPhysicalDeviceFeatures(m_physicalDevice, &deviceFeatures);
bx::memSet(&m_deviceFeatures, 0, sizeof(VkPhysicalDeviceFeatures));
m_deviceFeatures.independentBlend = deviceFeatures.independentBlend;

[cforfang]

While I think that would work, I'm not 100% if not enabling things like textureCompressionETC2 / textureCompressionASTC_LDR / textureCompressionBC might either (a) literally turn off support in the drivers, or (b) cause validation layers to complain.

It might be simpler to just enable everything available except robustBufferAccess as I don't know of any other features with a performance impact.