After I implemented the render pass wrapper, seeing how the
RenderPassCreateInfo structure and its dependencies were HUGE compared
to the actual tiny and lean RenderPass, I felt uneasy dragging their
definition along to every place where a RenderPass gets used. It's not
as bad with the others, but as new extensions are implemented I expect
that to get the same.
This change makes it easier for me to accept that Image.h / Buffer.h
depends on Memory.h. There isn't a real measurable difference when
building Magnum itself (50 ms out of 7 seconds for the Vk library
alone), but that's because most of the code (and tests) needs the
CreateInfo structures anyway.
You won't believe it, but it took me over a month of sitting on the
shitter until this design idea materialized out of [..] air. The whole
story, in order:
- Vulkan doesn't allow one VkDeviceMemory to be mapped more than once.
This is rather sad, because since Vulkan best practices suggest to
allocate a large block and suballocate from that, the engine needs
an extra layer that "emulates" mapping the suballocations for the
users but behind the scenes it inevitably has to map the whole
VkDeviceMemory anyway and keep it mapped for as long as any of the
sub-mappings is active.
- Because if it would map just a certain suballocation and then the
user would want to map another suballocation, it would have to
discard the original mapping and create a new one spanning both
suballocations and that has a risk of suddenly being in a different
VM block, making all pointers to the previous mapping invalid.
- The Vulkan Memory Allocator implements this approach of mapping the
whole thing and because of all the bookkeeping it doesn't give a
direct access to the underlying VkDeviceMemory, making it rather
hard to integrate.
Here I realized that:
- Most allocations won't need to be mapped ever, so the hiding and
obfuscation done by VMA isn't needed for those --- and we want
interoperability with 3rd party code, so preventing access to
VkDeviceMemory is out of question.
- There's KHR_dedicated_allocation, which (probably?) wasn't around
when VMA was originally designed. The extension was created because
a dedicated allocation actually *does* make sense in certain
cases and on certain architectures. Providing a way to make those
thus shouldn't be something "temporary, until a real allocator
exists" but rather a well-designed API that's there to stay.
- Except for iGPUs, the usual way to populate a GPU buffer would be to
first copy the data to some host-accessible scratch buffer and then
do a GPU-side copy of that buffer to a device-local memory. The
scratch buffer is very likely to have a vastly different
suballocation scheme than GPU buffers (grow & discard everything
once it's all uploaded, for example) so again trying to put the two
under the same allocator umbrella doesn't make sense.
Thus:
- To avoid implementing a full-blown allocator right from the start,
we'll first provide convenience APIs only for dedicated allocations
-- making it possible to transfer memory ownership to an
Image/Buffer so it can be treated the same way as in GL, and later
having the Image/Buffer constructor implicitly allocate a dedicated
VkDeviceMemory.
- This default allocation will be subsequently equipped with
KHR_dedicated_allocation bits.
- Thanks to the extensible/layered nature of the design, the user is
still capable of being completely in control of allocations,
managing VkDeviceMemory sub-allocations by hand.
Finally, once allocator APIs are figured out, the default Buffer/Image
behavior gets switched from a dedicated allocation to using an
allocator, and dedicated allocation will be only used if the
KHR_dedicated_allocation bit is requested.
Vladimír Vondruš