ref: https://devtalk.nvidia.com/default/topic/1070081/cuda-programming-and-performance/reverse-lut-for-lop3-lut/

video_core: memory_manager: Flush/invalidate asynchronously when possible.

Update httplib to 0.5.5

shader/texture: Allow 2D shadow arrays and simplify code

- Minor perf improvement.

maxwell_3d: Unify draw methods

gl_rasterizer: Use the least generic OpenGL draw function possible

shader_conversion: add conversion I2F for Short

vk_shader_decompiler: Fix vertex id and instance id

query_cache: Implement a query cache and query 21 (samples passed)

GUI: Add aspect ratio dropdown

Fix CMake build errors for certain SDL2 installations

nvhost_gpu: implement ChannelSetTimeslice

IUserLocalCommunicationService: stub function Initialize2

402: function name should be Initialize2 (7.0.0+) not SetOperationMode
Follow by: https://switchbrew.org/wiki/LDN_services#IUserLocalCommunicationService

Revert "video_core: memory_manager: Use GPU interface for cache functions."

Shadow sampler 2D arrays are supported on OpenGL, so there's no reason
to forbid these. Enable textureLod usage on these.

Minor style changes.

Pass instanced state of a draw invocation as an argument instead of
having two separate virtual methods.

Abstract the current OpenGL implementation into the VideoCommon
namespace and reimplement it on top of that. Doing this avoids repeating
code and logic in the Vulkan implementation.

Use a custom cache instead of relying on a ranged cache.

Instead of waiting immediately for executed commands, defer the query
until the guest CPU reads it. This way we get closer to what the guest
program is doing.

To archive this we have to build a dependency queue, because host APIs
(like OpenGL and Vulkan) use ranged queries instead of counters like
NVN.

Waiting for queries implicitly uses fences and this requires a command
being queued, otherwise the driver will lock waiting until a timeout. To
fix this when there are no commands queued, we explicitly call glFlush.