* When serialising out initial states of textures, we make a duplicate to
copy into. Since we use the storage APIs we need a sized format, so let's
convert the format early so the stored internal format is sized. This
doesn't cause a behaviour change as the impl is free to choose any sized
format and we're just reading back the choice, and making it explicit.
* If we do a partial replay of a frame, a query might be left open. We
track this and make sure we close any that were left open before the next
replay.
* This fixes issues where textures wouldn't render because they weren't
mipmap complete, and although I was point sampling a single mip the point
sampler was referencing the mipchain.
* We create fake backbuffer as SRGB if the original default backbuffer was
SRGB in the program.
* All of our output backbuffers are also SRGB, and we enable
GL_FRAMEBUFFER_SRGB when writing to them.
* Add the 'fake' srgb curve applied to linear data to display it as if it
were srgb, which tends to look intuitively correct even if it's not 100%
accurate. See the behaviour existing in D3D11 already.
* When we do glCopyImageSubData, this seems to require a complete texture
across all mips - ie. the full mip chain (up to MAX_LEVEL) must be set.
This requirement holds even if we only copy the mips that are present.
* However, for example attaching an image to a framebuffer, it doesn't
have to be mipmap complete even if the min/mag filters want mipmaps.
So in this case we have to force the texture to be complete enough for
our purposes, and we do this just by setting MAX_LEVEL to clamp the mips
* The general idea is that while idle we can handle these as normal -
If we ignore the map we ignore the map and mark the buffer dirty etc.
This is fine because modified but unflushed regions become undefined, and
we just let these become the modified values.
* While capturing a frame, we do everything to set up as normal, but then
instead of comparing the modified shadow buffer to an unmodified version,
detecting the modified range and making one big Unmap() chunk, we make a
chunk per Flush() call with the flushed range, and ignore the unmap.
* When glVertexAttrib*Pointer is called it 'latches' the current binding
of GL_ARRAY_BUFFER (it sets the attrib binding and vertex buffer on the
VAO internally).
* So when reading, we haven't serialised all the buffer bindings, so we
need to bind the right buffer by hand.
* If no program or pipeline is bound, need to make sure we don't try and
query a non-existant pipeline. This could happen if e.g. state is totally
cleared at the start of the frame.
baldurk