It's implemented only in the Vulkan renderer, partly because there's a
lot of experimenting going on with it, but the glyphs do get transferred
to the GPU (checked in render doc). No rendering is done yet: still
thinking about whether to do a quick-and-dirty test, or to add HarfBuzz
immediately, and the design surrounding that.
R8G8B8A8 was hard-coded by accident when creating Vulkan_LoadTexArray
(or probably even the original Vulkan_LoadTex). This wasn't a problem
while everything was loaded in that format, but attempting to load an R8
texture didn't go so well. The same format as the image itself is used
now (correctly so).
I have recently learned that pre-multiplied alpha is the correct way to
do compositing, which is pretty much what the 2d pass does (actually,
all passes, but...). This required ensuring the color factor passed to
the fragment shader is pre-multiplied (a little silly for cshifts as
they used to be pre-multiplied but were un-pre-multiplied early in QF's
history and I don't feel like fixing that right now as it affects all
renderers), and also pre-multiplying alpha when converting from 8-bit
palette to rgba as the palette entry for transparent has that funky pink
(which is used in full-brights).
I will need to do more work to improve the 2d allocation, but rounding
up the requested sizes to the next power of two proved to be excessively
wasteful: I was able to allocate spots for only half of the sub-pics I
needed (though I did still need to double the number of pixels in the
end).
The software renderer uses Bresenham's line slice algorithm as presented
by Michael Abrash in his Graphics Programming Black Book Special Edition
with the serial numbers filed off (as such, more just so *I* can read
the code easily), along with the Chen-Sutherland line clipping
algorithm. The other renderers were more or less trivial in comparison.
Most were pretty easy and fairly logical, but gib's regex was a bit of a
pain until I figured out the real problem was the conditional
assignments.
However, libs/gamecode/test/test-conv4 fails when optimizing due to gcc
using vcvttps2dq (which is nice, actually) for vector forms, but not the
single equivalent other times. I haven't decided what to do with the
test (I might abandon it as it does seem to be UD).
This does mean that the gl and sw renderers can no longer call
S_ExtraUpdate, but really, they shouldn't be anyway. And I seem to
remember it not really helping (been way too long since quake ran that
slowly for me).
The texture animation data is compacted into a small struct for each
texture, resulting in much less data access when animating the texture.
More importantly, no looping over the list of frames. I plan on
migrating this to at least the other hardware renderers.
I found a test map with no texture data. Even after fixing the bsp
loader, vulkan didn't like it. Now vulkan is happy. The "Missing" text
is full-bright magenta on a dim grey background so it should be visible
in any lighting conditions.
Conflagrant Rodent has a sub-model with 0 faces (double bit error?)
causing simply counting faces to get out of sync with actual model
starts thus breaking *all* brush models that come after it (including
other maps). Thus be a little less lazy in figuring out model start
faces.
The models are broken up into N sub-(sub-)models, one for each texture,
but all faces using the same texture are drawn as an instance, making
for both reduced draw calls and reduced index buffer use (and thus,
hopefully, reduced bandwidth). While texture animations are broken, this
does mark a significant milestone towards implementing shadows as it
should now be possible to use multiple threads (with multiple index and
entid buffers) to render the depth buffers for all the lights.
This allows the use of an entity id to index into the entity data and
fetch the transform and colormod data in the vertex shader, thus making
instanced rendering possible. Non-world brush entities are still not
rendered, but the world entity is using both the entity data buffer and
entid buffer.
Sub-models and instance models need an instance data buffer, but this
gets the basics working (and the proof of concept). Using arrays like
this actually simplified a lot of the code, and will make it easy to get
transparency without turbulence (just another queue).
The gl water warp ones have been useless since very early on due to not
doing water warp in gl (vertex warping just didn't work well), and the
recent water warp implementation doesn't need those hacks. The rest of
the removed flags just aren't needed for anything. SURF_DRAWNOALPHA
might get renamed, but should be useful for translucent bsp surfaces
(eg, vines in ad_tears).
One more step towards BSP thread-safety. This one brought with it a very
noticeable speed boost (ie, not lost in the noise) thanks to the face
visframes being in tightly packed groups instead of 128 bytes apart,
though the sw render's boost is lost in the noise (but it's very
fill-rate limited).
This is next critical step to making BSP rendering thread-safe.
visframe was replaced with cluster (not used yet) in anticipation of BSP
cluster reconstruction (which will be necessary for dealing with large
maps like ad_tears).
The main goal was to get visframe out of mnode_t to make it thread-safe
(each thread can have its own visframe array), but moving the plane info
into mnode_t made for better data access patters when traversing the bsp
tree as the plane is right there with the child indices. Nicely, the
size of mnode_t is the same as before (64 bytes due to alignment), with
4 bytes wasted.
Performance-wise, there seems to be very little difference. Maybe
slightly slower.
The unfortunate thing about the change is the plane distance is negated,
possibly leading to some confusion, particularly since the box and
sphere culling functions were affected. However, this is so point-plane
distance calculations can be done with a single 4d dot product.
The map uses 41% of a 4k light map scrap, and 512 texture descriptors
wasn't enough for vulkan. Ouch. I do need to get cvars on these things,
but this will do for now (decades later...)
This was one of the biggest reasons I had trouble understanding the bsp
display list code, but it turns out it was for dealing with GLES's
16-bit limit on vertex indices. Since vulkan uses 32-bit indices,
there's no need for the extra layer of indirection. I'm pretty sure it
was that lack of understanding that prevented me from removing it when I
first converted the glsl bsp code to vulkan (ie, that 16-bit indices
were the only reason for elements_t).
It's hard to tell whether the change makes much difference to
performance, though it seems it might (noisy stats even over 50 timedemo
loops) and the better data localization indicate it should at least be
just as good if not better. However, the reason for the change is
simplifying the data structures so I can make bsp rendering thread-safe
in preparation for rendering shadow maps.
And maybe a nano-optimization. Switching from (~side + 1) to (-side)
seems to give glsl a very tiny speed boost, but certainly doesn't hurt.
Looking at some assembly output for the three cases, the two hacks seem
to generate the same code as each other, but 3 instructions vs 6 for ?:.
While ?: is more generically robust, the hacks are tuned for the
knowledge side is either 0 or 1. The later xor might alter things, but
at least I now know that the hack (either version) is worthwhile.
With experience, I have found that trying to continue after a validation
error tends to result in a segfault or some other nastiness, and
Sys_Shutdown (and the full shutdown sequence) is triggered for any error
signal (segfault, abort, etc) so just exit(1).
Some very much needed comments :P Still, nicely, I now have a much
better understanding of how the display lists are created (10 years
is a long time to remember how intricate code works (I do remember
fighting to get it working back then))
This makes it much easier to see just what is being done to build a
polygon to be passed to the GPU, and it served as a test for the
lightmap st changes since Vulkan currently never used them.
Many modders use negative lights for interesting effects, but vulkan
doesn't like the result of a negative int treated as unsigned when it
comes to texture sizes.
However, this time it doesn't modify the light array when it sorts the
lights by size since the lights are now located before the renderer gets
to see them, and having the fix up the light leafs array would be too
painful (and probably the completely wrong thing to do anyway: the light
array should be treated as constant by the renderer). 1.6GB of memory
for gmsp3v2's lights (a little better than marcher: more smaller lights?).
For reference:
gmsp3v2: shadow maps: 8330 layers in 29 images: 1647706112
marcher: shadow maps: 2440 layers in 11 images: 2358575104
For now, at least (I have some ideas to possibly reduce the numbers and
also to avoid the need for actual limits). I've seen gmsp3v2 use over
500 lights at once (it has over 1300), and I spent too long figuring out
that weird light behavior was due to the limit being hit and lights
getting dropped (and even longer figuring out that more weird behavior
was due to the lack of shadows and the world being too bright in the
first place).
Since the staging buffer allocates the command buffers it uses, it
needs to free them when it is freed. I think I was confused by the
validation layers not complaining about unfreed buffers when shutting
down, but that's because destroying the pool (during program shutdown,
when the validation layers would complain) frees all the buffers. Thus,
due to staging buffers being created and destroyed during the level load
process, (rather large) command buffers were piling up like imps in a
Doom level.
In the process, it was necessary to rearrange some of the shutdown code
because vulkan_vid_render_shutdown destroys the shared command pool, but
the pool is required for freeing the command buffers, but there was a
minor mess of long-lived staging buffers being freed afterwards. That
didn't end particularly well.
This is meant for a "permanent" tear-down before freeing the memory
holding the VM state or at program shutdown. As a consequence, builtin
sub-systems registering resources are now required to pass a "destroy"
function pointer that will be called just before the memory holding
those resources is freed by the VM resource manager (ie, the manager
owns the resource memory block, but each subsystem is responsible for
cleaning up any resources held within that block).
This even enhances thread-safety in rua_obj (there are some problems
with cmd, cvar, and gib).
This gives a rather significant speed boost to timedemo demo1: from
about 2300-2360fps up to 2520-2600fps, at least when using
multi-texture.
Since it was necessary for testing the scrap, gl got the ability to set
the console background texture, too.
While it takes one extra step to grab the marksurface pointer,
R_MarkLeaves and R_MarkLights (the two actual users) seem to be either
the same speed or fractionally faster (by a few microseconds). I imagine
the loss gone to the extra fetch is made up for by better bandwidth
while traversing the leafs array (mleaf_t now fits in a single cache
line, so leafs are cache-aligned since hunk allocations are aligned).
Unfortunately, the animations are pre-baked (by the loader) blocking
run-time determined animations (IK etc). However, this at least gets
everything working so the basics can be verified (the shader posed some
issue resulting in horror movies ;).
Brush models looked a little too tricky due to the very different style
of command queue, so that's left for now, but alias, iqm and sprite
entities are now labeled. The labels are made up of the lower 5 hex
digits of the entity address, the position, and colored by the
normalized position vector. Not sure that's the best choice as it does
mean the color changes as the entity moves, and can be quite subtle
between nearby entities, but it still helps identify the entities in the
command buffer.
And, as I suspected, I've got multiple draw calls for the one ogre. Now
to find out why.
The bones aren't animated yet (and I realized I made the mistake of
thinking the bone buffer was per-model when it's really per-instance (I
think this mistake is in the rest of QF, too)), skin rendering is a
mess, need to default vertex attributes that aren't in the model...
Still, it's quite satisfying seeing Mr Fixit on screen again :)
I wound up moving the pipeline spec in with the rest of the pipelines as
the system isn't really ready for separating them.
The plists can now be accessed by name and the forward render pass
config is available (but not used, or tested beyond syntax). I was going
to have the IQM pipeline spec separate but ran into limitations in the
system (which needs a lot of polish, really).
That @inherit is pretty useful :) This makes it much easier to see how
different pipelines differ or how they are the similar. It also makes it
much clearer which sub-pass they're for.
I was wondering why scaled-down quake-guy was dimmer than full-size
quake-guy. And the per-fragment normalization gives the illusion of
smoothness if you don't look at his legs (and even then...).
I'm not sure what's up with the weird lighting that results from dynamic
lights being directional (sunlight works nicely in marcher, but it has a
unit vector for position).
The parsing of light data from maps is now in the client library, and
basic light management is in scene. Putting the light loading code into
the Vulkan renderer was a mistake I've wanted to correct for a while.
The client code still needs a bit of cleanup, but the basics are working
nicely.
This replaces *_NewMap with *_NewScene and adds SCR_NewScene to handle
loading a new map (for quake) in the renderer, and will eventually be
how any new scene is loaded.
This leaves only the one conditional in the shader code, that being the
distance check. It doesn't seem to make any noticeable difference to
performance, but other than explosion sprites being blue, lighting
quality seems to have improved. However, I really need to get shadows
working: marcher is just silly-bright without them, and light levels
changing as I move around is a bit disconcerting (but reasonable as
those lights' leaf nodes go in and out of visibility).
Id Software had pretty much nothing to do with the vulkan renderer (they
still get credit for code that's heavily based on the original quake
code, of course).
It's not used yet, and thus may have some incorrect settings, but I
decided that I will probably want it at some stage for qwaq. It's
essentially was was in the original spec, but updated for some of the
niceties added to parsing since I removed it back then. It's also in its
own file.
Despite the base IQM specification not supporting blend-shapes, I think
IQM will become the basis for QF's generic model representation (at
least for the more advanced renderers). After my experience with .mu
models (KSP) and unity mesh objects (both normal and skinned), and
reviewing the IQM spec, it looks like with the addition of support for
blend-shapes, IQM is actually pretty good.
This is just the preliminary work to get standard IQM models loading in
vulkan (seems to work, along with unloading), and they very basics into
the renderer (most likely not working: not tested yet). The rest of the
renderer seems to be unaffected, though, which is good.
The resource subsystem creates buffers, images, buffer views and image
views in a single batch operation, using a single memory object to back
all the buffers and images. I had been doing this by hand for a while,
but got tired of jumping through all those vulkan hoops. While it's
still a little tedious to set up the arrays for QFV_CreateResource (and
they need to be kept around for QFV_DestroyResource), it really eases
calculation of memory object size and sub-resource offsets. And
destroying all the objects is just one call to QFV_DestroyResource.
Vulkan doesn't appreciate the empty buffers that result from the model
not having any textures or surfaces that can be rendered (rightfully so,
for such a bare-metal api).
I doubt the calls were ever actually made in a normal map due to the
node actually being a node when breaking out of the loop, but when I
experimented with an empty world model (no nodes, one infinite empty
leaf) I found that visit_leaf was getting called twice instead of once.
Since it is updated every frame, it needs to be as fast as possible for
the cpu code. This seems to make a difference of about 10us (~130 ->
~120) when testing in marcher. Not a huge change, but the timing
calculation was wrapped around the entire base world pass, so there was
a fair bit of overhead from bsp traversal etc.
It makes a significant difference to level load times (approximately
halves them for demo1 and demo2). Nicely, it turns out I had implemented
the rest of the staging buffer code (in particular, flushing) correctly
in that it seems there's no corruption any of the data.
The misinterpretations were due to either the cvar not being accessed
directly by the engine, but via only the callback, or the cvars were
accesssed only by progs (in which case, they should be float). The
remainder are a potential enum (hud gravity) and a "too hard basket"
(rcon password: need to figure out how I want to handle secret strings).
This is an extremely extensive patch as it hits every cvar, and every
usage of the cvars. Cvars no longer store the value they control,
instead, they use a cexpr value object to reference the value and
specify the value's type (currently, a null type is used for strings).
Non-string cvars are passed through cexpr, allowing expressions in the
cvars' settings. Also, cvars have returned to an enhanced version of the
original (id quake) registration scheme.
As a minor benefit, relevant code having direct access to the
cvar-controlled variables is probably a slight optimization as it
removed a pointer dereference, and the variables can be located for data
locality.
The static cvar descriptors are made private as an additional safety
layer, though there's nothing stopping external modification via
Cvar_FindVar (which is needed for adding listeners).
While not used yet (partly due to working out the design), cvars can
have a validation function.
Registering a cvar allows a primary listener (and its data) to be
specified: it will always be called first when the cvar is modified. The
combination of proper listeners and direct access to the controlled
variable greatly simplifies the more complex cvar interactions as much
less null checking is required, and there's no need for one cvar's
callback to call another's.
nq-x11 is known to work at least well enough for the demos. More testing
will come.
This allows for easy (and safe) printing of cexpr values where the type
supports it. Types that don't support printing would be due to being too
complex or possibly write-only (eg, password strings, when strings are
supported directly).
Surprisingly, only two, but they were caught by the different value
fields being used, thus the cvar was checked in multiple places. I
imagine that's not really all that common, so there may be some
inconsistencies between default value and use.
This allows a single render pass description to be used for both
on-screen and off-screen targets. While Vulkan does allow a VkRenderPass
to be used with any compatible frame buffer, and vkparse caches a
VkRenderPass created from the same description, this allows the same
description to be used for a compatible off-screen target without any
dependence on the swapchain. However, there is a problem in the caching
when it comes to targeting outputs with different formats.
As I had suspected, it's due to a synchronization problem between the
scrap and drawing. There's actually a double problem in that data
uploaded to the scrap isn't flushed until the first frame is rendered
causing a quick init-shutdown sequence to take at least five seconds due
to the staging buffer waiting (and timing out) on a stuck fence.
Rendering just one frame "fixes" the problem (draw was one of the
earliest subsystems to get going in vulkan).
Surprisingly, only two, but they were caught by the different value
fields being used, thus the cvar was checked in multiple places. I
imagine that's not really all that common, so there may be some
inconsistencies between default value and use.
Since it is updated every frame, it needs to be as fast as possible for
the cpu code. This seems to make a difference of about 10us (~130 ->
~120) when testing in marcher. Not a huge change, but the timing
calculation was wrapped around the entire base world pass, so there was
a fair bit of overhead from bsp traversal etc.
Really, this won't make all that much difference because alias models
with more than one skin are quite rare, and those with animated skin
groups are even rarer. However, for those models that do have more than
one skin, it will allow for reduced allocation overheads, and when
supported (glsl, vulkan, maybe gl), loading all the skins into an array
texture (since all skins are the same size, though external skins may
vary), but that's not implemented yet, this just wraps the old one skin
at a time code.
While looking at the deferred attachment images with using a template in
mind, I noticed that the opaque attachment was using 8-bit color. The
problem is, it's meant to be HDRI with the compose pass crunching it
down to LDRI. Switching to 16-bit float does seem to have made a subtle
difference (hey, it's still quake data, not much HDRI in there).
That certainly makes it nicer to work with large sets, and shows one way
to be careful with allocated resources: don't allocate them in the
inherited data and use a template that needs a few things filled in to
be valid. Also, it seems that overriding values in sub-structures "just
works" :)
