This makes a possible improvement to e1m3, only barely affects ad_tears,
but makes about 30% difference to gmsp3v2 (21fps to 27, and from 3300
leafs to 2700).
The cascade_shadow and cube_shadow names are no longer relevant thanks
to the staging images, and the output field for render passes is
optional in general and irrelevant for shadow maps.
The rendering of the shadow maps now takes the culling information into
account resulting in a drastic reduction of work. There's still more
work to be done, but demo1 peaks at over 1000fps at 640x480, gmsp3v2 now
gets 14fps (1920x1080) near the front gate (used to be 3, then 6),
ad_tears is up to 3fps, but marcher is still unhappy, but it has
infinite radius lights, so needs more culling work (clipped light
volumes will help, I think). Also, culling lights for which nothing has
moved within their volumes will help somewhat (though not as much for
most id maps, I suspect).
Using the translucency pass made it easy to have depth-tested
translucent "solid" light volumes instead of always visible lines (which
are still an option as that's useful too). Most importantly, being able
to see the surfaces helped no end in figuring out that my hulls were
created with counter-clockwise windings instead of quake's usual
clockwise windings and thus my hulls were being rendered inside-out in
the occlusion pass.
The results of the occlusion queries give the lights that don't have a
visible hull, but unfortunately that includes any lights which the
camera is inside, but simple distance checks sort that out (with a
fudge-factor for the icosahedron vertices (1.583 (3(2+p)/(2+3p), p is
golden ratio)).
My efforts (especially the collect zone (what was I thinking)) got
tracy's knickers in a twist resulting in vanishing zones in the server.
It looks like there are some synchronisation issues between cpu and gpu,
but I'm not *too* worried about it at this stage.
The info isn't used yet, but this shows that vulkan's occlusion queries
are at least somewhat useful. However, the technique isn't perfect:
infinite radius lights (1/r and 1/r^2) are difficult to cull, and all
lights can poke through thin enough walls, and then lights containing
the camera get culled incorrectly (will need a separate test). Still, it
looks like it will help once everything is tied together.
And make it callable directly (needed to be able to submit the command
buffer separately from the main commands (though this does mess with
tracy a little).
They weren't rendering properly at all due to the matrix updates getting
overwritten by the light data (I'd forgotten to advance the packet data
pointer).
This doesn't make much of a difference on the GPU, but it drastically
cuts down CPU usage, especially for ad_tears: shadow map drawing is down
from 16.3ms to 3.7ms thanks to no having to run the alias model queues
as often.
Batching shadow map rendering needs be able to reference matrices for
multiple lights in a single batch, but the only input is the view index,
so use that to look up the matrix index rather than using it to index
the matrices directly (modulo the base index that's still there).
Actually, only 29 are used because nvidia's drivers segfault when there
are more than 29 views (regardless of the exact bit pattern in the view
mask). This will allow rendering shadow maps in large batches, which
should make for better GPU utilization.
Even that's getting pretty big, but with the quanta at 128, that's a
maximum of 8 different image sizes (which is nice for my planned
"staging image" idea).
Interestingly, this caused a reduction in memory use for some maps (but
did increase marcher's again, but not as much as the bogus rounding
did). The idea was to use sparse bindings to remap shadow map layers,
but it turns out sparse bindings are insanely slow (beyond unusable).
However, the reduction in the number of shadow map images seems to be
worth it.
Since switching to the 1.2 api as a requirement, might as well use the
relevant structs instead of extension struct (for multiview). Came up
when double-checking the max views property due to running into what
appears to be an nvidia bug where > 29 views (any bit pattern) cause a
segfault when creating the pipeline.
I had missed that upping max lights to 2048 meant that up to 12288
matrices are needed for all the possible lights. This made it so the
light type could not be encoded in id_data, but the shaders never used
it anyway. This leaves one bit free.
I'd added some developer output to see how the layers were distributed
between images and found the image widths to be... odd. It turns out I
was double-adding the shadow_quanta. Oops. Results in ~164MB less memory
used by marcher (for 32 pixel quanta).
This allows "large" updates to be done in a single staging buffer packet
instead of one packet per quad (or slice). Currently, they're batched
into groups of 64 (not really enough for conchars, but that's only at
init-time, so not all that bad). Nicely, this seems to simplify the
staging code.
Fixes#65.
When looking at a struct and seeing "count" and "size", I had to hunt to
see what "size" really meant. Cherno is very much right about size vs
count being bytes vs number of objects.
load_conchars and load_crosshairs were using create_quad directly (due
to make_static_quad having the wrong parameters), but this spread the
handling of which buffer and index where used through the code. Thus fix
make_static_quad to take the x, y offsets (like make_dyn_quad) and then
use it in load_conchars and load_crosshairs.
While QFV_PacketScatterBuffer works on only one destination buffer, it
turns out it's still useful for scattering to multiple buffers, just
with multiple calls. This makes it pretty easy to combine multiple
buffer updates into a single staging buffer packet, resulting in
reducing lighting's packet use from up to 7 to just one, drastically
reducing the pressure on the stating buffer packet pool, and thus
reducing the chances of QFV_PacketAcquire stalling.
This relies on my fork of tracy: https://github.com/taniwha/tracy
on the wip-c-vulkan branch. Everything is still rather flaky though.
This necessitated the jump to vulkan 1.2 as a requirement.
This gets the dynamic data closer to the gpu, so should make a
difference when there's a lot going on. However, for simple tests, it
made no difference.
I'm still not happy with it being a compile time constant, but this
takes care of the interlock between frames in flight... for now: it's
fragile and really needs the excessive small-packet use in draw and
lighting to be cleaned up.
After discussion with Darian, I've decided to go with one big staging
buffer (with lots of packets) shared between FiF as the large size will,
in the end, be more flexible.
Tracy is a frame profiler: https://github.com/wolfpld/tracy
This uses Tracy's C API to instrument the code (already added in several
places). It turns out there is something very weird with the fence
behavior between the staging buffers and render commands as the
inter-frame delay occurs in a very strangle place (in the draw code's
packet acquisition rather than the fence waiter that's there for that
purpose). I suspect some tangled dependencies.
I'm not sure what's up, but arm gcc thinks the array isn't properly
initialized even though x86_64 gcc does. Maybe something with padding.
At least c23 makes it easy to 0-initialize VLAs.
I'm actually surprised anything worked, though I guess it was just the
one entry getting corrupted (and not 32, but I figured allocate slots
for all of the dynamic lights just in case). Or none, really, since
larger scenes (ie, those with multiple lights that fit in the same image
size) would result in not all the maps getting used and thus one spare
for dynamic lights.
This seems excessive, but gmsp3v2 map has 1399 lights. Worse, it has a
lot of different light sizes that go up by small increments (generally
around 10) resulting in 33 shadow map images (1 too many). Quantizing
the sizes to 32 drops this nicely to 20, and reduces memory consumption
slightly too (image buffer overhead, I guess).
