Repository navigation
Releases
Kizuna v1.21
fritchii_ui.cpp: close IconTinted (merge resolution dropped its brace)
Commit 1046197, workflow run 36.
Extract the zip and run lichii.exe, or use Help > Check for Updates in Kizuna.
Kizuna v1.20
Port two stale-memory fixes for area changes (Lichii zortos293#1458, zortos293#1459)
After Castle Town -> Hyrule Castle, EDEN_NX2_TABLE_POINTER_TRACE showed a
compute dispatch reading its table pointers (storage slots 24-27, through
constant buffer 4) from the previous area's holders, whose memory by then
held other data, and the ground lost its textures until the next area.
- zortos293#1458: releasing a span frees every device range inside it. Ranges made
for subranges (211 maps at offsets) stayed on the old pages after their
memory was destroyed. - zortos293#1459: a deleted buffer is cleared from every channel's bindings, not
only the current one's.
Commit c24b81e, workflow run 31.
Extract the zip and run lichii.exe, or use Help > Check for Updates in Kizuna.
Kizuna v1.19
Allow device memory overallocation on AMD drivers (port of Lichii zortos293#1335)
AMD's driver fails device-local allocations once the heap is full unless
the app allows overallocation (VK_AMD_memory_overallocation_behavior),
which adds system memory as a backup heap. With Discord in a call on an
RX 6600, Windows cut the emulator's budget from 7379 to ~4260 MiB, image
allocations failed with VK_ERROR_OUT_OF_DEVICE_MEMORY even on the host
retry, and the session ended. Lichii build 177, which has zortos293#1335, kept
running in the same call.
Commit 6f8c90f, workflow run 30.
Extract the zip and run lichii.exe, or use Help > Check for Updates in Kizuna.
Kizuna v1.18
Walk only the dirty tracker words when uploading buffers (clean_words on)
A sample profile of the split's back thread in Hyrule Field showed 15% of it
walking clean words of the memory tracker at each storage buffer bind
(SynchronizeBuffer -> ForEachModifiedRange). The base already carries the
fix, clean_words, switched off with the rest of the Lichii zortos293#161 family.
With EDEN_NX2_PERF_ON=clean_words on a Ryzen 5 5600: bind stage buffers
3.74 -> 1.83 us a draw, compute synchronize 0.93 -> 0.37, the back's idle
time 131 -> 209 ms/s, and no wrong geometry in Kokiri Forest. It is now a
built-in switch; EDEN_NX2_PERF_OFF=clean_words turns it off.
KIZUNA.md also records what was tried and not kept: Lichii zortos293#1486 (not
ported, reverted upstream in zortos293#1651), EXP53, EXP12, EXP17/EXP18.
Commit b5dec56, workflow run 29.
Extract the zip and run lichii.exe, or use Help > Check for Updates in Kizuna.
Kizuna v1.17
Reuse the occlusion query lists flushed at each fence
The last piece of Lichii zortos293#1516 that Kizuna's base did not have: the list of
occlusion queries flushed at a fence was allocated again at each fence with
reports pending (Ocarina makes thousands a second); the lists the fencing
thread is done with are now reused. The rest of zortos293#1516 was already in the
base, so this build is expected to change little.
Commit c343057, workflow run 28.
Extract the zip and run lichii.exe, or use Help > Check for Updates in Kizuna.
Kizuna v1.16
Less video memory for Ocarina's textures, calmer texture collector
The v1.15 log showed where Hyrule Field's 19-25 FPS and its 3-5 s freezes
came from: video memory. Next to the other programs open, the card gave the
emulator about 5 GB and it used a little more, so Windows paged video memory
(the game waited up to 3.5 s for the GPU) and the texture collector deleted
and re-decoded 60-150 textures a second.
- Textures whose size is not a multiple of 4 (most of Ocarina's in handheld
mode: 1365, 683, 341, 171) were kept uncompressed on AMD. They are now
kept compressed (BC) at a slightly larger size, 3 to 4 times smaller in
memory; the two terrain arrays alone go from 1.8 GB to about 0.5 GB.
EDEN_NX2_ASTC_UNEVEN_PAD=0 restores the old behaviour. - The collector no longer deletes textures a third of a second after their
last use unless an allocation really failed; it waits 120 frames. - From Lichii: mesh-emulation buffers are kept until the GPU has used them
(zortos293#1452); running out of Windows memory during a zero-page pass no longer
closes the emulator (zortos293#1451).
Commit f87d571, workflow run 27.
Extract the zip and run lichii.exe, or use Help > Check for Updates in Kizuna.
Kizuna v1.15
Keep compiled shaders between sessions, and fewer freezes in new areas
AMD's driver keeps only about 1 GB of compiled shaders, so each session
compiled again everything earlier sessions had used: the boot shader builds
took 5.5 to 7.5 minutes and grew every session, and places visited before
compiled their shaders again, freezing for each.
Kizuna now keeps the compiled shaders itself (VK_KHR_pipeline_binary), in
shader/kizuna_pipeline_binaries.bin (at most 2 GB). The first session after
updating compiles as before and fills the file; from the next one on,
shaders compiled once load from it in a few milliseconds instead of being
compiled again. A driver update starts the file over. Deleting the file
resets it; EDEN_NX2_PIPELINE_BINARIES=0 turns it off.
Fewer draws wait for new shaders when entering an area:
- Depth, HDR and shadow targets the game re-creates when an area loads now
count as per-frame targets from their first frame when an earlier target
had the same size and format (21 such waits held the game 4.3 s at the
Kakariko church). Text and UI layers still wait. - Reflection probes drawn every few frames (not every frame) no longer
hold the game for their new shaders. - The log lists the targets whose draws still wait.
Commit c04ce5d, workflow run 26.
Extract the zip and run lichii.exe, or use Help > Check for Updates in Kizuna.
Kizuna v1.14
Fix freezes from slow disk reads when shaders are set up in new areas
The v1.13 log showed where the freezes entering new areas went: setting up
each new shader from the prepared files took up to 1.3 s on the GPU thread
while it did nothing but wait for the disk. Those files are 2.3 GB and
were read in small pieces at scattered places, from a nearly full disk.
The prepared files are now read once, from start to end, at low disk
priority right after the game loads. Windows keeps them in its file cache
(free RAM it gives back to programs when they need it), so setting up a
shader reads from memory instead of waiting for the disk.
Also turned back off: the separate shader cache file v1.8 enabled on AMD.
The driver stores almost nothing in it, nothing loaded faster, and it was
written to disk every 30 seconds.
Commit 8c8e9ec, workflow run 25.
Extract the zip and run lichii.exe, or use Help > Check for Updates in Kizuna.
Kizuna v1.13
Log where slow prepared-shader setups spend their time
Entering a new area, setting up shaders from the prepared files sometimes
took the GPU thread hundreds of milliseconds each (4.3 s for 20 in one
case), which shows as a freeze even when no draw waits for a compile. The
log now breaks each slow one down (game state, record checks, creation,
CPU time, page faults, bytes read), so the next fix can target the right
part.
Commit 580b94e, workflow run 24.
Extract the zip and run lichii.exe, or use Help > Check for Updates in Kizuna.
Kizuna v1.12
No background shader compiles on AMD while playing
After a shader preparation, three threads kept compiling all 120,124
shaders in the driver while you played, for hours. On AMD this only hurt:
the driver's cache keeps about 1 GB, so every shader compiled in the
background pushed out one you had actually used (which then had to compile
again, with a freeze), and those threads competed with the shaders the
game needed right then.
On AMD's driver this background step is now off: each shader is compiled
the first time the game needs it, from the prepared files, and the
preparation screen itself is unchanged.
Commit db2bcaf, workflow run 23.
Extract the zip and run lichii.exe, or use Help > Check for Updates in Kizuna.