On 1.2.1 the light target is scaled to the output by an upscaler of the game's own (twelve texelFetch taps,
fragment shader e6f7ae35b8479ef3), not by the FXAA draw vk_dlaa_skip_draw knows. It addresses its source in the
scene's pixels, so DLSS's output-sized picture swapped in for the light target (tex_info) showed only its top-left
part, two thirds at DLSS quality, stretched over the screen: the "floating camera" of #1106. #660 (169d057) made the
swap reach that draw: before it, DLSS ran on another target and its picture never reached the screen on 1.2.1.
DLSS already writes its picture into the scaling pass's target, so on 1.2.1 a draw in a pass larger than the light
target, without depth, that samples the light target is now skipped (vk_exec_draw_samples). The HUD drawn later into
the same target samples other textures and stays. 1.4.3 keeps the swap, where a tonemap can be the scaling draw.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
evaluate's preset_dirty loop stopped at slot 3 of the five D.sizes entries, so a preset change left the features in
slots 3 and 4 created with the old model: they kept running until the slot was evicted. The loop now covers the whole
array through DLSS_SIZES instead of a literal.
Taken from pull request #1057 (t04dJ14n9) alone; the PR's dlss_prewarm commits are not in this change.
(cherry picked from commit a76bdf5e08)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ported onto vk_dlaa_plan (#794): the scene end in the inventory also records its submission and op (end_fence,
end_pos) so chunked replays see it at the same point. DLSS-only: everything sits behind dl_active().
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
DLSS kept the replay on one thread: its decisions follow the passes in replay order, its jitter and camera are noted
draw by draw, and its evaluation reads the per-object motion list the scene's draws build. Now the scan takes DLSS's
decisions on its own state (vk_dlaa_plan, split out of vk_dlaa_pass) and notes the scene camera in replay order
(vk_dlaa_scan_draw), and the copy of the lit scene and the evaluation are records of the chunk that holds the op
(vk_dlaa_exec). A chunk's draws are jittered while they come before the op the scan ended the scene at. The
evaluation waits until the chunks before it are recorded, and the chunks after it wait for it (a hold), since they
sample its output. Chunks note their draws for the per-object motion one at a time (mo_lock in vk_exec_draw.c); the
uniform overrides of a motion draw and shader_for's last-frame variant flag are per thread, as the other chunks' draws
go on meanwhile; the FXAA shader table takes concurrent lookups. fxaa = 0 runs in chunks too.
With dlaa = 1 at 4K (DLSS Quality) on 1.2.1 and an RTX 5070, replay_workers = 4 takes the frame rate from 70 to 78
fps where the CPU is the limit, with no one-frame glitch in 241 consecutive frames.
(cherry picked from commit 04f8bb1b57)
replay_workers = N (0, the default, keeps the sequential replay) records a submission of enough draws in chunks that start between passes, on N threads and in twice as many chunks, and submits them in order. In the town at 4K on an RTX 5070, 4 threads take the replay from 8.2 to 4.5 ms a frame and the frame rate from 71 to 79.
A sequential scan on the submitting thread applies only what is queue state (fixed-function and bind ops, render targets, render enable and conditions, samples-passed counting) and keeps a copy of it where each chunk starts, with how far its lists got: inline uniform updates, RENDER_COND results, contact shadow marches and the ops that drop uniform copies. A chunk takes those from the lists instead of from the chunks before it, starts with every image in GENERAL and ends with a full barrier, so its recording depends only on the copy and the lists.
Each thread records with a context of its own: the replay's memos, descriptor state, occlusion query segments and per-command-buffer state moved from file statics into VkExec and VkExecDraw, and the tables the contexts share (stand-ins, depth twins, views, mirrors, timestamps, completion callbacks, crumbs) take locks or atomics. The scan makes render targets live in replay order and notes the op (vk_texture_revive), since a chunk can sample a storage before the chunk that renders it got there or after the scan went past a later target there: a chunk's texture slots resolve as of its own op, and the chunks after a texture with no entry yet wait for the chunk that makes it. Uniform copies are dropped at the op that writes them (vk_usnap_drop_at).
Contact shadows plan their march in replay order (vk_contact_plan) and the chunk that holds its op records it; the HDR source keeps the copy that comes last in replay order. DLSS and fxaa = 0 act on the frame draw by draw and keep the replay sequential. TOTK_REPLAY_STATE_CHECK=1 compares the state after the last chunk with the scan's.
A draw or dispatch whose shaders write storage buffers drops the uniform copies they cover from its own op on, as the other GPU writes do (scan_writes). The scan still translates no shader: it reads the shaders the recording contexts translated from a table shared with them (vk_exec_draw_writes), and an op whose shader no context translated yet holds the chunks after it until it is recorded. Before, the chunk that recorded such a write dropped the copies for every position at that moment, so a later chunk recorded first bound the copy taken at the submit call: on an RX 7800 XT on RADV the grass near the camera came out dark or unshaded in some runs with 4 threads.
TOTK_TRACE=perf counts the chunks held per submission by reason and logs when each chunk of the last submission began. The settings menu has the thread count under Advanced, Scheduling, with the count this run uses under it.
The recorder thread begins, ends and submits the submission's open part (vk_rec.c rec_end_cb), so a chunk's mirror prologue stays off it: vk_sync_part waits until the recorder is idle, as the part pool allows one thread at a time, then records, ends and submits the prologue itself. Through the recorder the prologue could reach the queue before it was ended, and the GPU lost the device at the title screen.
(cherry picked from commit b94e2e9642)
tools/check_platform.py failed on vk_dlaa.c: it read the NGX core's folder from the registry, loaded it with LoadLibraryA and GetProcAddress and named the executable's folder with GetModuleFileNameW. platform.h gets plat_ngx_core_path (the registry folder plus _nvngx.dll on Windows, libnvidia-ngx.so.1 on Linux), plat_exe_dir_wide (the folder as NGX's wchar_t string; the UTF-8 to UTF-32 conversion moves from vk_dlaa.c to linux.c) and plat_lib_error, and the core loads through plat_lib_open and plat_lib_sym on both hosts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SPIR-V leaves a register read before any write undefined and Mesa folds undefined values freely. The unconditional patch 0014 guarded against that and was taken out in #701 because the whole picture shook with DLSS on NVIDIA. Patch 0016 brings the zero start back behind a zero_locals job flag, which vk_caps_driver turns on for RADV, ANV and NVK. TOTK_ZERO_LOCALS overrides it, and gfx_shader_set_host mixes it into the translator identity so the variant has its own cache keys.
FXAA_FS gains the flag-on hash of the game's FXAA shader, so vk_dlaa_skip_draw knows it on Mesa too. A game run with TOTK_ZERO_LOCALS=1 on an RX 9070 produced that hash; a TOTK_FXAA=0 run logged no skip with either hash, so the switch itself is not confirmed.
All 29736 stored 1.4.3 inputs translate with the flag off and on, 0 spirv-val failures. FXAA hashes to 5c43ae18948d52cb with the flag off and aad42238801d31d8 with it on. Flag-on output matches the old translator byte for byte on a sample of 312 shaders. Not run on a Mesa driver.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The inventory's preview light target is drawn without depth before the bloom chain; a pass like that with nothing armed ends the scene once the bloom follows, so DLSS no longer runs on the finished menu picture. Arming on that target stays allowed, because at some camera angles the field's effects composite comes before the first pass with the scene depth. The G-buffer width is the widest of the frame, a small pass ends the scene only into a scene colour target, and no pair narrower than the widest G-buffer arms: with the scope open the half-size effects pass armed and DLSS scaled the effects into the light target.
Merged on top of #566 and #594, by hand in vk_dlaa.c:
- One dlss_preset key: default keeps #594's model per mode (K for DLAA, Quality and Balanced, M for Performance, L
for Ultra Performance); k, j or any other letter a..m takes that model in every mode, live, recreating the features
as #534 does. The menu texts say so, with pt-BR and zh-Hans added.
- The mask pass stays on #594's camera motion (the identity and the CPU-made reprojection in dlaa_mv.comp's two
blocks) and adds #534's outputs; it runs before the per-object pass. Where the reactive mask test
(TOTK_DLSS_REACTIVE) ran, DLSS gets that mask instead of the characters'.
- #566 dropped the scaler_skipped/scaler_missed report, so #534's reset of it goes too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Merged on top of #566. The camera motion is #594's: its blocks in camb also feed the per-object pass, and
dlaa_mv.comp stays as on main, fed the identity and the reprojection made on the CPU in double. #566's
vk_usnap_host goes, vk_usnap_bytes serves the same. #566's scaling-draw swap, camera rune guard and output-size
scaling pass stay in vk_dlaa.c. Translator patch 0012 becomes 0015 after dev's 0012-0014.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Solid scene draws whose position reads an object's own uniform blocks (bones c5, place c7, c6) are drawn a second time
with a vertex shader variant (translator patch 0012) that computes last frame's position from last frame's blocks; the
game writes those into two copies that take turns, and the uniform snapshot keeps both (vk_mvprev.c). The draws are
listed during the replay and recorded through the recorder thread with repeated state skipped. Blocks whose last-frame
copies measured wrong (c9 and c11 of walls and cliffs) keep this frame's; positions without an object block (terrain,
grass) keep the camera's motion, which measured as good.
The camera motion is reprojected in double precision on the CPU: the float inverse made noise on far surfaces and a
still camera that DLSS followed into a wobble. DLSS gets Ultra Performance (a third of the output, also in the F1 menu)
and NVIDIA's recommended preset per mode (K, M for Performance, L for Ultra Performance; TOTK_DLSS_PRESET overrides).
Test switches: TOTK_MV_CHECK measures the motion vectors and DLSS's output against the picture, TOTK_DLAA_TIMES times
DLSS and the motion passes on the GPU. A reactive mask pass for transparents is included but off (TOTK_DLSS_REACTIVE):
its mask covers sky and fog too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
NVIDIA's DLSS guide asks for 8 x (output / scene)^2 jitter phases: 18 in Quality, 24 in Balanced, 32 in
Performance. The sequence was 16 phases in every mode; it keeps 16 as the least (DLAA).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1.4.3 sets its render size once, at start (main+0x7000e67028 calls
main+0x585bd0(renderer, force)). The function-entry hook catches that call's
renderer, and a TOTK_DLSS change writes the new scene size to the patch page and
has the next nvnQueuePresentTexture call the setup again with force set, on the
game's thread between frames, with the guest registers kept. The heap is sized
for the largest scene a mode asks for. The DLSS texture bias is part of the
sampler cache key and follows the mode and the dlaa switch. The game's scaling
pass is told from the in-place FXAA by its missing depth attachment, not by the
sizes, which change a frame ahead of the drawing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The camera view draws a half-size RGBA16F target with a half-size D32F depth after the lighting (852x480 under a
1704x960 scene). It matched the light target pattern, so the DLAA point re-armed on it and DLSS scaled it into the
scene's own light target, which left every object black. An armed pair is no longer replaced by a smaller one, and
the scaling pass DLSS writes into must have the output size.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With a DLSS mode the game's own scaling draw covered DLSS's picture with the jittered scene. At a 1280x720 scene and
1920x1080 output the scaling is done by the game's tonemap, not by the FXAA shader vk_dlaa_skip_draw looks for, so no
hash list could catch it. From DLSS's evaluation to the frame's end, passes larger than the light target now sample
DLSS's output in its place (vk_dlaa_swap_img, applied in vk_exec_draw.c tex_info), whichever shader scales.
The motion-vector shader inverted the view-projection in 32-bit floats. With a still camera (identical matrices each
frame) the rounding still gave a few hundredths of a pixel of motion that changed with every jitter step; DLSS
followed it and the whole picture wobbled. mv_pass now makes last frame's view-projection times the inverse of this
frame's in double precision from the uniform snapshot (vk_usnap_host), and the shader only applies it.
Measured on 1.4.3 indoors, still camera, mean per-pixel change between consecutive frames on static areas:
DLSS quality 0.18 -> 0.06, DLAA 0.14 -> 0.05. A camera pan at 10 px per frame shows no ghosting.
The motion vectors follow the camera only, so a character, animal or held item that moves on its own drags a faint
copy of itself through DLSS's history. TotK's G-buffer target 0 carries an object class in green: above 64/255 for
characters, creatures and held items (Link 253, others 125, held items 160), 0 for terrain and 32 for foliage, checked
in the field, in Goron City and on horseback. The motion-vector pass now also writes DLSS's bias mask
(DLSS.Input.Bias.Current.Color.Mask) where such an object is now or was the frame before, read through the motion
vector from the previous frame's class mask, so the history that holds the moving object is weighed down there.
dlss_char_mask (0, 35, 65 or 100 percent, default 65; F1 > Screen > "DLSS character trails", live) sets the strength.
A 640x360 pass of the G-buffer's shape, seen while loading, no longer replaces the scene's class buffer in a frame.
The status line also prints after 300 frames, with the frames that had the mask.
dlss_preset (default, k or j; F1 > Screen > "DLSS model", live) sets the render preset hints when a DLSS feature is
created. default leaves the choice to NGX (preset K for DLAA, Quality and Balanced in the 310 releases); j is the
transformer model NVIDIA describes as slightly less ghosting at the cost of more flicker. A change recreates the
features at the next DLAA run. FrameTimeDeltaInMsec now carries the time since the previous evaluation instead of a
fixed 16.67 ms. The menu and totk.ini.example name the DLSS model file of the platform in use.
vk_dlaa.c reached the NGX core only through Windows: the registry, LoadLibraryA and GetModuleFileNameW. On Linux the
NVIDIA driver installs the core as libnvidia-ngx.so.1 with the same Vulkan entry points; it is opened with plat_lib_open.
The parameter object's vtable follows the Itanium C++ ABI there, each overload group in declaration order rather than
MSVC's reverse order, so the slot enum has a Linux variant; vtable_ok confirms it at start. NGX gets the executable's
folder as a 32-bit wchar_t string, converted from UTF-8 without the C locale, and loads libnvidia-ngx-dlss.so.<version>
from there. The Windows path is unchanged.
resolution.c sized the scene for a DLSS mode only on Windows, where it looks for nvngx_dlss.dll; on Linux it now looks
for libnvidia-ngx-dlss.so.<version> next to totk, so quality, balanced and performance work there too. The no-op
stubs vk_dlaa.c had for other platforms go, as Linux now builds the real code.
The DLSS modes already apply only with dlaa on (3de0839), so the DLL check no longer looks at dlaa, and off Windows,
where vk_dlaa.c is no-op stubs, a DLSS mode no longer shrinks the scene. fxaa = 0 names the FXAA shader it skipped,
or reports a minute of frames without it, so a translator change that moves FXAA_FS shows on any GPU.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With a DLSS mode (a scene smaller than the output) DLSS writes the scaled picture into the scaling pass's target and
vk_dlaa_skip_draw skips the game's own scaling draw, known by its fragment shader's SPIR-V hash. Translator patch
0011 (the shadow filter spread) changed that hash, so the draw ran again and covered DLSS's picture with the scaled
jittered scene: the whole picture shook by a fraction of a pixel each frame, most visible on lit flat surfaces in
shrines. The FXAA switch (fxaa = 0) stopped working for the same reason.
Add the hash of the current translator, and log once when DLSS scaled a frame whose scaling draw was not found, so
the next translator change that moves the hash shows in the log.
Measured on 1.4.3, Ukouh Shrine, output 3200x1800, dlss = quality, mean absolute change between consecutive frames
with a still camera: 1.40 before, 0.07 after (0.054 with dlaa = 0). Outdoors on stone: 4.29 before, 1.08 after.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A DLSS mode (quality, balanced, performance) shrank the scene to 67, 58 or 50 % of the output and sharpened the
material textures' mip bias to match, even when nvngx_dlss.dll was missing and NGX reported DLSS as not available.
The game's plain scaling then stretched the smaller scene, and the negative bias made textures shimmer in motion.
With dlaa = 1 the scene size now waits on nvngx_dlss.dll next to totk.exe (checked before the device exists, as
the size is decided then) and follows render_scale without it. The mip bias applies only when NGX came up with
DLSS. dlaa = 0 keeps the documented size with the game's scaling, without the bias.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
It includes windows.h and loads NGX as a Windows DLL, and build_game.py leaves only *_win.* files out of the Linux
build, which then stopped at it. Elsewhere DLAA, the DLSS scaling and the FXAA switch now do nothing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
vk_dlaa_wanted answered 1 on every machine, so each start on an NVIDIA driver loaded the NGX core and enabled its
device extensions with DLAA off; dlaa now applies after a restart. vk_dlaa_skip_draw hashed the fragment shader of
every draw through an unlocked 8192-entry table, which 1.4.3's shader variants outgrow; with fxaa on and DLSS not
scaling there is nothing to skip, so it returns before the lookup. Also gives the DLSS rows their Brazilian
Portuguese and Simplified Chinese text.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TOTK_DLSS (dlaa|quality|balanced|performance, read at start) sets the scene to
67, 58 or 50 % of the output. DLSS then runs at the begin of the game's scaling
pass (the FXAA shader drawing the light target into the output-sized target,
before the HUD) and its result replaces that draw; the HUD is drawn afterwards
and never jittered or passed through DLSS. The jitter now ends where the scene
ends (the bloom), separately from where DLSS runs. Scaled modes add a mip bias
of log2(scale) to repeating, mipmapped material samplers.
With a scene smaller than the output, fxaa = 0 no longer skips the game's
scaling draw (it left the 3D picture out).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
DLAA runs DLSS at the scene's own size on the R11G11B10F light target, right before the bloom chain's first pass,
so bloom, post and HUD read the anti-aliased picture. NGX is loaded from the driver's NGX core at run time; no NVIDIA
SDK code is compiled in, and nvngx_dlss.dll goes next to totk.exe. The scene's draws get a Halton(2,3) viewport
jitter that steps after each DLAA run; camera motion vectors come from the scene depth reprojected through this and
last frame's view-projection of the scene camera block (c4), read from the uniform snapshot. One DLSS feature per
scene size keeps dynamic resolution from stalling. dlaa and fxaa switch live from the settings menu; fxaa skips the
draw of the game's FXAA pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>