KytyPS5 U59 int18.1 AMD test 5 — the AMD driver crash fixed at its source, and faster CPU side

This is a test build, mainly for AMD Radeon owners whose game crashed (amdvlk64.dll+0x22240fc) or showed broken
graphics with int18. NVIDIA users are welcome to try it too: Sky Garden runs about 12% faster than with int18. The
first launch rebuilds the shader and program caches once, so the first minutes can stutter.

AMD graphics cards

Majed-allababedi found where the AMD driver crash comes from and sent three fixes (pull requests #25, #26 and #27).
Thank you!

• Driver crash at amdvlk64.dll+0x22240fc (#25): the crash was not in the game's shaders but in one of Kyty's
  own texture-tiling compute shaders. AMD's compiler crashed on the way it picked per-size values. The shader now
  picks them one by one; the result is the same bytes, the driver no longer crashes building it.
• Depth-compare sampling of a missing texture (#26): when a shader samples a depth texture that is not there,
  Kyty now binds a depth-format placeholder, as Vulkan requires. Before, it was a color placeholder, which AMD may
  turn into garbage.
• Pixel-shader inputs (#27): when a pixel shader reads a flat and a smooth value from the same vertex output,
  Kyty moved one of them to a spare slot that the vertex shader never wrote. NVIDIA tends to read zero there, AMD
  reads garbage (a likely cause of the pink and particle glitches). The vertex shader now writes that slot too, and
  pixel-shader inputs that are never read are no longer requested.

AMD test 4's workaround (building some pipelines without optimization) is not in this build: if #25 is the cause,
it is no longer needed. If you still crash, please say so and send your _kyty.txt.

Everyone

• Faster command processing: measured on an RTX 3090 (Ryzen 9 7950X3D), Sky Garden went from 32.8 fps with int18
  to 35.1 with the draw-run changes below (+7%), and the sequencer prefetch adds another 5-8% (launches in A B B A
  order: Sky Garden 34.9 -> 36.5 fps, pirate island 29.3 -> 31.6, clock tower 38.4 -> 40.9).
  • KYTY_CP_SEQ_PREFETCH=1 (new default): while the emulator waits for the GPU at a "wait for idle", it prepares
    the draws that follow. KYTY_CP_SEQ_PREFETCH=0 turns it off.
  • Draw runs through draws that write buffers (KYTY_DRAW_RUN_WRITES), fewer repeated checks of images and pages
    (KYTY_DRAW_RUN_IMAGE_EPOCH, KYTY_DRAW_PREP_CLEAN_SINCE), larger descriptor pools.
• Fixes from chenxiao07's fork: an image write-back no longer overwrites bytes the CPU wrote since (KYTY_IMAGE_WRITEBACK_GUARD),
  a CPU write keeps the GPU's other bytes of an image (KYTY_IMAGE_GPU_KEEP), invalid guest data is skipped instead
  of closing the emulator (KYTY_SKIP_INVALID), and the present thread no longer waits for the renderer's lock.

amd5b: Linux file reads fixed

On Linux, a file read into guest memory that the emulator had write-protected (GPU write tracking) stopped at
the first protected page and reported the short count as success: a worldmap level file read 72,704 of 271,864
bytes (the game's XML assert, crash in the loader thread) and a model file 768 bytes (crash in the model streamer
thread after skipping the first cutscene). The Windows build already finished such reads through a host buffer;
the Linux build now does the same. Windows is unchanged apart from this note.

How to test

Copy your working folder, unzip this build over it and start the game as usual. Play the first levels for a while.

• AMD: does it still crash, and do the graphics look right (no pink areas, no colorful particles)? If anything
  goes wrong, send _kyty.txt from the emulator folder and say which level.
• Astro's Playroom on AMD: do the clouds still break into blocks?
• Everyone: is it faster or slower than int18 for you?

---

KytyPS5 U59 int18 — AMD fixes, upstream sync, rumble for other controllers

int18 is int17 plus the fixes from the three AMD test builds, a sync with upstream Kyty, and fixes for
missing-preset and controller problems. The first launch after the update rebuilds the shader and program caches
once (the shader code generator changed), so the first minutes can stutter.

What's new in int18

AMD graphics cards

• Title-screen hang fixed: on AMD cards the GPU could stop responding ("device lost") at the same title-screen
  draw. A particle mesh shader waits for its own LDS writes inside branches that only some waves take; Kyty turned
  each wait into a workgroup-wide memory barrier, which AMD's driver handles like a workgroup barrier that the other
  waves never reach. The wait now orders only the wave's own memory (subgroup scope). KYTY_LDS_WAITCNT_BARRIER
  = subgroup (default), workgroup (the old behaviour) or 0 (no barrier).
• Water and particle corruption fixed: AMD runs vertex, mesh and pixel shaders with 64 lanes per subgroup, so one
  host subgroup held two of the PS5's 32-lane waves, and shaders read the other wave's lanes and masks (stretched
  water, pink particle clouds, loops that never end). Each wave now stays in its own 32 lanes
  (KYTY_WAVE32_CLUSTERS=auto, the default; 0 turns it off, force applies it everywhere). The log names each
  shader this applies to. NVIDIA is unchanged.
• Large compute shared memory: some Astro Bot compute shaders (GPU skinning) ask for 48 KiB of shared memory;
  AMD's driver allows 32 KiB, so Kyty cut it to 32 KiB and meshes with more than 682 bones could stretch. Such a
  shader now keeps its shared memory in a video memory buffer (from upstream Kyty, with Kyty's own buffer handling).
  KYTY_LDS_DEVICE_BUFFER = auto (default: only above the device limit), 0 (the old cut) or force;
  KYTY_LDS_DEVICE_BUFFER_MB (default 1024) caps the buffer. NVIDIA is unchanged (48 KiB fits).
• Helper lanes no longer spin: in pixel shaders, helper lanes skip the compare-exchange loops of emulated atomics,
  which could loop forever on AMD (KYTY_PS_HELPER_ATOMICS_SKIP=0 restores the old code).
• Less video memory without the preset: KYTY_FUNCTION_ARRAY_SHRINK is now on by default (it was only set by
  u59-preset.json). Vertex and pixel shaders emulate shared memory with a large per-lane array that AMD backs with
  memory for every lane (about 2 MiB per wave); the shrink cuts it to what the shader can reach (at most 352 dwords in
  every Astro Bot 1.018 shader; about 3.9 GiB less video memory on an RTX 3090). =0 turns it off and says so in the
  log. Two RX 6900 XT players whose launcher had not applied the preset crashed inside the driver without it.
• Crash reports name the module and pipeline: a host crash now writes its fault report into the log file
  (_kyty.txt), not only the console, with the module and offset of the crash (for example a driver DLL), each
  call-chain frame's module, and what the thread was doing (a graphics or compute pipeline build, with the guest
  shader hashes). Graphics pipeline builds are always traced in the log.
• Diagnostics for testers (off or unchanged by default): KYTY_DEVICE_FAULT_DIAGNOSTICS=1 prints a device
  report and enables GPU breadcrumbs (KYTY_GPU_BREADCRUMBS), which name the GPU operations in flight after a device
  loss; KYTY_COMPUTE_WAVE64=0, KYTY_PS_PER_VERTEX=0, KYTY_CLIP_GUARD=0, KYTY_VOLATILE_LOADS=1 and
  KYTY_LOOP_GUARD_SHADERS=all (with KYTY_LOOP_GUARD=<n>) switch off or add single shader features to narrow
  down a driver problem.

Everyone

• Rumble for other controllers: Xbox and other non-DualSense pads now rumble from Astro Bot's haptics (the game
  plays its vibration as audio for the DualSense; Kyty turns it into rumble with the per-channel gains). Thanks to
  Linky362 (PR #20).
• Black lighting without the preset fixed: Astro Bot without the lighting patches no longer renders black when
  u59-preset.json is not applied (launcher without the file, or kyty_emulator started directly):
  KYTY_SRT_VARIANT_READS is now on by default (0 turns it off). The launcher warns when u59-preset.json is
  missing or invalid (it looks next to launcher.exe and one folder up), and the emulator prints a warning at startup
  when the preset's settings are absent. Keep using the preset: without it the game runs much slower (about 20
  instead of 32 fps on an RTX 3090 in the Sky Garden). A failed ray tracing pipeline build, which turns lighting off
  until restart, is now shown as a WARNING in the console.
• Upstream Kyty sync: the 151 upstream commits since the last sync were reviewed; 62 were taken and 20 partly
  (new shader instructions and formats, native 64-bit integers and 8/16-bit storage, store-completion waits, GLC/DLC
  memory coherence, D16 and formatted buffer access, kernel memory fixes, the main-thread window title). Upstream
  changes that conflict with Kyty U59's own designs (resource tracking, MAD rounding, readback sizes) were left out.
  No frame rate change was measured (Sky Garden, casino and pirate island A/B within noise). 8/16-bit storage is now
  a required GPU feature, so very old drivers without it cannot start.
• Graphics fixes from upstream: colour exports routed through CB_SHADER_MASK (KYTY_CB_SHADER_MASK_EXPORTS,
  on through the preset), linear render-target mip chains, rendered contents kept across 1D and 2D texture aliases,
  and streamed texture mips uploaded from partly committed sparse reservations (fixes the exit "failed to read mapped
  guest image backing").
• GPU stores to untracked memory: a shader store through a buffer descriptor computed on the GPU, to memory Kyty
  had not cached on the GPU yet, was dropped. Such stores are now logged and copied back right after the dispatch
  (KYTY_BDA_STORE_LOG, default on; =0 restores the old drop).
• Smaller fixes: a depth image alias is found again for a smaller layer prefix (no second copy), trop...

KytyPS5 U59 int18.1 AMD test 5 — the AMD driver crash fixed at its source, and faster CPU side

This is a test build, mainly for AMD Radeon owners whose game crashed (amdvlk64.dll+0x22240fc) or showed broken
graphics with int18. NVIDIA users are welcome to try it too: Sky Garden runs about 12% faster than with int18. The
first launch rebuilds the shader and program caches once, so the first minutes can stutter.

AMD graphics cards

Majed-allababedi found where the AMD driver crash comes from and sent three fixes (pull requests #25, #26 and #27).
Thank you!

• Driver crash at amdvlk64.dll+0x22240fc (#25): the crash was not in the game's shaders but in one of Kyty's
  own texture-tiling compute shaders. AMD's compiler crashed on the way it picked per-size values. The shader now
  picks them one by one; the result is the same bytes, the driver no longer crashes building it.
• Depth-compare sampling of a missing texture (#26): when a shader samples a depth texture that is not there,
  Kyty now binds a depth-format placeholder, as Vulkan requires. Before, it was a color placeholder, which AMD may
  turn into garbage.
• Pixel-shader inputs (#27): when a pixel shader reads a flat and a smooth value from the same vertex output,
  Kyty moved one of them to a spare slot that the vertex shader never wrote. NVIDIA tends to read zero there, AMD
  reads garbage (a likely cause of the pink and particle glitches). The vertex shader now writes that slot too, and
  pixel-shader inputs that are never read are no longer requested.

AMD test 4's workaround (building some pipelines without optimization) is not in this build: if #25 is the cause,
it is no longer needed. If you still crash, please say so and send your _kyty.txt.

Everyone

• Faster command processing: measured on an RTX 3090 (Ryzen 9 7950X3D), Sky Garden went from 32.8 fps with int18
  to 35.1 with the draw-run changes below (+7%), and the sequencer prefetch adds another 5-8% (launches in A B B A
  order: Sky Garden 34.9 -> 36.5 fps, pirate island 29.3 -> 31.6, clock tower 38.4 -> 40.9).
  • KYTY_CP_SEQ_PREFETCH=1 (new default): while the emulator waits for the GPU at a "wait for idle", it prepares
    the draws that follow. KYTY_CP_SEQ_PREFETCH=0 turns it off.
  • Draw runs through draws that write buffers (KYTY_DRAW_RUN_WRITES), fewer repeated checks of images and pages
    (KYTY_DRAW_RUN_IMAGE_EPOCH, KYTY_DRAW_PREP_CLEAN_SINCE), larger descriptor pools.
• Fixes from chenxiao07's fork: an image write-back no longer overwrites bytes the CPU wrote since (KYTY_IMAGE_WRITEBACK_GUARD),
  a CPU write keeps the GPU's other bytes of an image (KYTY_IMAGE_GPU_KEEP), invalid guest data is skipped instead
  of closing the emulator (KYTY_SKIP_INVALID), and the present thread no longer waits for the renderer's lock.

How to test

Copy your working folder, unzip this build over it and start the game as usual. Play the first levels for a while.

• AMD: does it still crash, and do the graphics look right (no pink areas, no colorful particles)? If anything
  goes wrong, send _kyty.txt from the emulator folder and say which level.
• Astro's Playroom on AMD: do the clouds still break into blocks?
• Everyone: is it faster or slower than int18 for you?

---

KytyPS5 U59 int18 — AMD fixes, upstream sync, rumble for other controllers

int18 is int17 plus the fixes from the three AMD test builds, a sync with upstream Kyty, and fixes for
missing-preset and controller problems. The first launch after the update rebuilds the shader and program caches
once (the shader code generator changed), so the first minutes can stutter.

What's new in int18

AMD graphics cards

• Title-screen hang fixed: on AMD cards the GPU could stop responding ("device lost") at the same title-screen
  draw. A particle mesh shader waits for its own LDS writes inside branches that only some waves take; Kyty turned
  each wait into a workgroup-wide memory barrier, which AMD's driver handles like a workgroup barrier that the other
  waves never reach. The wait now orders only the wave's own memory (subgroup scope). KYTY_LDS_WAITCNT_BARRIER
  = subgroup (default), workgroup (the old behaviour) or 0 (no barrier).
• Water and particle corruption fixed: AMD runs vertex, mesh and pixel shaders with 64 lanes per subgroup, so one
  host subgroup held two of the PS5's 32-lane waves, and shaders read the other wave's lanes and masks (stretched
  water, pink particle clouds, loops that never end). Each wave now stays in its own 32 lanes
  (KYTY_WAVE32_CLUSTERS=auto, the default; 0 turns it off, force applies it everywhere). The log names each
  shader this applies to. NVIDIA is unchanged.
• Large compute shared memory: some Astro Bot compute shaders (GPU skinning) ask for 48 KiB of shared memory;
  AMD's driver allows 32 KiB, so Kyty cut it to 32 KiB and meshes with more than 682 bones could stretch. Such a
  shader now keeps its shared memory in a video memory buffer (from upstream Kyty, with Kyty's own buffer handling).
  KYTY_LDS_DEVICE_BUFFER = auto (default: only above the device limit), 0 (the old cut) or force;
  KYTY_LDS_DEVICE_BUFFER_MB (default 1024) caps the buffer. NVIDIA is unchanged (48 KiB fits).
• Helper lanes no longer spin: in pixel shaders, helper lanes skip the compare-exchange loops of emulated atomics,
  which could loop forever on AMD (KYTY_PS_HELPER_ATOMICS_SKIP=0 restores the old code).
• Less video memory without the preset: KYTY_FUNCTION_ARRAY_SHRINK is now on by default (it was only set by
  u59-preset.json). Vertex and pixel shaders emulate shared memory with a large per-lane array that AMD backs with
  memory for every lane (about 2 MiB per wave); the shrink cuts it to what the shader can reach (at most 352 dwords in
  every Astro Bot 1.018 shader; about 3.9 GiB less video memory on an RTX 3090). =0 turns it off and says so in the
  log. Two RX 6900 XT players whose launcher had not applied the preset crashed inside the driver without it.
• Crash reports name the module and pipeline: a host crash now writes its fault report into the log file
  (_kyty.txt), not only the console, with the module and offset of the crash (for example a driver DLL), each
  call-chain frame's module, and what the thread was doing (a graphics or compute pipeline build, with the guest
  shader hashes). Graphics pipeline builds are always traced in the log.
• Diagnostics for testers (off or unchanged by default): KYTY_DEVICE_FAULT_DIAGNOSTICS=1 prints a device
  report and enables GPU breadcrumbs (KYTY_GPU_BREADCRUMBS), which name the GPU operations in flight after a device
  loss; KYTY_COMPUTE_WAVE64=0, KYTY_PS_PER_VERTEX=0, KYTY_CLIP_GUARD=0, KYTY_VOLATILE_LOADS=1 and
  KYTY_LOOP_GUARD_SHADERS=all (with KYTY_LOOP_GUARD=<n>) switch off or add single shader features to narrow
  down a driver problem.

Everyone

• Rumble for other controllers: Xbox and other non-DualSense pads now rumble from Astro Bot's haptics (the game
  plays its vibration as audio for the DualSense; Kyty turns it into rumble with the per-channel gains). Thanks to
  Linky362 (PR #20).
• Black lighting without the preset fixed: Astro Bot without the lighting patches no longer renders black when
  u59-preset.json is not applied (launcher without the file, or kyty_emulator started directly):
  KYTY_SRT_VARIANT_READS is now on by default (0 turns it off). The launcher warns when u59-preset.json is
  missing or invalid (it looks next to launcher.exe and one folder up), and the emulator prints a warning at startup
  when the preset's settings are absent. Keep using the preset: without it the game runs much slower (about 20
  instead of 32 fps on an RTX 3090 in the Sky Garden). A failed ray tracing pipeline build, which turns lighting off
  until restart, is now shown as a WARNING in the console.
• Upstream Kyty sync: the 151 upstream commits since the last sync were reviewed; 62 were taken and 20 partly
  (new shader instructions and formats, native 64-bit integers and 8/16-bit storage, store-completion waits, GLC/DLC
  memory coherence, D16 and formatted buffer access, kernel memory fixes, the main-thread window title). Upstream
  changes that conflict with Kyty U59's own designs (resource tracking, MAD rounding, readback sizes) were left out.
  No frame rate change was measured (Sky Garden, casino and pirate island A/B within noise). 8/16-bit storage is now
  a required GPU feature, so very old drivers without it cannot start.
• Graphics fixes from upstream: colour exports routed through CB_SHADER_MASK (KYTY_CB_SHADER_MASK_EXPORTS,
  on through the preset), linear render-target mip chains, rendered contents kept across 1D and 2D texture aliases,
  and streamed texture mips uploaded from partly committed sparse reservations (fixes the exit "failed to read mapped
  guest image backing").
• GPU stores to untracked memory: a shader store through a buffer descriptor computed on the GPU, to memory Kyty
  had not cached on the GPU yet, was dropped. Such stores are now logged and copied back right after the dispatch
  (KYTY_BDA_STORE_LOG, default on; =0 restores the old drop).
• Smaller fixes: a depth image alias is found again for a smaller layer prefix (no second copy), trophy unlocks
  are copied once from the old _SaveData location (issue #18), and unknown socket options return the error codes the
  game expects on Windows.

New defaults in int18

Switch	Default	Off / old behaviour	What it does
KYTY_LDS_WAITCNT_BARRIER	subgroup	workgroup (int17), 0	Scope of the barrier after a shader waits for its LDS writes
KYTY_WAVE32_CLUSTERS	auto	0	32-lane waves stay in their own lanes on 64-lane host subgroups (AMD)
KYTY_LDS_DEVICE_BUFFER	auto	0...

KytyPS5 U59 int18 — AMD fixes, upstream sync, rumble for other controllers

int18 is int17 plus the fixes from the three AMD test builds, a sync with upstream Kyty, and fixes for
missing-preset and controller problems. The first launch after the update rebuilds the shader and program caches
once (the shader code generator changed), so the first minutes can stutter.

What's new in int18

AMD graphics cards

• Title-screen hang fixed: on AMD cards the GPU could stop responding ("device lost") at the same title-screen
  draw. A particle mesh shader waits for its own LDS writes inside branches that only some waves take; Kyty turned
  each wait into a workgroup-wide memory barrier, which AMD's driver handles like a workgroup barrier that the other
  waves never reach. The wait now orders only the wave's own memory (subgroup scope). KYTY_LDS_WAITCNT_BARRIER
  = subgroup (default), workgroup (the old behaviour) or 0 (no barrier).
• Water and particle corruption fixed: AMD runs vertex, mesh and pixel shaders with 64 lanes per subgroup, so one
  host subgroup held two of the PS5's 32-lane waves, and shaders read the other wave's lanes and masks (stretched
  water, pink particle clouds, loops that never end). Each wave now stays in its own 32 lanes
  (KYTY_WAVE32_CLUSTERS=auto, the default; 0 turns it off, force applies it everywhere). The log names each
  shader this applies to. NVIDIA is unchanged.
• Large compute shared memory: some Astro Bot compute shaders (GPU skinning) ask for 48 KiB of shared memory;
  AMD's driver allows 32 KiB, so Kyty cut it to 32 KiB and meshes with more than 682 bones could stretch. Such a
  shader now keeps its shared memory in a video memory buffer (from upstream Kyty, with Kyty's own buffer handling).
  KYTY_LDS_DEVICE_BUFFER = auto (default: only above the device limit), 0 (the old cut) or force;
  KYTY_LDS_DEVICE_BUFFER_MB (default 1024) caps the buffer. NVIDIA is unchanged (48 KiB fits).
• Helper lanes no longer spin: in pixel shaders, helper lanes skip the compare-exchange loops of emulated atomics,
  which could loop forever on AMD (KYTY_PS_HELPER_ATOMICS_SKIP=0 restores the old code).
• Less video memory without the preset: KYTY_FUNCTION_ARRAY_SHRINK is now on by default (it was only set by
  u59-preset.json). Vertex and pixel shaders emulate shared memory with a large per-lane array that AMD backs with
  memory for every lane (about 2 MiB per wave); the shrink cuts it to what the shader can reach (at most 352 dwords in
  every Astro Bot 1.018 shader; about 3.9 GiB less video memory on an RTX 3090). =0 turns it off and says so in the
  log. Two RX 6900 XT players whose launcher had not applied the preset crashed inside the driver without it.
• Crash reports name the module and pipeline: a host crash now writes its fault report into the log file
  (_kyty.txt), not only the console, with the module and offset of the crash (for example a driver DLL), each
  call-chain frame's module, and what the thread was doing (a graphics or compute pipeline build, with the guest
  shader hashes). Graphics pipeline builds are always traced in the log.
• Diagnostics for testers (off or unchanged by default): KYTY_DEVICE_FAULT_DIAGNOSTICS=1 prints a device
  report and enables GPU breadcrumbs (KYTY_GPU_BREADCRUMBS), which name the GPU operations in flight after a device
  loss; KYTY_COMPUTE_WAVE64=0, KYTY_PS_PER_VERTEX=0, KYTY_CLIP_GUARD=0, KYTY_VOLATILE_LOADS=1 and
  KYTY_LOOP_GUARD_SHADERS=all (with KYTY_LOOP_GUARD=<n>) switch off or add single shader features to narrow
  down a driver problem.

Everyone

• Rumble for other controllers: Xbox and other non-DualSense pads now rumble from Astro Bot's haptics (the game
  plays its vibration as audio for the DualSense; Kyty turns it into rumble with the per-channel gains). Thanks to
  Linky362 (PR #20).
• Black lighting without the preset fixed: Astro Bot without the lighting patches no longer renders black when
  u59-preset.json is not applied (launcher without the file, or kyty_emulator started directly):
  KYTY_SRT_VARIANT_READS is now on by default (0 turns it off). The launcher warns when u59-preset.json is
  missing or invalid (it looks next to launcher.exe and one folder up), and the emulator prints a warning at startup
  when the preset's settings are absent. Keep using the preset: without it the game runs much slower (about 20
  instead of 32 fps on an RTX 3090 in the Sky Garden). A failed ray tracing pipeline build, which turns lighting off
  until restart, is now shown as a WARNING in the console.
• Upstream Kyty sync: the 151 upstream commits since the last sync were reviewed; 62 were taken and 20 partly
  (new shader instructions and formats, native 64-bit integers and 8/16-bit storage, store-completion waits, GLC/DLC
  memory coherence, D16 and formatted buffer access, kernel memory fixes, the main-thread window title). Upstream
  changes that conflict with Kyty U59's own designs (resource tracking, MAD rounding, readback sizes) were left out.
  No frame rate change was measured (Sky Garden, casino and pirate island A/B within noise). 8/16-bit storage is now
  a required GPU feature, so very old drivers without it cannot start.
• Graphics fixes from upstream: colour exports routed through CB_SHADER_MASK (KYTY_CB_SHADER_MASK_EXPORTS,
  on through the preset), linear render-target mip chains, rendered contents kept across 1D and 2D texture aliases,
  and streamed texture mips uploaded from partly committed sparse reservations (fixes the exit "failed to read mapped
  guest image backing").
• GPU stores to untracked memory: a shader store through a buffer descriptor computed on the GPU, to memory Kyty
  had not cached on the GPU yet, was dropped. Such stores are now logged and copied back right after the dispatch
  (KYTY_BDA_STORE_LOG, default on; =0 restores the old drop).
• Smaller fixes: a depth image alias is found again for a smaller layer prefix (no second copy), trophy unlocks
  are copied once from the old _SaveData location (issue #18), and unknown socket options return the error codes the
  game expects on Windows.

New defaults in int18

Switch	Default	Off / old behaviour	What it does
KYTY_LDS_WAITCNT_BARRIER	subgroup	workgroup (int17), 0	Scope of the barrier after a shader waits for its LDS writes
KYTY_WAVE32_CLUSTERS	auto	0	32-lane waves stay in their own lanes on 64-lane host subgroups (AMD)
KYTY_LDS_DEVICE_BUFFER	auto	0	Compute shared memory above the device limit in a video memory buffer
KYTY_PS_HELPER_ATOMICS_SKIP	on	0	Pixel-shader helper lanes skip emulated atomic loops
KYTY_FUNCTION_ARRAY_SHRINK	on	0	Per-lane shader arrays cut to their reachable size
KYTY_SRT_VARIANT_READS	on	0	Lighting shaders run without the preset
KYTY_BDA_STORE_LOG	on	0	GPU stores to untracked memory are copied back
KYTY_CB_SHADER_MASK_EXPORTS (preset)	1	0	Colour exports routed by CB_SHADER_MASK

Checked

• All automated tests pass, including the new AMD-path tests (32-lane clusters and the shared-memory buffer, run on
  an RTX 3090 with the AMD limits forced).
• Astro Bot 1.018 on the test PC (RTX 3090, Ryzen 9 7950X3D): normal boot to the title screen and galaxy map with a
  cold shader cache, no console errors.
• The AMD fixes were tested by players on Radeon RX 6000 cards with the AMD test builds; please report how int18
  runs on your AMD card.

Known issues

• AMD graphics cards: the int17 note (device lost with the patches off, slow water in Go-Go Archipelago) should be
  fixed; if you still see it, please send the console text and _kyty.txt (they now name the shader or driver
  module).
• The int17 known issues below still apply otherwise.

KytyPS5 U59 int17 pre-release — faster heavy levels, performance-mode occlusion, many fixes

Pre-release. This is int16.1 plus the newest speed work and fixes (draw batching in the clock tower and Ape
Escape, GPU-built indirect mesh draws, background ray tracing pipelines, and ports from chenxiao07's fork). If
anything works worse than in int16.1, please report it (with your GPU and CPU model) and use int16.1.

What's new in int17

• Galaxy map fixed: the red trails, pillars and blocky red areas around the volcano (the hidden world) and the
  red band at the bottom of the screen are gone. The game fast-clears one of its 64-bit render targets every frame and
  Kyty decoded that clear with a 32-bit decoder, so the clear was dropped and old frames piled up. The same fix covers
  R16F, RG16F, R16, RG16, R8, RG8 and B10G11R11 register clears (KYTY_CLEAR_REGISTER_WIDE=0 restores the old
  behaviour). This bug was also in upstream Kyty.

• Performance mode by default: occlusion queries off. The game asks the GPU, thousands of times per frame,
  whether objects are hidden behind others. Answering those queries accurately costs about half the frame rate in
  the heaviest levels. int17 now answers "visible" to all of them by default (KYTY_GPU_OCCLUSION=0, as other
  Kyty forks do), which roughly doubles the frame rate in the snow level, the clock tower and Ape Escape.
  On the test PC (RTX 3090, Ryzen 9 7950X3D; a test build without the release build's profile optimization, which
  adds about 15%): snow 45 fps, clock tower 35, Ape Escape 38, Sky Garden 31, volcano 59.
  Possible side effects: the game may draw objects or effects that it would normally hide behind walls (rarely
  visible, and it can cost GPU time where a lot is hidden).

• Accurate occlusion toggle: Global settings > Graphics & display > "Accurate occlusion queries (slower; fixes
  rare visibility issues)" turns the accurate mode back on (it passes `--gpu-occlus...

KytyPS5 U59 int16.1 AMD test 4 — a workaround for the AMD driver crash a minute into play

This is a test build for AMD Radeon owners whose game crashes a minute or so into play (AMD test 3 kit:
"crash in amdvlk64.dll+0x22240fc"). Everyone else: please keep using int16.1.

What this build is

Every AMD test 3 run crashed in the same place: a null read inside AMD's Windows Vulkan driver (amdvlk64.dll), in
the part that compiles pipelines with full optimization. It happened on an RX 6900 XT and an RX 6800 XT, with every
setting the kit tried (wave32 clusters off as well). In issue #22's log, the optimized builds of one pixel shader
crashed while the unoptimized builds of the same shader pairs worked. AMD's offline compiler builds every Astro Bot
shader without trouble, and no single shader feature sets the crashing shaders apart (their near-identical
neighbours compile fine), so this build works around the driver instead of changing the shaders:

• Known crashing shaders built unoptimized (on by default, AMD only): pipelines that contain one of the pixel
  shaders seen in a crash are never built with full optimization. That is four pixel shaders across the game
  versions, so only a few pipelines skip the optimizer. The rest are built as before. KYTY_PIPELINE_NO_OPT_SHADERS=<hash,hash,...>
  replaces the list, and =none turns it off.

• Crash guard for optimized builds (on by default, AMD only): if the driver still crashes while optimizing a
  pipeline, the emulator catches it and uses the unoptimized build of that pipeline instead. It also writes the
  pipeline's shaders to _PipelineCache\<title>.noopt.txt for your GPU and driver, so later starts skip the optimizer
  for them from the beginning. The log then says "Pipeline optimization: the driver faulted ...".

  • A driver that crashed may be left in a bad state, so after such a crash the emulator stops saving the Vulkan
    pipeline cache for that session.
  • After 4 crashes in one session it stops building optimized pipelines altogether.
  • If the game still hangs or crashes later in such a session, use the next switch.

  KYTY_PIPELINE_OPT_FAULT_GUARD=0 turns the guard off.

• No optimized pipelines at all: KYTY_PIPELINE_OPTIMIZE=0. No pipeline is ever built with full optimization.
  This is the definitive test of "the optimized build crashes, the unoptimized one works". It may cost GPU speed,
  because unoptimized shader code can run slower, so it is a test switch and not the default.

• Everything else is the same as AMD test 3. The wave32 fix stays on: the clusters-off run crashed the same way, and
  without it the screen shows pink and colorful glitches.

How to test (about 15 minutes; you play two short runs)

Do the same as for AMD test 3: copy your working folder, unzip this build over it, check the launcher once and close
it, then double-click AMD-Test-Kit.cmd. The kit runs:

• noopt-list (you play): this build as it is.
• optimize-off (you play): no optimized pipelines at all (KYTY_PIPELINE_OPTIMIZE=0). Tell us if it stutters or
  runs slower than noopt-list.
• fix: the title screen only, hands off.

Play the first level (near the water) until the kit closes the game. Send the zip it makes, and also your
_PipelineCache\<title>.noopt.txt if it exists: it lists the pipelines the guard caught. -Only clusters-off
(and the other AMD test 3 runs, -Only noopt-none to reproduce the crash on purpose) still work.

Astro's Playroom: clouds or smoke broken into blocks

Same as in AMD test 3: start Playroom by hand twice from a command prompt in the emulator folder. The first time, run
set KYTY_DPP_SKIP_INACTIVE=0 and then launcher.exe. The second time, from a new command prompt, run
set KYTY_WAVE32_CLUSTERS=0 and then launcher.exe. Look at the same clouds each time, and send both _kyty.txt
files with a note on whether the clouds changed. AMD-TEST-KIT-README.txt has the steps.

---

KytyPS5 U59 int16.1 AMD test 3 — shared memory above the AMD limit, and a crash a minute into play

This is a test build for AMD Radeon owners who got past the title screen with AMD test 2. Everyone else: please
keep using int16.1.

What this build is

AMD test 2 fixed the "device lost" at the title screen. Its logs still say "Clamping LDS": one of the game's compute
shaders asks for 48 KiB of shared memory (LDS), NVIDIA allows 48 KiB, AMD 32 KiB, and the emulator cut it to 32 KiB,
so everything that shader kept above 32 KiB read as zero. Some players also see a crash about a minute into play,
inside a DLL (a driver or a Vulkan overlay). This build:

• LDS in a GPU buffer (on by default): a compute shader that needs more shared memory than the card allows keeps
  it in a GPU buffer instead, one region per workgroup (the approach of upstream b0bbaef). It only applies where
  the limit is exceeded, so NVIDIA cards are unchanged. KYTY_LDS_DEVICE_BUFFER=0 restores the AMD test 2 clamp,
  =force puts every compute shader's LDS in the buffer (a test). The log now says "its LDS lives in a device
  buffer" instead of "Clamping LDS", once per shader, with the shader's hash.
• Crash reports name the module and the pipeline: an unhandled crash now prints the DLL it happened in and the
  offset (for example pc=0x00007ff9fc1640fc (amdvlk64.dll+0x...)), what the thread was doing (for example
  "graphics pipeline build (VS, PS) 0x... 0x..."), the registers and the call chain, also into the log file
  (before, that part went only to the console window). Every graphics pipeline build is logged
  (PipelineTrace: graphics build begin/done, and the unoptimized and optimized builds of fast-first).
• Shader shared-memory arrays shrunk by default: pixel and vertex shaders emulate LDS with a 32 KiB array per
  pixel; KYTY_FUNCTION_ARRAY_SHRINK cuts it to what the shader can reach (every such Astro Bot pixel shader to at
  most 352 dwords). It was on only through u59-preset.json; now it is the default (=0 turns it off), so a
  missing preset cannot leave AMD reserving 32 KiB of scratch per pixel. Its lines now also reach _kyty.txt.
• Diagnostic switches for that crash: it came right after the same vertex/pixel shader pair was compiled; that
  pixel shader reads raw per-vertex attributes (fragment barycentrics). KYTY_PS_PER_VERTEX=0 replaces those reads
  by interpolated values (slightly wrong shading there), KYTY_CLIP_GUARD=0 drops the extra clip plane vertex
  shaders get. Both default on.
• Wave32 log lines: each wave32 shader that shares a 64-wide AMD subgroup is logged once with the cross-lane
  operations it uses (for the block-shaped cloud corruption seen in Astro's Playroom).
• Everything else is the same as AMD test 2 (subgroup LDS barrier, wave32 clusters, GPU breadcrumbs).

How to test (about 35 minutes; you play five short runs)

1. Copy your working int16.1 folder (with its _Patches, _PipelineCache and _SaveData) to a new folder, then
   unzip this build into the new folder and let it overwrite the files.
2. Start launcher.exe from the new folder once, check that Astro Bot and your usual patches are listed, then close
   the launcher again. Do not start the game from it.
3. Double-click AMD-Test-Kit.cmd. Before each run marked "YOU PLAY" it waits for Enter: start the game, Dive In,
   and play the first level until the kit closes the game (about 4 minutes after the intro video). After each played
   run it asks what you saw; press Enter if nothing looked wrong, or type a few words.
   • lds-fix: this build as it is.
   • pervertex-off: the per-vertex diagnostic switch above.
   • overlays-off: no Vulkan overlay layers (Steam, Epic, GOG Galaxy, OBS, fossilize), for the crash a minute into
     play.
   • fastfirst-off: pipelines built optimized at once instead of a quick first build (KYTY_PIPELINE_FAST_FIRST=0);
     AMD's quick builds took about 250 ms each in an AMD test 2 log.
   • lds-clamp: the AMD test 2 clamp again, to compare.
   • fix: the title screen only, hands off.
4. When it says "Send this file: ...AMD-Test-Results-.zip", send us that zip, also when every run went fine:
   the logs name the shaders that ask for more shared memory than the card has. Playing another game with this build
   (Astro's Playroom) and seeing broken effects? Send its _kyty.txt too.

Astro's Playroom: clouds or smoke broken into blocks

The kit only runs Astro Bot. If Playroom's clouds or smoke break up into blocks on your AMD card, start Playroom
twice by hand, once with each setting, look at the same clouds, and send the _kyty.txt of each start with a note
on whether the clouds changed. Open a command prompt in the emulator folder (type cmd in the Explorer address bar)
and run:

set KYTY_DPP_SKIP_INACTIVE=0
launcher.exe

Start Playroom, look at the clouds, close the game and the launcher, and keep a copy of _kyty.txt. Then close that
command prompt, open a new one in the same folder, and run:

set KYTY_WAVE32_CLUSTERS=0
launcher.exe

The setting only applies to the launcher started from that command prompt.

You can also just play: this build should behave like AMD test 2 or better. The kit changes nothing outside its
results folder. The zip contains the emulator's console logs, your notes, your OS, GPU, driver, CPU and RAM, and
Windows' records of display-driver resets during the test. AMD-TEST-KIT-README.txt has the details and options.

---

KytyPS5 U59 int16.1 AMD test 2 — the fix for the "device lost" after the intro video on AMD Radeon cards

This is a test build for AMD Radeon owners whose game stops with "device lost" / ErrorDeviceLost, usually right
after the intro video, at the title screen. Everyone else: please keep using int16.1.

What this build is

AMD test 1 showed which draw the GPU was stuck in: the title screen's particle draw (a me...

KytyPS5 U59 int16.1 AMD test 3 — shared memory above the AMD limit, and a crash a minute into play

This is a test build for AMD Radeon owners who got past the title screen with AMD test 2. Everyone else: please
keep using int16.1.

What this build is

AMD test 2 fixed the "device lost" at the title screen. Its logs still say "Clamping LDS": one of the game's compute
shaders asks for 48 KiB of shared memory (LDS), NVIDIA allows 48 KiB, AMD 32 KiB, and the emulator cut it to 32 KiB,
so everything that shader kept above 32 KiB read as zero. Some players also see a crash about a minute into play,
inside a DLL (a driver or a Vulkan overlay). This build:

• LDS in a GPU buffer (on by default): a compute shader that needs more shared memory than the card allows keeps
  it in a GPU buffer instead, one region per workgroup (the approach of upstream b0bbaef). It only applies where
  the limit is exceeded, so NVIDIA cards are unchanged. KYTY_LDS_DEVICE_BUFFER=0 restores the AMD test 2 clamp,
  =force puts every compute shader's LDS in the buffer (a test). The log now says "its LDS lives in a device
  buffer" instead of "Clamping LDS", once per shader, with the shader's hash.
• Crash reports name the module and the pipeline: an unhandled crash now prints the DLL it happened in and the
  offset (for example pc=0x00007ff9fc1640fc (amdvlk64.dll+0x...)), what the thread was doing (for example
  "graphics pipeline build (VS, PS) 0x... 0x..."), the registers and the call chain, also into the log file
  (before, that part went only to the console window). Every graphics pipeline build is logged
  (PipelineTrace: graphics build begin/done, and the unoptimized and optimized builds of fast-first).
• Shader shared-memory arrays shrunk by default: pixel and vertex shaders emulate LDS with a 32 KiB array per
  pixel; KYTY_FUNCTION_ARRAY_SHRINK cuts it to what the shader can reach (every such Astro Bot pixel shader to at
  most 352 dwords). It was on only through u59-preset.json; now it is the default (=0 turns it off), so a
  missing preset cannot leave AMD reserving 32 KiB of scratch per pixel. Its lines now also reach _kyty.txt.
• Diagnostic switches for that crash: it came right after the same vertex/pixel shader pair was compiled; that
  pixel shader reads raw per-vertex attributes (fragment barycentrics). KYTY_PS_PER_VERTEX=0 replaces those reads
  by interpolated values (slightly wrong shading there), KYTY_CLIP_GUARD=0 drops the extra clip plane vertex
  shaders get. Both default on.
• Wave32 log lines: each wave32 shader that shares a 64-wide AMD subgroup is logged once with the cross-lane
  operations it uses (for the block-shaped cloud corruption seen in Astro's Playroom).
• Everything else is the same as AMD test 2 (subgroup LDS barrier, wave32 clusters, GPU breadcrumbs).

How to test (about 35 minutes; you play five short runs)

1. Copy your working int16.1 folder (with its _Patches, _PipelineCache and _SaveData) to a new folder, then
   unzip this build into the new folder and let it overwrite the files.
2. Start launcher.exe from the new folder once, check that Astro Bot and your usual patches are listed, then close
   the launcher again. Do not start the game from it.
3. Double-click AMD-Test-Kit.cmd. Before each run marked "YOU PLAY" it waits for Enter: start the game, Dive In,
   and play the first level until the kit closes the game (about 4 minutes after the intro video). After each played
   run it asks what you saw; press Enter if nothing looked wrong, or type a few words.
   • lds-fix: this build as it is.
   • pervertex-off: the per-vertex diagnostic switch above.
   • overlays-off: no Vulkan overlay layers (Steam, Epic, GOG Galaxy, OBS, fossilize), for the crash a minute into
     play.
   • fastfirst-off: pipelines built optimized at once instead of a quick first build (KYTY_PIPELINE_FAST_FIRST=0);
     AMD's quick builds took about 250 ms each in an AMD test 2 log.
   • lds-clamp: the AMD test 2 clamp again, to compare.
   • fix: the title screen only, hands off.
4. When it says "Send this file: ...AMD-Test-Results-.zip", send us that zip, also when every run went fine:
   the logs name the shaders that ask for more shared memory than the card has. Playing another game with this build
   (Astro's Playroom) and seeing broken effects? Send its _kyty.txt too.

You can also just play: this build should behave like AMD test 2 or better. The kit changes nothing outside its
results folder. The zip contains the emulator's console logs, your notes, your OS, GPU, driver, CPU and RAM, and
Windows' records of display-driver resets during the test. AMD-TEST-KIT-README.txt has the details and options.

---

KytyPS5 U59 int16.1 AMD test 2 — the fix for the "device lost" after the intro video on AMD Radeon cards

This is a test build for AMD Radeon owners whose game stops with "device lost" / ErrorDeviceLost, usually right
after the intro video, at the title screen. Everyone else: please keep using int16.1.

What this build is

AMD test 1 showed which draw the GPU was stuck in: the title screen's particle draw (a mesh shader). That shader
waits at a barrier for part of its workgroup that never arrives, because of an extra memory barrier the emulator
adds after shared-memory (LDS) writes. This build has the fix, plus a second fix found while checking it:

• LDS barrier fix (on by default): the barrier added at S_WAITCNT lgkmcnt(0) is now limited to the wave's own
  lanes (subgroup scope) instead of the whole workgroup. KYTY_LDS_WAITCNT_BARRIER=workgroup restores int16.1,
  KYTY_LDS_WAITCNT_BARRIER=0 removes the barrier (the workaround from the GitHub issue, which shows pink particle
  clouds on the title screen).
• Wave32 fix (on by default, AMD only): AMD runs vertex, mesh and pixel shaders 64 lanes wide, so two 32-lane
  game waves share one. Each wave now keeps to its own 32 lanes (KYTY_WAVE32_CLUSTERS=0 turns this off).
• On NVIDIA both fixes leave the image and the speed unchanged.
• Everything else is the same as AMD test 1 (int16.1 plus the GPU breadcrumbs and the helper-lane fix).

How to test (about 15 minutes, hands off)

1. Copy your working int16.1 folder (with its _Patches, _PipelineCache and _SaveData) to a new folder, then
   unzip this build into the new folder and let it overwrite the files.
2. Start launcher.exe from the new folder once, check that Astro Bot and your usual patches are listed, then close
   the launcher again. Do not start the game from it.
3. Double-click AMD-Test-Kit.cmd and leave the PC alone. The game opens and closes by itself 6 times. Two of the
   runs bring back the old setting on purpose, so the screen may go black for a few seconds when the GPU resets;
   that is expected.
4. When it says "Send this file: ...AMD-Test-Results-.zip", send us that zip (and, if you can, a note on whether
   the title screen showed pink clouds in the runs that kept going).

You can also just play: this build should get past the title screen on its own. The kit changes nothing outside its
results folder. The zip contains the emulator's console logs, your OS, GPU, driver, CPU and RAM, and Windows' records
of display-driver resets during the test. AMD-TEST-KIT-README.txt has the details and options.

---

KytyPS5 U59 int16.1 AMD test — for AMD Radeon cards that crash ("device lost") after the intro video

This is a test build for AMD Radeon owners whose game stops with "device lost" / ErrorDeviceLost, usually right
after the intro video. Everyone else: please keep using int16.1.

What this build is

int16.1 plus crash diagnostics for AMD graphics cards and an automatic test kit. We have no AMD card ourselves, so
one test session on your PC tells us which part of the emulator makes the GPU stop responding.

• GPU breadcrumbs: after a device loss the console now names what the GPU was running (queue, operation and
  shader). They switch on with KYTY_DEVICE_FAULT_DIAGNOSTICS=1, which the test kit sets for you.
• One new fix on by default: pixel-shader helper lanes skip the compare-exchange loops that emulate some stores
  and atomics (KYTY_PS_HELPER_ATOMICS_SKIP=0 restores the old code). It is safe on NVIDIA.
• New option KYTY_PIPELINE_FAST_FIRST=nvidia: quick unoptimized pipeline builds only on NVIDIA.
• Everything else is the same as int16.1.

How to test (about 35 minutes, hands off)

1. Copy your working int16.1 folder (with its _Patches, _PipelineCache and _SaveData) to a new folder, then
   unzip this build into the new folder and let it overwrite the files.
2. Start launcher.exe from the new folder once, check that Astro Bot and your usual patches are listed, then close
   the launcher again. Do not start the game from it.
3. Double-click AMD-Test-Kit.cmd and leave the PC alone. The game opens and closes by itself about 19 times, each
   time with one setting changed. The screen may go black for a few seconds when the GPU resets; that is expected.
4. When it says "Send this file: ...AMD-Test-Results-.zip", send us that zip.

The kit changes nothing outside its results folder. The zip contains the emulator's console logs, your OS, GPU,
driver, CPU and RAM, and Windows' records of display-driver resets during the test. AMD-TEST-KIT-README.txt has
the details and options.

---

int16.1 release notes

int16.1 is the int16 pre-release plus low-risk fixes and a much faster default for heavy levels. If anything works
worse than in the int16 pre-release or int15, please report it (with your GPU and CPU model).

What's new in int16.1

• Galaxy map fixed: the red trails, pillars and blocky red areas around the volcano (the hidden world) and the
  red band at the bottom of the screen are gone. The game fast-clears one of its 64-bit render targets every frame and
  Kyty...

KytyPS5 U59 int16.1 AMD test 2 — the fix for the "device lost" after the intro video on AMD Radeon cards

This is a test build for AMD Radeon owners whose game stops with "device lost" / ErrorDeviceLost, usually right
after the intro video, at the title screen. Everyone else: please keep using int16.1.

What this build is

AMD test 1 showed which draw the GPU was stuck in: the title screen's particle draw (a mesh shader). That shader
waits at a barrier for part of its workgroup that never arrives, because of an extra memory barrier the emulator
adds after shared-memory (LDS) writes. This build has the fix, plus a second fix found while checking it:

• LDS barrier fix (on by default): the barrier added at S_WAITCNT lgkmcnt(0) is now limited to the wave's own
  lanes (subgroup scope) instead of the whole workgroup. KYTY_LDS_WAITCNT_BARRIER=workgroup restores int16.1,
  KYTY_LDS_WAITCNT_BARRIER=0 removes the barrier (the workaround from the GitHub issue, which shows pink particle
  clouds on the title screen).
• Wave32 fix (on by default, AMD only): AMD runs vertex, mesh and pixel shaders 64 lanes wide, so two 32-lane
  game waves share one. Each wave now keeps to its own 32 lanes (KYTY_WAVE32_CLUSTERS=0 turns this off).
• On NVIDIA both fixes leave the image and the speed unchanged.
• Everything else is the same as AMD test 1 (int16.1 plus the GPU breadcrumbs and the helper-lane fix).

How to test (about 15 minutes, hands off)

1. Copy your working int16.1 folder (with its _Patches, _PipelineCache and _SaveData) to a new folder, then
   unzip this build into the new folder and let it overwrite the files.
2. Start launcher.exe from the new folder once, check that Astro Bot and your usual patches are listed, then close
   the launcher again. Do not start the game from it.
3. Double-click AMD-Test-Kit.cmd and leave the PC alone. The game opens and closes by itself 6 times. Two of the
   runs bring back the old setting on purpose, so the screen may go black for a few seconds when the GPU resets;
   that is expected.
4. When it says "Send this file: ...AMD-Test-Results-.zip", send us that zip (and, if you can, a note on whether
   the title screen showed pink clouds in the runs that kept going).

You can also just play: this build should get past the title screen on its own. The kit changes nothing outside its
results folder. The zip contains the emulator's console logs, your OS, GPU, driver, CPU and RAM, and Windows' records
of display-driver resets during the test. AMD-TEST-KIT-README.txt has the details and options.

---

KytyPS5 U59 int16.1 AMD test — for AMD Radeon cards that crash ("device lost") after the intro video

This is a test build for AMD Radeon owners whose game stops with "device lost" / ErrorDeviceLost, usually right
after the intro video. Everyone else: please keep using int16.1.

What this build is

int16.1 plus crash diagnostics for AMD graphics cards and an automatic test kit. We have no AMD card ourselves, so
one test session on your PC tells us which part of the emulator makes the GPU stop responding.

• GPU breadcrumbs: after a device loss the console now names what the GPU was running (queue, operation and
  shader). They switch on with KYTY_DEVICE_FAULT_DIAGNOSTICS=1, which the test kit sets for you.
• One new fix on by default: pixel-shader helper lanes skip the compare-exchange loops that emulate some stores
  and atomics (KYTY_PS_HELPER_ATOMICS_SKIP=0 restores the old code). It is safe on NVIDIA.
• New option KYTY_PIPELINE_FAST_FIRST=nvidia: quick unoptimized pipeline builds only on NVIDIA.
• Everything else is the same as int16.1.

How to test (about 35 minutes, hands off)

1. Copy your working int16.1 folder (with its _Patches, _PipelineCache and _SaveData) to a new folder, then
   unzip this build into the new folder and let it overwrite the files.
2. Start launcher.exe from the new folder once, check that Astro Bot and your usual patches are listed, then close
   the launcher again. Do not start the game from it.
3. Double-click AMD-Test-Kit.cmd and leave the PC alone. The game opens and closes by itself about 19 times, each
   time with one setting changed. The screen may go black for a few seconds when the GPU resets; that is expected.
4. When it says "Send this file: ...AMD-Test-Results-.zip", send us that zip.

The kit changes nothing outside its results folder. The zip contains the emulator's console logs, your OS, GPU,
driver, CPU and RAM, and Windows' records of display-driver resets during the test. AMD-TEST-KIT-README.txt has
the details and options.

---

int16.1 release notes

int16.1 is the int16 pre-release plus low-risk fixes and a much faster default for heavy levels. If anything works
worse than in the int16 pre-release or int15, please report it (with your GPU and CPU model).

What's new in int16.1

• Galaxy map fixed: the red trails, pillars and blocky red areas around the volcano (the hidden world) and the
  red band at the bottom of the screen are gone. The game fast-clears one of its 64-bit render targets every frame and
  Kyty decoded that clear with a 32-bit decoder, so the clear was dropped and old frames piled up. The same fix covers
  R16F, RG16F, R16, RG16, R8, RG8 and B10G11R11 register clears (KYTY_CLEAR_REGISTER_WIDE=0 restores the old
  behaviour). This bug was also in upstream Kyty.

• Performance mode by default: occlusion queries off. The game asks the GPU, thousands of times per frame,
  whether objects are hidden behind others. Answering those queries accurately costs about half the frame rate in
  the heaviest levels. int16.1 now answers "visible" to all of them by default (KYTY_GPU_OCCLUSION=0, as other
  Kyty forks do), which roughly doubles the frame rate in the snow level, the clock tower and Ape Escape.
  On the test PC (RTX 3090, Ryzen 9 7950X3D; a test build without the release build's profile optimization, which
  adds about 15%): snow 41 fps, clock tower 31, Ape Escape 31, Sky Garden 30, volcano 59 (accurate mode: snow 20).
  Possible side effects: the game may draw objects or effects that it would normally hide behind walls (rarely
  visible, and it can cost GPU time where a lot is hidden).

• Accurate occlusion toggle: Global settings > Graphics & display > "Accurate occlusion queries (slower; fixes
  rare visibility issues)" turns the accurate mode back on (it passes --gpu-occlusion on, which overrides the
  preset). Use it if you see objects through walls or other visibility problems, and please report where. The
  accurate mode includes the volcano lava fix below.

• Volcano lava fixed (accurate occlusion mode): with the camera tilted down in the volcano level, the lava lake
  disappeared (flat orange fog). The emulator paired the game's occlusion queries by a wrong address rule; it now
  pairs them by address, which also keeps the occlusion gate on in Sky Garden.

• Hang watchdog no longer kills the game: on RTX 50 cards (where the watchdog is on by default) a 5 s stall could
  end the emulator with "Unhandled host exception ... RtlVirtualUnwind". The watchdog's stack dump is now guarded.

• Out of video memory no longer ends the emulator: an image that does not fit goes to system memory (slower; the
  console prints "Vulkan: video memory full ..."), and a buffer that does not fit is retried.

• File id fix: Astro Bot file ids are assigned one per file; six pairs of game files shared an id and one of each
  pair could be read in place of the other (one of them a galaxy-map icon).

• Audio: sounds the game sends to the DualSense speaker now play at 30% on the PC speakers (they were up to 19 dB
  louder than music and effects). Four new volume sliders (master, game sound, music, controller speaker on PC) are in
  the launcher's global Audio settings. The console also prints audio port diagnostics (AudioDiag lines).

• "AMD CPU patch" renamed to "Intel CPU compatibility (AMD instruction patch)" in the launcher's Compatibility
  settings. It is the same setting (saved configurations and --amd-cpu keep working): on Intel CPUs it emulates the
  PS5's AMD-only instructions (EXTRQ/INSERTQ) and AMD's VRSQRTPS results. It is not needed on AMD CPUs.

• Smaller fixes ported from chenxiao07's fork: a clean process exit after a fatal error (no more unkillable
  emulator processes), a faster guest memory search while streaming, and faster release shader translation.

• Level title text fixed (fixed after the int16 pre-release): the level name is drawn as in int14 again.

New defaults and how to turn each off

Switch	Default	Off / old behaviour	What it does
KYTY_GPU_OCCLUSION (preset)	0 (performance mode)	1, or the launcher's "Accurate occlusion queries" (--gpu-occlusion on)	Occlusion queries answered "visible" instead of counted on the GPU
KYTY_IMAGE_SYSMEM_FALLBACK	on	0	Images that do not fit in video memory go to system memory
KYTY_APR_HASHED_IDS	off	1 restores the old hashed ids	One file id per file path
KYTY_FATAL_EXIT_PROCESS	off	1 restores the old exit path	Fatal errors end the process with TerminateProcess
KYTY_IR_VALIDATE	off in release	1 runs it	Shader IR validation
KYTY_AUDIO_DIAG	on	0	Audio port open/close lines in the console (256 lines max)
Audio volumes	master/main/music 100%, pad speaker on PC 30%	launcher sliders, or KYTY_AUDIO_MASTER_VOLUME, KYTY_AUDIO_MAIN_VOLUME, KYTY_AUDIO_MUSIC_VOLUME, KYTY_AUDIO_PAD_SPEAKER_ON_MAIN_VOLUME (percent, 0-200)

Optional (off by default): KYTY_WIDE_FILL_CLEAR=1, KYTY_GUEST_STORAGE_REPEAT=1, KYTY_READBACK_MERGE_GAP_KB=64,
and the test tool KYTY_VRAM_LIMIT_MB=<MiB> (imitates a smaller GPU).

From int16

• Astro Bot without the "non RT patch": the game's own ti...

KytyPS5 U59 int16.1 AMD test — for AMD Radeon cards that crash ("device lost") after the intro video

This is a test build for AMD Radeon owners whose game stops with "device lost" / ErrorDeviceLost, usually right
after the intro video. Everyone else: please keep using int16.1.

What this build is

int16.1 plus crash diagnostics for AMD graphics cards and an automatic test kit. We have no AMD card ourselves, so
one test session on your PC tells us which part of the emulator makes the GPU stop responding.

• GPU breadcrumbs: after a device loss the console now names what the GPU was running (queue, operation and
  shader). They switch on with KYTY_DEVICE_FAULT_DIAGNOSTICS=1, which the test kit sets for you.
• One new fix on by default: pixel-shader helper lanes skip the compare-exchange loops that emulate some stores
  and atomics (KYTY_PS_HELPER_ATOMICS_SKIP=0 restores the old code). It is safe on NVIDIA.
• New option KYTY_PIPELINE_FAST_FIRST=nvidia: quick unoptimized pipeline builds only on NVIDIA.
• Everything else is the same as int16.1.

How to test (about 35 minutes, hands off)

1. Copy your working int16.1 folder (with its _Patches, _PipelineCache and _SaveData) to a new folder, then
   unzip this build into the new folder and let it overwrite the files.
2. Start launcher.exe from the new folder once, check that Astro Bot and your usual patches are listed, then close
   the launcher again. Do not start the game from it.
3. Double-click AMD-Test-Kit.cmd and leave the PC alone. The game opens and closes by itself about 19 times, each
   time with one setting changed. The screen may go black for a few seconds when the GPU resets; that is expected.
4. When it says "Send this file: ...AMD-Test-Results-.zip", send us that zip.

The kit changes nothing outside its results folder. The zip contains the emulator's console logs, your OS, GPU,
driver, CPU and RAM, and Windows' records of display-driver resets during the test. AMD-TEST-KIT-README.txt has
the details and options.

---

int16.1 release notes

int16.1 is the int16 pre-release plus low-risk fixes and a much faster default for heavy levels. If anything works
worse than in the int16 pre-release or int15, please report it (with your GPU and CPU model).

What's new in int16.1

• Galaxy map fixed: the red trails, pillars and blocky red areas around the volcano (the hidden world) and the
  red band at the bottom of the screen are gone. The game fast-clears one of its 64-bit render targets every frame and
  Kyty decoded that clear with a 32-bit decoder, so the clear was dropped and old frames piled up. The same fix covers
  R16F, RG16F, R16, RG16, R8, RG8 and B10G11R11 register clears (KYTY_CLEAR_REGISTER_WIDE=0 restores the old
  behaviour). This bug was also in upstream Kyty.

• Performance mode by default: occlusion queries off. The game asks the GPU, thousands of times per frame,
  whether objects are hidden behind others. Answering those queries accurately costs about half the frame rate in
  the heaviest levels. int16.1 now answers "visible" to all of them by default (KYTY_GPU_OCCLUSION=0, as other
  Kyty forks do), which roughly doubles the frame rate in the snow level, the clock tower and Ape Escape.
  On the test PC (RTX 3090, Ryzen 9 7950X3D; a test build without the release build's profile optimization, which
  adds about 15%): snow 41 fps, clock tower 31, Ape Escape 31, Sky Garden 30, volcano 59 (accurate mode: snow 20).
  Possible side effects: the game may draw objects or effects that it would normally hide behind walls (rarely
  visible, and it can cost GPU time where a lot is hidden).

• Accurate occlusion toggle: Global settings > Graphics & display > "Accurate occlusion queries (slower; fixes
  rare visibility issues)" turns the accurate mode back on (it passes --gpu-occlusion on, which overrides the
  preset). Use it if you see objects through walls or other visibility problems, and please report where. The
  accurate mode includes the volcano lava fix below.

• Volcano lava fixed (accurate occlusion mode): with the camera tilted down in the volcano level, the lava lake
  disappeared (flat orange fog). The emulator paired the game's occlusion queries by a wrong address rule; it now
  pairs them by address, which also keeps the occlusion gate on in Sky Garden.

• Hang watchdog no longer kills the game: on RTX 50 cards (where the watchdog is on by default) a 5 s stall could
  end the emulator with "Unhandled host exception ... RtlVirtualUnwind". The watchdog's stack dump is now guarded.

• Out of video memory no longer ends the emulator: an image that does not fit goes to system memory (slower; the
  console prints "Vulkan: video memory full ..."), and a buffer that does not fit is retried.

• File id fix: Astro Bot file ids are assigned one per file; six pairs of game files shared an id and one of each
  pair could be read in place of the other (one of them a galaxy-map icon).

• Audio: sounds the game sends to the DualSense speaker now play at 30% on the PC speakers (they were up to 19 dB
  louder than music and effects). Four new volume sliders (master, game sound, music, controller speaker on PC) are in
  the launcher's global Audio settings. The console also prints audio port diagnostics (AudioDiag lines).

• "AMD CPU patch" renamed to "Intel CPU compatibility (AMD instruction patch)" in the launcher's Compatibility
  settings. It is the same setting (saved configurations and --amd-cpu keep working): on Intel CPUs it emulates the
  PS5's AMD-only instructions (EXTRQ/INSERTQ) and AMD's VRSQRTPS results. It is not needed on AMD CPUs.

• Smaller fixes ported from chenxiao07's fork: a clean process exit after a fatal error (no more unkillable
  emulator processes), a faster guest memory search while streaming, and faster release shader translation.

• Level title text fixed (fixed after the int16 pre-release): the level name is drawn as in int14 again.

New defaults and how to turn each off

Switch	Default	Off / old behaviour	What it does
KYTY_GPU_OCCLUSION (preset)	0 (performance mode)	1, or the launcher's "Accurate occlusion queries" (--gpu-occlusion on)	Occlusion queries answered "visible" instead of counted on the GPU
KYTY_IMAGE_SYSMEM_FALLBACK	on	0	Images that do not fit in video memory go to system memory
KYTY_APR_HASHED_IDS	off	1 restores the old hashed ids	One file id per file path
KYTY_FATAL_EXIT_PROCESS	off	1 restores the old exit path	Fatal errors end the process with TerminateProcess
KYTY_IR_VALIDATE	off in release	1 runs it	Shader IR validation
KYTY_AUDIO_DIAG	on	0	Audio port open/close lines in the console (256 lines max)
Audio volumes	master/main/music 100%, pad speaker on PC 30%	launcher sliders, or KYTY_AUDIO_MASTER_VOLUME, KYTY_AUDIO_MAIN_VOLUME, KYTY_AUDIO_MUSIC_VOLUME, KYTY_AUDIO_PAD_SPEAKER_ON_MAIN_VOLUME (percent, 0-200)

Optional (off by default): KYTY_WIDE_FILL_CLEAR=1, KYTY_GUEST_STORAGE_REPEAT=1, KYTY_READBACK_MERGE_GAP_KB=64,
and the test tool KYTY_VRAM_LIMIT_MB=<MiB> (imitates a smaller GPU).

From int16

• Astro Bot without the "non RT patch": the game's own tiled lighting, GI probes and ray traced shadows now run,
  so the two patches are no longer needed. Without them the robots used to be black and the lighting missing, because
  the lighting shaders contain ray-tracing instructions and were skipped. Those instructions now run in software on
  any GPU (NVIDIA, AMD and Intel).
• How to play without the patches: open the launcher, select Astro Bot, open the cheats window and untick
  "Select the existing non-tiled deferred-lighting renderer" and "Disable GI probes and lighting shaders". If you
  see problems, tick them again; this build also runs fine with them.
• Speed: on the test PC (RTX 3090, Ryzen 9 7950X3D) the full graphics cost about 3% in Sky Garden (about 33.6 fps
  against 34.8 fps with the patches) and practically nothing at the snow level start (21.1 against 21.2 fps).
• First launch: the first start after an update builds the large lighting shaders, which can freeze the picture
  for up to about 20 seconds once.

Also in int15

• No more "unsupported sampled depth image" stops on fast GPUs, no crash when leaving extra levels in Sky Garden,
  about 20 upstream shader fixes, a Demon's Souls fix for an unknown image format, and thread priority changes.
• From int14: shader precompile after updates, faster first pipelines, a better automatic GPU choice, fewer loading
  crashes, Demon's Souls character creation, and the adaptive trigger fixes.

Checked

• All automated tests pass.
• Astro Bot 1.018 on one PC (RTX 3090, Ryzen 9 7950X3D), patches off: Sky Garden, snow, clock tower, Ape Escape and
  the volcano with the default (performance) occlusion mode, and snow plus the volcano lava with the camera tilted
  down in accurate mode, with no console errors.
• Not tested by us on AMD or Intel GPUs or on version 1.007; player reports from RTX 50
  and AMD cards are listed under known issues.

Installing

1. Download KytyPS5-U59-Windows-x64.zip and extract it to a new folder.
2. Open launcher.exe. Your existing game list and settings are picked up automatically.
3. Game patches go in a _Patches folder next to the launcher; saves are kept per folder in _SaveData.

Known issues

• Grass flicker: small grass/moss clumps on sand (e.g. the crash-site hub) can flicker between two looks from
  frame to frame. Also in earlier builds; being looked at.
• RTX 50 series: the first launch after installing or updating could crash once while the ray tracing shaders
  were being built. The hang watchdog fix should cure the crash we know of; if it still happens, start the game again
  and p...

KytyPS5 U59 int16.1 — Astro Bot with full lighting, GI and ray tracing, performance-mode occlusion

int16.1 is the int16 pre-release plus low-risk fixes and a much faster default for heavy levels. If anything works
worse than in the int16 pre-release or int15, please report it (with your GPU and CPU model).

What's new in int16.1

• Galaxy map fixed: the red trails, pillars and blocky red areas around the volcano (the hidden world) and the
  red band at the bottom of the screen are gone. The game fast-clears one of its 64-bit render targets every frame and
  Kyty decoded that clear with a 32-bit decoder, so the clear was dropped and old frames piled up. The same fix covers
  R16F, RG16F, R16, RG16, R8, RG8 and B10G11R11 register clears (KYTY_CLEAR_REGISTER_WIDE=0 restores the old
  behaviour). This bug was also in upstream Kyty.

• Performance mode by default: occlusion queries off. The game asks the GPU, thousands of times per frame,
  whether objects are hidden behind others. Answering those queries accurately costs about half the frame rate in
  the heaviest levels. int16.1 now answers "visible" to all of them by default (KYTY_GPU_OCCLUSION=0, as other
  Kyty forks do), which roughly doubles the frame rate in the snow level, the clock tower and Ape Escape.
  On the test PC (RTX 3090, Ryzen 9 7950X3D; a test build without the release build's profile optimization, which
  adds about 15%): snow 41 fps, clock tower 31, Ape Escape 31, Sky Garden 30, volcano 59 (accurate mode: snow 20).
  Possible side effects: the game may draw objects or effects that it would normally hide behind walls (rarely
  visible, and it can cost GPU time where a lot is hidden).

• Accurate occlusion toggle: Global settings > Graphics & display > "Accurate occlusion queries (slower; fixes
  rare visibility issues)" turns the accurate mode back on (it passes --gpu-occlusion on, which overrides the
  preset). Use it if you see objects through walls or other visibility problems, and please report where. The
  accurate mode includes the volcano lava fix below.

• Volcano lava fixed (accurate occlusion mode): with the camera tilted down in the volcano level, the lava lake
  disappeared (flat orange fog). The emulator paired the game's occlusion queries by a wrong address rule; it now
  pairs them by address, which also keeps the occlusion gate on in Sky Garden.

• Hang watchdog no longer kills the game: on RTX 50 cards (where the watchdog is on by default) a 5 s stall could
  end the emulator with "Unhandled host exception ... RtlVirtualUnwind". The watchdog's stack dump is now guarded.

• Out of video memory no longer ends the emulator: an image that does not fit goes to system memory (slower; the
  console prints "Vulkan: video memory full ..."), and a buffer that does not fit is retried.

• File id fix: Astro Bot file ids are assigned one per file; six pairs of game files shared an id and one of each
  pair could be read in place of the other (one of them a galaxy-map icon).

• Audio: sounds the game sends to the DualSense speaker now play at 30% on the PC speakers (they were up to 19 dB
  louder than music and effects). Four new volume sliders (master, game sound, music, controller speaker on PC) are in
  the launcher's global Audio settings. The console also prints audio port diagnostics (AudioDiag lines).

• "AMD CPU patch" renamed to "Intel CPU compatibility (AMD instruction patch)" in the launcher's Compatibility
  settings. It is the same setting (saved configurations and --amd-cpu keep working): on Intel CPUs it emulates the
  PS5's AMD-only instructions (EXTRQ/INSERTQ) and AMD's VRSQRTPS results. It is not needed on AMD CPUs.

• Smaller fixes ported from chenxiao07's fork: a clean process exit after a fatal error (no more unkillable
  emulator processes), a faster guest memory search while streaming, and faster release shader translation.

• Level title text fixed (fixed after the int16 pre-release): the level name is drawn as in int14 again.

New defaults and how to turn each off

Switch	Default	Off / old behaviour	What it does
KYTY_GPU_OCCLUSION (preset)	0 (performance mode)	1, or the launcher's "Accurate occlusion queries" (--gpu-occlusion on)	Occlusion queries answered "visible" instead of counted on the GPU
KYTY_IMAGE_SYSMEM_FALLBACK	on	0	Images that do not fit in video memory go to system memory
KYTY_APR_HASHED_IDS	off	1 restores the old hashed ids	One file id per file path
KYTY_FATAL_EXIT_PROCESS	off	1 restores the old exit path	Fatal errors end the process with TerminateProcess
KYTY_IR_VALIDATE	off in release	1 runs it	Shader IR validation
KYTY_AUDIO_DIAG	on	0	Audio port open/close lines in the console (256 lines max)
Audio volumes	master/main/music 100%, pad speaker on PC 30%	launcher sliders, or KYTY_AUDIO_MASTER_VOLUME, KYTY_AUDIO_MAIN_VOLUME, KYTY_AUDIO_MUSIC_VOLUME, KYTY_AUDIO_PAD_SPEAKER_ON_MAIN_VOLUME (percent, 0-200)

Optional (off by default): KYTY_WIDE_FILL_CLEAR=1, KYTY_GUEST_STORAGE_REPEAT=1, KYTY_READBACK_MERGE_GAP_KB=64,
and the test tool KYTY_VRAM_LIMIT_MB=<MiB> (imitates a smaller GPU).

From int16

• Astro Bot without the "non RT patch": the game's own tiled lighting, GI probes and ray traced shadows now run,
  so the two patches are no longer needed. Without them the robots used to be black and the lighting missing, because
  the lighting shaders contain ray-tracing instructions and were skipped. Those instructions now run in software on
  any GPU (NVIDIA, AMD and Intel).
• How to play without the patches: open the launcher, select Astro Bot, open the cheats window and untick
  "Select the existing non-tiled deferred-lighting renderer" and "Disable GI probes and lighting shaders". If you
  see problems, tick them again; this build also runs fine with them.
• Speed: on the test PC (RTX 3090, Ryzen 9 7950X3D) the full graphics cost about 3% in Sky Garden (about 33.6 fps
  against 34.8 fps with the patches) and practically nothing at the snow level start (21.1 against 21.2 fps).
• First launch: the first start after an update builds the large lighting shaders, which can freeze the picture
  for up to about 20 seconds once.

Also in int15

• No more "unsupported sampled depth image" stops on fast GPUs, no crash when leaving extra levels in Sky Garden,
  about 20 upstream shader fixes, a Demon's Souls fix for an unknown image format, and thread priority changes.
• From int14: shader precompile after updates, faster first pipelines, a better automatic GPU choice, fewer loading
  crashes, Demon's Souls character creation, and the adaptive trigger fixes.

Checked

• All automated tests pass.
• Astro Bot 1.018 on one PC (RTX 3090, Ryzen 9 7950X3D), patches off: Sky Garden, snow, clock tower, Ape Escape and
  the volcano with the default (performance) occlusion mode, and snow plus the volcano lava with the camera tilted
  down in accurate mode, with no console errors.
• Not tested by us on AMD or Intel GPUs or on version 1.007; player reports from RTX 50
  and AMD cards are listed under known issues.

Installing

1. Download KytyPS5-U59-Windows-x64.zip and extract it to a new folder.
2. Open launcher.exe. Your existing game list and settings are picked up automatically.
3. Game patches go in a _Patches folder next to the launcher; saves are kept per folder in _SaveData.

Known issues

• Grass flicker: small grass/moss clumps on sand (e.g. the crash-site hub) can flicker between two looks from
  frame to frame. Also in earlier builds; being looked at.
• RTX 50 series: the first launch after installing or updating could crash once while the ray tracing shaders
  were being built. The hang watchdog fix should cure the crash we know of; if it still happens, start the game again
  and please send the console text.
• AMD graphics cards: with the two patches turned off, the GPU can stop responding ("device lost"), and the water
  in Go-Go Archipelago can make the frame rate drop sharply. On AMD, keep the patches on, or use int15.
• Second (red) galaxy: one player (RTX 5080, Intel CPU) crashed while arriving at Go-Go Archipelago for the first
  time (the loader read through a bad pointer). We could not reproduce it on our PC. A save from just before the red
  galaxy would help.
• One crash during startup was seen once in testing and could not be reproduced. If the console shows a crash, please
  send it; it now names the code path.
• The software ray tracing changes behaviour for other games too (for example Demon's Souls); please report any new
  problem there.
• Microsoft Defender may flag launcher.exe (Trojan:Win32/Bearfoos.A!ml, a machine-learning verdict on the
  unsigned launcher). It is built from this repository's source by the GitHub workflow.

Testers wanted

Please report (GPU, CPU, driver, what you did, the console text) especially from:

• RTX 50 series cards (first launch after the update, ray tracing shaders, hang watchdog);
• AMD Radeon cards (device lost with the patches off, water in Go-Go Archipelago);
• Intel CPUs (with "Intel CPU compatibility" on and off; the red galaxy's first arrival);
• 8-12 GB graphics cards (video memory full messages, stutter, frame rate with the defaults).

Switches for testing: KYTY_RT_SOFTWARE=0 KYTY_RT_STUB=1 keeps the lighting but lets every ray miss (no ray-traced
shadows); KYTY_RT_SOFTWARE=0 alone skips those shaders again (only with the patches on). The bundled preset's
KYTY_SRT_VARIANT_READS=1 is required. See U59-README.md in the download for the full list of changes and switches.

KytyPS5 U59 int17 pre-release — faster heavy levels, performance-mode occlusion, many fixes

Pre-release. This is int16.1 plus the newest speed work and fixes (draw batching in the clock tower and Ape
Escape, GPU-built indirect mesh draws, background ray tracing pipelines, and ports from chenxiao07's fork). If
anything works worse than in int16.1, please report it (with your GPU and CPU model) and use int16.1.

What's new in int17

• Galaxy map fixed: the red trails, pillars and blocky red areas around the volcano (the hidden world) and the
  red band at the bottom of the screen are gone. The game fast-clears one of its 64-bit render targets every frame and
  Kyty decoded that clear with a 32-bit decoder, so the clear was dropped and old frames piled up. The same fix covers
  R16F, RG16F, R16, RG16, R8, RG8 and B10G11R11 register clears (KYTY_CLEAR_REGISTER_WIDE=0 restores the old
  behaviour). This bug was also in upstream Kyty.

• Performance mode by default: occlusion queries off. The game asks the GPU, thousands of times per frame,
  whether objects are hidden behind others. Answering those queries accurately costs about half the frame rate in
  the heaviest levels. int17 now answers "visible" to all of them by default (KYTY_GPU_OCCLUSION=0, as other
  Kyty forks do), which roughly doubles the frame rate in the snow level, the clock tower and Ape Escape.
  On the test PC (RTX 3090, Ryzen 9 7950X3D; a test build without the release build's profile optimization, which
  adds about 15%): snow 45 fps, clock tower 35, Ape Escape 38, Sky Garden 31, volcano 59.
  Possible side effects: the game may draw objects or effects that it would normally hide behind walls (rarely
  visible, and it can cost GPU time where a lot is hidden).

• Accurate occlusion toggle: Global settings > Graphics & display > "Accurate occlusion queries (slower; fixes
  rare visibility issues)" turns the accurate mode back on (it passes --gpu-occlusion on, which overrides the
  preset). Use it if you see objects through walls or other visibility problems, and please report where. The
  accurate mode is also faster than in int16: render passes are no longer split at every query, draws keep running
  in batches across the game's occlusion proxies, and it includes the volcano lava fix below.
  Same PC and build, accurate mode: snow 25 fps, clock tower 22, Ape Escape 19 (the int16 pre-release built the same
  way: about 17.5, 13 and 13).

• Faster indirect draws in both modes: indirect draws are prepared by the worker threads and continue draw
  batches, and indirect mesh draws are built on the GPU instead of waiting for the GPU to hand their arguments back.

• Volcano lava fixed (accurate occlusion mode): with the camera tilted down in the volcano level, the lava lake
  disappeared (flat orange fog). The emulator paired the game's occlusion queries by a wrong address rule; it now pairs them by address, which
  also keeps the occlusion gate on in Sky Garden.

• Hang watchdog no longer kills the game: on RTX 50 cards (where the watchdog is on by default) a 5 s stall could
  end the emulator with "Unhandled host exception ... RtlVirtualUnwind". The watchdog's stack dump is now guarded.

• Ray tracing shaders build in the background: the large software ray tracing pipelines no longer freeze the game
  on first use; their effect starts a moment later instead. If a build fails, ray tracing turns itself off for the
  session instead of stopping the emulator.

• Out of video memory no longer ends the emulator: an image that does not fit goes to system memory (slower, the
  console prints "Vulkan: video memory full ..."), and a buffer that does not fit is retried.

• Galaxy map / file fix: Astro Bot file ids are assigned one per file; six pairs of game files shared an id and one
  of each pair could be read in place of the other (one of them a galaxy-map icon).

• Audio: sounds the game sends to the DualSense speaker now play at 30% on the PC speakers (they were up to 19 dB
  louder than music and effects). Four new volume sliders (master, game sound, music, controller speaker on PC) are in
  the launcher's global Audio settings. AudioPropagation (3D sound) is implemented as an empty scene.

• 3D object audio now plays: a large part of Astro Bot's sound goes to the PS5's 3D audio object ports (the
  ambisonics scene, a stereo pair and positioned per-sound effects), which Kyty used to drop. They are now mixed
  into the game sound. Launcher: global Audio settings > "3D audio objects (game's positional sounds)" switches
  them off (--audio-objects off, KYTY_AUDIO_OBJECTS=0), and a "3D audio objects" volume slider
  (--audio-objects-volume, KYTY_AUDIO_OBJECTS_VOLUME, 0-200%) sets their level.

• Louder stereo output: on a stereo device Kyty now downmixes the game's 7.1 output itself (front at full level)
  instead of SDL, which played it about 13 dB quieter. KYTY_AUDIO_STEREO_DOWNMIX=0 restores SDL's downmix.

• Ported from chenxiao07's fork: protection kept on remapped memory views (fixes a "frozen model" class of bugs),
  small fibers on their own stacks (Demon's Souls job-thread crashes), indirect draws' start instance (Demon's Souls
  missing pillars and rock faces), faster guest memory search while streaming, a clean process exit after a fatal
  error, and faster release shader translation.

• "AMD CPU patch" renamed to "Intel CPU compatibility (AMD instruction patch)" in the launcher's Compatibility
  settings. It is the same setting (saved configurations and --amd-cpu keep working): on Intel CPUs it emulates the
  PS5's AMD-only instructions (EXTRQ/INSERTQ) and AMD's VRSQRTPS results. It is not needed on AMD CPUs. The console
  lines now read "AMD instruction patch for Intel CPUs: ...".

• Read-only shader buffers (AMD graphics cards): shaders that never write their storage buffers now declare them
  read-only, so AMD's driver can load them once per wave instead of per thread (from serbru20066666's fork, which
  measured a lighting shader in Astro's Playroom at 3.45 -> 2.76 ms on an RX 6800 XT). No change on NVIDIA. Not
  tested by us on AMD: if you see new graphics problems on an AMD card, try KYTY_READONLY_BUFFERS=0 and please
  report the result. The first launch rebuilds the shader cache.

New defaults and how to turn each off

Set these in the launcher's environment / preset (tools/u59-preset.json style, NAME=value). Use the off value if
you suspect a change, and please report what it fixed.

Switch	Default	Off / old behaviour	What it does
KYTY_GPU_OCCLUSION (preset)	0 (performance mode)	1, or the launcher's "Accurate occlusion queries" (--gpu-occlusion on)	Occlusion queries answered "visible" instead of counted on the GPU
KYTY_OCCLUSION_SPLIT	on	0	An occlusion query inside a render pass ends only the query, not the pass
KYTY_DRAW_RUN_QUIET_OPS	on	0	Occlusion dumps and CPU-timeline writes no longer end a draw batch
KYTY_DRAW_PREP_INDIRECT	on	0	Worker threads prepare single indirect draws
KYTY_DRAW_RUN_INDIRECT	on	0	Prepared indirect draws continue draw batches
KYTY_DRAW_RUN_QUIET_SYNTHETIC	on	0	In performance mode, the always-visible query results no longer end draw batches (about +1.5%)
KYTY_DEPTH_FEEDBACK_LAZY	on	0	Depth feedback handling only for draws that sample the depth image
KYTY_NATIVE_INDIRECT_MESH	on	0	Indirect mesh draws built on the GPU (no CPU readback wait)
KYTY_RT_PIPELINE_ASYNC	on	0	Ray tracing pipelines built in the background (KYTY_RT_PIPELINE_WAIT_MS, default 250)
KYTY_RT_SOFTWARE	auto (preset)	0	Software ray tracing (0 skips the RT shaders; only with the patches on)
KYTY_RT_FUNCTION	1	0	Ray tracing node test as one shared shader function (smaller shaders)
KYTY_IMAGE_SYSMEM_FALLBACK	on	0	Images that do not fit in video memory go to system memory
KYTY_REMAP_TRACKER_PROTECTION	on	0	Restore write tracking on remapped memory views
KYTY_FIBER_GUEST_STACK	small fibers	0 (all = every fiber)	Fibers under 256 KiB run on their own guest stacks
KYTY_START_INSTANCE_SGPR	on	0	Indirect draws pass their start instance to the shader
KYTY_READONLY_BUFFERS	on	0	Storage buffers a shader never writes are declared read-only (faster on AMD)
KYTY_APR_HASHED_IDS	off	1 restores the old hashed ids	One file id per file path
KYTY_FATAL_EXIT_PROCESS	off	1 restores the old exit path	Fatal errors end the process with TerminateProcess
KYTY_IR_VALIDATE	off in release	1 runs it	Shader IR validation
KYTY_AUDIO_DIAG	on	0	Audio port open/close lines in the console (256 lines max)
KYTY_AUDIO_PROPAGATION_RENDER	direct sound	silent	AudioPropagation sources play their direct sound
Audio volumes	master/main/music 100%, pad speaker on PC 30%	launcher sliders, or KYTY_AUDIO_MASTER_VOLUME, KYTY_AUDIO_MAIN_VOLUME, KYTY_AUDIO_MUSIC_VOLUME, KYTY_AUDIO_PAD_SPEAKER_ON_MAIN_VOLUME (percent, 0-200)
KYTY_AUDIO_OBJECTS	1	0, or untick "3D audio objects" (--audio-objects off)	AudioOut2 3D object ports mixed into the game sound
KYTY_AUDIO_STEREO_DOWNMIX	on	0	7.1 output downmixed to stereo by Kyty, not SDL

Optional (off by default): KYTY_WIDE_FILL_CLEAR=1, KYTY_GUEST_STORAGE_REPEAT=1, KYTY_READBACK_MERGE_GAP_KB=64,
and the test tool KYTY_VRAM_LIMIT_MB=<MiB> (imitates a
smaller GPU). The accurate-mode switches above (KYTY_OCCLUSION_SPLIT, KYTY_DRAW_RUN_QUIET_OPS) only matter with accurate
occlusion on.

From int16.1

• Level title text fixed: the level name shown when Astro Bot enters a level is drawn as in int14 again.

From int16

• **Astro Bot without the "non RT...

KytyPS5 U59 int15 — Crash fixes, upstream shader fixes and less shader stutter

This is the new main release. It includes everything from the int13 and int14 pre-releases. Keep using the Astro Bot
"non RT patch" game patches with this build; the pre-release int16 runs the game's full graphics without them.

What's new since int14

• Fewer crashes on fast GPUs: on fast graphics cards Astro Bot raises its resolution, reuses memory between
  images, and the emulator stopped with "unsupported sampled depth image". These images are now read correctly.
• No crash when leaving extra levels: leaving an extra level in Sky Garden could stop the emulator on an image
  clear it rejected. Such a clear is now skipped, and the console prints one TextureCache: ClearImage skipped ...
  line with the details. Please send that line if you see it.
• Upstream fixes: about 20 shader translation fixes, a fix for Demon's Souls stopping on an unknown image format,
  thread priorities that keep busy game threads from starving the graphics thread, and trophy notifications.

From int14

• Shader precompile: the emulator records the shaders it has built, and after an emulator update it rebuilds the
  ones you have already seen in the background, so areas you have visited do not stutter again. The first time you
  play an area works as before.
• Faster first pipelines: a quick version of each new pipeline is used until the optimized one is ready. In Sky
  Garden with an empty cache, the game waited 0.3 s in total for pipelines instead of 1.4 s.
• Automatic GPU choice: on auto, the emulator now prefers your dedicated graphics card, and the console shows the
  chosen GPU ("Kyty GPU: ..."). If a game crashes for you on auto, select your graphics card in the launcher.
• Fewer loading crashes, Vulkan validation no longer stops the game, Demon's Souls character creation works, and
  the adaptive trigger fixes.

Checked

• All automated tests pass.
• Astro Bot on the test PC (RTX 3090, Ryzen 9 7950X3D): Sky Garden 34.8 fps (int14: 34.1) and the snow level
  21.2 fps (int14: 21.7), within run-to-run noise.

Installing

1. Download KytyPS5-U59-Windows-x64.zip and extract it to a new folder.
2. Open launcher.exe. Your existing game list and settings are picked up automatically.
3. Game patches go in a _Patches folder next to the launcher; saves are kept per folder in _SaveData.

If a game crashes, please send the console text with your GPU and CPU model.

Known issues

• Microsoft Defender may flag launcher.exe (Trojan:Win32/Bearfoos.A!ml, a machine-learning verdict on the
  unsigned launcher). It is built from this repository's source by the GitHub workflow.
• A game patch made for another game version can crash the game. Use patch files that match your version exactly.

See U59-README.md in the download for the full list of changes and switches.