Watertoon
LightNews LightNews
watertoon.bsky.social
Watertoon
@watertoon.bsky.social
110 followers 26 following 110 posts
I reverse engineer game engines among other software memery.
Posts Replies Media Videos
watertoon.bsky.social
Minecraft McNugget. (The frontmost McNugget was actually 2 McNuggets crafted together)
The $2 lizard on a McDonalds Minecraft themed box. The $2 lizard facing an opened Minecraft 10 piece McNugget box with a small pile of McNuggets inside.
April 3, 2025 at 11:49 AM
watertoon.bsky.social
Switch 2 backwards compatibility is hard, but not intractable. The api boundaries between application and sdk are well-defined despite both running in the same userland process. So Nintendo could replace the entire NVN api transparently and then use extra Switch 2 power for data conversion/caching.
April 2, 2025 at 4:44 PM
watertoon.bsky.social
Great to see PhysX 5.6 open-source the GPU CUDA code. I still have reservations certain features like deformable (soft) bodies currently require a NVidia GPU, but PhysX 5 is close to my top choice for a game physics api I am planning (the internal struggle to not just write my own physics engine).
March 27, 2025 at 7:51 AM
watertoon.bsky.social
Real unimpressive tech demo #1.
Screenshot of a testing application to load a PNG and draw it through a shader. The image displayed is of Alvin and The Chipmunks: Chipwrecked (Features 30 Hit Songs!) for the Nintendo Wii, but the debug camera is positioned far away.
January 6, 2025 at 8:14 AM
watertoon.bsky.social
Happy new year. This year I will use my custom in-progress game engine to start producing real unimpressive tech demos that can be seen and heard with graphics I poorly art and sounds I screech alongside unapologetic technical babble I'm attracted to as an overbearing dim-witted moth. In good fun.
A partial list of foundational awn sub libraries reading: Awn is an in-progress collection of real-time focused software libraries. These libraries provide a number of services that manage the underlying CPU architecture, platform, and graphics api for the user. >awn::mem - Provides virtual and heap-based CPU and GPU memory management. >awn::sys - Provides thread and system environment management (managing Win32 or [redacted]). >awn::async - Provides asynchronous task scheduling and dependency graph job scheduling. >awn::res - Provides synchronous and asynchronous file resource management. >awn::gfx - Provides a high-level graphics api (managing Vulkan or [redacted]), designed to interoperate with [redacted] formats. >awn::audio - Provides an audio server api for audio in, out, [redacted]. >awn::render - Provides display and layer management (managing DirectX Graphics Infrastructure, DirectComposition, and DirectX12, or [redacted]). >awn::para - Provides serializable parameter class autogen, databases, blackboards, [redacted], variable framerates, etc. >awn::hid - Provides host input device management for keyboards and mice. >awn::frm - Provides a bootstrap for libraries and implements a module-based application framework.
January 1, 2025 at 10:19 AM
watertoon.bsky.social
The new 1.4 NV Vulkan driver added vkGetImageViewHandle64NVX. Currently combined image handles are implied as 2 packed numbers in a 32-bit uint, 20:12, image_id:sampler_id. A (new?) flag exists that increases the image id to a full 32-bits, raising the max image cap from 1,048,576 to 4,294,967,296.
Partial screenshot of Ghidra decompiler output for combined image sampler handle behavior for vkGetImageViewHandle64NVX in nvogl64.dll from the Windows 553.45 NVidia Vulkan Beta driver, showcasing new driver behavior for increasing the image limit for combined image samplers from an implied 20-bits to an implied 32-bits.
December 3, 2024 at 2:46 AM
watertoon.bsky.social
I investigated VK_NVX_image_view_handle (rev2) since I wanted to use NV_bindless_texture (I must macro descriptor indexing vs NV_bindless_texture for Vulkan vs NVN already) and concluded it returns the same data as using vkGetDescriptorEXT, it seems I could also get the raw sampler handle this way.
A not-fully labelled screenshot of Ghidra decompiler output for the implementation of vkGetImageViewHandleNVX in the 64-bit Windows 553.40 NVidia driver. Showcases the handling for retrieving a 32-bit handle to a Vulkan combined image, sampled image, and storage image. A not-fully labelled screenshot of Ghidra decompiler output for the implementation of vkGetDescriptorEXT in the 64-bit Windows 553.40 NVidia driver. Showcases the handling for retrieving a 32-bit handle to a Vulkan sampler, combined image, sampled image, storage image, and input attachment.
November 30, 2024 at 6:21 AM
watertoon.bsky.social
Made a primitive drawer for the Gamecube in a day to illustrate to someone how ridiculously long it takes to support modern graphics api's properly. That someone is me. I saw this and it fried my brain for the past 8 hours relieving all my PC graphics programming trauma I am not done with just yet.
A screenshot of Google Drive for a Gamecube executable dayed March 2023 featuring a testing port of my game engine library, awn. It implements a small drawing routine using the GX api.
November 23, 2024 at 4:07 PM
watertoon.bsky.social
NintendoWare g3d2 encompasses bfres v9.1.1, v10.0.0, and v10.2.0 so far (bfres being a 3d model archive file format). Only ModuleSystem titles feature g3d2 though, every other g3d title I've seen still only uses g3d capped at bfres v9.1.0.
A screenshot of strings from the executable of The Legend Of Zelda: Tears of the Kingdom version 1.2.1, a title on the unofficially named "ModuleSystem" engine. There is a selection highlighting "SDK MW+Nintendo+NintendoWare_G3d2-15_3_1-Release", indicating the executable contains this library. A screenshot of strings from the executable of Super Mario Party Jamboree, a title on the recent Bezel Engine 2. A selection highlights "SDK MW+Nintendo+NintendoWare_G3d-17_5_4-Release", indicating the executable contains this library. Note; despite the later SDK version from The Legend of Zelda: Tears of the Kingdom, the library is g3d and not g3d2.
November 13, 2024 at 9:16 PM
watertoon.bsky.social
By at least Switch Sports, Nintendo EPD's main graphics library, agl, moved to using "pipeline" state objects, likely during the addition of Vulkan as a backend (outside NVN). In NSO Playtest, there no longer seems to be pipeline objects. This could possibly be a result of recent Vulkan extensions.
November 11, 2024 at 12:26 AM
watertoon.bsky.social
From between nnsdk 16.2.0 to 18.3.4. nn::os::ReaderWriterLockType was rebalanced and further optimized internally. I believe effectively up to 1023 read and write acquisitions can now safely happen, up from 255. Games use RWLocks plenty, so it's good work to see.
A snippet of the 16.2.0 nnSdk nso from Mario Wonder and the 18.3.4 nnSdk nso from the NSO Playtest. Showcasing notable changes to nn::os::AcquireReadLock, such as a new writer read lock counter, now separate from the reader count found in the atomic state, and the previously branchless read acquisition introducing a branch when there are write waiters, likely saving a small bit of work overall. An info dump on the changes, reading: nn::os::ReaderWriterLockType: >64-bit atomic state was redone from bits 32:15:1:8:8, write thread handle, reader count, is write locked, read waiter count, write waiter count to bits 32:10:10:10:1:1, write thread handle, reader count, read waiter count, write waiter count, is write locked, reserved >this rebalances max reader count from 32767 to 1023, while raising read waiter and write waiter thread counts from 255 to 1023 >I believe recursive read acquisitions would be a bug, as a read acquisition blocks if a writer is waiting for readers, a recursive reader could block on a writer waiter never releasing the read lock for the writer >effectively I believe 1023 threads can now safely simultaneously use the lock, up from 255 >acquiring read lock as writer uses a separate unique int counter now instead of updating atomic state >in general underlying logic was optimized and simplified in some regards
November 7, 2024 at 5:12 PM
watertoon.bsky.social
Me when the new g3d2 version in the NSO Playtest software added meshlet support on the console without mesh shaders.
Screenshot of decompiled code from the NSO Playtest software for the unofficially named function nn::g3d2::ResBfresMeshlet::Setup where an array of graphics buffers are being initialized and set with the debug label "g3d_meshlet".
November 5, 2024 at 1:28 PM
watertoon.bsky.social
The nnSdk in Switch games has a math library under nn::util, while Nintendo EPD has a math lib under sead. In EPD games the matrix types can interact. It seems a sead 3x4 row-major matrix is equivalent to an nn::util 4x3 column-major matrix, I'm just going to pretend sead is column major 4x3 lol.
A visual for 3 matrix structs. sead::Matrix34f, a 3 row 4 column row-major matrix type. nn::util::MatrixRowMajor4x3f, a 4 row 3 column row-major matrix type. And nn::util::MatrixColumnMajor4x3f, a 4 row 3 column column major type.
November 2, 2024 at 3:21 PM
watertoon.bsky.social
Unironically how I went about learning C++ and Operating Systems was by writing an entire cooperative thread scheduler on Windows to emulate the Nintendo Switch's cooperative threading model. Which I then scrapped in my lib recently once managing single vs cross process threads became too annoying.
The public interface for a custom Win32 fiber based cooperative thread scheduler. A snippet of the main Win32 fiber scheduling function for a custom cooperative thread scheduler.
October 20, 2024 at 7:04 PM