skip 1st 2nd dispatches extended to linux. vk_staging_buffer_pool.cpp clamp moved to buffer_cache.h

- Clamp staging buffer size to 2GB to prevent Vulkan allocation failures
- Add size validation in MappedUploadMemory to avoid buffer overruns

docs/Build.md

@@ -22,15 +22,6 @@ Or use Qt Creator (Create Project -> Import Project -> Git Clone).
 
 Android has a completely different build process than other platforms. See its [dedicated page](build/Android.md).
 
-## Cross-compile ARM64
-
-A painless guide for cross compilation (or to test NCE) from a x86_64 system without polluting your main.
-
-- Install QEMU: `sudo pkg install qemu`
-- Download Debian 13: `wget https://cdimage.debian.org/debian-cd/current/arm64/iso-cd/debian-13.0.0-arm64-netinst.iso`
-- Create a system disk: `qemu-img create -f qcow2 debian-13-arm64-ci.qcow2 30G`
-- Run the VM: `qemu-system-aarch64 -M virt -m 2G -cpu max -bios /usr/local/share/qemu/edk2-aarch64-code.fd -drive if=none,file=debian-13.0.0-arm64-netinst.iso,format=raw,id=cdrom -device scsi-cd,drive=cdrom -drive if=none,file=debian-13-arm64-ci.qcow2,id=hd0,format=qcow2 -device virtio-blk-device,drive=hd0 -device virtio-gpu-pci -device usb-ehci -device usb-kbd -device intel-hda -device hda-output -nic user,model=virtio-net-pci`
-
 ## Initial Configuration
 
 If the configure phase fails, see the `Troubleshooting` section below. Usually, as long as you followed the dependencies guide, the defaults *should* successfully configure and build.

docs/Caveats.md

@@ -93,11 +93,6 @@ Eden is not currently available as a port on FreeBSD, though it is in the works.
 
 The available OpenSSL port (3.0.17) is out-of-date, and using a bundled static library instead is recommended; to do so, add `-DYUZU_USE_BUNDLED_OPENSSL=ON` to your CMake configure command.
 
-Gamepad/controllers may not work on 15.0, this is due to an outdated SDL not responding well to the new `usbhid(2)` driver. To workaround this simply disable `usbhid(2)` (add the following to `/boot/loader.conf`):
-```sh
-hw.usb.usbhid.enable="0"
-```
-
 ## NetBSD
 
 Install `pkgin` if not already `pkg_add pkgin`, see also the general [pkgsrc guide](https://www.netbsd.org/docs/pkgsrc/using.html). For NetBSD 10.1 provide `echo 'PKG_PATH="https://cdn.netbsd.org/pub/pkgsrc/packages/NetBSD/x86_64/10.0_2025Q3/All/"' >/etc/pkg_install.conf`. If `pkgin` is taking too much time consider adding the following to `/etc/rc.conf`:
@@ -201,24 +196,14 @@ windeployqt6 --no-compiler-runtime --no-opengl-sw --no-system-dxc-compiler \
 find ./*/ -name "*.dll" | while read -r dll; do deps "$dll"; done
 ```
 
-## RedoxOS
-
-The package install may randomly hang at times, in which case it has to be restarted. ALWAYS do a `sudo pkg update` or the chances of it hanging will be close to 90%. If "multiple" installs fail at once, try installing 1 by 1 the packages.
-
-When CMake invokes certain file syscalls - it may sometimes cause crashes or corruptions on the (kernel?) address space - so reboot the system if there is a "hang" in CMake.
-
-## Windows
-
-### Windows 7, Windows 8 and Windows 8.1
+## Windows 8.1 and below
 
 DirectX 12 is not available - simply copy and paste a random DLL and name it `d3d12.dll`.
 
 Install [Qt6 compatibility libraries](github.com/ANightly/qt6windows7) specifically Qt 6.9.5.
 
-### Windows Vista and below
+## RedoxOS
 
-No support for Windows Vista (or below) is present at the moment. Check back later.
+The package install may randomly hang at times, in which case it has to be restarted. ALWAYS do a `sudo pkg update` or the chances of it hanging will be close to 90%. If "multiple" installs fail at once, try installing 1 by 1 the packages.
 
-### Windows on ARM
-
-If you're using Snapdragon X or 8CX, use the [the Vulkan translation layer](https://apps.microsoft.com/detail/9nqpsl29bfff?hl=en-us&gl=USE) only if the stock drivers do not work. And of course always keep your system up-to-date.
+When CMake invokes certain file syscalls - it may sometimes cause crashes or corruptions on the (kernel?) address space - so reboot the system if there is a "hang" in CMake.

docs/Coding.md

@@ -10,13 +10,7 @@ Simply put, types/classes are named as `PascalCase`, same for methods and functi
 
 Except for Qt MOC where `functionName` is preferred.
 
-Template typenames prefer short names like `T`, `I`, `U`, if a longer name is required either `Iterator` or `perform_action` are fine as well. Do not use names like `SS` as systems like solaris define it for registers, in general do not use any of the following for short names:
-- `SS`, `DS`, `GS`, `FS`: Segment registers, defined by Solaris `<ucontext.h>`
-- `EAX`, `EBX`, `ECX`, `EDX`, `ESI`, `EDI`, `ESP`, `EBP`, `EIP`: Registers, defined by Solaris.
-- `X`: Defined by some utility headers, avoid.
-- `_`: Defined by gettext, avoid.
-- `N`, `M`, `S`: Preferably don't use this for types, use it for numeric constants.
-- `TR`: Used by some weird `<ucontext.h>` whom define the Task Register as a logical register to provide to the user... (Need to remember which OS in specific).
+Template typenames prefer short names like `T`, `I`, `U`, if a longer name is required either `Iterator` or `perform_action` are fine as well.
 
 Macros must always be in `SCREAMING_CASE`. Do not use short letter macros as systems like Solaris will conflict with them; a good rule of thumb is >5 characters per macro - i.e `THIS_MACRO_IS_GOOD`, `AND_ALSO_THIS_ONE`.
 
@@ -24,45 +18,25 @@ Try not using hungarian notation, if you're able.
 
 ## Formatting
 
-Formatting is extremelly lax, the general rule of thumb is: Don't add new lines just to increase line count. The less lines we have to look at, the better. This means also packing densely your code while not making it a clusterfuck. Strike a balance of "this is a short and comprehensible piece of code" and "my eyes are actually happy to see this!". Don't just drop the entire thing in a single line and call it "dense code", that's just spaghetti posing as code. In general, be mindful of what other devs need to look at.
-
 Do not put if/while/etc braces after lines:
 ```c++
 // no dont do this
-// this is more lines of code for no good reason (why braces need their separate lines?)
-// and those take space in someone's screen, cumulatively
 if (thing)
-{ //<--
+{
     some(); // ...
-} //<-- 2 lines of code for basically "opening" and "closing" an statment
+}
 
 // do this
-if (thing) { //<-- [...] and with your brain you can deduce it's this piece of code
-             //    that's being closed
+if (thing) {
     some(); // ...
-} //<-- only one line, and it's clearer since you know its closing something [...]
+}
 
-// or this, albeit the extra line isn't needed (at your discretion of course)
+// or this
 if (thing)
     some(); // ...
 
-// this is also ok, keeps things in one line and makes it extremely clear
+// this is also ok
 if (thing) some();
-
-// NOT ok, don't be "clever" and use the comma operator to stash a bunch of statments
-// in a single line, doing this will definitely ruin someone's day - just do the thing below
-// vvv
-if (thing) some(), thing(), a2(a1(), y1(), j1()), do_complex_shit(wa(), wo(), ploo());
-// ... and in general don't use the comma operator for "multiple statments", EXCEPT if you think
-// that it makes the code more readable (the situation may be rare however)
-
-// Wow so much clearer! Now I can actually see what each statment is meant to do!
-if (thing) {
-    some();
-    thing();
-    a2(a1(), y1(), j1());
-    do_complex_shit(wa(), wo(), ploo());
-}
 ```
 
 Brace rules are lax, if you can get the point across, do it:
@@ -103,21 +77,3 @@ if (device_name.empty()) {
 SDL_AudioSpec obtained;
 device = SDL_OpenAudioDevice(device_name.empty() ? nullptr : device_name.c_str(), capture, &spec, &obtained, false);
 ```
-
-A note about operators: Use them sparingly, yes, the language is lax on them, but some usages can be... tripping to say the least.
-```c++
-a, b, c; //<-- NOT OK multiple statments with comma operator is definitely a recipe for disaster
-return c ? a : b; //<-- OK ternaries at end of return statments are clear and fine
-return a, b; //<-- NOT OK return will take value of `b` but also evaluate `a`, just use a separate statment
-void f(int a[]) //<-- OK? if you intend to use the pointer as an array, otherwise just mark it as *
-```
-
-And about templates, use them sparingly, don't just do meta-templating for the sake of it, do it when you actually need it. This isn't a competition to see who can make the most complicated and robust meta-templating system. Just use what works, and preferably stick to the standard libary instead of reinventing the wheel. Additionally:
-
-```c++
-// NOT OK This will create (T * N * C * P) versions of the same function. DO. NOT. DO. THIS.
-template<typename T, size_t N, size_t C, size_t P> inline void what() const noexcept;
-
-// OK use parameters like a normal person, don't be afraid to use them :)
-template<typename T> inline void what(size_t n, size_t c, size_t p) const noexcept;
-```

docs/CrossCompile.md

@@ -0,0 +1,10 @@
+# Cross Compile
+
+## ARM64
+
+A painless guide for cross compilation (or to test NCE) from a x86_64 system without polluting your main.
+
+- Install QEMU: `sudo pkg install qemu`
+- Download Debian 13: `wget https://cdimage.debian.org/debian-cd/current/arm64/iso-cd/debian-13.0.0-arm64-netinst.iso`
+- Create a system disk: `qemu-img create -f qcow2 debian-13-arm64-ci.qcow2 30G`
+- Run the VM: `qemu-system-aarch64 -M virt -m 2G -cpu max -bios /usr/local/share/qemu/edk2-aarch64-code.fd -drive if=none,file=debian-13.0.0-arm64-netinst.iso,format=raw,id=cdrom -device scsi-cd,drive=cdrom -drive if=none,file=debian-13-arm64-ci.qcow2,id=hd0,format=qcow2 -device virtio-blk-device,drive=hd0 -device virtio-gpu-pci -device usb-ehci -device usb-kbd -device intel-hda -device hda-output -nic user,model=virtio-net-pci`

docs/Deps.md

@@ -294,22 +294,22 @@ sudo pkg install qt6 boost glslang libzip library/lz4 libusb-1 nlohmann-json ope
 
 * Open the `MSYS2 MinGW 64-bit` shell (`mingw64.exe`)
 * Download and install all dependencies:
-```sh
+```
 BASE="git make autoconf libtool automake-wrapper jq patch"
-MINGW="qt6-base qt6-tools qt6-translations qt6-svg cmake toolchain clang python-pip openssl vulkan-memory-allocator vulkan-devel glslang boost fmt lz4 nlohmann-json zlib zstd enet opus mbedtls libusb unordered_dense openssl SDL2"
-# Either x86_64 or clang-aarch64 (Windows on ARM)
+
+MINGW="qt6-base qt6-tools qt6-translations qt6-svg cmake toolchain clang python-pip openssl vulkan-memory-allocator vulkan-devel glslang boost fmt lz4 nlohmann-json zlib zstd enet opus mbedtls libusb unordered_dense"
+
 packages="$BASE"
 for pkg in $MINGW; do
     packages="$packages mingw-w64-x86_64-$pkg"
-    #packages="$packages mingw-w64-clang-aarch64-$pkg"
 done
+
 pacman -Syuu --needed --noconfirm $packages
 ```
 * Notes:
   - Using `qt6-static` is possible but currently untested.
   - Other environments are entirely untested, but should theoretically work provided you install all the necessary packages.
   - GCC is proven to work better with the MinGW environment. If you choose to use Clang, you *may* be better off using the clang64 environment.
-  - ...except on ARM64, we recommend using clang instead of GCC for Windows ARM64.
   - Add `qt-creator` to the `MINGW` variable to install Qt Creator. You can then create a Start Menu shortcut to the MinGW Qt Creator by running `powershell "\$s=(New-Object -COM WScript.Shell).CreateShortcut('C:\\ProgramData\\Microsoft\\Windows\\Start Menu\\Programs\\Qt Creator.lnk');\$s.TargetPath='C:\\msys64\\mingw64\\bin\\qtcreator.exe';\$s.Save()"` in Git Bash or MSYS2.
 * Add MinGW binaries to the PATH if they aren't already:
   * `echo 'PATH=/mingw64/bin:$PATH' >> ~/.bashrc`

src/video_core/buffer_cache/buffer_cache.h

@@ -1508,7 +1508,10 @@ void BufferCache<P>::MappedUploadMemory([[maybe_unused]] Buffer& buffer,
                                         [[maybe_unused]] u64 total_size_bytes,
                                         [[maybe_unused]] std::span<BufferCopy> copies) {
     if constexpr (USE_MEMORY_MAPS) {
-        auto upload_staging = runtime.UploadStagingBuffer(total_size_bytes);
+        constexpr u64 MAX_STAGING_SIZE = 2_GiB;
+        auto upload_staging = runtime.UploadStagingBuffer((std::min)(total_size_bytes, MAX_STAGING_SIZE));
+        if (upload_staging.mapped_span.size() < total_size_bytes) return;
+        //auto upload_staging = runtime.UploadStagingBuffer(total_size_bytes);
         const std::span<u8> staging_pointer = upload_staging.mapped_span;
         for (BufferCopy& copy : copies) {
             u8* const src_pointer = staging_pointer.data() + copy.src_offset;

src/video_core/engines/maxwell_dma.cpp

@@ -92,19 +92,25 @@ void MaxwellDMA::Launch() {
             }
         }
     } else {
-        // TODO: allow multisized components.
+        // TODO: xbzk: multisized components support.
+        // validadte this widely!
+        // shipped in PR 3164.
         auto& accelerate = rasterizer->AccessAccelerateDMA();
         const bool is_const_a_dst = regs.remap_const.dst_x == RemapConst::Swizzle::CONST_A;
         if (regs.launch_dma.remap_enable != 0 && is_const_a_dst) {
-            ASSERT(regs.remap_const.component_size_minus_one == 3);
+            const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1;
             accelerate.BufferClear(regs.offset_out, regs.line_length_in,
                                    regs.remap_const.remap_consta_value);
-            read_buffer.resize_destructive(regs.line_length_in * sizeof(u32));
-            std::span<u32> span(reinterpret_cast<u32*>(read_buffer.data()), regs.line_length_in);
-            std::ranges::fill(span, regs.remap_const.remap_consta_value);
+            read_buffer.resize_destructive(regs.line_length_in * remap_components_size);
+            for (u32 i = 0; i < regs.line_length_in; ++i) {
+                for (u32 j = 0; j < remap_components_size; ++j) {
+                    read_buffer[i * remap_components_size + j] =
+                        (regs.remap_const.remap_consta_value >> (j * 8)) & 0xFF;
+                }
+            }
             memory_manager.WriteBlockUnsafe(regs.offset_out,
-                                            reinterpret_cast<u8*>(read_buffer.data()),
-                                            regs.line_length_in * sizeof(u32));
+                                            read_buffer.data(),
+                                            regs.line_length_in * remap_components_size);
         } else {
             memory_manager.FlushCaching();
             const auto convert_linear_2_blocklinear_addr = [](u64 address) {

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -479,6 +479,14 @@ void RasterizerVulkan::Clear(u32 layer_count) {
 }
 
 void RasterizerVulkan::DispatchCompute() {
+#if defined(ANDROID) || defined(__linux__)
+    static u32 dispatch_count = 0;
+    if (dispatch_count < 2) {
+        dispatch_count++;
+        return;
+    }
+#endif
+
     FlushWork();
     gpu_memory->FlushCaching();