Wii U support (#1097)

* Wii U support

* [WiiU] Combined Dockerfile

* [WiiU] Combined Dockerfile

* [WiiU] Combined Dockerfile

* Add Jenkins support

* wiiu: fix scissor clamp

* wiiu: improve button remapping

* wiiu: fix scaling issues

* Update Dockerfile after merge

* Pull assets before build

* Only stop container once

* Adjust logging sinks

* wiiu: Change button mapping to match PC version

* wiiu: Implement controller changes

* wiiu: Update BUILDING.md

Co-authored-by: qurious-pixel <62252937+qurious-pixel@users.noreply.github.com>
Co-authored-by: David Chavez <david@dcvz.io>

BUILDING.md

@@ -167,6 +167,27 @@ cmake --build build-switch --target soh_nro
 # To develop the project open the repository in VSCode (or your preferred editor)
 ```
 
+## Wii U
+1. Requires that your build machine is setup with the tools necessary for your platform above
+2. Requires that you have the Wii U build tools installed 
+3. Clone the Ship of Harkinian repository
+4. Place one or more [compatible](#compatible-roms) roms in the `OTRExporter` directory with namings of your choice
+
+```bash
+cd Shipwright
+# Setup cmake project for your host machine
+cmake -H. -Bbuild-cmake -GNinja
+# Extract assets & generate OTR (run this anytime you need to regenerate OTR)
+cmake --build build-cmake --target ExtractAssets
+# Setup cmake project for building for Wii U
+cmake -H. -Bbuild-wiiu -GNinja -DCMAKE_TOOLCHAIN_FILE=/opt/devkitpro/cmake/WiiU.cmake # -DCMAKE_BUILD_TYPE:STRING=Release (if you're packaging)
+# Build project and generate rpx
+cmake --build build-wiiu --target soh
+
+# Now you can run the executable in ./build-wiiu/soh/soh.rpx
+# To develop the project open the repository in VSCode (or your preferred editor)
+```
+
 # Compatible Roms
 ```
 OOT_PAL_GC      checksum 0x09465AC3

CMakeLists.txt

@@ -122,7 +122,7 @@ add_subdirectory(ZAPDTR/ZAPD ${CMAKE_BINARY_DIR}/ZAPD)
 add_subdirectory(ZAPDTR/ZAPDUtils ${CMAKE_BINARY_DIR}/ZAPDUtils)
 add_subdirectory(OTRExporter)
 add_subdirectory(soh)
-if(NOT CMAKE_SYSTEM_NAME MATCHES "Darwin|NintendoSwitch")
+if(NOT CMAKE_SYSTEM_NAME MATCHES "Darwin|NintendoSwitch|CafeOS")
 add_subdirectory(OTRGui)
 endif()
 
@@ -179,13 +179,13 @@ add_custom_target(CreateOSXIcons
 add_dependencies(soh CreateOSXIcons)
 endif()
 
-if(CMAKE_SYSTEM_NAME MATCHES "Windows|NintendoSwitch")
+if(CMAKE_SYSTEM_NAME MATCHES "Windows|NintendoSwitch|CafeOS")
 INSTALL(FILES ${CMAKE_SOURCE_DIR}/README.md DESTINATION . COMPONENT ship RENAME readme.txt )
 endif()
 
 if(CMAKE_SYSTEM_NAME MATCHES "Linux")
 	set(CPACK_GENERATOR "External")
-elseif(CMAKE_SYSTEM_NAME MATCHES "Windows|NintendoSwitch")
+elseif(CMAKE_SYSTEM_NAME MATCHES "Windows|NintendoSwitch|CafeOS")
     set(CPACK_GENERATOR "ZIP")
 elseif(CMAKE_SYSTEM_NAME MATCHES "Darwin")
     set(CPACK_GENERATOR "Bundle")

Dockerfile

@@ -56,12 +56,13 @@ RUN \
 	echo "deb [signed-by=/usr/local/share/keyring/devkitpro-pub.gpg] https://apt.devkitpro.org stable main" > /etc/apt/sources.list.d/devkitpro.list && \
 	apt-get update -y && \
 	apt-get install -y devkitpro-pacman && \
-	yes | dkp-pacman -Syu switch-dev switch-portlibs --noconfirm
+	yes | dkp-pacman -Syu switch-dev switch-portlibs wiiu-dev wiiu-portlibs --noconfirm
 
 ENV DEVKITPRO=/opt/devkitpro
 ENV DEVKITARM=/opt/devkitpro/devkitARM
 ENV DEVKITPPC=/opt/devkitpro/devkitPPC
-ENV PATH=$PATH:/opt/devkitpro/portlibs/switch/bin/
+ENV PATH=$PATH:/opt/devkitpro/portlibs/switch/bin/:$DEVKITPPC/bin
+ENV WUT_ROOT=$DEVKITPRO/wut
 
 RUN mkdir /soh
 WORKDIR /soh

Jenkinsfile

@@ -185,6 +185,43 @@ pipeline {
                         }
                     }
                 }
+                stage ('Build Wii U') {
+                    agent {
+                        label "SoH-Linux-Builders"
+                    }
+                    steps {
+                        checkout([
+                            $class: 'GitSCM',
+                            branches: scm.branches,
+                            doGenerateSubmoduleConfigurations: scm.doGenerateSubmoduleConfigurations,
+                            extensions: scm.extensions,
+                            userRemoteConfigs: scm.userRemoteConfigs
+                        ])
+                        catchError(buildResult: 'FAILURE', stageResult: 'FAILURE') {
+                            unstash 'assets'
+                            sh '''
+                            if docker ps -aq --filter "name=sohwiiucont" | grep -q .; then docker rm -f sohwiiucont; fi
+                            docker build . -t sohwiiu
+                            docker run --name sohwiiucont -dit --rm -v $(pwd):/soh sohwiiu /bin/bash
+                            docker exec sohwiiucont scripts/wiiu/build.sh
+                            
+                            mv build-wiiu/soh/*.rpx soh.rpx
+                            mv README.md readme.txt
+                            
+                            7z a soh-wiiu.7z soh.rpx readme.txt
+                            
+                            '''
+                        }
+                        archiveArtifacts artifacts: 'soh-wiiu.7z', followSymlinks: false, onlyIfSuccessful: true
+                    }
+                    post {
+                        always {
+                            sh 'sudo docker container stop sohwiiucont'
+                            sh 'docker images --quiet --filter=dangling=true | xargs --no-run-if-empty docker rmi' // Clean dangling docker images
+                            step([$class: 'WsCleanup']) // Clean workspace
+                        }
+                    }
+                }
             }
         }
     }

StormLib/src/FileStream.cpp

@@ -34,7 +34,12 @@
 // Local functions - platform-specific functions
 
 #ifndef STORMLIB_WINDOWS
+
+#ifndef STORMLIB_WIIU // WIIU doesn't support thread_local
 static thread_local DWORD dwLastError = ERROR_SUCCESS;
+#else
+static DWORD dwLastError = ERROR_SUCCESS;
+#endif
 
 DWORD GetLastError()
 {

StormLib/src/SBaseCommon.cpp

@@ -2005,7 +2005,7 @@ void ConvertTMPQHeader(void *header, uint16_t version)
 	TMPQHeader * theHeader = (TMPQHeader *)header;
 
     // Swap header part version 1
-    if(version == MPQ_FORMAT_VERSION_1)
+    if(version >= MPQ_FORMAT_VERSION_1)
     {
 	    theHeader->dwID = SwapUInt32(theHeader->dwID);
 	    theHeader->dwHeaderSize = SwapUInt32(theHeader->dwHeaderSize);
@@ -2018,21 +2018,21 @@ void ConvertTMPQHeader(void *header, uint16_t version)
 	    theHeader->dwBlockTableSize = SwapUInt32(theHeader->dwBlockTableSize);
     }
 
-	if(version == MPQ_FORMAT_VERSION_2)
+	if(version >= MPQ_FORMAT_VERSION_2)
 	{
 		theHeader->HiBlockTablePos64 = SwapUInt64(theHeader->HiBlockTablePos64);
         theHeader->wHashTablePosHi = SwapUInt16(theHeader->wHashTablePosHi);
 		theHeader->wBlockTablePosHi = SwapUInt16(theHeader->wBlockTablePosHi);
     }
 
-    if(version == MPQ_FORMAT_VERSION_3)
+    if(version >= MPQ_FORMAT_VERSION_3)
     {
         theHeader->ArchiveSize64 = SwapUInt64(theHeader->ArchiveSize64);
         theHeader->BetTablePos64 = SwapUInt64(theHeader->BetTablePos64);
         theHeader->HetTablePos64 = SwapUInt64(theHeader->HetTablePos64);
     }
 
-    if(version == MPQ_FORMAT_VERSION_4)
+    if(version >= MPQ_FORMAT_VERSION_4)
 	{
         theHeader->HashTableSize64    = SwapUInt64(theHeader->HashTableSize64);
         theHeader->BlockTableSize64   = SwapUInt64(theHeader->BlockTableSize64);

StormLib/src/SFileAttributes.cpp

@@ -180,7 +180,7 @@ static DWORD LoadAttributesFile(TMPQArchive * ha, LPBYTE pbAttrFile, DWORD cbAtt
         if((pbAttrPtr + cbArraySize) > pbAttrFileEnd)
             return ERROR_FILE_CORRUPT;
 
-        BSWAP_ARRAY32_UNSIGNED(ArrayCRC32, cbCRC32Size);
+        BSWAP_ARRAY32_UNSIGNED(ArrayCRC32, cbArraySize);
         for(i = 0; i < dwAttributesEntries; i++)
             ha->pFileTable[i].dwCrc32 = ArrayCRC32[i];
         pbAttrPtr += cbArraySize;
@@ -196,7 +196,7 @@ static DWORD LoadAttributesFile(TMPQArchive * ha, LPBYTE pbAttrFile, DWORD cbAtt
         if((pbAttrPtr + cbArraySize) > pbAttrFileEnd)
             return ERROR_FILE_CORRUPT;
 
-        BSWAP_ARRAY64_UNSIGNED(ArrayFileTime, cbFileTimeSize);
+        BSWAP_ARRAY64_UNSIGNED(ArrayFileTime, cbArraySize);
         for(i = 0; i < dwAttributesEntries; i++)
             ha->pFileTable[i].FileTime = ArrayFileTime[i];
         pbAttrPtr += cbArraySize;

StormLib/src/SFilePatchArchives.cpp

@@ -480,7 +480,7 @@ static bool IsMatchingPatchFile(
         {
             // Load the patch header
             SFileReadFile(hFile, &PatchHeader, sizeof(MPQ_PATCH_HEADER), &dwTransferred, NULL);
-            BSWAP_ARRAY32_UNSIGNED(pPatchHeader, sizeof(DWORD) * 6);
+            BSWAP_ARRAY32_UNSIGNED(&PatchHeader, sizeof(DWORD) * 6);
 
             // If the file contains an incremental patch,
             // compare the "MD5 before patching" with the base file MD5

StormLib/src/StormLib.h

@@ -631,8 +631,8 @@ typedef struct _TMPQHash
 
 #else
 
-    BYTE   Platform;
     BYTE   Reserved;
+    BYTE   Platform;
     USHORT lcLocale;
 
 #endif

StormLib/src/StormPort.h

@@ -254,6 +254,34 @@
 
 #endif
 
+//-----------------------------------------------------------------------------
+// Defines for Wii U platform
+
+#if !defined(STORMLIB_PLATFORM_DEFINED) && defined(__WIIU__)
+
+  #include <sys/types.h>
+  #include <sys/stat.h>
+  #include <fcntl.h>
+  #include <unistd.h>
+  #include <stdint.h>
+  #include <stdlib.h>
+  #include <stdio.h>
+  #include <stdarg.h>
+  #include <string.h>
+  #include <strings.h>
+  #include <ctype.h>
+  #include <assert.h>
+  #include <errno.h>
+  #include <malloc.h>
+
+  #undef STORMLIB_LITTLE_ENDIAN                     // Wii U is always big endian
+
+  #define STORMLIB_MAC                              // Use Mac compatible code
+  #define STORMLIB_WIIU
+  #define STORMLIB_PLATFORM_DEFINED
+
+#endif
+
 //-----------------------------------------------------------------------------
 // Assumption: If the platform is not defined, assume a Linux-like platform
 

ZAPDTR/ZAPD/CMakeLists.txt

@@ -465,6 +465,12 @@ elseif(CMAKE_SYSTEM_NAME STREQUAL "NintendoSwitch")
                PNG::PNG
                Threads::Threads
        )
+elseif(CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+    set(ADDITIONAL_LIBRARY_DEPENDENCIES
+        "ZAPDUtils;"
+        "libultraship;"
+        PNG::PNG
+    )
 else()
 	set(THREADS_PREFER_PTHREAD_FLAG ON)
 	find_package(Threads REQUIRED)
@@ -478,7 +484,7 @@ else()
 	)
 endif()
 
-if(CMAKE_SYSTEM_NAME STREQUAL "NintendoSwitch")
+if(CMAKE_SYSTEM_NAME MATCHES "NintendoSwitch|CafeOS")
 add_library(pathconf OBJECT pathconf.c)
 target_link_libraries(${PROJECT_NAME} PRIVATE "${ADDITIONAL_LIBRARY_DEPENDENCIES}" $<TARGET_OBJECTS:pathconf> )
 else()

ZAPDTR/ZAPDUtils/Utils/BinaryReader.cpp

@@ -130,7 +130,13 @@ float BinaryReader::ReadSingle()
 	stream->Read((char*)&result, sizeof(float));
 
 	if (endianness != Endianness::Native)
-		result = BitConverter::ToFloatBE((uint8_t*)&result, 0);
+	{
+		float tmp;
+		char* dst = (char*)&tmp;
+		char* src = (char*)&result;
+		dst[3] = src[0]; dst[2] = src[1]; dst[1] = src[2]; dst[0] = src[3];
+		result = tmp;
+	}
 
 	if (std::isnan(result))
 		throw std::runtime_error("BinaryReader::ReadSingle(): Error reading stream");
@@ -145,7 +151,14 @@ double BinaryReader::ReadDouble()
 	stream->Read((char*)&result, sizeof(double));
 
 	if (endianness != Endianness::Native)
-		result = BitConverter::ToDoubleBE((uint8_t*)&result, 0);
+	{
+		double tmp;
+		char* dst = (char*)&tmp;
+		char* src = (char*)&result;
+		dst[7] = src[0]; dst[6] = src[1]; dst[5] = src[2]; dst[4] = src[3];
+		dst[3] = src[4]; dst[2] = src[5]; dst[1] = src[6]; dst[0] = src[7];
+		result = tmp;
+	}
 
 	if (std::isnan(result))
 		throw std::runtime_error("BinaryReader::ReadDouble(): Error reading stream");

ZAPDTR/ZAPDUtils/Utils/BinaryWriter.cpp

@@ -107,7 +107,13 @@ void BinaryWriter::Write(uint64_t value)
 void BinaryWriter::Write(float value)
 {
 	if (endianness != Endianness::Native)
-		value = BitConverter::ToFloatBE((uint8_t*)&value, 0);
+	{
+		float tmp;
+		char* dst = (char*)&tmp;
+		char* src = (char*)&value;
+		dst[3] = src[0]; dst[2] = src[1]; dst[1] = src[2]; dst[0] = src[3];
+		value = tmp;
+	}
 
 	stream->Write((char*)&value, sizeof(float));
 }
@@ -115,7 +121,14 @@ void BinaryWriter::Write(float value)
 void BinaryWriter::Write(double value)
 {
 	if (endianness != Endianness::Native)
-		value = BitConverter::ToDoubleBE((uint8_t*)&value, 0);
+	{
+		double tmp;
+		char* dst = (char*)&tmp;
+		char* src = (char*)&value;
+		dst[7] = src[0]; dst[6] = src[1]; dst[5] = src[2]; dst[4] = src[3];
+		dst[3] = src[4]; dst[2] = src[5]; dst[1] = src[6]; dst[0] = src[7];
+		value = tmp;
+	}
 
 	stream->Write((char*)&value, sizeof(double));
 }

ZAPDTR/ZAPDUtils/Utils/StringHelper.cpp

@@ -64,7 +64,11 @@ void StringHelper::ReplaceOriginal(std::string& str, const std::string& from, co
 
 bool StringHelper::StartsWith(const std::string& s, const std::string& input)
 {
+#if __cplusplus >= 202002L
+	return s.starts_with(input.c_str());
+#else
 	return s.rfind(input, 0) == 0;
+#endif
 }
 
 bool StringHelper::Contains(const std::string& s, const std::string& input)

libultraship/libultraship/CMakeLists.txt

@@ -86,11 +86,17 @@ set(Source_Files__Controller
     "InputEditor.h"
     "KeyboardController.cpp"
     "KeyboardController.h"
-    "SDLController.cpp"
-    "SDLController.h"
     "UltraController.h"
     "DummyController.h"
 )
+
+if (NOT CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+list(APPEND Source_Files__Controller
+    "SDLController.cpp"
+    "SDLController.h"
+)
+endif()
+
 source_group("Source Files\\Controller" FILES ${Source_Files__Controller})
 
 set(Source_Files__Controller__Attachment
@@ -150,16 +156,21 @@ source_group("Source Files\\Lib" FILES ${Source_Files__Lib})
 set(Source_Files__Lib__Fast3D
     "Lib/Fast3D/gfx_cc.cpp"
     "Lib/Fast3D/gfx_cc.h"
-    "Lib/Fast3D/gfx_opengl.cpp"
-    "Lib/Fast3D/gfx_opengl.h"
     "Lib/Fast3D/gfx_pc.cpp"
     "Lib/Fast3D/gfx_pc.h"
     "Lib/Fast3D/gfx_rendering_api.h"
     "Lib/Fast3D/gfx_screen_config.h"
-    "Lib/Fast3D/gfx_sdl.h"
-    "Lib/Fast3D/gfx_sdl2.cpp"
     "Lib/Fast3D/gfx_window_manager_api.h"
 )
+
+if (NOT CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+list(APPEND Source_Files__Lib__Fast3D
+    "Lib/Fast3D/gfx_opengl.cpp"
+    "Lib/Fast3D/gfx_opengl.h"
+    "Lib/Fast3D/gfx_sdl.h"
+    "Lib/Fast3D/gfx_sdl2.cpp"
+)
+endif()
 source_group("Source Files\\Lib\\Fast3D" FILES ${Source_Files__Lib__Fast3D})
 
 if (CMAKE_SYSTEM_NAME STREQUAL "Windows")
@@ -179,14 +190,19 @@ set(Source_Files__Lib__Fast3D__extra
     "Lib/Fast3D/gfx_glx.cpp"
     "Lib/Fast3D/gfx_glx.h"
 )
+elseif (CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+set(Source_Files__Lib__Fast3D__extra
+    "Lib/Fast3D/gfx_wiiu.cpp"
+    "Lib/Fast3D/gfx_wiiu.h"
+    "Lib/Fast3D/gfx_gx2.cpp"
+    "Lib/Fast3D/gfx_gx2.h"
+    "Lib/Fast3D/gx2_shader_gen.c"
+    "Lib/Fast3D/gx2_shader_gen.h"
+)
 endif()
 source_group("Source Files\\Lib\\Fast3D\\extra" FILES ${Source_Files__Lib__Fast3D__extra})
 
 set(Source_Files__Lib__ImGui
-    "Lib/ImGui/backends/imgui_impl_opengl3.cpp"
-    "Lib/ImGui/backends/imgui_impl_opengl3.h"
-    "Lib/ImGui/backends/imgui_impl_sdl.cpp"
-    "Lib/ImGui/backends/imgui_impl_sdl.h"
     "Lib/ImGui/imconfig.h"
     "Lib/ImGui/imgui.cpp"
     "Lib/ImGui/imgui.h"
@@ -199,6 +215,16 @@ set(Source_Files__Lib__ImGui
     "Lib/ImGui/imstb_textedit.h"
     "Lib/ImGui/imstb_truetype.h"
 )
+
+if (NOT CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+list(APPEND Source_Files__Lib__ImGui
+    "Lib/ImGui/backends/imgui_impl_opengl3.cpp"
+    "Lib/ImGui/backends/imgui_impl_opengl3.h"
+    "Lib/ImGui/backends/imgui_impl_sdl.cpp"
+    "Lib/ImGui/backends/imgui_impl_sdl.h"
+)
+endif()
+
 if (CMAKE_SYSTEM_NAME STREQUAL "Windows")
 set(Source_Files__Lib__ImGui__Windows
     "Lib/ImGui/backends/imgui_impl_dx11.cpp"
@@ -206,8 +232,19 @@ set(Source_Files__Lib__ImGui__Windows
     "Lib/ImGui/backends/imgui_impl_win32.cpp"
     "Lib/ImGui/backends/imgui_impl_win32.h"
 )
+elseif (CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+set(Source_Files__Lib__ImGui__WiiU
+    "Lib/ImGui/backends/wiiu/imgui_impl_wiiu.cpp"
+    "Lib/ImGui/backends/wiiu/imgui_impl_wiiu.h"
+    "Lib/ImGui/backends/wiiu/imgui_impl_gx2.cpp"
+    "Lib/ImGui/backends/wiiu/imgui_impl_gx2.h"
+)
 endif ()
-source_group("Source Files\\Lib\\ImGui" FILES ${Source_Files__Lib__ImGui} ${Source_Files__Lib__ImGui__Windows})
+source_group("Source Files\\Lib\\ImGui" FILES
+    ${Source_Files__Lib__ImGui}
+    ${Source_Files__Lib__ImGui__Windows}
+    ${Source_Files__Lib__ImGui__WiiU}
+)
 
 set(Source_Files__Lib__Mercury
     "Lib/Mercury/Mercury.cpp"
@@ -316,6 +353,18 @@ set(Source_Files__NintendoSwitch
 source_group("Source Files\\NintendoSwitch" FILES ${Source_Files__NintendoSwitch})
 endif()
 
+if (CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+set(Source_Files__CafeOS
+    "WiiUController.cpp"
+    "WiiUController.h"
+    "WiiUGamepad.cpp"
+    "WiiUGamepad.h"
+    "WiiUImpl.cpp"
+    "WiiUImpl.h"
+)
+source_group("Source Files\\CafeOS" FILES ${Source_Files__CafeOS})
+endif()
+
 set(ALL_FILES
     ${Header_Files__Resources__Factories}
     ${Header_Files__Resources__Files}
@@ -332,6 +381,7 @@ set(ALL_FILES
     ${Source_Files__Lib__Fast3D__extra}
     ${Source_Files__Lib__ImGui}
     ${Source_Files__Lib__ImGui__Windows}
+    ${Source_Files__Lib__ImGui__WiiU}
     ${Source_Files__Lib__Mercury}
     ${Source_Files__Lib__stb}
     ${Source_Files__Lib__dr_libs}
@@ -343,6 +393,7 @@ set(ALL_FILES
     ${Source_Files__Resources__mpq}
     ${Source_Files__Darwin}
     ${Source_Files__NintendoSwitch}
+    ${Source_Files__CafeOS}
 )
 
 ################################################################################
@@ -445,9 +496,21 @@ if (CMAKE_SYSTEM_NAME STREQUAL "Windows")
 			STORMLIB_NO_AUTO_LINK
 		)
 	endif()
-endif()
 
-if ("${CMAKE_CXX_COMPILER_ID}" MATCHES "GNU|Clang|AppleClang")
+elseif (CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+	target_compile_definitions(${PROJECT_NAME} PRIVATE
+            "$<$<CONFIG:Debug>:"
+                "_DEBUG"
+            ">"
+            "$<$<CONFIG:Release>:"
+                "NDEBUG"
+            ">"
+            "SPDLOG_ACTIVE_LEVEL=3;"
+            "SPDLOG_NO_THREAD_ID;"
+            "SPDLOG_NO_TLS;"
+            "STBI_NO_THREAD_LOCALS;"
+	)
+elseif ("${CMAKE_CXX_COMPILER_ID}" MATCHES "GNU|Clang|AppleClang")
 	target_compile_definitions(${PROJECT_NAME} PRIVATE
 		"$<$<CONFIG:Debug>:"
 				"_DEBUG"
@@ -462,7 +525,7 @@ endif()
 ################################################################################
 # Compile and link options
 ################################################################################
-if (NOT CMAKE_SYSTEM_NAME MATCHES "Windows|NintendoSwitch")
+if (NOT CMAKE_SYSTEM_NAME MATCHES "Windows|NintendoSwitch|CafeOS")
     find_package(SDL2)
     find_package(GLEW)
     find_package(X11)
@@ -485,7 +548,7 @@ if (NOT CMAKE_SYSTEM_NAME MATCHES "Windows|NintendoSwitch")
         set(GLEW-INCLUDE  ${GLEW_INCLUDE_DIRS})
     endif()
     set(SDL2-INCLUDE ${SDL2_INCLUDE_DIRS})
-elseif (CMAKE_SYSTEM_NAME MATCHES "NintendoSwitch")
+elseif (CMAKE_SYSTEM_NAME MATCHES "NintendoSwitch|CafeOS")
     find_package(SDL2)
 else()
     set(GLEW-INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/Lib/GLEW/)
@@ -627,6 +690,15 @@ elseif(CMAKE_SYSTEM_NAME STREQUAL "NintendoSwitch")
         SDL2::SDL2
         Threads::Threads
     )
+elseif(CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+    find_package(SDL2 REQUIRED)
+    target_link_libraries(${PROJECT_NAME}
+        storm
+        SDL2::SDL2-static
+    )
+    target_include_directories(${PROJECT_NAME} PRIVATE
+        ${DEVKITPRO}/portlibs/wiiu/include/
+    )
 else()
     target_link_libraries(${PROJECT_NAME}
         SDL2::SDL2

libultraship/libultraship/ControlDeck.cpp

@@ -3,11 +3,17 @@
 #include "Window.h"
 #include "Controller.h"
 #include "DummyController.h"
-#include "KeyboardController.h"
-#include "SDLController.h"
 #include <Utils/StringHelper.h>
 #include "Cvar.h"
 
+#ifndef __WIIU__
+#include "KeyboardController.h"
+#include "SDLController.h"
+#else
+#include "WiiUGamepad.h"
+#include "WiiUController.h"
+#endif
+
 namespace Ship {
 
     void ControlDeck::Init(uint8_t* bits) {
@@ -20,6 +26,7 @@ namespace Ship {
         virtualDevices.clear();
         physicalDevices.clear();
 
+#ifndef __WIIU__
         for (int i = 0; i < SDL_NumJoysticks(); i++) {
             if (SDL_IsGameController(i)) {
                 auto sdl = std::make_shared<SDLController>(i);
@@ -30,6 +37,20 @@ namespace Ship {
 
         physicalDevices.push_back(std::make_shared<DummyController>("Auto", "Auto", true));
         physicalDevices.push_back(std::make_shared<KeyboardController>());
+#else
+        physicalDevices.push_back(std::make_shared<DummyController>("Auto", "Auto", true));
+
+        auto gamepad = std::make_shared<Ship::WiiUGamepad>();
+        gamepad->Open();
+        physicalDevices.push_back(gamepad);
+
+        for (int i = 0; i < 4; i++) {
+            auto controller = std::make_shared<Ship::WiiUController>((WPADChan) i);
+            controller->Open();
+            physicalDevices.push_back(controller);
+        }
+#endif
+
         physicalDevices.push_back(std::make_shared<DummyController>("Disconnected", "None", false));
 
         for (const auto& device : physicalDevices) {

libultraship/libultraship/Controller.cpp

@@ -21,7 +21,9 @@ namespace Ship {
 	void Controller::Read(OSContPad* pad, int32_t virtualSlot) {
 		ReadFromSource(virtualSlot);
 
+#ifndef __WIIU__
 		SDL_PumpEvents();
+#endif
 
 		// Button Inputs
 		pad->button |= getPressedButtons(virtualSlot) & 0xFFFF;

libultraship/libultraship/GlobalCtx2.cpp

@@ -11,6 +11,8 @@
 #include "OSXFolderManager.h"
 #elif defined(__SWITCH__)
 #include "SwitchImpl.h"
+#elif defined(__WIIU__)
+#include "WiiUImpl.h"
 #endif
 
 namespace Ship {
@@ -59,7 +61,6 @@ namespace Ship {
     void GlobalCtx2::InitWindow() {
         InitLogging();
         Config = std::make_shared<Mercury>(GetPathRelativeToAppDirectory("shipofharkinian.json"));
-        Config->reload();
 
         MainPath = Config->getString("Game.Main Archive", GetPathRelativeToAppDirectory("oot.otr"));
         PatchesPath = Config->getString("Game.Patches Archive", GetAppDirectoryPath() + "/mods");
@@ -73,6 +74,8 @@ namespace Ship {
             MessageBox(nullptr, L"Main OTR file not found!", L"Uh oh", MB_OK);
 #elif defined(__SWITCH__)
             printf("Main OTR file not found!\n");
+#elif defined(__WIIU__)
+            Ship::WiiU::ThrowMissingOTR(MainPath.c_str());
 #else
             SPDLOG_ERROR("Main OTR file not found!");
 #endif
@@ -85,28 +88,36 @@ namespace Ship {
 
     void GlobalCtx2::InitLogging() {
         try {
-            auto logPath = GetPathRelativeToAppDirectory(("logs/" + GetName() + ".log").c_str());
-
             // Setup Logging
             spdlog::init_thread_pool(8192, 1);
+            std::vector<spdlog::sink_ptr> Sinks;
+
             auto SohConsoleSink = std::make_shared<spdlog::sinks::soh_sink_mt>();
             SohConsoleSink->set_level(spdlog::level::trace);
-#if defined(__linux__)
+            Sinks.push_back(SohConsoleSink);
+
+#if (!defined(_WIN32) && !defined(__WIIU__)) || defined(_DEBUG)
             auto ConsoleSink = std::make_shared<spdlog::sinks::stdout_color_sink_mt>();
             ConsoleSink->set_level(spdlog::level::trace);
+            Sinks.push_back(ConsoleSink);
 #endif
+
+#ifndef __WIIU__
+            auto logPath = GetPathRelativeToAppDirectory(("logs/" + GetName() + ".log").c_str());
             auto FileSink = std::make_shared<spdlog::sinks::rotating_file_sink_mt>(logPath, 1024 * 1024 * 10, 10);
             FileSink->set_level(spdlog::level::trace);
-            std::vector<spdlog::sink_ptr> Sinks{
+            Sinks.push_back(FileSink);
-#if defined(__linux__)
-                ConsoleSink,
 #endif
-            	FileSink,
+
-            	SohConsoleSink
-            };
             Logger = std::make_shared<spdlog::async_logger>(GetName(), Sinks.begin(), Sinks.end(), spdlog::thread_pool(), spdlog::async_overflow_policy::block);
             GetLogger()->set_level(spdlog::level::trace);
+
+#ifndef __WIIU__
             GetLogger()->set_pattern("[%Y-%m-%d %H:%M:%S.%e] [%@] [%l] %v");
+#else
+            GetLogger()->set_pattern("[%s:%#] [%l] %v");
+#endif
+
             spdlog::register_logger(GetLogger());
             spdlog::set_default_logger(GetLogger());
         }

libultraship/libultraship/ImGuiImpl.cpp

@@ -30,6 +30,16 @@
 #include "Lib/spdlog/include/spdlog/common.h"
 #include "UltraController.h"
 
+#ifdef __WIIU__
+#include <gx2/registers.h> // GX2SetViewport / GX2SetScissor
+
+#include "Lib/ImGui/backends/wiiu/imgui_impl_gx2.h"
+#include "Lib/ImGui/backends/wiiu/imgui_impl_wiiu.h"
+
+#include "Lib/Fast3D/gfx_wiiu.h"
+#include "Lib/Fast3D/gfx_gx2.h"
+#endif
+
 #if __APPLE__
 #include <SDL_hints.h>
 #else
@@ -101,7 +111,11 @@ namespace SohImGui {
     bool statsWindowOpen;
 
     const char* filters[3] = {
+#ifdef __WIIU__
+        "",
+#else
         "Three-Point",
+#endif
         "Linear",
         "None"
     };
@@ -110,7 +124,11 @@ namespace SohImGui {
 #ifdef _WIN32
         { "dx11", "DirectX" },
 #endif
+#ifndef __WIIU__
         { "sdl", "OpenGL" }
+#else
+        { "wiiu", "GX2" }
+#endif
     };
 
 
@@ -178,10 +196,16 @@ namespace SohImGui {
 
     void ImGuiWMInit() {
         switch (impl.backend) {
+#ifdef __WIIU__
+        case Backend::GX2:
+            ImGui_ImplWiiU_Init();
+            break;
+#else
         case Backend::SDL:
             SDL_SetHint(SDL_HINT_TOUCH_MOUSE_EVENTS, "1");
             ImGui_ImplSDL2_InitForOpenGL(static_cast<SDL_Window*>(impl.sdl.window), impl.sdl.context);
             break;
+#endif
 #if defined(ENABLE_DX11) || defined(ENABLE_DX12)
         case Backend::DX11:
             ImGui_ImplWin32_Init(impl.dx11.window);
@@ -195,6 +219,11 @@ namespace SohImGui {
 
     void ImGuiBackendInit() {
         switch (impl.backend) {
+#ifdef __WIIU__
+        case Backend::GX2:
+            ImGui_ImplGX2_Init();
+            break;
+#else
         case Backend::SDL:
         #if defined(__APPLE__)
             ImGui_ImplOpenGL3_Init("#version 410 core");
@@ -202,6 +231,7 @@ namespace SohImGui {
             ImGui_ImplOpenGL3_Init("#version 120");
         #endif
             break;
+#endif
 
 #if defined(ENABLE_DX11) || defined(ENABLE_DX12)
         case Backend::DX11:
@@ -215,9 +245,17 @@ namespace SohImGui {
 
     void ImGuiProcessEvent(EventImpl event) {
         switch (impl.backend) {
+#ifdef __WIIU__
+        case Backend::GX2:
+            if (!ImGui_ImplWiiU_ProcessInput((ImGui_ImplWiiU_ControllerInput*)event.gx2.input)) {
+                
+            }
+            break;
+#else
         case Backend::SDL:
             ImGui_ImplSDL2_ProcessEvent(static_cast<const SDL_Event*>(event.sdl.event));
             break;
+#endif
 #if defined(ENABLE_DX11) || defined(ENABLE_DX12)
         case Backend::DX11:
             ImGui_ImplWin32_WndProcHandler(static_cast<HWND>(event.win32.handle), event.win32.msg, event.win32.wparam, event.win32.lparam);
@@ -230,9 +268,14 @@ namespace SohImGui {
 
     void ImGuiWMNewFrame() {
         switch (impl.backend) {
+#ifdef __WIIU__
+        case Backend::GX2:
+            break;
+#else
         case Backend::SDL:
             ImGui_ImplSDL2_NewFrame(static_cast<SDL_Window*>(impl.sdl.window));
             break;
+#endif
 #if defined(ENABLE_DX11) || defined(ENABLE_DX12)
         case Backend::DX11:
             ImGui_ImplWin32_NewFrame();
@@ -245,9 +288,16 @@ namespace SohImGui {
 
     void ImGuiBackendNewFrame() {
         switch (impl.backend) {
+#ifdef __WIIU__
+        case Backend::GX2:
+            io->DeltaTime = (float) frametime / 1000.0f / 1000.0f;
+            ImGui_ImplGX2_NewFrame();
+            break;
+#else
         case Backend::SDL:
             ImGui_ImplOpenGL3_NewFrame();
             break;
+#endif
 #if defined(ENABLE_DX11) || defined(ENABLE_DX12)
         case Backend::DX11:
             ImGui_ImplDX11_NewFrame();
@@ -260,9 +310,20 @@ namespace SohImGui {
 
     void ImGuiRenderDrawData(ImDrawData* data) {
         switch (impl.backend) {
+#ifdef __WIIU__
+        case Backend::GX2:
+            ImGui_ImplGX2_RenderDrawData(data);
+
+            // Reset viewport and scissor for drawing the keyboard
+            GX2SetViewport(0.0f, 0.0f, io->DisplaySize.x, io->DisplaySize.y, 0.0f, 1.0f);
+            GX2SetScissor(0, 0, io->DisplaySize.x, io->DisplaySize.y);
+            ImGui_ImplWiiU_DrawKeyboardOverlay();
+            break;
+#else
         case Backend::SDL:
             ImGui_ImplOpenGL3_RenderDrawData(data);
             break;
+#endif
 #if defined(ENABLE_DX11) || defined(ENABLE_DX12)
         case Backend::DX11:
             ImGui_ImplDX11_RenderDrawData(data);
@@ -395,9 +456,19 @@ namespace SohImGui {
         Ship::Switch::SetupFont(io->Fonts);
     #endif
 
+    #ifdef __WIIU__
+        // Scale everything by 2 for the Wii U
+        ImGui::GetStyle().ScaleAllSizes(2.0f);
+        io->FontGlobalScale = 2.0f;
+
+        // Setup display sizes
+        io->DisplaySize.x = window_impl.gx2.width;
+        io->DisplaySize.y =  window_impl.gx2.height;
+    #endif
+
         lastBackendID = GetBackendID(GlobalCtx2::GetInstance()->GetConfig());
         if (CVar_GetS32("gOpenMenuBar", 0) != 1) {
-            #ifdef __SWITCH__
+            #if defined(__SWITCH__) || defined(__WIIU__)
             SohImGui::overlay->TextDrawNotification(30.0f, true, "Press - to access enhancements menu");
             #else
             SohImGui::overlay->TextDrawNotification(30.0f, true, "Press F1 to access enhancements menu");
@@ -412,6 +483,18 @@ namespace SohImGui {
         if (UseViewports()) {
             io->ConfigFlags |= ImGuiConfigFlags_ViewportsEnable;
         }
+
+    #ifdef __SWITCH__
+        bool enableControllerNavigation = true;
+    #else
+        bool enableControllerNavigation = CVar_GetS32("gControlNav", 0);
+    #endif
+        if (enableControllerNavigation && CVar_GetS32("gOpenMenuBar", 0)) {
+            io->ConfigFlags |= ImGuiConfigFlags_NavEnableGamepad | ImGuiConfigFlags_NavEnableKeyboard;
+        } else {
+            io->ConfigFlags &= ~ImGuiConfigFlags_NavEnableGamepad;
+        }
+
         console->Init();
         overlay->Init();
         controller->Init();
@@ -616,7 +699,11 @@ namespace SohImGui {
             ImGui::SetCursorPosX(ImGui::GetCursorPosX() - 7.0f);
         }
         if (PlusMinusButton) {
+        #ifdef __WIIU__
+            ImGui::PushItemWidth(ImGui::GetWindowSize().x - 79.0f * 2);
+        #else
             ImGui::PushItemWidth(ImGui::GetWindowSize().x - 79.0f);
+        #endif
         }
         if (ImGui::SliderFloat(id, &val, min, max, format))
         {
@@ -846,12 +933,8 @@ namespace SohImGui {
         #else
             bool enableControllerNavigation = CVar_GetS32("gControlNav", 0);
         #endif
-            if (enableControllerNavigation) {
+            if (enableControllerNavigation && CVar_GetS32("gOpenMenuBar", 0)) {
-                if (CVar_GetS32("gOpenMenuBar", 0)) {
+                io->ConfigFlags |= ImGuiConfigFlags_NavEnableGamepad | ImGuiConfigFlags_NavEnableKeyboard;
-                    io->ConfigFlags |=ImGuiConfigFlags_NavEnableGamepad | ImGuiConfigFlags_NavEnableKeyboard;
-                } else {
-                    io->ConfigFlags &= ~ImGuiConfigFlags_NavEnableGamepad;
-                }
             } else {
                 io->ConfigFlags &= ~ImGuiConfigFlags_NavEnableGamepad;
             }
@@ -875,13 +958,18 @@ namespace SohImGui {
             if (DefaultAssets.contains("Game_Icon")) {
             #ifdef __SWITCH__
                 ImVec2 iconSize = ImVec2(20.0f, 20.0f);
+                float posScale = 1.0f;
+            #elif defined(__WIIU__)
+                ImVec2 iconSize = ImVec2(16.0f * 2, 16.0f * 2);
+                float posScale = 2.0f;
             #else
                 ImVec2 iconSize = ImVec2(16.0f, 16.0f);
+                float posScale = 1.0f;
             #endif
-                ImGui::SetCursorPos(ImVec2(5, 2.5f));
+                ImGui::SetCursorPos(ImVec2(5, 2.5f) * posScale);
                 ImGui::Image(GetTextureByID(DefaultAssets["Game_Icon"]->textureId), iconSize);
                 ImGui::SameLine();
-                ImGui::SetCursorPos(ImVec2(25, 0));
+                ImGui::SetCursorPos(ImVec2(25, 0) * posScale);
             }
 
             static ImVec2 windowPadding(8.0f, 8.0f);
@@ -957,9 +1045,11 @@ namespace SohImGui {
                     Tooltip("Multiplies your output resolution by the value inputted, as a more intensive but effective form of anti-aliasing");
                     gfx_current_dimensions.internal_mul = CVar_GetFloat("gInternalResolution", 1);
                 #endif
+                #ifndef __WIIU__
                     PaddedEnhancementSliderInt("MSAA: %d", "##IMSAA", "gMSAAValue", 1, 8, "", 1, false, true, false);
                     Tooltip("Activates multi-sample anti-aliasing when above 1x up to 8x for 8 samples for every pixel");
                     gfx_msaa_level = CVar_GetS32("gMSAAValue", 1);
+                #endif
 
                     if (impl.backend == Backend::DX11)
                     {
@@ -1434,7 +1524,7 @@ namespace SohImGui {
 
                 const char* fps_cvar = "gInterpolationFPS";
                 {
-                #ifdef __SWITCH__
+                #if defined(__SWITCH__) || defined(__WIIU__)
                     int minFps = 20;
                     int maxFps = 60;
                 #else
@@ -1444,6 +1534,12 @@ namespace SohImGui {
 
                     int val = CVar_GetS32(fps_cvar, minFps);
                     val = MAX(MIN(val, maxFps), 20);
+
+                #ifdef __WIIU__
+                    // only support divisors of 60 on the Wii U
+                    val = 60 / (60 / val);
+                #endif
+
                     int fps = val;
 
                     if (fps == 20)
@@ -1458,15 +1554,28 @@ namespace SohImGui {
                     std::string MinusBTNFPSI = " - ##FPSInterpolation";
                     std::string PlusBTNFPSI = " + ##FPSInterpolation";
                     if (ImGui::Button(MinusBTNFPSI.c_str())) {
+                    #ifdef __WIIU__
+                        if (val >= 60) val = 30;
+                        else val = 20;
+                    #else
                         val--;
+                    #endif
                         CVar_SetS32(fps_cvar, val);
                         needs_save = true;
                     }
                     ImGui::SameLine();
                     ImGui::SetCursorPosX(ImGui::GetCursorPosX() - 7.0f);
+                #ifdef __WIIU__
+                    ImGui::PushItemWidth(ImGui::GetWindowSize().x - 79.0f * 2);
+                #else
                     ImGui::PushItemWidth(ImGui::GetWindowSize().x - 79.0f);
+                #endif
                     if (ImGui::SliderInt("##FPSInterpolation", &val, minFps, maxFps, "", ImGuiSliderFlags_AlwaysClamp))
                     {
+                    #ifdef __WIIU__
+                        // only support divisors of 60 on the Wii U
+                        val = 60 / (60 / val);
+                    #endif
                         if (val > 360)
                         {
                             val = 360;
@@ -1490,7 +1599,12 @@ namespace SohImGui {
                     ImGui::SameLine();
                     ImGui::SetCursorPosX(ImGui::GetCursorPosX() - 7.0f);
                     if (ImGui::Button(PlusBTNFPSI.c_str())) {
+                    #ifdef __WIIU__
+                        if (val <= 20) val = 30;
+                        else val = 60;
+                    #else
                         val++;
+                    #endif
                         CVar_SetS32(fps_cvar, val);
                         needs_save = true;
                     }
@@ -1791,14 +1905,18 @@ namespace SohImGui {
             ImGui::PushStyleColor(ImGuiCol_Border, ImVec4(0, 0, 0, 0));
             ImGui::Begin("Debug Stats", &statsWindowOpen, ImGuiWindowFlags_NoFocusOnAppearing);
 
-#ifdef _WIN32
+#if defined(_WIN32)
             ImGui::Text("Platform: Windows");
-#elif __APPLE__
+#elif defined(__APPLE__)
             ImGui::Text("Platform: macOS");
 #elif defined(__SWITCH__)
             ImGui::Text("Platform: Nintendo Switch");
-#else
+#elif defined(__WIIU__)
+            ImGui::Text("Platform: Nintendo Wii U");
+#elif defined(__linux__)
             ImGui::Text("Platform: Linux");
+#else
+            ImGui::Text("Platform: Unknown");
 #endif
             ImGui::Text("Status: %.3f ms/frame (%.1f FPS)", 1000.0f / framerate, framerate);
             ImGui::End();
@@ -2281,6 +2399,13 @@ namespace SohImGui {
             return gfx_d3d11_get_texture_by_id(id);
         }
 #endif
+#ifdef __WIIU__
+        if (impl.backend == Backend::GX2)
+        {
+            return gfx_gx2_texture_for_imgui(id);
+        }
+#endif
+
         return reinterpret_cast<ImTextureID>(id);
     }
 

libultraship/libultraship/ImGuiImpl.h

@@ -24,7 +24,8 @@ struct GameAsset {
 namespace SohImGui {
     enum class Backend {
         DX11,
-        SDL
+        SDL,
+        GX2,
     };
 
     enum class Dialogues {
@@ -45,6 +46,10 @@ namespace SohImGui {
                 void* window;
                 void* context;
             } sdl;
+            struct {
+                uint32_t width;
+                uint32_t height;
+            } gx2;
         };
     } WindowImpl;
 
@@ -58,6 +63,9 @@ namespace SohImGui {
         struct {
             void* event;
         } sdl;
+        struct {
+            void* input;
+        } gx2;
     } EventImpl;
 
     extern WindowImpl impl;

libultraship/libultraship/InputEditor.cpp

@@ -45,6 +45,9 @@ namespace Ship {
 			if(btn != -1) {
 				backend->SetButtonMapping(CurrentPort, n64Btn, btn);
 				BtnReading = -1;
+
+				// avoid immediately triggering another button during gamepad nav
+				ImGui::SetKeyboardFocusHere(0);
 			}
 		}
 
@@ -145,9 +148,17 @@ namespace Ship {
 
 				ImGui::SetCursorPosX(ImGui::GetCursorPosX() + 5);
 
-				ImGui::BeginChild("##MSInput", ImVec2(90, 50), false);
+#ifdef __WIIU__
+				ImGui::BeginChild("##MSInput", ImVec2(90 * 2, 50 * 2), false);
+#else
+				ImGui::BeginChild("##MSInput", ImVec2(90 , 50), false);
+#endif
 				ImGui::Text("Deadzone");
+			#ifdef __WIIU__
+				ImGui::PushItemWidth(80 * 2);
+			#else
 				ImGui::PushItemWidth(80);
+			#endif
 				ImGui::InputFloat("##MDZone", &profile->AxisDeadzones[0] /* This is the SDL value for left stick X axis */, 1.0f, 0.0f, "%.0f");
 				ImGui::PopItemWidth();
 				ImGui::EndChild();
@@ -176,13 +187,25 @@ namespace Ship {
 
 				ImGui::SameLine();
 				ImGui::SetCursorPosX(ImGui::GetCursorPosX() + 5);
+#ifdef __WIIU__
+				ImGui::BeginChild("##CSInput", ImVec2(90 * 2, 85 * 2), false);
+#else
 				ImGui::BeginChild("##CSInput", ImVec2(90, 85), false);
+#endif
 					ImGui::Text("Deadzone");
+			#ifdef __WIIU__
+					ImGui::PushItemWidth(80 * 2);
+			#else
 					ImGui::PushItemWidth(80);
-					ImGui::InputFloat("##MDZone", &profile->AxisDeadzones[2] /* This is the SDL value for left stick X axis */, 1.0f, 0.0f, "%.0f");
+			#endif
+					ImGui::InputFloat("##MDZone", &profile->AxisDeadzones[2] /* This is the SDL value for right stick X axis */, 1.0f, 0.0f, "%.0f");
 					ImGui::PopItemWidth();
 					ImGui::Text("Sensitivity");
+			#ifdef __WIIU__
+					ImGui::PushItemWidth(80 * 2);
+			#else
 					ImGui::PushItemWidth(80);
+			#endif
 					ImGui::InputFloat("##MSensitivity", &profile->AxisSensitivities[2] /* This is the SDL value for right stick X axis */, 1.0f, 0.0f, "%.0f");
 					profile->AxisSensitivities[3] = profile->AxisSensitivities[2];
 					ImGui::PopItemWidth();
@@ -195,15 +218,20 @@ namespace Ship {
 		}
 
 		if(Backend->CanGyro()) {
+		#ifndef __WIIU__
 			ImGui::SameLine();
-
+		#endif
 			SohImGui::BeginGroupPanel("Gyro Options", ImVec2(175, 20));
 				float cursorX = ImGui::GetCursorPosX() + 5;
 				ImGui::SetCursorPosX(cursorX);
 				ImGui::Checkbox("Enable Gyro", &profile->UseGyro);
 				ImGui::SetCursorPosX(cursorX);
 				ImGui::Text("Gyro Sensitivity: %d%%", static_cast<int>(100.0f * profile->GyroData[GYRO_SENSITIVITY]));
+			#ifdef __WIIU__
+				ImGui::PushItemWidth(135.0f * 2);
+			#else
 				ImGui::PushItemWidth(135.0f);
+			#endif
 				ImGui::SetCursorPosX(cursorX);
 				ImGui::SliderFloat("##GSensitivity", &profile->GyroData[GYRO_SENSITIVITY], 0.0f, 1.0f, "");
 				ImGui::PopItemWidth();
@@ -218,13 +246,25 @@ namespace Ship {
 
 				ImGui::SameLine();
 				ImGui::SetCursorPosX(ImGui::GetCursorPosX() + 5);
+			#ifdef __WIIU__
+				ImGui::BeginChild("##GyInput", ImVec2(90 * 2, 85 * 2), false);
+			#else
 				ImGui::BeginChild("##GyInput", ImVec2(90, 85), false);
+			#endif
 				ImGui::Text("Drift X");
+			#ifdef __WIIU__
+				ImGui::PushItemWidth(80 * 2);
+			#else
 				ImGui::PushItemWidth(80);
+			#endif
 				ImGui::InputFloat("##GDriftX", &profile->GyroData[DRIFT_X], 1.0f, 0.0f, "%.1f");
 				ImGui::PopItemWidth();
 				ImGui::Text("Drift Y");
+			#ifdef __WIIU__
+				ImGui::PushItemWidth(80 * 2);
+			#else
 				ImGui::PushItemWidth(80);
+			#endif
 				ImGui::InputFloat("##GDriftY", &profile->GyroData[DRIFT_Y], 1.0f, 0.0f, "%.1f");
 				ImGui::PopItemWidth();
 				ImGui::EndChild();
@@ -250,6 +290,8 @@ namespace Ship {
 		ImGui::SetCursorPosX(cursor.x);
 	#ifdef __SWITCH__
 		ImGui::SetCursorPosY(cursor.y + 167);
+	#elif defined(__WIIU__)
+		ImGui::SetCursorPosY(cursor.y + 120 * 2);
 	#else
 		ImGui::SetCursorPosY(cursor.y + 120);
 	#endif
@@ -261,7 +303,11 @@ namespace Ship {
 				ImGui::SetCursorPosX(cursorX);
 				ImGui::Text("Rumble Force: %d%%", static_cast<int>(100.0f * profile->RumbleStrength));
 				ImGui::SetCursorPosX(cursorX);
+			#ifdef __WIIU__
+				ImGui::PushItemWidth(135.0f * 2);
+			#else
 				ImGui::PushItemWidth(135.0f);
+			#endif
 				ImGui::SliderFloat("##RStrength", &profile->RumbleStrength, 0.0f, 1.0f, "");
 				ImGui::PopItemWidth();
 			}
@@ -279,6 +325,9 @@ namespace Ship {
 #ifdef __SWITCH__
 		ImVec2 minSize = ImVec2(641, 250);
 		ImVec2 maxSize = ImVec2(2200, 505);
+#elif defined(__WIIU__)
+		ImVec2 minSize = ImVec2(641 * 2, 250 * 2);
+		ImVec2 maxSize = ImVec2(1200 * 2, 290 * 2);
 #else
 		ImVec2 minSize = ImVec2(641, 250);
 		ImVec2 maxSize = ImVec2(1200, 290);

libultraship/libultraship/Lib/Fast3D/gfx_gx2.cpp

@@ -0,0 +1,826 @@
+#ifdef __WIIU__
+
+#include "../../Window.h"
+
+#include <stdint.h>
+#include <stdbool.h>
+#include <stdio.h>
+#include <malloc.h>
+
+#include <map>
+
+#ifndef _LANGUAGE_C
+#define _LANGUAGE_C
+#endif
+#include "PR/ultra64/gbi.h"
+
+#include "gfx_cc.h"
+#include "gfx_rendering_api.h"
+#include "../../GlobalCtx2.h"
+#include "gfx_pc.h"
+#include "gfx_wiiu.h"
+
+#include <gx2/texture.h>
+#include <gx2/draw.h>
+#include <gx2/clear.h>
+#include <gx2/state.h>
+#include <gx2/swap.h>
+#include <gx2/event.h>
+#include <gx2/utils.h>
+#include <gx2/mem.h>
+#include <gx2/registers.h>
+#include <gx2/display.h>
+#include "gx2_shader_gen.h"
+
+#include <proc_ui/procui.h>
+#include <coreinit/memory.h>
+
+#include "../ImGui/backends/wiiu/imgui_impl_gx2.h"
+
+#define ALIGN(x, align) (((x) + ((align) -1)) & ~((align) -1))
+
+struct ShaderProgram {
+    struct ShaderGroup group;
+    uint8_t num_inputs;
+    bool used_textures[2];
+    bool used_noise;
+    uint32_t window_params_offset;
+    uint32_t samplers_location[2];
+};
+
+struct Texture {
+    GX2Texture texture;
+    bool texture_uploaded;
+
+    GX2Sampler sampler;
+    bool sampler_set;
+
+    // For ImGui rendering
+    ImGui_ImplGX2_Texture imtex;
+};
+
+struct Framebuffer {
+    GX2ColorBuffer color_buffer;
+    bool colorBufferMem1;
+    GX2DepthBuffer depth_buffer;
+    bool depthBufferMem1;
+
+    GX2Texture texture;
+    GX2Sampler sampler;
+
+    // For ImGui rendering
+    ImGui_ImplGX2_Texture imtex;
+};
+
+static struct Framebuffer main_framebuffer;
+static GX2DepthBuffer depthReadBuffer;
+static struct Framebuffer *current_framebuffer;
+
+static std::map<std::pair<uint64_t, uint32_t>, struct ShaderProgram> shader_program_pool;
+static struct ShaderProgram *current_shader_program;
+
+static struct Texture *current_texture;
+static int current_tile;
+
+// 96 Mb (should be more than enough to draw everything without waiting for the GPU)
+#define DRAW_BUFFER_SIZE 0x6000000
+static uint8_t *draw_buffer = nullptr;
+static uint8_t *draw_ptr = nullptr;
+
+static uint32_t frame_count;
+static float current_noise_scale;
+static FilteringMode current_filter_mode = FILTER_LINEAR;
+
+static BOOL current_depth_test = TRUE;
+static BOOL current_depth_write = TRUE;
+static GX2CompareFunction current_depth_compare_function = GX2_COMPARE_FUNC_LESS;
+
+static float current_viewport_x = 0.0f;
+static float current_viewport_y = 0.0f;
+static float current_viewport_width = WIIU_DEFAULT_FB_WIDTH;
+static float current_viewport_height = WIIU_DEFAULT_FB_HEIGHT;
+
+static uint32_t current_scissor_x = 0;
+static uint32_t current_scissor_y = 0;
+static uint32_t current_scissor_width = WIIU_DEFAULT_FB_WIDTH;
+static uint32_t current_scissor_height = WIIU_DEFAULT_FB_HEIGHT;
+
+static bool current_zmode_decal = false;
+static bool current_use_alpha = false;
+
+static inline GX2SamplerVar *GX2GetPixelSamplerVar(const GX2PixelShader *shader, const char *name)
+{
+    for (uint32_t i = 0; i < shader->samplerVarCount; ++i) {
+        if (strcmp(name, shader->samplerVars[i].name) == 0) {
+            return &shader->samplerVars[i];
+        }
+    }
+
+    return nullptr;
+}
+
+static inline int32_t GX2GetPixelSamplerVarLocation(const GX2PixelShader *shader, const char *name)
+{
+    GX2SamplerVar *sampler = GX2GetPixelSamplerVar(shader, name);
+    return sampler ? sampler->location : -1;
+}
+
+static inline int32_t GX2GetPixelUniformVarOffset(const GX2PixelShader *shader, const char *name)
+{
+    GX2UniformVar *uniform = GX2GetPixelUniformVar(shader, name);
+    return uniform ? uniform->offset : -1;
+}
+
+static void gfx_gx2_init_framebuffer(struct Framebuffer *buffer, uint32_t width, uint32_t height) {
+    memset(&buffer->color_buffer, 0, sizeof(GX2ColorBuffer));
+    buffer->color_buffer.surface.use = GX2_SURFACE_USE_TEXTURE_COLOR_BUFFER_TV;
+    buffer->color_buffer.surface.dim = GX2_SURFACE_DIM_TEXTURE_2D;
+    buffer->color_buffer.surface.width = width;
+    buffer->color_buffer.surface.height = height;
+    buffer->color_buffer.surface.depth = 1;
+    buffer->color_buffer.surface.mipLevels = 1;
+    buffer->color_buffer.surface.format = GX2_SURFACE_FORMAT_UNORM_R8_G8_B8_A8;
+    buffer->color_buffer.surface.aa = GX2_AA_MODE1X;
+    buffer->color_buffer.surface.tileMode = GX2_TILE_MODE_DEFAULT;
+    buffer->color_buffer.viewNumSlices = 1;
+
+    memset(&buffer->depth_buffer, 0, sizeof(GX2DepthBuffer));
+    buffer->depth_buffer.surface.use = GX2_SURFACE_USE_DEPTH_BUFFER | GX2_SURFACE_USE_TEXTURE;
+    buffer->depth_buffer.surface.dim = GX2_SURFACE_DIM_TEXTURE_2D;
+    buffer->depth_buffer.surface.width = width;
+    buffer->depth_buffer.surface.height = height;
+    buffer->depth_buffer.surface.depth = 1;
+    buffer->depth_buffer.surface.mipLevels = 1;
+    buffer->depth_buffer.surface.format = GX2_SURFACE_FORMAT_FLOAT_R32;
+    buffer->depth_buffer.surface.aa = GX2_AA_MODE1X;
+    buffer->depth_buffer.surface.tileMode = GX2_TILE_MODE_DEFAULT;
+    buffer->depth_buffer.viewNumSlices = 1;
+    buffer->depth_buffer.depthClear = 1.0f;
+}
+
+static struct GfxClipParameters gfx_gx2_get_clip_parameters(void) {
+    return { false, false };
+}
+
+static void gfx_gx2_set_uniforms(struct ShaderProgram *prg) {
+    if (prg->used_noise) {
+        float window_params_array[2] = { current_noise_scale, (float) frame_count };
+
+        GX2SetPixelUniformReg(prg->window_params_offset, 2, window_params_array);
+    }
+}
+
+static void gfx_gx2_unload_shader(struct ShaderProgram *old_prg) {
+    current_shader_program = nullptr;
+}
+
+static void gfx_gx2_load_shader(struct ShaderProgram *new_prg) {
+    current_shader_program = new_prg;
+
+    GX2SetFetchShader(&new_prg->group.fetchShader);
+    GX2SetVertexShader(&new_prg->group.vertexShader);
+    GX2SetPixelShader(&new_prg->group.pixelShader);
+
+    gfx_gx2_set_uniforms(new_prg);
+}
+
+static struct ShaderProgram* gfx_gx2_create_and_load_new_shader(uint64_t shader_id0, uint32_t shader_id1) {
+    struct CCFeatures cc_features;
+    gfx_cc_get_features(shader_id0, shader_id1, &cc_features);
+
+    struct ShaderProgram* prg = &shader_program_pool[std::make_pair(shader_id0, shader_id1)];
+
+    printf("Generating shader: %016llx-%08x\n", shader_id0, shader_id1);
+    if (gx2GenerateShaderGroup(&prg->group, &cc_features) != 0) {
+        printf("Failed to generate shader\n");
+        current_shader_program = nullptr;
+        return nullptr;
+    }
+
+    prg->num_inputs = cc_features.num_inputs;
+    prg->used_textures[0] = cc_features.used_textures[0];
+    prg->used_textures[1] = cc_features.used_textures[1];
+
+    gfx_gx2_load_shader(prg);
+
+    prg->window_params_offset = GX2GetPixelUniformVarOffset(&prg->group.pixelShader, "window_params");
+    prg->samplers_location[0] = GX2GetPixelSamplerVarLocation(&prg->group.pixelShader, "uTex0");
+    prg->samplers_location[1] = GX2GetPixelSamplerVarLocation(&prg->group.pixelShader, "uTex1");
+
+    prg->used_noise = cc_features.opt_alpha && cc_features.opt_noise;
+
+    printf("Generated and loaded shader\n");
+
+    return prg;
+}
+
+static struct ShaderProgram *gfx_gx2_lookup_shader(uint64_t shader_id0, uint32_t shader_id1) {
+    auto it = shader_program_pool.find(std::make_pair(shader_id0, shader_id1));
+    return it == shader_program_pool.end() ? nullptr : &it->second;
+}
+
+static void gfx_gx2_shader_get_info(struct ShaderProgram *prg, uint8_t *num_inputs, bool used_textures[2]) {
+    *num_inputs = prg->num_inputs;
+    used_textures[0] = prg->used_textures[0];
+    used_textures[1] = prg->used_textures[1];
+}
+
+static uint32_t gfx_gx2_new_texture(void) {
+    // some 32-bit trickery :P
+    struct Texture *tex = (struct Texture *) calloc(1, sizeof(struct Texture));
+
+    tex->imtex.Texture = &tex->texture;
+    tex->imtex.Sampler = &tex->sampler;
+
+    return (uint32_t) tex;
+}
+
+static void gfx_gx2_delete_texture(uint32_t texture_id) {
+    struct Texture *tex = (struct Texture *) texture_id;
+
+    if (tex->texture.surface.image) {
+        free(tex->texture.surface.image);
+    }
+
+    free((void *) tex);
+}
+
+static void gfx_gx2_select_texture(int tile, uint32_t texture_id) {
+    struct Texture *tex = (struct Texture *) texture_id;
+    current_texture = tex;
+    current_tile = tile;
+
+    if (current_shader_program) {
+        uint32_t sampler_location = current_shader_program->samplers_location[tile];
+
+        if (tex->texture_uploaded) {
+            GX2SetPixelTexture(&tex->texture, sampler_location);
+        }
+
+        if (tex->sampler_set) {
+            GX2SetPixelSampler(&tex->sampler, sampler_location);
+        }
+    }
+}
+
+static void gfx_gx2_upload_texture(const uint8_t *rgba32_buf, uint32_t width, uint32_t height) {
+    struct Texture *tex = current_texture;
+    assert(tex);
+
+    if ((tex->texture.surface.width != width) ||
+        (tex->texture.surface.height != height) || 
+        !tex->texture.surface.image) {
+
+        if (tex->texture.surface.image) {
+            free(tex->texture.surface.image);
+            tex->texture.surface.image = nullptr;
+        }
+
+        memset(&tex->texture, 0, sizeof(GX2Texture));
+        tex->texture.surface.use = GX2_SURFACE_USE_TEXTURE;
+        tex->texture.surface.dim = GX2_SURFACE_DIM_TEXTURE_2D;
+        tex->texture.surface.width = width;
+        tex->texture.surface.height = height;
+        tex->texture.surface.depth = 1;
+        tex->texture.surface.mipLevels = 1;
+        tex->texture.surface.format = GX2_SURFACE_FORMAT_UNORM_R8_G8_B8_A8;
+        tex->texture.surface.aa = GX2_AA_MODE1X;
+        tex->texture.surface.tileMode = GX2_TILE_MODE_LINEAR_ALIGNED;
+        tex->texture.viewFirstMip = 0;
+        tex->texture.viewNumMips = 1;
+        tex->texture.viewFirstSlice = 0;
+        tex->texture.viewNumSlices = 1;
+        tex->texture.compMap = GX2_COMP_MAP(GX2_SQ_SEL_R, GX2_SQ_SEL_G, GX2_SQ_SEL_B, GX2_SQ_SEL_A);
+
+        GX2CalcSurfaceSizeAndAlignment(&tex->texture.surface);
+        GX2InitTextureRegs(&tex->texture);
+
+        tex->texture.surface.image = memalign(tex->texture.surface.alignment, tex->texture.surface.imageSize);
+    }
+
+    uint8_t* buf = (uint8_t *) tex->texture.surface.image;
+    assert(buf);
+
+    for (uint32_t y = 0; y < height; ++y) {
+        memcpy(buf + (y * tex->texture.surface.pitch * 4), rgba32_buf + (y * width * 4), width * 4);
+    }
+
+    GX2Invalidate(GX2_INVALIDATE_MODE_CPU_TEXTURE, tex->texture.surface.image, tex->texture.surface.imageSize);
+
+    if (current_shader_program) {
+        GX2SetPixelTexture(&tex->texture, current_shader_program->samplers_location[current_tile]);
+    }
+
+    tex->texture_uploaded = true;
+}
+
+static GX2TexClampMode gfx_cm_to_gx2(uint32_t val) {
+    switch (val) {
+        case G_TX_NOMIRROR | G_TX_CLAMP:
+            return GX2_TEX_CLAMP_MODE_CLAMP;
+        case G_TX_MIRROR | G_TX_WRAP:
+            return GX2_TEX_CLAMP_MODE_MIRROR;
+        case G_TX_MIRROR | G_TX_CLAMP:
+            return GX2_TEX_CLAMP_MODE_MIRROR_ONCE;
+        case G_TX_NOMIRROR | G_TX_WRAP:
+            return GX2_TEX_CLAMP_MODE_WRAP;
+    }
+
+    return GX2_TEX_CLAMP_MODE_WRAP;
+}
+
+static void gfx_gx2_set_sampler_parameters(int tile, bool linear_filter, uint32_t cms, uint32_t cmt) {
+    struct Texture *tex = current_texture;
+    assert(tex);
+
+    current_tile = tile;
+
+    GX2InitSampler(&tex->sampler, GX2_TEX_CLAMP_MODE_CLAMP,
+        (linear_filter && current_filter_mode == FILTER_LINEAR) ? 
+        GX2_TEX_XY_FILTER_MODE_LINEAR : GX2_TEX_XY_FILTER_MODE_POINT);
+
+    GX2InitSamplerClamping(&tex->sampler, gfx_cm_to_gx2(cms), gfx_cm_to_gx2(cmt), GX2_TEX_CLAMP_MODE_WRAP);
+
+    if (current_shader_program) {
+        GX2SetPixelSampler(&tex->sampler, current_shader_program->samplers_location[tile]);
+    }
+
+    tex->sampler_set = true;
+}
+
+static void gfx_gx2_set_depth_test_and_mask(bool depth_test, bool z_upd) {
+    current_depth_test = depth_test || z_upd;
+    current_depth_write = z_upd;
+    current_depth_compare_function = depth_test ? GX2_COMPARE_FUNC_LEQUAL : GX2_COMPARE_FUNC_ALWAYS;
+
+    GX2SetDepthOnlyControl(current_depth_test, current_depth_write, current_depth_compare_function);
+}
+
+static void gfx_gx2_set_zmode_decal(bool zmode_decal) {
+    current_zmode_decal = zmode_decal;
+    if (zmode_decal) {
+        GX2SetPolygonControl(GX2_FRONT_FACE_CCW, FALSE, FALSE, TRUE,
+                             GX2_POLYGON_MODE_TRIANGLE, GX2_POLYGON_MODE_TRIANGLE,
+                             TRUE, TRUE, FALSE);
+        GX2SetPolygonOffset(-2.0f, -2.0f, -2.0f, -2.0f, 0.0f);
+    } else {
+        GX2SetPolygonControl(GX2_FRONT_FACE_CCW, FALSE, FALSE, FALSE,
+                             GX2_POLYGON_MODE_TRIANGLE, GX2_POLYGON_MODE_TRIANGLE,
+                             FALSE, FALSE, FALSE);
+        GX2SetPolygonOffset(0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
+    }
+}
+
+static void gfx_gx2_set_viewport(int x, int y, int width, int height) {
+    uint32_t buffer_height = current_framebuffer->color_buffer.surface.height;
+
+    current_viewport_x = x;
+    current_viewport_y = buffer_height - y - height;
+    current_viewport_width = width;
+    current_viewport_height = height;
+
+    GX2SetViewport(current_viewport_x, current_viewport_y, current_viewport_width, current_viewport_height, 0.0f, 1.0f);
+}
+
+static void gfx_gx2_set_scissor(int x, int y, int width, int height) {
+    uint32_t buffer_height = current_framebuffer->color_buffer.surface.height;
+    uint32_t buffer_width = current_framebuffer->color_buffer.surface.width;
+
+    current_scissor_x = std::min((uint32_t) width, (uint32_t) x);
+    current_scissor_y = std::min((uint32_t) height, buffer_height - y - height);
+    current_scissor_width = std::min((uint32_t) width, buffer_width);
+    current_scissor_height = std::min((uint32_t) height, buffer_height);
+
+    GX2SetScissor(current_scissor_x, current_scissor_y, current_scissor_width, current_scissor_height);
+}
+
+static void gfx_gx2_set_use_alpha(bool use_alpha) {
+    current_use_alpha = use_alpha;
+    GX2SetColorControl(GX2_LOGIC_OP_COPY, use_alpha ? 0xff : 0, FALSE, TRUE);
+}
+
+static void gfx_gx2_draw_triangles(float buf_vbo[], size_t buf_vbo_len, size_t buf_vbo_num_tris) {
+    if (!current_shader_program) {
+        return;
+    }
+
+    size_t vbo_len = sizeof(float) * buf_vbo_len;
+
+    if (draw_ptr + vbo_len >= draw_buffer + DRAW_BUFFER_SIZE) {
+        printf("Waiting on GPU!!!\n");
+        GX2DrawDone();
+        draw_ptr = draw_buffer;
+    }
+
+    float* new_vbo = (float *) draw_ptr;
+    draw_ptr += ALIGN(vbo_len, GX2_VERTEX_BUFFER_ALIGNMENT);
+
+    OSBlockMove(new_vbo, buf_vbo, vbo_len, FALSE);
+    GX2Invalidate(GX2_INVALIDATE_MODE_CPU_ATTRIBUTE_BUFFER, new_vbo, vbo_len);
+
+    GX2SetAttribBuffer(0, vbo_len, current_shader_program->group.stride, new_vbo);
+    GX2DrawEx(GX2_PRIMITIVE_MODE_TRIANGLES, 3 * buf_vbo_num_tris, 0, 1);
+}
+
+static void gfx_gx2_init(void) {
+    // Init the default framebuffer
+    gfx_gx2_init_framebuffer(&main_framebuffer, WIIU_DEFAULT_FB_WIDTH, WIIU_DEFAULT_FB_HEIGHT);
+
+    GX2CalcSurfaceSizeAndAlignment(&main_framebuffer.color_buffer.surface);
+    GX2InitColorBufferRegs(&main_framebuffer.color_buffer);
+
+    main_framebuffer.color_buffer.surface.image = gfx_wiiu_alloc_mem1(main_framebuffer.color_buffer.surface.imageSize, main_framebuffer.color_buffer.surface.alignment);
+    assert(main_framebuffer.color_buffer.surface.image);
+
+    GX2CalcSurfaceSizeAndAlignment(&main_framebuffer.depth_buffer.surface);
+    GX2InitDepthBufferRegs(&main_framebuffer.depth_buffer);
+
+    main_framebuffer.depth_buffer.surface.image = gfx_wiiu_alloc_mem1(main_framebuffer.depth_buffer.surface.imageSize, main_framebuffer.depth_buffer.surface.alignment);
+    assert(main_framebuffer.depth_buffer.surface.image);
+
+    main_framebuffer.imtex.Texture = &main_framebuffer.texture;
+    main_framebuffer.imtex.Sampler = &main_framebuffer.sampler;
+
+    // create a linear aligned copy of the depth buffer to read pixels to
+    memcpy(&depthReadBuffer, &main_framebuffer.depth_buffer, sizeof(GX2DepthBuffer));
+
+    depthReadBuffer.surface.tileMode = GX2_TILE_MODE_LINEAR_ALIGNED;
+    depthReadBuffer.surface.width = 1;
+    depthReadBuffer.surface.height = 1;
+
+    GX2CalcSurfaceSizeAndAlignment(&depthReadBuffer.surface);
+
+    depthReadBuffer.surface.image = gfx_wiiu_alloc_mem1(depthReadBuffer.surface.alignment, depthReadBuffer.surface.imageSize);
+    assert(depthReadBuffer.surface.image);
+    GX2Invalidate(GX2_INVALIDATE_MODE_CPU | GX2_INVALIDATE_MODE_DEPTH_BUFFER, depthReadBuffer.surface.image, depthReadBuffer.surface.imageSize);
+
+    GX2SetColorBuffer(&main_framebuffer.color_buffer, GX2_RENDER_TARGET_0);
+    GX2SetDepthBuffer(&main_framebuffer.depth_buffer);
+
+    current_framebuffer = &main_framebuffer;
+
+    // allocate draw buffer
+    draw_buffer = (uint8_t *) memalign(GX2_VERTEX_BUFFER_ALIGNMENT, DRAW_BUFFER_SIZE);
+    assert(draw_buffer);
+    draw_ptr = draw_buffer;
+
+    GX2SetRasterizerClipControl(TRUE, FALSE);
+
+    GX2SetBlendControl(GX2_RENDER_TARGET_0,
+                       GX2_BLEND_MODE_SRC_ALPHA,
+                       GX2_BLEND_MODE_INV_SRC_ALPHA,
+                       GX2_BLEND_COMBINE_MODE_ADD,
+                       FALSE,
+                       GX2_BLEND_MODE_ZERO,
+                       GX2_BLEND_MODE_ZERO,
+                       GX2_BLEND_COMBINE_MODE_ADD);
+}
+
+void gfx_gx2_shutdown(void) {
+    if (has_foreground) {
+        GX2DrawDone();
+
+        if (depthReadBuffer.surface.image) {
+            gfx_wiiu_free_mem1(depthReadBuffer.surface.image);
+            depthReadBuffer.surface.image = nullptr;
+        }
+
+        if (main_framebuffer.color_buffer.surface.image) {
+            gfx_wiiu_free_mem1(main_framebuffer.color_buffer.surface.image);
+            main_framebuffer.color_buffer.surface.image = nullptr;
+        }
+
+        if (main_framebuffer.depth_buffer.surface.image) {
+            gfx_wiiu_free_mem1(main_framebuffer.depth_buffer.surface.image);
+            main_framebuffer.depth_buffer.surface.image = nullptr;
+        }
+    }
+
+    if (draw_buffer) {
+        free(draw_buffer);
+        draw_buffer = nullptr;
+        draw_ptr = nullptr;
+    }
+}
+
+static void gfx_gx2_on_resize(void) {
+}
+
+static void gfx_gx2_start_frame(void) {
+    // Restore state since ImGui modified it when rendering
+    GX2SetViewport(current_viewport_x, current_viewport_y, current_viewport_width, current_viewport_height, 0.0f, 1.0f);
+    GX2SetScissor(current_scissor_x, current_scissor_y, current_scissor_width, current_scissor_height);
+
+    GX2SetColorControl(GX2_LOGIC_OP_COPY, current_use_alpha ? 0xff : 0, FALSE, TRUE);
+
+    GX2SetBlendControl(GX2_RENDER_TARGET_0,
+                       GX2_BLEND_MODE_SRC_ALPHA,
+                       GX2_BLEND_MODE_INV_SRC_ALPHA,
+                       GX2_BLEND_COMBINE_MODE_ADD,
+                       FALSE,
+                       GX2_BLEND_MODE_ZERO,
+                       GX2_BLEND_MODE_ZERO,
+                       GX2_BLEND_COMBINE_MODE_ADD);
+
+    GX2SetDepthOnlyControl(current_depth_test, current_depth_write, current_depth_compare_function);
+
+    if (current_zmode_decal) {
+        GX2SetPolygonControl(GX2_FRONT_FACE_CCW, FALSE, FALSE, TRUE,
+                             GX2_POLYGON_MODE_TRIANGLE, GX2_POLYGON_MODE_TRIANGLE,
+                             TRUE, TRUE, FALSE);
+        GX2SetPolygonOffset(-2.0f, -2.0f, -2.0f, -2.0f, 0.0f);
+    } else {
+        GX2SetPolygonControl(GX2_FRONT_FACE_CCW, FALSE, FALSE, FALSE,
+                             GX2_POLYGON_MODE_TRIANGLE, GX2_POLYGON_MODE_TRIANGLE,
+                             FALSE, FALSE, FALSE);
+        GX2SetPolygonOffset(0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
+    }
+
+    frame_count++;
+}
+
+static void gfx_gx2_end_frame(void) {
+    draw_ptr = draw_buffer;
+
+    GX2CopyColorBufferToScanBuffer(&main_framebuffer.color_buffer, GX2_SCAN_TARGET_TV);
+    GX2CopyColorBufferToScanBuffer(&main_framebuffer.color_buffer, GX2_SCAN_TARGET_DRC);
+}
+
+static void gfx_gx2_finish_render(void) {
+}
+
+static int gfx_gx2_create_framebuffer(void) {
+    struct Framebuffer *buffer = (struct Framebuffer *) calloc(1, sizeof(struct Framebuffer));
+    assert(buffer);
+
+    GX2InitSampler(&buffer->sampler, GX2_TEX_CLAMP_MODE_WRAP, GX2_TEX_XY_FILTER_MODE_LINEAR);
+
+    buffer->imtex.Texture = &buffer->texture;
+    buffer->imtex.Sampler = &buffer->sampler;
+
+    // some more 32-bit shenanigans :D
+    return (int) buffer;
+}
+
+static void gfx_gx2_update_framebuffer_parameters(int fb, uint32_t width, uint32_t height, uint32_t msaa_level, bool opengl_invert_y, bool render_target, bool has_depth_buffer, bool can_extract_depth) {
+    struct Framebuffer *buffer = (struct Framebuffer *) fb;
+
+    // we don't support updating the main buffer (fb 0)
+    if (!buffer) {
+        return;
+    }
+
+    if (buffer->texture.surface.width == width &&
+        buffer->texture.surface.height == height) {
+        return;
+    }
+
+    // make sure the GPU no longer writes to the buffer
+    GX2DrawDone();
+
+    if (buffer->texture.surface.image) {
+        if (buffer->colorBufferMem1) {
+            gfx_wiiu_free_mem1(buffer->texture.surface.image);
+        } else {
+            free(buffer->texture.surface.image);
+        }
+        buffer->texture.surface.image = nullptr;
+    }
+
+    if (buffer->depth_buffer.surface.image) {
+        if (buffer->depthBufferMem1) {
+            gfx_wiiu_free_mem1(buffer->depth_buffer.surface.image);
+        } else {
+            free(buffer->depth_buffer.surface.image);
+        }
+        buffer->depth_buffer.surface.image = nullptr;
+    }
+
+    gfx_gx2_init_framebuffer(buffer, width, height);
+
+    GX2CalcSurfaceSizeAndAlignment(&buffer->depth_buffer.surface);
+    GX2InitDepthBufferRegs(&buffer->depth_buffer);
+
+    buffer->depth_buffer.surface.image = gfx_wiiu_alloc_mem1(buffer->depth_buffer.surface.imageSize, buffer->depth_buffer.surface.alignment);
+    // fall back to mem2
+    if (!buffer->depth_buffer.surface.image) {
+        buffer->depth_buffer.surface.image = memalign(buffer->depth_buffer.surface.alignment, buffer->depth_buffer.surface.imageSize);
+        buffer->depthBufferMem1 = false;
+    } else {
+        buffer->depthBufferMem1 = true;
+    }
+    assert(buffer->depth_buffer.surface.image);
+
+    GX2CalcSurfaceSizeAndAlignment(&buffer->color_buffer.surface);
+    GX2InitColorBufferRegs(&buffer->color_buffer);
+
+    memset(&buffer->texture, 0, sizeof(GX2Texture));
+    buffer->texture.surface.use = GX2_SURFACE_USE_TEXTURE;
+    buffer->texture.surface.dim = GX2_SURFACE_DIM_TEXTURE_2D;
+    buffer->texture.surface.width = width;
+    buffer->texture.surface.height = height;
+    buffer->texture.surface.depth = 1;
+    buffer->texture.surface.mipLevels = 1;
+    buffer->texture.surface.format = GX2_SURFACE_FORMAT_UNORM_R8_G8_B8_A8;
+    buffer->texture.surface.aa = GX2_AA_MODE1X;
+    buffer->texture.surface.tileMode = GX2_TILE_MODE_DEFAULT;
+    buffer->texture.viewFirstMip = 0;
+    buffer->texture.viewNumMips = 1;
+    buffer->texture.viewFirstSlice = 0;
+    buffer->texture.viewNumSlices = 1;
+    buffer->texture.compMap = GX2_COMP_MAP(GX2_SQ_SEL_R, GX2_SQ_SEL_G, GX2_SQ_SEL_B, GX2_SQ_SEL_A);
+
+    GX2CalcSurfaceSizeAndAlignment(&buffer->texture.surface);
+    GX2InitTextureRegs(&buffer->texture);
+
+    // the texture and color buffer share a buffer
+    assert(buffer->color_buffer.surface.imageSize == buffer->texture.surface.imageSize);
+
+    buffer->texture.surface.image = gfx_wiiu_alloc_mem1(buffer->texture.surface.imageSize, buffer->texture.surface.alignment);
+    // fall back to mem2
+    if (!buffer->texture.surface.image) {
+        buffer->texture.surface.image = memalign(buffer->texture.surface.alignment, buffer->texture.surface.imageSize);
+        buffer->colorBufferMem1 = false;
+    } else {
+        buffer->colorBufferMem1 = true;
+    }
+    assert(buffer->texture.surface.image);
+
+    buffer->color_buffer.surface.image = buffer->texture.surface.image;
+}
+
+void gfx_gx2_start_draw_to_framebuffer(int fb, float noise_scale) {
+    struct Framebuffer *buffer = (struct Framebuffer *) fb;
+
+    // fb 0 = main buffer
+    if (!buffer) {
+        buffer = &main_framebuffer;
+    }
+
+    if (noise_scale != 0.0f) {
+        current_noise_scale = 1.0f / noise_scale;
+    }
+
+    GX2SetColorBuffer(&buffer->color_buffer, GX2_RENDER_TARGET_0);
+    GX2SetDepthBuffer(&buffer->depth_buffer);
+
+    current_framebuffer = buffer;
+}
+
+void gfx_gx2_clear_framebuffer(void) {
+    struct Framebuffer *buffer = current_framebuffer;
+
+    GX2ClearColor(&buffer->color_buffer, 0.0f, 0.0f, 0.0f, 1.0f);
+    GX2ClearDepthStencilEx(&buffer->depth_buffer, 
+        buffer->depth_buffer.depthClear,
+        buffer->depth_buffer.stencilClear, GX2_CLEAR_FLAGS_BOTH);
+    
+    gfx_wiiu_set_context_state();
+}
+
+void gfx_gx2_resolve_msaa_color_buffer(int fb_id_target, int fb_id_source) {
+    struct Framebuffer *src_buffer = (struct Framebuffer *) fb_id_source;
+    struct Framebuffer *target_buffer = (struct Framebuffer *) fb_id_target;
+
+    // fb 0 = main buffer
+    if (!src_buffer) {
+        src_buffer = &main_framebuffer;
+    }
+    if (!target_buffer) {
+        target_buffer = &main_framebuffer;
+    }
+
+    if (src_buffer->color_buffer.surface.aa == GX2_AA_MODE1X) {
+        GX2CopySurface(&src_buffer->color_buffer.surface,
+            src_buffer->color_buffer.viewMip,
+            src_buffer->color_buffer.viewFirstSlice,
+            &target_buffer->color_buffer.surface,
+            target_buffer->color_buffer.viewMip,
+            target_buffer->color_buffer.viewFirstSlice);
+    } else {
+        GX2ResolveAAColorBuffer(&src_buffer->color_buffer,
+            &target_buffer->color_buffer.surface,
+            target_buffer->color_buffer.viewMip,
+            target_buffer->color_buffer.viewFirstSlice);
+    }
+}
+
+void *gfx_gx2_get_framebuffer_texture_id(int fb_id) {
+    struct Framebuffer *buffer = (struct Framebuffer *) fb_id;
+
+    // fb 0 = main buffer
+    if (!buffer) {
+        buffer = &main_framebuffer;
+    }
+
+    return &buffer->imtex;
+}
+
+void gfx_gx2_select_texture_fb(int fb) {
+    struct Framebuffer *buffer = (struct Framebuffer *) fb;
+    assert(buffer);
+
+    assert(current_shader_program);
+    uint32_t location = current_shader_program->samplers_location[0];
+    GX2SetPixelTexture(&buffer->texture, location);
+    GX2SetPixelSampler(&buffer->sampler, location);
+}
+
+static std::unordered_map<std::pair<float, float>, uint16_t, hash_pair_ff> gfx_gx2_get_pixel_depth(int fb_id, const std::set<std::pair<float, float>>& coordinates) {
+    struct Framebuffer *buffer = (struct Framebuffer *) fb_id;
+
+    // fb 0 = main buffer
+    if (!buffer) {
+        buffer = &main_framebuffer;
+    }
+
+    std::unordered_map<std::pair<float, float>, uint16_t, hash_pair_ff> res;
+
+    for (const auto& c : coordinates) {
+        // bug? coordinates sometimes read from oob
+        if ((c.first < 0.0f) || (c.first > (float) buffer->depth_buffer.surface.width)
+            || (c.second < 0.0f) || (c.second > (float) buffer->depth_buffer.surface.height)) {
+            res.emplace(c, 0);
+            continue;
+        }
+
+        GX2Invalidate(GX2_INVALIDATE_MODE_CPU | GX2_INVALIDATE_MODE_DEPTH_BUFFER, depthReadBuffer.surface.image, depthReadBuffer.surface.imageSize);
+
+        // copy the pixel to the depthReadBuffer
+        GX2Rect srcRect = { 
+            (int32_t) c.first,
+            (int32_t) buffer->depth_buffer.surface.height - (int32_t) c.second,
+            (int32_t) c.first + 1,
+            (int32_t) (buffer->depth_buffer.surface.height - (int32_t) c.second) + 1
+        };
+        GX2Point dstPoint = { 0, 0 };
+        GX2CopySurfaceEx(&buffer->depth_buffer.surface, 0, 0, &depthReadBuffer.surface, 0, 0, 1, &srcRect, &dstPoint);
+        GX2DrawDone();
+
+        gfx_wiiu_set_context_state();
+
+        // read the pixel from the depthReadBuffer
+        uint32_t tmp = __builtin_bswap32(*(uint32_t *)depthReadBuffer.surface.image);
+        float val = *(float *)&tmp;
+
+        res.emplace(c, val * 65532.0f);
+    }
+
+    return res;
+}
+
+void gfx_gx2_set_texture_filter(FilteringMode mode) {
+    // three-point is not implemented in the shaders yet
+    if (mode == FILTER_THREE_POINT) {
+        mode = FILTER_LINEAR;
+    }
+
+    current_filter_mode = mode;
+    gfx_texture_cache_clear();
+}
+
+FilteringMode gfx_gx2_get_texture_filter(void) {
+    return current_filter_mode;
+}
+
+ImGui_ImplGX2_Texture* gfx_gx2_texture_for_imgui(uint32_t texture_id) {
+    struct Texture *tex = (struct Texture *) texture_id;
+    return &tex->imtex;
+}
+
+struct GfxRenderingAPI gfx_gx2_api = {
+    gfx_gx2_get_clip_parameters,
+    gfx_gx2_unload_shader,
+    gfx_gx2_load_shader,
+    gfx_gx2_create_and_load_new_shader,
+    gfx_gx2_lookup_shader,
+    gfx_gx2_shader_get_info,
+    gfx_gx2_new_texture,
+    gfx_gx2_select_texture,
+    gfx_gx2_upload_texture,
+    gfx_gx2_set_sampler_parameters,
+    gfx_gx2_set_depth_test_and_mask,
+    gfx_gx2_set_zmode_decal,
+    gfx_gx2_set_viewport,
+    gfx_gx2_set_scissor,
+    gfx_gx2_set_use_alpha,
+    gfx_gx2_draw_triangles,
+    gfx_gx2_init,
+    gfx_gx2_on_resize,
+    gfx_gx2_start_frame,
+    gfx_gx2_end_frame,
+    gfx_gx2_finish_render,
+    gfx_gx2_create_framebuffer,
+    gfx_gx2_update_framebuffer_parameters,
+    gfx_gx2_start_draw_to_framebuffer,
+    gfx_gx2_clear_framebuffer,
+    gfx_gx2_resolve_msaa_color_buffer,
+    gfx_gx2_get_pixel_depth,
+    gfx_gx2_get_framebuffer_texture_id,
+    gfx_gx2_select_texture_fb,
+    gfx_gx2_delete_texture,
+    gfx_gx2_set_texture_filter,
+    gfx_gx2_get_texture_filter
+};
+
+#endif

libultraship/libultraship/Lib/Fast3D/gfx_gx2.h

@@ -0,0 +1,12 @@
+#ifndef GFX_GX2_H
+#define GFX_GX2_H
+
+#include "gfx_rendering_api.h"
+
+void gfx_gx2_shutdown(void);
+
+struct ImGui_ImplGX2_Texture* gfx_gx2_texture_for_imgui(uint32_t texture_id);
+
+extern struct GfxRenderingAPI gfx_gx2_api;
+
+#endif

libultraship/libultraship/Lib/Fast3D/gfx_pc.cpp

@@ -1382,12 +1382,25 @@ static void gfx_sp_tri1(uint8_t vtx1_idx, uint8_t vtx2_idx, uint8_t vtx3_idx, bo
             buf_vbo[buf_vbo_len++] = u / tex_width[t];
             buf_vbo[buf_vbo_len++] = v / tex_height[t];
 
-            if (tm & (1 << 2 * t)) {
+            bool clampS = tm & (1 << 2 * t);
+            bool clampT = tm & (1 << 2 * t + 1);
+
+            if (clampS) {
                 buf_vbo[buf_vbo_len++] = (tex_width2[t] - 0.5f) / tex_width[t];
             }
-            if (tm & (1 << 2 * t + 1)) {
+#ifdef __WIIU__
+            else {
+                buf_vbo[buf_vbo_len++] = 0.0f;
+            }
+#endif
+            if (clampT) {
                 buf_vbo[buf_vbo_len++] = (tex_height2[t] - 0.5f) / tex_height[t];
             }
+#ifdef __WIIU__
+            else {
+                buf_vbo[buf_vbo_len++] = 0.0f;
+            }
+#endif
         }
 
         if (use_fog) {
@@ -1464,6 +1477,12 @@ static void gfx_sp_tri1(uint8_t vtx1_idx, uint8_t vtx2_idx, uint8_t vtx3_idx, bo
                     buf_vbo[buf_vbo_len++] = color->r / 255.0f;
                     buf_vbo[buf_vbo_len++] = color->g / 255.0f;
                     buf_vbo[buf_vbo_len++] = color->b / 255.0f;
+#ifdef __WIIU__
+                    // padding
+                    if (!use_alpha) {
+                        buf_vbo[buf_vbo_len++] = 1.0f;
+                    }
+#endif
                 }
                 else {
                     if (use_fog && color == &v_arr[i]->color) {
@@ -2658,8 +2677,10 @@ void gfx_init(struct GfxWindowManagerAPI *wapi, struct GfxRenderingAPI *rapi, co
     gfx_current_dimensions.internal_mul = CVar_GetFloat("gInternalResolution", 1);
 #endif
     gfx_msaa_level = CVar_GetS32("gMSAAValue", 1);
+#ifndef __WIIU__ // Wii U overrides dimentions in gfx_wapi->init to match framebuffer size
     gfx_current_dimensions.width = width;
     gfx_current_dimensions.height = height;
+#endif
     game_framebuffer = gfx_rapi->create_framebuffer();
     game_framebuffer_msaa_resolved = gfx_rapi->create_framebuffer();
 

libultraship/libultraship/Lib/Fast3D/gfx_wiiu.cpp

@@ -0,0 +1,467 @@
+#ifdef __WIIU__
+
+#include <stdio.h>
+#include <time.h>
+#include <malloc.h>
+
+#include <coreinit/time.h>
+#include <coreinit/thread.h>
+#include <coreinit/foreground.h>
+#include <coreinit/memory.h>
+#include <coreinit/memheap.h>
+#include <coreinit/memdefaultheap.h>
+#include <coreinit/memexpheap.h>
+#include <coreinit/memfrmheap.h>
+
+#include <gx2/state.h>
+#include <gx2/context.h>
+#include <gx2/display.h>
+#include <gx2/event.h>
+#include <gx2/swap.h>
+#include <gx2/mem.h>
+#include <gx2r/mem.h>
+
+#include <whb/proc.h>
+#include <proc_ui/procui.h>
+#include <proc_ui/memory.h>
+
+#include <vpad/input.h>
+#include <padscore/kpad.h>
+
+#ifndef _LANGUAGE_C
+#define _LANGUAGE_C
+#endif
+#include "PR/ultra64/gbi.h"
+
+#include "gfx_window_manager_api.h"
+#include "gfx_pc.h"
+#include "gfx_gx2.h"
+#include "gfx_wiiu.h"
+
+#include "Lib/ImGui/backends/wiiu/imgui_impl_wiiu.h"
+#include "../../WiiUImpl.h"
+#include "../../ImGuiImpl.h"
+#include "../../Hooks.h"
+
+static MEMHeapHandle heap_MEM1 = nullptr;
+static MEMHeapHandle heap_foreground = nullptr;
+
+bool has_foreground = false;
+static void *mem1_storage = nullptr;
+static void *command_buffer_pool = nullptr;
+GX2ContextState *context_state = nullptr;
+
+static GX2TVRenderMode tv_render_mode;
+static void *tv_scan_buffer = nullptr;
+static uint32_t tv_scan_buffer_size = 0;
+static uint32_t tv_width;
+static uint32_t tv_height;
+
+static GX2DrcRenderMode drc_render_mode;
+static void *drc_scan_buffer = nullptr;
+static uint32_t drc_scan_buffer_size = 0;
+
+static int frame_divisor = 1;
+
+// for ImGui DeltaTime
+// (initialized to 1 to not trigger imguis assert on initial draw)
+uint32_t frametime = 1;
+
+bool gfx_wiiu_init_mem1(void) {
+    MEMHeapHandle heap = MEMGetBaseHeapHandle(MEM_BASE_HEAP_MEM1);
+    uint32_t size;
+    void *base;
+
+    size = MEMGetAllocatableSizeForFrmHeapEx(heap, 4);
+    if (!size) {
+        printf("%s: MEMGetAllocatableSizeForFrmHeapEx == 0", __FUNCTION__);
+        return false;
+    }
+
+    base = MEMAllocFromFrmHeapEx(heap, size, 4);
+    if (!base) {
+        printf("%s: MEMAllocFromFrmHeapEx(heap, 0x%X, 4) failed", __FUNCTION__, size);
+        return false;
+    }
+
+    heap_MEM1 = MEMCreateExpHeapEx(base, size, 0);
+    if (!heap_MEM1) {
+        printf("%s: MEMCreateExpHeapEx(%p, 0x%X, 0) failed", __FUNCTION__, base, size);
+        return false;
+    }
+
+    return true;
+}
+
+void gfx_wiiu_destroy_mem1(void) {
+    MEMHeapHandle heap = MEMGetBaseHeapHandle(MEM_BASE_HEAP_MEM1);
+
+    if (heap_MEM1) {
+        MEMDestroyExpHeap(heap_MEM1);
+        heap_MEM1 = NULL;
+    }
+}
+
+bool gfx_wiiu_init_foreground(void) {
+    MEMHeapHandle heap = MEMGetBaseHeapHandle(MEM_BASE_HEAP_FG);
+    uint32_t size;
+    void *base;
+
+    size = MEMGetAllocatableSizeForFrmHeapEx(heap, 4);
+    if (!size) {
+        printf("%s: MEMAllocFromFrmHeapEx(heap, 0x%X, 4)", __FUNCTION__, size);
+        return false;
+    }
+
+    base = MEMAllocFromFrmHeapEx(heap, size, 4);
+    if (!base) {
+        printf("%s: MEMGetAllocatableSizeForFrmHeapEx == 0", __FUNCTION__);
+        return false;
+    }
+
+    heap_foreground = MEMCreateExpHeapEx(base, size, 0);
+    if (!heap_foreground) {
+        printf("%s: MEMCreateExpHeapEx(%p, 0x%X, 0)", __FUNCTION__, base, size);
+        return false;
+    }
+
+    return true;
+}
+
+void gfx_wiiu_destroy_foreground(void) {
+    MEMHeapHandle foreground = MEMGetBaseHeapHandle(MEM_BASE_HEAP_FG);
+
+    if (heap_foreground) {
+        MEMDestroyExpHeap(heap_foreground);
+        heap_foreground = NULL;
+    }
+
+    MEMFreeToFrmHeap(foreground, MEM_FRM_HEAP_FREE_ALL);
+}
+
+void *gfx_wiiu_alloc_mem1(uint32_t size, uint32_t alignment) {
+    void *block;
+
+    if (!heap_MEM1) {
+        return NULL;
+    }
+
+    if (alignment < 4) {
+        alignment = 4;
+    }
+
+    block = MEMAllocFromExpHeapEx(heap_MEM1, size, alignment);
+    return block;
+}
+
+void gfx_wiiu_free_mem1(void *block) {
+    if (!heap_MEM1) {
+        return;
+    }
+
+    MEMFreeToExpHeap(heap_MEM1, block);
+}
+
+void *gfx_wiiu_alloc_foreground(uint32_t size, uint32_t alignment) {
+    void *block;
+
+    if (!heap_foreground) {
+        return NULL;
+    }
+
+    if (alignment < 4) {
+        alignment = 4;
+    }
+
+    block = MEMAllocFromExpHeapEx(heap_foreground, size, alignment);
+    return block;
+}
+
+void gfx_wiiu_free_foreground(void *block) {
+    if (!heap_foreground) {
+        return;
+    }
+
+    MEMFreeToExpHeap(heap_foreground, block);
+}
+
+static uint32_t gfx_wiiu_proc_callback_acquired(void *context) {
+    has_foreground = true;
+
+    bool result = gfx_wiiu_init_foreground();
+    assert(result);
+
+    tv_scan_buffer = gfx_wiiu_alloc_foreground(tv_scan_buffer_size, GX2_SCAN_BUFFER_ALIGNMENT);
+    assert(tv_scan_buffer);
+
+    GX2Invalidate(GX2_INVALIDATE_MODE_CPU, tv_scan_buffer, tv_scan_buffer_size);
+    GX2SetTVBuffer(tv_scan_buffer, tv_scan_buffer_size, tv_render_mode, GX2_SURFACE_FORMAT_UNORM_R8_G8_B8_A8, GX2_BUFFERING_MODE_DOUBLE);
+
+    drc_scan_buffer = gfx_wiiu_alloc_foreground(drc_scan_buffer_size, GX2_SCAN_BUFFER_ALIGNMENT);
+    assert(drc_scan_buffer);
+
+    GX2Invalidate(GX2_INVALIDATE_MODE_CPU, drc_scan_buffer, drc_scan_buffer_size);
+    GX2SetDRCBuffer(drc_scan_buffer, drc_scan_buffer_size, drc_render_mode, GX2_SURFACE_FORMAT_UNORM_R8_G8_B8_A8, GX2_BUFFERING_MODE_DOUBLE);
+
+    return 0;
+}
+
+static uint32_t gfx_wiiu_proc_callback_released(void* context) {
+    if (tv_scan_buffer) {
+        gfx_wiiu_free_foreground(tv_scan_buffer);
+        tv_scan_buffer = nullptr;
+    }
+
+    if (drc_scan_buffer) {
+        gfx_wiiu_free_foreground(drc_scan_buffer);
+        drc_scan_buffer = nullptr;
+    }
+
+    gfx_wiiu_destroy_foreground();
+
+    has_foreground = false;
+
+    return 0;
+}
+
+static void gfx_wiiu_init(const char *game_name, bool start_in_fullscreen, uint32_t width, uint32_t height) {
+    WHBProcInit();
+
+    uint32_t mem1_addr, mem1_size;
+    OSGetMemBound(OS_MEM1, &mem1_addr, &mem1_size);
+    mem1_storage = memalign(0x40, mem1_size);
+    assert(mem1_storage);
+
+    ProcUISetMEM1Storage(mem1_storage, mem1_size);
+
+    bool result = gfx_wiiu_init_mem1();
+    assert(result);
+
+    command_buffer_pool = memalign(GX2_COMMAND_BUFFER_ALIGNMENT, 0x400000);
+    assert(command_buffer_pool);
+
+    uint32_t initAttribs[] = {
+        GX2_INIT_CMD_BUF_BASE, (uintptr_t) command_buffer_pool,
+        GX2_INIT_CMD_BUF_POOL_SIZE, 0x400000,
+        GX2_INIT_ARGC, 0,
+        GX2_INIT_ARGV, 0,
+        GX2_INIT_END
+    };
+    GX2Init(initAttribs);
+
+    switch(GX2GetSystemTVScanMode()) {
+    case GX2_TV_SCAN_MODE_480I:
+    case GX2_TV_SCAN_MODE_480P:
+        tv_render_mode = GX2_TV_RENDER_MODE_WIDE_480P;
+        tv_width = 854;
+        tv_height = 480;
+        break;
+    case GX2_TV_SCAN_MODE_1080I:
+    case GX2_TV_SCAN_MODE_1080P:
+        tv_render_mode = GX2_TV_RENDER_MODE_WIDE_1080P;
+        tv_width = 1920;
+        tv_height = 1080;
+        break;
+    case GX2_TV_SCAN_MODE_720P:
+    default:
+        tv_render_mode = GX2_TV_RENDER_MODE_WIDE_720P;
+        tv_width = 1280;
+        tv_height = 720;
+        break;
+    }
+
+    drc_render_mode = GX2GetSystemDRCScanMode();
+
+    uint32_t unk;
+    GX2CalcTVSize(tv_render_mode, GX2_SURFACE_FORMAT_UNORM_R8_G8_B8_A8, GX2_BUFFERING_MODE_DOUBLE, &tv_scan_buffer_size, &unk);
+    GX2CalcDRCSize(drc_render_mode, GX2_SURFACE_FORMAT_UNORM_R8_G8_B8_A8, GX2_BUFFERING_MODE_DOUBLE, &drc_scan_buffer_size, &unk);
+
+    ProcUIRegisterCallback(PROCUI_CALLBACK_ACQUIRE, gfx_wiiu_proc_callback_acquired, nullptr, 100);
+    ProcUIRegisterCallback(PROCUI_CALLBACK_RELEASE, gfx_wiiu_proc_callback_released, nullptr, 100);
+
+    gfx_wiiu_proc_callback_acquired(nullptr);
+
+    context_state = (GX2ContextState *) memalign(GX2_CONTEXT_STATE_ALIGNMENT, sizeof(GX2ContextState));
+    assert(context_state);
+
+    GX2SetupContextStateEx(context_state, TRUE);
+    GX2SetContextState(context_state);
+
+    GX2SetTVScale(WIIU_DEFAULT_FB_WIDTH, WIIU_DEFAULT_FB_HEIGHT);
+    GX2SetDRCScale(WIIU_DEFAULT_FB_WIDTH, WIIU_DEFAULT_FB_HEIGHT);
+
+    GX2SetSwapInterval(frame_divisor);
+
+    gfx_current_dimensions.width = gfx_current_game_window_viewport.width = WIIU_DEFAULT_FB_WIDTH;
+    gfx_current_dimensions.height = gfx_current_game_window_viewport.height = WIIU_DEFAULT_FB_HEIGHT;
+
+    SohImGui::WindowImpl window_impl;
+    window_impl.backend = SohImGui::Backend::GX2;
+    window_impl.gx2.width = WIIU_DEFAULT_FB_WIDTH;
+    window_impl.gx2.height = WIIU_DEFAULT_FB_HEIGHT;
+    SohImGui::Init(window_impl);
+}
+
+static void gfx_wiiu_shutdown(void) {
+    if (has_foreground) {
+        gfx_wiiu_proc_callback_released(nullptr);
+        gfx_wiiu_destroy_mem1();
+    }
+
+    GX2Shutdown();
+
+    if (context_state) {
+        free(context_state);
+        context_state = nullptr;
+    }
+
+    if (command_buffer_pool) {
+        free(command_buffer_pool);
+        command_buffer_pool = nullptr;
+    }
+
+    ProcUISetMEM1Storage(nullptr, 0);
+    free(mem1_storage);
+}
+
+void gfx_wiiu_set_context_state(void) {
+    GX2SetContextState(context_state);
+}
+
+static void gfx_wiiu_set_fullscreen_changed_callback(void (*on_fullscreen_changed)(bool is_now_fullscreen)) {
+}
+
+static void gfx_wiiu_set_fullscreen(bool enable) {
+}
+
+static void gfx_wiiu_show_cursor(bool hide) {
+}
+
+static void gfx_wiiu_set_keyboard_callbacks(bool (*on_key_down)(int scancode), bool (*on_key_up)(int scancode), void (*on_all_keys_up)(void)) {
+}
+
+static void gfx_wiiu_main_loop(void (*run_one_game_iter)(void)) {
+    while (WHBProcIsRunning()) {
+        run_one_game_iter();
+    }
+
+    Ship::ExecuteHooks<Ship::ExitGame>();
+    Ship::WiiU::Exit();
+
+    gfx_gx2_shutdown();
+    gfx_wiiu_shutdown();
+    WHBProcShutdown();
+}
+
+static void gfx_wiiu_get_dimensions(uint32_t *width, uint32_t *height) {
+    *width = WIIU_DEFAULT_FB_WIDTH;
+    *height = WIIU_DEFAULT_FB_HEIGHT;
+}
+
+static void gfx_wiiu_handle_events(void) {
+    Ship::WiiU::Update();
+
+    ImGui_ImplWiiU_ControllerInput input{};
+
+    VPADReadError vpad_error;
+    input.vpad = Ship::WiiU::GetVPADStatus(&vpad_error);
+    if (vpad_error != VPAD_READ_SUCCESS) {
+        input.vpad = nullptr;
+    }
+
+    KPADError kpad_error;
+    for (int i = 0; i < 4; i++) {
+        input.kpad[i] = Ship::WiiU::GetKPADStatus((WPADChan) i, &kpad_error);
+        if (kpad_error != KPAD_ERROR_OK) {
+            input.kpad[i] = nullptr;
+        }
+    }
+
+    SohImGui::EventImpl event_impl;
+    event_impl.gx2.input = &input;
+    SohImGui::Update(event_impl);
+}
+
+static bool gfx_wiiu_start_frame(void) {
+    uint32_t swap_count, flip_count;
+    OSTime last_flip, last_vsync;
+    uint32_t wait_count = 0;
+
+    while (true) {
+        GX2GetSwapStatus(&swap_count, &flip_count, &last_flip, &last_vsync);
+
+        if (flip_count >= swap_count) {
+            break;
+        }
+
+        if (wait_count >= 10) {
+            // GPU timed out, drop frame
+            return false;
+        }
+
+        wait_count++;
+        GX2WaitForVsync();
+    }
+
+    return true;
+}
+
+static void gfx_wiiu_swap_buffers_begin(void) {
+    GX2SwapScanBuffers();
+    GX2Flush();
+
+    gfx_wiiu_set_context_state();
+
+    GX2SetTVEnable(TRUE);
+    GX2SetDRCEnable(TRUE);
+}
+
+static void gfx_wiiu_swap_buffers_end(void) {
+    static OSTick tick = 0;
+    frametime = OSTicksToMicroseconds(OSGetSystemTick() - tick);
+    tick = OSGetSystemTick();
+}
+
+static double gfx_wiiu_get_time(void) {
+    return 0.0;
+}
+
+static void gfx_wiiu_set_target_fps(int fps) {
+    // use the nearest divisor
+    int divisor = 60 / fps;
+    if (divisor < 1) {
+        divisor = 1;
+    }
+
+    if (frame_divisor != divisor) {
+        GX2SetSwapInterval(divisor);
+        frame_divisor = divisor;
+    }
+}
+
+static void gfx_wiiu_set_maximum_frame_latency(int latency) {
+}
+
+static float gfx_wiiu_get_detected_hz(void) {
+    return 0;
+}
+
+struct GfxWindowManagerAPI gfx_wiiu = {
+    gfx_wiiu_init,
+    gfx_wiiu_set_keyboard_callbacks,
+    gfx_wiiu_set_fullscreen_changed_callback,
+    gfx_wiiu_set_fullscreen,
+    gfx_wiiu_show_cursor,
+    gfx_wiiu_main_loop,
+    gfx_wiiu_get_dimensions,
+    gfx_wiiu_handle_events,
+    gfx_wiiu_start_frame,
+    gfx_wiiu_swap_buffers_begin,
+    gfx_wiiu_swap_buffers_end,
+    gfx_wiiu_get_time,
+    gfx_wiiu_set_target_fps,
+    gfx_wiiu_set_maximum_frame_latency,
+    gfx_wiiu_get_detected_hz,
+};
+
+#endif

libultraship/libultraship/Lib/Fast3D/gfx_wiiu.h

@@ -0,0 +1,38 @@
+#ifndef GFX_WIIU_H
+#define GFX_WIIU_H
+#ifdef __WIIU__
+
+#include <vpad/input.h>
+#include <padscore/kpad.h>
+
+#include "gfx_window_manager_api.h"
+
+// make the default fb always 1080p to not mess with scaling
+#define WIIU_DEFAULT_FB_WIDTH  1920
+#define WIIU_DEFAULT_FB_HEIGHT 1080
+
+extern bool has_foreground;
+extern uint32_t frametime;
+
+bool gfx_wiiu_init_mem1(void);
+
+void gfx_wiiu_destroy_mem1(void);
+
+bool gfx_wiiu_init_foreground(void);
+
+void gfx_wiiu_destroy_foreground(void);
+
+void *gfx_wiiu_alloc_mem1(uint32_t size, uint32_t alignment);
+
+void gfx_wiiu_free_mem1(void *block);
+
+void *gfx_wiiu_alloc_foreground(uint32_t size, uint32_t alignment);
+
+void gfx_wiiu_free_foreground(void *block);
+
+void gfx_wiiu_set_context_state(void);
+
+extern struct GfxWindowManagerAPI gfx_wiiu;
+
+#endif
+#endif

libultraship/libultraship/Lib/Fast3D/gx2_shader_gen.c

@@ -0,0 +1,810 @@
+#ifdef __WIIU__
+
+#include "gx2_shader_gen.h"
+#include "gx2_shader_inl.h"
+
+#include <malloc.h>
+#include <gx2/mem.h>
+
+#define ROUNDUP(x, align) (((x) + ((align) -1)) & ~((align) -1))
+
+#define FRAG_COORD_REG _R0
+#define TEXEL_REG      _R1
+#define FOG_REG        _R3
+#define GRAYSCALE_REG  _R4
+
+enum {
+    SHADER_TEXINFO0 = SHADER_COMBINED + 1,
+    SHADER_TEXINFO1,
+};
+
+static uint8_t get_reg(struct CCFeatures *cc_features, uint8_t c) {
+    if (c == SHADER_0) {
+        return ALU_SRC_0;
+    }
+    if (c == SHADER_1) {
+        return ALU_SRC_1;
+    }
+
+    if (c == SHADER_COMBINED) {
+        return TEXEL_REG;
+    }
+
+    if (c >= SHADER_INPUT_1 && c <= SHADER_INPUT_7) {
+        return _R(5 + (c - SHADER_INPUT_1));
+    }
+
+    uint8_t input_last = (cc_features->num_inputs + 5) - 1;
+
+    if (c == SHADER_TEXEL0 || c == SHADER_TEXEL0A) {
+        // reuse unused regs
+        if (!cc_features->opt_noise) {
+            return FRAG_COORD_REG;
+        } else if (!cc_features->opt_fog) {
+            return FOG_REG;
+        } else if (!cc_features->opt_grayscale) {
+            return GRAYSCALE_REG;
+        }
+
+        return _R(input_last + 1);
+    }
+    if (c == SHADER_TEXEL1 || c == SHADER_TEXEL1A) {
+        // if the shader doesn't use texture 0 we can reuse it for texture 1
+        if (!cc_features->used_textures[0]) {
+            return get_reg(cc_features, SHADER_TEXEL0);
+        }
+        // reuse unused regs which tex 0 doesn't use yet
+        else if (!cc_features->opt_fog && get_reg(cc_features, SHADER_TEXEL0) != FOG_REG) {
+            return FOG_REG;
+        } else if (!cc_features->opt_grayscale && get_reg(cc_features, SHADER_TEXEL0) != GRAYSCALE_REG) {
+            return GRAYSCALE_REG;
+        }
+
+        return _R(input_last + 2);
+    }
+
+    // reuse the regs above
+    if (c == SHADER_TEXINFO0) {
+        return get_reg(cc_features, SHADER_TEXEL0);
+    }
+    if (c == SHADER_TEXINFO1) {
+        return get_reg(cc_features, SHADER_TEXEL1);
+    }
+
+    return 0;
+}
+
+static uint8_t get_num_regs(struct CCFeatures *cc_features) {
+    uint8_t input_count = cc_features->num_inputs + 5;
+
+    uint8_t last_tex_reg;
+    if (cc_features->used_textures[1]) {
+        last_tex_reg = get_reg(cc_features, SHADER_TEXEL1) + 1;
+    } else if (cc_features->used_textures[0]) {
+        last_tex_reg = get_reg(cc_features, SHADER_TEXEL0) + 1;
+    } else {
+        return input_count;
+    }
+    
+    return (last_tex_reg < input_count) ? input_count : last_tex_reg;
+}
+
+#define ADD_INSTR(...) \
+    uint64_t tmp[] = {__VA_ARGS__}; \
+    memcpy(*alu_ptr, tmp, sizeof(tmp)); \
+    *alu_ptr += sizeof(tmp) / sizeof(uint64_t)
+
+static inline void add_tex_clamp_S_T(struct CCFeatures *cc_features, uint64_t **alu_ptr, uint8_t tex) {
+    uint8_t texinfo_reg = get_reg(cc_features, (tex == 0) ? SHADER_TEXINFO0 : SHADER_TEXINFO1);
+    uint8_t texcoord_reg = (tex == 0) ? _R1 : _R2;
+
+    ADD_INSTR(
+        /* R127.xy = (float) texinfo.xy */
+        ALU_INT_TO_FLT(_R127, _x, texinfo_reg, _x) SCL_210
+        ALU_LAST,
+
+        ALU_INT_TO_FLT(_R127, _y, texinfo_reg, _y) SCL_210
+        ALU_LAST,
+
+        /* R127.xy = 0.5f / texSize */
+        ALU_RECIP_IEEE(__, _x, _R127, _x) SCL_210
+        ALU_LAST,
+
+        ALU_MUL_IEEE(_R127, _x, ALU_SRC_PS, _x, ALU_SRC_0_5, _x),
+        ALU_RECIP_IEEE(__, _y, _R127, _y) SCL_210
+        ALU_LAST,
+
+        ALU_MUL_IEEE(_R127, _y, ALU_SRC_PS, _y, ALU_SRC_0_5, _x)
+        ALU_LAST,
+
+        /* texCoord.xy = clamp(texCoord.xy, R127.xy, texClamp.xy) */
+        ALU_MAX(__, _x, texcoord_reg, _x, _R127, _x),
+        ALU_MAX(__, _y, texcoord_reg, _y, _R127, _y)
+        ALU_LAST,
+
+        ALU_MIN(texcoord_reg, _x, ALU_SRC_PV, _x, texcoord_reg, _z),
+        ALU_MIN(texcoord_reg, _y, ALU_SRC_PV, _y, texcoord_reg, _w)
+        ALU_LAST,
+    );
+}
+
+static inline void add_tex_clamp_S(struct CCFeatures *cc_features, uint64_t **alu_ptr, uint8_t tex) {
+    uint8_t texinfo_reg = get_reg(cc_features, (tex == 0) ? SHADER_TEXINFO0 : SHADER_TEXINFO1);
+    uint8_t texcoord_reg = (tex == 0) ? _R1 : _R2;
+
+    ADD_INSTR(
+        /* R127.x = (float) texinfo.x */
+        ALU_INT_TO_FLT(_R127, _x, texinfo_reg, _x) SCL_210
+        ALU_LAST,
+
+        /* R127.x = 0.5f / texSize */
+        ALU_RECIP_IEEE(__, _x, _R127, _x) SCL_210
+        ALU_LAST,
+
+        ALU_MUL_IEEE(_R127, _x, ALU_SRC_PS, _x, ALU_SRC_0_5, _x)
+        ALU_LAST,
+
+        /* texCoord.xy = clamp(texCoord.xy, R127.xy, texClamp.xy) */
+        ALU_MAX(__, _x, texcoord_reg, _x, _R127, _x)
+        ALU_LAST,
+
+        ALU_MIN(texcoord_reg, _x, ALU_SRC_PV, _x, texcoord_reg, _z)
+        ALU_LAST,
+    );
+}
+
+static inline void add_tex_clamp_T(struct CCFeatures *cc_features, uint64_t **alu_ptr, uint8_t tex) {
+    uint8_t texinfo_reg = get_reg(cc_features, (tex == 0) ? SHADER_TEXINFO0 : SHADER_TEXINFO1);
+    uint8_t texcoord_reg = (tex == 0) ? _R1 : _R2;
+
+    ADD_INSTR(
+        /* R127.y = (float) texinfo.y */
+        ALU_INT_TO_FLT(_R127, _y, texinfo_reg, _y) SCL_210
+        ALU_LAST,
+
+        /* R127.y = 0.5f / texSize */
+        ALU_RECIP_IEEE(__, _x, _R127, _y) SCL_210
+        ALU_LAST,
+
+        ALU_MUL_IEEE(_R127, _y, ALU_SRC_PS, _x, ALU_SRC_0_5, _x)
+        ALU_LAST,
+
+        /* texCoord.xy = clamp(texCoord.xy, R127.xy, texClamp.xy) */
+        ALU_MAX(__, _y, texcoord_reg, _y, _R127, _y)
+        ALU_LAST,
+
+        ALU_MIN(texcoord_reg, _y, ALU_SRC_PV, _y, texcoord_reg, _w)
+        ALU_LAST,
+    );
+}
+
+static inline void add_mov(struct CCFeatures *cc_features, uint64_t **alu_ptr, uint8_t src, bool single) {
+    bool src_alpha = (src == SHADER_TEXEL0A) || (src == SHADER_TEXEL1A);
+    src = get_reg(cc_features, src);
+
+    /* texel = src */
+    if (single) {
+        ADD_INSTR(
+            ALU_MOV(TEXEL_REG, _w, src, _w)
+            ALU_LAST,
+        );
+    } else {
+        ADD_INSTR(
+            ALU_MOV(TEXEL_REG, _x, src, src_alpha ? _w :_x),
+            ALU_MOV(TEXEL_REG, _y, src, src_alpha ? _w :_y),
+            ALU_MOV(TEXEL_REG, _z, src, src_alpha ? _w :_z)
+            ALU_LAST,
+        );
+    }
+}
+
+static inline void add_mul(struct CCFeatures *cc_features, uint64_t **alu_ptr, uint8_t src0, uint8_t src1, bool single) {
+    bool src0_alpha = (src0 == SHADER_TEXEL0A) || (src0 == SHADER_TEXEL1A);
+    bool src1_alpha = (src1 == SHADER_TEXEL0A) || (src1 == SHADER_TEXEL1A);
+    src0 = get_reg(cc_features, src0);
+    src1 = get_reg(cc_features, src1);
+
+    /* texel = src0 * src1 */
+    if (single) {
+        ADD_INSTR(
+            ALU_MUL(TEXEL_REG, _w, src0, _w, src1, _w)
+            ALU_LAST,
+        );
+    } else {
+        ADD_INSTR(
+            ALU_MUL(TEXEL_REG, _x, src0, src0_alpha ? _w : _x, src1, src1_alpha ? _w : _x),
+            ALU_MUL(TEXEL_REG, _y, src0, src0_alpha ? _w : _y, src1, src1_alpha ? _w : _y),
+            ALU_MUL(TEXEL_REG, _z, src0, src0_alpha ? _w : _z, src1, src1_alpha ? _w : _z)
+            ALU_LAST,
+        );
+    }
+}
+
+static inline void add_mix(struct CCFeatures *cc_features, uint64_t **alu_ptr, uint8_t src0, uint8_t src1, uint8_t src2, uint8_t src3, bool single) {
+    bool src0_alpha = (src0 == SHADER_TEXEL0A) || (src0 == SHADER_TEXEL1A);
+    bool src1_alpha = (src1 == SHADER_TEXEL0A) || (src1 == SHADER_TEXEL1A);
+    bool src2_alpha = (src2 == SHADER_TEXEL0A) || (src2 == SHADER_TEXEL1A);
+    bool src3_alpha = (src3 == SHADER_TEXEL0A) || (src3 == SHADER_TEXEL1A);
+    src0 = get_reg(cc_features, src0);
+    src1 = get_reg(cc_features, src1);
+    src2 = get_reg(cc_features, src2);
+    src3 = get_reg(cc_features, src3);
+
+    /* texel = (src0 - src1) * src2 - src3 */
+    if (single) {
+        ADD_INSTR(
+            ALU_ADD(__, _w, src0, _w, src1 _NEG, _w)
+            ALU_LAST,
+
+            ALU_MULADD(TEXEL_REG, _w, ALU_SRC_PV, _w, src2, _w, src3, _w)
+            ALU_LAST,
+        );
+    } else {
+        ADD_INSTR(
+            ALU_ADD(__, _x, src0, src0_alpha ? _w : _x, src1 _NEG, src1_alpha ? _w : _x),
+            ALU_ADD(__, _y, src0, src0_alpha ? _w : _y, src1 _NEG, src1_alpha ? _w : _y),
+            ALU_ADD(__, _z, src0, src0_alpha ? _w : _z, src1 _NEG, src1_alpha ? _w : _z)
+            ALU_LAST,
+
+            ALU_MULADD(TEXEL_REG, _x, ALU_SRC_PV, _x, src2, src2_alpha ? _w : _x, src3, src3_alpha ? _w : _x),
+            ALU_MULADD(TEXEL_REG, _y, ALU_SRC_PV, _y, src2, src2_alpha ? _w : _y, src3, src3_alpha ? _w : _y),
+            ALU_MULADD(TEXEL_REG, _z, ALU_SRC_PV, _z, src2, src2_alpha ? _w : _z, src3, src3_alpha ? _w : _z)
+            ALU_LAST,
+        );
+    }
+}
+#undef ADD_INSTR
+
+static void append_tex_clamp(struct CCFeatures *cc_features, uint64_t **alu_ptr, uint8_t tex, bool s, bool t) {
+    if (s && t) {
+        add_tex_clamp_S_T(cc_features, alu_ptr, tex);
+    } else if (s) {
+        add_tex_clamp_S(cc_features, alu_ptr, tex);
+    } else {
+        add_tex_clamp_T(cc_features, alu_ptr, tex);
+    }
+}
+
+static void append_formula(struct CCFeatures *cc_features, uint64_t **alu_ptr, uint8_t c[2][4], bool do_single, bool do_multiply, bool do_mix, bool only_alpha) {
+    if (do_single) {
+        add_mov(cc_features, alu_ptr, c[only_alpha][3], only_alpha);
+    } else if (do_multiply) {
+        add_mul(cc_features, alu_ptr, c[only_alpha][0], c[only_alpha][2], only_alpha);
+    } else if (do_mix) {
+        add_mix(cc_features, alu_ptr, c[only_alpha][0], c[only_alpha][1], c[only_alpha][2], c[only_alpha][1], only_alpha);
+    } else {
+        add_mix(cc_features, alu_ptr, c[only_alpha][0], c[only_alpha][1], c[only_alpha][2], c[only_alpha][3], only_alpha);
+    }
+}
+
+static const uint64_t noise_instructions[] = {
+    /* R127 = floor(gl_FragCoord.xy * window_params.x) */
+    ALU_MUL(__, _x, FRAG_COORD_REG, _x, _C(0), _x),
+    ALU_MUL(__, _y, FRAG_COORD_REG, _y, _C(0), _x)
+    ALU_LAST,
+
+    ALU_FLOOR(_R127, _x, ALU_SRC_PV, _x),
+    ALU_FLOOR(_R127, _y, ALU_SRC_PV, _y)
+    ALU_LAST,
+
+    /* R127 = sin(vec3(R127.x, R127.y, window_params.y)) */
+    ALU_MULADD(_R127, _x, _R127, _x, ALU_SRC_LITERAL, _x, ALU_SRC_0_5, _x),
+    ALU_MULADD(_R127, _y, _R127, _y, ALU_SRC_LITERAL, _x, ALU_SRC_0_5, _x),
+    ALU_MULADD(_R127, _z, _C(0), _y, ALU_SRC_LITERAL, _x, ALU_SRC_0_5, _x)
+    ALU_LAST,
+    ALU_LITERAL(0x3E22F983 /* 0.1591549367f (radians -> revolutions) */),
+
+    ALU_FRACT(__, _x, _R127, _x),
+    ALU_FRACT(__, _y, _R127, _y),
+    ALU_FRACT(__, _z, _R127, _z)
+    ALU_LAST,
+
+    ALU_MULADD(_R127, _x, ALU_SRC_PV, _x, ALU_SRC_LITERAL, _x, ALU_SRC_LITERAL, _y),
+    ALU_MULADD(_R127, _y, ALU_SRC_PV, _y, ALU_SRC_LITERAL, _x, ALU_SRC_LITERAL, _y),
+    ALU_MULADD(_R127, _z, ALU_SRC_PV, _z, ALU_SRC_LITERAL, _x, ALU_SRC_LITERAL, _y)
+    ALU_LAST,
+    ALU_LITERAL2(0x40C90FDB /* 6.283185482f (tau) */, 0xC0490FDB /* -3.141592741f (-pi) */),
+
+    ALU_MUL(_R127, _x, ALU_SRC_PV, _x, ALU_SRC_LITERAL, _x),
+    ALU_MUL(_R127, _y, ALU_SRC_PV, _y, ALU_SRC_LITERAL, _x),
+    ALU_MUL(_R127, _z, ALU_SRC_PV, _z, ALU_SRC_LITERAL, _x)
+    ALU_LAST,
+    ALU_LITERAL(0x3E22F983 /* 0.1591549367f (radians -> revolutions) */),
+
+    ALU_SIN(_R127, _x, _R127, _x) SCL_210
+    ALU_LAST,
+
+    ALU_SIN(_R127, _y, _R127, _y) SCL_210
+    ALU_LAST,
+
+    ALU_SIN(_R127, _z, _R127, _z) SCL_210
+    ALU_LAST,
+
+    /* R127.x = dot(R127.xyz, vec3(12.9898, 78.233, 37.719)); */
+    ALU_DOT4(_R127, _x, _R127, _x, ALU_SRC_LITERAL, _x),
+    ALU_DOT4(__, _y, _R127, _y, ALU_SRC_LITERAL, _y),
+    ALU_DOT4(__, _z, _R127, _z, ALU_SRC_LITERAL, _z),
+    ALU_DOT4(__, _w, ALU_SRC_LITERAL, _w, ALU_SRC_0, _x)
+    ALU_LAST,
+    ALU_LITERAL4(0x414FD639 /* 12.9898f */, 0x429C774C /* 78.233f */, 0x4216E042 /* 37.719f */, 0x80000000 /* -0.0f */),
+
+    /* R127.x = fract(sin(R127.x) * 143758.5453); */
+    ALU_MULADD(_R127, _x, _R127, _x, ALU_SRC_LITERAL, _x, ALU_SRC_0_5, _x)
+    ALU_LAST,
+    ALU_LITERAL(0x3E22F983 /* 0.1591549367f (radians -> revolutions) */),
+
+    ALU_FRACT(__, _x, _R127, _x)
+    ALU_LAST,
+
+    ALU_MULADD(_R127, _x, ALU_SRC_PV, _x, ALU_SRC_LITERAL, _x, ALU_SRC_LITERAL, _y)
+    ALU_LAST,
+    ALU_LITERAL2(0x40C90FDB /* 6.283185482f (tau) */, 0xC0490FDB /* -3.141592741f (-pi) */),
+
+    ALU_SIN(_R127, _x, _R127, _x) SCL_210
+    ALU_LAST,
+
+    ALU_MUL(__, _x, _R127, _x, ALU_SRC_LITERAL, _x)
+    ALU_LAST,
+    ALU_LITERAL(0x480C63A3 /* 143758.5453f */),
+
+    ALU_FRACT( _R127, _x, ALU_SRC_PV, _x)
+    ALU_LAST,
+
+    /* texel.a *= floor(R127.x + 0.5); */
+    ALU_ADD(__, _x, _R127, _x, ALU_SRC_0_5, _x)
+    ALU_LAST,
+
+    ALU_FLOOR(__, _x, ALU_SRC_PV, _x)
+    ALU_LAST,
+
+    ALU_MUL(TEXEL_REG, _w, TEXEL_REG, _w, ALU_SRC_PV, _x)
+    ALU_LAST,
+};
+
+static GX2UniformVar uniformVars[] = {
+    { "window_params", GX2_SHADER_VAR_TYPE_FLOAT2, 1, 0, -1, },
+};
+
+static GX2SamplerVar samplerVars[] = {
+    { "uTex0", GX2_SAMPLER_VAR_TYPE_SAMPLER_2D, 0 },
+    { "uTex1", GX2_SAMPLER_VAR_TYPE_SAMPLER_2D, 1 },
+};
+
+#define ADD_INSTR(...) \
+    do { \
+    uint64_t tmp[] = {__VA_ARGS__}; \
+    memcpy(cur_buf, tmp, sizeof(tmp)); \
+    cur_buf += sizeof(tmp) / sizeof(uint64_t); \
+    } while (0)
+
+static int generatePixelShader(GX2PixelShader *psh, struct CCFeatures *cc_features) {
+    static const size_t max_program_buf_size = 512 * sizeof(uint64_t);
+    uint64_t *program_buf = memalign(GX2_SHADER_PROGRAM_ALIGNMENT, max_program_buf_size);
+    if (!program_buf) {
+        return -1;
+    }
+
+    memset(program_buf, 0, max_program_buf_size);
+
+    // start placing alus at offset 32
+    static const uint32_t base_alu_offset = 32;
+    uint64_t *cur_buf = NULL;
+
+    // check if we need to clamp
+    bool texclamp[2] = { false, false };
+    for (int i = 0; i < 2; i++) {
+        if (cc_features->used_textures[i]) {
+            if (cc_features->clamp[i][0] || cc_features->clamp[i][1]) {
+                texclamp[i] = true;
+            }
+        }
+    }
+
+    uint32_t texclamp_alu_offset = base_alu_offset;
+    uint32_t texclamp_alu_size = 0;
+    uint32_t texclamp_alu_cnt = 0;
+
+    if (texclamp[0] || texclamp[1]) {
+        // texclamp alu
+        cur_buf = program_buf + texclamp_alu_offset;
+
+        for (int i = 0; i < 2; i++) {
+            if (cc_features->used_textures[i] && texclamp[i]) {
+                append_tex_clamp(cc_features, &cur_buf, i, cc_features->clamp[i][0], cc_features->clamp[i][1]);
+            }
+        }
+
+        texclamp_alu_size = (uintptr_t) cur_buf - ((uintptr_t) (program_buf + texclamp_alu_offset));
+        texclamp_alu_cnt = texclamp_alu_size / sizeof(uint64_t);
+    }
+
+    // main alu0
+    uint32_t main_alu0_offset = texclamp_alu_offset + texclamp_alu_cnt;
+    cur_buf = program_buf + main_alu0_offset;
+
+    for (int c = 0; c < (cc_features->opt_2cyc ? 2 : 1); c++) {
+        append_formula(cc_features, &cur_buf, cc_features->c[c], cc_features->do_single[c][0], cc_features->do_multiply[c][0], cc_features->do_mix[c][0], false);
+        if (cc_features->opt_alpha) {
+            append_formula(cc_features, &cur_buf, cc_features->c[c], cc_features->do_single[c][1], cc_features->do_multiply[c][1], cc_features->do_mix[c][1], true);
+        }
+    }
+
+    if (cc_features->opt_fog) {
+        ADD_INSTR(
+            /* texel.rgb = mix(texel.rgb, vFog.rgb, vFog.a); */
+            ALU_ADD(__, _x, FOG_REG, _x, _R1 _NEG, _x),
+            ALU_ADD(__, _y, FOG_REG, _y, _R1 _NEG, _y),
+            ALU_ADD(__, _z, FOG_REG, _z, _R1 _NEG, _z)
+            ALU_LAST,
+
+            ALU_MULADD(TEXEL_REG, _x, ALU_SRC_PV, _x, FOG_REG, _w, TEXEL_REG, _x),
+            ALU_MULADD(TEXEL_REG, _y, ALU_SRC_PV, _y, FOG_REG, _w, TEXEL_REG, _y),
+            ALU_MULADD(TEXEL_REG, _z, ALU_SRC_PV, _z, FOG_REG, _w, TEXEL_REG, _z)
+            ALU_LAST,
+        );
+    }
+
+    if (cc_features->opt_texture_edge && cc_features->opt_alpha) {
+        ADD_INSTR(
+            /* if (texel.a > 0.19) texel.a = 1.0; else discard; */
+            ALU_KILLGT(__, _x, ALU_SRC_LITERAL, _x, TEXEL_REG, _w),
+            ALU_MOV(TEXEL_REG, _w, ALU_SRC_1, _x)
+            ALU_LAST,
+            ALU_LITERAL(0x3e428f5c /*0.19f*/),
+        );
+    }
+
+    const uint32_t main_alu0_size = (uintptr_t) cur_buf - ((uintptr_t) (program_buf + main_alu0_offset));
+    const uint32_t main_alu0_cnt = main_alu0_size / sizeof(uint64_t);
+
+    // main alu1
+    // place the following instructions into a new alu, in case the other alu uses KILL
+    const uint32_t main_alu1_offset = main_alu0_offset + main_alu0_cnt;
+    cur_buf = program_buf + main_alu1_offset;
+
+    if (cc_features->opt_alpha && cc_features->opt_noise) {
+        memcpy(cur_buf, noise_instructions, sizeof(noise_instructions));
+        cur_buf += sizeof(noise_instructions) / sizeof(uint64_t);
+    }
+
+    if (cc_features->opt_grayscale) {
+        ADD_INSTR(
+            /* texel.r + texel.g + texel.b */
+            ALU_ADD(__, _x, TEXEL_REG, _x, TEXEL_REG, _y)
+            ALU_LAST,
+
+            ALU_ADD(__, _x, ALU_SRC_PV, _x, TEXEL_REG, _z)
+            ALU_LAST,
+
+            /* PV.x / 3 */
+            ALU_MUL_IEEE(__, _x, ALU_SRC_PV, _x, ALU_SRC_LITERAL, _x)
+            ALU_LAST,
+            ALU_LITERAL(0x3eaaaaab /*0.3333333433f*/),
+
+            /* texel.rgb = mix(texel.rgb, vGrayscaleColor.rgb * intensity, vGrayscaleColor.a); */
+            ALU_MULADD(_R127, _x, GRAYSCALE_REG, _x, ALU_SRC_PV, _x, _R1 _NEG, _x),
+            ALU_MULADD(_R127, _y, GRAYSCALE_REG, _y, ALU_SRC_PV, _x, _R1 _NEG, _y),
+            ALU_MULADD(_R127, _z, GRAYSCALE_REG, _z, ALU_SRC_PV, _x, _R1 _NEG, _z)
+            ALU_LAST,
+
+            ALU_MULADD(TEXEL_REG, _x, ALU_SRC_PV, _x, GRAYSCALE_REG, _w, TEXEL_REG, _x),
+            ALU_MULADD(TEXEL_REG, _y, ALU_SRC_PV, _y, GRAYSCALE_REG, _w, TEXEL_REG, _y),
+            ALU_MULADD(TEXEL_REG, _z, ALU_SRC_PV, _z, GRAYSCALE_REG, _w, TEXEL_REG, _z)
+            ALU_LAST,
+        );
+    }
+
+    if (cc_features->opt_alpha) {
+        if (cc_features->opt_alpha_threshold) {
+            ADD_INSTR(
+                /* if (texel.a < 8.0 / 256.0) discard; */
+                ALU_KILLGT(__, _x, ALU_SRC_LITERAL, _x, TEXEL_REG, _w)
+                ALU_LAST,
+                ALU_LITERAL(0x3d000000 /*0.03125f*/),
+            );
+        }
+
+        if (cc_features->opt_invisible) {
+            ADD_INSTR(
+                /* texel.a = 0.0; */
+                ALU_MOV(TEXEL_REG, _w, ALU_SRC_0, _x)
+                ALU_LAST,
+            );
+        }
+    }
+
+    const uint32_t main_alu1_size = (uintptr_t) cur_buf - ((uintptr_t) (program_buf + main_alu1_offset));
+    const uint32_t main_alu1_cnt = main_alu1_size / sizeof(uint64_t);
+
+    // tex
+    uint32_t num_textures = cc_features->used_textures[0] + cc_features->used_textures[1];
+    uint32_t num_texinfo = texclamp[0] + texclamp[1];
+
+    uint32_t texinfo_offset = ROUNDUP(main_alu1_offset + main_alu1_cnt, 16);
+    uint32_t cur_tex_offset = texinfo_offset;
+
+    for (int i = 0; i < 2; i++) {
+        if (cc_features->used_textures[i] && texclamp[i]) {
+            uint8_t dst_reg = get_reg(cc_features, (i == 0) ? SHADER_TEXINFO0 : SHADER_TEXINFO1);
+
+            uint64_t texinfo_buf[] = {
+                TEX_GET_TEXTURE_INFO(dst_reg, _x, _y, _m, _m, _R1, _0, _0, _0, _0,  _t(i), _s(i))
+            };
+
+            memcpy(program_buf + cur_tex_offset, texinfo_buf, sizeof(texinfo_buf));
+            cur_tex_offset += sizeof(texinfo_buf) / sizeof(uint64_t);
+        }
+    }
+
+    uint32_t texsample_offset = cur_tex_offset;
+
+    for (int i = 0; i < 2; i++) {
+        if (cc_features->used_textures[i]) {
+            uint8_t texcoord_reg = (i == 0) ? _R1 : _R2;
+            uint8_t dst_reg = get_reg(cc_features, (i == 0) ? SHADER_TEXEL0 : SHADER_TEXEL1);
+
+            uint64_t tex_buf[] = {
+                TEX_SAMPLE(dst_reg, _x, _y, _z, _w, texcoord_reg, _x, _y, _0, _x, _t(i), _s(i))
+            };
+
+            memcpy(program_buf + cur_tex_offset, tex_buf, sizeof(tex_buf));
+            cur_tex_offset += sizeof(tex_buf) / sizeof(uint64_t);
+        }
+    }
+
+    // make sure we didn't overflow the buffer
+    const uint32_t total_program_size = cur_tex_offset * sizeof(uint64_t);
+    assert(total_program_size <= max_program_buf_size);
+
+    // cf
+    uint32_t cur_cf_offset = 0;
+
+    // if we use texclamp place those alus first
+    if (texclamp[0] || texclamp[1]) {
+        program_buf[cur_cf_offset++] = TEX(texinfo_offset, num_texinfo);
+        program_buf[cur_cf_offset++] = ALU(texclamp_alu_offset, texclamp_alu_cnt);
+    }
+
+    if (num_textures > 0) {
+        program_buf[cur_cf_offset++] = TEX(texsample_offset, num_textures) VALID_PIX;
+    }
+
+    program_buf[cur_cf_offset++] = ALU(main_alu0_offset, main_alu0_cnt);
+
+    if (main_alu1_cnt > 0) {
+        program_buf[cur_cf_offset++] = ALU(main_alu1_offset, main_alu1_cnt);
+    }
+
+    if (cc_features->opt_alpha) {
+        program_buf[cur_cf_offset++] = EXP_DONE(PIX0, TEXEL_REG, _x, _y, _z, _w) END_OF_PROGRAM;
+    } else {
+        program_buf[cur_cf_offset++] = EXP_DONE(PIX0, TEXEL_REG, _x, _y, _z, _1) END_OF_PROGRAM;
+    }
+
+    // regs
+    const uint32_t num_ps_inputs = 4 + cc_features->num_inputs;
+
+    psh->regs.sq_pgm_resources_ps = get_num_regs(cc_features); // num_gprs
+    psh->regs.sq_pgm_exports_ps = 2; // export_mode
+    psh->regs.spi_ps_in_control_0 = (num_ps_inputs + 1) // num_interp
+        | (1 << 8) // position_ena
+        | (1 << 26) // persp_gradient_ena
+        | (1 << 28); // baryc_sample_cntl
+
+    psh->regs.num_spi_ps_input_cntl = num_ps_inputs + 1;
+
+    // frag pos
+    psh->regs.spi_ps_input_cntls[0] = 0 | (1 << 8);
+
+    // inputs
+    for (int i = 0; i < num_ps_inputs; i++) {
+        psh->regs.spi_ps_input_cntls[i + 1] = i | (1 << 8);
+    }
+
+    psh->regs.cb_shader_mask = 0xf; // output0_enable
+    psh->regs.cb_shader_control = 1; // rt0_enable
+    psh->regs.db_shader_control = (1 << 4) // z_order
+        | (1 << 6); // kill_enable
+    
+    // program
+    psh->size = total_program_size;
+    psh->program = program_buf;
+
+    psh->mode = GX2_SHADER_MODE_UNIFORM_REGISTER;
+
+    // uniform vars
+    psh->uniformVars = uniformVars;
+    psh->uniformVarCount = sizeof(uniformVars) / sizeof(GX2UniformVar);
+
+    // samplers
+    psh->samplerVars = samplerVars;
+    psh->samplerVarCount = sizeof(samplerVars) / sizeof(GX2SamplerVar);
+
+    return 0;
+}
+
+static GX2AttribVar attribVars[] = {
+    { "aVtxPos",         GX2_SHADER_VAR_TYPE_FLOAT4, 0, 0 },
+    { "aTexCoord0",      GX2_SHADER_VAR_TYPE_FLOAT4, 0, 1 },
+    { "aTexCoord1",      GX2_SHADER_VAR_TYPE_FLOAT4, 0, 2 },
+    { "aFog",            GX2_SHADER_VAR_TYPE_FLOAT4, 0, 3 },
+    { "aGrayscaleColor", GX2_SHADER_VAR_TYPE_FLOAT4, 0, 4 },
+    { "aInput1",         GX2_SHADER_VAR_TYPE_FLOAT4, 0, 5 },
+    { "aInput2",         GX2_SHADER_VAR_TYPE_FLOAT4, 0, 6 },
+    { "aInput3",         GX2_SHADER_VAR_TYPE_FLOAT4, 0, 7 },
+    { "aInput4",         GX2_SHADER_VAR_TYPE_FLOAT4, 0, 8 },
+    { "aInput5",         GX2_SHADER_VAR_TYPE_FLOAT4, 0, 9 },
+    { "aInput6",         GX2_SHADER_VAR_TYPE_FLOAT4, 0, 10 },
+    { "aInput7",         GX2_SHADER_VAR_TYPE_FLOAT4, 0, 11 },
+};
+
+static int generateVertexShader(GX2VertexShader *vsh, struct CCFeatures *cc_features) {
+    static const size_t max_program_buf_size = 16 * sizeof(uint64_t);
+    uint64_t *program_buf = memalign(GX2_SHADER_PROGRAM_ALIGNMENT, max_program_buf_size);
+    if (!program_buf) {
+        return -1;
+    }
+
+    const uint32_t num_ps_inputs = 4 + cc_features->num_inputs;
+
+    uint64_t *cur_buf = program_buf;
+
+    // aVtxPos
+    ADD_INSTR(
+        CALL_FS NO_BARRIER,
+        EXP_DONE(POS0, _R1, _x, _y, _z, _w),
+    );
+
+    // params
+    for (int i = 0; i < num_ps_inputs - 1; i++) {
+        ADD_INSTR(
+            EXP(PARAM(i), _R(i + 2), _x, _y, _z, _w) NO_BARRIER,
+        );
+    }
+
+    // last param
+    ADD_INSTR(
+        (EXP_DONE(PARAM(num_ps_inputs - 1), _R(num_ps_inputs + 1), _x, _y, _z, _w) NO_BARRIER)
+        END_OF_PROGRAM,
+    );
+
+    const uint32_t program_size = (uintptr_t) cur_buf - ((uintptr_t) program_buf);
+    assert(program_size <= max_program_buf_size);
+
+    // regs
+    vsh->regs.sq_pgm_resources_vs = (num_ps_inputs + 2) // num_gprs
+        | (1 << 8); // stack_size
+
+    // num outputs minus 1
+    vsh->regs.spi_vs_out_config = ((num_ps_inputs - 1) << 1);
+
+    vsh->regs.num_spi_vs_out_id = 3;
+    memset(vsh->regs.spi_vs_out_id, 0xff, sizeof(vsh->regs.spi_vs_out_id));
+    vsh->regs.spi_vs_out_id[0] = (0) | (1 << 8) | (2 << 16) | (3 << 24);
+    vsh->regs.spi_vs_out_id[1] = (4) | (5 << 8) | (6 << 16) | (7 << 24);
+    vsh->regs.spi_vs_out_id[2] = (8) | (9 << 8) | (10 << 16) | (0xff << 24);
+
+    vsh->regs.sq_vtx_semantic_clear = ~((1 << 12) - 1);
+    vsh->regs.num_sq_vtx_semantic = 12;
+    memset(vsh->regs.sq_vtx_semantic, 0xff, sizeof(vsh->regs.sq_vtx_semantic));
+    // aVtxPos
+    vsh->regs.sq_vtx_semantic[0] = 0;
+    // aTexCoord0
+    vsh->regs.sq_vtx_semantic[1] = 1;
+    // aTexCoord1
+    vsh->regs.sq_vtx_semantic[2] = 2;
+    // aFog
+    vsh->regs.sq_vtx_semantic[3] = 3;
+    // aGrayscaleColor
+    vsh->regs.sq_vtx_semantic[4] = 4;
+    // aInput1
+    vsh->regs.sq_vtx_semantic[5] = 5;
+    // aInput2
+    vsh->regs.sq_vtx_semantic[6] = 6;
+    // aInput3
+    vsh->regs.sq_vtx_semantic[7] = 7;
+    // aInput4
+    vsh->regs.sq_vtx_semantic[8] = 8;
+    // aInput5
+    vsh->regs.sq_vtx_semantic[9] = 9;
+    // aInput6
+    vsh->regs.sq_vtx_semantic[10] = 10;
+    // aInput7
+    vsh->regs.sq_vtx_semantic[11] = 11;
+
+    vsh->regs.vgt_vertex_reuse_block_cntl = 14; // vtx_reuse_depth
+    vsh->regs.vgt_hos_reuse_depth = 16; // reuse_depth
+
+    // program
+    vsh->program = program_buf;
+    vsh->size = program_size;
+
+    vsh->mode = GX2_SHADER_MODE_UNIFORM_REGISTER;
+
+    // attribs
+    vsh->attribVarCount = sizeof(attribVars) / sizeof(GX2AttribVar);
+    vsh->attribVars = attribVars;
+
+    return 0;
+}
+#undef ADD_INSTR
+
+int gx2GenerateShaderGroup(struct ShaderGroup *group, struct CCFeatures *cc_features) {
+    memset(group, 0, sizeof(struct ShaderGroup));
+
+    // generate the pixel shader
+    if (generatePixelShader(&group->pixelShader, cc_features) != 0) {
+        gx2FreeShaderGroup(group);
+        return -1;
+    }
+
+    // generate the vertex shader
+    if (generateVertexShader(&group->vertexShader, cc_features) != 0) {
+        gx2FreeShaderGroup(group);
+        return -1;
+    }
+
+    uint32_t attribOffset = 0;
+
+    // aVtxPos
+    group->attributes[group->numAttributes++] = 
+        (GX2AttribStream) { 0, 0, attribOffset, GX2_ATTRIB_FORMAT_FLOAT_32_32_32_32, GX2_ATTRIB_INDEX_PER_VERTEX, 0, GX2_COMP_SEL(_x, _y, _z, _w), GX2_ENDIAN_SWAP_DEFAULT };
+    attribOffset += 4 * sizeof(float);
+
+    for (int i = 0; i < 2; i++) {
+        if (cc_features->used_textures[i]) {
+            // aTexCoordX
+            group->attributes[group->numAttributes++] = 
+                (GX2AttribStream) { 1 + i, 0, attribOffset, GX2_ATTRIB_FORMAT_FLOAT_32_32_32_32, GX2_ATTRIB_INDEX_PER_VERTEX, 0, GX2_COMP_SEL(_x, _y, _z, _w), GX2_ENDIAN_SWAP_DEFAULT };
+            attribOffset += 4 * sizeof(float);
+        }
+    }
+
+    // aFog
+    if (cc_features->opt_fog) {
+        group->attributes[group->numAttributes++] = 
+            (GX2AttribStream) { 3, 0, attribOffset, GX2_ATTRIB_FORMAT_FLOAT_32_32_32_32, GX2_ATTRIB_INDEX_PER_VERTEX, 0, GX2_COMP_SEL(_x, _y, _z, _w), GX2_ENDIAN_SWAP_DEFAULT };
+        attribOffset += 4 * sizeof(float);
+    }
+
+    // aGrayscaleColor
+    if (cc_features->opt_grayscale) {
+        group->attributes[group->numAttributes++] = 
+            (GX2AttribStream) { 4, 0, attribOffset, GX2_ATTRIB_FORMAT_FLOAT_32_32_32_32, GX2_ATTRIB_INDEX_PER_VERTEX, 0, GX2_COMP_SEL(_x, _y, _z, _w), GX2_ENDIAN_SWAP_DEFAULT };
+        attribOffset += 4 * sizeof(float);
+    }
+
+    // aInput
+    for (int i = 0; i < cc_features->num_inputs; i++) {
+        group->attributes[group->numAttributes++] = 
+            (GX2AttribStream) { 5 + i, 0, attribOffset, GX2_ATTRIB_FORMAT_FLOAT_32_32_32_32, GX2_ATTRIB_INDEX_PER_VERTEX, 0, GX2_COMP_SEL(_x, _y, _z, _w), GX2_ENDIAN_SWAP_DEFAULT };
+        attribOffset += 4 * sizeof(float);
+    }
+
+    group->stride = attribOffset;
+
+    // init the fetch shader
+    group->fetchShader.size = GX2CalcFetchShaderSizeEx(group->numAttributes, GX2_FETCH_SHADER_TESSELLATION_NONE, GX2_TESSELLATION_MODE_DISCRETE);
+    group->fetchShader.program = memalign(GX2_SHADER_PROGRAM_ALIGNMENT, group->fetchShader.size);
+    if (!group->fetchShader.program) {
+        gx2FreeShaderGroup(group);
+        return -1;
+    }
+
+    GX2InitFetchShaderEx(&group->fetchShader, group->fetchShader.program, group->numAttributes, group->attributes, GX2_FETCH_SHADER_TESSELLATION_NONE, GX2_TESSELLATION_MODE_DISCRETE);
+
+    // invalidate all programs
+    GX2Invalidate(GX2_INVALIDATE_MODE_CPU_SHADER, group->vertexShader.program, group->vertexShader.size);
+    GX2Invalidate(GX2_INVALIDATE_MODE_CPU_SHADER, group->pixelShader.program, group->pixelShader.size);
+    GX2Invalidate(GX2_INVALIDATE_MODE_CPU_SHADER, group->fetchShader.program, group->fetchShader.size);
+
+    return 0;
+}
+
+void gx2FreeShaderGroup(struct ShaderGroup *group) {
+    free(group->vertexShader.program);
+    free(group->pixelShader.program);
+    free(group->fetchShader.program);
+}
+
+#endif

libultraship/libultraship/Lib/Fast3D/gx2_shader_gen.h

@@ -0,0 +1,27 @@
+#pragma once
+
+#include "gfx_cc.h"
+#include <gx2/shaders.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ShaderGroup {
+    GX2VertexShader vertexShader;
+    GX2PixelShader pixelShader;
+    GX2FetchShader fetchShader;
+
+    uint32_t stride;
+
+    uint32_t numAttributes;
+    GX2AttribStream attributes[13];
+};
+
+int gx2GenerateShaderGroup(struct ShaderGroup *group, struct CCFeatures *cc_features);
+
+void gx2FreeShaderGroup(struct ShaderGroup *group);
+
+#ifdef __cplusplus
+}
+#endif

libultraship/libultraship/Lib/Fast3D/gx2_shader_inl.h

@@ -0,0 +1,624 @@
+/*  
+ *  Copyright (C) 2014-2016 - Ali Bouhlel
+ *  Copyright (C) 2022 - GaryOderNichts
+ *
+ *  Licensed as MIT.
+ */
+
+#ifndef GX2_SHADER_INL_H
+#define GX2_SHADER_INL_H
+
+#define to_QWORD(w0, w1) (((uint64_t)(w0) << 32ull) | (w1))
+#define to_LE(x) (__builtin_bswap32(x))
+
+/* CF */
+#define CF_DWORD0(addr) to_LE(addr)
+
+#define CF_DWORD1(popCount, cfConst, cond, count, callCount, inst) \
+   to_LE(popCount | (cfConst << 3) | (cond << 8) | (count << 10) | (callCount << 13) | (inst << 23) | (1 << 31))
+
+#define CF_ALU_WORD0(addr, kcacheBank0, kcacheBank1, kcacheMode0) \
+   to_LE(addr | (kcacheBank0 << 22) | (kcacheBank1 << 26) | (kcacheMode0 << 30))
+#define CF_ALU_WORD1(kcacheMode1, kcacheAddr0, kcacheAddr1, count, altConst, inst) \
+   to_LE(kcacheMode1 | (kcacheAddr0 << 2) | (kcacheAddr1 << 10) | (count << 18) | (altConst << 25) | (inst << 26) | (1 << 31))
+
+#define CF_EXP_WORD0(dstReg_and_type, srcReg, srcRel, indexGpr, elemSize)\
+   to_LE(dstReg_and_type | (srcReg << 15) | (srcRel << 22) | (indexGpr << 23) | (elemSize << 30))
+
+#define CF_EXP_WORD1(srcSelX, srcSelY, srcSelZ, srcSelW, validPixelMode, inst) \
+   to_LE(srcSelX | (srcSelY << 3) | (srcSelZ << 6) | (srcSelW << 9) | (validPixelMode << 22) | (inst << 23) | (1 << 31))
+
+#define CF_ALLOC_EXPORT_WORD0(arrayBase, type, dstReg, dstRel, indexGpr, elemSize) \
+   to_LE(arrayBase | (type << 13) | (dstReg << 15) | (dstRel << 22) | (indexGpr << 23) | (elemSize << 30))
+
+#define CF_ALLOC_EXPORT_WORD1_BUF(arraySize, writeMask, inst) \
+   to_LE(arraySize | (writeMask << 12) | (inst << 23) | (1 << 31))
+
+#define ALU_SRC_KCACHE0_BASE  0x80
+#define ALU_SRC_KCACHE1_BASE  0xA0
+#define CF_KCACHE_BANK_LOCK_1 0x1
+#define CB1                   0x1
+#define CB2                   0x2
+#define _0_15                 CF_KCACHE_BANK_LOCK_1
+
+#define KC0(x) (x + ALU_SRC_KCACHE0_BASE)
+#define KC1(x) (x + ALU_SRC_KCACHE1_BASE)
+
+#define NO_BARRIER      & (~to_QWORD(0,to_LE(1 << 31)))
+#define END_OF_PROGRAM  | to_QWORD(0,to_LE(1 << 21))
+#define VALID_PIX       | to_QWORD(0,to_LE(1 << 22))
+#define WHOLE_QUAD_MODE | to_QWORD(0,to_LE(1 << 30))
+#define BURSTCNT(x)     | to_QWORD(0,to_LE(x << 17))
+#define WRITE(x)        (x >> 2)
+#define ARRAY_SIZE(x)   x
+#define ELEM_SIZE(x)    x
+#define KCACHE0(bank, mode) | to_QWORD(CF_ALU_WORD0(0, bank, 0, mode), 0)
+#define KCACHE1(bank, mode) | to_QWORD(CF_ALU_WORD0(0, 0, bank, 0), CF_ALU_WORD1(mode,0, 0, 0, 0, 0))
+
+#define DEACTIVATE               1
+#define UPDATE_EXEC_MASK(mode)   | to_QWORD(0, to_LE(mode << 2))
+#define UPDATE_PRED              | to_QWORD(0, to_LE(1ull << 3))
+#define CLAMP                    | to_QWORD(0, to_LE(1ull << 31))
+#define ALU_LAST                 | to_QWORD(to_LE(1ull << 31), 0)
+
+/* ALU */
+
+#define ALU_WORD0(src0Sel, src0Rel, src0Chan, src0Neg, src1Sel, src1Rel, src1Chan, src1Neg, indexMode, predSel) \
+   to_LE(src0Sel | ((src0Rel) << 9) | ((src0Chan) << 10) | ((src0Neg) << 12) | ((src1Sel) << 13) | ((src1Rel) << 22) \
+                 | ((src1Chan) << 23) | ((src1Neg) << 25) | ((indexMode) << 26) | ((predSel) << 29))
+
+#define ALU_WORD1_OP2(src0Abs, src1Abs, updateExecuteMask, updatePred, writeMask, omod, inst, encoding, bankSwizzle, dstGpr, dstRel, dstChan, clamp) \
+      to_LE(src0Abs | (src1Abs << 1) | (updateExecuteMask << 2) | (updatePred << 3) | (writeMask << 4) | (omod << 5) | (inst << 7) | \
+                         (encoding << 15) | (bankSwizzle << 18) | ((dstGpr&0x7F) << 21) | (dstRel << 28) | ((dstChan&0x3) << 29) | (clamp << 31))
+
+#define ALU_WORD1_OP3(src2Sel, src2Rel, src2Chan, src2Neg, inst, bankSwizzle, dstGpr, dstRel, dstChan, clamp) \
+      to_LE(src2Sel | (src2Rel << 9) | (src2Chan << 10) | (src2Neg << 12) | (inst << 13) | \
+     (bankSwizzle << 18) | ((dstGpr&0x7F) << 21) | (dstRel << 28) | ((dstChan&0x3) << 29) | (clamp << 31))
+
+/* TEX */
+#define TEX_WORD0(inst, bcFracMode, fetchWholeQuad, resourceID, srcReg, srcRel, altConst) \
+   to_LE(inst | (bcFracMode << 5) | (fetchWholeQuad << 7) | (resourceID << 8) | (srcReg << 16) | (srcRel << 23) | (altConst << 24))
+
+#define TEX_WORD1(dstReg, dstRel, dstSelX, dstSelY, dstSelZ, dstSelW, lodBias, coordTypeX, coordTypeY, coordTypeZ, coordTypeW) \
+   to_LE(dstReg | (dstRel << 7) | (dstSelX << 9) | (dstSelY << 12) | (dstSelZ << 15) | (dstSelW << 18) | \
+   (lodBias << 21) | (coordTypeX << 28) | (coordTypeY << 29) | (coordTypeZ << 30) | (coordTypeW << 31))
+
+#define TEX_WORD2(offsetX, offsetY, offsetZ, samplerID, srcSelX, srcSelY, srcSelZ, srcSelW) \
+   to_LE(offsetX | (offsetY << 5) | (offsetZ << 10) | (samplerID << 15) | (srcSelX << 20) | (srcSelY << 23) | (srcSelZ << 26) | (srcSelW << 29))
+
+#define VTX_WORD0(inst, type, buffer_id, srcReg, srcSelX, mega) \
+   to_LE(inst | (type << 5) | (buffer_id << 8) | (srcReg << 16) | (srcSelX << 24) | (mega << 26))
+
+#define VTX_WORD1(dstReg, dstSelX, dstSelY, dstSelZ, dstSelW) \
+   to_LE(dstReg | (dstSelX << 9) | (dstSelY << 12) | (dstSelZ << 15) | (dstSelW << 18) | (1 << 21))
+
+#define VTX_WORD2(offset, ismega) \
+   to_LE(offset| (ismega << 19))
+
+#define _x 0
+#define _y 1
+#define _z 2
+#define _w 3
+#define _0 4
+#define _1 5
+#define _m 7 /*mask*/
+
+#define _xyzw 0b1111
+#define _xy__ 0b0011
+
+#define GX2_COMP_SEL(c0, c1, c2, c3) (((c0) << 24) | ((c1) << 16) | ((c2) << 8) | (c3))
+
+#define ALU_LITERAL(v)  to_QWORD(to_LE(v), 0)
+#define ALU_LITERAL2(v0,v1)  to_QWORD(to_LE(v0), to_LE(v1))
+#define ALU_LITERAL3(v0,v1,v2)  ALU_LITERAL2(v0,v1),ALU_LITERAL(v2)
+#define ALU_LITERAL4(v0,v1,v2,v3)  ALU_LITERAL2(v0,v1),ALU_LITERAL2(v2,v3)
+#define ALU_LITERAL5(v0,v1,v2,v3,v5)  ALU_LITERAL4(v0,v1,v2,v3),ALU_LITERAL(v4)
+
+/* SRCx_SEL special constants */
+#define ALU_SRC_1_DBL_L     0xF4
+#define ALU_SRC_1_DBL_M     0xF5
+#define ALU_SRC_0_5_DBL_L   0xF6
+#define ALU_SRC_0_5_DBL_M   0xF7
+#define ALU_SRC_0           0xF8
+#define ALU_SRC_1           0xF9
+#define ALU_SRC_1_INT       0xFA
+#define ALU_SRC_M_1_INT     0xFB
+#define ALU_SRC_0_5         0xFC
+#define ALU_SRC_LITERAL     0xFD
+#define ALU_SRC_PV          0xFE
+#define ALU_SRC_PS          0xFF
+
+#define _NEG                | (1 << 12)
+#define _ABS                | (1 << 13)
+
+#define ALU_OMOD_OFF          0x0
+#define ALU_OMOD_M2           0x1
+#define ALU_OMOD_M4           0x2
+#define ALU_OMOD_D2           0x3
+
+#define ALU_VEC_012           0x0
+#define ALU_VEC_021           0x1
+#define ALU_VEC_120           0x2
+#define ALU_VEC_102           0x3
+#define ALU_VEC_201           0x4
+#define ALU_VEC_210           0x5
+#define VEC_012               | to_QWORD(0, to_LE(ALU_VEC_012 << 18))
+#define VEC_021               | to_QWORD(0, to_LE(ALU_VEC_021 << 18))
+#define VEC_120               | to_QWORD(0, to_LE(ALU_VEC_120 << 18))
+#define VEC_102               | to_QWORD(0, to_LE(ALU_VEC_102 << 18))
+#define VEC_201               | to_QWORD(0, to_LE(ALU_VEC_201 << 18))
+#define VEC_210               | to_QWORD(0, to_LE(ALU_VEC_210 << 18))
+
+#define VALID_PIX       | to_QWORD(0,to_LE(1 << 22))
+
+#define ALU_SCL_210           0x0
+#define ALU_SCL_122           0x1
+#define ALU_SCL_212           0x2
+#define ALU_SCL_221           0x3
+
+#define SCL_210               | to_QWORD(0, to_LE(ALU_SCL_210 << 18))
+#define SCL_122               | to_QWORD(0, to_LE(ALU_SCL_122 << 18))
+#define SCL_212               | to_QWORD(0, to_LE(ALU_SCL_212 << 18))
+#define SCL_221               | to_QWORD(0, to_LE(ALU_SCL_221 << 18))
+
+#define FETCH_TYPE(x) x
+#define MINI(x) ((x) - 1)
+#define MEGA(x) (MINI(x) | 0x80000000)
+#define OFFSET(x) x
+
+#define VERTEX_DATA     0
+#define INSTANCE_DATA   1
+#define NO_INDEX_OFFSET 2
+
+/* CF defines */
+#define CF_COND_ACTIVE      0x0
+#define CF_COND_FALSE       0x1
+#define CF_COND_BOOL        0x2
+#define CF_COND_NOT_BOOL    0x3
+
+/* TEX defines */
+#define TEX_UNNORMALIZED    0x0
+#define TEX_NORMALIZED      0x1
+
+/* instructions */
+/* CF */
+#define CF_INST_TEX              0x01
+#define CF_INST_VTX              0x02
+#define CF_INST_LOOP_END         0x05
+#define CF_INST_LOOP_START_DX10  0x06
+#define CF_INST_JUMP             0x0A
+#define CF_INST_ELSE             0x0D
+#define CF_INST_POP              0x0E
+#define CF_INST_CALL_FS          0x13
+#define CF_INST_EMIT_VERTEX      0x15
+#define CF_INST_MEM_RING         0x26
+
+#define CF_INST_ALU              0x08
+#define CF_INST_ALU_PUSH_BEFORE  0x09
+#define CF_INST_ALU_POP_AFTER    0x0A
+#define CF_INST_ALU_POP2_AFTER   0x0B
+#define CF_INST_ALU_BREAK        0x0E
+#define CF_INST_ALU_ELSE_AFTER   0x0F
+/* ALU */
+#define OP2_INST_ADD             0x0
+#define OP2_INST_MUL             0x1
+#define OP2_INST_MUL_IEEE        0x2
+#define OP2_INST_MIN             0x04
+#define OP2_INST_MAX             0x03
+#define OP2_INST_MAX_DX10        0x05
+#define OP2_INST_FRACT           0x10
+#define OP2_INST_SETGT           0x09
+#define OP2_INST_SETE_DX10       0x0C
+#define OP2_INST_SETGT_DX10      0x0D
+#define OP2_INST_SETGE_DX10      0x0E
+#define OP2_INST_FLOOR           0x14
+#define OP2_INST_MOVA_INT        0x18
+#define OP2_INST_MOV             0x19
+#define OP2_INST_NOP             0x1A
+#define OP2_INST_PRED_SETGT      0x21
+#define OP2_INST_KILLGT          0x2D
+#define OP2_INST_AND_INT         0x30
+#define OP2_INST_OR_INT          0x31
+#define OP2_INST_NOT_INT         0x33
+#define OP2_INST_ADD_INT         0x34
+#define OP2_INST_SETE_INT        0x3A
+#define OP2_INST_SETGT_INT       0x3B
+#define OP2_INST_SETGE_INT       0x3C
+#define OP2_INST_SETNE_INT       0x3D
+#define OP2_INST_PRED_SETE_INT   0x42
+#define OP2_INST_PRED_SETGT_INT  0x43
+#define OP2_INST_PRED_SETGE_INT  0x44
+#define OP2_INST_PRED_SETNE_INT  0x45
+#define OP2_INST_KILLE_INT       0x46
+#define OP2_INST_KILLGT_INT      0x47
+#define OP2_INST_KILLGE_INT      0x48
+#define OP2_INST_KILLNE_INT      0x49
+#define OP2_INST_DOT4            0x50
+#define OP2_INST_DOT4_IEEE       0x51
+#define OP2_INST_EXP_IEEE        0x61
+#define OP2_INST_LOG_CLAMPED     0x62
+#define OP2_INST_RECIP_IEEE      0x66
+#define OP2_INST_RECIPSQRT_IEEE  0x69
+#define OP2_INST_SQRT_IEEE       0x6A
+#define OP2_INST_FLT_TO_INT      0x6B
+#define OP2_INST_INT_TO_FLT      0x6C
+#define OP2_INST_SIN             0x6E
+#define OP2_INST_COS             0x6F
+#define OP2_INST_LSHR_INT        0x71
+#define OP2_INST_MULLO_INT       0x73
+#define OP2_INST_LSHL_INT        0x72
+#define OP2_INST_FLT_TO_UINT     0x79
+
+#define OP3_INST_MULADD          0x10
+#define OP3_INST_MULADD_D2       0x13
+#define OP3_INST_CNDGT           0x19
+#define OP3_INST_CNDE_INT        0x1C
+/* EXP */
+#define CF_INST_EXP      0x27
+#define CF_INST_EXP_DONE 0x28
+
+/* TEX */
+#define TEX_INST_LD                0x3
+#define TEX_INST_GET_TEXTURE_INFO  0x4
+#define TEX_INST_GET_GRADIENTS_H   0x07
+#define TEX_INST_GET_GRADIENTS_V   0x08
+#define TEX_INST_SAMPLE            0x10
+/* VTX */
+#define VTX_INST_FETCH  0x0
+
+/* EXPORT_TYPE */
+#define EXPORT_TYPE_PIXEL  0x0
+#define EXPORT_TYPE_POS    0x1
+#define EXPORT_TYPE_PARAM  0x2
+
+#define EXPORT_ARRAY_BASE_POS(id)      (0x3C + id)   /* [0, 3] */
+#define EXPORT_ARRAY_BASE_PARAM(id)    id          /* [0, 31] */
+#define EXPORT_ARRAY_BASE_PIX(id)      id
+
+/* exports */
+#define POS(id)   EXPORT_ARRAY_BASE_POS(id)   | (EXPORT_TYPE_POS   << 13)
+#define PARAM(id) EXPORT_ARRAY_BASE_PARAM(id) | (EXPORT_TYPE_PARAM << 13)
+#define PIX(id)   EXPORT_ARRAY_BASE_PIX(id)   | (EXPORT_TYPE_PIXEL << 13)
+#define POS0   POS(0)
+#define PARAM0 PARAM(0)
+#define PARAM1 PARAM(1)
+#define PARAM2 PARAM(2)
+#define PARAM3 PARAM(3)
+#define PARAM4 PARAM(4)
+#define PARAM5 PARAM(5)
+#define PARAM6 PARAM(6)
+#define PARAM7 PARAM(7)
+#define PARAM8 PARAM(8)
+#define PARAM9 PARAM(9)
+#define PARAM10 PARAM(10)
+#define PARAM11 PARAM(11)
+#define PIX0   PIX(0)
+
+/* registers */
+#define __     (0x80) /* invalid regitser (write mask off) */
+#define _R(x)  x
+#define _R0    _R(0x0)
+#define _R1    _R(0x1)
+#define _R2    _R(0x2)
+#define _R3    _R(0x3)
+#define _R4    _R(0x4)
+#define _R5    _R(0x5)
+#define _R6    _R(0x6)
+#define _R7    _R(0x7)
+#define _R8    _R(0x8)
+#define _R9    _R(0x9)
+#define _R10    _R(0xA)
+#define _R11    _R(0xB)
+#define _R12    _R(0xC)
+#define _R13    _R(0xD)
+#define _R14    _R(0xE)
+#define _R15    _R(0xF)
+#define _R16    _R(0x10)
+#define _R17    _R(0x11)
+#define _R18    _R(0x12)
+#define _R19    _R(0x13)
+#define _R20    _R(0x14)
+#define _R21    _R(0x15)
+#define _R22    _R(0x16)
+#define _R23    _R(0x17)
+#define _R24    _R(0x18)
+#define _R25    _R(0x19)
+#define _R26    _R(0x1A)
+#define _R27    _R(0x1B)
+#define _R28    _R(0x1C)
+#define _R29    _R(0x1D)
+#define _R30    _R(0x1E)
+#define _R31    _R(0x1F)
+
+#define _R120    _R(0x78)
+#define _R121    _R(0x79)
+#define _R122    _R(0x7A)
+#define _R123    _R(0x7B)
+#define _R124    _R(0x7C)
+#define _R125    _R(0x7D)
+#define _R126    _R(0x7E)
+#define _R127    _R(0x7F)
+
+/* texture */
+#define _t(x)  x
+#define _t0    _t(0x0)
+#define _t1    _t(0x1)
+
+/* sampler */
+#define _s(x)  x
+#define _s0    _s(0x0)
+#define _s1    _s(0x1)
+
+#define _b(x)  x
+
+/* const files */
+#define ALU_SRC_CONST_FILE_BASE 0x100
+#define _C(x) (ALU_SRC_CONST_FILE_BASE + (x))
+
+#define CALL_FS to_QWORD(CF_DWORD0(0), CF_DWORD1(0,0,0,0,0,CF_INST_CALL_FS))
+
+#define TEX(addr, cnt) to_QWORD(CF_DWORD0(addr), CF_DWORD1(0x0, 0x0, CF_COND_ACTIVE, (cnt - 1), 0x0, CF_INST_TEX))
+#define VTX(addr, cnt) to_QWORD(CF_DWORD0(addr), CF_DWORD1(0x0, 0x0, CF_COND_ACTIVE, (cnt - 1), 0x0, CF_INST_VTX))
+#define LOOP_END(addr) to_QWORD(CF_DWORD0(addr), CF_DWORD1(0x0, 0x0, CF_COND_ACTIVE, 0x0, 0x0, CF_INST_LOOP_END))
+#define LOOP_START_DX10(addr) to_QWORD(CF_DWORD0(addr), CF_DWORD1(0x0, 0x0, CF_COND_ACTIVE, 0x0, 0x0, CF_INST_LOOP_START_DX10))
+#define JUMP(popCount, addr) to_QWORD(CF_DWORD0(addr), CF_DWORD1(popCount, 0x0, CF_COND_ACTIVE, 0x0, 0x0, CF_INST_JUMP))
+#define ELSE(popCount, addr) to_QWORD(CF_DWORD0(addr), CF_DWORD1(popCount, 0x0, CF_COND_ACTIVE, 0x0, 0x0, CF_INST_ELSE))
+#define POP(popCount, addr) to_QWORD(CF_DWORD0(addr), CF_DWORD1(popCount, 0x0, CF_COND_ACTIVE, 0x0, 0x0, CF_INST_POP))
+
+#define ALU(addr, cnt) to_QWORD(CF_ALU_WORD0(addr, 0x0, 0x0, 0x0), CF_ALU_WORD1(0x0, 0x0, 0x0, (cnt - 1), 0x0, CF_INST_ALU))
+#define ALU_PUSH_BEFORE(addr, cnt) to_QWORD(CF_ALU_WORD0(addr, 0x0, 0x0, 0x0), CF_ALU_WORD1(0x0, 0x0, 0x0, (cnt - 1), 0x0, CF_INST_ALU_PUSH_BEFORE))
+#define ALU_POP_AFTER(addr, cnt) to_QWORD(CF_ALU_WORD0(addr, 0x0, 0x0, 0x0), CF_ALU_WORD1(0x0, 0x0, 0x0, (cnt - 1), 0x0, CF_INST_ALU_POP_AFTER))
+#define ALU_POP2_AFTER(addr, cnt) to_QWORD(CF_ALU_WORD0(addr, 0x0, 0x0, 0x0), CF_ALU_WORD1(0x0, 0x0, 0x0, (cnt - 1), 0x0, CF_INST_ALU_POP2_AFTER))
+#define ALU_BREAK(addr, cnt) to_QWORD(CF_ALU_WORD0(addr, 0x0, 0x0, 0x0), CF_ALU_WORD1(0x0, 0x0, 0x0, (cnt - 1), 0x0, CF_INST_ALU_BREAK))
+#define ALU_ELSE_AFTER(addr, cnt) to_QWORD(CF_ALU_WORD0(addr, 0x0, 0x0, 0x0), CF_ALU_WORD1(0x0, 0x0, 0x0, (cnt - 1), 0x0, CF_INST_ALU_ELSE_AFTER))
+
+#define EXP_DONE(dstReg_and_type, srcReg, srcSelX, srcSelY, srcSelZ, srcSelW) to_QWORD(CF_EXP_WORD0(dstReg_and_type, srcReg, 0x0, 0x0, 0x0), \
+   CF_EXP_WORD1(srcSelX, srcSelY, srcSelZ, srcSelW, 0x0, CF_INST_EXP_DONE))
+
+#define EXP(dstReg_and_type, srcReg, srcSelX, srcSelY, srcSelZ, srcSelW) to_QWORD(CF_EXP_WORD0(dstReg_and_type, srcReg, 0x0, 0x0, 0x0), \
+   CF_EXP_WORD1(srcSelX, srcSelY, srcSelZ, srcSelW, 0x0, CF_INST_EXP))
+
+#define MEM_RING(arrayBase, dstReg, writeMask, arraySize, elemSize) \
+   to_QWORD(CF_ALLOC_EXPORT_WORD0(arrayBase, 0x00, dstReg, 0x00, 0x00, elemSize), \
+   CF_ALLOC_EXPORT_WORD1_BUF(arraySize, writeMask, CF_INST_MEM_RING))
+
+#define EMIT_VERTEX to_QWORD(0, CF_DWORD1(0, 0, 0, 0, 0, CF_INST_EMIT_VERTEX))
+
+#define ALU_OP2(inst, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, omod) \
+   to_QWORD(ALU_WORD0(((src0Sel) & ((1 << 13) - 1)), 0x0, src0Chan, 0x0, ((src1Sel) & ((1 << 13) - 1)), 0x0, src1Chan, 0x0, 0x0, 0x0), \
+   ALU_WORD1_OP2(((src0Sel) >> 13), ((src1Sel) >> 13), 0x0, 0x0, (((dstGpr&__) >> 7) ^ 0x1), omod, inst, 0x0, 0x0, dstGpr, 0x0, dstChan, 0x0))
+
+#define ALU_OP3(inst, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, src2Sel, src2Chan) \
+   to_QWORD(ALU_WORD0(src0Sel, 0x0, src0Chan, 0x0, src1Sel, 0x0, src1Chan, 0x0, 0x0, 0x0), \
+   ALU_WORD1_OP3(src2Sel, 0x0, src2Chan, 0x0, inst, 0x0, dstGpr, 0x0, dstChan, 0x0))
+
+#define ALU_NOP(dstGpr, dstChan) \
+   ALU_OP2(OP2_INST_NOP, dstGpr, dstChan, ALU_SRC_PV, dstChan, ALU_SRC_PV, dstChan, ALU_OMOD_OFF)
+
+#define ALU_ADD(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_ADD, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_ADD_x2(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_ADD, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_M2)
+
+#define ALU_ADD_D2(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_ADD, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_D2)
+
+#define ALU_MUL(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_MUL, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_MUL_x2(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_MUL, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_M2)
+
+#define ALU_MUL_x4(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_MUL, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_M4)
+
+#define ALU_MUL_IEEE(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_MUL_IEEE, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_MUL_IEEE_x2(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_MUL_IEEE, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_M2)
+
+#define ALU_MUL_IEEE_x4(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_MUL_IEEE, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_M4)
+
+#define ALU_FRACT(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_FRACT, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_FLOOR(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_FLOOR, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_SQRT_IEEE(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_SQRT_IEEE, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_SQRT_IEEE_D2(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_SQRT_IEEE, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_D2)
+
+#define ALU_MOVA_INT(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_MOVA_INT, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_MOV(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_MOV, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_MOV_D2(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_MOV, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_D2)
+
+#define ALU_MOV_x2(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_MOV, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_M2)
+
+#define ALU_MOV_x4(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_MOV, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_M4)
+
+#define ALU_DOT4_IEEE(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_DOT4_IEEE, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_DOT4(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_DOT4, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_PRED_SETGT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_PRED_SETGT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_KILLGT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_KILLGT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_SETE_DX10(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_SETE_DX10, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_SETGT_DX10(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_SETGT_DX10, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_SETGE_DX10(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_SETGE_DX10, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_SETGT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_SETGT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_PRED_SETE_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_PRED_SETE_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_PRED_SETGT_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_PRED_SETGT_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_PRED_SETGE_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_PRED_SETGE_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_KILLGE_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_KILLGE_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_SETGT_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_SETGT_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_SETGE_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_SETGE_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_ADD_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_ADD_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_PRED_SETNE_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_PRED_SETNE_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_MIN(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_MIN, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_MAX(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_MAX, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_MAX_DX10(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_MAX_DX10, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_LSHR_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_LSHR_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_MULLO_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_MULLO_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_LSHL_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_LSHL_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_AND_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_AND_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_SETE_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_SETE_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_KILLE_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_KILLE_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_KILLGT_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_KILLGT_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_KILLNE_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_KILLNE_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_SETNE_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_SETNE_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_OR_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan) \
+   ALU_OP2(OP2_INST_OR_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, ALU_OMOD_OFF)
+
+#define ALU_INT_TO_FLT(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_INT_TO_FLT, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_FLT_TO_UINT(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_FLT_TO_UINT, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_FLT_TO_INT(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_FLT_TO_INT, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_EXP_IEEE(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_EXP_IEEE, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_LOG_CLAMPED(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_LOG_CLAMPED, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_RECIP_IEEE(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_RECIP_IEEE, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_RECIPSQRT_IEEE(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_RECIPSQRT_IEEE, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_SIN(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_SIN, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_COS(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_COS, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+#define ALU_COS_D2(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_COS, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_D2)
+
+#define ALU_NOT_INT(dstGpr, dstChan, src0Sel, src0Chan) \
+   ALU_OP2(OP2_INST_NOT_INT, dstGpr, dstChan, src0Sel, src0Chan, ALU_SRC_0, 0x0, ALU_OMOD_OFF)
+
+
+#define ALU_MULADD(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, src2Sel, src2Chan) \
+   ALU_OP3(OP3_INST_MULADD, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, src2Sel, src2Chan)
+
+#define ALU_MULADD_D2(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, src2Sel, src2Chan) \
+   ALU_OP3(OP3_INST_MULADD_D2, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, src2Sel, src2Chan)
+
+#define ALU_CNDGT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, src2Sel, src2Chan) \
+   ALU_OP3(OP3_INST_CNDGT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, src2Sel, src2Chan)
+
+#define ALU_CNDE_INT(dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, src2Sel, src2Chan) \
+   ALU_OP3(OP3_INST_CNDE_INT, dstGpr, dstChan, src0Sel, src0Chan, src1Sel, src1Chan, src2Sel, src2Chan)
+
+#define TEX_LD(dstReg, dstSelX, dstSelY, dstSelZ, dstSelW, srcReg, srcSelX, srcSelY, srcSelZ, srcSelW, resourceID, samplerID)\
+   to_QWORD(TEX_WORD0(TEX_INST_LD, 0x0, 0x0, resourceID, srcReg, 0x0, 0x0), \
+   TEX_WORD1(dstReg, 0x0, dstSelX, dstSelY, dstSelZ, dstSelW, 0x0, TEX_UNNORMALIZED, TEX_UNNORMALIZED, TEX_UNNORMALIZED, TEX_UNNORMALIZED)), \
+   to_QWORD(TEX_WORD2(0x0, 0x0, 0x0, samplerID, srcSelX, srcSelY, srcSelZ, srcSelW), 0x00000000)
+
+#define TEX_SAMPLE(dstReg, dstSelX, dstSelY, dstSelZ, dstSelW, srcReg, srcSelX, srcSelY, srcSelZ, srcSelW, resourceID, samplerID)\
+   to_QWORD(TEX_WORD0(TEX_INST_SAMPLE, 0x0, 0x0, resourceID, srcReg, 0x0, 0x0), \
+   TEX_WORD1(dstReg, 0x0, dstSelX, dstSelY, dstSelZ, dstSelW, 0x0, TEX_NORMALIZED, TEX_NORMALIZED, TEX_NORMALIZED, TEX_NORMALIZED)), \
+   to_QWORD(TEX_WORD2(0x0, 0x0, 0x0, samplerID, _x, _y, _0, _x), 0x00000000)
+
+#define TEX_GET_GRADIENTS_H(dstReg, dstSelX, dstSelY, dstSelZ, dstSelW, srcReg, srcSelX, srcSelY, srcSelZ, srcSelW, resourceID, samplerID)\
+   to_QWORD(TEX_WORD0(TEX_INST_GET_GRADIENTS_H, 0x0, 0x0, resourceID, srcReg, 0x0, 0x0), \
+   TEX_WORD1(dstReg, 0x0, dstSelX, dstSelY, dstSelZ, dstSelW, 0x0, TEX_NORMALIZED, TEX_NORMALIZED, TEX_NORMALIZED, TEX_NORMALIZED)), \
+   to_QWORD(TEX_WORD2(0x0, 0x0, 0x0, samplerID, _x, _y, _z, _x), 0x00000000)
+
+#define TEX_GET_GRADIENTS_V(dstReg, dstSelX, dstSelY, dstSelZ, dstSelW, srcReg, srcSelX, srcSelY, srcSelZ, srcSelW, resourceID, samplerID)\
+   to_QWORD(TEX_WORD0(TEX_INST_GET_GRADIENTS_V, 0x0, 0x0, resourceID, srcReg, 0x0, 0x0), \
+   TEX_WORD1(dstReg, 0x0, dstSelX, dstSelY, dstSelZ, dstSelW, 0x0, TEX_NORMALIZED, TEX_NORMALIZED, TEX_NORMALIZED, TEX_NORMALIZED)), \
+   to_QWORD(TEX_WORD2(0x0, 0x0, 0x0, samplerID, _x, _y, _z, _x), 0x00000000)
+
+#define TEX_GET_TEXTURE_INFO(dstReg, dstSelX, dstSelY, dstSelZ, dstSelW, srcReg, srcSelX, srcSelY, srcSelZ, srcSelW, resourceID, samplerID)\
+   to_QWORD(TEX_WORD0(TEX_INST_GET_TEXTURE_INFO, 0x0, 0x0, resourceID, srcReg, 0x0, 0x0), \
+   TEX_WORD1(dstReg, 0x0, dstSelX, dstSelY, dstSelZ, dstSelW, 0x0, TEX_NORMALIZED, TEX_NORMALIZED, TEX_NORMALIZED, TEX_NORMALIZED)), \
+   to_QWORD(TEX_WORD2(0x0, 0x0, 0x0, samplerID, srcSelX, srcSelY, srcSelZ, srcSelW), 0x00000000)
+
+
+#define VTX_FETCH(dstReg, dstSelX, dstSelY, dstSelZ, dstSelW, srcReg, srcSelX, buffer_id, type, mega, offset) \
+   to_QWORD(VTX_WORD0(VTX_INST_FETCH, type, buffer_id, srcReg, srcSelX, mega), VTX_WORD1(dstReg, dstSelX, dstSelY, dstSelZ, dstSelW)) , \
+   to_QWORD(VTX_WORD2(offset, (mega >> 31)), 0x00000000)
+
+#define _x2(v)        v, v
+#define _x4(v)   _x2(v), _x2(v)
+#define _x8(v)   _x4(v), _x4(v)
+#define _x16(v)  _x8(v), _x8(v)
+
+#define _x9(v)   _x8(v), v
+#define _x30(v) _x16(v), _x8(v), _x4(v),_x2(v)
+#define _x31(v) _x30(v), v
+
+#endif /* GX2_SHADER_INL_H */

libultraship/libultraship/Lib/ImGui/backends/wiiu/imgui_impl_gx2.cpp

@@ -0,0 +1,334 @@
+// dear imgui: Renderer Backend for the Nintendo Wii U using GX2
+#include "imgui.h"
+#include "imgui_impl_gx2.h"
+#include <stdio.h>
+#include <stdint.h>     // intptr_t
+#include <malloc.h>     // memalign
+
+// GX2 includes
+#include <whb/gfx.h>
+#include <gx2/registers.h>
+#include <gx2/draw.h>
+#include <gx2/utils.h>
+#include <gx2/mem.h>
+#include <gx2r/surface.h>
+
+// Include shader data
+#include "shaders/shader.h"
+
+// GX2 Data
+struct ImGui_ImplGX2_Data
+{
+    uint32_t VertexBufferSize;
+    void* VertexBuffer;
+    uint32_t IndexBufferSize;
+    void* IndexBuffer;
+
+    ImGui_ImplGX2_Texture* FontTexture;
+
+    WHBGfxShaderGroup* ShaderGroup;
+
+    ImGui_ImplGX2_Data() { memset(this, 0, sizeof(*this)); }
+};
+
+// Backend data stored in io.BackendRendererUserData
+static ImGui_ImplGX2_Data* ImGui_ImplGX2_GetBackendData()
+{
+    return ImGui::GetCurrentContext() ? (ImGui_ImplGX2_Data*)ImGui::GetIO().BackendRendererUserData : NULL;
+}
+
+// Functions
+bool    ImGui_ImplGX2_Init()
+{
+    ImGuiIO& io = ImGui::GetIO();
+    IM_ASSERT(io.BackendRendererUserData == NULL && "Already initialized a renderer backend!");
+
+    ImGui_ImplGX2_Data* bd = IM_NEW(ImGui_ImplGX2_Data)();
+    io.BackendRendererUserData = (void*)bd;
+    io.BackendRendererName = "imgui_impl_gx2";
+
+    io.BackendFlags |= ImGuiBackendFlags_RendererHasVtxOffset;  // We can honor the ImDrawCmd::VtxOffset field, allowing for large meshes.
+
+    return true;
+}
+
+void    ImGui_ImplGX2_Shutdown()
+{
+    ImGui_ImplGX2_Data* bd = ImGui_ImplGX2_GetBackendData();
+    IM_ASSERT(bd != NULL && "No renderer backend to shutdown, or already shutdown?");
+    ImGuiIO& io = ImGui::GetIO();
+
+    ImGui_ImplGX2_DestroyDeviceObjects();
+    io.BackendRendererName = NULL;
+    io.BackendRendererUserData = NULL;
+    IM_DELETE(bd);
+}
+
+void    ImGui_ImplGX2_NewFrame()
+{
+    ImGui_ImplGX2_Data* bd = ImGui_ImplGX2_GetBackendData();
+    IM_ASSERT(bd != NULL && "Did you call ImGui_ImplGX2_Init()?");
+
+    if (!bd->ShaderGroup)
+        ImGui_ImplGX2_CreateDeviceObjects();
+}
+
+static void ImGui_ImplGX2_SetupRenderState(ImDrawData* draw_data, int fb_width, int fb_height)
+{
+    ImGui_ImplGX2_Data* bd = ImGui_ImplGX2_GetBackendData();
+
+    // Setup render state: alpha-blending enabled, no face culling, no depth testing
+    GX2SetColorControl(GX2_LOGIC_OP_COPY, 0xFF, FALSE, TRUE);
+    GX2SetBlendControl(GX2_RENDER_TARGET_0,
+                GX2_BLEND_MODE_SRC_ALPHA,
+                GX2_BLEND_MODE_INV_SRC_ALPHA,
+                GX2_BLEND_COMBINE_MODE_ADD,
+                TRUE,
+                GX2_BLEND_MODE_ONE,
+                GX2_BLEND_MODE_INV_SRC_ALPHA,
+                GX2_BLEND_COMBINE_MODE_ADD);
+    GX2SetCullOnlyControl(GX2_FRONT_FACE_CCW, FALSE, FALSE);
+    GX2SetDepthOnlyControl(FALSE, FALSE, GX2_COMPARE_FUNC_NEVER);
+
+    // Setup viewport, orthographic projection matrix
+    // Our visible imgui space lies from draw_data->DisplayPos (top left) to draw_data->DisplayPos+data_data->DisplaySize (bottom right). DisplayPos is (0,0) for single viewport apps.
+    GX2SetViewport(0, 0, (float)fb_width, (float)fb_height, 0.0f, 1.0f);
+
+    GX2SetFetchShader(&bd->ShaderGroup->fetchShader);
+    GX2SetVertexShader(bd->ShaderGroup->vertexShader);
+    GX2SetPixelShader(bd->ShaderGroup->pixelShader);
+
+    float L = draw_data->DisplayPos.x;
+    float R = draw_data->DisplayPos.x + draw_data->DisplaySize.x;
+    float T = draw_data->DisplayPos.y;
+    float B = draw_data->DisplayPos.y + draw_data->DisplaySize.y;
+    const float ortho_projection[4][4] =
+    {
+        { 2.0f/(R-L),   0.0f,         0.0f,   0.0f },
+        { 0.0f,         2.0f/(T-B),   0.0f,   0.0f },
+        { 0.0f,         0.0f,        -1.0f,   0.0f },
+        { (R+L)/(L-R),  (T+B)/(B-T),  0.0f,   1.0f },
+    };
+
+    GX2SetVertexUniformReg(0, sizeof(ortho_projection) / sizeof(float), &ortho_projection[0][0]);
+}
+
+void    ImGui_ImplGX2_RenderDrawData(ImDrawData* draw_data)
+{
+    // Avoid rendering when minimized, scale coordinates for retina displays (screen coordinates != framebuffer coordinates)
+    int fb_width = (int)(draw_data->DisplaySize.x * draw_data->FramebufferScale.x);
+    int fb_height = (int)(draw_data->DisplaySize.y * draw_data->FramebufferScale.y);
+    if (fb_width <= 0 || fb_height <= 0)
+        return;
+
+    ImGui_ImplGX2_Data* bd = ImGui_ImplGX2_GetBackendData();
+
+    ImGui_ImplGX2_SetupRenderState(draw_data, fb_width, fb_height);
+
+    // Will project scissor/clipping rectangles into framebuffer space
+    ImVec2 clip_off = draw_data->DisplayPos;         // (0,0) unless using multi-viewports
+    ImVec2 clip_scale = draw_data->FramebufferScale; // (1,1) unless using retina display which are often (2,2)
+
+    // Create continuous vertex/index buffers
+    uint32_t vtx_buffer_size = (uint32_t)draw_data->TotalVtxCount * (int)sizeof(ImDrawVert);
+    uint32_t idx_buffer_size = (uint32_t)draw_data->TotalIdxCount * (int)sizeof(ImDrawIdx);
+
+    // Grow buffers if needed
+    if (bd->VertexBufferSize < vtx_buffer_size)
+    {
+        bd->VertexBufferSize = vtx_buffer_size;
+        free(bd->VertexBuffer);
+        bd->VertexBuffer = memalign(GX2_VERTEX_BUFFER_ALIGNMENT, vtx_buffer_size);
+    }
+    if (bd->IndexBufferSize < idx_buffer_size)
+    {
+        bd->IndexBufferSize = idx_buffer_size;
+        free(bd->IndexBuffer);
+        bd->IndexBuffer = memalign(GX2_INDEX_BUFFER_ALIGNMENT, idx_buffer_size);
+    }
+
+    // Copy data into continuous buffers
+    uint8_t* vtx_dst = (uint8_t*)bd->VertexBuffer;
+    uint8_t* idx_dst = (uint8_t*)bd->IndexBuffer;
+    for (int n = 0; n < draw_data->CmdListsCount; n++)
+    {
+        const ImDrawList* cmd_list = draw_data->CmdLists[n];
+
+        memcpy(vtx_dst, cmd_list->VtxBuffer.Data, cmd_list->VtxBuffer.Size * sizeof(ImDrawVert));
+        vtx_dst += cmd_list->VtxBuffer.Size * sizeof(ImDrawVert);
+
+        memcpy(idx_dst, cmd_list->IdxBuffer.Data, cmd_list->IdxBuffer.Size * sizeof(ImDrawIdx));
+        idx_dst += cmd_list->IdxBuffer.Size * sizeof(ImDrawIdx);
+    }
+
+    // Flush memory
+    GX2Invalidate(GX2_INVALIDATE_MODE_CPU_ATTRIBUTE_BUFFER, bd->VertexBuffer, vtx_buffer_size);
+    GX2Invalidate(GX2_INVALIDATE_MODE_CPU, bd->IndexBuffer, idx_buffer_size);
+
+    GX2SetAttribBuffer(0, vtx_buffer_size, sizeof(ImDrawVert), bd->VertexBuffer);
+
+    // Render command lists
+    // (Because we merged all buffers into a single one, we maintain our own offset into them)
+    int global_vtx_offset = 0;
+    int global_idx_offset = 0;
+    for (int n = 0; n < draw_data->CmdListsCount; n++)
+    {
+        const ImDrawList* cmd_list = draw_data->CmdLists[n];
+
+        for (int cmd_i = 0; cmd_i < cmd_list->CmdBuffer.Size; cmd_i++)
+        {
+            const ImDrawCmd* pcmd = &cmd_list->CmdBuffer[cmd_i];
+            if (pcmd->UserCallback != NULL)
+            {
+                // User callback, registered via ImDrawList::AddCallback()
+                // (ImDrawCallback_ResetRenderState is a special callback value used by the user to request the renderer to reset render state.)
+                if (pcmd->UserCallback == ImDrawCallback_ResetRenderState)
+                    ImGui_ImplGX2_SetupRenderState(draw_data, fb_width, fb_height);
+                else
+                    pcmd->UserCallback(cmd_list, pcmd);
+            }
+            else
+            {
+                // Project scissor/clipping rectangles into framebuffer space
+                ImVec2 clip_min((pcmd->ClipRect.x - clip_off.x) * clip_scale.x, (pcmd->ClipRect.y - clip_off.y) * clip_scale.y);
+                ImVec2 clip_max((pcmd->ClipRect.z - clip_off.x) * clip_scale.x, (pcmd->ClipRect.w - clip_off.y) * clip_scale.y);
+                if (clip_max.x <= clip_min.x || clip_max.y <= clip_min.y)
+                    continue;
+
+                // Discard invalid draws to avoid GPU crash
+                if (clip_min.x < 0.0f || clip_min.y < 0.0f || clip_max.x > fb_width || clip_max.y > fb_height || !pcmd->ElemCount)
+                    continue;
+
+                // Apply scissor/clipping rectangle
+                GX2SetScissor((uint32_t)clip_min.x, (uint32_t)clip_min.y, (uint32_t)(clip_max.x - clip_min.x), (uint32_t)(clip_max.y - clip_min.y));
+
+                // Bind texture, Draw
+                ImGui_ImplGX2_Texture* tex = (ImGui_ImplGX2_Texture*) pcmd->GetTexID();
+                IM_ASSERT(tex && "TextureID cannot be NULL");
+
+                GX2SetPixelTexture(tex->Texture, 0);
+                GX2SetPixelSampler(tex->Sampler, 0);
+
+                GX2DrawIndexedEx(GX2_PRIMITIVE_MODE_TRIANGLES, pcmd->ElemCount,
+                    sizeof(ImDrawIdx) == 2 ? GX2_INDEX_TYPE_U16 : GX2_INDEX_TYPE_U32,
+                    (uint8_t*) bd->IndexBuffer + (pcmd->IdxOffset + global_idx_offset) * sizeof(ImDrawIdx),
+                    global_vtx_offset + pcmd->VtxOffset, 1);
+            }
+        }
+        global_idx_offset += cmd_list->IdxBuffer.Size;
+        global_vtx_offset += cmd_list->VtxBuffer.Size;
+    }
+}
+
+bool ImGui_ImplGX2_CreateFontsTexture()
+{
+    ImGuiIO& io = ImGui::GetIO();
+    ImGui_ImplGX2_Data* bd = ImGui_ImplGX2_GetBackendData();
+
+    if (bd->FontTexture)
+    {
+        return false;
+    }
+
+    // Build texture atlas
+    unsigned char* src_pixels;
+    int width, height;
+    io.Fonts->GetTexDataAsRGBA32(&src_pixels, &width, &height);   // Load as RGBA 32-bit (75% of the memory is wasted, but default font is so small) because it is more likely to be compatible with user's existing shaders.
+
+    bd->FontTexture = IM_NEW(ImGui_ImplGX2_Texture)();
+
+    GX2Texture* tex = IM_NEW(GX2Texture)();
+    memset(tex, 0, sizeof(GX2Texture));
+    bd->FontTexture->Texture = tex;
+
+    tex->surface.dim = GX2_SURFACE_DIM_TEXTURE_2D;
+    tex->surface.use = GX2_SURFACE_USE_TEXTURE;
+    tex->surface.width = width;
+    tex->surface.height = height;
+    tex->surface.depth = 1;
+    tex->surface.mipLevels = 1;
+    tex->surface.format = GX2_SURFACE_FORMAT_UNORM_R8_G8_B8_A8;
+    tex->surface.aa = GX2_AA_MODE1X;
+    tex->surface.tileMode = GX2_TILE_MODE_LINEAR_ALIGNED;
+    tex->viewNumSlices = 1;
+    tex->viewNumMips = 1;
+    // swapped for endianness
+    tex->compMap = GX2_COMP_MAP(GX2_SQ_SEL_A, GX2_SQ_SEL_B, GX2_SQ_SEL_G, GX2_SQ_SEL_R);
+
+    GX2RCreateSurface(&tex->surface, GX2R_RESOURCE_BIND_TEXTURE | GX2R_RESOURCE_USAGE_CPU_WRITE | GX2R_RESOURCE_USAGE_GPU_READ);
+    GX2InitTextureRegs(tex);
+
+    unsigned char* dst_pixels = (unsigned char*) GX2RLockSurfaceEx(&tex->surface, 0, GX2R_RESOURCE_BIND_NONE);
+
+    for (int y = 0; y < height; y++) {
+        memcpy(dst_pixels + (y * tex->surface.pitch * 4), src_pixels + (y * width * 4), width * 4);
+    }
+
+    GX2RUnlockSurfaceEx(&tex->surface, 0, GX2R_RESOURCE_BIND_NONE);
+
+    bd->FontTexture->Sampler = IM_NEW(GX2Sampler)();
+    GX2InitSampler(bd->FontTexture->Sampler, GX2_TEX_CLAMP_MODE_CLAMP, GX2_TEX_XY_FILTER_MODE_LINEAR);
+
+    // Store our identifier
+    io.Fonts->SetTexID((ImTextureID) bd->FontTexture);
+
+    return true;
+}
+
+void ImGui_ImplGX2_DestroyFontsTexture()
+{
+    ImGuiIO& io = ImGui::GetIO();
+    ImGui_ImplGX2_Data* bd = ImGui_ImplGX2_GetBackendData();
+    if (bd->FontTexture)
+    {
+        GX2RDestroySurfaceEx(&bd->FontTexture->Texture->surface, GX2R_RESOURCE_BIND_NONE);
+        io.Fonts->SetTexID(0);
+        IM_DELETE(bd->FontTexture->Texture);
+        IM_DELETE(bd->FontTexture->Sampler);
+        IM_DELETE(bd->FontTexture);
+        bd->FontTexture = NULL;
+    }
+}
+
+bool    ImGui_ImplGX2_CreateDeviceObjects()
+{
+    ImGui_ImplGX2_Data* bd = ImGui_ImplGX2_GetBackendData();
+    bd->ShaderGroup = IM_NEW(WHBGfxShaderGroup)();
+
+    if (!WHBGfxLoadGFDShaderGroup(bd->ShaderGroup, 0, shader_gsh))
+    {
+        IM_DELETE(bd->ShaderGroup);
+        return false;
+    }
+
+    WHBGfxInitShaderAttribute(bd->ShaderGroup, "Position", 0, 0, GX2_ATTRIB_FORMAT_FLOAT_32_32);
+    WHBGfxInitShaderAttribute(bd->ShaderGroup, "UV", 0, 8, GX2_ATTRIB_FORMAT_FLOAT_32_32);
+    WHBGfxInitShaderAttribute(bd->ShaderGroup, "Color", 0, 16, GX2_ATTRIB_TYPE_8_8_8_8);
+
+    if (!WHBGfxInitFetchShader(bd->ShaderGroup))
+    {
+        IM_DELETE(bd->ShaderGroup);
+        return false;
+    }
+
+    ImGui_ImplGX2_CreateFontsTexture();
+
+    return true;
+}
+
+void    ImGui_ImplGX2_DestroyDeviceObjects()
+{
+    ImGui_ImplGX2_Data* bd = ImGui_ImplGX2_GetBackendData();
+
+    free(bd->VertexBuffer);
+    bd->VertexBuffer = NULL;
+
+    free(bd->IndexBuffer);
+    bd->IndexBuffer = NULL;
+
+    WHBGfxFreeShaderGroup(bd->ShaderGroup);
+    IM_DELETE(bd->ShaderGroup);
+    bd->ShaderGroup = NULL;
+
+    ImGui_ImplGX2_DestroyFontsTexture();
+}

libultraship/libultraship/Lib/ImGui/backends/wiiu/imgui_impl_gx2.h

@@ -0,0 +1,25 @@
+// dear imgui: Renderer Backend for the Nintendo Wii U using GX2
+#pragma once
+#include "imgui.h"      // IMGUI_IMPL_API
+
+// GX2 Texture / contains a texture and sampler
+// Can be used as a ImTextureID with the GX2 backend
+struct ImGui_ImplGX2_Texture
+{
+    struct GX2Texture* Texture;
+    struct GX2Sampler* Sampler;
+
+    ImGui_ImplGX2_Texture() { memset(this, 0, sizeof(*this)); }
+};
+
+// Backend API
+IMGUI_IMPL_API bool     ImGui_ImplGX2_Init();
+IMGUI_IMPL_API void     ImGui_ImplGX2_Shutdown();
+IMGUI_IMPL_API void     ImGui_ImplGX2_NewFrame();
+IMGUI_IMPL_API void     ImGui_ImplGX2_RenderDrawData(ImDrawData* draw_data);
+
+// (Optional) Called by Init/NewFrame/Shutdown
+IMGUI_IMPL_API bool     ImGui_ImplGX2_CreateFontsTexture();
+IMGUI_IMPL_API void     ImGui_ImplGX2_DestroyFontsTexture();
+IMGUI_IMPL_API bool     ImGui_ImplGX2_CreateDeviceObjects();
+IMGUI_IMPL_API void     ImGui_ImplGX2_DestroyDeviceObjects();

libultraship/libultraship/Lib/ImGui/backends/wiiu/imgui_impl_wiiu.cpp

@@ -0,0 +1,294 @@
+// dear imgui: Platform Backend for the Wii U
+#include "imgui.h"
+#include "imgui_impl_wiiu.h"
+#include <stdlib.h> // malloc/free
+
+// Software keyboard
+#include <nn/swkbd.h>
+
+// Wii U Data
+struct ImGui_ImplWiiU_Data
+{
+    nn::swkbd::CreateArg CreateArg;
+    nn::swkbd::AppearArg AppearArg;
+    nn::swkbd::ControllerType LastController;
+
+    bool WantedTextInput;
+    bool WasTouched;
+
+    ImGui_ImplWiiU_Data()   { memset((void*)this, 0, sizeof(*this)); }
+};
+
+// Backend data stored in io.BackendPlatformUserData
+static ImGui_ImplWiiU_Data* ImGui_ImplWiiU_GetBackendData()
+{
+    return ImGui::GetCurrentContext() ? (ImGui_ImplWiiU_Data*)ImGui::GetIO().BackendPlatformUserData : NULL;
+}
+
+bool     ImGui_ImplWiiU_Init()
+{
+    ImGuiIO& io = ImGui::GetIO();
+    IM_ASSERT(io.BackendPlatformUserData == NULL && "Already initialized a platform backend!");
+
+    // Setup backend data
+    ImGui_ImplWiiU_Data* bd = IM_NEW(ImGui_ImplWiiU_Data)();
+    io.BackendPlatformUserData = (void*)bd;
+    io.BackendPlatformName = "imgui_impl_wiiu";
+    io.BackendFlags |= ImGuiBackendFlags_HasGamepad;
+
+    // Initialize and create software keyboard
+    nn::swkbd::CreateArg createArg;
+
+    createArg.workMemory = malloc(nn::swkbd::GetWorkMemorySize(0));
+    createArg.fsClient = (FSClient*) malloc(sizeof(FSClient));
+    if (!createArg.workMemory || !createArg.fsClient)
+    {
+        free(createArg.workMemory);
+        free(createArg.fsClient);
+        return false;
+    }
+
+    FSAddClient(createArg.fsClient, FS_ERROR_FLAG_NONE);
+
+    if (!nn::swkbd::Create(createArg))
+        return false;
+
+    nn::swkbd::AppearArg appearArg;
+    bd->CreateArg = createArg;
+    bd->AppearArg = appearArg;
+
+    return true;
+}
+
+void     ImGui_ImplWiiU_Shutdown()
+{
+    ImGui_ImplWiiU_Data* bd = ImGui_ImplWiiU_GetBackendData();
+    IM_ASSERT(bd != NULL && "No platform backend to shutdown, or already shutdown?");
+    ImGuiIO& io = ImGui::GetIO();
+
+    // Destroy software keyboard
+    nn::swkbd::Destroy();
+    free(bd->CreateArg.workMemory);
+    bd->CreateArg.workMemory = NULL;
+
+    if (bd->CreateArg.fsClient)
+    {
+        FSDelClient(bd->CreateArg.fsClient, FS_ERROR_FLAG_NONE);
+        free(bd->CreateArg.fsClient);
+        bd->CreateArg.fsClient = NULL;
+    }
+
+    io.BackendPlatformName = NULL;
+    io.BackendPlatformUserData = NULL;
+    IM_DELETE(bd);
+}
+
+static void ImGui_ImplWiiU_UpdateKeyboardInput(ImGui_ImplWiiU_ControllerInput* input)
+{
+    ImGuiIO& io = ImGui::GetIO();
+
+    VPADGetTPCalibratedPoint(VPAD_CHAN_0, &input->vpad->tpNormal, &input->vpad->tpNormal);
+
+    nn::swkbd::ControllerInfo controllerInfo;
+    controllerInfo.vpad = input->vpad;
+    for (int i = 0; i < 4; i++)
+        controllerInfo.kpad[i] = input->kpad[i];
+
+    nn::swkbd::Calc(controllerInfo);
+
+    if (nn::swkbd::IsNeedCalcSubThreadFont())
+        nn::swkbd::CalcSubThreadFont();
+
+    if (nn::swkbd::IsNeedCalcSubThreadPredict())
+        nn::swkbd::CalcSubThreadPredict();
+
+    if (nn::swkbd::IsDecideOkButton(NULL))
+    {
+        // Add entered text
+        const char16_t* string = nn::swkbd::GetInputFormString();
+        for (int i = 0; *string; string++)
+            io.AddInputCharacterUTF16(string[i]);
+
+        // close keyboard
+        nn::swkbd::DisappearInputForm();
+    }
+
+    if (nn::swkbd::IsDecideCancelButton(NULL))
+        nn::swkbd::DisappearInputForm();
+}
+
+static void ImGui_ImplWiiU_UpdateTouchInput(ImGui_ImplWiiU_ControllerInput* input)
+{
+    if (!input->vpad)
+        return;
+
+    ImGui_ImplWiiU_Data* bd = ImGui_ImplWiiU_GetBackendData();
+    ImGuiIO& io = ImGui::GetIO();
+
+    VPADTouchData touch;
+    VPADGetTPCalibratedPoint(VPAD_CHAN_0, &touch, &input->vpad->tpNormal);
+
+    if (touch.touched)
+    {
+        float scale_x = (io.DisplaySize.x / io.DisplayFramebufferScale.x) / 1280.0f;
+        float scale_y = (io.DisplaySize.y / io.DisplayFramebufferScale.y) / 720.0f;
+        io.AddMousePosEvent(touch.x * scale_x, touch.y * scale_y);
+    }
+
+    if (touch.touched != bd->WasTouched)
+    {
+        io.AddMouseButtonEvent(ImGuiMouseButton_Left, touch.touched);
+        bd->WasTouched = touch.touched;
+        bd->LastController = nn::swkbd::ControllerType::DrcGamepad;
+    }
+}
+
+#define IM_CLAMP(V, MN, MX)     ((V) < (MN) ? (MN) : (V) > (MX) ? (MX) : (V))
+
+static void ImGui_ImplWiiU_UpdateControllerInput(ImGui_ImplWiiU_ControllerInput* input)
+{
+    ImGui_ImplWiiU_Data* bd = ImGui_ImplWiiU_GetBackendData();
+    ImGuiIO& io = ImGui::GetIO();
+
+    // SoH removal to make opening the menu easier
+    // if ((io.ConfigFlags & ImGuiConfigFlags_NavEnableGamepad) == 0)
+    //     return;
+
+    uint32_t vpad_buttons = input->vpad ? input->vpad->hold : 0;
+    uint32_t wpad_buttons = 0;
+    uint32_t classic_buttons = 0;
+    uint32_t pro_buttons = 0;
+
+    float stick_l_x = input->vpad ? input->vpad->leftStick.x : 0.0f;
+    float stick_l_y = input->vpad ? input->vpad->leftStick.y : 0.0f;
+    float stick_r_x = input->vpad ? input->vpad->rightStick.x : 0.0f;
+    float stick_r_y = input->vpad ? input->vpad->rightStick.y : 0.0f;
+
+    for (int i = 0; i < 4; i++)
+    {
+        KPADStatus* kpad = input->kpad[i];
+        if (!kpad)
+            continue;
+
+        switch (kpad->extensionType)
+        {
+        case WPAD_EXT_CORE:
+        case WPAD_EXT_NUNCHUK:
+        case WPAD_EXT_MPLUS:
+        case WPAD_EXT_MPLUS_NUNCHUK:
+            wpad_buttons |= kpad->hold;
+            break;
+        case WPAD_EXT_CLASSIC:
+        case WPAD_EXT_MPLUS_CLASSIC:
+            classic_buttons |= kpad->classic.hold;
+            if (classic_buttons & WPAD_CLASSIC_BUTTON_Y)
+                bd->LastController = (nn::swkbd::ControllerType) i;
+
+            stick_l_x += kpad->classic.leftStick.x;
+            stick_l_y += kpad->classic.leftStick.y;
+            stick_r_x += kpad->classic.rightStick.x;
+            stick_r_y += kpad->classic.rightStick.y;
+            break;
+        case WPAD_EXT_PRO_CONTROLLER:
+            pro_buttons |= kpad->pro.hold;
+            if (pro_buttons & WPAD_PRO_BUTTON_Y)
+                bd->LastController = (nn::swkbd::ControllerType) i;
+
+            stick_l_x += kpad->pro.leftStick.x;
+            stick_l_y += kpad->pro.leftStick.y;
+            stick_r_x += kpad->pro.rightStick.x;
+            stick_r_y += kpad->pro.rightStick.y;
+            break;
+        }
+    }
+
+    if (vpad_buttons & VPAD_BUTTON_Y)
+        bd->LastController = nn::swkbd::ControllerType::DrcGamepad;
+
+    io.AddKeyEvent(ImGuiKey_GamepadStart, (vpad_buttons & VPAD_BUTTON_PLUS) || (wpad_buttons & WPAD_BUTTON_PLUS) || (classic_buttons & WPAD_CLASSIC_BUTTON_PLUS) || (pro_buttons & WPAD_PRO_BUTTON_PLUS));
+    io.AddKeyEvent(ImGuiKey_GamepadBack, (vpad_buttons & VPAD_BUTTON_MINUS) || (wpad_buttons & WPAD_BUTTON_MINUS) || (classic_buttons & WPAD_CLASSIC_BUTTON_MINUS) || (pro_buttons & WPAD_PRO_BUTTON_MINUS));
+    io.AddKeyEvent(ImGuiKey_GamepadFaceLeft, (vpad_buttons & VPAD_BUTTON_X) || (classic_buttons & WPAD_CLASSIC_BUTTON_X) || (pro_buttons & WPAD_PRO_BUTTON_X));
+    io.AddKeyEvent(ImGuiKey_GamepadFaceRight, (vpad_buttons & VPAD_BUTTON_B) || (wpad_buttons & WPAD_BUTTON_B) || (classic_buttons & WPAD_CLASSIC_BUTTON_B) || (pro_buttons & WPAD_PRO_BUTTON_B));
+    io.AddKeyEvent(ImGuiKey_GamepadFaceUp, (vpad_buttons & VPAD_BUTTON_Y) || (classic_buttons & WPAD_CLASSIC_BUTTON_Y) || (pro_buttons & WPAD_PRO_BUTTON_Y));
+    io.AddKeyEvent(ImGuiKey_GamepadFaceDown, (vpad_buttons & VPAD_BUTTON_A) || (wpad_buttons & WPAD_BUTTON_A) || (classic_buttons & WPAD_CLASSIC_BUTTON_A) || (pro_buttons & WPAD_PRO_BUTTON_A));
+    io.AddKeyEvent(ImGuiKey_GamepadDpadLeft, (vpad_buttons & VPAD_BUTTON_LEFT) || (wpad_buttons & WPAD_BUTTON_LEFT) || (classic_buttons & WPAD_CLASSIC_BUTTON_LEFT) || (pro_buttons & WPAD_PRO_BUTTON_LEFT));
+    io.AddKeyEvent(ImGuiKey_GamepadDpadRight, (vpad_buttons & VPAD_BUTTON_RIGHT) || (wpad_buttons & WPAD_BUTTON_RIGHT) || (classic_buttons & WPAD_CLASSIC_BUTTON_RIGHT) || (pro_buttons & WPAD_PRO_BUTTON_RIGHT));
+    io.AddKeyEvent(ImGuiKey_GamepadDpadUp, (vpad_buttons & VPAD_BUTTON_UP) || (wpad_buttons & WPAD_BUTTON_UP) || (classic_buttons & WPAD_CLASSIC_BUTTON_UP) || (pro_buttons & WPAD_PRO_BUTTON_UP));
+    io.AddKeyEvent(ImGuiKey_GamepadDpadDown, (vpad_buttons & VPAD_BUTTON_DOWN) || (wpad_buttons & WPAD_BUTTON_DOWN) || (classic_buttons & WPAD_CLASSIC_BUTTON_DOWN) || (pro_buttons & WPAD_PRO_BUTTON_DOWN));
+    io.AddKeyEvent(ImGuiKey_GamepadL1, (vpad_buttons & VPAD_BUTTON_L) || (classic_buttons & WPAD_CLASSIC_BUTTON_L) || (pro_buttons & WPAD_PRO_TRIGGER_L));
+    io.AddKeyEvent(ImGuiKey_GamepadR1, (vpad_buttons & VPAD_BUTTON_R) || (classic_buttons & WPAD_CLASSIC_BUTTON_R) || (pro_buttons & WPAD_PRO_TRIGGER_R));
+    io.AddKeyEvent(ImGuiKey_GamepadL2, (vpad_buttons & VPAD_BUTTON_ZL) || (classic_buttons & WPAD_CLASSIC_BUTTON_ZL) || (pro_buttons & WPAD_PRO_TRIGGER_ZL));
+    io.AddKeyEvent(ImGuiKey_GamepadR2, (vpad_buttons & VPAD_BUTTON_ZR) || (classic_buttons & WPAD_CLASSIC_BUTTON_ZR) || (pro_buttons & WPAD_PRO_TRIGGER_ZR));
+    io.AddKeyEvent(ImGuiKey_GamepadL3, (vpad_buttons & VPAD_BUTTON_STICK_L) || (pro_buttons & WPAD_PRO_BUTTON_STICK_L));
+    io.AddKeyEvent(ImGuiKey_GamepadR3, (vpad_buttons & VPAD_BUTTON_STICK_R) || (pro_buttons & WPAD_PRO_BUTTON_STICK_R));
+
+    stick_l_x = IM_CLAMP(stick_l_x, -1.0f, 1.0f);
+    io.AddKeyAnalogEvent(ImGuiKey_GamepadLStickLeft, stick_l_x < -0.1f, (stick_l_x < -0.1f) ? (stick_l_x * -1.0f) : 0.0f);
+    io.AddKeyAnalogEvent(ImGuiKey_GamepadLStickRight, stick_l_x > 0.1f, (stick_l_x > 0.1f) ? stick_l_x : 0.0f);
+
+    stick_l_y = IM_CLAMP(stick_l_y, -1.0f, 1.0f);
+    io.AddKeyAnalogEvent(ImGuiKey_GamepadLStickUp, stick_l_y > 0.1f, (stick_l_y > 0.1f) ? stick_l_y : 0.0f);
+    io.AddKeyAnalogEvent(ImGuiKey_GamepadLStickDown, stick_l_y < -0.1f, (stick_l_y < -0.1f) ? (stick_l_y * -1.0f) : 0.0f);
+
+    stick_r_x = IM_CLAMP(stick_r_x, -1.0f, 1.0f);
+    io.AddKeyAnalogEvent(ImGuiKey_GamepadRStickLeft, stick_r_x < -0.1f, (stick_r_x < -0.1f) ? (stick_r_x * -1.0f) : 0.0f);
+    io.AddKeyAnalogEvent(ImGuiKey_GamepadRStickRight, stick_r_x > 0.1f, (stick_r_x > 0.1f) ? stick_r_x : 0.0f);
+
+    stick_r_y = IM_CLAMP(stick_r_y, -1.0f, 1.0f);
+    io.AddKeyAnalogEvent(ImGuiKey_GamepadRStickUp, stick_r_y > 0.1f, (stick_r_y > 0.1f) ? stick_r_y : 0.0f);
+    io.AddKeyAnalogEvent(ImGuiKey_GamepadRStickDown, stick_r_y < -0.1f, (stick_r_y < -0.1f) ? (stick_r_y * -1.0f) : 0.0f);
+}
+
+bool     ImGui_ImplWiiU_ProcessInput(ImGui_ImplWiiU_ControllerInput* input)
+{
+    ImGui_ImplWiiU_Data* bd = ImGui_ImplWiiU_GetBackendData();
+    IM_ASSERT(bd != NULL && "Did you call ImGui_ImplWiiU_Init()?");
+    ImGuiIO& io = ImGui::GetIO();
+
+    // Show keyboard if wanted
+    if (io.WantTextInput && !bd->WantedTextInput) 
+    {
+        // Open the keyboard for the controller which requested the text input
+        bd->AppearArg.keyboardArg.configArg.controllerType = bd->LastController;
+
+        if (nn::swkbd::GetStateInputForm() == nn::swkbd::State::Hidden)
+            nn::swkbd::AppearInputForm(bd->AppearArg);
+    }
+    bd->WantedTextInput = io.WantTextInput;
+
+    // Update keyboard input
+    if (nn::swkbd::GetStateInputForm() != nn::swkbd::State::Hidden)
+    {
+        ImGui_ImplWiiU_UpdateKeyboardInput(input);
+        return true;
+    }
+
+    // Update touch screen
+    ImGui_ImplWiiU_UpdateTouchInput(input);
+
+    // Update gamepads
+    ImGui_ImplWiiU_UpdateControllerInput(input);
+
+    return false;
+}
+
+void     ImGui_ImplWiiU_DrawKeyboardOverlay(ImGui_ImplWiiU_KeyboardOverlayType type)
+{
+    ImGui_ImplWiiU_Data* bd = ImGui_ImplWiiU_GetBackendData();
+    IM_ASSERT(bd != NULL && "Did you call ImGui_ImplWiiU_Init()?");
+
+    if (nn::swkbd::GetStateInputForm() != nn::swkbd::State::Hidden)
+    {
+        if (type == ImGui_KeyboardOverlay_Auto)
+        {
+            if (bd->LastController == nn::swkbd::ControllerType::DrcGamepad)
+                nn::swkbd::DrawDRC();
+            else
+                nn::swkbd::DrawTV();
+        }
+        else if (type == ImGui_KeyboardOverlay_DRC)
+            nn::swkbd::DrawDRC();
+        else if (type == ImGui_KeyboardOverlay_TV)
+            nn::swkbd::DrawTV();
+    }
+}

libultraship/libultraship/Lib/ImGui/backends/wiiu/imgui_impl_wiiu.h

@@ -0,0 +1,29 @@
+// dear imgui: Platform Backend for the Wii U
+#pragma once
+#include "imgui.h"      // IMGUI_IMPL_API
+
+// GamePad Input
+#include <vpad/input.h>
+// Controller Input
+#include <padscore/kpad.h>
+
+struct ImGui_ImplWiiU_ControllerInput
+{
+   VPADStatus* vpad = nullptr;
+   KPADStatus* kpad[4] = { nullptr };
+};
+
+enum ImGui_ImplWiiU_KeyboardOverlayType
+{
+   //! Draw for the DRC
+   ImGui_KeyboardOverlay_DRC,
+   //! Draw for the TV
+   ImGui_KeyboardOverlay_TV,
+   //! Draw for the controller which requested the keyboard
+   ImGui_KeyboardOverlay_Auto
+};
+
+IMGUI_IMPL_API bool     ImGui_ImplWiiU_Init();
+IMGUI_IMPL_API void     ImGui_ImplWiiU_Shutdown();
+IMGUI_IMPL_API bool     ImGui_ImplWiiU_ProcessInput(ImGui_ImplWiiU_ControllerInput* input);
+IMGUI_IMPL_API void     ImGui_ImplWiiU_DrawKeyboardOverlay(ImGui_ImplWiiU_KeyboardOverlayType type = ImGui_KeyboardOverlay_Auto);

libultraship/libultraship/Lib/ImGui/backends/wiiu/shaders/build-shaders.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+
+# to build shaders you need to place a copy of latte-assembler into the current directory
+# latte-assembler is part of decaf-emu <https://github.com/decaf-emu/decaf-emu>
+
+# shader
+./latte-assembler assemble --vsh=shader.vsh --psh=shader.psh shader.gsh
+xxd -i shader.gsh > shader.h

libultraship/libultraship/Lib/ImGui/backends/wiiu/shaders/shader.h

@@ -0,0 +1,148 @@
+unsigned char shader_gsh[] = {
+  0x47, 0x66, 0x78, 0x32, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x07,
+  0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x42, 0x4c, 0x4b, 0x7b,
+  0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x01, 0xd8, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x04, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x01, 0xff, 0xff, 0x01, 0x00,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xf8, 0x00, 0x00, 0x00, 0x03,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02,
+  0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff,
+  0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff,
+  0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff,
+  0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff,
+  0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff,
+  0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff,
+  0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff,
+  0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff,
+  0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff,
+  0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0xff, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x01, 0x70,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0xd0, 0x60, 0x01, 0x34,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x03, 0xd0, 0x60, 0x01, 0x48, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0xca, 0x70, 0x01, 0x78, 0x00, 0x00, 0x00, 0x1d, 0x00, 0x00, 0x00, 0x01,
+  0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xca, 0x70, 0x01, 0x80,
+  0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0xca, 0x70, 0x01, 0x8c, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x01, 0xca, 0x70, 0x01, 0x90, 0x00, 0x00, 0x00, 0x0b,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x50, 0x72, 0x6f, 0x6a,
+  0x4d, 0x74, 0x78, 0x00, 0x50, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e,
+  0x00, 0x00, 0x00, 0x00, 0x55, 0x56, 0x00, 0x00, 0x43, 0x6f, 0x6c, 0x6f,
+  0x72, 0x00, 0x00, 0x00, 0xd0, 0x60, 0x00, 0xe8, 0xd0, 0x60, 0x01, 0x08,
+  0xca, 0x70, 0x01, 0x34, 0xca, 0x70, 0x01, 0x48, 0xca, 0x70, 0x01, 0x58,
+  0xca, 0x70, 0x01, 0x68, 0x7d, 0x42, 0x4c, 0x4b, 0x00, 0x00, 0x00, 0x28,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x98, 0xd0, 0x60, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x20, 0xd0, 0x60, 0x01, 0x78, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x06, 0xd0, 0x60, 0x01, 0x98, 0x42, 0x4c, 0x4b, 0x7b,
+  0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x01, 0x70, 0x00, 0x00, 0x00, 0x01,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x09,
+  0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x34, 0xa0, 0x3c, 0xa0, 0x00, 0x00,
+  0x88, 0x06, 0x00, 0x94, 0x00, 0x40, 0x01, 0x00, 0x08, 0x09, 0x80, 0x13,
+  0x01, 0xc0, 0x01, 0x00, 0x88, 0x06, 0x20, 0x14, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xfd, 0x60, 0x20, 0x00,
+  0x80, 0x00, 0x00, 0x00, 0xfd, 0x64, 0xa0, 0x00, 0x80, 0x00, 0x00, 0x20,
+  0xfd, 0x68, 0x20, 0x01, 0x80, 0x00, 0x00, 0x40, 0xfd, 0x6c, 0xa0, 0x81,
+  0x80, 0x00, 0x00, 0x60, 0x00, 0x00, 0x80, 0x3f, 0x00, 0x00, 0x80, 0x3f,
+  0x00, 0x00, 0x80, 0x3f, 0x00, 0x00, 0x80, 0x3f, 0x01, 0x24, 0x20, 0x00,
+  0xfe, 0x00, 0xe2, 0x0f, 0x01, 0x24, 0xa0, 0x00, 0xfe, 0x04, 0xe2, 0x2f,
+  0x01, 0x24, 0x20, 0x01, 0xfe, 0x08, 0xe2, 0x4f, 0x01, 0x24, 0xa0, 0x81,
+  0xfe, 0x0c, 0xe2, 0x6f, 0x01, 0x00, 0x20, 0x00, 0xfe, 0x00, 0x22, 0x00,
+  0x01, 0x00, 0xa0, 0x00, 0xfe, 0x04, 0x22, 0x20, 0x01, 0x00, 0x20, 0x01,
+  0xfe, 0x08, 0x22, 0x40, 0x01, 0x00, 0xa0, 0x81, 0xfe, 0x0c, 0x22, 0x60,
+  0x42, 0x4c, 0x4b, 0x7b, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x01, 0x2c,
+  0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02,
+  0x00, 0x00, 0x00, 0x02, 0x14, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01, 0x01,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x10,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x90, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x01, 0xd0, 0x60, 0x00, 0xe8, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0xca, 0x70, 0x00, 0xf4, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00,
+  0x54, 0x65, 0x78, 0x74, 0x75, 0x72, 0x65, 0x00, 0xd0, 0x60, 0x00, 0xd4,
+  0xca, 0x70, 0x00, 0xe8, 0x7d, 0x42, 0x4c, 0x4b, 0x00, 0x00, 0x00, 0x28,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xfc, 0xd0, 0x60, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x08, 0xd0, 0x60, 0x00, 0xf4, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x02, 0xd0, 0x60, 0x00, 0xfc, 0x42, 0x4c, 0x4b, 0x7b,
+  0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x01, 0x90, 0x00, 0x00, 0x00, 0x03,
+  0x00, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00, 0xc0, 0x80,
+  0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0xa0, 0x00, 0x00, 0x00, 0x00,
+  0x88, 0x06, 0x20, 0x94, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x80, 0x01,
+  0x90, 0x00, 0x00, 0x00, 0x00, 0x24, 0x00, 0x01, 0x90, 0x00, 0x00, 0x20,
+  0x00, 0x28, 0x80, 0x00, 0x90, 0x00, 0x00, 0x40, 0x00, 0x2c, 0x00, 0x80,
+  0x90, 0x00, 0x00, 0x60, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x10, 0x0d, 0xf0,
+  0x00, 0x00, 0x80, 0x10, 0x00, 0x00, 0x00, 0x00, 0x42, 0x4c, 0x4b, 0x7b,
+  0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04,
+  0x00, 0x00, 0x00, 0x00
+};
+unsigned int shader_gsh_len = 1732;

libultraship/libultraship/Lib/ImGui/backends/wiiu/shaders/shader.psh

@@ -0,0 +1,25 @@
+; $MODE = "UniformRegister"
+
+; $NUM_SPI_PS_INPUT_CNTL = 2
+; Frag_UV R0
+; $SPI_PS_INPUT_CNTL[0].SEMANTIC = 0
+; $SPI_PS_INPUT_CNTL[0].DEFAULT_VAL = 1
+; Frag_Color R1
+; $SPI_PS_INPUT_CNTL[1].SEMANTIC = 1
+; $SPI_PS_INPUT_CNTL[1].DEFAULT_VAL = 1
+
+; $SAMPLER_VARS[0].name = "Texture"
+; $SAMPLER_VARS[0].type = "SAMPLER2D"
+; $SAMPLER_VARS[0].location = 0
+
+; Note: R1 is swapped for endianness
+
+00 TEX: ADDR(48) CNT(1) VALID_PIX
+    0  SAMPLE R0, R0.xy0x, t0, s0
+01 ALU: ADDR(32) CNT(4)
+    1  x: MUL R0.x, R0.x, R1.w
+       y: MUL R0.y, R0.y, R1.z
+       z: MUL R0.z, R0.z, R1.y
+       w: MUL R0.w, R0.w, R1.x
+02 EXP_DONE: PIX0, R0
+END_OF_PROGRAM

libultraship/libultraship/Lib/ImGui/backends/wiiu/shaders/shader.vsh

@@ -0,0 +1,47 @@
+; $MODE = "UniformRegister"
+
+; $SPI_VS_OUT_CONFIG.VS_EXPORT_COUNT = 1
+; $NUM_SPI_VS_OUT_ID = 1
+; uv
+; $SPI_VS_OUT_ID[0].SEMANTIC_0 = 0
+; color
+; $SPI_VS_OUT_ID[0].SEMANTIC_1 = 1
+
+; C0
+; $UNIFORM_VARS[0].name = "ProjMtx"
+; $UNIFORM_VARS[0].type = "mat4"
+; $UNIFORM_VARS[0].count = 1
+; $UNIFORM_VARS[0].block = -1
+; $UNIFORM_VARS[0].offset = 0
+
+; R1
+; $ATTRIB_VARS[0].name = "Position"
+; $ATTRIB_VARS[0].type = "vec2"
+; $ATTRIB_VARS[0].location = 0
+; R2
+; $ATTRIB_VARS[1].name = "UV"
+; $ATTRIB_VARS[1].type = "vec2"
+; $ATTRIB_VARS[1].location = 1
+; R3
+; $ATTRIB_VARS[2].name = "Color"
+; $ATTRIB_VARS[2].type = "vec4"
+; $ATTRIB_VARS[2].location = 2
+
+00 CALL_FS NO_BARRIER
+01 ALU: ADDR(32) CNT(14)
+    0  x: MUL    ____,   1.0f, C3.x
+       y: MUL    ____,   1.0f, C3.y
+       z: MUL    ____,   1.0f, C3.z
+       w: MUL    ____,   1.0f, C3.w
+    1  x: MULADD R127.x, R1.y, C1.x, PV0.x
+       y: MULADD R127.y, R1.y, C1.y, PV0.y
+       z: MULADD R127.z, R1.y, C1.z, PV0.z
+       w: MULADD R127.w, R1.y, C1.w, PV0.w
+    2  x: MULADD R1.x,   R1.x, C0.x, PV0.x
+       y: MULADD R1.y,   R1.x, C0.y, PV0.y
+       z: MULADD R1.z,   R1.x, C0.z, PV0.z
+       w: MULADD R1.w,   R1.x, C0.w, PV0.w
+02 EXP_DONE: POS0, R1
+03 EXP: PARAM0, R2.xy00 NO_BARRIER
+04 EXP_DONE: PARAM1, R3 NO_BARRIER
+END_OF_PROGRAM

libultraship/libultraship/Lib/ImGui/imgui_draw.cpp

@@ -761,7 +761,12 @@ void ImDrawList::AddPolyline(const ImVec2* points, const int points_count, ImU32
 
         // Temporary buffer
         // The first <points_count> items are normals at each line point, then after that there are either 2 or 4 temp points for each line point
+#ifdef __WIIU__
+        // avoid using alloca on the Wii U
+        ImVec2 temp_normals[points_count * ((use_texture || !thick_line) ? 3 : 5) * sizeof(ImVec2)]; //-V630
+#else
         ImVec2* temp_normals = (ImVec2*)alloca(points_count * ((use_texture || !thick_line) ? 3 : 5) * sizeof(ImVec2)); //-V630
+#endif
         ImVec2* temp_points = temp_normals + points_count;
 
         // Calculate normals (tangents) for each line segment
@@ -1009,7 +1014,12 @@ void ImDrawList::AddConvexPolyFilled(const ImVec2* points, const int points_coun
         }
 
         // Compute normals
+#ifdef __WIIU__
+        // Avoid using alloca on the Wii U
+        ImVec2 temp_normals[points_count * sizeof(ImVec2)]; //-V630
+#else
         ImVec2* temp_normals = (ImVec2*)alloca(points_count * sizeof(ImVec2)); //-V630
+#endif
         for (int i0 = points_count - 1, i1 = 0; i1 < points_count; i0 = i1++)
         {
             const ImVec2& p0 = points[i0];

libultraship/libultraship/Lib/Mercury/Mercury.cpp

@@ -111,6 +111,12 @@ void Mercury::reload() {
         return;
     }
     std::ifstream ifs(this->path_);
+
+#ifdef __WIIU__
+    alignas(0x40) char buffer[8192];
+    ifs.rdbuf()->pubsetbuf(buffer, sizeof(buffer));
+#endif
+
     try {
         this->rjson = json::parse(ifs);
         this->vjson = this->rjson.flatten();
@@ -122,5 +128,11 @@ void Mercury::reload() {
 
 void Mercury::save() const {
     std::ofstream file(this->path_);
+
+#ifdef __WIIU__
+    alignas(0x40) char buffer[8192];
+    file.rdbuf()->pubsetbuf(buffer, sizeof(buffer));
+#endif
+
     file << this->vjson.unflatten().dump(4);
 }

libultraship/libultraship/Lib/spdlog/include/spdlog/details/os-inl.h

@@ -286,7 +286,7 @@ SPDLOG_INLINE int utc_minutes_offset(const std::tm &tm)
     return offset;
 #else
 
-#    if defined(sun) || defined(__sun) || defined(_AIX) || (!defined(_BSD_SOURCE) && !defined(_GNU_SOURCE))
+#    if defined(sun) || defined(__sun) || defined(_AIX) || (!defined(_BSD_SOURCE) && !defined(_GNU_SOURCE)) || defined(__WIIU__)
     // 'tm_gmtoff' field is BSD extension and it's missing on SunOS/Solaris
     struct helper
     {

libultraship/libultraship/WiiUController.cpp

@@ -0,0 +1,509 @@
+#ifdef __WIIU__
+#include "WiiUController.h"
+#include "GlobalCtx2.h"
+#include "Window.h"
+#include "ImGuiImpl.h"
+
+#include "WiiUImpl.h"
+
+namespace Ship {
+    WiiUController::WiiUController(WPADChan chan) : Controller(), chan(chan) {
+        connected = false;
+        extensionType = (WPADExtensionType) -1;
+
+        controllerName = std::string("Wii U Controller ") + std::to_string((int) chan) + " (Disconnected)";
+    }
+
+    bool WiiUController::Open() {
+        KPADError error;
+        KPADStatus* status = Ship::WiiU::GetKPADStatus(chan, &error);
+        if (!status || error != KPAD_ERROR_OK) {
+            Close();
+            return false;
+        }
+
+        connected = true;
+        extensionType = (WPADExtensionType) status->extensionType;
+
+        // Create a GUID and name based on extension and channel
+        GUID = std::string("WiiU") + GetControllerExtensionName() + std::to_string((int) chan);
+        controllerName = std::string("Wii U ") + GetControllerExtensionName() + std::string(" ") + std::to_string((int) chan);
+
+        return true;
+    }
+
+    void WiiUController::Close() {
+        connected = false;
+        extensionType = (WPADExtensionType) -1;
+
+        for (int i = 0; i < MAXCONTROLLERS; i++) {
+            getPressedButtons(i) = 0;
+            getLeftStickX(i) = 0;
+            getLeftStickY(i) = 0;
+            getRightStickX(i) = 0;
+            getRightStickY(i) = 0;
+            getGyroX(i) = 0;
+            getGyroY(i) = 0;
+        }
+    }
+
+    void WiiUController::ReadFromSource(int32_t virtualSlot) {
+        auto profile = getProfile(virtualSlot);
+
+        KPADError error;
+        KPADStatus* status = Ship::WiiU::GetKPADStatus(chan, &error);
+        if (!status) {
+            Close();
+            return;
+        }
+
+        // Make sure the controller type doesn't change after opening
+        if (status->extensionType != extensionType) {
+            Close();
+            return;
+        }
+
+        getPressedButtons(virtualSlot) = 0;
+        getLeftStickX(virtualSlot) = 0;
+        getLeftStickY(virtualSlot) = 0;
+        getRightStickX(virtualSlot) = 0;
+        getRightStickY(virtualSlot) = 0;
+        getGyroX(virtualSlot) = 0;
+        getGyroY(virtualSlot) = 0;
+
+        if (error != KPAD_ERROR_OK) {
+            return;
+        }
+
+        int16_t stickX = 0;
+        int16_t stickY = 0;
+        int16_t camX = 0;
+        int16_t camY = 0;
+
+        switch (extensionType) {
+            case WPAD_EXT_PRO_CONTROLLER:
+                for (uint32_t i = WPAD_PRO_BUTTON_UP; i <= WPAD_PRO_STICK_R_EMULATION_UP; i <<= 1) {
+                    if (profile->Mappings.contains(i)) {
+                        // check if the stick is mapped to an analog stick
+                        if (i >= WPAD_PRO_STICK_L_EMULATION_LEFT) {
+                            float axisX = i >= WPAD_PRO_STICK_R_EMULATION_LEFT ? status->pro.rightStick.x : status->pro.leftStick.x;
+                            float axisY = i >= WPAD_PRO_STICK_R_EMULATION_LEFT ? status->pro.rightStick.y : status->pro.leftStick.y;
+
+                            if (profile->Mappings[i] == BTN_STICKRIGHT || profile->Mappings[i] == BTN_STICKLEFT) {
+                                stickX = axisX * 85;
+                                continue;
+                            } else if (profile->Mappings[i] == BTN_STICKDOWN || profile->Mappings[i] == BTN_STICKUP) {
+                                stickY = axisY * 85;
+                                continue;
+                            } else if (profile->Mappings[i] == BTN_VSTICKRIGHT || profile->Mappings[i] == BTN_VSTICKLEFT) {
+                                camX = axisX * 85;
+                                continue;
+                            } else if (profile->Mappings[i] == BTN_VSTICKDOWN || profile->Mappings[i] == BTN_VSTICKUP) {
+                                camY = axisY * 85;
+                                continue;
+                            }
+                        }
+
+                        if (status->pro.hold & i) {
+                            getPressedButtons(virtualSlot) |= profile->Mappings[i];
+                        }
+                    }
+                }
+                break;
+            case WPAD_EXT_CLASSIC:
+            case WPAD_EXT_MPLUS_CLASSIC:
+                for (uint32_t i = WPAD_CLASSIC_BUTTON_UP; i <= WPAD_CLASSIC_STICK_R_EMULATION_UP; i <<= 1) {
+                    if (profile->Mappings.contains(i)) {
+                        // check if the stick is mapped to an analog stick
+                        if (i >= WPAD_CLASSIC_STICK_L_EMULATION_LEFT) {
+                            float axisX = i >= WPAD_CLASSIC_STICK_R_EMULATION_LEFT ? status->classic.rightStick.x : status->classic.leftStick.x;
+                            float axisY = i >= WPAD_CLASSIC_STICK_R_EMULATION_LEFT ? status->classic.rightStick.y : status->classic.leftStick.y;
+
+                            if (profile->Mappings[i] == BTN_STICKRIGHT || profile->Mappings[i] == BTN_STICKLEFT) {
+                                stickX = axisX * 85;
+                                continue;
+                            } else if (profile->Mappings[i] == BTN_STICKDOWN || profile->Mappings[i] == BTN_STICKUP) {
+                                stickY = axisY * 85;
+                                continue;
+                            } else if (profile->Mappings[i] == BTN_VSTICKRIGHT || profile->Mappings[i] == BTN_VSTICKLEFT) {
+                                camX = axisX * 85;
+                                continue;
+                            } else if (profile->Mappings[i] == BTN_VSTICKDOWN || profile->Mappings[i] == BTN_VSTICKUP) {
+                                camY = axisY * 85;
+                                continue;
+                            }
+                        }
+
+                        if (status->classic.hold & i) {
+                            getPressedButtons(virtualSlot) |= profile->Mappings[i];
+                        }
+                    }
+                }
+                break;
+            case WPAD_EXT_NUNCHUK:
+            case WPAD_EXT_MPLUS_NUNCHUK:
+            case WPAD_EXT_MPLUS:
+            case WPAD_EXT_CORE:
+                for (uint32_t i = WPAD_BUTTON_LEFT; i <= WPAD_BUTTON_HOME; i <<= 1) {
+                    if (profile->Mappings.contains(i)) {
+                        if (status->hold & i) {
+                            getPressedButtons(virtualSlot) |= profile->Mappings[i];
+                        }
+                    }
+                }
+                stickX += status->nunchuck.stick.x * 85;
+                stickY += status->nunchuck.stick.y * 85;
+                break;
+        }
+
+        if (stickX || stickY) {
+            NormalizeStickAxis(virtualSlot, stickX, stickY, profile->AxisDeadzones[0], false);
+        }
+
+        if (camX || camY) {
+            NormalizeStickAxis(virtualSlot, camX, camY, profile->AxisDeadzones[2], true);
+        }
+    }
+
+    void WiiUController::WriteToSource(int32_t virtualSlot, ControllerCallback* controller) {
+        if (getProfile(virtualSlot)->UseRumble) {
+            WPADControlMotor(chan, controller->rumble);
+        }
+    }
+
+    void WiiUController::ClearRawPress() {
+        // Clear already triggered buttons
+        KPADError error;
+        KPADStatus* status = Ship::WiiU::GetKPADStatus(chan, &error);
+        if (status) {
+            switch (extensionType) {
+                case WPAD_EXT_PRO_CONTROLLER:
+                    status->pro.trigger = 0;
+                    break;
+                case WPAD_EXT_CLASSIC:
+                case WPAD_EXT_MPLUS_CLASSIC:
+                    status->classic.trigger = 0;
+                    break;
+                case WPAD_EXT_NUNCHUK:
+                case WPAD_EXT_MPLUS_NUNCHUK:
+                case WPAD_EXT_MPLUS:
+                case WPAD_EXT_CORE:
+                    status->trigger = 0;
+                    break;
+            }
+        }
+    }
+
+    int32_t WiiUController::ReadRawPress() {
+        KPADError error;
+        KPADStatus* status = Ship::WiiU::GetKPADStatus(chan, &error);
+        if (!status || error != KPAD_ERROR_OK) {
+            return -1;
+        }
+
+        switch (extensionType) {
+            case WPAD_EXT_PRO_CONTROLLER:
+                for (uint32_t i = WPAD_PRO_BUTTON_UP; i <= WPAD_PRO_STICK_R_EMULATION_UP; i <<= 1) {
+                    if (status->pro.trigger & i) {
+                        return i;
+                    }
+                }
+
+                if (status->pro.leftStick.x > 0.7f) {
+                    return WPAD_PRO_STICK_L_EMULATION_RIGHT;
+                }
+                if (status->pro.leftStick.x < -0.7f) {
+                    return WPAD_PRO_STICK_L_EMULATION_LEFT;
+                }
+                if (status->pro.leftStick.y > 0.7f) {
+                    return WPAD_PRO_STICK_L_EMULATION_UP;
+                }
+                if (status->pro.leftStick.y < -0.7f) {
+                    return WPAD_PRO_STICK_L_EMULATION_DOWN;
+                }
+
+                if (status->pro.rightStick.x > 0.7f) {
+                    return WPAD_PRO_STICK_R_EMULATION_RIGHT;
+                }
+                if (status->pro.rightStick.x < -0.7f) {
+                    return WPAD_PRO_STICK_R_EMULATION_LEFT;
+                }
+                if (status->pro.rightStick.y > 0.7f) {
+                    return WPAD_PRO_STICK_R_EMULATION_UP;
+                }
+                if (status->pro.rightStick.y < -0.7f) {
+                    return WPAD_PRO_STICK_R_EMULATION_DOWN;
+                }
+                break;
+            case WPAD_EXT_CLASSIC:
+            case WPAD_EXT_MPLUS_CLASSIC:
+                for (uint32_t i = WPAD_CLASSIC_BUTTON_UP; i <= WPAD_CLASSIC_STICK_R_EMULATION_UP; i <<= 1) {
+                    if (status->classic.trigger & i) {
+                        return i;
+                    }
+                }
+
+                if (status->classic.leftStick.x > 0.7f) {
+                    return WPAD_CLASSIC_STICK_L_EMULATION_RIGHT;
+                }
+                if (status->classic.leftStick.x < -0.7f) {
+                    return WPAD_CLASSIC_STICK_L_EMULATION_LEFT;
+                }
+                if (status->classic.leftStick.y > 0.7f) {
+                    return WPAD_CLASSIC_STICK_L_EMULATION_UP;
+                }
+                if (status->classic.leftStick.y < -0.7f) {
+                    return WPAD_CLASSIC_STICK_L_EMULATION_DOWN;
+                }
+
+                if (status->classic.rightStick.x > 0.7f) {
+                    return WPAD_CLASSIC_STICK_R_EMULATION_RIGHT;
+                }
+                if (status->classic.rightStick.x < -0.7f) {
+                    return WPAD_CLASSIC_STICK_R_EMULATION_LEFT;
+                }
+                if (status->classic.rightStick.y > 0.7f) {
+                    return WPAD_CLASSIC_STICK_R_EMULATION_UP;
+                }
+                if (status->classic.rightStick.y < -0.7f) {
+                    return WPAD_CLASSIC_STICK_R_EMULATION_DOWN;
+                }
+                break;
+            case WPAD_EXT_NUNCHUK:
+            case WPAD_EXT_MPLUS_NUNCHUK:
+            case WPAD_EXT_MPLUS:
+            case WPAD_EXT_CORE:
+                for (uint32_t i = WPAD_BUTTON_LEFT; i <= WPAD_BUTTON_HOME; i <<= 1) {
+                    if (status->trigger & i) {
+                        return i;
+                    }
+                }
+                break;
+        }
+
+        return -1;
+    }
+
+    const std::string WiiUController::GetButtonName(int32_t virtualSlot, int n64Button) {
+        std::map<int32_t, int32_t>& Mappings = getProfile(virtualSlot)->Mappings;
+        const auto find = std::find_if(Mappings.begin(), Mappings.end(), [n64Button](const std::pair<int32_t, int32_t>& pair) {
+            return pair.second == n64Button;
+        });
+
+        if (find == Mappings.end()) return "Unknown";
+
+        uint32_t btn = find->first;
+
+        switch (extensionType) {
+            case WPAD_EXT_PRO_CONTROLLER:
+                switch (btn) {
+                    case WPAD_PRO_BUTTON_A: return "A";
+                    case WPAD_PRO_BUTTON_B: return "B";
+                    case WPAD_PRO_BUTTON_X: return "X";
+                    case WPAD_PRO_BUTTON_Y: return "Y";
+                    case WPAD_PRO_BUTTON_LEFT: return "D-pad Left";
+                    case WPAD_PRO_BUTTON_RIGHT: return "D-pad Right";
+                    case WPAD_PRO_BUTTON_UP: return "D-pad Up";
+                    case WPAD_PRO_BUTTON_DOWN: return "D-pad Down";
+                    case WPAD_PRO_TRIGGER_ZL: return "ZL";
+                    case WPAD_PRO_TRIGGER_ZR: return "ZR";
+                    case WPAD_PRO_TRIGGER_L: return "L";
+                    case WPAD_PRO_TRIGGER_R: return "R";
+                    case WPAD_PRO_BUTTON_PLUS: return "+ (START)";
+                    case WPAD_PRO_BUTTON_MINUS: return "- (SELECT)";
+                    case WPAD_PRO_BUTTON_STICK_R: return "Stick Button R";
+                    case WPAD_PRO_BUTTON_STICK_L: return "Stick Button L";
+                    case WPAD_PRO_STICK_R_EMULATION_LEFT: return "Right Stick Left";
+                    case WPAD_PRO_STICK_R_EMULATION_RIGHT: return "Right Stick Right";
+                    case WPAD_PRO_STICK_R_EMULATION_UP: return "Right Stick Up";
+                    case WPAD_PRO_STICK_R_EMULATION_DOWN: return "Right Stick Down";
+                    case WPAD_PRO_STICK_L_EMULATION_LEFT: return "Left Stick Left";
+                    case WPAD_PRO_STICK_L_EMULATION_RIGHT: return "Left Stick Right";
+                    case WPAD_PRO_STICK_L_EMULATION_UP: return "Left Stick Up";
+                    case WPAD_PRO_STICK_L_EMULATION_DOWN: return "Left Stick Down";
+                }
+                break;
+            case WPAD_EXT_CLASSIC:
+            case WPAD_EXT_MPLUS_CLASSIC:
+                switch (btn) {
+                    case WPAD_CLASSIC_BUTTON_A: return "A";
+                    case WPAD_CLASSIC_BUTTON_B: return "B";
+                    case WPAD_CLASSIC_BUTTON_X: return "X";
+                    case WPAD_CLASSIC_BUTTON_Y: return "Y";
+                    case WPAD_CLASSIC_BUTTON_LEFT: return "D-pad Left";
+                    case WPAD_CLASSIC_BUTTON_RIGHT: return "D-pad Right";
+                    case WPAD_CLASSIC_BUTTON_UP: return "D-pad Up";
+                    case WPAD_CLASSIC_BUTTON_DOWN: return "D-pad Down";
+                    case WPAD_CLASSIC_BUTTON_ZL: return "ZL";
+                    case WPAD_CLASSIC_BUTTON_ZR: return "ZR";
+                    case WPAD_CLASSIC_BUTTON_L: return "L";
+                    case WPAD_CLASSIC_BUTTON_R: return "R";
+                    case WPAD_CLASSIC_BUTTON_PLUS: return "+ (START)";
+                    case WPAD_CLASSIC_BUTTON_MINUS: return "- (SELECT)";
+                    case WPAD_CLASSIC_STICK_R_EMULATION_LEFT: return "Right Stick Left";
+                    case WPAD_CLASSIC_STICK_R_EMULATION_RIGHT: return "Right Stick Right";
+                    case WPAD_CLASSIC_STICK_R_EMULATION_UP: return "Right Stick Up";
+                    case WPAD_CLASSIC_STICK_R_EMULATION_DOWN: return "Right Stick Down";
+                    case WPAD_CLASSIC_STICK_L_EMULATION_LEFT: return "Left Stick Left";
+                    case WPAD_CLASSIC_STICK_L_EMULATION_RIGHT: return "Left Stick Right";
+                    case WPAD_CLASSIC_STICK_L_EMULATION_UP: return "Left Stick Up";
+                    case WPAD_CLASSIC_STICK_L_EMULATION_DOWN: return "Left Stick Down";
+                }
+                break;
+            case WPAD_EXT_NUNCHUK:
+            case WPAD_EXT_MPLUS_NUNCHUK:
+            case WPAD_EXT_MPLUS:
+            case WPAD_EXT_CORE:
+                switch (btn) {
+                    case WPAD_BUTTON_A: return "A";
+                    case WPAD_BUTTON_B: return "B";
+                    case WPAD_BUTTON_1: return "1";
+                    case WPAD_BUTTON_2: return "2";
+                    case WPAD_BUTTON_LEFT: return "D-pad Left";
+                    case WPAD_BUTTON_RIGHT: return "D-pad Right";
+                    case WPAD_BUTTON_UP: return "D-pad Up";
+                    case WPAD_BUTTON_DOWN: return "D-pad Down";
+                    case WPAD_BUTTON_Z: return "Z";
+                    case WPAD_BUTTON_C: return "C";
+                    case WPAD_BUTTON_PLUS: return "+ (START)";
+                    case WPAD_BUTTON_MINUS: return "- (SELECT)";
+                }
+                break;
+        }
+
+        return "Unknown";
+    }
+
+    const std::string WiiUController::GetControllerName() {
+        return controllerName;
+    }
+
+    void WiiUController::NormalizeStickAxis(int32_t virtualSlot, float x, float y, uint16_t threshold, bool isRightStick) {
+        auto profile = getProfile(virtualSlot);
+
+        //create scaled circular dead-zone in range {-15 ... +15}
+        auto len = sqrt(x * x + y * y);
+        if (len < threshold) {
+            len = 0;
+        }
+        else if (len > 85.0) {
+            len = 85.0 / len;
+        }
+        else {
+            len = (len - threshold) * 85.0 / (85.0 - threshold) / len;
+        }
+        x *= len;
+        y *= len;
+
+        //bound diagonals to an octagonal range {-68 ... +68}
+        if (x != 0.0 && y != 0.0) {
+            auto slope = y / x;
+            auto edgex = copysign(85.0 / (fabs(slope) + 16.0 / 69.0), x);
+            auto edgey = copysign(std::min(fabs(edgex * slope), 85.0 / (1.0 / fabs(slope) + 16.0 / 69.0)), y);
+            edgex = edgey / slope;
+
+            auto scale = sqrt(edgex * edgex + edgey * edgey) / 85.0;
+            x *= scale;
+            y *= scale;
+        }
+
+        if (isRightStick) {
+            getRightStickX(virtualSlot) = x * profile->AxisSensitivities[2];
+            getRightStickY(virtualSlot) = y * profile->AxisSensitivities[3];
+        } else {
+            getLeftStickX(virtualSlot) = x * profile->AxisSensitivities[0];
+            getLeftStickY(virtualSlot) = y * profile->AxisSensitivities[1];
+        }
+    }
+
+    void WiiUController::CreateDefaultBinding(int32_t virtualSlot) {
+        auto profile = getProfile(virtualSlot);
+        profile->Mappings.clear();
+
+        profile->UseRumble = true;
+        profile->RumbleStrength = 1.0f;
+        profile->UseGyro = false;
+
+        switch (extensionType) {
+            case WPAD_EXT_PRO_CONTROLLER:
+                profile->Mappings[WPAD_PRO_STICK_R_EMULATION_RIGHT] = BTN_CRIGHT;
+                profile->Mappings[WPAD_PRO_STICK_R_EMULATION_LEFT] = BTN_CLEFT;
+                profile->Mappings[WPAD_PRO_STICK_R_EMULATION_DOWN] = BTN_CDOWN;
+                profile->Mappings[WPAD_PRO_STICK_R_EMULATION_UP] = BTN_CUP;
+                profile->Mappings[WPAD_PRO_TRIGGER_ZR] = BTN_R;
+                profile->Mappings[WPAD_PRO_TRIGGER_L] = BTN_L;
+                profile->Mappings[WPAD_PRO_BUTTON_RIGHT] = BTN_DRIGHT;
+                profile->Mappings[WPAD_PRO_BUTTON_LEFT] = BTN_DLEFT;
+                profile->Mappings[WPAD_PRO_BUTTON_DOWN] = BTN_DDOWN;
+                profile->Mappings[WPAD_PRO_BUTTON_UP] = BTN_DUP;
+                profile->Mappings[WPAD_PRO_BUTTON_PLUS] = BTN_START;
+                profile->Mappings[WPAD_PRO_TRIGGER_ZL] = BTN_Z;
+                profile->Mappings[WPAD_PRO_BUTTON_B] = BTN_B;
+                profile->Mappings[WPAD_PRO_BUTTON_A] = BTN_A;
+                profile->Mappings[WPAD_PRO_STICK_L_EMULATION_RIGHT] = BTN_STICKRIGHT;
+                profile->Mappings[WPAD_PRO_STICK_L_EMULATION_LEFT] = BTN_STICKLEFT;
+                profile->Mappings[WPAD_PRO_STICK_L_EMULATION_DOWN] = BTN_STICKDOWN;
+                profile->Mappings[WPAD_PRO_STICK_L_EMULATION_UP] = BTN_STICKUP;
+                break;
+            case WPAD_EXT_CLASSIC:
+            case WPAD_EXT_MPLUS_CLASSIC:
+                profile->Mappings[WPAD_CLASSIC_STICK_R_EMULATION_RIGHT] = BTN_CRIGHT;
+                profile->Mappings[WPAD_CLASSIC_STICK_R_EMULATION_LEFT] = BTN_CLEFT;
+                profile->Mappings[WPAD_CLASSIC_STICK_R_EMULATION_DOWN] = BTN_CDOWN;
+                profile->Mappings[WPAD_CLASSIC_STICK_R_EMULATION_UP] = BTN_CUP;
+                profile->Mappings[WPAD_CLASSIC_BUTTON_ZR] = BTN_R;
+                profile->Mappings[WPAD_CLASSIC_BUTTON_L] = BTN_L;
+                profile->Mappings[WPAD_CLASSIC_BUTTON_RIGHT] = BTN_DRIGHT;
+                profile->Mappings[WPAD_CLASSIC_BUTTON_LEFT] = BTN_DLEFT;
+                profile->Mappings[WPAD_CLASSIC_BUTTON_DOWN] = BTN_DDOWN;
+                profile->Mappings[WPAD_CLASSIC_BUTTON_UP] = BTN_DUP;
+                profile->Mappings[WPAD_CLASSIC_BUTTON_PLUS] = BTN_START;
+                profile->Mappings[WPAD_CLASSIC_BUTTON_ZL] = BTN_Z;
+                profile->Mappings[WPAD_CLASSIC_BUTTON_B] = BTN_B;
+                profile->Mappings[WPAD_CLASSIC_BUTTON_A] = BTN_A;
+                profile->Mappings[WPAD_CLASSIC_STICK_L_EMULATION_RIGHT] = BTN_STICKRIGHT;
+                profile->Mappings[WPAD_CLASSIC_STICK_L_EMULATION_LEFT] = BTN_STICKLEFT;
+                profile->Mappings[WPAD_CLASSIC_STICK_L_EMULATION_DOWN] = BTN_STICKDOWN;
+                profile->Mappings[WPAD_CLASSIC_STICK_L_EMULATION_UP] = BTN_STICKUP;
+                break;
+            case WPAD_EXT_NUNCHUK:
+            case WPAD_EXT_MPLUS_NUNCHUK:
+            case WPAD_EXT_MPLUS:
+            case WPAD_EXT_CORE:
+                profile->Mappings[WPAD_BUTTON_1] = BTN_R;
+                profile->Mappings[WPAD_BUTTON_2] = BTN_L;
+                profile->Mappings[WPAD_BUTTON_RIGHT] = BTN_DRIGHT;
+                profile->Mappings[WPAD_BUTTON_LEFT] = BTN_DLEFT;
+                profile->Mappings[WPAD_BUTTON_DOWN] = BTN_DDOWN;
+                profile->Mappings[WPAD_BUTTON_UP] = BTN_DUP;
+                profile->Mappings[WPAD_BUTTON_PLUS] = BTN_START;
+                profile->Mappings[WPAD_BUTTON_MINUS] = BTN_Z;
+                profile->Mappings[WPAD_BUTTON_B] = BTN_B;
+                profile->Mappings[WPAD_BUTTON_A] = BTN_A;
+                break;
+        }
+
+        for (int i = 0; i < 4; i++) {
+            profile->AxisSensitivities[i] = 1.0f;
+            profile->AxisDeadzones[i] = 0.0f;
+        }
+    }
+
+    std::string WiiUController::GetControllerExtensionName() {
+        switch (extensionType) {
+            case WPAD_EXT_PRO_CONTROLLER:
+                return "ProController";
+            case WPAD_EXT_CLASSIC:
+            case WPAD_EXT_MPLUS_CLASSIC:
+                return "ClassicController";
+            case WPAD_EXT_NUNCHUK:
+            case WPAD_EXT_MPLUS_NUNCHUK:
+            case WPAD_EXT_MPLUS:
+            case WPAD_EXT_CORE:
+                return "WiiRemote";
+        }
+
+        return "Controller";
+    }
+}
+#endif

libultraship/libultraship/WiiUController.h

@@ -0,0 +1,38 @@
+#pragma once
+#include "Controller.h"
+#include <string>
+
+#include <padscore/wpad.h>
+
+namespace Ship {
+    class WiiUController : public Controller {
+        public:
+            WiiUController(WPADChan chan);
+            bool Open();
+            void Close();
+
+            void ReadFromSource(int32_t virtualSlot) override;
+            void WriteToSource(int32_t virtualSlot, ControllerCallback* controller) override;
+            bool Connected() const override { return connected; };
+            bool CanGyro() const override { return false; }
+            bool CanRumble() const override { return true; };
+
+            void ClearRawPress() override;
+            int32_t ReadRawPress() override;
+
+            const std::string GetButtonName(int32_t virtualSlot, int n64Button) override;
+            const std::string GetControllerName() override;
+
+        protected:
+            void NormalizeStickAxis(int32_t virtualSlot, float x, float y, uint16_t threshold, bool isRightStick);
+            void CreateDefaultBinding(int32_t virtualSlot) override;
+
+        private:
+            std::string GetControllerExtensionName();
+            std::string controllerName;
+
+            bool connected;
+            WPADChan chan;
+            WPADExtensionType extensionType;
+    };
+}

libultraship/libultraship/WiiUGamepad.cpp

@@ -0,0 +1,329 @@
+#ifdef __WIIU__
+#include "WiiUGamepad.h"
+#include "GlobalCtx2.h"
+#include "ImGuiImpl.h"
+
+#include "WiiUImpl.h"
+
+namespace Ship {
+    WiiUGamepad::WiiUGamepad() : Controller(), connected(true), rumblePatternStrength(1.0f) {
+        memset(rumblePattern, 0xff, sizeof(rumblePattern));
+
+        GUID = "WiiUGamepad";
+    }
+
+    bool WiiUGamepad::Open() {
+        VPADReadError error;
+        VPADStatus* status = Ship::WiiU::GetVPADStatus(&error);
+        if (!status || error == VPAD_READ_INVALID_CONTROLLER) {
+            Close();
+            return false;
+        }
+
+        return true;
+    }
+
+    void WiiUGamepad::Close() {
+        connected = false;
+
+        for (int i = 0; i < MAXCONTROLLERS; i++) {
+            getPressedButtons(i) = 0;
+            getLeftStickX(i) = 0;
+            getLeftStickY(i) = 0;
+            getRightStickX(i) = 0;
+            getRightStickY(i) = 0;
+            getGyroX(i) = 0;
+            getGyroY(i) = 0;
+        }
+    }
+
+    void WiiUGamepad::ReadFromSource(int32_t virtualSlot) {
+        auto profile = getProfile(virtualSlot);
+
+        VPADReadError error;
+        VPADStatus* status = Ship::WiiU::GetVPADStatus(&error);
+        if (!status) {
+            Close();
+            return;
+        }
+
+        getPressedButtons(virtualSlot) = 0;
+        getLeftStickX(virtualSlot) = 0;
+        getLeftStickY(virtualSlot) = 0;
+        getRightStickX(virtualSlot) = 0;
+        getRightStickY(virtualSlot) = 0;
+        getGyroX(virtualSlot) = 0;
+        getGyroY(virtualSlot) = 0;
+
+        if (error != VPAD_READ_SUCCESS) {
+            return; 
+        }
+
+        int16_t stickX = 0;
+        int16_t stickY = 0;
+        int16_t camX = 0;
+        int16_t camY = 0;
+
+        for (uint32_t i = VPAD_BUTTON_SYNC; i <= VPAD_STICK_L_EMULATION_LEFT; i <<= 1) {
+            if (profile->Mappings.contains(i)) {
+                // check if the stick is mapped to an analog stick
+                if (i >= VPAD_STICK_R_EMULATION_DOWN) {
+                    float axisX = i >= VPAD_STICK_L_EMULATION_DOWN ? status->leftStick.x : status->rightStick.x;
+                    float axisY = i >= VPAD_STICK_L_EMULATION_DOWN ? status->leftStick.y : status->rightStick.y;
+
+                    if (profile->Mappings[i] == BTN_STICKRIGHT || profile->Mappings[i] == BTN_STICKLEFT) {
+                        stickX = axisX * 85;
+                        continue;
+                    } else if (profile->Mappings[i] == BTN_STICKDOWN || profile->Mappings[i] == BTN_STICKUP) {
+                        stickY = axisY * 85;
+                        continue;
+                    } else if (profile->Mappings[i] == BTN_VSTICKRIGHT || profile->Mappings[i] == BTN_VSTICKLEFT) {
+                        camX = axisX * 85;
+                        continue;
+                    } else if (profile->Mappings[i] == BTN_VSTICKDOWN || profile->Mappings[i] == BTN_VSTICKUP) {
+                        camY = axisY * 85;
+                        continue;
+                    }
+                }
+
+                if (status->hold & i) {
+                    getPressedButtons(virtualSlot) |= profile->Mappings[i];
+                }
+            }
+        }
+
+        if (stickX || stickY) {
+            NormalizeStickAxis(virtualSlot, stickX, stickY, profile->AxisDeadzones[0], false);
+        }
+
+        if (camX || camY) {
+            NormalizeStickAxis(virtualSlot, camX, camY, profile->AxisDeadzones[2], true);
+        }
+
+        if (profile->UseGyro) {
+            float gyroX = status->gyro.x * -8.0f;
+            float gyroY = status->gyro.z * 8.0f;
+
+            float gyro_drift_x = profile->GyroData[DRIFT_X] / 100.0f;
+            float gyro_drift_y = profile->GyroData[DRIFT_Y] / 100.0f;
+            const float gyro_sensitivity = profile->GyroData[GYRO_SENSITIVITY];
+
+            if (gyro_drift_x == 0) {
+                gyro_drift_x = gyroX;
+            }
+
+            if (gyro_drift_y == 0) {
+                gyro_drift_y = gyroY;
+            }
+
+            profile->GyroData[DRIFT_X] = gyro_drift_x * 100.0f;
+            profile->GyroData[DRIFT_Y] = gyro_drift_y * 100.0f;
+
+            getGyroX(virtualSlot) = gyroX - gyro_drift_x;
+            getGyroY(virtualSlot) = gyroY - gyro_drift_y;
+
+            getGyroX(virtualSlot) *= gyro_sensitivity;
+            getGyroY(virtualSlot) *= gyro_sensitivity;
+        }
+    }
+
+    void WiiUGamepad::WriteToSource(int32_t virtualSlot, ControllerCallback* controller) {
+        auto profile = getProfile(virtualSlot);
+
+        if (profile->UseRumble) {
+            int patternSize = sizeof(rumblePattern) * 8;
+
+            // update rumble pattern if strength changed
+            if (rumblePatternStrength != profile->RumbleStrength) {
+                rumblePatternStrength = profile->RumbleStrength;
+                if (rumblePatternStrength > 1.0f) {
+                    rumblePatternStrength = 1.0f;
+                } else if (rumblePatternStrength < 0.0f) {
+                    rumblePatternStrength = 0.0f;
+                }
+
+                memset(rumblePattern, 0, sizeof(rumblePattern));
+
+                // distribute wanted amount of bits equally in pattern
+                float scale = (rumblePatternStrength * (1.0f - 0.3f)) + 0.3f;
+                int bitcnt = patternSize * scale;
+                for (int i = 0; i < bitcnt; i++) {
+                    int bitpos = ((i * patternSize) / bitcnt) % patternSize;
+                    rumblePattern[bitpos / 8] |= 1 << (bitpos % 8);
+                }
+            }
+
+            VPADControlMotor(VPAD_CHAN_0, rumblePattern, controller->rumble ? patternSize : 0);
+        }
+    }
+
+    void WiiUGamepad::ClearRawPress() {
+        // Clear already triggered buttons
+        VPADReadError error;
+        VPADStatus* status = Ship::WiiU::GetVPADStatus(&error);
+        if (status) {
+            status->trigger = 0;
+        }
+    }
+
+    int32_t WiiUGamepad::ReadRawPress() {
+        VPADReadError error;
+        VPADStatus* status = Ship::WiiU::GetVPADStatus(&error);
+        if (!status || error != VPAD_READ_SUCCESS) {
+            return -1;
+        }
+
+        for (uint32_t i = VPAD_BUTTON_SYNC; i <= VPAD_BUTTON_STICK_L; i <<= 1) {
+            if (status->trigger & i) {
+                return i;
+            }
+        }
+
+        if (status->leftStick.x > 0.7f) {
+            return VPAD_STICK_L_EMULATION_RIGHT;
+        }
+        if (status->leftStick.x < -0.7f) {
+            return VPAD_STICK_L_EMULATION_LEFT;
+        }
+        if (status->leftStick.y > 0.7f) {
+            return VPAD_STICK_L_EMULATION_UP;
+        }
+        if (status->leftStick.y < -0.7f) {
+            return VPAD_STICK_L_EMULATION_DOWN;
+        }
+
+        if (status->rightStick.x > 0.7f) {
+            return VPAD_STICK_R_EMULATION_RIGHT;
+        }
+        if (status->rightStick.x < -0.7f) {
+            return VPAD_STICK_R_EMULATION_LEFT;
+        }
+        if (status->rightStick.y > 0.7f) {
+            return VPAD_STICK_R_EMULATION_UP;
+        }
+        if (status->rightStick.y < -0.7f) {
+            return VPAD_STICK_R_EMULATION_DOWN;
+        }
+
+        return -1;
+    }
+
+    const std::string WiiUGamepad::GetButtonName(int32_t virtualSlot, int n64Button) {
+        std::map<int32_t, int32_t>& Mappings = getProfile(virtualSlot)->Mappings;
+        const auto find = std::find_if(Mappings.begin(), Mappings.end(), [n64Button](const std::pair<int32_t, int32_t>& pair) {
+            return pair.second == n64Button;
+        });
+
+        if (find == Mappings.end()) return "Unknown";
+
+        uint32_t btn = find->first;
+        switch (btn) {
+            case VPAD_BUTTON_A: return "A";
+            case VPAD_BUTTON_B: return "B";
+            case VPAD_BUTTON_X: return "X";
+            case VPAD_BUTTON_Y: return "Y";
+            case VPAD_BUTTON_LEFT: return "D-pad Left";
+            case VPAD_BUTTON_RIGHT: return "D-pad Right";
+            case VPAD_BUTTON_UP: return "D-pad Up";
+            case VPAD_BUTTON_DOWN: return "D-pad Down";
+            case VPAD_BUTTON_ZL: return "ZL";
+            case VPAD_BUTTON_ZR: return "ZR";
+            case VPAD_BUTTON_L: return "L";
+            case VPAD_BUTTON_R: return "R";
+            case VPAD_BUTTON_PLUS: return "+ (START)";
+            case VPAD_BUTTON_MINUS: return "- (SELECT)";
+            case VPAD_BUTTON_STICK_R: return "Stick Button R";
+            case VPAD_BUTTON_STICK_L: return "Stick Button L";
+            case VPAD_STICK_R_EMULATION_LEFT: return "Right Stick Left";
+            case VPAD_STICK_R_EMULATION_RIGHT: return "Right Stick Right";
+            case VPAD_STICK_R_EMULATION_UP: return "Right Stick Up";
+            case VPAD_STICK_R_EMULATION_DOWN: return "Right Stick Down";
+            case VPAD_STICK_L_EMULATION_LEFT: return "Left Stick Left";
+            case VPAD_STICK_L_EMULATION_RIGHT: return "Left Stick Right";
+            case VPAD_STICK_L_EMULATION_UP: return "Left Stick Up";
+            case VPAD_STICK_L_EMULATION_DOWN: return "Left Stick Down";
+        }
+
+        return "Unknown";
+    }
+
+    const std::string WiiUGamepad::GetControllerName() {
+        return Connected() ? "Wii U GamePad" : "Wii U GamePad (Disconnected)";
+    }
+
+    void WiiUGamepad::NormalizeStickAxis(int32_t virtualSlot, float x, float y, uint16_t threshold, bool isRightStick) {
+        auto profile = getProfile(virtualSlot);
+
+        //create scaled circular dead-zone in range {-15 ... +15}
+        auto len = sqrt(x * x + y * y);
+        if (len < threshold) {
+            len = 0;
+        }
+        else if (len > 85.0) {
+            len = 85.0 / len;
+        }
+        else {
+            len = (len - threshold) * 85.0 / (85.0 - threshold) / len;
+        }
+        x *= len;
+        y *= len;
+
+        //bound diagonals to an octagonal range {-68 ... +68}
+        if (x != 0.0 && y != 0.0) {
+            auto slope = y / x;
+            auto edgex = copysign(85.0 / (fabs(slope) + 16.0 / 69.0), x);
+            auto edgey = copysign(std::min(fabs(edgex * slope), 85.0 / (1.0 / fabs(slope) + 16.0 / 69.0)), y);
+            edgex = edgey / slope;
+
+            auto scale = sqrt(edgex * edgex + edgey * edgey) / 85.0;
+            x *= scale;
+            y *= scale;
+        }
+
+        if (isRightStick) {
+            getRightStickX(virtualSlot) = x * profile->AxisSensitivities[2];
+            getRightStickY(virtualSlot) = y * profile->AxisSensitivities[3];
+        } else {
+            getLeftStickX(virtualSlot) = x * profile->AxisSensitivities[0];
+            getLeftStickY(virtualSlot) = y * profile->AxisSensitivities[1];
+        }
+    }
+
+    void WiiUGamepad::CreateDefaultBinding(int32_t virtualSlot) {
+        auto profile = getProfile(virtualSlot);
+        profile->Mappings.clear();
+
+        profile->UseRumble = true;
+        profile->RumbleStrength = 1.0f;
+        profile->UseGyro = false;
+
+        profile->Mappings[VPAD_STICK_R_EMULATION_RIGHT] = BTN_CRIGHT;
+        profile->Mappings[VPAD_STICK_R_EMULATION_LEFT] = BTN_CLEFT;
+        profile->Mappings[VPAD_STICK_R_EMULATION_DOWN] = BTN_CDOWN;
+        profile->Mappings[VPAD_STICK_R_EMULATION_UP] = BTN_CUP;
+        profile->Mappings[VPAD_BUTTON_ZR] = BTN_R;
+        profile->Mappings[VPAD_BUTTON_L] = BTN_L;
+        profile->Mappings[VPAD_BUTTON_RIGHT] = BTN_DRIGHT;
+        profile->Mappings[VPAD_BUTTON_LEFT] = BTN_DLEFT;
+        profile->Mappings[VPAD_BUTTON_DOWN] = BTN_DDOWN;
+        profile->Mappings[VPAD_BUTTON_UP] = BTN_DUP;
+        profile->Mappings[VPAD_BUTTON_PLUS] = BTN_START;
+        profile->Mappings[VPAD_BUTTON_ZL] = BTN_Z;
+        profile->Mappings[VPAD_BUTTON_B] = BTN_B;
+        profile->Mappings[VPAD_BUTTON_A] = BTN_A;
+        profile->Mappings[VPAD_STICK_L_EMULATION_RIGHT] = BTN_STICKRIGHT;
+        profile->Mappings[VPAD_STICK_L_EMULATION_LEFT] = BTN_STICKLEFT;
+        profile->Mappings[VPAD_STICK_L_EMULATION_DOWN] = BTN_STICKDOWN;
+        profile->Mappings[VPAD_STICK_L_EMULATION_UP] = BTN_STICKUP;
+
+        for (int i = 0; i < 4; i++) {
+            profile->AxisSensitivities[i] = 1.0f;
+            profile->AxisDeadzones[i] = 0.0f;
+        }
+
+        profile->GyroData[GYRO_SENSITIVITY] = 1.0f;
+        profile->GyroData[DRIFT_X] = 0.0f;
+        profile->GyroData[DRIFT_Y] = 0.0f;
+    }
+}
+#endif

libultraship/libultraship/WiiUGamepad.h

@@ -0,0 +1,34 @@
+#pragma once
+#include "Controller.h"
+#include <string>
+
+namespace Ship {
+    class WiiUGamepad : public Controller {
+        public:
+            WiiUGamepad();
+
+            bool Open();
+            void Close();
+
+            void ReadFromSource(int32_t virtualSlot) override;
+            void WriteToSource(int32_t virtualSlot, ControllerCallback* controller) override;
+            bool Connected() const override { return connected; };
+            bool CanGyro() const override { return true; }
+            bool CanRumble() const override { return true; };
+
+            void ClearRawPress() override;
+            int32_t ReadRawPress() override;
+
+            const std::string GetButtonName(int32_t virtualSlot, int n64Button) override;
+            const std::string GetControllerName() override;
+
+        protected:
+            void NormalizeStickAxis(int32_t virtualSlot, float x, float y, uint16_t threshold, bool isRightStick);
+            void CreateDefaultBinding(int32_t virtualSlot) override;
+
+        private:
+            bool connected = true;
+            float rumblePatternStrength;
+            uint8_t rumblePattern[15];
+    };
+}

libultraship/libultraship/WiiUImpl.cpp

@@ -0,0 +1,134 @@
+#ifdef __WIIU__
+#include "WiiUImpl.h"
+
+#include <stdio.h>
+#include <unistd.h>
+#include <sys/iosupport.h>
+
+#include <whb/log.h>
+#include <whb/log_udp.h>
+#include <coreinit/debug.h>
+
+#include "Window.h"
+
+namespace Ship {
+namespace WiiU {
+
+static bool hasVpad = false;
+static VPADReadError vpadError;
+static VPADStatus vpadStatus;
+
+static bool hasKpad[4] = { false };
+static KPADError kpadError[4] = { KPAD_ERROR_OK };
+static KPADStatus kpadStatus[4];
+
+#ifdef _DEBUG
+extern "C" {
+void __wrap_abort() {
+    printf("Abort called.\n");
+    // force a stack trace
+    *(uint32_t*)0xdeadc0de = 0xcafebabe;
+    while(1);
+}
+
+static ssize_t wiiu_log_write(struct _reent* r, void* fd, const char* ptr, size_t len) {
+    char buf[1024];
+    snprintf(buf, sizeof(buf), "%*.*s", len, len, ptr);
+    OSReport(buf);
+    WHBLogWritef("%*.*s", len, len, ptr);
+    return len;
+}
+
+static const devoptab_t dotab_stdout = {
+    .name = "stdout_whb",
+    .write_r = wiiu_log_write,
+};
+};
+#endif
+
+void Init() {
+#ifdef _DEBUG
+    WHBLogUdpInit();
+    WHBLogPrint("Hello World!");
+
+    devoptab_list[STD_OUT] = &dotab_stdout;
+    devoptab_list[STD_ERR] = &dotab_stdout;
+#endif
+
+    // make sure the required folders exist
+    mkdir("/vol/external01/wiiu/", 0755);
+    mkdir("/vol/external01/wiiu/apps/", 0755);
+    mkdir("/vol/external01/wiiu/apps/soh/", 0755);
+
+    chdir("/vol/external01/wiiu/apps/soh/");
+
+    KPADInit();
+    WPADEnableURCC(true);
+}
+
+void Exit() {
+    KPADShutdown();
+
+    WHBLogUdpDeinit();
+}
+
+void ThrowMissingOTR(const char* otrPath) {
+    // TODO handle this better in the future
+    OSFatal("Main OTR file not found!");
+}
+
+void Update() {
+    bool rescan = false;
+
+    VPADRead(VPAD_CHAN_0, &vpadStatus, 1, &vpadError);
+    if (vpadError == VPAD_READ_SUCCESS) {
+        if (!hasVpad) {
+            rescan = true;
+        }
+
+        hasVpad = true;
+    } else if (vpadError != VPAD_READ_NO_SAMPLES) {
+        if (hasVpad) {
+            rescan = true;
+        }
+
+        hasVpad = false;
+    }
+
+    for (int i = 0; i < 4; i++) {
+        KPADReadEx((KPADChan) i, &kpadStatus[i], 1, &kpadError[i]);
+        if (kpadError[i] == KPAD_ERROR_OK && kpadStatus[i].extensionType != 255) {
+            if (!hasKpad[i]) {
+                rescan = true;
+            }
+
+            hasKpad[i] = true;
+        } else if (kpadError[i] != KPAD_ERROR_NO_SAMPLES) {
+            if (hasKpad[i]) {
+                rescan = true;
+            }
+
+            hasKpad[i] = false;
+        }
+    }
+
+    // rescan devices if connection state changed
+    if (rescan) {
+		Ship::GlobalCtx2::GetInstance()->GetWindow()->GetControlDeck()->ScanPhysicalDevices();
+    }
+}
+
+VPADStatus *GetVPADStatus(VPADReadError *error) {
+    *error = vpadError;
+    return hasVpad ? &vpadStatus : nullptr;
+}
+
+KPADStatus *GetKPADStatus(WPADChan chan, KPADError *error) {
+    *error = kpadError[chan];
+    return hasKpad[chan] ? &kpadStatus[chan] : nullptr;
+}
+
+};
+};
+
+#endif

libultraship/libultraship/WiiUImpl.h

@@ -0,0 +1,22 @@
+#pragma once
+
+#include <vpad/input.h>
+#include <padscore/kpad.h>
+
+namespace Ship {
+namespace WiiU {
+
+void Init();
+
+void Exit();
+
+void ThrowMissingOTR(const char* otrPath);
+
+void Update();
+
+VPADStatus *GetVPADStatus(VPADReadError *error);
+
+KPADStatus *GetKPADStatus(WPADChan chan, KPADError *error);
+
+};
+};

libultraship/libultraship/Window.cpp

@@ -33,6 +33,7 @@ extern "C" {
     int32_t osContInit(OSMesgQueue* mq, uint8_t* controllerBits, OSContStatus* status) {
         *controllerBits = 0;
 
+#ifndef __WIIU__
         if (SDL_Init(SDL_INIT_GAMECONTROLLER) != 0) {
             SPDLOG_ERROR("Failed to initialize SDL game controllers ({})", SDL_GetError());
             exit(EXIT_FAILURE);
@@ -47,6 +48,7 @@ extern "C" {
             SPDLOG_ERROR("Failed add SDL game controller mappings from \"{}\" ({})", controllerDb, SDL_GetError());
         }
     #endif
+#endif
 
         Ship::GlobalCtx2::GetInstance()->GetWindow()->GetControlDeck()->Init(controllerBits);
 

libultraship/libultraship/WindowShim.cpp

@@ -6,6 +6,8 @@
 #include "Lib/Fast3D/gfx_opengl.h"
 #include "Lib/Fast3D/gfx_direct3d11.h"
 #include "Lib/Fast3D/gfx_direct3d12.h"
+#include "Lib/Fast3D/gfx_wiiu.h"
+#include "Lib/Fast3D/gfx_gx2.h"
 #include "Lib/Fast3D/gfx_window_manager_api.h"
 
 #include <string>
@@ -38,6 +40,10 @@ void SetWindowManager(struct GfxWindowManagerAPI** WmApi, struct GfxRenderingAPI
     *RenderingApi = &gfx_direct3d11_api;
     *WmApi = &gfx_dxgi_api;
 #endif
+#ifdef __WIIU__
+    *RenderingApi = &gfx_gx2_api;
+    *WmApi = &gfx_wiiu;
+#endif
 
     // Config can override
 #ifdef ENABLE_DX11

scripts/wiiu/build.sh

@@ -0,0 +1,4 @@
+#!/bin/bash
+
+cmake -H. -Bbuild-wiiu -GNinja -DCMAKE_TOOLCHAIN_FILE=/opt/devkitpro/cmake/WiiU.cmake -DCMAKE_BUILD_TYPE:STRING=Release
+cmake --build build-wiiu --target soh --config Release

soh/CMakeLists.txt

@@ -1696,9 +1696,20 @@ if (CMAKE_SYSTEM_NAME STREQUAL "Windows")
 			STORMLIB_NO_AUTO_LINK
 		)
 	endif()
-endif()
+elseif (CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
-
+	target_compile_definitions(${PROJECT_NAME} PRIVATE
-if ("${CMAKE_CXX_COMPILER_ID}" MATCHES "GNU|Clang|AppleClang")
+		"$<$<CONFIG:Debug>:"
+				"_DEBUG"
+		">"
+		"$<$<CONFIG:Release>:"
+			"NDEBUG"
+		">"
+        "SPDLOG_ACTIVE_LEVEL=3;"
+		"SPDLOG_NO_THREAD_ID;"
+        "SPDLOG_NO_TLS;"
+		"STBI_NO_THREAD_LOCALS;"
+	)
+elseif ("${CMAKE_CXX_COMPILER_ID}" MATCHES "GNU|Clang|AppleClang")
 	target_compile_definitions(${PROJECT_NAME} PRIVATE
 		"$<$<CONFIG:Debug>:"
 				"_DEBUG"
@@ -1825,6 +1836,26 @@ if (CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang|AppleClang")
         target_link_options(${PROJECT_NAME} PRIVATE
             -pthread
         )
+    elseif (CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+        target_compile_options(${PROJECT_NAME} PRIVATE
+            -O2
+
+            # disable some warnings to not clutter output
+            -Wno-multichar
+            -Wno-return-type
+            -Wno-narrowing
+            -Wno-switch-outside-range
+            $<$<COMPILE_LANGUAGE:C>:
+                -Wno-incompatible-pointer-types
+                -Wno-discarded-array-qualifiers
+                -Wno-discarded-qualifiers
+                -Wno-int-conversion
+                -Wno-implicit-function-declaration
+                -Wno-builtin-declaration-mismatch
+                -Wno-switch-unreachable
+                -Wno-stringop-overflow
+            >
+        )
     else()
         if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64")
 		set(CPU_OPTION -msse2 -mfpmath=sse)
@@ -1908,6 +1939,18 @@ elseif(CMAKE_SYSTEM_NAME STREQUAL "NintendoSwitch")
         -lglad
         Threads::Threads
     )
+elseif(CMAKE_SYSTEM_NAME STREQUAL "CafeOS")
+    find_package(SDL2 REQUIRED)
+    set(ADDITIONAL_LIBRARY_DEPENDENCIES
+        "libultraship;"
+        "ZAPDUtils;"
+        SDL2::SDL2-static
+
+        "$<$<CONFIG:Debug>:-Wl,--wrap=abort>"
+    )
+    target_include_directories(${PROJECT_NAME} PRIVATE
+        ${DEVKITPRO}/portlibs/wiiu/include/
+    )
 else()
     find_package(SDL2)
     set(THREADS_PREFER_PTHREAD_FLAG ON)
@@ -1961,7 +2004,7 @@ if (CMAKE_SYSTEM_NAME STREQUAL "Windows")
     endif()
 endif()
 
-if(NOT CMAKE_SYSTEM_NAME STREQUAL "NintendoSwitch")
+if(NOT CMAKE_SYSTEM_NAME MATCHES "NintendoSwitch|CafeOS")
 INSTALL(TARGETS soh DESTINATION . COMPONENT ship)
 endif()
 
@@ -1970,7 +2013,7 @@ execute_process(COMMAND ${CURL} -sSfL https://raw.githubusercontent.com/gabomdq/
 
 if("${CMAKE_SYSTEM_NAME}" STREQUAL "Darwin")
 INSTALL(FILES ${CMAKE_BINARY_DIR}/gamecontrollerdb.txt DESTINATION ../MacOS COMPONENT ship)
-elseif(NOT "${CMAKE_SYSTEM_NAME}" STREQUAL "NintendoSwitch")
+elseif(NOT "${CMAKE_SYSTEM_NAME}" MATCHES "NintendoSwitch|CafeOS")
 INSTALL(FILES ${CMAKE_BINARY_DIR}/gamecontrollerdb.txt DESTINATION . COMPONENT ship)
 endif()
 
@@ -1982,9 +2025,9 @@ install(CODE "
     ")
 endif()
 
-if(CMAKE_SYSTEM_NAME STREQUAL "NintendoSwitch")
+if(CMAKE_SYSTEM_NAME MATCHES "NintendoSwitch|CafeOS")
     if (NOT TARGET pathconf)
-        add_library(pathconf OBJECT switch/pathconf.c)
+        add_library(pathconf OBJECT platform/pathconf.c)
     endif()
     target_link_libraries(${PROJECT_NAME} PRIVATE "${ADDITIONAL_LIBRARY_DEPENDENCIES}" $<TARGET_OBJECTS:pathconf> )
 else()
@@ -2006,4 +2049,10 @@ nx_create_nro(soh
 
 INSTALL(FILES ${CMAKE_CURRENT_BINARY_DIR}/soh.nro DESTINATION . COMPONENT ship)
 
+elseif(CMAKE_SYSTEM_NAME MATCHES "CafeOS")
+
+wut_create_rpx(${PROJECT_NAME})
+
+INSTALL(FILES ${CMAKE_CURRENT_BINARY_DIR}/soh.rpx DESTINATION . COMPONENT ship)
+
 endif()

soh/include/functions.h

@@ -60,7 +60,9 @@ void Locale_ResetRegion(void);
 u32 func_80001F48(void);
 u32 func_80001F8C(void);
 u32 Locale_IsRegionNative(void);
-#if !defined(__APPLE__) && !defined(__SWITCH__)
+#ifdef __WIIU__
+void _assert(const char* exp, const char* file, s32 line);
+#elif !defined(__APPLE__) && !defined(__SWITCH__)
 void __assert(const char* exp, const char* file, s32 line);
 #endif
 #if defined(__APPLE__) && defined(NDEBUG)

soh/include/macros.h

@@ -233,7 +233,11 @@ extern GraphicsContext* __gfxCtx;
 
 #define VTX_T(x,y,z,s,t,cr,cg,cb,a) { { x, y, z }, 0, { s, t }, { cr, cg, cb, a } }
 
+#ifdef __WIIU__
+#define ASSERT(expression) (void)((!!(expression)) || (_assert(#expression, __FILE__, (unsigned)(__LINE__)), 0))
+#else
 #define ASSERT(expression) (void)((!!(expression)) || (__assert(#expression, __FILE__, (unsigned)(__LINE__)), 0))
+#endif
 
 #define gDPSetTileCustom(pkt, fmt, siz, width, height, pal, cms, cmt, masks, maskt, shifts, shiftt)                    \
     do {                                                                                                               \

soh/soh/OTRGlobals.cpp

@@ -49,6 +49,8 @@
 
 #ifdef __SWITCH__
 #include "SwitchImpl.h"
+#elif defined(__WIIU__)
+#include "WiiUImpl.h"
 #endif
 
 #include <Audio.h>
@@ -171,6 +173,8 @@ extern "C" void OTRExtScanner() {
 extern "C" void InitOTR() {
 #ifdef __SWITCH__
     Ship::Switch::Init(Ship::PreInitPhase);
+#elif defined(__WIIU__)
+    Ship::WiiU::Init();
 #endif
     OTRGlobals::Instance = new OTRGlobals();
     SaveManager::Instance = new SaveManager();
@@ -235,6 +239,7 @@ extern "C" void Graph_ProcessFrame(void (*run_one_game_iter)(void)) {
 }
 
 extern "C" void Graph_StartFrame() {
+#ifndef __WIIU__
     // Why -1?
     int32_t dwScancode = OTRGlobals::Instance->context->GetWindow()->lastScancode;
     OTRGlobals::Instance->context->GetWindow()->lastScancode = -1;
@@ -292,6 +297,7 @@ extern "C" void Graph_StartFrame() {
             break;
         }
     }
+#endif
     OTRGlobals::Instance->context->GetWindow()->StartFrame();
 }
 

soh/soh/SaveManager.cpp

@@ -152,6 +152,12 @@ void SaveManager::Init() {
     // If the global save file exist, load it. Otherwise, create it.
     if (std::filesystem::exists(sGlobalPath)) {
         std::ifstream input(sGlobalPath);
+
+#ifdef __WIIU__
+        alignas(0x40) char buffer[8192];
+        input.rdbuf()->pubsetbuf(buffer, sizeof(buffer));
+#endif
+
         nlohmann::json globalBlock;
         input >> globalBlock;
 
@@ -474,6 +480,12 @@ void SaveManager::SaveFile(int fileNum) {
     }
 
     std::ofstream output(GetFileName(fileNum));
+
+#ifdef __WIIU__
+    alignas(0x40) char buffer[8192];
+    output.rdbuf()->pubsetbuf(buffer, sizeof(buffer));
+#endif
+
     output << std::setw(4) << baseBlock << std::endl;
 
     InitMeta(fileNum);
@@ -486,6 +498,12 @@ void SaveManager::SaveGlobal() {
     globalBlock["zTargetSetting"] = gSaveContext.zTargetSetting;
     globalBlock["language"] = gSaveContext.language;
     std::ofstream output("Save/global.sav");
+
+#ifdef __WIIU__
+    alignas(0x40) char buffer[8192];
+    output.rdbuf()->pubsetbuf(buffer, sizeof(buffer));
+#endif
+
     output << std::setw(4) << globalBlock << std::endl;
 }
 
@@ -494,6 +512,12 @@ void SaveManager::LoadFile(int fileNum) {
     InitFile(false);
 
     std::ifstream input(GetFileName(fileNum));
+
+#ifdef __WIIU__
+    alignas(0x40) char buffer[8192];
+    input.rdbuf()->pubsetbuf(buffer, sizeof(buffer));
+#endif
+
     nlohmann::json saveBlock;
     input >> saveBlock;
     if (!saveBlock.contains("version")) {
@@ -1125,10 +1149,41 @@ void SaveManager::LoadStruct(const std::string& name, LoadStructFunc func) {
     }
 }
 
+#ifdef __WIIU__
+// std::filesystem::copy_file doesn't work properly with the Wii U's toolchain atm
+int copy_file(const char* src, const char* dst)
+{
+    alignas(0x40) uint8_t buf[4096];
+    FILE* r = fopen(src, "r");
+    if (!r) {
+        return -1;
+    }
+    FILE* w = fopen(dst, "w");
+    if (!w) {
+        return -2;
+    }
+
+    size_t res;
+    while ((res = fread(buf, 1, sizeof(buf), r)) > 0) {
+        if (fwrite(buf, 1, res, w) != res) {
+            break;
+        }
+    }
+
+    fclose(r);
+    fclose(w);
+    return res >= 0 ? 0 : res;
+}
+#endif
+
 void SaveManager::CopyZeldaFile(int from, int to) {
     assert(std::filesystem::exists(GetFileName(from)));
     DeleteZeldaFile(to);
+#ifdef __WIIU__
+    assert(copy_file(GetFileName(from).c_str(), GetFileName(to).c_str()) == 0);
+#else
     std::filesystem::copy_file(GetFileName(from), GetFileName(to));
+#endif
     fileMetaInfo[to].valid = true;
     fileMetaInfo[to].deaths = fileMetaInfo[from].deaths;
     for (int i = 0; i < ARRAY_COUNT(fileMetaInfo[to].playerName); i++) {
@@ -1457,6 +1512,12 @@ void SaveManager::ConvertFromUnversioned() {
 #define SLOT_OFFSET(index) (SRAM_HEADER_SIZE + 0x10 + (index * SLOT_SIZE))
 
     std::ifstream input("oot_save.sav", std::ios::binary);
+
+#ifdef __WIIU__
+    alignas(0x40) char buffer[8192];
+    input.rdbuf()->pubsetbuf(buffer, sizeof(buffer));
+#endif
+
     std::vector<char> data(std::istreambuf_iterator<char>(input), {});
     input.close();
 

soh/src/boot/assert.c

@@ -1,7 +1,11 @@
 #include "global.h"
 
 #ifndef __SWITCH__
+#ifdef __WIIU__
+void _assert(const char* exp, const char* file, s32 line) {
+#else
 void __assert(const char* exp, const char* file, s32 line) {
+#endif
     char msg[256];
 
     osSyncPrintf("Assertion failed: %s, file %s, line %d, thread %d\n", exp, file, line, osGetThreadId(NULL));

soh/src/code/z_bgcheck.c

@@ -397,6 +397,16 @@ s32 CollisionPoly_LineVsPoly(CollisionPoly* poly, Vec3s* vtxList, Vec3f* posA, V
         (poly->normal.x * posB->x + poly->normal.y * posB->y + poly->normal.z * posB->z) * COLPOLY_NORMAL_FRAC +
         plane.originDist;
 
+#ifdef __WIIU__
+    // on some platforms this ends up as very small numbers due to rounding issues
+    if (IS_ZERO(planeDistA)) {
+        planeDistA = 0.0f;
+    }
+    if (IS_ZERO(planeDistB)) {
+        planeDistB = 0.0f;
+    }
+#endif
+
     planeDistDelta = planeDistA - planeDistB;
     if ((planeDistA >= 0.0f && planeDistB >= 0.0f) || (planeDistA < 0.0f && planeDistB < 0.0f) ||
         (chkOneFace && planeDistA < 0.0f && planeDistB > 0.0f) || IS_ZERO(planeDistDelta)) {