Merge branch 'main' into function-pointers

CMakePresets.json

@@ -0,0 +1,50 @@
+{
+    "version": 8,
+    "configurePresets": [
+        {
+            "name": "windows-base",
+            "hidden": true,
+            "generator": "Ninja",
+            "binaryDir": "${sourceDir}/out/build/${presetName}",
+            "installDir": "${sourceDir}/out/install/${presetName}",
+            "cacheVariables": {
+                "CMAKE_C_COMPILER": "clang-cl.exe",
+                "CMAKE_CXX_COMPILER": "clang-cl.exe",
+                "CMAKE_TOOLCHAIN_FILE": {
+                    "value": "$env{VCPKG_ROOT}/scripts/buildsystems/vcpkg.cmake",
+                    "type": "FILEPATH"
+                }
+            },
+            "condition": {
+                "type": "equals",
+                "lhs": "${hostSystemName}",
+                "rhs": "Windows"
+            },
+            "toolset": "ClangCL"
+        },
+        {
+            "name": "x64-Clang-Debug",
+            "displayName": "Debug",
+            "inherits": "windows-base",
+            "architecture": {
+                "value": "x64",
+                "strategy": "external"
+            },
+            "cacheVariables": {
+                "CMAKE_BUILD_TYPE": "Debug",
+                "VCPKG_TARGET_TRIPLET": {
+                    "value": "x64-windows-static",
+                    "type": "STRING"
+                }
+            }
+        },
+        {
+            "name": "x64-Clang-Release",
+            "displayName": "Release",
+            "inherits": "x64-Clang-Debug",
+            "cacheVariables": {
+                "CMAKE_BUILD_TYPE": "Release"
+            }
+        }
+    ]
+}

CMakeSettings.json

@@ -1,40 +0,0 @@
-{
-    "configurations": [
-        {
-            "name": "x64-Clang-Debug",
-            "generator": "Ninja",
-            "configurationType": "Debug",
-            "buildRoot": "${projectDir}\\out\\build\\${name}",
-            "installRoot": "${projectDir}\\out\\install\\${name}",
-            "cmakeCommandArgs": "",
-            "buildCommandArgs": "",
-            "ctestCommandArgs": "",
-            "inheritEnvironments": [ "clang_cl_x64_x64" ],
-            "variables": [
-                {
-                    "name": "VCPKG_TARGET_TRIPLET",
-                    "value": "x64-windows-static",
-                    "type": "STRING"
-                }
-            ]
-        },
-        {
-            "name": "x64-Clang-Release",
-            "generator": "Ninja",
-            "configurationType": "RelWithDebInfo",
-            "buildRoot": "${projectDir}\\out\\build\\${name}",
-            "installRoot": "${projectDir}\\out\\install\\${name}",
-            "cmakeCommandArgs": "",
-            "buildCommandArgs": "",
-            "ctestCommandArgs": "",
-            "inheritEnvironments": [ "clang_cl_x64_x64" ],
-            "variables": [
-                {
-                    "name": "VCPKG_TARGET_TRIPLET",
-                    "value": "x64-windows-static",
-                    "type": "STRING"
-                }
-            ]
-        }
-    ]
-}

UnleashedRecomp/CMakeLists.txt

@@ -79,6 +79,7 @@ set(SWA_UI_CXX_SOURCES
 )
 
 set(SWA_CXX_SOURCES
+    "app.cpp"
     "main.cpp"
     "misc_impl.cpp"
     "stdafx.cpp"
@@ -188,6 +189,7 @@ function(compile_pixel_shader FILE_PATH)
 endfunction()
 
 compile_vertex_shader(copy_vs)
+compile_pixel_shader(gamma_correction_ps)
 compile_pixel_shader(imgui_ps)
 compile_vertex_shader(imgui_vs)
 compile_pixel_shader(movie_ps)

UnleashedRecomp/app.cpp

@@ -0,0 +1,19 @@
+#include <kernel/function.h>
+#include <ui/window.h>
+#include <app.h>
+
+double g_deltaTime;
+
+// CApplication::Update
+PPC_FUNC_IMPL(__imp__sub_822C1130);
+PPC_FUNC(sub_822C1130)
+{
+    g_deltaTime = ctx.f1.f64;
+
+    SDL_PumpEvents();
+    SDL_FlushEvents(SDL_FIRSTEVENT, SDL_LASTEVENT);
+
+    Window::Update();
+
+    __imp__sub_822C1130(ctx, base);
+}

UnleashedRecomp/app.h

@@ -0,0 +1,3 @@
+#pragma once
+
+extern double g_deltaTime;

UnleashedRecomp/gpu/shader/gamma_correction_ps.hlsl

@@ -0,0 +1,24 @@
+#include "../../../thirdparty/ShaderRecomp/ShaderRecomp/shader_common.hlsli"
+
+#ifdef __spirv__
+
+#define g_Gamma vk::RawBufferLoad<float3>(g_PushConstants.SharedConstants + 0)
+#define g_TextureDescriptorIndex vk::RawBufferLoad<uint>(g_PushConstants.SharedConstants + 12)
+
+#else
+
+cbuffer SharedConstants : register(b2, space4)
+{
+    float3 g_Gamma : packoffset(c0.x);
+    uint g_TextureDescriptorIndex : packoffset(c0.w);
+};
+
+#endif
+
+float4 main(in float4 position : SV_Position) : SV_Target
+{
+    Texture2D<float4> texture = g_Texture2DDescriptorHeap[g_TextureDescriptorIndex];
+    float4 color = texture.Load(int3(position.xy, 0));
+    color.rgb = pow(color.rgb, g_Gamma);
+    return color;
+}

UnleashedRecomp/gpu/video.cpp

@@ -15,6 +15,8 @@
 
 #include "shader/copy_vs.hlsl.dxil.h"
 #include "shader/copy_vs.hlsl.spirv.h"
+#include "shader/gamma_correction_ps.hlsl.dxil.h"
+#include "shader/gamma_correction_ps.hlsl.spirv.h"
 #include "shader/imgui_ps.hlsl.dxil.h"
 #include "shader/imgui_ps.hlsl.spirv.h"
 #include "shader/imgui_vs.hlsl.dxil.h"
@@ -171,11 +173,20 @@ static std::unique_ptr<RenderSwapChain> g_swapChain;
 static bool g_swapChainValid;
 static bool g_needsResize;
 
+static constexpr RenderFormat BACKBUFFER_FORMAT = RenderFormat::B8G8R8A8_UNORM;
+
 static std::unique_ptr<RenderCommandSemaphore> g_acquireSemaphores[NUM_FRAMES];
 static std::unique_ptr<RenderCommandSemaphore> g_renderSemaphores[NUM_FRAMES];
 static uint32_t g_backBufferIndex;
 static GuestSurface* g_backBuffer;
 
+static std::unique_ptr<RenderTexture> g_intermediaryBackBufferTexture;
+static uint32_t g_intermediaryBackBufferTextureWidth;
+static uint32_t g_intermediaryBackBufferTextureHeight;
+static uint32_t g_intermediaryBackBufferTextureDescriptorIndex;
+
+static std::unique_ptr<RenderPipeline> g_gammaCorrectionPipeline;
+
 struct std::unique_ptr<RenderDescriptorSet> g_textureDescriptorSet;
 struct std::unique_ptr<RenderDescriptorSet> g_samplerDescriptorSet;
 
@@ -1080,7 +1091,7 @@ static void CreateImGuiBackend()
     pipelineDesc.pipelineLayout = g_imPipelineLayout.get();
     pipelineDesc.vertexShader = vertexShader.get();
     pipelineDesc.pixelShader = pixelShader.get();
-    pipelineDesc.renderTargetFormat[0] = RenderFormat::B8G8R8A8_UNORM;
+    pipelineDesc.renderTargetFormat[0] = BACKBUFFER_FORMAT;
     pipelineDesc.renderTargetBlend[0] = RenderBlendDesc::AlphaBlend();
     pipelineDesc.renderTargetCount = 1;
     pipelineDesc.inputElements = inputElements;
@@ -1121,7 +1132,7 @@ static void CreateHostDevice()
     g_copyCommandList = g_device->createCommandList(RenderCommandListType::COPY);
     g_copyCommandFence = g_device->createCommandFence();
 
-    g_swapChain = g_queue->createSwapChain(Window::s_handle, Config::TripleBuffering ? 3 : 2, RenderFormat::B8G8R8A8_UNORM);
+    g_swapChain = g_queue->createSwapChain(Window::s_handle, Config::TripleBuffering ? 3 : 2, BACKBUFFER_FORMAT);
     g_swapChain->setVsyncEnabled(Config::VSync);
     g_swapChainValid = !g_swapChain->needsResize();
 
@@ -1253,6 +1264,17 @@ static void CreateHostDevice()
     }
 
     CreateImGuiBackend();
+
+    auto gammaCorrectionShader = CREATE_SHADER(gamma_correction_ps);
+
+    RenderGraphicsPipelineDesc desc;
+    desc.pipelineLayout = g_pipelineLayout.get();
+    desc.vertexShader = copyShader.get();
+    desc.pixelShader = gammaCorrectionShader.get();
+    desc.renderTargetFormat[0] = BACKBUFFER_FORMAT;
+    desc.renderTargetBlend[0] = RenderBlendDesc::Copy();
+    desc.renderTargetCount = 1;
+    g_gammaCorrectionPipeline = g_device->createGraphicsPipeline(desc);
 }
 
 static void WaitForGPU()
@@ -1290,7 +1312,7 @@ static void BeginCommandList()
     g_depthStencil = nullptr;
     g_framebuffer = nullptr;
 
-    g_pipelineState.renderTargetFormat = g_backBuffer->format;
+    g_pipelineState.renderTargetFormat = BACKBUFFER_FORMAT;
     g_pipelineState.depthStencilFormat = RenderFormat::UNKNOWN;
 
     g_swapChainValid &= !g_swapChain->needsResize();
@@ -1307,9 +1329,40 @@ static void BeginCommandList()
         g_swapChainValid = g_swapChain->acquireTexture(g_acquireSemaphores[g_frame].get(), &g_backBufferIndex);
 
     if (g_swapChainValid)
-        g_backBuffer->texture = g_swapChain->getTexture(g_backBufferIndex);
+    {
+        bool applyingGammaCorrection = Config::Xbox360ColourCorrection || abs(Config::Brightness - 0.5f) > 0.001f;
+
+        if (applyingGammaCorrection)
+        {
+            uint32_t width = g_swapChain->getWidth();
+            uint32_t height = g_swapChain->getHeight();
+
+            if (g_intermediaryBackBufferTextureWidth != width ||
+                g_intermediaryBackBufferTextureHeight != height)
+            {
+                if (g_intermediaryBackBufferTextureDescriptorIndex == NULL)
+                    g_intermediaryBackBufferTextureDescriptorIndex = g_textureDescriptorAllocator.allocate();
+
+                WaitForGPU(); // Fine to wait for GPU, this'll only happen during resize.
+
+                g_intermediaryBackBufferTexture = g_device->createTexture(RenderTextureDesc::Texture2D(width, height, 1, BACKBUFFER_FORMAT, RenderTextureFlag::RENDER_TARGET));
+                g_textureDescriptorSet->setTexture(g_intermediaryBackBufferTextureDescriptorIndex, g_intermediaryBackBufferTexture.get(), RenderTextureLayout::SHADER_READ);
+
+                g_intermediaryBackBufferTextureWidth = width;
+                g_intermediaryBackBufferTextureHeight = height;
+            }
+
+            g_backBuffer->texture = g_intermediaryBackBufferTexture.get();
+        }
+        else
+        {
+            g_backBuffer->texture = g_swapChain->getTexture(g_backBufferIndex);
+        }
+    }
     else
+    {
         g_backBuffer->texture = g_backBuffer->textureHolder.get();
+    }
 
     g_backBuffer->layout = RenderTextureLayout::UNKNOWN;
 
@@ -1339,8 +1392,8 @@ static uint32_t CreateDevice(uint32_t a1, uint32_t a2, uint32_t a3, uint32_t a4,
     g_backBuffer = g_userHeap.AllocPhysical<GuestSurface>(ResourceType::RenderTarget);
     g_backBuffer->width = 1280;
     g_backBuffer->height = 720;
-    g_backBuffer->format = RenderFormat::B8G8R8A8_UNORM;
+    g_backBuffer->format = BACKBUFFER_FORMAT;
-    g_backBuffer->textureHolder = g_device->createTexture(RenderTextureDesc::Texture2D(16, 16, 1, g_backBuffer->format, RenderTextureFlag::RENDER_TARGET));
+    g_backBuffer->textureHolder = g_device->createTexture(RenderTextureDesc::Texture2D(1, 1, 1, BACKBUFFER_FORMAT, RenderTextureFlag::RENDER_TARGET));
 
     BeginCommandList();
 
@@ -1639,12 +1692,84 @@ static void Present()
     g_renderQueue.enqueue(cmd);
 }
 
+static void SetRootDescriptor(const UploadAllocation& allocation, size_t index)
+{
+    auto& commandList = g_commandLists[g_frame];
+
+    if (g_vulkan)
+        commandList->setGraphicsPushConstants(0, &allocation.deviceAddress, 8 * index, 8);
+    else
+        commandList->setGraphicsRootDescriptor(allocation.buffer->at(allocation.offset), index);
+}
+
 static void ProcPresent(const RenderCommand& cmd)
 {
     if (g_swapChainValid)
-        AddBarrier(g_backBuffer, RenderTextureLayout::PRESENT);
+    {
+        auto swapChainTexture = g_swapChain->getTexture(g_backBufferIndex);
+        if (g_backBuffer->texture == g_intermediaryBackBufferTexture.get())
+        {
+            struct
+            {
+                float gammaR;
+                float gammaG;
+                float gammaB;
+                uint32_t textureDescriptorIndex;
+            } constants;
 
-    FlushBarriers();
+            if (Config::Xbox360ColourCorrection)
+            {
+                constants.gammaR = 1.2f;
+                constants.gammaG = 1.17f;
+                constants.gammaB = 0.98f;
+            }
+            else
+            {
+                constants.gammaR = 1.0f;
+                constants.gammaG = 1.0f;
+                constants.gammaB = 1.0f;
+            }
+
+            float offset = (Config::Brightness - 0.5f) * 1.2f;
+
+            constants.gammaR = 1.0f / std::clamp(constants.gammaR + offset, 0.1f, 4.0f);
+            constants.gammaG = 1.0f / std::clamp(constants.gammaG + offset, 0.1f, 4.0f);
+            constants.gammaB = 1.0f / std::clamp(constants.gammaB + offset, 0.1f, 4.0f);
+            constants.textureDescriptorIndex = g_intermediaryBackBufferTextureDescriptorIndex;
+
+            auto& framebuffer = g_backBuffer->framebuffers[swapChainTexture];
+            if (!framebuffer)
+            {
+                RenderFramebufferDesc desc;
+                desc.colorAttachments = const_cast<const RenderTexture**>(&swapChainTexture);
+                desc.colorAttachmentsCount = 1;
+                framebuffer = g_device->createFramebuffer(desc);
+            }
+
+            RenderTextureBarrier srcBarriers[] =
+            {
+                RenderTextureBarrier(g_intermediaryBackBufferTexture.get(), RenderTextureLayout::SHADER_READ),
+                RenderTextureBarrier(swapChainTexture, RenderTextureLayout::COLOR_WRITE)
+            };
+
+            auto& commandList = g_commandLists[g_frame];
+            commandList->barriers(RenderBarrierStage::GRAPHICS, srcBarriers, std::size(srcBarriers));
+            commandList->setGraphicsPipelineLayout(g_pipelineLayout.get());
+            commandList->setPipeline(g_gammaCorrectionPipeline.get());
+            commandList->setGraphicsDescriptorSet(g_textureDescriptorSet.get(), 0);
+            SetRootDescriptor(g_uploadAllocators[g_frame].allocate<false>(&constants, sizeof(constants), 0x100), 2);
+            commandList->setFramebuffer(framebuffer.get());
+            commandList->setViewports(RenderViewport(0.0f, 0.0f, g_intermediaryBackBufferTextureWidth, g_intermediaryBackBufferTextureHeight));
+            commandList->setScissors(RenderRect(0, 0, g_intermediaryBackBufferTextureWidth, g_intermediaryBackBufferTextureHeight));
+            commandList->drawInstanced(6, 1, 0, 0);
+            commandList->barriers(RenderBarrierStage::GRAPHICS, RenderTextureBarrier(swapChainTexture, RenderTextureLayout::PRESENT));
+        }
+        else
+        {
+            AddBarrier(g_backBuffer, RenderTextureLayout::PRESENT);
+            FlushBarriers();
+        }
+    }
 
     auto& commandList = g_commandLists[g_frame];
     commandList->end();
@@ -2537,16 +2662,6 @@ static void ProcSetSamplerState(const RenderCommand& cmd)
     }
 }
 
-static void SetRootDescriptor(const UploadAllocation& allocation, size_t index)
-{
-    auto& commandList = g_commandLists[g_frame];
-
-    if (g_vulkan)
-        commandList->setGraphicsPushConstants(0, &allocation.deviceAddress, 8 * index, 8);
-    else
-        commandList->setGraphicsRootDescriptor(allocation.buffer->at(allocation.offset), index);
-}
-
 static void ProcSetVertexShaderConstants(const RenderCommand& cmd)
 {
     SetRootDescriptor(cmd.setVertexShaderConstants.allocation, 0);

UnleashedRecomp/patches/fps_patches.cpp

@@ -2,6 +2,7 @@
 #include <api/SWA.h>
 #include <ui/window.h>
 #include <cfg/config.h>
+#include <app.h>
 
 float m_lastLoadingFrameDelta = 0.0f;
 std::chrono::steady_clock::time_point m_lastLoadingFrameTime;
@@ -36,14 +37,16 @@ static double ComputeLerpFactor(double t, double deltaTime)
     return 1.0 - pow(1.0 - t, (30.0 + bias) / (fps + bias));
 }
 
-void CameraLerpFixMidAsmHook(PPCRegister& t, PPCRegister& deltaTime)
+// It's important to use global delta time here instead of function provided
+// delta time, as it might be time scaled and not match with 30 FPS behavior.
+void CameraLerpFixMidAsmHook(PPCRegister& t)
 {
-    t.f64 = ComputeLerpFactor(t.f64, deltaTime.f64);
+    t.f64 = ComputeLerpFactor(t.f64, g_deltaTime);
 }
 
-void CameraTargetSideOffsetLerpFixMidAsmHook(PPCVRegister& v13, PPCVRegister& v62, PPCRegister& deltaTime)
+void CameraTargetSideOffsetLerpFixMidAsmHook(PPCVRegister& v13, PPCVRegister& v62)
 {
-    float factor = float(ComputeLerpFactor(double(v13.f32[0] * v62.f32[0]), deltaTime.f64));
+    float factor = float(ComputeLerpFactor(double(v13.f32[0] * v62.f32[0]), g_deltaTime));
 
     for (size_t i = 0; i < 4; i++)
     {

UnleashedRecomp/patches/video_patches.cpp

@@ -41,6 +41,9 @@ void CameraBoostAspectRatioMidAsmHook(PPCRegister& r31, PPCRegister& f0)
 
 void CSDAspectRatioMidAsmHook(PPCRegister& f1, PPCRegister& f2)
 {
+    if (Config::UIScaleMode == EUIScaleMode::Stretch)
+        return;
+
     auto newAspectRatio = (float)Window::s_width / (float)Window::s_height;
 
     if (newAspectRatio > m_baseAspectRatio)

UnleashedRecomp/ui/window.cpp

@@ -1,7 +1,6 @@
 #include "window.h"
 #include "sdl_listener.h"
 #include <cfg/config.h>
-#include <kernel/function.h>
 #include <SDL_syswm.h>
 
 bool m_isFullscreenKeyReleased = true;
@@ -175,15 +174,3 @@ void Window::Update()
         Config::WindowHeight = Window::s_height;
     }
 }
-
-// CApplication::Update
-PPC_FUNC_IMPL(__imp__sub_822C1130);
-PPC_FUNC(sub_822C1130)
-{
-    SDL_PumpEvents();
-    SDL_FlushEvents(SDL_FIRSTEVENT, SDL_LASTEVENT);
-
-    Window::Update();
-
-    __imp__sub_822C1130(ctx, base);
-}

UnleashedRecomp/ui/window.h

@@ -63,7 +63,7 @@ public:
         if (!title)
         {
             title = Config::Language == ELanguage::Japanese
-                ? "Sonic World Adventure"
+                ? "SONIC WORLD ADVENTURE"
                 : "SONIC UNLEASHED";
         }
 

UnleashedRecompLib/config/SWA.toml

@@ -213,52 +213,52 @@ jump_address = 0x8247DD3C
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247DD48 # Slope
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247DDFC # Dash path binormal
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247E280 # Target front offset
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247E300 # Target
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247E3C8 # Target
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247E4A0 # Target up positive
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247E4E8 # Target up negative
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247E9FC # Position
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247EA84 # Position up negative
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247EA50 # Position up positive
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraDeltaTimeFixMidAsmHook"
@@ -269,7 +269,7 @@ jump_address = 0x8247ED34
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247ED40 # Distance offset 1
-registers = ["f0", "f21"]
+registers = ["f0"]
 
 [[midasm_hook]]
 name = "CameraDeltaTimeFixMidAsmHook"
@@ -280,12 +280,12 @@ jump_address = 0x8247ED58
 [[midasm_hook]]
 name = "CameraLerpFixMidAsmHook"
 address = 0x8247ED64 # Distance offset 2
-registers = ["f13", "f21"]
+registers = ["f13"]
 
 [[midasm_hook]]
 name = "CameraTargetSideOffsetLerpFixMidAsmHook"
 address = 0x8247F12C # Target side offset
-registers = ["v13", "v62", "f21"]
+registers = ["v13", "v62"]
 
 # 2D camera HFR fixes
 [[midasm_hook]]

thirdparty/PowerRecomp

@@ -1 +1 @@
-Subproject commit a7c970d32497b7f1624a2871fdd67435d724b1ef
+Subproject commit d8676283fd1e8990b415cc0e4810c6db895b9fba