Merge branch 'main' into function-pointers

UnleashedRecomp/gpu/video.cpp

@@ -2023,12 +2023,37 @@ static void ProcStretchRect(const RenderCommand& cmd)
     AddBarrier(args.texture, RenderTextureLayout::SHADER_READ);
 }
 
+static void SetDefaultViewport(GuestDevice* device, GuestSurface* surface)
+{
+    if (surface != nullptr)
+    {
+        RenderCommand cmd;
+        cmd.type = RenderCommandType::SetViewport;
+        cmd.setViewport.x = 0.0f;
+        cmd.setViewport.y = 0.0f;
+        cmd.setViewport.width = float(surface->width);
+        cmd.setViewport.height = float(surface->height);
+        cmd.setViewport.minDepth = 0.0f;
+        cmd.setViewport.maxDepth = 1.0f;
+        g_renderQueue.enqueue(cmd);
+
+        device->viewport.x = 0.0f;
+        device->viewport.y = 0.0f;
+        device->viewport.width = float(surface->width);
+        device->viewport.height = float(surface->height);
+        device->viewport.minZ = 0.0f;
+        device->viewport.maxZ = 1.0f;
+    }
+}
+
 static void SetRenderTarget(GuestDevice* device, uint32_t index, GuestSurface* renderTarget) 
 {
     RenderCommand cmd;
     cmd.type = RenderCommandType::SetRenderTarget;
     cmd.setRenderTarget.renderTarget = renderTarget;
     g_renderQueue.enqueue(cmd);
+
+    SetDefaultViewport(device, renderTarget);
 }
 
 static void ProcSetRenderTarget(const RenderCommand& cmd)
@@ -2049,6 +2074,8 @@ static void SetDepthStencilSurface(GuestDevice* device, GuestSurface* depthStenc
     cmd.type = RenderCommandType::SetDepthStencilSurface;
     cmd.setDepthStencilSurface.depthStencil = depthStencil;
     g_renderQueue.enqueue(cmd);
+
+    SetDefaultViewport(device, depthStencil);
 }
 
 static void ProcSetDepthStencilSurface(const RenderCommand& cmd)
@@ -2996,6 +3023,8 @@ static void SetVertexDeclaration(GuestDevice* device, GuestVertexDeclaration* ve
     cmd.type = RenderCommandType::SetVertexDeclaration;
     cmd.setVertexDeclaration.vertexDeclaration = vertexDeclaration;
     g_renderQueue.enqueue(cmd);
+
+    device->vertexDeclaration = g_memory.MapVirtual(vertexDeclaration);
 }
 
 static void ProcSetVertexDeclaration(const RenderCommand& cmd)
@@ -3693,9 +3722,13 @@ static void SetResolution(be<uint32_t>* device)
     device[47] = height == 0 ? 720 : height;
 }
 
-static uint32_t StubFunction()
+// The game does some weird stuff to render targets if they are above 
+// 1024x1024 resolution, setting this bool at address 20 seems to avoid all that.
+PPC_FUNC(sub_82E9F048)
 {
-    return 0;
+    PPC_STORE_U8(ctx.r4.u32 + 20, 1);
+    PPC_STORE_U32(ctx.r4.u32 + 44, PPC_LOAD_U32(ctx.r4.u32 + 8)); // Width
+    PPC_STORE_U32(ctx.r4.u32 + 48, PPC_LOAD_U32(ctx.r4.u32 + 12)); // Height
 }
 
 static GuestShader* g_movieVertexShader;
@@ -3890,7 +3923,6 @@ GUEST_FUNCTION_HOOK(sub_82C00910, D3DXFillVolumeTexture);
 GUEST_FUNCTION_HOOK(sub_82E43FC8, MakePictureData);
 
 GUEST_FUNCTION_HOOK(sub_82E9EE38, SetResolution);
-GUEST_FUNCTION_HOOK(sub_82BE77B0, StubFunction);
 
 GUEST_FUNCTION_HOOK(sub_82AE2BF8, ScreenShaderInit);
 

UnleashedRecomp/kernel/imports.cpp

@@ -664,7 +664,7 @@ void RtlRaiseException_x()
 void KfReleaseSpinLock(uint32_t* spinLock)
 {
     //printf("!!! STUB !!! KfReleaseSpinLock\n");
-    *spinLock = 0;
+    InterlockedExchange((volatile long*)spinLock, 0);
 }
 
 void KfAcquireSpinLock(uint32_t* spinLock)
@@ -711,7 +711,7 @@ void VdGetSystemCommandBuffer()
 void KeReleaseSpinLockFromRaisedIrql(uint32_t* spinLock)
 {
     //printf("!!! STUB !!! KeReleaseSpinLockFromRaisedIrql\n");
-    *spinLock = 0;
+    InterlockedExchange((volatile long*)spinLock, 0);
 }
 
 void KeAcquireSpinLockAtRaisedIrql(uint32_t* spinLock)

UnleashedRecomp/patches/fps_patches.cpp

@@ -3,6 +3,9 @@
 #include "ui/window.h"
 #include "config.h"
 
+float m_lastLoadingFrameDelta = 0.0f;
+std::chrono::steady_clock::time_point m_lastLoadingFrameTime;
+
 void HighFrameRateDeltaTimeFixMidAsmHook(PPCRegister& f1)
 {
     // Having 60 FPS threshold ensures we still retain
@@ -58,3 +61,15 @@ void Camera2DSlopeLerpFixMidAsmHook(PPCRegister& t, PPCRegister& deltaTime)
 {
     t.f64 = ComputeLerpFactor(t.f64, deltaTime.f64 / 60.0);
 }
+
+void LoadingScreenSpeedFixMidAsmHook(PPCRegister& r4)
+{
+    auto now = std::chrono::high_resolution_clock::now();
+
+    m_lastLoadingFrameDelta = std::min(std::chrono::duration<float>(now - m_lastLoadingFrameTime).count(), 1.0f / 15.0f);
+    m_lastLoadingFrameTime = now;
+
+    auto pDeltaTime = (be<float>*)g_memory.Translate(r4.u32);
+
+    *pDeltaTime = m_lastLoadingFrameDelta;
+}

UnleashedRecomp/stdafx.h

@@ -12,6 +12,7 @@
 #include <vector>
 #include <string>
 #include <cassert>
+#include <chrono>
 #include <xbox.h>
 #include <xxhash.h>
 #include <ankerl/unordered_dense.h>

UnleashedRecompLib/config/SWA.toml

@@ -417,3 +417,8 @@ jump_address_on_true = 0x827D20EC
 name = "ParticleTestDrawIndexedPrimitiveMidAsmHook"
 address = 0x827D25AC
 registers = ["r7"]
+
+[[midasm_hook]]
+name = "LoadingScreenSpeedFixMidAsmHook"
+address = 0x824DAB60
+registers = ["r4"]

thirdparty/ShaderRecomp

@@ -1 +1 @@
-Subproject commit b7c037224665ac826490c142b800b11dfc4acf49
+Subproject commit 30f598604767602e3afce56b947e99dba2b51211