Merge branch 'shadps4-emu:main' into unity-flip-fix

.github/workflows/build.yml

@@ -111,10 +111,10 @@ jobs:
     - name: Setup Qt
       uses: jurplel/install-qt-action@v4
       with:
-        version: 6.7.3
+        version: 6.8.2
         host: windows
         target: desktop
-        arch: win64_msvc2019_64
+        arch: win64_msvc2022_64
         archives: qtbase qttools
         modules: qtmultimedia
 
@@ -228,7 +228,7 @@ jobs:
     - name: Setup Qt
       uses: jurplel/install-qt-action@v4
       with:
-        version: 6.7.3
+        version: 6.8.2
         host: mac
         target: desktop
         arch: clang_64

REUSE.toml

@@ -3,6 +3,7 @@ version = 1
 [[annotations]]
 path = [
     "REUSE.toml",
+    "crowdin.yml",
     "CMakeSettings.json",
     ".github/FUNDING.yml",
     ".github/shadps4.png",

crowdin.yml

@@ -0,0 +1,3 @@
+files:
+  - source: /src/qt_gui/translations/en_US.ts
+    translation: /%original_path%/%locale_with_underscore%.ts

documents/building-windows.md

@@ -25,7 +25,7 @@ Once you are within the installer:
 
 Beware, this requires you to create a Qt account. If you do not want to do this, please follow the MSYS2/MinGW compilation method instead.
 
-1. Under the current, non beta version of Qt (at the time of writing 6.7.3), select the option `MSVC 2022 64-bit` or similar, as well as `QT Multimedia`.  
+1. Under the current, non beta version of Qt (at the time of writing 6.8.2), select the option `MSVC 2022 64-bit` or similar, as well as `QT Multimedia`.  
    If you are on Windows on ARM / Qualcomm Snapdragon Elite X, select `MSVC 2022 ARM64` instead.
 
    Go through the installation normally. If you know what you are doing, you may unselect individual components that eat up too much disk space.
@@ -35,7 +35,7 @@ Beware, this requires you to create a Qt account. If you do not want to do this,
 Once you are finished, you will have to configure Qt within Visual Studio:
 
 1. Tools -> Options -> Qt -> Versions
-2. Add a new Qt version and navigate it to the correct folder. Should look like so: `C:\Qt\6.7.3\msvc2022_64`
+2. Add a new Qt version and navigate it to the correct folder. Should look like so: `C:\Qt\6.8.2\msvc2022_64`
 3. Enable the default checkmark on the new version you just created.
 
 ### (Prerequisite) Download [**Git for Windows**](https://git-scm.com/download/win)
@@ -55,7 +55,7 @@ Go through the Git for Windows installation as normal
 3. If you want to build shadPS4 with the Qt Gui:
    1. Click x64-Clang-Release and select "Manage Configurations"
    2. Look for "CMake command arguments" and add to the text field  
-    `-DENABLE_QT_GUI=ON -DCMAKE_PREFIX_PATH=C:\Qt\6.7.3\msvc2022_64`  
+    `-DENABLE_QT_GUI=ON -DCMAKE_PREFIX_PATH=C:\Qt\6.8.2\msvc2022_64`  
     (Change Qt path if you've installed it to non-default path)
    3. Press CTRL+S to save and wait a moment for CMake generation
 4. Change the project to build to shadps4.exe
@@ -64,7 +64,7 @@ Go through the Git for Windows installation as normal
 Your shadps4.exe will be in `C:\path\to\source\Build\x64-Clang-Release\`
 
 To automatically populate the necessary files to run shadPS4.exe, run in a command prompt or terminal:  
-`C:\Qt\6.7.3\msvc2022_64\bin\windeployqt6.exe "C:\path\to\shadps4.exe"`  
+`C:\Qt\6.8.2\msvc2022_64\bin\windeployqt6.exe "C:\path\to\shadps4.exe"`  
 (Change Qt path if you've installed it to non-default path)
 
 ## Option 2: MSYS2/MinGW

src/core/libraries/gnmdriver/gnmdriver.cpp

@@ -2307,8 +2307,36 @@ s32 PS4_SYSV_ABI sceGnmUpdateGsShader(u32* cmdbuf, u32 size, const u32* gs_regs)
     return ORBIS_OK;
 }
 
-int PS4_SYSV_ABI sceGnmUpdateHsShader() {
+int PS4_SYSV_ABI sceGnmUpdateHsShader(u32* cmdbuf, u32 size, const u32* hs_regs, u32 ls_hs_config) {
-    LOG_ERROR(Lib_GnmDriver, "(STUBBED) called");
+    LOG_TRACE(Lib_GnmDriver, "called");
+
+    if (!cmdbuf || size <= 0x1c) {
+        return -1;
+    }
+
+    if (!hs_regs) {
+        LOG_ERROR(Lib_GnmDriver, "Null pointer passed as argument");
+        return -1;
+    }
+
+    if (hs_regs[1] != 0) {
+        LOG_ERROR(Lib_GnmDriver, "Invalid shader address");
+        return -1;
+    }
+
+    cmdbuf = PM4CmdSetData::SetShReg(cmdbuf, 0x108u, hs_regs[0],
+                                     0u); // SPI_SHADER_PGM_LO_HS/SPI_SHADER_PGM_HI_HS
+    cmdbuf = PM4CmdSetData::SetShReg(cmdbuf, 0x10au, hs_regs[2],
+                                     hs_regs[3]); // SPI_SHADER_PGM_RSRC1_HS/SPI_SHADER_PGM_RSRC1_LS
+    cmdbuf = WritePacket<PM4ItOpcode::Nop>(
+        cmdbuf, PM4ShaderType::ShaderGraphics, 0xc01e0286u, hs_regs[5],
+        hs_regs[6]); // VGT_HOS_MAX_TESS_LEVEL/VGT_HOS_MIN_TESS_LEVEL update
+    cmdbuf = WritePacket<PM4ItOpcode::Nop>(cmdbuf, PM4ShaderType::ShaderGraphics, 0xc01e02dbu,
+                                           hs_regs[4]); // VGT_TF_PARAM update
+    cmdbuf = WritePacket<PM4ItOpcode::Nop>(cmdbuf, PM4ShaderType::ShaderGraphics, 0xc01e02d6u,
+                                           ls_hs_config); // VGT_LS_HS_CONFIG update
+
+    WriteTrailingNop<11>(cmdbuf);
     return ORBIS_OK;
 }
 

src/core/libraries/gnmdriver/gnmdriver.h

@@ -227,7 +227,7 @@ int PS4_SYSV_ABI sceGnmUnregisterAllResourcesForOwner();
 int PS4_SYSV_ABI sceGnmUnregisterOwnerAndResources();
 int PS4_SYSV_ABI sceGnmUnregisterResource();
 s32 PS4_SYSV_ABI sceGnmUpdateGsShader(u32* cmdbuf, u32 size, const u32* gs_regs);
-int PS4_SYSV_ABI sceGnmUpdateHsShader();
+int PS4_SYSV_ABI sceGnmUpdateHsShader(u32* cmdbuf, u32 size, const u32* ps_regs, u32 ls_hs_config);
 s32 PS4_SYSV_ABI sceGnmUpdatePsShader(u32* cmdbuf, u32 size, const u32* ps_regs);
 s32 PS4_SYSV_ABI sceGnmUpdatePsShader350(u32* cmdbuf, u32 size, const u32* ps_regs);
 s32 PS4_SYSV_ABI sceGnmUpdateVsShader(u32* cmdbuf, u32 size, const u32* vs_regs,

src/qt_gui/cheats_patches.cpp

@@ -568,7 +568,7 @@ void CheatsPatches::downloadCheats(const QString& source, const QString& gameSer
                             } else {
                                 QMessageBox::warning(this, tr("Error"),
                                                      QString(tr("Failed to download file:") +
-                                                             "%1\n\n" + tr("Error:") + "%2")
+                                                             "%1\n\n" + tr("Error") + ":%2")
                                                          .arg(fileUrl)
                                                          .arg(fileReply->errorString()));
                             }
@@ -644,7 +644,7 @@ void CheatsPatches::downloadCheats(const QString& source, const QString& gameSer
                             } else {
                                 QMessageBox::warning(this, tr("Error"),
                                                      QString(tr("Failed to download file:") +
-                                                             "%1\n\n" + tr("Error:") + "%2")
+                                                             "%1\n\n" + tr("Error") + ":%2")
                                                          .arg(fileUrl)
                                                          .arg(fileReply->errorString()));
                             }
@@ -843,7 +843,7 @@ void CheatsPatches::compatibleVersionNotice(const QString repository) {
     foreach (const QString& xmlFile, xmlFiles) {
         QFile file(dir.filePath(xmlFile));
         if (!file.open(QIODevice::ReadOnly | QIODevice::Text)) {
-            QMessageBox::warning(this, tr("ERROR"),
+            QMessageBox::warning(this, tr("Error"),
                                  QString(tr("Failed to open file:") + "\n%1").arg(xmlFile));
             continue;
         }
@@ -871,7 +871,7 @@ void CheatsPatches::compatibleVersionNotice(const QString repository) {
         }
 
         if (xmlReader.hasError()) {
-            QMessageBox::warning(this, tr("ERROR"),
+            QMessageBox::warning(this, tr("Error"),
                                  QString(tr("XML ERROR:") + "\n%1").arg(xmlReader.errorString()));
         }
 
@@ -926,7 +926,7 @@ void CheatsPatches::createFilesJson(const QString& repository) {
     foreach (const QString& xmlFile, xmlFiles) {
         QFile file(dir.filePath(xmlFile));
         if (!file.open(QIODevice::ReadOnly | QIODevice::Text)) {
-            QMessageBox::warning(this, tr("ERROR"),
+            QMessageBox::warning(this, tr("Error"),
                                  QString(tr("Failed to open file:") + "\n%1").arg(xmlFile));
             continue;
         }
@@ -944,7 +944,7 @@ void CheatsPatches::createFilesJson(const QString& repository) {
         }
 
         if (xmlReader.hasError()) {
-            QMessageBox::warning(this, tr("ERROR"),
+            QMessageBox::warning(this, tr("Error"),
                                  QString(tr("XML ERROR:") + "\n%1").arg(xmlReader.errorString()));
         }
         filesObject[xmlFile] = titleIdsArray;
@@ -952,7 +952,7 @@ void CheatsPatches::createFilesJson(const QString& repository) {
 
     QFile jsonFile(dir.absolutePath() + "/files.json");
     if (!jsonFile.open(QIODevice::WriteOnly)) {
-        QMessageBox::warning(this, tr("ERROR"), tr("Failed to open files.json for writing"));
+        QMessageBox::warning(this, tr("Error"), tr("Failed to open files.json for writing"));
         return;
     }
 
@@ -1155,7 +1155,7 @@ void CheatsPatches::addPatchesToLayout(const QString& filePath) {
     QString fullPath = dir.filePath(folderPath);
 
     if (!dir.exists(fullPath)) {
-        QMessageBox::warning(this, tr("ERROR"),
+        QMessageBox::warning(this, tr("Error"),
                              QString(tr("Directory does not exist:") + "\n%1").arg(fullPath));
         return;
     }
@@ -1165,7 +1165,7 @@ void CheatsPatches::addPatchesToLayout(const QString& filePath) {
 
     QFile jsonFile(filesJsonPath);
     if (!jsonFile.open(QIODevice::ReadOnly)) {
-        QMessageBox::warning(this, tr("ERROR"), tr("Failed to open files.json for reading."));
+        QMessageBox::warning(this, tr("Error"), tr("Failed to open files.json for reading."));
         return;
     }
 
@@ -1189,7 +1189,7 @@ void CheatsPatches::addPatchesToLayout(const QString& filePath) {
 
             if (!xmlFile.open(QIODevice::ReadOnly)) {
                 QMessageBox::warning(
-                    this, tr("ERROR"),
+                    this, tr("Error"),
                     QString(tr("Failed to open file:") + "\n%1").arg(xmlFile.fileName()));
                 continue;
             }

src/qt_gui/pkg_viewer.cpp

@@ -18,17 +18,8 @@ PKGViewer::PKGViewer(std::shared_ptr<GameInfoClass> game_info_get, QWidget* pare
     treeWidget = new QTreeWidget(this);
     treeWidget->setColumnCount(9);
     QStringList headers;
-    headers << "Name"
+    headers << tr("Name") << tr("Serial") << tr("Installed") << tr("Size") << tr("Category")
-            << "Serial"
+            << tr("Type") << tr("App Ver") << tr("FW") << tr("Region") << tr("Flags") << tr("Path");
-            << "Installed"
-            << "Size"
-            << "Category"
-            << "Type"
-            << "App Ver"
-            << "FW"
-            << "Region"
-            << "Flags"
-            << "Path";
     treeWidget->setHeaderLabels(headers);
     treeWidget->header()->setDefaultAlignment(Qt::AlignCenter);
     treeWidget->setContextMenuPolicy(Qt::CustomContextMenu);
@@ -36,7 +27,7 @@ PKGViewer::PKGViewer(std::shared_ptr<GameInfoClass> game_info_get, QWidget* pare
     this->setCentralWidget(treeWidget);
     QMenuBar* menuBar = new QMenuBar(this);
     menuBar->setContextMenuPolicy(Qt::PreventContextMenu);
-    QMenu* fileMenu = menuBar->addMenu(tr("&File"));
+    QMenu* fileMenu = menuBar->addMenu(tr("File"));
     QAction* openFolderAct = new QAction(tr("Open Folder"), this);
     fileMenu->addAction(openFolderAct);
     this->setMenuBar(menuBar);
@@ -114,15 +105,15 @@ void PKGViewer::ProcessPKGInfo() {
             return;
         }
         psf.Open(package.sfo);
-        QString title_name =
+        QString title_name = QString::fromStdString(
-            QString::fromStdString(std::string{psf.GetString("TITLE").value_or("Unknown")});
+            std::string{psf.GetString("TITLE").value_or(std::string{tr("Unknown").toStdString()})});
-        QString title_id =
+        QString title_id = QString::fromStdString(std::string{
-            QString::fromStdString(std::string{psf.GetString("TITLE_ID").value_or("Unknown")});
+            psf.GetString("TITLE_ID").value_or(std::string{tr("Unknown").toStdString()})});
         QString app_type = GameListUtils::GetAppType(psf.GetInteger("APP_TYPE").value_or(0));
-        QString app_version =
+        QString app_version = QString::fromStdString(std::string{
-            QString::fromStdString(std::string{psf.GetString("APP_VER").value_or("Unknown")});
+            psf.GetString("APP_VER").value_or(std::string{tr("Unknown").toStdString()})});
-        QString title_category =
+        QString title_category = QString::fromStdString(std::string{
-            QString::fromStdString(std::string{psf.GetString("CATEGORY").value_or("Unknown")});
+            psf.GetString("CATEGORY").value_or(std::string{tr("Unknown").toStdString()})});
         QString pkg_size = GameListUtils::FormatSize(package.GetPkgHeader().pkg_size);
         pkg_content_flag = package.GetPkgHeader().pkg_content_flags;
         QString flagss = "";
@@ -134,7 +125,7 @@ void PKGViewer::ProcessPKGInfo() {
             }
         }
 
-        QString fw_ = "Unknown";
+        QString fw_ = tr("Unknown");
         if (const auto fw_int_opt = psf.GetInteger("SYSTEM_VER"); fw_int_opt.has_value()) {
             const u32 fw_int = *fw_int_opt;
             if (fw_int == 0) {
@@ -221,6 +212,6 @@ void PKGViewer::ProcessPKGInfo() {
     // Update status bar.
     statusBar->clearMessage();
     int numPkgs = m_pkg_list.size();
-    QString statusMessage = QString::number(numPkgs) + " Package.";
+    QString statusMessage = QString::number(numPkgs) + " " + tr("Package");
     statusBar->showMessage(statusMessage);
 }

src/qt_gui/translations/sv_SE.ts

@@ -1,8 +1,8 @@
 <?xml version="1.0" encoding="utf-8"?>
+<!-- SPDX-FileCopyrightText: Copyright 2025 shadPS4 Emulator Project
+     SPDX-License-Identifier: GPL-2.0-or-later -->
 <!DOCTYPE TS>
 <TS version="2.1" language="sv_SE">
-	<!-- SPDX-FileCopyrightText: Copyright 2025 shadPS4 Emulator Project
-         SPDX-License-Identifier: GPL-2.0-or-later -->
 <context>
     <name>AboutDialog</name>
     <message>
@@ -244,14 +244,6 @@
         <source>Can&apos;t apply cheats before the game is started</source>
         <translation>Kan inte tillämpa fusk innan spelet är startat</translation>
     </message>
-    <message>
-        <source>Error:</source>
-        <translation>Fel:</translation>
-    </message>
-    <message>
-        <source>ERROR</source>
-        <translation>FEL</translation>
-    </message>
     <message>
         <source>Close</source>
         <translation>Stäng</translation>
@@ -386,10 +378,6 @@
         <source>Unable to update compatibility data! Try again later.</source>
         <translation>Kunde inte uppdatera kompatibilitetsdata! Försök igen senare.</translation>
     </message>
-    <message>
-        <source>Unable to open compatibility.json for writing.</source>
-        <translation type="vanished">Kunde inte öppna compatibility.json för skrivning.</translation>
-    </message>
     <message>
         <source>Unknown</source>
         <translation>Okänt</translation>
@@ -673,10 +661,6 @@
         <source>Game can be completed with playable performance and no major glitches</source>
         <translation>Spelet kan spelas klart med spelbar prestanda och utan större problem</translation>
     </message>
-    <message>
-        <source>Click to go to issue</source>
-        <translation type="vanished">Klicka för att gå till problem</translation>
-    </message>
     <message>
         <source>Last updated</source>
         <translation>Senast uppdaterad</translation>
@@ -1196,14 +1180,66 @@
         <source>Open Folder</source>
         <translation>Öppna mapp</translation>
     </message>
-    <message>
-        <source>&amp;File</source>
-        <translation>&amp;Arkiv</translation>
-    </message>
     <message>
         <source>PKG ERROR</source>
         <translation>PKG-FEL</translation>
     </message>
+    <message>
+        <source>Name</source>
+        <translation type="unfinished">Namn</translation>
+    </message>
+    <message>
+        <source>Serial</source>
+        <translation type="unfinished">Serienummer</translation>
+    </message>
+    <message>
+        <source>Installed</source>
+        <translation type="unfinished"></translation>
+    </message>
+    <message>
+        <source>Size</source>
+        <translation type="unfinished">Storlek</translation>
+    </message>
+    <message>
+        <source>Category</source>
+        <translation type="unfinished"></translation>
+    </message>
+    <message>
+        <source>Type</source>
+        <translation type="unfinished"></translation>
+    </message>
+    <message>
+        <source>App Ver</source>
+        <translation type="unfinished"></translation>
+    </message>
+    <message>
+        <source>FW</source>
+        <translation type="unfinished"></translation>
+    </message>
+    <message>
+        <source>Region</source>
+        <translation type="unfinished">Region</translation>
+    </message>
+    <message>
+        <source>Flags</source>
+        <translation type="unfinished"></translation>
+    </message>
+    <message>
+        <source>Path</source>
+        <translation type="unfinished">Sökväg</translation>
+    </message>
+    <message>
+        <source>File</source>
+        <translation type="unfinished">Arkiv</translation>
+    </message>
+    <message>
+        <source>Unknown</source>
+        <translation type="unfinished">Okänt</translation>
+    </message>
+    <message>
+        <source>Package</source>
+        <translation type="unfinished"></translation>
+    </message>
 </context>
 <context>
     <name>SettingsDialog</name>
@@ -1255,10 +1291,6 @@
         <source>Show Splash</source>
         <translation>Visa startskärm</translation>
     </message>
-    <message>
-        <source>ps4proCheckBox</source>
-        <translation type="vanished">Är PS4 Pro:\nGör att emulatorn agerar som en PS4 PRO, vilket kan aktivera speciella funktioner i spel som har stöd för det</translation>
-    </message>
     <message>
         <source>Enable Discord Rich Presence</source>
         <translation>Aktivera Discord Rich Presence</translation>
@@ -1323,10 +1355,6 @@
         <source>Graphics</source>
         <translation>Grafik</translation>
     </message>
-    <message>
-        <source>Gui</source>
-        <translation type="vanished">Gränssnitt</translation>
-    </message>
     <message>
         <source>User</source>
         <translation>Användare</translation>
@@ -1743,6 +1771,14 @@
         <source>GUIBackgroundImageGroupBox</source>
         <translation>Bakgrundsbild:\nKontrollerar opaciteten för spelets bakgrundsbild</translation>
     </message>
+    <message>
+        <source>Enable HDR</source>
+        <translation type="unfinished"></translation>
+    </message>
+    <message>
+        <source>enableHDRCheckBox</source>
+        <translation type="unfinished"></translation>
+    </message>
 </context>
 <context>
     <name>TrophyViewer</name>

src/shader_recompiler/backend/spirv/emit_spirv_instructions.h

@@ -120,10 +120,8 @@ Id EmitUndefU32(EmitContext& ctx);
 Id EmitUndefU64(EmitContext& ctx);
 Id EmitLoadSharedU32(EmitContext& ctx, Id offset);
 Id EmitLoadSharedU64(EmitContext& ctx, Id offset);
-Id EmitLoadSharedU128(EmitContext& ctx, Id offset);
 void EmitWriteSharedU32(EmitContext& ctx, Id offset, Id value);
 void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value);
-void EmitWriteSharedU128(EmitContext& ctx, Id offset, Id value);
 Id EmitSharedAtomicIAdd32(EmitContext& ctx, Id offset, Id value);
 Id EmitSharedAtomicUMax32(EmitContext& ctx, Id offset, Id value);
 Id EmitSharedAtomicSMax32(EmitContext& ctx, Id offset, Id value);

src/shader_recompiler/backend/spirv/emit_spirv_shared_memory.cpp

@@ -38,24 +38,6 @@ Id EmitLoadSharedU64(EmitContext& ctx, Id offset) {
     }
 }
 
-Id EmitLoadSharedU128(EmitContext& ctx, Id offset) {
-    const Id shift_id{ctx.ConstU32(2U)};
-    const Id base_index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
-    std::array<Id, 4> values{};
-    for (u32 i = 0; i < 4; ++i) {
-        const Id index{i == 0 ? base_index : ctx.OpIAdd(ctx.U32[1], base_index, ctx.ConstU32(i))};
-        if (ctx.info.has_emulated_shared_memory) {
-            const Id pointer{ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32,
-                                               ctx.u32_zero_value, index)};
-            values[i] = ctx.OpLoad(ctx.U32[1], pointer);
-        } else {
-            const Id pointer{ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, index)};
-            values[i] = ctx.OpLoad(ctx.U32[1], pointer);
-        }
-    }
-    return ctx.OpCompositeConstruct(ctx.U32[4], values);
-}
-
 void EmitWriteSharedU32(EmitContext& ctx, Id offset, Id value) {
     const Id shift{ctx.ConstU32(2U)};
     const Id word_offset{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift)};
@@ -88,20 +70,4 @@ void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value) {
     }
 }
 
-void EmitWriteSharedU128(EmitContext& ctx, Id offset, Id value) {
-    const Id shift{ctx.ConstU32(2U)};
-    const Id base_index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift)};
-    for (u32 i = 0; i < 4; ++i) {
-        const Id index{i == 0 ? base_index : ctx.OpIAdd(ctx.U32[1], base_index, ctx.ConstU32(i))};
-        if (ctx.info.has_emulated_shared_memory) {
-            const Id pointer{ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32,
-                                               ctx.u32_zero_value, index)};
-            ctx.OpStore(pointer, ctx.OpCompositeExtract(ctx.U32[1], value, i));
-        } else {
-            const Id pointer{ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, index)};
-            ctx.OpStore(pointer, ctx.OpCompositeExtract(ctx.U32[1], value, i));
-        }
-    }
-}
-
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -813,6 +813,8 @@ void EmitContext::DefineSharedMemory() {
     if (!info.uses_shared) {
         return;
     }
+    ASSERT(info.stage == Stage::Compute);
+
     const u32 max_shared_memory_size = profile.max_shared_memory_size;
     u32 shared_memory_size = runtime_info.cs_info.shared_memory_size;
     if (shared_memory_size == 0) {

src/shader_recompiler/info.h

@@ -233,7 +233,8 @@ struct Info {
     }
 
     void AddBindings(Backend::Bindings& bnd) const {
-        const auto total_buffers = buffers.size() + (has_readconst ? 1 : 0);
+        const auto total_buffers =
+            buffers.size() + (has_readconst ? 1 : 0) + (has_emulated_shared_memory ? 1 : 0);
         bnd.buffer += total_buffers;
         bnd.unified += total_buffers + images.size() + samplers.size();
         bnd.user_data += ud_mask.NumRegs();

src/shader_recompiler/ir/ir_emitter.cpp

@@ -308,8 +308,6 @@ Value IREmitter::LoadShared(int bit_size, bool is_signed, const U32& offset) {
         return Inst<U32>(Opcode::LoadSharedU32, offset);
     case 64:
         return Inst(Opcode::LoadSharedU64, offset);
-    case 128:
-        return Inst(Opcode::LoadSharedU128, offset);
     default:
         UNREACHABLE_MSG("Invalid bit size {}", bit_size);
     }
@@ -323,9 +321,6 @@ void IREmitter::WriteShared(int bit_size, const Value& value, const U32& offset)
     case 64:
         Inst(Opcode::WriteSharedU64, offset, value);
         break;
-    case 128:
-        Inst(Opcode::WriteSharedU128, offset, value);
-        break;
     default:
         UNREACHABLE_MSG("Invalid bit size {}", bit_size);
     }

src/shader_recompiler/ir/microinstruction.cpp

@@ -78,7 +78,6 @@ bool Inst::MayHaveSideEffects() const noexcept {
     case Opcode::BufferAtomicSwap32:
     case Opcode::DataAppend:
     case Opcode::DataConsume:
-    case Opcode::WriteSharedU128:
     case Opcode::WriteSharedU64:
     case Opcode::WriteSharedU32:
     case Opcode::SharedAtomicIAdd32:

src/shader_recompiler/ir/opcodes.inc

@@ -32,10 +32,8 @@ OPCODE(EmitPrimitive,                                       Void,
 // Shared memory operations
 OPCODE(LoadSharedU32,                                       U32,            U32,                                                                            )
 OPCODE(LoadSharedU64,                                       U32x2,          U32,                                                                            )
-OPCODE(LoadSharedU128,                                      U32x4,          U32,                                                                            )
 OPCODE(WriteSharedU32,                                      Void,           U32,            U32,                                                            )
 OPCODE(WriteSharedU64,                                      Void,           U32,            U32x2,                                                          )
-OPCODE(WriteSharedU128,                                     Void,           U32,            U32x4,                                                          )
 
 // Shared atomic operations
 OPCODE(SharedAtomicIAdd32,                                  U32,            U32,            U32,                                                            )

src/shader_recompiler/ir/passes/hull_shader_transform.cpp

@@ -225,10 +225,8 @@ private:
         switch (use.user->GetOpcode()) {
         case IR::Opcode::LoadSharedU32:
         case IR::Opcode::LoadSharedU64:
-        case IR::Opcode::LoadSharedU128:
         case IR::Opcode::WriteSharedU32:
-        case IR::Opcode::WriteSharedU64:
+        case IR::Opcode::WriteSharedU64: {
-        case IR::Opcode::WriteSharedU128: {
             u32 counter = inst->Flags<u32>();
             inst->SetFlags<u32>(counter + inc);
             // Stop here
@@ -435,12 +433,9 @@ void HullShaderTransform(IR::Program& program, RuntimeInfo& runtime_info) {
             }
 
             case IR::Opcode::WriteSharedU32:
-            case IR::Opcode::WriteSharedU64:
+            case IR::Opcode::WriteSharedU64: {
-            case IR::Opcode::WriteSharedU128: {
                 IR::IREmitter ir{*block, IR::Block::InstructionList::s_iterator_to(inst)};
-                const u32 num_dwords = opcode == IR::Opcode::WriteSharedU32
+                const u32 num_dwords = opcode == IR::Opcode::WriteSharedU32 ? 1 : 2;
-                                           ? 1
-                                           : (opcode == IR::Opcode::WriteSharedU64 ? 2 : 4);
                 const IR::U32 addr{inst.Arg(0)};
                 const IR::U32 data{inst.Arg(1).Resolve()};
 
@@ -480,15 +475,12 @@ void HullShaderTransform(IR::Program& program, RuntimeInfo& runtime_info) {
                 break;
             }
 
-            case IR::Opcode::LoadSharedU32: {
+            case IR::Opcode::LoadSharedU32:
-            case IR::Opcode::LoadSharedU64:
+            case IR::Opcode::LoadSharedU64: {
-            case IR::Opcode::LoadSharedU128:
                 IR::IREmitter ir{*block, IR::Block::InstructionList::s_iterator_to(inst)};
                 const IR::U32 addr{inst.Arg(0)};
                 const AttributeRegion region = GetAttributeRegionKind(&inst, info, runtime_info);
-                const u32 num_dwords = opcode == IR::Opcode::LoadSharedU32
+                const u32 num_dwords = opcode == IR::Opcode::LoadSharedU32 ? 1 : 2;
-                                           ? 1
-                                           : (opcode == IR::Opcode::LoadSharedU64 ? 2 : 4);
                 ASSERT_MSG(region == AttributeRegion::InputCP ||
                                region == AttributeRegion::OutputCP,
                            "Unhandled read of patchconst attribute in hull shader");
@@ -562,14 +554,11 @@ void DomainShaderTransform(IR::Program& program, RuntimeInfo& runtime_info) {
             IR::IREmitter ir{*block, IR::Block::InstructionList::s_iterator_to(inst)};
             const auto opcode = inst.GetOpcode();
             switch (inst.GetOpcode()) {
-            case IR::Opcode::LoadSharedU32: {
+            case IR::Opcode::LoadSharedU32:
-            case IR::Opcode::LoadSharedU64:
+            case IR::Opcode::LoadSharedU64: {
-            case IR::Opcode::LoadSharedU128:
                 const IR::U32 addr{inst.Arg(0)};
                 AttributeRegion region = GetAttributeRegionKind(&inst, info, runtime_info);
-                const u32 num_dwords = opcode == IR::Opcode::LoadSharedU32
+                const u32 num_dwords = opcode == IR::Opcode::LoadSharedU32 ? 1 : 2;
-                                           ? 1
-                                           : (opcode == IR::Opcode::LoadSharedU64 ? 2 : 4);
                 const auto GetInput = [&](IR::U32 addr, u32 off_dw) -> IR::F32 {
                     if (region == AttributeRegion::OutputCP) {
                         return ReadTessControlPointAttribute(
@@ -611,10 +600,8 @@ void TessellationPreprocess(IR::Program& program, RuntimeInfo& runtime_info) {
                 switch (inst.GetOpcode()) {
                 case IR::Opcode::LoadSharedU32:
                 case IR::Opcode::LoadSharedU64:
-                case IR::Opcode::LoadSharedU128:
                 case IR::Opcode::WriteSharedU32:
-                case IR::Opcode::WriteSharedU64:
+                case IR::Opcode::WriteSharedU64: {
-                case IR::Opcode::WriteSharedU128: {
                     IR::Value addr = inst.Arg(0);
                     auto read_const_buffer = IR::BreadthFirstSearch(
                         addr, [](IR::Inst* maybe_tess_const) -> std::optional<IR::Inst*> {

src/shader_recompiler/ir/passes/ir_passes.h

@@ -20,7 +20,7 @@ void FlattenExtendedUserdataPass(IR::Program& program);
 void ResourceTrackingPass(IR::Program& program);
 void CollectShaderInfoPass(IR::Program& program);
 void LowerBufferFormatToRaw(IR::Program& program);
-void LowerSharedMemToRegisters(IR::Program& program);
+void LowerSharedMemToRegisters(IR::Program& program, const RuntimeInfo& runtime_info);
 void RingAccessElimination(const IR::Program& program, const RuntimeInfo& runtime_info,
                            Stage stage);
 void TessellationPreprocess(IR::Program& program, RuntimeInfo& runtime_info);

src/shader_recompiler/ir/passes/lower_shared_mem_to_registers.cpp

@@ -1,38 +1,81 @@
 // SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
-#include <boost/container/small_vector.hpp>
+#include <unordered_map>
+
+#include "shader_recompiler/ir/ir_emitter.h"
 #include "shader_recompiler/ir/program.h"
 
 namespace Shader::Optimization {
 
-void LowerSharedMemToRegisters(IR::Program& program) {
+static bool IsSharedMemoryInst(const IR::Inst& inst) {
-    boost::container::small_vector<IR::Inst*, 8> ds_writes;
+    const auto opcode = inst.GetOpcode();
-    Info& info{program.info};
+    return opcode == IR::Opcode::LoadSharedU32 || opcode == IR::Opcode::LoadSharedU64 ||
+           opcode == IR::Opcode::WriteSharedU32 || opcode == IR::Opcode::WriteSharedU64;
+}
+
+static u32 GetSharedMemImmOffset(const IR::Inst& inst) {
+    const auto* address = inst.Arg(0).InstRecursive();
+    ASSERT(address->GetOpcode() == IR::Opcode::IAdd32);
+    const auto ir_offset = address->Arg(1);
+    ASSERT_MSG(ir_offset.IsImmediate());
+    const auto offset = ir_offset.U32();
+    // Typical usage is the compiler spilling registers into shared memory, with 256 bytes between
+    // each register to account for 4 bytes per register times 64 threads per group. Ensure that
+    // this assumption holds, as if it does not this approach may need to be revised.
+    ASSERT_MSG(offset % 256 == 0, "Unexpected shared memory offset alignment: {}", offset);
+    return offset;
+}
+
+static void ConvertSharedMemToVgpr(IR::IREmitter& ir, IR::Inst& inst, const IR::VectorReg vgpr) {
+    switch (inst.GetOpcode()) {
+    case IR::Opcode::LoadSharedU32:
+        inst.ReplaceUsesWithAndRemove(ir.GetVectorReg(vgpr));
+        break;
+    case IR::Opcode::LoadSharedU64:
+        inst.ReplaceUsesWithAndRemove(
+            ir.CompositeConstruct(ir.GetVectorReg(vgpr), ir.GetVectorReg(vgpr + 1)));
+        break;
+    case IR::Opcode::WriteSharedU32:
+        ir.SetVectorReg(vgpr, IR::U32{inst.Arg(1)});
+        inst.Invalidate();
+        break;
+    case IR::Opcode::WriteSharedU64: {
+        const auto value = inst.Arg(1);
+        ir.SetVectorReg(vgpr, IR::U32{ir.CompositeExtract(value, 0)});
+        ir.SetVectorReg(vgpr, IR::U32{ir.CompositeExtract(value, 1)});
+        inst.Invalidate();
+        break;
+    }
+    default:
+        UNREACHABLE_MSG("Unknown shared memory opcode: {}", inst.GetOpcode());
+    }
+}
+
+void LowerSharedMemToRegisters(IR::Program& program, const RuntimeInfo& runtime_info) {
+    u32 next_vgpr_num = runtime_info.num_allocated_vgprs;
+    std::unordered_map<u32, IR::VectorReg> vgpr_map;
+    const auto get_vgpr = [&next_vgpr_num, &vgpr_map](const u32 offset) {
+        const auto [it, is_new] = vgpr_map.try_emplace(offset);
+        if (is_new) {
+            ASSERT_MSG(next_vgpr_num < 256, "Out of VGPRs");
+            const auto new_vgpr = static_cast<IR::VectorReg>(next_vgpr_num++);
+            it->second = new_vgpr;
+        }
+        return it->second;
+    };
+
     for (IR::Block* const block : program.blocks) {
         for (IR::Inst& inst : block->Instructions()) {
-            const auto opcode = inst.GetOpcode();
+            if (!IsSharedMemoryInst(inst)) {
-            if (opcode == IR::Opcode::WriteSharedU32 || opcode == IR::Opcode::WriteSharedU64) {
-                ds_writes.emplace_back(&inst);
                 continue;
             }
-            if (opcode == IR::Opcode::LoadSharedU32 || opcode == IR::Opcode::LoadSharedU64) {
+            const auto offset = GetSharedMemImmOffset(inst);
-                // Search for write instruction with same offset
+            const auto vgpr = get_vgpr(offset);
-                const IR::Inst* prod = inst.Arg(0).InstRecursive();
+            IR::IREmitter ir{*block, IR::Block::InstructionList::s_iterator_to(inst)};
-                const auto it = std::ranges::find_if(ds_writes, [&](const IR::Inst* write) {
+            ConvertSharedMemToVgpr(ir, inst, vgpr);
-                    const IR::Inst* write_prod = write->Arg(0).InstRecursive();
-                    return write_prod->Arg(1).U32() == prod->Arg(1).U32();
-                });
-                ASSERT(it != ds_writes.end());
-                // Replace data read with value written.
-                inst.ReplaceUsesWithAndRemove((*it)->Arg(1));
-            }
         }
     }
-    // We should have eliminated everything. Invalidate data write instructions.
-    for (const auto inst : ds_writes) {
-        inst->Invalidate();
-    }
 }
 
 } // namespace Shader::Optimization

src/shader_recompiler/recompiler.cpp

@@ -65,6 +65,10 @@ IR::Program TranslateProgram(std::span<const u32> code, Pools& pools, Info& info
     // Run optimization passes
     const auto stage = program.info.stage;
 
+    if (stage == Stage::Fragment) {
+        // Before SSA pass, as it will rewrite to VGPR load/store.
+        Shader::Optimization::LowerSharedMemToRegisters(program, runtime_info);
+    }
     Shader::Optimization::SsaRewritePass(program.post_order_blocks);
     Shader::Optimization::IdentityRemovalPass(program.blocks);
     if (info.l_stage == LogicalStage::TessellationControl) {
@@ -82,9 +86,6 @@ IR::Program TranslateProgram(std::span<const u32> code, Pools& pools, Info& info
     }
     Shader::Optimization::ConstantPropagationPass(program.post_order_blocks);
     Shader::Optimization::RingAccessElimination(program, runtime_info, stage);
-    if (stage != Stage::Compute) {
-        Shader::Optimization::LowerSharedMemToRegisters(program);
-    }
     Shader::Optimization::ConstantPropagationPass(program.post_order_blocks);
     Shader::Optimization::FlattenExtendedUserdataPass(program);
     Shader::Optimization::ResourceTrackingPass(program);

src/video_core/amdgpu/liverpool.h

@@ -143,6 +143,11 @@ struct Liverpool {
             const u32 num_dwords = bininfo.length / sizeof(u32);
             return std::span{code, num_dwords};
         }
+
+        [[nodiscard]] u32 NumVgprs() const {
+            // Each increment allocates 4 registers, where 0 = 4 registers.
+            return (settings.num_vgprs + 1) * 4;
+        }
     };
 
     struct HsTessFactorClamp {

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -84,7 +84,7 @@ const Shader::RuntimeInfo& PipelineCache::BuildRuntimeInfo(Stage stage, LogicalS
     const auto BuildCommon = [&](const auto& program) {
         info.num_user_data = program.settings.num_user_regs;
         info.num_input_vgprs = program.settings.vgpr_comp_cnt;
-        info.num_allocated_vgprs = program.settings.num_vgprs * 4;
+        info.num_allocated_vgprs = program.NumVgprs();
         info.fp_denorm_mode32 = program.settings.fp_denorm_mode32;
         info.fp_round_mode32 = program.settings.fp_round_mode32;
     };

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -535,6 +535,7 @@ void Rasterizer::BindBuffers(const Shader::Info& stage, Shader::Backend::Binding
             .descriptorType = vk::DescriptorType::eStorageBuffer,
             .pBufferInfo = &buffer_infos.back(),
         });
+        ++binding.buffer;
     }
 
     // Bind the flattened user data buffer as a UBO so it's accessible to the shader