GS/DX12: Fix recreating swapchain failing on vsync mode changes

Make sure uniform buffers are 16 bytes aligned.
Make sure full size is 32 bytes with padding, was 28 previously which can be bad for cpu cache.

.github/workflows/scripts/linux/build-dependencies-qt.sh

@@ -25,7 +25,7 @@ LIBBACKTRACE=ad106d5fdd5d960bd33fae1c48a351af567fd075
 LIBJPEGTURBO=3.1.2
 LIBPNG=1.6.53
 LIBWEBP=1.6.0
-NVENC=11.1.5.3
+NVENC=13.0.19.0
 SDL=SDL3-3.4.0
 QT=6.10.1
 QTAPNG=1.3.0
@@ -57,7 +57,7 @@ e4ab7009bf0629fd11982d4c2aa83964cf244cffba7347ecd39019a9e38c4564  libwebp-$LIBWE
 082cbf5f429e0d80820f68dc2b507a94d4cc1b4e70817b119bbb8ec6a69584b8  $SDL.tar.gz
 452a1a290bd0cf18737fad0057dc17b7fdf10a73eda2d6d4f31ba04fda25ef2c  libpng-$LIBPNG-apng.patch.gz
 537512904744b35e232912055ccf8ec66d768639ff3abe5788d90d792ec5f48b  lz4-$LZ4.tar.gz
-2974b91062197e0527dffa3aadd8fe3bfa6681ae45f5ff9181bc0ca6479abd59  nv-codec-headers-$NVENC.tar.gz
+13da39edb3a40ed9713ae390ca89faa2f1202c9dda869ef306a8d4383e242bee  nv-codec-headers-$NVENC.tar.gz
 c465aa56757e7746ac707f582b6e2d51546569a4a2488c1172fb543aa5fdfc2c  vulkan-sdk-$VULKAN.tar.gz
 eb33e51f49a15e023950cd7825ca74a4a2b43db8354825ac24fc1b7ee09e6fa3  zstd-$ZSTD.tar.gz
 5a6226f7e23db51fdc3223121eba53f3f5447cf0cc4d6cb82a3a2df7a65d265d  qtbase-everywhere-src-$QT.tar.xz

pcsx2-qt/Settings/BIOSSettingsWidget.ui

@@ -135,7 +135,7 @@
    <item>
     <widget class="QGroupBox" name="groupBox_2">
      <property name="title">
-      <string>Options and Patches</string>
+      <string>Fast Boot Options</string>
      </property>
      <layout class="QVBoxLayout" name="verticalLayout">
       <item>

pcsx2-qt/Translations/pcsx2-qt_en.ts

@@ -2223,7 +2223,7 @@ Leaderboard Position: {1} of {2}</source>
     </message>
     <message>
         <location filename="../Settings/BIOSSettingsWidget.ui" line="138"/>
-        <source>Options and Patches</source>
+        <source>Fast Boot Options</source>
         <translation type="unfinished"></translation>
     </message>
     <message>
@@ -6804,11 +6804,6 @@ The URL was: %1</source>
         <source>BIOS Selection</source>
         <translation type="unfinished"></translation>
     </message>
-    <message>
-        <location filename="../../pcsx2/ImGui/FullscreenUI.cpp" line="9512"/>
-        <source>Options and Patches</source>
-        <translation type="unfinished"></translation>
-    </message>
     <message>
         <location filename="../../pcsx2/ImGui/FullscreenUI.cpp" line="9513"/>
         <source>Skips the intro screen, and bypasses region checks.</source>
@@ -9942,6 +9937,11 @@ Do you want to shutdown anyway and IRREVERSIBLY CORRUPT YOUR MEMORY CARD?</sourc
         <source>Shows the number of dumped and loaded texture replacements on the OSD.</source>
         <translation type="unfinished"></translation>
     </message>
+    <message>
+        <location filename="../../pcsx2/ImGui/FullscreenUI.cpp" line="9512"/>
+        <source>Fast Boot Options</source>
+        <translation type="unfinished"></translation>
+    </message>
     <message>
         <location filename="../../pcsx2/ImGui/FullscreenUI.cpp" line="9604"/>
         <source>Bilinear Dirty Upscale</source>
@@ -11893,7 +11893,7 @@ This action cannot be undone.</source>
         <translation type="unfinished"></translation>
     </message>
     <message>
-        <location filename="../../pcsx2/GS/Renderers/Vulkan/GSDeviceVK.cpp" line="5055"/>
+        <location filename="../../pcsx2/GS/Renderers/Vulkan/GSDeviceVK.cpp" line="5052"/>
         <source>Spin GPU During Readbacks is enabled, but calibrated timestamps are unavailable.  This might be really slow.</source>
         <translation type="unfinished"></translation>
     </message>

pcsx2/GS/Renderers/Common/GSDevice.cpp

@@ -899,14 +899,13 @@ void GSDevice::ShadeBoost()
 	if (ResizeRenderTarget(&m_target_tmp, m_current->GetWidth(), m_current->GetHeight(), false, false))
 	{
 		// predivide to avoid the divide (multiply) in the shader
-		const float params[4] = {
+		const GSVector4 params(
 			static_cast<float>(GSConfig.ShadeBoost_Brightness) * (1.0f / 50.0f),
 			static_cast<float>(GSConfig.ShadeBoost_Contrast) * (1.0f / 50.0f),
 			static_cast<float>(GSConfig.ShadeBoost_Saturation) * (1.0f / 50.0f),
-			static_cast<float>(GSConfig.ShadeBoost_Gamma) * (1.0f / 50.0f),
-		};
+			static_cast<float>(GSConfig.ShadeBoost_Gamma) * (1.0f / 50.0f));
 
-		DoShadeBoost(m_current, m_target_tmp, params);
+		DoShadeBoost(m_current, m_target_tmp, params.v);
 
 		m_current = m_target_tmp;
 	}

pcsx2/GS/Renderers/DX11/GSDevice11.cpp

@@ -1397,7 +1397,7 @@ void GSDevice11::PresentRect(GSTexture* sTex, const GSVector4& sRect, GSTexture*
 	const float right = dRect.z * 2 / ds.x - 1.0f;
 	const float bottom = 1.0f - dRect.w * 2 / ds.y;
 
-	GSVertexPT1 vertices[] =
+	const GSVertexPT1 vertices[] =
 	{
 		{GSVector4(left, top, 0.5f, 1.0f), GSVector2(sRect.x, sRect.y)},
 		{GSVector4(right, top, 0.5f, 1.0f), GSVector2(sRect.z, sRect.y)},
@@ -1427,13 +1427,14 @@ void GSDevice11::PresentRect(GSTexture* sTex, const GSVector4& sRect, GSTexture*
 void GSDevice11::UpdateCLUTTexture(GSTexture* sTex, float sScale, u32 offsetX, u32 offsetY, GSTexture* dTex, u32 dOffset, u32 dSize)
 {
 	// match merge cb
-	struct Uniforms
+	struct alignas(16) Uniforms
 	{
 		float scale;
 		float pad1[3];
 		u32 offsetX, offsetY, dOffset;
+		u32 pad2;
 	};
-	const Uniforms cb = {sScale, {}, offsetX, offsetY, dOffset};
+	const Uniforms cb = {sScale, {}, offsetX, offsetY, dOffset, 0};
 	m_ctx->UpdateSubresource(m_merge.cb.get(), 0, nullptr, &cb, 0, 0);
 
 	const GSVector4 dRect(0, 0, dSize, 1);
@@ -1444,14 +1445,15 @@ void GSDevice11::UpdateCLUTTexture(GSTexture* sTex, float sScale, u32 offsetX, u
 void GSDevice11::ConvertToIndexedTexture(GSTexture* sTex, float sScale, u32 offsetX, u32 offsetY, u32 SBW, u32 SPSM, GSTexture* dTex, u32 DBW, u32 DPSM)
 {
 	// match merge cb
-	struct Uniforms
+	struct alignas(16) Uniforms
 	{
 		float scale;
 		float pad1[3];
 		u32 SBW, DBW, SPSM;
+		u32 pad2;
 	};
 
-	const Uniforms cb = {sScale, {}, SBW, DBW, SPSM};
+	const Uniforms cb = {sScale, {}, SBW, DBW, SPSM, 0};
 	m_ctx->UpdateSubresource(m_merge.cb.get(), 0, nullptr, &cb, 0, 0);
 
 	const GSVector4 dRect(0, 0, dTex->GetWidth(), dTex->GetHeight());
@@ -1461,7 +1463,7 @@ void GSDevice11::ConvertToIndexedTexture(GSTexture* sTex, float sScale, u32 offs
 
 void GSDevice11::FilteredDownsampleTexture(GSTexture* sTex, GSTexture* dTex, u32 downsample_factor, const GSVector2i& clamp_min, const GSVector4& dRect)
 {
-	struct Uniforms
+	struct alignas(16) Uniforms
 	{
 		float weight;
 		float step_multiplier;
@@ -2063,13 +2065,13 @@ void GSDevice11::RenderImGui()
 
 	UpdateImGuiTextures();
 
-	const float L = 0.0f;
+	constexpr float L = 0.0f;
 	const float R = static_cast<float>(m_window_info.surface_width);
-	const float T = 0.0f;
+	constexpr float T = 0.0f;
 	const float B = static_cast<float>(m_window_info.surface_height);
 
 	// clang-format off
-  const float ortho_projection[4][4] =
+	const GSVector4 ortho_projection[4] =
 	{
 		{ 2.0f/(R-L),   0.0f,           0.0f,       0.0f },
 		{ 0.0f,         2.0f/(T-B),     0.0f,       0.0f },

pcsx2/GS/Renderers/DX12/GSDevice12.cpp

@@ -938,6 +938,7 @@ void GSDevice12::SetVSyncMode(GSVSyncMode mode, bool allow_present_throttle)
 
 	if (GetSwapChainBufferCount() != old_buffer_count)
 	{
+		ExecuteCommandList(true);
 		DestroySwapChain();
 		if (!CreateSwapChain())
 			pxFailRel("Failed to recreate swap chain after vsync change.");
@@ -1635,14 +1636,14 @@ void GSDevice12::UpdateCLUTTexture(
 	GSTexture* sTex, float sScale, u32 offsetX, u32 offsetY, GSTexture* dTex, u32 dOffset, u32 dSize)
 {
 	// match merge cb
-	struct Uniforms
+	struct alignas(16) Uniforms
 	{
 		float scale;
 		float pad1[3];
 		u32 offsetX, offsetY, dOffset;
 		u32 pad2;
 	};
-	const Uniforms cb = {sScale, {}, offsetX, offsetY, dOffset};
+	const Uniforms cb = {sScale, {}, offsetX, offsetY, dOffset, 0};
 	SetUtilityRootSignature();
 	SetUtilityPushConstants(&cb, sizeof(cb));
 
@@ -1656,14 +1657,15 @@ void GSDevice12::ConvertToIndexedTexture(
 	GSTexture* sTex, float sScale, u32 offsetX, u32 offsetY, u32 SBW, u32 SPSM, GSTexture* dTex, u32 DBW, u32 DPSM)
 {
 	// match merge cb
-	struct Uniforms
+	struct alignas(16) Uniforms
 	{
 		float scale;
 		float pad1[3];
 		u32 SBW, DBW, SPSM;
+		u32 pad2;
 	};
 
-	const Uniforms cb = {sScale, {}, SBW, DBW, SPSM};
+	const Uniforms cb = {sScale, {}, SBW, DBW, SPSM, 0};
 	SetUtilityRootSignature();
 	SetUtilityPushConstants(&cb, sizeof(cb));
 
@@ -1675,7 +1677,7 @@ void GSDevice12::ConvertToIndexedTexture(
 
 void GSDevice12::FilteredDownsampleTexture(GSTexture* sTex, GSTexture* dTex, u32 downsample_factor, const GSVector2i& clamp_min, const GSVector4& dRect)
 {
-	struct Uniforms
+	struct alignas(16) Uniforms
 	{
 		float weight;
 		float step_multiplier;
@@ -2184,7 +2186,7 @@ void GSDevice12::RenderImGui()
 	const float B = static_cast<float>(m_window_info.surface_height);
 
 	// clang-format off
-  const float ortho_projection[4][4] =
+	const GSVector4 ortho_projection[4] =
 	{
 		{ 2.0f/(R-L),   0.0f,           0.0f,       0.0f },
 		{ 0.0f,         2.0f/(T-B),     0.0f,       0.0f },

pcsx2/GS/Renderers/Vulkan/GSDeviceVK.cpp

@@ -3131,7 +3131,7 @@ void GSDeviceVK::UpdateCLUTTexture(
 	GSTexture* sTex, float sScale, u32 offsetX, u32 offsetY, GSTexture* dTex, u32 dOffset, u32 dSize)
 {
 	// Super annoying, but apparently NVIDIA doesn't like floats/ints packed together in the same vec4?
-	struct Uniforms
+	struct alignas(16) Uniforms
 	{
 		u32 offsetX, offsetY, dOffset, pad1;
 		float scale;
@@ -3150,7 +3150,7 @@ void GSDeviceVK::UpdateCLUTTexture(
 void GSDeviceVK::ConvertToIndexedTexture(
 	GSTexture* sTex, float sScale, u32 offsetX, u32 offsetY, u32 SBW, u32 SPSM, GSTexture* dTex, u32 DBW, u32 DPSM)
 {
-	struct Uniforms
+	struct alignas(16) Uniforms
 	{
 		u32 SBW;
 		u32 DBW;
@@ -3171,7 +3171,7 @@ void GSDeviceVK::ConvertToIndexedTexture(
 
 void GSDeviceVK::FilteredDownsampleTexture(GSTexture* sTex, GSTexture* dTex, u32 downsample_factor, const GSVector2i& clamp_min, const GSVector4& dRect)
 {
-	struct Uniforms
+	struct alignas(16) Uniforms
 	{
 		GSVector2i clamp_min;
 		int downsample_factor;
@@ -4434,16 +4434,13 @@ void GSDeviceVK::RenderImGui()
 
 	UpdateImGuiTextures();
 
-	const float uniforms[2][2] = {{
-									  2.0f / static_cast<float>(m_window_info.surface_width),
-									  2.0f / static_cast<float>(m_window_info.surface_height),
-								  },
-		{
-			-1.0f,
-			-1.0f,
-		}};
+	const GSVector4 uniforms(
+		2.0f / static_cast<float>(m_window_info.surface_width),
+		2.0f / static_cast<float>(m_window_info.surface_height),
+		-1.0f,
+		-1.0f);
 
-	SetUtilityPushConstants(uniforms, sizeof(uniforms));
+	SetUtilityPushConstants(&uniforms, sizeof(uniforms));
 	SetPipeline(m_imgui_pipeline);
 
 	if (m_utility_sampler != m_linear_sampler)

pcsx2/ImGui/FullscreenUI.cpp

@@ -4461,7 +4461,7 @@ void FullscreenUI::DrawBIOSSettingsPage()
 			});
 	}
 
-	MenuHeading(FSUI_CSTR("Options and Patches"));
+	MenuHeading(FSUI_CSTR("Fast Boot Options"));
 	DrawToggleSetting(bsi, FSUI_ICONSTR(ICON_FA_FORWARD_FAST, "Fast Boot"), FSUI_CSTR("Skips the intro screen, and bypasses region checks."),
 		"EmuCore", "EnableFastBoot", true);
 
@@ -9509,7 +9509,7 @@ TRANSLATE_NOOP("FullscreenUI", "Resets configuration to defaults (excluding cont
 TRANSLATE_NOOP("FullscreenUI", "BIOS Configuration");
 TRANSLATE_NOOP("FullscreenUI", "Changes the BIOS image used to start future sessions.");
 TRANSLATE_NOOP("FullscreenUI", "BIOS Selection");
-TRANSLATE_NOOP("FullscreenUI", "Options and Patches");
+TRANSLATE_NOOP("FullscreenUI", "Fast Boot Options");
 TRANSLATE_NOOP("FullscreenUI", "Skips the intro screen, and bypasses region checks.");
 TRANSLATE_NOOP("FullscreenUI", "Speed Control");
 TRANSLATE_NOOP("FullscreenUI", "Sets the speed when running without fast forwarding.");