GS/DX/VK: Properly align uniform buffers.

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.

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/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

@@ -1635,14 +1635,14 @@ 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 +1656,15 @@ 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 +1676,7 @@ 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 +2185,7 @@ 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.");