GS: Add Pre-Round Sprite hack

This attempts to preproduce hardware behaviour, but falls down in a bunch of cases, hence the hack.

pcsx2-qt/Settings/GraphicsSettingsWidget.cpp

@@ -100,6 +100,7 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsWindow* dialog, QWidget*
 	SettingWidgetBinder::BindWidgetToIntSetting(
 		sif, m_ui.screenshotFormat, "EmuCore/GS", "ScreenshotFormat", static_cast<int>(GSScreenshotFormat::PNG));
 	SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.screenshotQuality, "EmuCore/GS", "ScreenshotQuality", 50);
+	SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.PreRoundSprites, "EmuCore/GS", "preround_sprites", false);
 	SettingWidgetBinder::BindWidgetToFloatSetting(sif, m_ui.stretchY, "EmuCore/GS", "StretchY", 100.0f);
 	SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.cropLeft, "EmuCore/GS", "CropLeft", 0);
 	SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.cropTop, "EmuCore/GS", "CropTop", 0);

pcsx2-qt/Settings/GraphicsSettingsWidget.ui

@@ -446,6 +446,16 @@
            </property>
           </widget>
          </item>
+         <item row="3" column="1">
+          <widget class="QCheckBox" name="PreRoundSprites">
+           <property name="toolTip">
+            <string>Attempts to pre-round sprite texel coordinates to resolve rounding issues. Helpful for games such as Beyond Good and Evil, and Manhunt</string>
+           </property>
+           <property name="text">
+            <string>Pre-Round Sprites</string>
+           </property>
+          </widget>
+         </item>
         </layout>
        </item>
       </layout>

pcsx2/Config.h

@@ -638,6 +638,7 @@ struct Pcsx2Config
 					PreloadFrameWithGSData : 1,
 					Mipmap : 1,
 					HWMipmap : 1,
+					PreRoundSprites : 1,
 					ManualUserHacks : 1,
 					UserHacks_AlignSpriteX : 1,
 					UserHacks_CPUFBConversion : 1,

pcsx2/GS/GSState.cpp

@@ -1683,7 +1683,96 @@ void GSState::FlushPrim()
 		// Sometimes hardware doesn't get affected, likely due to the difference in how GPU's handle textures (Persona minimap).
 		if (PRIM->TME && (GSUtil::GetPrimClass(PRIM->PRIM) == GS_PRIM_CLASS::GS_SPRITE_CLASS || m_vt.m_eq.z))
 		{
-			if (!PRIM->FST) // STQ's
+			if (PRIM->FST && GSConfig.PreRoundSprites) // UV's
+			{
+				// UV's on sprites only alter the final UV, I believe the problem is we're drawing too much (drawing stops earlier than we do)
+				// But this fixes Beyond Good adn Evil water and Dark Cloud 2's UI
+				if (GSUtil::GetPrimClass(PRIM->PRIM) == GS_PRIM_CLASS::GS_SPRITE_CLASS)
+				{
+					for (u32 i = 0; i < m_index.tail; i += 2)
+					{
+						GSVertex* v1 = &m_vertex.buff[m_index.buff[i]];
+						GSVertex* v2 = &m_vertex.buff[m_index.buff[i + 1]];
+
+						GSVertex* vu1;
+						GSVertex* vu2;
+						GSVertex* vv1;
+						GSVertex* vv2;
+						if (v1->U > v2->U)
+						{
+							vu2 = v1;
+							vu1 = v2;
+						}
+						else
+						{
+							vu2 = v2;
+							vu1 = v1;
+						}
+
+						if (v1->V > v2->V)
+						{
+							vv2 = v1;
+							vv1 = v2;
+						}
+						else
+						{
+							vv2 = v2;
+							vv1 = v1;
+						}
+
+						GSVector2 pos_range(vu2->XYZ.X - vu1->XYZ.X, vv2->XYZ.Y - vv1->XYZ.Y);
+						const GSVector2 uv_range(vu2->U - vu1->U, vv2->V - vv1->V);
+						if (pos_range.x == 0)
+							pos_range.x = 1;
+						if (pos_range.y == 0)
+							pos_range.y = 1;
+						const GSVector2 grad(uv_range / pos_range);
+						bool isclamp_w = m_context->CLAMP.WMS > 0 && m_context->CLAMP.WMS < 3;
+						bool isclamp_h = m_context->CLAMP.WMT > 0 && m_context->CLAMP.WMT < 3;
+						int max_w = (m_context->CLAMP.WMS == 2) ? m_context->CLAMP.MAXU : ((1 << m_context->TEX0.TW) - 1);
+						int max_h = (m_context->CLAMP.WMT == 2) ? m_context->CLAMP.MAXV : ((1 << m_context->TEX0.TH) - 1);
+						int width = vu2->U >> 4;
+						int height = vv2->V >> 4;
+
+						if (m_context->CLAMP.WMS == 3)
+							width = (width & m_context->CLAMP.MINU) | m_context->CLAMP.MAXU;
+
+						if (m_context->CLAMP.WMT == 3)
+							height = (height & m_context->CLAMP.MINV) | m_context->CLAMP.MAXV;
+
+						if ((isclamp_w && width <= max_w && grad.x != 1.0f) || !isclamp_w)
+						{
+							if (vu2->U >= 0x1 && (!(vu2->U & 0xf) || grad.x <= 1.0f))
+							{
+								vu2->U = (vu2->U - 1);
+								/*if (!isclamp_w && ((vu2->XYZ.X - m_context->XYOFFSET.OFX) >> 4) < m_context->scissor.in.z)
+									vu2->XYZ.X -= 1;*/
+							}
+						}
+						else
+							vu2->U &= ~0xf;
+
+						// Dark Cloud 2 tries to address wider than the TBW when in CLAMP mode (maybe some weird behaviour in HW?)
+						if ((vu2->U >> 4) > (int)(m_context->TEX0.TBW * 64) && isclamp_w && (vu2->U >> 4) - 8 <= (int)(m_context->TEX0.TBW * 64))
+						{
+							vu2->U = (m_context->TEX0.TBW * 64) << 4;
+						}
+
+						if ((isclamp_h && height <= max_h && grad.y != 1.0f) || !isclamp_h)
+						{
+							if (vv2->V >= 0x1 && (!(vv2->V & 0xf) || grad.y <= 1.0f))
+							{
+								vv2->V = (vv2->V - 1);
+								/*if (!isclamp_h && ((vv2->XYZ.Y - m_context->XYOFFSET.OFY) >> 4) < m_context->scissor.in.w)
+									vv2->XYZ.Y -= 1;*/
+							}
+						}
+						else
+							vv2->V &= ~0xf;
+					}
+				}
+			}
+			else if (!PRIM->FST) // STQ's
 			{
 				const bool is_sprite = GSUtil::GetPrimClass(PRIM->PRIM) == GS_PRIM_CLASS::GS_SPRITE_CLASS;
 				// ST's have the lowest 9 bits (or greater depending on exponent difference) rounding down (from hardware tests).
@@ -1736,7 +1825,6 @@ void GSState::FlushPrim()
 		if (unused > 0)
 		{
 			memcpy(m_vertex.buff, buff, sizeof(GSVertex) * unused);
-
 			m_vertex.tail = unused;
 			m_vertex.next = next > head ? next - head : 0;
 
@@ -3486,7 +3574,6 @@ __forceinline void GSState::VertexKick(u32 skip)
 	const GSVector4i new_v1(m_v.m[1]);
 
 	GSVector4i* RESTRICT tailptr = (GSVector4i*)&m_vertex.buff[tail];
-
 	tailptr[0] = new_v0;
 	tailptr[1] = new_v1;
 

pcsx2/GameDatabase.cpp

@@ -377,6 +377,7 @@ static const char* s_gs_hw_fix_names[] = {
 	"estimateTextureRegion",
 	"PCRTCOffsets",
 	"PCRTCOverscan",
+	"preRoundSprites",
 	"trilinearFiltering",
 	"skipDrawStart",
 	"skipDrawEnd",
@@ -423,6 +424,7 @@ bool GameDatabaseSchema::isUserHackHWFix(GSHWFixId id)
 		case GSHWFixId::Deinterlace:
 		case GSHWFixId::Mipmap:
 		case GSHWFixId::TexturePreloading:
+		case GSHWFixId::PreRoundSprites:
 		case GSHWFixId::TrilinearFiltering:
 		case GSHWFixId::MinimumBlendingLevel:
 		case GSHWFixId::MaximumBlendingLevel:
@@ -779,6 +781,10 @@ void GameDatabaseSchema::GameEntry::applyGSHardwareFixes(Pcsx2Config::GSOptions&
 				config.PCRTCOverscan = (value > 0);
 				break;
 
+			case GSHWFixId::PreRoundSprites:
+				config.PreRoundSprites = (value > 0);
+				break;
+
 			case GSHWFixId::Mipmap:
 				config.HWMipmap = (value > 0);
 				break;

pcsx2/GameDatabase.h

@@ -60,6 +60,7 @@ namespace GameDatabaseSchema
 		EstimateTextureRegion,
 		PCRTCOffsets,
 		PCRTCOverscan,
+		PreRoundSprites,
 
 		// integer settings
 		TrilinearFiltering,

pcsx2/Pcsx2Config.cpp

@@ -645,6 +645,7 @@ Pcsx2Config::GSOptions::GSOptions()
 	PreloadFrameWithGSData = false;
 	Mipmap = true;
 	HWMipmap = true;
+	PreRoundSprites = false;
 
 	ManualUserHacks = false;
 	UserHacks_AlignSpriteX = false;
@@ -847,6 +848,7 @@ void Pcsx2Config::GSOptions::LoadSave(SettingsWrapper& wrap)
 	SettingsWrapBitBool(OsdShowVideoCapture);
 	SettingsWrapBitBool(OsdShowInputRec);
 
+	SettingsWrapBitBoolEx(PreRoundSprites, "preround_sprites");
 	SettingsWrapBitBool(HWSpinGPUForReadbacks);
 	SettingsWrapBitBool(HWSpinCPUForReadbacks);
 	SettingsWrapBitBoolEx(GPUPaletteConversion, "paltex");