GS: Add optimisations for common GIF transfer sequences

PCSX2 · Sep 26, 2023 · 53b930b · 53b930b
1 parent fef282f
commit 53b930b
Show file tree

Hide file tree

Showing 3 changed files with 537 additions and 7 deletions.
diff --git a/pcsx2/GS/GSRegs.h b/pcsx2/GS/GSRegs.h
@@ -87,6 +87,15 @@ enum GIF_REG_COMPLEX
 {
 	GIF_REG_STQRGBAXYZF2 = 0x00,
 	GIF_REG_STQRGBAXYZ2 = 0x01,
+	GIF_REG_UVRGBAXYZ2 = 0x02,
+	GIF_REG_RGBAXYZ2 = 0x03,
+	GIF_REG_RGBAXYZF2 = 0x04,
+	GIF_REG_STQXYZ2 = 0x05,
+	GIF_REG_STQXYZF2 = 0x06,
+	GIF_REG_UVXYZ2 = 0x07,
+	GIF_REG_UVXYZF2 = 0x08,
+	GIF_REG_RGBAUVXYZF2 = 0x09,
+	GIF_REG_UVRGBAXYZF2 = 0x0A
 };
 
 enum GIF_A_D_REG
@@ -1126,7 +1135,16 @@ struct alignas(32) GIFPath
 		TYPE_UNKNOWN,
 		TYPE_ADONLY,
 		TYPE_STQRGBAXYZF2,
-		TYPE_STQRGBAXYZ2
+		TYPE_STQRGBAXYZ2,
+		TYPE_UVRGBAXYZ2,
+		TYPE_RGBAXYZF2,
+		TYPE_UVXYZF2,
+		TYPE_STQXYZF2,
+		TYPE_RGBAXYZ2,
+		TYPE_UVXYZ2,
+		TYPE_STQXYZ2,
+		TYPE_RGBAUVXYZF2,
+		TYPE_UVRGBAXYZF2,
 	};
 
 	__forceinline void SetTag(const void* mem)
@@ -1156,7 +1174,7 @@ struct alignas(32) GIFPath
 
 		if (tag.FLG == GIF_FLG_PACKED)
 		{
-			if (regs.eq8(GSVector4i(0x0e0e0e0e)).mask() == (1 << nreg) - 1)
+			if (regs.eq8(GSVector4i::cxpr(0x0e0e0e0e)).mask() == (1 << nreg) - 1)
 			{
 				type = TYPE_ADONLY;
 			}
@@ -1167,21 +1185,69 @@ struct alignas(32) GIFPath
 					case 1:
 						break;
 					case 2:
+						if (regs.eq8(GSVector4i::cxpr(0x00000401)).mask() == (1 << 2) - 1)
+							type = TYPE_RGBAXYZF2;
+						else if (regs.eq8(GSVector4i::cxpr(0x0000040)).mask() == (1 << 2) - 1)
+							type = TYPE_STQXYZF2;
+						else if (regs.eq8(GSVector4i::cxpr(0x00000402)).mask() == (1 << 2) - 13)
+							type = TYPE_UVXYZF2;
+						else if (regs.eq8(GSVector4i::cxpr(0x00000501)).mask() == (1 << 2) - 1)
+							type = TYPE_RGBAXYZ2;
+						else if (regs.eq8(GSVector4i::cxpr(0x00000502)).mask() == (1 << 2) - 1)
+							type = TYPE_STQXYZ2;
+						else if (regs.eq8(GSVector4i::cxpr(0x00000503)).mask() == (1 << 2) - 1)
+							type = TYPE_UVXYZ2;
 						break;
 					case 3:
 						// many games, TODO: formats mixed with NOPs (xeno2: 040f010f02, 04010f020f, mgs3: 04010f0f02, 0401020f0f, 04010f020f)
-						if (regs.U32[0] == 0x00040102)
+						if (regs.eq8(GSVector4i::cxpr(0x00040102)).mask() == (1 << 3) - 1)
 							type = TYPE_STQRGBAXYZF2;
 						// GoW (has other crazy formats, like ...030503050103)
-						if (regs.U32[0] == 0x00050102)
+						else if (regs.eq8(GSVector4i::cxpr(0x00050102)).mask() == (1 << 3) - 1)
 							type = TYPE_STQRGBAXYZ2;
 						// TODO: common types with UV instead
+						else if (regs.eq8(GSVector4i::cxpr(0x00050103)).mask() == (1 << 3) - 1)
+							type = TYPE_UVRGBAXYZ2;
+						else if (regs.eq8(GSVector4i::cxpr(0x00040103)).mask() == (1 << 3) - 1)
+							type = TYPE_UVRGBAXYZF2;
+						else if (regs.eq8(GSVector4i::cxpr(0x00040301)).mask() == (1 << 3) - 1)
+							type = TYPE_RGBAUVXYZF2;
 						break;
 					case 4:
+						if (regs.eq8(GSVector4i(0x04030403)).mask() == (1 << 4) - 1)
+						{
+							type = TYPE_UVXYZF2;
+							nreg = 2;
+							nloop *= 2;
+						}
+						else if (regs.eq8(GSVector4i::cxpr(0x05030503)).mask() == (1 << 4) - 1)
+						{
+							type = TYPE_UVXYZ2;
+							nreg = 2;
+							nloop *= 2;
+						}
 						break;
 					case 5:
 						break;
 					case 6:
+						if (regs.U32[0] == 0x03040103 && regs.U32[1] == 0x00000401)
+						{
+							type = TYPE_UVRGBAXYZF2;
+							nreg = 3;
+							nloop *= 2;
+						}
+						else if (regs.U32[0] == 0x01040301 && regs.U32[1] == 0x00000403)
+						{
+							type = TYPE_RGBAUVXYZF2;
+							nreg = 3;
+							nloop *= 2;
+						}
+						else if (regs.U32[0] == 0x03050103 && regs.U32[1] == 0x00000501)
+						{
+							type = TYPE_UVRGBAXYZ2;
+							nreg = 3;
+							nloop *= 2;
+						}
 						break;
 					case 7:
 						break;