From 127433628df436853a76c12b4fc76858e734b4ed Mon Sep 17 00:00:00 2001
From: TellowKrinkle <tellowkrinkle@gmail.com>
Date: Tue, 6 Apr 2021 23:15:40 -0500
Subject: [PATCH] GS: AVX2 ReadAndExpandBlock4(HH|HL)_32

---
 pcsx2/GS/GSBlock.h | 75 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 75 insertions(+)

diff --git a/pcsx2/GS/GSBlock.h b/pcsx2/GS/GSBlock.h
index b1f1786574..94adafee73 100644
--- a/pcsx2/GS/GSBlock.h
+++ b/pcsx2/GS/GSBlock.h
@@ -2054,6 +2054,42 @@ public:
 	{
 		//printf("ReadAndExpandBlock4HL_32\n");
 
+#if _M_SSE >= 0x501
+
+		const GSVector8i* s = (const GSVector8i*)src;
+
+		GSVector8i p0, p1, p2, p3;
+		LoadPalVecs(pal, p0, p1, p2, p3);
+		GSVector8i mask(0x0f0f0f0f);
+
+		GSVector8i v0, v1, v2, v3;
+
+		for (int i = 0; i < 2; i++)
+		{
+			GSVector8i* d0 = reinterpret_cast<GSVector8i*>(dst);
+			GSVector8i* d1 = reinterpret_cast<GSVector8i*>(dst + dstpitch);
+			GSVector8i* d2 = reinterpret_cast<GSVector8i*>(dst + dstpitch * 2);
+			GSVector8i* d3 = reinterpret_cast<GSVector8i*>(dst + dstpitch * 3);
+
+			v0 = s[i * 4 + 0] >> 24;
+			v1 = s[i * 4 + 1] >> 24;
+			v2 = s[i * 4 + 2] >> 24;
+			v3 = s[i * 4 + 3] >> 24;
+
+			GSVector8i::sw128(v0, v1);
+			GSVector8i::sw64(v0, v1);
+			GSVector8i::sw128(v2, v3);
+			GSVector8i::sw64(v2, v3);
+
+			GSVector8i all = v0.ps32(v1).pu16(v2.ps32(v3)) & mask;
+
+			ReadClut4(p0, p1, p2, p3, all, *d0, *d1, *d2, *d3);
+
+			dst += dstpitch * 4;
+		}
+
+#else
+
 		const GSVector4i* s = (const GSVector4i*)src;
 
 		GSVector4i p0, p1, p2, p3;
@@ -2080,6 +2116,8 @@ public:
 
 			dst += dstpitch * 2;
 		}
+
+#endif
 	}
 
 	// TODO: ReadAndExpandBlock4HL_16
@@ -2088,6 +2126,41 @@ public:
 	{
 		//printf("ReadAndExpandBlock4HH_32\n");
 
+#if _M_SSE >= 0x501
+
+		const GSVector8i* s = (const GSVector8i*)src;
+
+		GSVector8i p0, p1, p2, p3;
+		LoadPalVecs(pal, p0, p1, p2, p3);
+
+		GSVector8i v0, v1, v2, v3;
+
+		for (int i = 0; i < 2; i++)
+		{
+			GSVector8i* d0 = reinterpret_cast<GSVector8i*>(dst);
+			GSVector8i* d1 = reinterpret_cast<GSVector8i*>(dst + dstpitch);
+			GSVector8i* d2 = reinterpret_cast<GSVector8i*>(dst + dstpitch * 2);
+			GSVector8i* d3 = reinterpret_cast<GSVector8i*>(dst + dstpitch * 3);
+
+			v0 = s[i * 4 + 0] >> 28;
+			v1 = s[i * 4 + 1] >> 28;
+			v2 = s[i * 4 + 2] >> 28;
+			v3 = s[i * 4 + 3] >> 28;
+
+			GSVector8i::sw128(v0, v1);
+			GSVector8i::sw64(v0, v1);
+			GSVector8i::sw128(v2, v3);
+			GSVector8i::sw64(v2, v3);
+
+			GSVector8i all = v0.ps32(v1).pu16(v2.ps32(v3));
+
+			ReadClut4(p0, p1, p2, p3, all, *d0, *d1, *d2, *d3);
+
+			dst += dstpitch * 4;
+		}
+
+#else
+
 		const GSVector4i* s = (const GSVector4i*)src;
 
 		GSVector4i p0, p1, p2, p3;
@@ -2113,6 +2186,8 @@ public:
 
 			dst += dstpitch * 2;
 		}
+
+#endif
 	}
 
 	// TODO: ReadAndExpandBlock4HH_16