From bdc7dc2cd80bb8ac7d15d66ace4a1caa3b39bfa8 Mon Sep 17 00:00:00 2001
From: TellowKrinkle <tellowkrinkle@gmail.com>
Date: Tue, 6 Apr 2021 01:17:40 -0500
Subject: [PATCH] GS: ReadColumn8 AVX2 path

---
 pcsx2/GS/GSBlock.cpp  |  3 +++
 pcsx2/GS/GSBlock.h    | 33 ++++++++++++++++++++-------------
 pcsx2/GS/GSVector8i.h |  6 ++++++
 3 files changed, 29 insertions(+), 13 deletions(-)

diff --git a/pcsx2/GS/GSBlock.cpp b/pcsx2/GS/GSBlock.cpp
index ccb25fd7fc..8d15138455 100644
--- a/pcsx2/GS/GSBlock.cpp
+++ b/pcsx2/GS/GSBlock.cpp
@@ -20,6 +20,9 @@ CONSTINIT const GSVector4i GSBlock::m_r16mask(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6,
 CONSTINIT const GSVector4i GSBlock::m_r8mask(0, 4, 2, 6, 8, 12, 10, 14, 1, 5, 3, 7, 9, 13, 11, 15);
 CONSTINIT const GSVector4i GSBlock::m_r4mask(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
 
+CONSTINIT const GSVector4i GSBlock::m_avx2_r8mask1(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15);
+CONSTINIT const GSVector4i GSBlock::m_avx2_r8mask2(1, 5, 9, 13, 0, 4, 8, 12, 3, 7, 11, 15, 2, 6, 10, 14);
+
 CONSTINIT const GSVector4i GSBlock::m_uw8hmask0(0, 0, 0, 0, 1, 1, 1, 1, 8, 8, 8, 8, 9, 9, 9, 9);
 CONSTINIT const GSVector4i GSBlock::m_uw8hmask1(2, 2, 2, 2, 3, 3, 3, 3, 10, 10, 10, 10, 11, 11, 11, 11);
 CONSTINIT const GSVector4i GSBlock::m_uw8hmask2(4, 4, 4, 4, 5, 5, 5, 5, 12, 12, 12, 12, 13, 13, 13, 13);
diff --git a/pcsx2/GS/GSBlock.h b/pcsx2/GS/GSBlock.h
index 3d3f7aa8a6..d9457e6d87 100644
--- a/pcsx2/GS/GSBlock.h
+++ b/pcsx2/GS/GSBlock.h
@@ -25,6 +25,9 @@ class GSBlock
 	static const GSVector4i m_r8mask;
 	static const GSVector4i m_r4mask;
 
+	static const GSVector4i m_avx2_r8mask1;
+	static const GSVector4i m_avx2_r8mask2;
+
 	static const GSVector4i m_uw8hmask0;
 	static const GSVector4i m_uw8hmask1;
 	static const GSVector4i m_uw8hmask2;
@@ -512,30 +515,34 @@ public:
 
 		//for(int j = 0; j < 64; j++) ((u8*)src)[j] = (u8)j;
 
-#if 0 //_M_SSE >= 0x501
+#if _M_SSE >= 0x501
 
 		const GSVector8i* s = (const GSVector8i*)src;
 
-		GSVector8i v0 = s[i * 2 + 0];
-		GSVector8i v1 = s[i * 2 + 1];
+		GSVector8i v0, v1;
 
-		GSVector8i::sw8(v0, v1);
-		GSVector8i::sw16(v0, v1);
-		GSVector8i::sw8(v0, v1);
-		GSVector8i::sw128(v0, v1);
-		GSVector8i::sw16(v0, v1);
+		if ((i & 1) == 0)
+		{
+			v0 = s[i * 2 + 0];
+			v1 = s[i * 2 + 1];
+		}
+		else
+		{
+			v1 = s[i * 2 + 0];
+			v0 = s[i * 2 + 1];
+		}
 
-		v0 = v0.acbd();
-		v1 = v1.acbd();
-		v1 = v1.yxwz();
+		GSVector8i v2 = v0.acbd().shuffle8(GSVector8i::broadcast128(m_avx2_r8mask1));
+		GSVector8i v3 = v1.acbd().shuffle8(GSVector8i::broadcast128(m_avx2_r8mask2));
+
+		v0 = v2.blend32<0xaa>(v3);
+		v1 = v3.blend32<0xaa>(v2);
 
 		GSVector8i::storel(&dst[dstpitch * 0], v0);
 		GSVector8i::storeh(&dst[dstpitch * 1], v0);
 		GSVector8i::storel(&dst[dstpitch * 2], v1);
 		GSVector8i::storeh(&dst[dstpitch * 3], v1);
 
-		// TODO: not sure if this is worth it, not in this form, there should be a shorter path
-
 #else
 
 		const GSVector4i* s = (const GSVector4i*)src;
diff --git a/pcsx2/GS/GSVector8i.h b/pcsx2/GS/GSVector8i.h
index 9044240d23..ec1693482c 100644
--- a/pcsx2/GS/GSVector8i.h
+++ b/pcsx2/GS/GSVector8i.h
@@ -316,6 +316,12 @@ public:
 		return GSVector8i(_mm256_blend_epi16(m, a, mask));
 	}
 
+	template <int mask>
+	__forceinline GSVector8i blend32(const GSVector8i& a) const
+	{
+		return GSVector8i(_mm256_blend_epi32(m, a, mask));
+	}
+
 	__forceinline GSVector8i blend(const GSVector8i& a, const GSVector8i& mask) const
 	{
 		return GSVector8i(_mm256_or_si256(_mm256_andnot_si256(mask, m), _mm256_and_si256(mask, a)));