Merge pull request #2141 from Tilka/position_cache

zfreeze: cache vertex positions

10 files changed, 123 insertions, 36 deletions

diff --git a/Source/Core/Common/x64Emitter.cpp b/Source/Core/Common/x64Emitter.cpp
index c1b703b174..1c1a8b9754 100644
--- a/Source/Core/Common/x64Emitter.cpp
+++ b/Source/Core/Common/x64Emitter.cpp
@@ -1823,6 +1823,7 @@ void XEmitter::PCMPGTD(X64Reg dest, const OpArg& arg)  {WriteSSEOp(0x66, 0x66, d
 
 void XEmitter::PEXTRW(X64Reg dest, const OpArg& arg, u8 subreg) {WriteSSEOp(0x66, 0xC5, dest, arg); Write8(subreg);}
 void XEmitter::PINSRW(X64Reg dest, const OpArg& arg, u8 subreg) {WriteSSEOp(0x66, 0xC4, dest, arg); Write8(subreg);}
+void XEmitter::PINSRD(X64Reg dest, const OpArg& arg, u8 subreg) {WriteSSE41Op(0x66, 0x3A22, dest, arg); Write8(subreg);}
 
 void XEmitter::PMADDWD(X64Reg dest, const OpArg& arg)  {WriteSSEOp(0x66, 0xF5, dest, arg); }
 void XEmitter::PSADBW(X64Reg dest, const OpArg& arg)   {WriteSSEOp(0x66, 0xF6, dest, arg);}
diff --git a/Source/Core/Common/x64Emitter.h b/Source/Core/Common/x64Emitter.h
index 7b6c9be703..c62254aec4 100644
--- a/Source/Core/Common/x64Emitter.h
+++ b/Source/Core/Common/x64Emitter.h
@@ -711,6 +711,7 @@ public:
 
 	void PEXTRW(X64Reg dest, const OpArg& arg, u8 subreg);
 	void PINSRW(X64Reg dest, const OpArg& arg, u8 subreg);
+	void PINSRD(X64Reg dest, const OpArg& arg, u8 subreg);
 
 	void PMADDWD(X64Reg dest, const OpArg& arg);
 	void PSADBW(X64Reg dest, const OpArg& arg);
diff --git a/Source/Core/VideoCommon/VertexLoader.cpp b/Source/Core/VideoCommon/VertexLoader.cpp
index 72aad88d7f..b8c8fab4de 100644
--- a/Source/Core/VideoCommon/VertexLoader.cpp
+++ b/Source/Core/VideoCommon/VertexLoader.cpp
@@ -14,6 +14,7 @@
 #include "VideoCommon/VertexLoader_Normal.h"
 #include "VideoCommon/VertexLoader_Position.h"
 #include "VideoCommon/VertexLoader_TextCoord.h"
+#include "VideoCommon/VertexLoaderManager.h"
 #include "VideoCommon/VideoCommon.h"
 #include "VideoCommon/VideoConfig.h"
 
@@ -24,6 +25,8 @@ u8* g_vertex_manager_write_ptr;
 static void LOADERDECL PosMtx_ReadDirect_UByte(VertexLoader* loader)
 {
 	u32 posmtx = DataReadU8() & 0x3f;
+	if (loader->m_counter < 3)
+		VertexLoaderManager::position_matrix_index[loader->m_counter] = posmtx;
 	DataWrite<u32>(posmtx);
 	PRIM_LOG("posmtx: %d, ", posmtx);
 }
@@ -316,7 +319,7 @@ int VertexLoader::RunVertices(DataReader src, DataReader dst, int count)
 	m_numLoadedVertices += count;
 	m_skippedVertices = 0;
 
-	for (int s = 0; s < count; s++)
+	for (m_counter = count - 1; m_counter >= 0; m_counter--)
 	{
 		m_tcIndex = 0;
 		m_colIndex = 0;
diff --git a/Source/Core/VideoCommon/VertexLoader.h b/Source/Core/VideoCommon/VertexLoader.h
index b53532a592..50ece53095 100644
--- a/Source/Core/VideoCommon/VertexLoader.h
+++ b/Source/Core/VideoCommon/VertexLoader.h
@@ -49,6 +49,7 @@ public:
 	int m_texmtxread;
 	bool m_vertexSkip;
 	int m_skippedVertices;
+	int m_counter;
 
 private:
 	// Pipeline.
diff --git a/Source/Core/VideoCommon/VertexLoaderARM64.cpp b/Source/Core/VideoCommon/VertexLoaderARM64.cpp
index 36b6aae0f5..35d3a989e7 100644
--- a/Source/Core/VideoCommon/VertexLoaderARM64.cpp
+++ b/Source/Core/VideoCommon/VertexLoaderARM64.cpp
@@ -165,6 +165,18 @@ int VertexLoaderARM64::ReadVertex(u64 attribute, int format, int count_in, int c
 		m_float_emit.ST1(32, 1, coords, EncodeRegTo64(scratch2_reg));
 	}
 
+	// Z-Freeze
+	if (native_format == &m_native_vtx_decl.position)
+	{
+		CMP(count_reg, 3);
+		FixupBranch dont_store = B(CC_GT);
+		MOVI2R(EncodeRegTo64(scratch2_reg), (u64)VertexLoaderManager::position_cache);
+		ORR(scratch1_reg, WSP, count_reg, ArithOption(count_reg, ST_LSL, 4));
+		ADD(EncodeRegTo64(scratch1_reg), EncodeRegTo64(scratch1_reg), EncodeRegTo64(scratch2_reg));
+		m_float_emit.STUR(write_size, coords, EncodeRegTo64(scratch1_reg), -16);
+		SetJumpTarget(dont_store);
+	}
+
 	native_format->components = count_out;
 	native_format->enable = true;
 	native_format->offset = m_dst_ofs;
@@ -342,6 +354,14 @@ void VertexLoaderARM64::GenerateVertexLoader()
 		LDRB(INDEX_UNSIGNED, scratch1_reg, src_reg, m_src_ofs);
 		AND(scratch1_reg, scratch1_reg, 0, 5);
 		STR(INDEX_UNSIGNED, scratch1_reg, dst_reg, m_dst_ofs);
+
+		// Z-Freeze
+		CMP(count_reg, 3);
+		FixupBranch dont_store = B(CC_GT);
+		MOVI2R(EncodeRegTo64(scratch2_reg), (u64)VertexLoaderManager::position_matrix_index - sizeof(u32));
+		STR(INDEX_UNSIGNED, scratch1_reg, EncodeRegTo64(scratch2_reg), 0);
+		SetJumpTarget(dont_store);
+
 		m_native_components |= VB_HAS_POSMTXIDX;
 		m_native_vtx_decl.posmtx.components = 4;
 		m_native_vtx_decl.posmtx.enable = true;
diff --git a/Source/Core/VideoCommon/VertexLoaderManager.cpp b/Source/Core/VideoCommon/VertexLoaderManager.cpp
index dbb421991c..46a0eeaab0 100644
--- a/Source/Core/VideoCommon/VertexLoaderManager.cpp
+++ b/Source/Core/VideoCommon/VertexLoaderManager.cpp
@@ -26,6 +26,9 @@
 namespace VertexLoaderManager
 {
 
+float position_cache[3][4];
+u32 position_matrix_index[3];
+
 typedef std::unordered_map<PortableVertexDeclaration, std::unique_ptr<NativeVertexFormat>> NativeVertexFormatMap;
 static NativeVertexFormatMap s_native_vertex_map;
 static NativeVertexFormat* s_current_vtx_fmt;
diff --git a/Source/Core/VideoCommon/VertexLoaderManager.h b/Source/Core/VideoCommon/VertexLoaderManager.h
index 99336fc333..92d7e6da8f 100644
--- a/Source/Core/VideoCommon/VertexLoaderManager.h
+++ b/Source/Core/VideoCommon/VertexLoaderManager.h
@@ -28,5 +28,10 @@ namespace VertexLoaderManager
 	// Resolved pointers to array bases. Used by vertex loaders.
 	extern u8 *cached_arraybases[12];
 	void UpdateVertexArrayPointers();
+
+	// Position cache for zfreeze (3 vertices, 4 floats each to allow SIMD overwrite).
+	// These arrays are in reverse order.
+	extern float position_cache[3][4];
+	extern u32 position_matrix_index[3];
 }
 
diff --git a/Source/Core/VideoCommon/VertexLoaderX64.cpp b/Source/Core/VideoCommon/VertexLoaderX64.cpp
index 9abde62b8b..a298d7e1dd 100644
--- a/Source/Core/VideoCommon/VertexLoaderX64.cpp
+++ b/Source/Core/VideoCommon/VertexLoaderX64.cpp
@@ -23,6 +23,11 @@ static const X64Reg base_reg = RBX;
 
 static const u8* memory_base_ptr = (u8*)&g_main_cp_state.array_strides;
 
+static OpArg MPIC(const void* ptr, X64Reg scale_reg, int scale = SCALE_1)
+{
+	return MComplex(base_reg, scale_reg, scale, (s32)((u8*)ptr - memory_base_ptr));
+}
+
 static OpArg MPIC(const void* ptr)
 {
 	return MDisp(base_reg, (s32)((u8*)ptr - memory_base_ptr));
@@ -77,7 +82,7 @@ OpArg VertexLoaderX64::GetVertexAddr(int array, u64 attribute)
 
 int VertexLoaderX64::ReadVertex(OpArg data, u64 attribute, int format, int count_in, int count_out, bool dequantize, u8 scaling_exponent, AttributeFormat* native_format)
 {
-	static const __m128i shuffle_lut[4][3] = {
+	static const __m128i shuffle_lut[5][3] = {
 		{_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFF00L),  // 1x u8
 		 _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFF01L, 0xFFFFFF00L),  // 2x u8
 		 _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFF02L, 0xFFFFFF01L, 0xFFFFFF00L)}, // 3x u8
@@ -90,6 +95,9 @@ int VertexLoaderX64::ReadVertex(OpArg data, u64 attribute, int format, int count
 		{_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0x0001FFFFL),  // 1x s16
 		 _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0x0203FFFFL, 0x0001FFFFL),  // 2x s16
 		 _mm_set_epi32(0xFFFFFFFFL, 0x0405FFFFL, 0x0203FFFFL, 0x0001FFFFL)}, // 3x s16
+		{_mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0x00010203L),  // 1x float
+		 _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0x04050607L, 0x00010203L),  // 2x float
+		 _mm_set_epi32(0xFFFFFFFFL, 0x08090A0BL, 0x04050607L, 0x00010203L)}, // 3x float
 	};
 	static const __m128 scale_factors[32] = {
 		_mm_set_ps1(1./(1u<< 0)), _mm_set_ps1(1./(1u<< 1)), _mm_set_ps1(1./(1u<< 2)), _mm_set_ps1(1./(1u<< 3)),
@@ -119,21 +127,6 @@ int VertexLoaderX64::ReadVertex(OpArg data, u64 attribute, int format, int count
 	if (attribute == DIRECT)
 		m_src_ofs += load_bytes;
 
-	if (format == FORMAT_FLOAT)
-	{
-		// Floats don't need to be scaled or converted,
-		// so we can just load/swap/store them directly
-		// and return early.
-		for (int i = 0; i < count_in; i++)
-		{
-			LoadAndSwap(32, scratch3, data);
-			MOV(32, dest, R(scratch3));
-			data.AddMemOffset(sizeof(float));
-			dest.AddMemOffset(sizeof(float));
-		}
-		return load_bytes;
-	}
-
 	if (cpu_info.bSSSE3)
 	{
 		if (load_bytes > 8)
@@ -194,13 +187,54 @@ int VertexLoaderX64::ReadVertex(OpArg data, u64 attribute, int format, int count
 			else
 				PSRLD(coords, 16);
 			break;
+		case FORMAT_FLOAT:
+			// Floats don't need to be scaled or converted,
+			// so we can just load/swap/store them directly
+			// and return early.
+			// (In SSSE3 we still need to store them.)
+			for (int i = 0; i < count_in; i++)
+			{
+				LoadAndSwap(32, scratch3, data);
+				MOV(32, dest, R(scratch3));
+				data.AddMemOffset(sizeof(float));
+				dest.AddMemOffset(sizeof(float));
+
+				// zfreeze
+				if (native_format == &m_native_vtx_decl.position)
+				{
+					if (cpu_info.bSSE4_1)
+					{
+						PINSRD(coords, R(scratch3), i);
+					}
+					else
+					{
+						PINSRW(coords, R(scratch3), 2 * i + 0);
+						SHR(32, R(scratch3), Imm8(16));
+						PINSRW(coords, R(scratch3), 2 * i + 1);
+					}
+				}
+			}
+
+			// zfreeze
+			if (native_format == &m_native_vtx_decl.position)
+			{
+				CMP(32, R(count_reg), Imm8(3));
+				FixupBranch dont_store = J_CC(CC_A);
+				LEA(32, scratch3, MScaled(count_reg, SCALE_4, -4));
+				MOVUPS(MPIC(VertexLoaderManager::position_cache, scratch3, SCALE_4), coords);
+				SetJumpTarget(dont_store);
+			}
+			return load_bytes;
 		}
 	}
 
-	CVTDQ2PS(coords, R(coords));
+	if (format != FORMAT_FLOAT)
+	{
+		CVTDQ2PS(coords, R(coords));
 
-	if (dequantize && scaling_exponent)
-		MULPS(coords, MPIC(&scale_factors[scaling_exponent]));
+		if (dequantize && scaling_exponent)
+			MULPS(coords, MPIC(&scale_factors[scaling_exponent]));
+	}
 
 	switch (count_out)
 	{
@@ -209,6 +243,16 @@ int VertexLoaderX64::ReadVertex(OpArg data, u64 attribute, int format, int count
 	case 3: MOVUPS(dest, coords); break;
 	}
 
+	// zfreeze
+	if (native_format == &m_native_vtx_decl.position)
+	{
+		CMP(32, R(count_reg), Imm8(3));
+		FixupBranch dont_store = J_CC(CC_A);
+		LEA(32, scratch3, MScaled(count_reg, SCALE_4, -4));
+		MOVUPS(MPIC(VertexLoaderManager::position_cache, scratch3, SCALE_4), coords);
+		SetJumpTarget(dont_store);
+	}
+
 	return load_bytes;
 }
 
@@ -384,6 +428,13 @@ void VertexLoaderX64::GenerateVertexLoader()
 		MOVZX(32, 8, scratch1, MDisp(src_reg, m_src_ofs));
 		AND(32, R(scratch1), Imm8(0x3F));
 		MOV(32, MDisp(dst_reg, m_dst_ofs), R(scratch1));
+
+		// zfreeze
+		CMP(32, R(count_reg), Imm8(3));
+		FixupBranch dont_store = J_CC(CC_A);
+		MOV(32, MPIC(VertexLoaderManager::position_matrix_index - 1, count_reg, SCALE_4), R(scratch1));
+		SetJumpTarget(dont_store);
+
 		m_native_components |= VB_HAS_POSMTXIDX;
 		m_native_vtx_decl.posmtx.components = 4;
 		m_native_vtx_decl.posmtx.enable = true;
diff --git a/Source/Core/VideoCommon/VertexLoader_Position.cpp b/Source/Core/VideoCommon/VertexLoader_Position.cpp
index ed6b4587b3..fd17fbf4d7 100644
--- a/Source/Core/VideoCommon/VertexLoader_Position.cpp
+++ b/Source/Core/VideoCommon/VertexLoader_Position.cpp
@@ -32,7 +32,12 @@ void LOADERDECL Pos_ReadDirect(VertexLoader* loader)
 	DataReader src(g_video_buffer_read_ptr, nullptr);
 
 	for (int i = 0; i < N; ++i)
-		dst.Write(PosScale(src.Read<T>(), scale));
+	{
+		float value = PosScale(src.Read<T>(), scale);
+		if (loader->m_counter < 3)
+			VertexLoaderManager::position_cache[loader->m_counter][i] = value;
+		dst.Write(value);
+	}
 
 	g_vertex_manager_write_ptr = dst.GetPointer();
 	g_video_buffer_read_ptr = src.GetPointer();
@@ -52,7 +57,12 @@ void LOADERDECL Pos_ReadIndex(VertexLoader* loader)
 	DataReader dst(g_vertex_manager_write_ptr, nullptr);
 
 	for (int i = 0; i < N; ++i)
-		dst.Write(PosScale(Common::FromBigEndian(data[i]), scale));
+	{
+		float value = PosScale(Common::FromBigEndian(data[i]), scale);
+		if (loader->m_counter < 3)
+			VertexLoaderManager::position_cache[loader->m_counter][i] = value;
+		dst.Write(value);
+	}
 
 	g_vertex_manager_write_ptr = dst.GetPointer();
 	LOG_VTX();
diff --git a/Source/Core/VideoCommon/VertexManagerBase.cpp b/Source/Core/VideoCommon/VertexManagerBase.cpp
index dcef793cec..e9e8ebf901 100644
--- a/Source/Core/VideoCommon/VertexManagerBase.cpp
+++ b/Source/Core/VideoCommon/VertexManagerBase.cpp
@@ -279,7 +279,6 @@ void VertexManager::DoState(PointerWrap& p)
 
 void VertexManager::CalculateZSlope(NativeVertexFormat* format)
 {
-	float vtx[9];
 	float out[12];
 	float viewOffset[2] = { xfmem.viewport.xOrig - bpmem.scissorOffset.x * 2,
 	                        xfmem.viewport.yOrig - bpmem.scissorOffset.y * 2};
@@ -290,31 +289,24 @@ void VertexManager::CalculateZSlope(NativeVertexFormat* format)
 	// Global matrix ID.
 	u32 mtxIdx = g_main_cp_state.matrix_index_a.PosNormalMtxIdx;
 	const PortableVertexDeclaration vert_decl = format->GetVertexDeclaration();
-	size_t posOff = vert_decl.position.offset;
-	size_t mtxOff = vert_decl.posmtx.offset;
 
 	// Make sure the buffer contains at least 3 vertices.
 	if ((s_pCurBufferPointer - s_pBaseBufferPointer) < (vert_decl.stride * 3))
 		return;
 
 	// Lookup vertices of the last rendered triangle and software-transform them
-	// This allows us to determine the depth slope, which will be used if z--freeze
+	// This allows us to determine the depth slope, which will be used if z-freeze
 	// is enabled in the following flush.
 	for (unsigned int i = 0; i < 3; ++i)
 	{
-		u8* vtx_ptr = s_pCurBufferPointer - vert_decl.stride * (3 - i);
-		vtx[0 + i * 3] = ((float*)(vtx_ptr + posOff))[0];
-		vtx[1 + i * 3] = ((float*)(vtx_ptr + posOff))[1];
-		if (vert_decl.position.components == 3)
-			vtx[2 + i * 3] = ((float*)(vtx_ptr + posOff))[2];
-		else
-			vtx[2 + i * 3] = 0;
-
 		// If this vertex format has per-vertex position matrix IDs, look it up.
 		if (vert_decl.posmtx.enable)
-			mtxIdx = *((u32*)(vtx_ptr + mtxOff));
+			mtxIdx = VertexLoaderManager::position_matrix_index[2 - i];
+
+		if (vert_decl.position.components == 2)
+			VertexLoaderManager::position_cache[2 - i][2] = 0;
 
-		VertexShaderManager::TransformToClipSpace(&vtx[i * 3], &out[i * 4], mtxIdx);
+		VertexShaderManager::TransformToClipSpace(&VertexLoaderManager::position_cache[2 - i][0], &out[i * 4], mtxIdx);
 
 		// Transform to Screenspace
 		float inv_w = 1.0f / out[3 + i * 4];