diff --git a/Utilities/Array.h b/Utilities/Array.h
index 02c80417c8..0f7df002de 100644
--- a/Utilities/Array.h
+++ b/Utilities/Array.h
@@ -141,7 +141,7 @@ public:
 	virtual void SetCount(const u32 count, bool memzero = true)
 	{
 		if(m_count >= count) return;
-		
+
 		_InsertRoomEnd(count - m_count);
 
 		if(memzero) memset(m_array + m_count - count, 0, sizeof(T) * (m_count - count));
diff --git a/rpcs3/Emu/Cell/PPCDecoder.h b/rpcs3/Emu/Cell/PPCDecoder.h
index 053a30c2d1..81e24087dd 100644
--- a/rpcs3/Emu/Cell/PPCDecoder.h
+++ b/rpcs3/Emu/Cell/PPCDecoder.h
@@ -381,7 +381,11 @@ public:
 		, m_parent(nullptr)
 		, m_opcode(-1)
 	{
-		memset(m_instrs, 0, sizeof(InstrCaller<TO>*) * count);
+		for(int i=0; i<count; ++i)
+		{
+			m_instrs[i] = error_func;
+		}
+
 		memset(m_instrs_info, 0, sizeof(InstrBase<TO>*) * count);
 	}
 
@@ -413,6 +417,14 @@ public:
 
 	void set_error_func(InstrCaller<TO>* error_func)
 	{
+		for(int i=0; i<count; ++i)
+		{
+			if(m_instrs[i] == m_error_func || !m_instrs[i])
+			{
+				m_instrs[i] = error_func;
+			}
+		}
+
 		m_error_func = error_func;
 	}
 
@@ -442,16 +454,7 @@ public:
 
 	void decode(TO* op, u32 entry, u32 code) const
 	{
-		InstrCaller<TO>* instr = m_instrs[entry];
-
-		if(instr)
-		{
-			(*instr)(op, code);
-		}
-		else if(m_error_func)
-		{
-			(*m_error_func)(op, code);
-		}
+		(*m_instrs[entry])(op, code);
 	}
 
 	virtual void operator ()(TO* op, u32 code) const
diff --git a/rpcs3/Emu/Cell/PPUInterpreter.h b/rpcs3/Emu/Cell/PPUInterpreter.h
index 86d012636c..9125b4a43a 100644
--- a/rpcs3/Emu/Cell/PPUInterpreter.h
+++ b/rpcs3/Emu/Cell/PPUInterpreter.h
@@ -3564,7 +3564,7 @@ private:
 	}
 	void FNEG(u32 frd, u32 frb, bool rc)
 	{
-		CPU.FPR[frd] = -CPU.FPR[frb];
+		CPU.FPR[frd] = ((u64&)CPU.FPR[frb]) ^ (1ULL << 63);
 		if(rc) UNK("fneg.");//CPU.UpdateCR1(CPU.FPR[frd]);
 	}
 	void FMR(u32 frd, u32 frb, bool rc)
diff --git a/rpcs3/Emu/GS/GL/FragmentProgram.cpp b/rpcs3/Emu/GS/GL/FragmentProgram.cpp
index a2678419ff..5ae2e41e9a 100644
--- a/rpcs3/Emu/GS/GL/FragmentProgram.cpp
+++ b/rpcs3/Emu/GS/GL/FragmentProgram.cpp
@@ -3,10 +3,38 @@
 
 void FragmentDecompilerThread::AddCode(wxString code)
 {
-	if(!src0.exec_if_eq && !src0.exec_if_gr && !src0.exec_if_le) return;
-	if(!src0.exec_if_eq || !src0.exec_if_gr || !src0.exec_if_le)
+	if(!src0.exec_if_eq && !src0.exec_if_gr && !src0.exec_if_lt) return;
+
+	wxString cond;
+
+	if(src0.exec_if_gr)
 	{
-		ConLog.Error("Bad cond! eq: %d  gr: %d  le: %d", src0.exec_if_eq, src0.exec_if_gr, src0.exec_if_le);
+		cond = ">";
+	}
+	else if(src0.exec_if_lt)
+	{
+		cond = "<";
+	}
+	else if(src0.exec_if_eq)
+	{
+		cond = "=";
+	}
+
+	if(src0.exec_if_eq)
+	{
+		cond += "=";
+	}
+	else
+	{
+		if(src0.exec_if_gr && src0.exec_if_lt)
+		{
+			cond = "!=";
+		}
+	}
+
+	if(cond)
+	{
+		ConLog.Error("cond! [eq: %d  gr: %d  lt: %d] (%s)", src0.exec_if_eq, src0.exec_if_gr, src0.exec_if_lt, cond);
 		Emu.Pause();
 		return;
 	}
@@ -29,7 +57,7 @@ void FragmentDecompilerThread::AddCode(wxString code)
 		}
 	}
 
-	code = AddReg(dst.dest_reg) + GetMask() + " = " + code + GetMask();
+	code = AddReg(dst.dest_reg, dst.fp16) + GetMask() + " = " + code + GetMask();
 
 	main += "\t" + code + ";\n";
 }
@@ -46,20 +74,29 @@ wxString FragmentDecompilerThread::GetMask()
 	return ret.IsEmpty() || ret == "xyzw" ? wxEmptyString : ("." + ret);
 }
 
-wxString FragmentDecompilerThread::AddReg(u32 index)
+wxString FragmentDecompilerThread::AddReg(u32 index, int fp16)
 {
 	//if(!index) return "gl_FragColor";
-	return m_parr.AddParam(index ? PARAM_NONE : PARAM_OUT, "vec4", wxString::Format("r%d", index), index ? -1 : 0);
+	return m_parr.AddParam(index ? PARAM_NONE : PARAM_OUT, "vec4",
+		wxString::Format((fp16 ? "h%d" : "r%d"), index), (index || fp16) ? -1 : 0);
 }
 
 wxString FragmentDecompilerThread::AddConst()
 {
-	return m_parr.AddParam(PARAM_CONST, "vec4", wxString::Format("fc%d", m_const_index++));
+	mem32_t data(m_addr + m_size + m_offset);
+
+	m_offset += 4 * 4;
+	u32 x = GetData(data[0]);
+	u32 y = GetData(data[1]);
+	u32 z = GetData(data[2]);
+	u32 w = GetData(data[3]);
+	return m_parr.AddParam("vec4", wxString::Format("fc%d", m_const_index++),
+		wxString::Format("vec4(%f, %f, %f, %f)", (float&)x, (float&)y, (float&)z, (float&)w));
 }
 
 wxString FragmentDecompilerThread::AddTex()
 {
-	return m_parr.AddParam(PARAM_CONST, "sampler2D", wxString::Format("tex_%d", dst.tex_num));
+	return m_parr.AddParam(PARAM_UNIFORM, "sampler2D", wxString::Format("tex%d", dst.tex_num));
 }
 
 template<typename T> wxString FragmentDecompilerThread::GetSRC(T src)
@@ -69,7 +106,7 @@ template<typename T> wxString FragmentDecompilerThread::GetSRC(T src)
 	switch(src.reg_type)
 	{
 	case 0: //tmp
-		ret += AddReg(src.tmp_reg_index);
+		ret += AddReg(src.tmp_reg_index, src.fp16);
 	break;
 
 	case 1: //input
@@ -102,7 +139,6 @@ template<typename T> wxString FragmentDecompilerThread::GetSRC(T src)
 	break;
 
 	case 2: //const
-		ConLog.Write("reg index = %d", src.tmp_reg_index);
 		ret += AddConst();
 	break;
 
@@ -157,6 +193,8 @@ void FragmentDecompilerThread::Task()
 		src1.HEX = GetData(data[2]);
 		src2.HEX = GetData(data[3]);
 
+		m_offset = 4 * 4;
+
 		switch(dst.opcode)
 		{
 		case 0x00: break; //NOP
@@ -166,24 +204,24 @@ void FragmentDecompilerThread::Task()
 		case 0x04: AddCode("(" + GetSRC(src0) + " * " + GetSRC(src1) + " + " + GetSRC(src2) + ")"); break; //MAD
 		case 0x05: AddCode("dot(" + GetSRC(src0) + ".xyz, " + GetSRC(src1) + ".xyz)"); break; // DP3
 		case 0x06: AddCode("dot(" + GetSRC(src0) + ", " + GetSRC(src1) + ")"); break; // DP4
-		//case 0x07: break; // DST
+		case 0x07: AddCode("distance(" + GetSRC(src0) + ", " + GetSRC(src1) + ")"); break; // DST
 		case 0x08: AddCode("min(" + GetSRC(src0) + ", " + GetSRC(src1) + ")"); break; // MIN
 		case 0x09: AddCode("max(" + GetSRC(src0) + ", " + GetSRC(src1) + ")"); break; // MAX
-		//case 0x0a: break; // SLT
-		//case 0x0b: break; // SGE
-		//case 0x0c: break; // SLE
-		//case 0x0d: break; // SGT
-		//case 0x0e: break; // SNE
-		//case 0x0f: break; // SEQ
+		case 0x0a: AddCode("vec4(lessThan(" + GetSRC(src0) + ", " + GetSRC(src1) + "))"); break; // SLT
+		case 0x0b: AddCode("vec4(greaterThanEqual(" + GetSRC(src0) + ", " + GetSRC(src1) + "))"); break; // SGE
+		case 0x0c: AddCode("vec4(lessThanEqual(" + GetSRC(src0) + ", " + GetSRC(src1) + "))"); break; // SLE
+		case 0x0d: AddCode("vec4(greaterThan(" + GetSRC(src0) + ", " + GetSRC(src1) + "))"); break; // SGT
+		case 0x0e: AddCode("vec4(notEqual(" + GetSRC(src0) + ", " + GetSRC(src1) + "))"); break; // SNE
+		case 0x0f: AddCode("vec4(equal(" + GetSRC(src0) + ", " + GetSRC(src1) + "))"); break; // SEQ
 
-		//case 0x10: break; // FRC
-		//case 0x11: break; // FLR
+		case 0x10: AddCode("fract(" + GetSRC(src0) + ")"); break; // FRC
+		case 0x11: AddCode("floor(" + GetSRC(src0) + ")"); break; // FLR
 		//case 0x12: break; // KIL
 		//case 0x13: break; // PK4
 		//case 0x14: break; // UP4
 		case 0x15: AddCode("ddx(" + GetSRC(src0) + ")"); break; // DDX
 		case 0x16: AddCode("ddy(" + GetSRC(src0) + ")"); break; // DDY
-		case 0x17: AddCode("texture(" + AddTex() + ", (" + GetSRC(src0) + ").xy)"); break; //TEX
+		case 0x17: AddCode("texture(" + AddTex() + ", " + GetSRC(src0) + ".xy)"); break; //TEX
 		//case 0x18: break; // TXP
 		//case 0x19: break; // TXD
 		case 0x1a: AddCode("1 / (" + GetSRC(src0) + ")"); break; // RCP
@@ -199,7 +237,7 @@ void FragmentDecompilerThread::Task()
 		case 0x23: AddCode("sin(" + GetSRC(src0) + ")"); break; // SIN
 		//case 0x24: break; // PK2
 		//case 0x25: break; // UP2
-		//case 0x26: break; // POW
+		case 0x26: AddCode("pow(" + GetSRC(src0) + ", " + GetSRC(src1) +")"); break; // POW
 		//case 0x27: break; // PKB
 		//case 0x28: break; // UPB
 		//case 0x29: break; // PK16
@@ -218,23 +256,23 @@ void FragmentDecompilerThread::Task()
 		//case 0x37: break; // TIMESWTEX
 		//case 0x38: break; // DP2
 		//case 0x39: break; // NRM
-		//case 0x3a: break; // DIV
-		//case 0x3b: break; // DIVSQ
+		case 0x3a: AddCode("(" + GetSRC(src0) + " / " + GetSRC(src1) + ")"); break; // DIV
+		case 0x3b: AddCode("(" + GetSRC(src0) + " / " + GetSRC(src1) + ")"); break; // DIVSQ
 		//case 0x3c: break; // LIF
-		//case 0x3d: break; // FENCT
-		//case 0x3e: break; // FENCB
+		case 0x3d: break; // FENCT
+		case 0x3e: break; // FENCB
 
 		default:
-			ConLog.Error("Unknown opcode 0x%x", dst.opcode);
+			ConLog.Error("Unknown opcode 0x%x (inst %d)", dst.opcode, m_size / (4 * 4));
 			Emu.Pause();
 		break;
 		}
 
-		m_size += 4 * 4;
+		m_size += m_offset;
 
 		if(dst.end) break;
 
-		data.SetOffset(4 * 4);
+		data.SetOffset(m_offset);
 	}
 
 	m_shader = BuildCode();
diff --git a/rpcs3/Emu/GS/GL/FragmentProgram.h b/rpcs3/Emu/GS/GL/FragmentProgram.h
index e3758cd303..09955cbc95 100644
--- a/rpcs3/Emu/GS/GL/FragmentProgram.h
+++ b/rpcs3/Emu/GS/GL/FragmentProgram.h
@@ -100,6 +100,7 @@ struct FragmentDecompilerThread : public ThreadBase
 	u32 m_addr;
 	u32& m_size;
 	u32 m_const_index;
+	u32 m_offset;
 
 	FragmentDecompilerThread(wxString& shader, ParamArray& parr, u32 addr, u32& size)
 		: ThreadBase(false, "Fragment Shader Decompiler Thread")
@@ -115,7 +116,7 @@ struct FragmentDecompilerThread : public ThreadBase
 	wxString GetMask();
 
 	void AddCode(wxString code);
-	wxString AddReg(u32 index);
+	wxString AddReg(u32 index, int fp16);
 	wxString AddConst();
 	wxString AddTex();
 
diff --git a/rpcs3/Emu/GS/GL/GLGSRender.cpp b/rpcs3/Emu/GS/GL/GLGSRender.cpp
index 4f1cd9973a..686f7f099f 100644
--- a/rpcs3/Emu/GS/GL/GLGSRender.cpp
+++ b/rpcs3/Emu/GS/GL/GLGSRender.cpp
@@ -1,8 +1,8 @@
 #include "stdafx.h"
 #include "GLGSRender.h"
 #include "Emu/Cell/PPCInstrTable.h"
-#define CMD_DEBUG 1
-#define DUMP_VERTEX_DATA 1
+#define CMD_DEBUG 0
+#define DUMP_VERTEX_DATA 0
 
 #if	CMD_DEBUG
 	#define CMD_LOG ConLog.Write
@@ -75,8 +75,8 @@ void GLGSFrame::SetViewport(int x, int y, u32 w, u32 h)
 }
 
 GLGSRender::GLGSRender()
-	: m_frame(NULL)
-	, m_rsx_thread(NULL)
+	: m_frame(nullptr)
+	, m_rsx_thread(nullptr)
 	, m_fp_buf_num(-1)
 	, m_vp_buf_num(-1)
 {
@@ -122,6 +122,12 @@ void GLRSXThread::Task()
 	glEnable(GL_TEXTURE_2D);
 	glSwapInterval(Ini.GSVSyncEnable.GetValue() ? 1 : 0);
 
+	for(u32 i=0; i<p.m_textures_count; ++i)
+	{
+		GLTexture& tex = p.m_frame->GetTexture(i);
+		tex.Create();
+	}
+
 	bool draw = true;
 	u32 drawed = 0;
 	u32 skipped = 0;
@@ -172,7 +178,7 @@ void GLRSXThread::Task()
 			continue;
 		}
 
-		ConLog.Write("addr = 0x%x", p.m_ioAddress + get);
+		//ConLog.Write("addr = 0x%x", p.m_ioAddress + get);
 		const u32 cmd = Memory.Read32(p.m_ioAddress + get);
 		const u32 count = (cmd >> 18) & 0x7ff;
 		//if(cmd == 0) continue;
@@ -251,6 +257,10 @@ void GLGSRender::Init(const u32 ioAddress, const u32 ioSize, const u32 ctrlAddre
 	m_localAddress = localAddress;
 	m_ctrl = (CellGcmControl*)Memory.GetMemFromAddr(m_ctrlAddress);
 
+	m_cur_vertex_prog = nullptr;
+	m_cur_shader_prog = nullptr;
+	m_cur_shader_prog_num = 0;
+
 	(m_rsx_thread = new GLRSXThread(this))->Start();
 }
 
@@ -309,7 +319,7 @@ void GLGSRender::EnableVertexData(bool indexed_draw)
 	wxFile dump("VertexDataArray.dump", wxFile::write);
 #endif
 
-	for(u32 i=0; i<16; ++i)
+	for(u32 i=0; i<m_vertex_count; ++i)
 	{
 		if(!m_vertex_data[i].IsEnabled()) continue;
 
@@ -409,7 +419,7 @@ void GLGSRender::EnableVertexData(bool indexed_draw)
 void GLGSRender::DisableVertexData()
 {
 	m_vdata.Clear();
-	for(u32 i=0; i<16; ++i)
+	for(u32 i=0; i<m_vertex_count; ++i)
 	{
 		if(!m_vertex_data[i].IsEnabled()) continue;
 		m_vertex_data[i].data.Clear();
@@ -419,7 +429,7 @@ void GLGSRender::DisableVertexData()
 
 void GLGSRender::LoadVertexData(u32 first, u32 count)
 {
-	for(u32 i=0; i<16; ++i)
+	for(u32 i=0; i<m_vertex_count; ++i)
 	{
 		if(!m_vertex_data[i].IsEnabled()) continue;
 
@@ -429,11 +439,10 @@ void GLGSRender::LoadVertexData(u32 first, u32 count)
 
 void GLGSRender::InitVertexData()
 {
-	for(u32 i=0; i<m_cur_vertex_prog->constants4.GetCount(); ++i)
+	for(u32 i=0; i<m_transform_constants.GetCount(); ++i)
 	{
-		const VertexProgram::Constant4& c = m_cur_vertex_prog->constants4[i];
+		const TransformConstant& c = m_transform_constants[i];
 		const wxString& name = wxString::Format("vc%d", c.id);
-		//const int l = glGetUniformLocation(m_program.id, name);
 		const int l = m_program.GetLocation(name);
 		checkForGlError("glGetUniformLocation " + name);
 
@@ -499,7 +508,9 @@ void GLGSRender::DoCmd(const u32 fcmd, const u32 cmd, mem32_t& args, const u32 c
 		CMD_LOG("index = %d, offset=0x%x, location=0x%x, cubemap=0x%x, dimension=0x%x, format=0x%x, mipmap=0x%x",
 			index, offset, location, cubemap, dimension, format, mipmap);
 
-		tex.SetOffset(GetAddress(offset, location));
+		u32 tex_addr = GetAddress(offset, location);
+		ConLog.Warning("texture addr = 0x%x", tex_addr);
+		tex.SetOffset(tex_addr);
 		tex.SetFormat(cubemap, dimension, format, mipmap);
 	}
 	break;
@@ -810,10 +821,11 @@ void GLGSRender::DoCmd(const u32 fcmd, const u32 cmd, mem32_t& args, const u32 c
 
 	case NV4097_SET_COLOR_CLEAR_VALUE:
 	{
-		const u8 a = (args[0] >> 24) & 0xff;
-		const u8 r = (args[0] >> 16) & 0xff;
-		const u8 g = (args[0] >> 8) & 0xff;
-		const u8 b = args[0] & 0xff;
+		const u32 color = args[0];
+		const u8 a = (color >> 24) & 0xff;
+		const u8 r = (color >> 16) & 0xff;
+		const u8 g = (color >> 8) & 0xff;
+		const u8 b = color & 0xff;
 		CMD_LOG("a=%d, r=%d, g=%d, b=%d", a, r, g, b);
 		glClearColor(
 			(float)r / 255.0f,
@@ -825,9 +837,9 @@ void GLGSRender::DoCmd(const u32 fcmd, const u32 cmd, mem32_t& args, const u32 c
 
 	case NV4097_SET_SHADER_PROGRAM:
 	{
-		m_shader_prog.Delete();
-		m_shader_prog.addr = GetAddress(args[0] & ~0x3, (args[0] & 0x3) - 1);
-		//m_program.Delete();
+		m_cur_shader_prog = &m_shader_progs[m_cur_shader_prog_num++];
+		m_cur_shader_prog->Delete();
+		m_cur_shader_prog->addr = GetAddress(args[0] & ~0x3, (args[0] & 0x3) - 1);
 	}
 	break;
 
@@ -850,13 +862,14 @@ void GLGSRender::DoCmd(const u32 fcmd, const u32 cmd, mem32_t& args, const u32 c
 
 	case NV4097_SET_TRANSFORM_PROGRAM_LOAD:
 	{
-		//m_program.Delete();
 		m_cur_vertex_prog = &m_vertex_progs[args[0]];
 		m_cur_vertex_prog->Delete();
 
 		if(count == 2)
 		{
 			const u32 start = args[1];
+			if(start)
+				ConLog.Warning("NV4097_SET_TRANSFORM_PROGRAM_LOAD: start = %d", start);
 		}
 	}
 	break;
@@ -899,12 +912,6 @@ void GLGSRender::DoCmd(const u32 fcmd, const u32 cmd, mem32_t& args, const u32 c
 			break;
 		}
 
-		if(!m_cur_vertex_prog)
-		{
-			ConLog.Warning("NV4097_SET_TRANSFORM_CONSTANT_LOAD: m_cur_vertex_prog == NULL");
-			break;
-		}
-
 		for(u32 id = args[0], i = 1; i<count; ++id)
 		{
 			const u32 x = args[i++];
@@ -912,15 +919,10 @@ void GLGSRender::DoCmd(const u32 fcmd, const u32 cmd, mem32_t& args, const u32 c
 			const u32 z = args[i++];
 			const u32 w = args[i++];
 
-			VertexProgram::Constant4 c;
-			c.id = id;
-			c.x = (float&)x;
-			c.y = (float&)y;
-			c.z = (float&)z;
-			c.w = (float&)w;
+			TransformConstant c(id, (float&)x, (float&)y, (float&)z, (float&)w);
 
 			CMD_LOG("SET_TRANSFORM_CONSTANT_LOAD[%d : %d] = (%f, %f, %f, %f)", i, id, c.x, c.y, c.z, c.w);
-			m_cur_vertex_prog->constants4.AddCpy(c);
+			m_transform_constants.AddCpy(c);
 		}
 	}
 	break;
@@ -1154,13 +1156,20 @@ void GLGSRender::DoCmd(const u32 fcmd, const u32 cmd, mem32_t& args, const u32 c
 
 bool GLGSRender::LoadProgram()
 {
-	m_fp_buf_num = m_prog_buffer.SearchFp(m_shader_prog);
+	if(!m_cur_shader_prog)
+	{
+		ConLog.Warning("LoadProgram: m_cur_shader_prog == NULL");
 
-	if(m_fp_buf_num == -1) m_shader_prog.Decompile();
+		return false;
+	}
+
+	m_fp_buf_num = m_prog_buffer.SearchFp(*m_cur_shader_prog);
+
+	if(m_fp_buf_num == -1) m_cur_shader_prog->Decompile();
 
 	if(!m_cur_vertex_prog)
 	{
-		ConLog.Warning("NV4097_SET_BEGIN_END: m_cur_vertex_prog == NULL");
+		ConLog.Warning("LoadProgram: m_cur_vertex_prog == NULL");
 
 		return false;
 	}
@@ -1182,11 +1191,11 @@ bool GLGSRender::LoadProgram()
 	if(m_fp_buf_num == -1)
 	{
 		ConLog.Warning("FP not found in buffer!");
-		m_shader_prog.Wait();
-		m_shader_prog.Compile();
+		m_cur_shader_prog->Wait();
+		m_cur_shader_prog->Compile();
 
 		wxFile f(wxGetCwd() + "/FragmentProgram.txt", wxFile::write);
-		f.Write(m_shader_prog.shader);
+		f.Write(m_cur_shader_prog->shader);
 	}
 
 	if(m_vp_buf_num == -1)
@@ -1203,10 +1212,14 @@ bool GLGSRender::LoadProgram()
 		m_program.id = m_prog_buffer.GetProg(m_fp_buf_num, m_vp_buf_num);
 	}
 
-	if(!m_program.id)
+	if(m_program.id)
 	{
-		m_program.Create(m_cur_vertex_prog->id, m_shader_prog.id);
-		m_prog_buffer.Add(m_program, m_shader_prog, *m_cur_vertex_prog);
+		m_program.Use();
+	}
+	else
+	{
+		m_program.Create(m_cur_vertex_prog->id, m_cur_shader_prog->id);
+		m_prog_buffer.Add(m_program, *m_cur_shader_prog, *m_cur_vertex_prog);
 
 		m_program.Use();
 
@@ -1270,7 +1283,12 @@ void GLGSRender::ExecCMD()
 			glDepthRangef(m_clip_min, m_clip_max);
 		}
 
-		for(u32 i=0; i<16; ++i)
+		if(m_indexed_array.m_count && m_draw_array_count)
+		{
+			ConLog.Warning("m_indexed_array.m_count && draw_array_count");
+		}
+
+		for(u32 i=0; i<m_textures_count; ++i)
 		{
 			GLTexture& tex = m_frame->GetTexture(i);
 			if(!tex.IsEnabled()) continue;
@@ -1280,17 +1298,11 @@ void GLGSRender::ExecCMD()
 			tex.Bind();
 			checkForGlError("tex.Bind");
 			m_program.SetTex(i);
-			checkForGlError("m_program.SetTex");
 			tex.Init();
 			checkForGlError("tex.Init");
 			tex.Save();
 		}
 
-		if(m_indexed_array.m_count && m_draw_array_count)
-		{
-			ConLog.Warning("m_indexed_array.m_count && draw_array_count");
-		}
-
 		if(m_indexed_array.m_count)
 		{
 			LoadVertexData(m_indexed_array.index_min, m_indexed_array.index_max - m_indexed_array.index_min + 1);
@@ -1347,13 +1359,20 @@ void GLGSRender::ExecCMD()
 
 void GLGSRender::Reset()
 {
-	m_program.Delete();
-	//m_program.id = 0;
-	m_shader_prog.id = 0;
+	m_program.UnUse();
+
+	if(m_cur_shader_prog)
+		m_cur_shader_prog->id = 0;
+
 	if(m_cur_vertex_prog)
 		m_cur_vertex_prog->id = 0;
 
-	memset(m_vertex_data, 0, sizeof(VertexData) * 16);
+	m_cur_shader_prog_num = 0;
+	m_transform_constants.Clear();
+	for(uint i=0; i<m_vertex_count; ++i)
+	{
+		m_vertex_data[i].Reset();
+	}
 
 	if(m_vbo.IsCreated())
 	{
diff --git a/rpcs3/Emu/GS/GL/GLGSRender.h b/rpcs3/Emu/GS/GL/GLGSRender.h
index e8aef91f04..22def2cbd4 100644
--- a/rpcs3/Emu/GS/GL/GLGSRender.h
+++ b/rpcs3/Emu/GS/GL/GLGSRender.h
@@ -38,6 +38,20 @@ public:
 	{
 	}
 
+	void Create()
+	{
+		if(!m_id)
+		{
+			glGenTextures(1, &m_id);
+			checkForGlError("GLTexture::Init() -> glGenTextures");
+			Bind();
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP);
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP);
+		}
+	}
+
 	void SetRect(const u32 width, const u32 height)
 	{
 		m_width = width;
@@ -68,21 +82,8 @@ public:
 
 	void Init()
 	{
-		if(!m_id)
-		{
-			glGenTextures(1, &m_id);
-			checkForGlError("GLTexture::Init() -> glGenTextures");
-			glBindTexture(GL_TEXTURE_2D, m_id);
-			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
-			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP);
-			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP);
-		}
-		else
-		{
-			glBindTexture(GL_TEXTURE_2D, m_id);
-		}
-
+		Bind();
+		ConLog.Warning("texture addr = 0x%x, width = %d, height = %d", m_offset, m_width, m_height);
 		//TODO: safe init
 		checkForGlError("GLTexture::Init() -> glBindTexture");
 
@@ -106,18 +107,17 @@ public:
 		default: ConLog.Error("Init tex error: Bad tex format (0x%x)", m_format); break;
 		}
 
-		//glBindTexture(GL_TEXTURE_2D, 0);
+		//Unbind();
 	}
 
 	void Save(const wxString& name)
 	{
 		if(!m_id || !m_offset) return;
 
-		ConLog.Write("start");
 		u32* alldata = new u32[m_width * m_height];
 
-		glBindTexture(GL_TEXTURE_2D, m_id);
-		glGetTexImage(GL_TEXTURE_2D, 0, GL_RGBA, GL_UNSIGNED_BYTE, alldata);
+		Bind();
+		glGetTexImage(GL_TEXTURE_2D, 0, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8, alldata);
 
 		u8* data = new u8[m_width * m_height * 3];
 		u8* alpha = new u8[m_width * m_height];
@@ -133,7 +133,6 @@ public:
 			*dst_a++ = *src++;
 		}
 
-		ConLog.Write("end");
 		wxImage out;
 		out.Create(m_width, m_height, data, alpha);
 		out.SaveFile(name, wxBITMAP_TYPE_PNG);
@@ -160,10 +159,34 @@ public:
 		glBindTexture(GL_TEXTURE_2D, m_id);
 	}
 
+	void Unbind()
+	{
+		glBindTexture(GL_TEXTURE_2D, 0);
+	}
+
 	void Enable(bool enable) { m_enabled = enable; }
 	bool IsEnabled() const { return m_enabled; }
 };
 
+struct TransformConstant
+{
+	u32 id;
+	float x, y, z, w;
+
+	TransformConstant()
+	{
+	}
+
+	TransformConstant(u32 id, float x, float y, float z, float w)
+		: id(id)
+		, x(x)
+		, y(y)
+		, z(z)
+		, w(w)
+	{
+	}
+};
+
 struct IndexArrayData
 {
 	Array<u8> m_data;
@@ -228,16 +251,25 @@ class GLGSRender
 	, public GSRender
 	, public ExecRSXCMDdata
 {
+public:
+	static const uint m_vertex_count = 16;
+	static const uint m_fragment_count = 16;
+	static const uint m_textures_count = 16;
+
 private:
 	GLRSXThread* m_rsx_thread;
 
 	IndexArrayData m_indexed_array;
 
-	ShaderProgram m_shader_prog;
-	VertexData m_vertex_data[16];
+	ShaderProgram m_shader_progs[m_fragment_count];
+	ShaderProgram* m_cur_shader_prog;
+	int m_cur_shader_prog_num;
+	VertexData m_vertex_data[m_vertex_count];
 	Array<u8> m_vdata;
-	VertexProgram m_vertex_progs[16];
+	VertexProgram m_vertex_progs[m_vertex_count];
 	VertexProgram* m_cur_vertex_prog;
+	Array<TransformConstant> m_transform_constants;
+
 	Program m_program;
 	int m_fp_buf_num;
 	int m_vp_buf_num;
diff --git a/rpcs3/Emu/GS/GL/Program.cpp b/rpcs3/Emu/GS/GL/Program.cpp
index bc90e28628..9d74358cc0 100644
--- a/rpcs3/Emu/GS/GL/Program.cpp
+++ b/rpcs3/Emu/GS/GL/Program.cpp
@@ -19,7 +19,9 @@ int Program::GetLocation(const wxString& name)
 	u32 pos = m_locations.Move(new Location());
 	m_locations[pos].name = name;
 
-	return m_locations[pos].loc = glGetUniformLocation(id, name);
+	m_locations[pos].loc = glGetUniformLocation(id, name);
+	checkForGlError(wxString::Format("glGetUniformLocation(0x%x, %s)", id, name));
+	return m_locations[pos].loc;
 }
 
 bool Program::IsCreated() const
@@ -72,6 +74,11 @@ void Program::Create(const u32 vp, const u32 fp)
 	}
 }
 
+void Program::UnUse()
+{
+	id = 0;
+}
+
 void Program::Use()
 {
 	glUseProgram(id);
@@ -80,8 +87,10 @@ void Program::Use()
 
 void Program::SetTex(u32 index)
 {
-	glUniform1i(GetLocation(wxString::Format("tex_%d", index)), index);
-	checkForGlError(wxString::Format("SetTex(%d)", index));
+	int loc = GetLocation(wxString::Format("tex%d", index));
+	checkForGlError(wxString::Format("GetLocation(tex%d)", index));
+	glUniform1i(loc, index);
+	checkForGlError(wxString::Format("SetTex(%d - %d)", index, loc));
 }
 
 void Program::Delete()
diff --git a/rpcs3/Emu/GS/GL/Program.h b/rpcs3/Emu/GS/GL/Program.h
index 4a376414b6..71fb597ff6 100644
--- a/rpcs3/Emu/GS/GL/Program.h
+++ b/rpcs3/Emu/GS/GL/Program.h
@@ -22,6 +22,7 @@ public:
 	bool IsCreated() const;
 	void Create(const u32 vp, const u32 fp);
 	void Use();
+	void UnUse();
 	void SetTex(u32 index);
 	void Delete();
 };
\ No newline at end of file
diff --git a/rpcs3/Emu/GS/GL/ProgramBuffer.cpp b/rpcs3/Emu/GS/GL/ProgramBuffer.cpp
index ee6a292e73..859bad2bce 100644
--- a/rpcs3/Emu/GS/GL/ProgramBuffer.cpp
+++ b/rpcs3/Emu/GS/GL/ProgramBuffer.cpp
@@ -8,7 +8,7 @@ int ProgramBuffer::SearchFp(ShaderProgram& fp)
 		if(memcmp(&m_buf[i].fp_data[0], &Memory[fp.addr], m_buf[i].fp_data.GetCount()) != 0) continue;
 
 		fp.id = m_buf[i].fp_id;
-		fp.shader = m_buf[i].fp_shader;
+		fp.shader = m_buf[i].fp_shader.GetPtr();
 
 		return i;
 	}
@@ -21,10 +21,10 @@ int ProgramBuffer::SearchVp(VertexProgram& vp)
 	for(u32 i=0; i<m_buf.GetCount(); ++i)
 	{
 		if(m_buf[i].vp_data.GetCount() != vp.data.GetCount()) continue;
-		if(memcmp(&m_buf[i].vp_data[0], &vp.data[0], vp.data.GetCount() * 4) != 0) continue;
+		if(memcmp(m_buf[i].vp_data.GetPtr(), vp.data.GetPtr(), vp.data.GetCount() * 4) != 0) continue;
 
 		vp.id = m_buf[i].vp_id;
-		vp.shader = m_buf[i].vp_shader;
+		vp.shader = m_buf[i].vp_shader.GetPtr();
 
 		return i;
 	}
@@ -35,18 +35,30 @@ int ProgramBuffer::SearchVp(VertexProgram& vp)
 bool ProgramBuffer::CmpVP(const u32 a, const u32 b) const
 {
 	if(m_buf[a].vp_data.GetCount() != m_buf[b].vp_data.GetCount()) return false;
-	return memcmp(&m_buf[a].vp_data[0], &m_buf[b].vp_data[0], m_buf[a].vp_data.GetCount() * 4) == 0;
+	return memcmp(m_buf[a].vp_data.GetPtr(), m_buf[b].vp_data.GetPtr(), m_buf[a].vp_data.GetCount() * 4) == 0;
 }
 
 bool ProgramBuffer::CmpFP(const u32 a, const u32 b) const
 {
 	if(m_buf[a].fp_data.GetCount() != m_buf[b].fp_data.GetCount()) return false;
-	return memcmp(&m_buf[a].fp_data[0], &m_buf[b].fp_data[0], m_buf[a].fp_data.GetCount()) == 0;
+	return memcmp(m_buf[a].fp_data.GetPtr(), m_buf[b].fp_data.GetPtr(), m_buf[a].fp_data.GetCount()) == 0;
 }
 
 u32 ProgramBuffer::GetProg(u32 fp, u32 vp) const
 {
-	if(fp == vp) return m_buf[fp].prog_id;
+	if(fp == vp)
+	{
+		/*
+		ConLog.Write("Get program (%d):", fp);
+		ConLog.Write("*** prog id = %d", m_buf[fp].prog_id);
+		ConLog.Write("*** vp id = %d", m_buf[fp].vp_id);
+		ConLog.Write("*** fp id = %d", m_buf[fp].fp_id);
+
+		ConLog.Write("*** vp shader = \n%s", m_buf[fp].vp_shader);
+		ConLog.Write("*** fp shader = \n%s", m_buf[fp].fp_shader);
+		*/
+		return m_buf[fp].prog_id;
+	}
 
 	for(u32 i=0; i<m_buf.GetCount(); ++i)
 	{
@@ -54,13 +66,15 @@ u32 ProgramBuffer::GetProg(u32 fp, u32 vp) const
 
 		if(CmpVP(vp, i) && CmpFP(fp, i))
 		{
+			/*
 			ConLog.Write("Get program (%d):", i);
 			ConLog.Write("*** prog id = %d", m_buf[i].prog_id);
 			ConLog.Write("*** vp id = %d", m_buf[i].vp_id);
 			ConLog.Write("*** fp id = %d", m_buf[i].fp_id);
 
-			ConLog.Write("*** vp shader = %s\n", m_buf[i].vp_shader);
-			ConLog.Write("*** fp shader = %s\n", m_buf[i].fp_shader);
+			ConLog.Write("*** vp shader = \n%s", m_buf[i].vp_shader);
+			ConLog.Write("*** fp shader = \n%s", m_buf[i].fp_shader);
+			*/
 			return m_buf[i].prog_id;
 		}
 	}
@@ -79,8 +93,8 @@ void ProgramBuffer::Add(Program& prog, ShaderProgram& fp, VertexProgram& vp)
 	ConLog.Write("*** vp data size = %d", vp.data.GetCount() * 4);
 	ConLog.Write("*** fp data size = %d", fp.size);
 
-	ConLog.Write("*** vp shader = %s\n", vp.shader);
-	ConLog.Write("*** fp shader = %s\n", fp.shader);
+	ConLog.Write("*** vp shader = \n%s", vp.shader);
+	ConLog.Write("*** fp shader = \n%s", fp.shader);
 
 	new_buf.prog_id = prog.id;
 	new_buf.vp_id = vp.id;
diff --git a/rpcs3/Emu/GS/GL/ProgramBuffer.h b/rpcs3/Emu/GS/GL/ProgramBuffer.h
index 988b46371c..8d66dca99b 100644
--- a/rpcs3/Emu/GS/GL/ProgramBuffer.h
+++ b/rpcs3/Emu/GS/GL/ProgramBuffer.h
@@ -8,8 +8,8 @@ struct BufferInfo
 	u32 vp_id;
 	Array<u8> fp_data;
 	Array<u32> vp_data;
-	wxString fp_shader;
-	wxString vp_shader;
+	ArrayString fp_shader;
+	ArrayString vp_shader;
 };
 
 struct ProgramBuffer
diff --git a/rpcs3/Emu/GS/GL/ShaderParam.h b/rpcs3/Emu/GS/GL/ShaderParam.h
index f90ba43acc..553d4278a6 100644
--- a/rpcs3/Emu/GS/GL/ShaderParam.h
+++ b/rpcs3/Emu/GS/GL/ShaderParam.h
@@ -5,18 +5,21 @@ enum ParamFlag
 {
 	PARAM_IN,
 	PARAM_OUT,
+	PARAM_UNIFORM,
 	PARAM_CONST,
 	PARAM_NONE,
 };
 
 struct ParamItem
 {
-	wxString name;
-	wxString location;
+	ArrayString name;
+	ArrayString location;
+	ArrayString value;
 
-	ParamItem(const wxString& _name, int _location)
+	ParamItem(const wxString& _name, int _location, const wxString& _value = wxEmptyString)
 		: name(_name)
 		, location(_location > -1 ? wxString::Format("layout (location = %d) ", _location) : "")
+		, value(_value)
 	{
 	}
 };
@@ -24,7 +27,7 @@ struct ParamItem
 struct ParamType
 {
 	const ParamFlag flag;
-	wxString type;
+	ArrayString type;
 	Array<ParamItem> items;
 
 	ParamType(const ParamFlag _flag, const wxString& _type)
@@ -37,7 +40,7 @@ struct ParamType
 	{
 		for(u32 i=0; i<items.GetCount(); ++i)
 		{
-			if(items[i].name.Cmp(name) == 0) return true;
+			if(name.Cmp(items[i].name.GetPtr()) == 0) return true;
 		}
 
 		return false;
@@ -49,10 +52,15 @@ struct ParamType
 
 		for(u32 n=0; n<items.GetCount(); ++n)
 		{
-			ret += items[n].location;
-			ret += type;
+			ret += items[n].location.GetPtr();
+			ret += type.GetPtr();
 			ret += " ";
-			ret += items[n].name;
+			ret += items[n].name.GetPtr();
+			if(items[n].value.GetCount())
+			{
+				ret += " = ";
+				ret += items[n].value.GetPtr();
+			}
 			ret += ";\n";
 		}
 
@@ -68,10 +76,10 @@ struct ParamArray
 	{
 		for(u32 i=0; i<params.GetCount(); ++i)
 		{
-			if(params[i].type.Cmp(type) == 0) return &params[i];
+			if(type.Cmp(params[i].type.GetPtr()) == 0) return &params[i];
 		}
 
-		return NULL;
+		return nullptr;
 	}
 
 	wxString GetParamFlag(const ParamFlag flag)
@@ -80,12 +88,32 @@ struct ParamArray
 		{
 		case PARAM_OUT:		return "out ";
 		case PARAM_IN:		return "in ";
-		case PARAM_CONST:	return "uniform ";
+		case PARAM_UNIFORM:	return "uniform ";
+		case PARAM_CONST:	return "const ";
 		}
 
 		return wxEmptyString;
 	}
 
+	wxString AddParam(wxString type, const wxString& name, const wxString& value)
+	{
+		type = GetParamFlag(PARAM_CONST) + type;
+		ParamType* t = SearchParam(type);
+
+		if(t)
+		{
+			if(!t->SearchName(name)) t->items.Move(new ParamItem(name, -1, value));
+		}
+		else
+		{
+			const u32 num = params.GetCount();
+			params.Move(new ParamType(PARAM_CONST, type));
+			params[num].items.Move(new ParamItem(name, -1, value));
+		}
+
+		return name;
+	}
+
 	wxString AddParam(const ParamFlag flag, wxString type, const wxString& name, int location = -1)
 	{
 		type = GetParamFlag(flag) + type;
diff --git a/rpcs3/Emu/GS/GL/VertexProgram.cpp b/rpcs3/Emu/GS/GL/VertexProgram.cpp
index 9938ad184f..7a9bcfcfb4 100644
--- a/rpcs3/Emu/GS/GL/VertexProgram.cpp
+++ b/rpcs3/Emu/GS/GL/VertexProgram.cpp
@@ -1,28 +1,36 @@
 #include "stdafx.h"
 #include "VertexProgram.h"
 
-wxString VertexDecompilerThread::GetVecMask()
+wxString VertexDecompilerThread::GetMask(bool is_sca)
 {
 	wxString ret = wxEmptyString;
 
-	if(d3.vec_writemask_x) ret += "x";
-	if(d3.vec_writemask_y) ret += "y";
-	if(d3.vec_writemask_z) ret += "z";
-	if(d3.vec_writemask_w) ret += "w";
+	if(is_sca)
+	{
+		if(d3.sca_writemask_x) ret += "x";
+		if(d3.sca_writemask_y) ret += "y";
+		if(d3.sca_writemask_z) ret += "z";
+		if(d3.sca_writemask_w) ret += "w";
+	}
+	else
+	{
+		if(d3.vec_writemask_x) ret += "x";
+		if(d3.vec_writemask_y) ret += "y";
+		if(d3.vec_writemask_z) ret += "z";
+		if(d3.vec_writemask_w) ret += "w";
+	}
 
 	return ret.IsEmpty() || ret == "xyzw" ? wxEmptyString : ("." + ret);
 }
 
+wxString VertexDecompilerThread::GetVecMask()
+{
+	return GetMask(false);
+}
+
 wxString VertexDecompilerThread::GetScaMask()
 {
-	wxString ret = wxEmptyString;
-
-	if(d3.sca_writemask_x) ret += "x";
-	if(d3.sca_writemask_y) ret += "y";
-	if(d3.sca_writemask_z) ret += "z";
-	if(d3.sca_writemask_w) ret += "w";
-
-	return ret.IsEmpty() || ret == "xyzw" ? wxEmptyString : ("." + ret);
+	return GetMask(true);
 }
 
 wxString VertexDecompilerThread::GetDST(bool isSca)
@@ -38,26 +46,25 @@ wxString VertexDecompilerThread::GetDST(bool isSca)
 	};
 
 	wxString ret = wxEmptyString;
-	u32 dst = isSca ? d3.sca_dst : d3.dst;
 
-	switch(dst)
+	switch(isSca ? 0x1f : d3.dst)
 	{
 	case 0x0: case 0x6:
-		ret += reg_table[dst];
+		ret += reg_table[d3.dst];
 	break;
 
 	case 0x1f:
-		ret += m_parr.AddParam(PARAM_NONE, "vec4", wxString::Format("tmp%d", d0.dst_tmp));
+		ret += m_parr.AddParam(PARAM_NONE, "vec4", wxString::Format("tmp%d", isSca ? d3.sca_dst_tmp : d0.dst_tmp));
 	break;
 
 	default:
-		if(dst < WXSIZEOF(reg_table))
+		if(d3.dst < WXSIZEOF(reg_table))
 		{
-			ret += m_parr.AddParam(PARAM_OUT, "vec4", reg_table[dst]);
+			ret += m_parr.AddParam(PARAM_OUT, "vec4", reg_table[d3.dst]);
 		}
 		else
 		{
-			ConLog.Error("Bad dst reg num: %d", dst);
+			ConLog.Error("Bad dst reg num: %d", d3.dst);
 			ret += m_parr.AddParam(PARAM_OUT, "vec4", "unk");
 		}
 	break;
@@ -97,7 +104,7 @@ wxString VertexDecompilerThread::GetSRC(const u32 n, bool isSca)
 		}
 	break;
 	case 3: //const
-		ret += m_parr.AddParam(PARAM_CONST, "vec4", wxString::Format("vc%d", d1.const_src));
+		ret += m_parr.AddParam(PARAM_UNIFORM, "vec4", wxString::Format("vc%d", d1.const_src));
 	break;
 
 	default:
@@ -106,8 +113,7 @@ wxString VertexDecompilerThread::GetSRC(const u32 n, bool isSca)
 	break;
 	}
 
-	static const char f[4] = {'x', 'z', 'w', 'y'};
-	static const char fSca[4] = {'x', 'z', 'y', 'w'};
+	static const char f[4] = {'x', 'y', 'z', 'w'};
 
 	wxString swizzle = wxEmptyString;
 
@@ -117,7 +123,7 @@ wxString VertexDecompilerThread::GetSRC(const u32 n, bool isSca)
 		assert(src[n].swz_z == src[n].swz_w);
 		assert(src[n].swz_x == src[n].swz_z);
 
-		ret += "." + fSca[src[n].swz_x];
+		ret += "." + f[src[n].swz_x];
 	}
 	else
 	{
@@ -144,34 +150,50 @@ wxString VertexDecompilerThread::GetSRC(const u32 n, bool isSca)
 	return ret;
 }
 
-void VertexDecompilerThread::AddVecCode(wxString code, bool src_mask)
+void VertexDecompilerThread::AddCode(bool is_sca, wxString code, bool src_mask)
 {
 	if(d0.cond == 0) return;
-	if(d0.cond != 7)
+	enum
 	{
-		ConLog.Error("Bad cond! %d", d0.cond);
-		Emu.Pause();
-		return;
+		lt = 0x1,
+		eq = 0x2,
+		gt = 0x4,
+	};
+
+	wxString cond;
+
+	if(d0.cond != (lt | gt | eq))
+	{
+		if(d0.cond & (lt | gt))
+		{
+			cond = "!=";
+		}
+		else
+		{
+			if(d0.cond & lt) cond = "<";
+			else if(d0.cond & gt) cond = ">";
+			else if(d0.cond & eq) cond = "=";
+
+			if(d0.cond & eq) cond += "=";
+		}
+
+		ConLog.Error("cond! %d (%d %s %d %d)", d0.cond, d0.dst_tmp, cond, d1.input_src, d1.const_src);
+		cond = wxString::Format("if(tmp%d.x %s 0) ", d0.dst_tmp, cond);
 	}
 
-	code = GetDST() + GetVecMask() + " = " + (src_mask ? code + GetVecMask() : code);
+	code = cond + GetDST(is_sca) + GetMask(is_sca) + " = " + (src_mask ? code + GetMask(is_sca) : code);
 
 	main += "\t" + code + ";\n";
 }
 
-void VertexDecompilerThread::AddScaCode(wxString code, bool src_mask)
+void VertexDecompilerThread::AddVecCode(const wxString& code, bool src_mask)
 {
-	if(d0.cond == 0) return;
-	if(d0.cond != 7)
-	{
-		ConLog.Error("Bad cond! %d", d0.cond);
-		Emu.Pause();
-		return;
-	}
+	AddCode(false, code, src_mask);
+}
 
-	code = GetDST(true) + GetScaMask() + " = " + (src_mask ? code + GetScaMask() : code);
-
-	main += "\t" + code + ";\n";
+void VertexDecompilerThread::AddScaCode(const wxString& code, bool src_mask)
+{
+	AddCode(true, code, src_mask);
 }
 
 wxString VertexDecompilerThread::BuildCode()
@@ -201,16 +223,18 @@ void VertexDecompilerThread::Task()
 		d2.HEX = m_data[i++];
 		d3.HEX = m_data[i++];
 
-		src[0].HEX = d2.src0l | (d1.src0h << 9);
-		src[1].HEX = d2.src1;
-		src[2].HEX = d3.src2l | (d2.src2h << 11);
+		src[0].src0l = d2.src0l;
+		src[0].src0h = d1.src0h;
+		src[1].src1 = d2.src1;
+		src[2].src2l = d3.src2l;
+		src[2].src2h = d2.src2h;
 
 		switch(d1.sca_opcode)
 		{
 		case 0x00: break; // NOP
 		case 0x01: AddScaCode(GetSRC(2, true)); break; // MOV
 		case 0x02: AddScaCode("1 / (" + GetSRC(2, true) + ")"); break; // RCP
-		case 0x03: AddScaCode("clamp(1 / (" + GetSRC(2, true) + "), 5.42101e-20, 1.884467e19"); break; // RCC
+		case 0x03: AddScaCode("clamp(1 / (" + GetSRC(2, true) + "), 5.42101e-20, 1.884467e19)"); break; // RCC
 		case 0x04: AddScaCode("inversesqrt(" + GetSRC(2, true) + ")"); break; // RSQ
 		case 0x05: AddScaCode("exp(" + GetSRC(2, true) + ")"); break; // EXP
 		case 0x06: AddScaCode("log(" + GetSRC(2, true) + ")"); break; // LOG
@@ -229,7 +253,6 @@ void VertexDecompilerThread::Task()
 		//case 0x13: break; // PSH : works differently (PSH o[1].x A0;)
 		//case 0x14: break; // POP : works differently (POP o[1].x;)
 
-
 		default:
 			ConLog.Error("Unknown vp sca_opcode 0x%x", d1.sca_opcode);
 			Emu.Pause();
@@ -241,7 +264,7 @@ void VertexDecompilerThread::Task()
 		case 0x00: break; //NOP
 		case 0x01: AddVecCode(GetSRC(0)); break; //MOV
 		case 0x02: AddVecCode("(" + GetSRC(0) + " * " + GetSRC(1) + ")"); break; //MUL
-		case 0x03: AddVecCode("(" + GetSRC(0) + " + " + GetSRC(1) + ")"); break; //ADD
+		case 0x03: AddVecCode("(" + GetSRC(0) + " + " + GetSRC(2) + ")"); break; //ADD
 		case 0x04: AddVecCode("(" + GetSRC(0) + " * " + GetSRC(1) + " + " + GetSRC(2) + ")"); break; //MAD
 		case 0x05: AddVecCode("dot(" + GetSRC(0) + ".xyz, " + GetSRC(1) + ".xyz)", false); break; //DP3
 		case 0x06: AddVecCode("(dot(" + GetSRC(0) + ".xyz, " + GetSRC(1) + ".xyz) + " + GetSRC(1) + ".w)", false); break; //DPH
@@ -249,18 +272,17 @@ void VertexDecompilerThread::Task()
 		case 0x08: AddVecCode("distance(" + GetSRC(0) + ", " + GetSRC(1) + ")"); break; //DST
 		case 0x09: AddVecCode("min(" + GetSRC(0) + ", " + GetSRC(1) + ")"); break; //MIN
 		case 0x0a: AddVecCode("max(" + GetSRC(0) + ", " + GetSRC(1) + ")"); break; //MAX
-		case 0x0b: AddVecCode("lessThan(" + GetSRC(0) + ", " + GetSRC(1) + ")"); break; //SLT
-		case 0x0c: AddVecCode("greaterThanEqual(" + GetSRC(0) + ", " + GetSRC(1) + ")"); break; //SGE
+		case 0x0b: AddVecCode("vec4(lessThan(" + GetSRC(0) + ", " + GetSRC(1) + "))"); break; //SLT
+		case 0x0c: AddVecCode("vec4(greaterThanEqual(" + GetSRC(0) + ", " + GetSRC(1) + "))"); break; //SGE
 		case 0x0e: AddVecCode("fract(" + GetSRC(0) + ")"); break; //FRC
 		case 0x0f: AddVecCode("floor(" + GetSRC(0) + ")"); break; //FLR
-		case 0x10: AddVecCode("equal(" + GetSRC(0) + ", " + GetSRC(1) + ")"); break; //SEQ
-		case 0x11: AddVecCode("vec4(0.0f, 0.0f, 0.0f, 0.0f)"); break; //SFL
-		case 0x12: AddVecCode("greaterThan(" + GetSRC(0) + ", " + GetSRC(1) + ")"); break; //SGT
-		case 0x13: AddVecCode("lessThanEqual(" + GetSRC(0) + ", " + GetSRC(1) + ")"); break; //SLE
-		case 0x14: AddVecCode("notEqual(" + GetSRC(0) + ", " + GetSRC(1) + ")"); break; //SNE
-		case 0x15: AddVecCode("vec4(1.0f, 1.0f, 1.0f, 1.0f)"); break; //STR
-		case 0x16: AddVecCode("sign(" + GetSRC(0) + ", " + GetSRC(1) + ")"); break; //SSG
-			
+		case 0x10: AddVecCode("vec4(equal(" + GetSRC(0) + ", " + GetSRC(1) + "))"); break; //SEQ
+		//case 0x11: AddVecCode("vec4(0.0f, 0.0f, 0.0f, 0.0f)"); break; //SFL
+		case 0x12: AddVecCode("vec4(greaterThan(" + GetSRC(0) + ", " + GetSRC(1) + "))"); break; //SGT
+		case 0x13: AddVecCode("vec4(lessThanEqual(" + GetSRC(0) + ", " + GetSRC(1) + "))"); break; //SLE
+		case 0x14: AddVecCode("vec4(notEqual(" + GetSRC(0) + ", " + GetSRC(1) + "))"); break; //SNE
+		//case 0x15: AddVecCode("vec4(1.0f, 1.0f, 1.0f, 1.0f)"); break; //STR
+		case 0x16: AddVecCode("sign(" + GetSRC(0) + ")"); break; //SSG
 
 		default:
 			ConLog.Error("Unknown vp opcode 0x%x", d1.vec_opcode);
@@ -276,7 +298,7 @@ void VertexDecompilerThread::Task()
 }
 
 VertexProgram::VertexProgram()
-	: m_decompiler_thread(NULL)
+	: m_decompiler_thread(nullptr)
 	, id(0)
 {
 }
@@ -358,13 +380,7 @@ void VertexProgram::Compile()
 void VertexProgram::Delete()
 {
 	data.Clear();
-	for(u32 i=0; i<parr.params.GetCount(); ++i)
-	{
-		parr.params[i].items.Clear();
-		parr.params[i].type.Clear();
-	}
 	parr.params.Clear();
-	constants4.Clear();
 	shader.Clear();
 
 	if(id)
@@ -374,6 +390,28 @@ void VertexProgram::Delete()
 	}
 }
 
+VertexData::VertexData()
+	: index(0)
+	, frequency(0)
+	, stride(0)
+	, size(0)
+	, type(0)
+	, addr(0)
+	, data()
+{
+}
+
+void VertexData::Reset()
+{
+	index = 0;
+	frequency = 0;
+	stride = 0;
+	size = 0;
+	type = 0;
+	addr = 0;
+	data.ClearF();
+}
+
 void VertexData::Load(u32 start, u32 count)
 {
 	const u32 tsize = GetTypeSize();
diff --git a/rpcs3/Emu/GS/GL/VertexProgram.h b/rpcs3/Emu/GS/GL/VertexProgram.h
index bc513c014a..1147631237 100644
--- a/rpcs3/Emu/GS/GL/VertexProgram.h
+++ b/rpcs3/Emu/GS/GL/VertexProgram.h
@@ -72,8 +72,7 @@ struct VertexDecompilerThread : public ThreadBase
 			u32	end				: 1;
 			u32 index_const		: 1;
 			u32 dst				: 5;
-			u32 sca_dst			: 5;
-			u32 sca_result		: 1;
+			u32 sca_dst_tmp		: 6;
 			u32 vec_writemask_w : 1;
 			u32 vec_writemask_z : 1;
 			u32 vec_writemask_y : 1;
@@ -93,17 +92,37 @@ struct VertexDecompilerThread : public ThreadBase
 
 	union SRC
 	{
-		u16 HEX;
+		union
+		{
+			u32 HEX;
+
+			struct
+			{
+				u32 src0l		: 9;
+				u32 src0h		: 8;
+			};
+
+			struct
+			{
+				u32 src1 : 17;
+			};
+
+			struct
+			{
+				u32 src2l		: 11;
+				u32 src2h		: 6;
+			};
+		};
 
 		struct
 		{
-			u16 reg_type	: 2;
-			u16 tmp_src		: 5;
-			u16 swz_w		: 2;
-			u16 swz_z		: 2;
-			u16 swz_y		: 2;
-			u16 swz_x		: 2;
-			u16 neg			: 1;
+			u32 reg_type	: 2;
+			u32 tmp_src		: 6;
+			u32 swz_w		: 2;
+			u32 swz_z		: 2;
+			u32 swz_y		: 2;
+			u32 swz_x		: 2;
+			u32 neg			: 1;
 		};
 	} src[3];
 
@@ -120,12 +139,14 @@ struct VertexDecompilerThread : public ThreadBase
 	{
 	}
 
+	wxString GetMask(bool is_sca);
 	wxString GetVecMask();
 	wxString GetScaMask();
-	wxString GetDST(bool isSca = false);
-	wxString GetSRC(const u32 n, bool isSca = false);
-	void AddVecCode(wxString code, bool src_mask = true);
-	void AddScaCode(wxString code, bool src_mask = true);
+	wxString GetDST(bool is_sca = false);
+	wxString GetSRC(const u32 n, bool is_sca = false);
+	void AddCode(bool is_sca, wxString code, bool src_mask = true);
+	void AddVecCode(const wxString& code, bool src_mask = true);
+	void AddScaCode(const wxString& code, bool src_mask = true);
 	wxString BuildCode();
 
 	virtual void Task();
@@ -140,14 +161,6 @@ struct VertexProgram
 	VertexProgram();
 	~VertexProgram();
 
-	struct Constant4
-	{
-		u32 id;
-		float x, y, z, w;
-	};
-
-	Array<Constant4> constants4;
-
 	Array<u32> data;
 	ParamArray parr;
 
@@ -174,11 +187,10 @@ struct VertexData
 
 	Array<u8> data;
 
-	VertexData() { memset(this, 0, sizeof(*this)); }
-	~VertexData() { data.Clear(); }
+	VertexData();
 
+	void Reset();
 	bool IsEnabled() { return size > 0; }
-
 	void Load(u32 start, u32 count);
 
 	u32 GetTypeSize();
diff --git a/rpcs3/Emu/SysCalls/FuncList.cpp b/rpcs3/Emu/SysCalls/FuncList.cpp
index 0dbc923703..83557bc3e3 100644
--- a/rpcs3/Emu/SysCalls/FuncList.cpp
+++ b/rpcs3/Emu/SysCalls/FuncList.cpp
@@ -1,7 +1,7 @@
 #include "stdafx.h"
 #include "SysCalls.h"
 
-#define FUNC_LOG_ERROR(x) ConLog.Error(x); return SC_ARGS_1
+#define FUNC_LOG_ERROR(x) ConLog.Error(x); return 0
 s64 SysCalls::DoFunc(const u32 id) 
 {
 	switch(id)