Merge branch 'const_buffer_rework'

11 files changed, 83 insertions, 415 deletions

diff --git a/Source/Core/VideoBackends/D3D/Src/PixelShaderCache.cpp b/Source/Core/VideoBackends/D3D/Src/PixelShaderCache.cpp
index 56b61e05cd..04a1722d04 100644
--- a/Source/Core/VideoBackends/D3D/Src/PixelShaderCache.cpp
+++ b/Source/Core/VideoBackends/D3D/Src/PixelShaderCache.cpp
@@ -14,15 +14,12 @@
 #include "Globals.h"
 #include "PixelShaderGen.h"
 #include "PixelShaderCache.h"
+#include "PixelShaderManager.h"
 
 #include "ConfigManager.h"
 
 extern int frameCount;
 
-// See comment near the bottom of this file.
-float psconstants[C_PENVCONST_END*4];
-bool pscbufchanged = true;
-
 namespace DX11
 {
 
@@ -339,15 +336,15 @@ ID3D11PixelShader* PixelShaderCache::GetClearProgram()
 ID3D11Buffer* &PixelShaderCache::GetConstantBuffer()
 {
 	// TODO: divide the global variables of the generated shaders into about 5 constant buffers to speed this up
-	if (pscbufchanged)
+	if (PixelShaderManager::dirty)
 	{
 		D3D11_MAPPED_SUBRESOURCE map;
 		D3D::context->Map(pscbuf, 0, D3D11_MAP_WRITE_DISCARD, 0, &map);
-		memcpy(map.pData, psconstants, sizeof(psconstants));
+		memcpy(map.pData, &PixelShaderManager::constants, sizeof(PixelShaderConstants));
 		D3D::context->Unmap(pscbuf, 0);
-		pscbufchanged = false;
+		PixelShaderManager::dirty = false;
 		
-		ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(psconstants));
+		ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(PixelShaderConstants));
 	}
 	return pscbuf;
 }
@@ -364,7 +361,7 @@ public:
 
 void PixelShaderCache::Init()
 {
-	unsigned int cbsize = ((sizeof(psconstants))&(~0xf))+0x10; // must be a multiple of 16
+	unsigned int cbsize = ((sizeof(PixelShaderConstants))&(~0xf))+0x10; // must be a multiple of 16
 	D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(cbsize, D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
 	D3D::device->CreateBuffer(&cbdesc, NULL, &pscbuf);
 	CHECK(pscbuf!=NULL, "Create pixel shader constant buffer");
@@ -536,50 +533,4 @@ bool PixelShaderCache::InsertByteCode(const PixelShaderUid &uid, const void* byt
 	return true;
 }
 
-// These are "callbacks" from VideoCommon and thus must be outside namespace DX11.
-// This will have to be changed when we merge.
-
-// HACK to avoid some invasive VideoCommon changes
-// these values are hardcoded, they depend on internal D3DCompile behavior; TODO: Solve this with D3DReflect or something
-// offset given in floats, table index is float4
-static const unsigned int ps_constant_offset_table[] = {
-	0, 4, 8, 12,					// C_COLORS, 16
-	16, 20, 24, 28,					// C_KCOLORS, 16
-	32,								// C_ALPHA, 4
-	36, 40, 44, 48, 52, 56, 60, 64,	// C_TEXDIMS, 32
-	68, 72,							// C_ZBIAS, 8
-	76, 80,							// C_INDTEXSCALE, 8
-	84, 88, 92, 96, 100, 104,		// C_INDTEXMTX, 24
-	108, 112, 116,					// C_FOG, 12
-	120, 124, 128, 132, 136,		// C_PLIGHTS0, 20
-	140, 144, 148, 152, 156,		// C_PLIGHTS1, 20
-	160, 164, 168, 172,	176,		// C_PLIGHTS2, 20
-	180, 184, 188, 192, 196,		// C_PLIGHTS3, 20		
-	200, 204, 208, 212, 216,		// C_PLIGHTS4, 20
-	220, 224, 228, 232, 236,		// C_PLIGHTS5, 20
-	240, 244, 248, 252,	256,		// C_PLIGHTS6, 20
-	260, 264, 268, 272, 276,		// C_PLIGHTS7, 20
-	280, 284, 288, 292				// C_PMATERIALS, 16	
-};
-void Renderer::SetPSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4)
-{
-	psconstants[ps_constant_offset_table[const_number]  ] = f1;
-	psconstants[ps_constant_offset_table[const_number]+1] = f2;
-	psconstants[ps_constant_offset_table[const_number]+2] = f3;
-	psconstants[ps_constant_offset_table[const_number]+3] = f4;
-	pscbufchanged = true;
-}
-
-void Renderer::SetPSConstant4fv(unsigned int const_number, const float* f)
-{
-	memcpy(&psconstants[ps_constant_offset_table[const_number]], f, sizeof(float)*4);
-	pscbufchanged = true;
-}
-
-void Renderer::SetMultiPSConstant4fv(unsigned int const_number, unsigned int count, const float* f)
-{
-	memcpy(&psconstants[ps_constant_offset_table[const_number]], f, sizeof(float)*4*count);
-	pscbufchanged = true;
-}
-
 }  // DX11
diff --git a/Source/Core/VideoBackends/D3D/Src/Render.h b/Source/Core/VideoBackends/D3D/Src/Render.h
index 1fddfdb244..39b5b90d81 100644
--- a/Source/Core/VideoBackends/D3D/Src/Render.h
+++ b/Source/Core/VideoBackends/D3D/Src/Render.h
@@ -51,15 +51,6 @@ public:
 	bool SaveScreenshot(const std::string &filename, const TargetRectangle &rc);
 
 	static bool CheckForResize();
-
-	void SetPSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4);
-	void SetPSConstant4fv(unsigned int const_number, const float *f);
-	void SetMultiPSConstant4fv(unsigned int const_number, unsigned int count, const float *f);
-
-	void SetVSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4);
-	void SetVSConstant4fv(unsigned int const_number, const float *f);
-	void SetMultiVSConstant3fv(unsigned int const_number, unsigned int count, const float *f);
-	void SetMultiVSConstant4fv(unsigned int const_number, unsigned int count, const float *f);
 };
 
 }
diff --git a/Source/Core/VideoBackends/D3D/Src/VertexShaderCache.cpp b/Source/Core/VideoBackends/D3D/Src/VertexShaderCache.cpp
index d9ea7e2a8c..28a03b3bca 100644
--- a/Source/Core/VideoBackends/D3D/Src/VertexShaderCache.cpp
+++ b/Source/Core/VideoBackends/D3D/Src/VertexShaderCache.cpp
@@ -12,14 +12,10 @@
 #include "D3DShader.h"
 #include "Globals.h"
 #include "VertexShaderCache.h"
+#include "VertexShaderManager.h"
 
 #include "ConfigManager.h"
 
-// See comment near the bottom of this file
-static unsigned int vs_constant_offset_table[C_VENVCONST_END];
-float vsconstants[C_VENVCONST_END*4];
-bool vscbufchanged = true;
-
 namespace DX11 {
 
 VertexShaderCache::VSCache VertexShaderCache::vshaders;
@@ -44,15 +40,15 @@ ID3D11Buffer* vscbuf = NULL;
 ID3D11Buffer* &VertexShaderCache::GetConstantBuffer()
 {
 	// TODO: divide the global variables of the generated shaders into about 5 constant buffers to speed this up
-	if (vscbufchanged)
+	if (VertexShaderManager::dirty)
 	{
 		D3D11_MAPPED_SUBRESOURCE map;
 		D3D::context->Map(vscbuf, 0, D3D11_MAP_WRITE_DISCARD, 0, &map);
-		memcpy(map.pData, vsconstants, sizeof(vsconstants));
+		memcpy(map.pData, &VertexShaderManager::constants, sizeof(VertexShaderConstants));
 		D3D::context->Unmap(vscbuf, 0);
-		vscbufchanged = false;
+		VertexShaderManager::dirty = false;
 		
-		ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(vsconstants));
+		ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(VertexShaderConstants));
 	}
 	return vscbuf;
 }
@@ -116,7 +112,7 @@ void VertexShaderCache::Init()
 		{ "COLOR", 0, DXGI_FORMAT_R8G8B8A8_UNORM, 0, 12, D3D11_INPUT_PER_VERTEX_DATA, 0 },
 	};
 
-	unsigned int cbsize = ((sizeof(vsconstants))&(~0xf))+0x10; // must be a multiple of 16
+	unsigned int cbsize = ((sizeof(VertexShaderConstants))&(~0xf))+0x10; // must be a multiple of 16
 	D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(cbsize, D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
 	HRESULT hr = D3D::device->CreateBuffer(&cbdesc, NULL, &vscbuf);
 	CHECK(hr==S_OK, "Create vertex shader constant buffer (size=%u)", cbsize);
@@ -141,19 +137,6 @@ void VertexShaderCache::Init()
 
 	Clear();
 
-	// these values are hardcoded, they depend on internal D3DCompile behavior
-	// TODO: Do this with D3DReflect or something instead
-	unsigned int k;
-	for (k = 0;k <  6;k++) vs_constant_offset_table[C_POSNORMALMATRIX+k]       =   0+4*k;
-	for (k = 0;k <  4;k++) vs_constant_offset_table[C_PROJECTION+k]            =  24+4*k;
-	for (k = 0;k <  4;k++) vs_constant_offset_table[C_MATERIALS+k]             =  40+4*k;
-	for (k = 0;k < 40;k++) vs_constant_offset_table[C_LIGHTS+k]                =  56+4*k;
-	for (k = 0;k < 24;k++) vs_constant_offset_table[C_TEXMATRICES+k]           = 216+4*k;
-	for (k = 0;k < 64;k++) vs_constant_offset_table[C_TRANSFORMMATRICES+k]     = 312+4*k;	
-	for (k = 0;k < 32;k++) vs_constant_offset_table[C_NORMALMATRICES+k]        = 568+4*k;	
-	for (k = 0;k < 64;k++) vs_constant_offset_table[C_POSTTRANSFORMMATRICES+k] = 696+4*k;
-	vs_constant_offset_table[C_DEPTHPARAMS] = 952;
-
 	if (!File::Exists(File::GetUserPath(D_SHADERCACHE_IDX)))
 		File::CreateDir(File::GetUserPath(D_SHADERCACHE_IDX).c_str());
 
@@ -277,39 +260,4 @@ bool VertexShaderCache::InsertByteCode(const VertexShaderUid &uid, D3DBlob* bcod
 	return true;
 }
 
-// These are "callbacks" from VideoCommon and thus must be outside namespace DX11.
-// This will have to be changed when we merge.
-
-// maps the constant numbers to float indices in the constant buffer
-void Renderer::SetVSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4)
-{
-	vsconstants[vs_constant_offset_table[const_number]  ] = f1;
-	vsconstants[vs_constant_offset_table[const_number]+1] = f2;
-	vsconstants[vs_constant_offset_table[const_number]+2] = f3;
-	vsconstants[vs_constant_offset_table[const_number]+3] = f4;
-	vscbufchanged = true;
-}
-
-void Renderer::SetVSConstant4fv(unsigned int const_number, const float* f)
-{
-	memcpy(&vsconstants[vs_constant_offset_table[const_number]], f, sizeof(float)*4);
-	vscbufchanged = true;
-}
-
-void Renderer::SetMultiVSConstant3fv(unsigned int const_number, unsigned int count, const float* f)
-{
-	for (unsigned int i = 0; i < count; i++)
-	{
-		memcpy(&vsconstants[vs_constant_offset_table[const_number+i]], f+3*i, sizeof(float)*3);
-		vsconstants[vs_constant_offset_table[const_number+i]+3] = 0.f;		
-	}
-	vscbufchanged = true;
-}
-
-void Renderer::SetMultiVSConstant4fv(unsigned int const_number, unsigned int count, const float* f)
-{
-	memcpy(&vsconstants[vs_constant_offset_table[const_number]], f, sizeof(float)*4*count);
-	vscbufchanged = true;
-}
-
 }  // namespace DX11
diff --git a/Source/Core/VideoBackends/OGL/CMakeLists.txt b/Source/Core/VideoBackends/OGL/CMakeLists.txt
index f3d7ce4e4f..3e1d2aaa60 100644
--- a/Source/Core/VideoBackends/OGL/CMakeLists.txt
+++ b/Source/Core/VideoBackends/OGL/CMakeLists.txt
@@ -3,7 +3,6 @@ set(SRCS	Src/FramebufferManager.cpp
 			Src/main.cpp
 			Src/NativeVertexFormat.cpp
 			Src/PerfQuery.cpp
-			Src/PixelShaderCache.cpp
 			Src/PostProcessing.cpp
 			Src/ProgramShaderCache.cpp
 			Src/RasterFont.cpp
@@ -12,7 +11,6 @@ set(SRCS	Src/FramebufferManager.cpp
 			Src/StreamBuffer.cpp
 			Src/TextureCache.cpp
 			Src/TextureConverter.cpp
-			Src/VertexShaderCache.cpp
 			Src/VertexManager.cpp)
 
 set(LIBS	videocommon
diff --git a/Source/Core/VideoBackends/OGL/OGL.vcxproj b/Source/Core/VideoBackends/OGL/OGL.vcxproj
index 69e23065c5..46113f53ce 100644
--- a/Source/Core/VideoBackends/OGL/OGL.vcxproj
+++ b/Source/Core/VideoBackends/OGL/OGL.vcxproj
@@ -151,7 +151,6 @@
     <ClCompile Include="Src\main.cpp" />
     <ClCompile Include="Src\NativeVertexFormat.cpp" />
     <ClCompile Include="Src\PerfQuery.cpp" />
-    <ClCompile Include="Src\PixelShaderCache.cpp" />
     <ClCompile Include="Src\PostProcessing.cpp" />
     <ClCompile Include="Src\ProgramShaderCache.cpp" />
     <ClCompile Include="Src\RasterFont.cpp" />
@@ -169,7 +168,6 @@
     <ClCompile Include="Src\TextureCache.cpp" />
     <ClCompile Include="Src\TextureConverter.cpp" />
     <ClCompile Include="Src\VertexManager.cpp" />
-    <ClCompile Include="Src\VertexShaderCache.cpp" />
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="Src\FramebufferManager.h" />
@@ -203,4 +201,4 @@
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>
-</Project>
\ No newline at end of file
+</Project>
diff --git a/Source/Core/VideoBackends/OGL/Src/PixelShaderCache.cpp b/Source/Core/VideoBackends/OGL/Src/PixelShaderCache.cpp
deleted file mode 100644
index 3fac62bd73..0000000000
--- a/Source/Core/VideoBackends/OGL/Src/PixelShaderCache.cpp
+++ /dev/null
@@ -1,104 +0,0 @@
-// Copyright 2013 Dolphin Emulator Project
-// Licensed under GPLv2
-// Refer to the license.txt file included.
-
-#include "Globals.h"
-
-#include "GLUtil.h"
-
-#include <cmath>
-
-#include "Statistics.h"
-#include "VideoConfig.h"
-#include "ImageWrite.h"
-#include "Common.h"
-#include "Render.h"
-#include "VertexShaderGen.h"
-#include "ProgramShaderCache.h"
-#include "PixelShaderManager.h"
-#include "OnScreenDisplay.h"
-#include "StringUtil.h"
-#include "FileUtil.h"
-#include "Debugger.h"
-
-namespace OGL
-{
-
-void SetPSConstant4fvByName(const char * name, unsigned int offset, const float *f, const unsigned int count = 1)
-{
-	ProgramShaderCache::PCacheEntry tmp = ProgramShaderCache::GetShaderProgram();
-	for (int a = 0; a < NUM_UNIFORMS; ++a)
-	{
-		if (!strcmp(name, UniformNames[a]))
-		{
-			if (tmp.shader.UniformLocations[a] == -1)
-				return;
-			else if (tmp.shader.UniformSize[a] <= offset)
-				return;
-			else
-			{
-				unsigned int maxcount= tmp.shader.UniformSize[a]-offset;
-				glUniform4fv(tmp.shader.UniformLocations[a] + offset, std::min(count, maxcount), f);
-				return;
-			}
-		}
-	}
-}
-
-// Renderer functions
-void Renderer::SetPSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4)
-{
-	float const f[4] = {f1, f2, f3, f4};
-
-	if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-	{
-		ProgramShaderCache::SetMultiPSConstant4fv(const_number, f, 1);
-		return;
-	}
-	for (unsigned int a = 0; a < 10; ++a)
-	{
-		if (const_number >= PSVar_Loc[a].reg && const_number < (PSVar_Loc[a].reg + PSVar_Loc[a].size))
-		{
-			unsigned int offset = const_number - PSVar_Loc[a].reg;
-			SetPSConstant4fvByName(PSVar_Loc[a].name, offset, f);
-			return;
-		}
-	}
-}
-
-void Renderer::SetPSConstant4fv(unsigned int const_number, const float *f)
-{
-	if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-	{
-		ProgramShaderCache::SetMultiPSConstant4fv(const_number, f, 1);
-		return;
-	}
-	for (unsigned int a = 0; a < 10; ++a)
-	{
-		if (const_number >= PSVar_Loc[a].reg && const_number < (PSVar_Loc[a].reg + PSVar_Loc[a].size))
-		{
-			unsigned int offset = const_number - PSVar_Loc[a].reg;
-			SetPSConstant4fvByName(PSVar_Loc[a].name, offset, f);
-			return;
-		}
-	}
-}
-
-void Renderer::SetMultiPSConstant4fv(unsigned int const_number, unsigned int count, const float *f)
-{
-	if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-	{
-		ProgramShaderCache::SetMultiPSConstant4fv(const_number, f, count);
-		return;
-	}
-	for (unsigned int a = 0; a < 10; ++a)
-	{
-		if (const_number >= PSVar_Loc[a].reg && const_number < (PSVar_Loc[a].reg + PSVar_Loc[a].size))
-		{
-			unsigned int offset = const_number - PSVar_Loc[a].reg;
-			SetPSConstant4fvByName(PSVar_Loc[a].name, offset, f, count);
-			return;
-		}
-	}
-}
-}  // namespace OGL
diff --git a/Source/Core/VideoBackends/OGL/Src/ProgramShaderCache.cpp b/Source/Core/VideoBackends/OGL/Src/ProgramShaderCache.cpp
index 94b9502d2f..51f41c063b 100644
--- a/Source/Core/VideoBackends/OGL/Src/ProgramShaderCache.cpp
+++ b/Source/Core/VideoBackends/OGL/Src/ProgramShaderCache.cpp
@@ -10,18 +10,16 @@
 #include "Statistics.h"
 #include "ImageWrite.h"
 #include "Render.h"
+#include "PixelShaderManager.h"
+#include "VertexShaderManager.h"
 
 namespace OGL
 {
 
 static const u32 UBO_LENGTH = 32*1024*1024;
 
-GLintptr ProgramShaderCache::s_vs_data_size;
-GLintptr ProgramShaderCache::s_ps_data_size;
-GLintptr ProgramShaderCache::s_vs_data_offset;
-u8 *ProgramShaderCache::s_ubo_buffer;
 u32 ProgramShaderCache::s_ubo_buffer_size;
-bool ProgramShaderCache::s_ubo_dirty;
+s32 ProgramShaderCache::s_ubo_align;
 
 static StreamBuffer *s_buffer;
 static int num_failures = 0;
@@ -36,6 +34,10 @@ UidChecker<VertexShaderUid,VertexShaderCode> ProgramShaderCache::vertex_uid_chec
 
 static char s_glsl_header[1024] = "";
 
+
+
+// Annoying sure, can be removed once we drop our UBO workaround
+
 const char *UniformNames[NUM_UNIFORMS] =
 {
 	// PIXEL SHADER UNIFORMS
@@ -61,6 +63,33 @@ const char *UniformNames[NUM_UNIFORMS] =
 	I_DEPTHPARAMS,
 };
 
+const static int PSVar_Loc[] = {
+	offsetof(PixelShaderConstants, colors)/16,
+	offsetof(PixelShaderConstants, kcolors)/16,
+	offsetof(PixelShaderConstants, alpha)/16,
+	offsetof(PixelShaderConstants, texdims)/16,
+	offsetof(PixelShaderConstants, zbias)/16,
+	offsetof(PixelShaderConstants, indtexscale)/16,
+	offsetof(PixelShaderConstants, indtexmtx)/16,
+	offsetof(PixelShaderConstants, fog)/16,
+	offsetof(PixelShaderConstants, plights)/16,
+	offsetof(PixelShaderConstants, pmaterials)/16,
+};
+
+const static int VSVar_Loc[] = {
+	offsetof(VertexShaderConstants, posnormalmatrix)/16,
+	offsetof(VertexShaderConstants, projection)/16,
+	offsetof(VertexShaderConstants, materials)/16,
+	offsetof(VertexShaderConstants, lights)/16,
+	offsetof(VertexShaderConstants, texmatrices)/16,
+	offsetof(VertexShaderConstants, transformmatrices)/16,
+	offsetof(VertexShaderConstants, normalmatrices)/16,
+	offsetof(VertexShaderConstants, posttransformmatrices)/16,
+	offsetof(VertexShaderConstants, depthparams)/16,
+};
+
+// End of UBO workaround
+
 void SHADER::SetProgramVariables()
 {
 	// glsl shader must be bind to set samplers
@@ -162,30 +191,43 @@ void SHADER::Bind()
 	}
 }
 
-
-void ProgramShaderCache::SetMultiPSConstant4fv(unsigned int offset, const float *f, unsigned int count)
-{
-	s_ubo_dirty = true;
-	memcpy(s_ubo_buffer+(offset*4*sizeof(float)), f, count*4*sizeof(float));
-}
-
-void ProgramShaderCache::SetMultiVSConstant4fv(unsigned int offset, const float *f, unsigned int count)
-{
-	s_ubo_dirty = true;
-	memcpy(s_ubo_buffer+(offset*4*sizeof(float))+s_vs_data_offset, f, count*4*sizeof(float));
-}
-
 void ProgramShaderCache::UploadConstants()
 {
-	if(s_ubo_dirty) {
-		s_buffer->Alloc(s_ubo_buffer_size);
-		size_t offset = s_buffer->Upload(s_ubo_buffer, s_ubo_buffer_size);
-		glBindBufferRange(GL_UNIFORM_BUFFER, 1, s_buffer->getBuffer(), offset, s_ps_data_size);
-		glBindBufferRange(GL_UNIFORM_BUFFER, 2, s_buffer->getBuffer(), offset + s_vs_data_offset, s_vs_data_size);
-		s_ubo_dirty = false;
+	if(g_ActiveConfig.backend_info.bSupportsGLSLUBO)
+	{
+		if(PixelShaderManager::dirty || VertexShaderManager::dirty)
+		{
+			s_buffer->Alloc(s_ubo_buffer_size);
+			
+			size_t offset = s_buffer->Upload((u8*)&PixelShaderManager::constants, ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align));
+			glBindBufferRange(GL_UNIFORM_BUFFER, 1, s_buffer->getBuffer(), offset, sizeof(PixelShaderConstants));
+			offset = s_buffer->Upload((u8*)&VertexShaderManager::constants, ROUND_UP(sizeof(VertexShaderConstants), s_ubo_align));
+			glBindBufferRange(GL_UNIFORM_BUFFER, 2, s_buffer->getBuffer(), offset, sizeof(VertexShaderConstants));
+	
+			PixelShaderManager::dirty = false;
+			VertexShaderManager::dirty = false;
+			
+			ADDSTAT(stats.thisFrame.bytesUniformStreamed, s_ubo_buffer_size);
+		}
+	}
+	else
+	{
+		// UBO workaround
+		// this must be updated per shader switch, so also update it when it's not dirty
+		for (unsigned int a = 0; a < 10; ++a)
+		{
+			if(last_entry->shader.UniformSize[a] > 0)
+				glUniform4fv(last_entry->shader.UniformLocations[a], last_entry->shader.UniformSize[a], (float*) &PixelShaderManager::constants + 4*PSVar_Loc[a]);
+		}
+		for (unsigned int a = 0; a < 9; ++a)
+		{
+			if(last_entry->shader.UniformSize[a+10] > 0)
+				glUniform4fv(last_entry->shader.UniformLocations[a+10], last_entry->shader.UniformSize[a+10], (float*) &VertexShaderManager::constants + 4*VSVar_Loc[a]);
+		}
 		
 		ADDSTAT(stats.thisFrame.bytesUniformStreamed, s_ubo_buffer_size);
 	}
+	
 }
 
 GLuint ProgramShaderCache::GetCurrentProgram(void)
@@ -419,22 +461,14 @@ void ProgramShaderCache::Init(void)
 	// then the UBO will fail.
 	if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
 	{
-		GLint Align;
-		glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &Align);
+		glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &s_ubo_align);
 
-		s_ps_data_size = C_PENVCONST_END * sizeof(float) * 4;
-		s_vs_data_size = C_VENVCONST_END * sizeof(float) * 4;
-		s_vs_data_offset = ROUND_UP(s_ps_data_size, Align);
-		s_ubo_buffer_size = ROUND_UP(s_ps_data_size, Align) + ROUND_UP(s_vs_data_size, Align);
+		s_ubo_buffer_size = ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align) + ROUND_UP(sizeof(VertexShaderConstants), s_ubo_align);
 
 		// We multiply by *4*4 because we need to get down to basic machine units.
 		// So multiply by four to get how many floats we have from vec4s
 		// Then once more to get bytes
 		s_buffer = new StreamBuffer(GL_UNIFORM_BUFFER, UBO_LENGTH);
-		
-		s_ubo_buffer = new u8[s_ubo_buffer_size];
-		memset(s_ubo_buffer, 0, s_ubo_buffer_size);
-		s_ubo_dirty = true;
 	}
 
 	// Read our shader cache, only if supported
@@ -509,8 +543,6 @@ void ProgramShaderCache::Shutdown(void)
 	{
 		delete s_buffer;
 		s_buffer = 0;
-		delete [] s_ubo_buffer;
-		s_ubo_buffer = 0;
 	}
 }
 
diff --git a/Source/Core/VideoBackends/OGL/Src/ProgramShaderCache.h b/Source/Core/VideoBackends/OGL/Src/ProgramShaderCache.h
index 6c1350fe34..d0c792aaa0 100644
--- a/Source/Core/VideoBackends/OGL/Src/ProgramShaderCache.h
+++ b/Source/Core/VideoBackends/OGL/Src/ProgramShaderCache.h
@@ -87,9 +87,6 @@ public:
 	
 	static bool CompileShader(SHADER &shader, const char* vcode, const char* pcode);
 	static GLuint CompileSingleShader(GLuint type, const char *code);
-
-	static void SetMultiPSConstant4fv(unsigned int offset, const float *f, unsigned int count);
-	static void SetMultiVSConstant4fv(unsigned int offset, const float *f, unsigned int count);
 	static void UploadConstants();
 
 	static void Init(void);
@@ -110,12 +107,8 @@ private:
 	static UidChecker<PixelShaderUid,PixelShaderCode> pixel_uid_checker;
 	static UidChecker<VertexShaderUid,VertexShaderCode> vertex_uid_checker;
 
-	static GLintptr s_vs_data_size;
-	static GLintptr s_ps_data_size;
-	static GLintptr s_vs_data_offset;
-	static u8 *s_ubo_buffer;
 	static u32 s_ubo_buffer_size;
-	static bool s_ubo_dirty;
+	static s32 s_ubo_align;
 };
 
 }  // namespace OGL
diff --git a/Source/Core/VideoBackends/OGL/Src/Render.h b/Source/Core/VideoBackends/OGL/Src/Render.h
index 4086ad3e87..8c6a8ca8f8 100644
--- a/Source/Core/VideoBackends/OGL/Src/Render.h
+++ b/Source/Core/VideoBackends/OGL/Src/Render.h
@@ -81,15 +81,6 @@ public:
 
 	bool SaveScreenshot(const std::string &filename, const TargetRectangle &rc);
 
-	void SetPSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4);
-	void SetPSConstant4fv(unsigned int const_number, const float *f);
-	void SetMultiPSConstant4fv(unsigned int const_number, unsigned int count, const float *f);
-
-	void SetVSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4);
-	void SetVSConstant4fv(unsigned int const_number, const float *f);
-	void SetMultiVSConstant3fv(unsigned int const_number, unsigned int count, const float *f);
-	void SetMultiVSConstant4fv(unsigned int const_number, unsigned int count, const float *f);
-
 private:
 	void UpdateEFBCache(EFBAccessType type, u32 cacheRectIdx, const EFBRectangle& efbPixelRc, const TargetRectangle& targetPixelRc, const u32* data);
 };
diff --git a/Source/Core/VideoBackends/OGL/Src/TextureConverter.cpp b/Source/Core/VideoBackends/OGL/Src/TextureConverter.cpp
index 693784054e..b2104af66a 100644
--- a/Source/Core/VideoBackends/OGL/Src/TextureConverter.cpp
+++ b/Source/Core/VideoBackends/OGL/Src/TextureConverter.cpp
@@ -324,7 +324,7 @@ int EncodeToRamFromTexture(u32 address,GLuint source_texture, bool bFromZBuffer,
 	};
 
 	texconv_shader.Bind();
-	glUniform4fv(texconv_shader.UniformLocations[C_COLORS], 2, params);
+	glUniform4fv(texconv_shader.UniformLocations[0], 2, params);
 
 	TargetRectangle scaledSource;
 	scaledSource.top = 0;
diff --git a/Source/Core/VideoBackends/OGL/Src/VertexShaderCache.cpp b/Source/Core/VideoBackends/OGL/Src/VertexShaderCache.cpp
deleted file mode 100644
index 8e2c048f8d..0000000000
--- a/Source/Core/VideoBackends/OGL/Src/VertexShaderCache.cpp
+++ /dev/null
@@ -1,130 +0,0 @@
-// Copyright 2013 Dolphin Emulator Project
-// Licensed under GPLv2
-// Refer to the license.txt file included.
-
-#include <math.h>
-
-#include "Globals.h"
-#include "VideoConfig.h"
-#include "Statistics.h"
-
-#include "GLUtil.h"
-
-#include "Render.h"
-#include "VertexShaderGen.h"
-#include "VertexShaderManager.h"
-#include "ProgramShaderCache.h"
-#include "VertexManager.h"
-#include "VertexLoader.h"
-#include "XFMemory.h"
-#include "ImageWrite.h"
-#include "FileUtil.h"
-#include "Debugger.h"
-
-namespace OGL
-{
-
-void SetVSConstant4fvByName(const char * name, unsigned int offset, const float *f, const unsigned int count = 1)
-{
-	ProgramShaderCache::PCacheEntry tmp = ProgramShaderCache::GetShaderProgram();
-	for (int a = 0; a < NUM_UNIFORMS; ++a)
-	{
-		if (!strcmp(name, UniformNames[a]))
-		{
-			if (tmp.shader.UniformLocations[a] == -1)
-				return;
-			else if (tmp.shader.UniformSize[a] <= offset)
-				return;
-			else
-			{
-				unsigned int maxcount= tmp.shader.UniformSize[a]-offset;
-				glUniform4fv(tmp.shader.UniformLocations[a] + offset, std::min(count, maxcount), f);
-				return;
-			}
-		}
-	}
-}
-
-void Renderer::SetVSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4)
-{
-	float const buf[4] = {f1, f2, f3, f4};
-
-	if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-	{
-		ProgramShaderCache::SetMultiVSConstant4fv(const_number, buf, 1);
-		return;
-	}
-	for (unsigned int a = 0; a < 9; ++a)
-	{
-		if (const_number >= VSVar_Loc[a].reg && const_number < ( VSVar_Loc[a].reg + VSVar_Loc[a].size))
-		{
-			unsigned int offset = const_number - VSVar_Loc[a].reg;
-			SetVSConstant4fvByName(VSVar_Loc[a].name, offset, buf);
-			return;
-		}
-	}
-}
-
-void Renderer::SetVSConstant4fv(unsigned int const_number, const float *f)
-{
-	if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-	{
-		ProgramShaderCache::SetMultiVSConstant4fv(const_number, f, 1);
-		return;
-	}
-	for (unsigned int a = 0; a < 9; ++a)
-	{
-		if (const_number >= VSVar_Loc[a].reg && const_number < ( VSVar_Loc[a].reg + VSVar_Loc[a].size))
-		{
-			unsigned int offset = const_number - VSVar_Loc[a].reg;
-			SetVSConstant4fvByName(VSVar_Loc[a].name, offset, f);
-			return;
-		}
-	}
-}
-
-void Renderer::SetMultiVSConstant4fv(unsigned int const_number, unsigned int count, const float *f)
-{
-	if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-	{
-		ProgramShaderCache::SetMultiVSConstant4fv(const_number, f, count);
-		return;
-	}
-	for (unsigned int a = 0; a < 9; ++a)
-	{
-		if (const_number >= VSVar_Loc[a].reg && const_number < ( VSVar_Loc[a].reg + VSVar_Loc[a].size))
-		{
-			unsigned int offset = const_number - VSVar_Loc[a].reg;
-			SetVSConstant4fvByName(VSVar_Loc[a].name, offset, f, count);
-			return;
-		}
-	}
-}
-
-void Renderer::SetMultiVSConstant3fv(unsigned int const_number, unsigned int count, const float *f)
-{
-	float buf[4 * C_VENVCONST_END];
-	for (unsigned int i = 0; i < count; i++)
-	{
-		buf[4*i  ] = *f++;
-		buf[4*i+1] = *f++;
-		buf[4*i+2] = *f++;
-		buf[4*i+3] = 0.f;
-	}
-	if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-	{
-		ProgramShaderCache::SetMultiVSConstant4fv(const_number, buf, count);
-		return;
-	}
-	for (unsigned int a = 0; a < 9; ++a)
-	{
-		if (const_number >= VSVar_Loc[a].reg && const_number < ( VSVar_Loc[a].reg + VSVar_Loc[a].size))
-		{
-			unsigned int offset = const_number - VSVar_Loc[a].reg;
-			SetVSConstant4fvByName(VSVar_Loc[a].name, offset, buf, count);
-			return;
-		}
-	}
-}
-
-}  // namespace OGL