ogl: implement useful constant buffer upload

this will remove the additional memcpy introduced in my last commit

9 files changed, 75 insertions, 110 deletions

diff --git a/Source/Core/VideoBackends/OGL/Src/PixelShaderCache.cpp b/Source/Core/VideoBackends/OGL/Src/PixelShaderCache.cpp
index 0635003dfd..ba6a0b2e97 100644
--- a/Source/Core/VideoBackends/OGL/Src/PixelShaderCache.cpp
+++ b/Source/Core/VideoBackends/OGL/Src/PixelShaderCache.cpp
@@ -27,20 +27,5 @@ namespace OGL
 // Renderer functions
 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;
-	}
-	
-	ProgramShaderCache::PCacheEntry tmp = ProgramShaderCache::GetShaderProgram();
-	for (unsigned int a = 0; a < 10; ++a)
-	{
-		u32 offset = PSVar_Loc[a].reg - const_number;
-		if(offset >= count) return;
-		u32 size = std::min(tmp.shader.UniformSize[a], count-offset);
-		if(size > 0)
-			glUniform4fv(tmp.shader.UniformLocations[a], size, f + 4*offset);
-	}
 }
 }  // namespace OGL
diff --git a/Source/Core/VideoBackends/OGL/Src/ProgramShaderCache.cpp b/Source/Core/VideoBackends/OGL/Src/ProgramShaderCache.cpp
index 94b9502d2f..0904ef41eb 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 +63,37 @@ const char *UniformNames[NUM_UNIFORMS] =
 	I_DEPTHPARAMS,
 };
 
+struct s_svar
+{
+	const unsigned int reg;
+	const unsigned int size;
+};
+
+const s_svar PSVar_Loc[] = { {C_COLORS, 4 },
+						{C_KCOLORS, 4 },
+						{C_ALPHA, 1 },
+						{C_TEXDIMS, 8 },
+						{C_ZBIAS, 2  },
+						{C_INDTEXSCALE, 2  },
+						{C_INDTEXMTX, 6 },
+						{C_FOG, 3 },
+						{C_PLIGHTS, 40 },
+						{C_PMATERIALS, 4 },
+						};
+
+const s_svar VSVar_Loc[] = {  {C_POSNORMALMATRIX, 6 },
+						{C_PROJECTION, 4  },
+						{C_MATERIALS, 4 },
+						{C_LIGHTS, 40 },
+						{C_TEXMATRICES, 24 },
+						{C_TRANSFORMMATRICES, 64  },
+						{C_NORMALMATRICES, 32  },
+						{C_POSTTRANSFORMMATRICES, 64 },
+						{C_DEPTHPARAMS, 1 },
+						};
+
+// End of UBO workaround
+
 void SHADER::SetProgramVariables()
 {
 	// glsl shader must be bind to set samplers
@@ -162,30 +195,43 @@ 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].reg);
+		}
+		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].reg);
+		}
 		
 		ADDSTAT(stats.thisFrame.bytesUniformStreamed, s_ubo_buffer_size);
 	}
+	
 }
 
 GLuint ProgramShaderCache::GetCurrentProgram(void)
@@ -419,22 +465,14 @@ 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 +547,6 @@ 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/VertexShaderCache.cpp b/Source/Core/VideoBackends/OGL/Src/VertexShaderCache.cpp
index 09f2368bad..c2c49b8505 100644
--- a/Source/Core/VideoBackends/OGL/Src/VertexShaderCache.cpp
+++ b/Source/Core/VideoBackends/OGL/Src/VertexShaderCache.cpp
@@ -26,20 +26,6 @@ namespace OGL
 
 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;
-	}
-	ProgramShaderCache::PCacheEntry tmp = ProgramShaderCache::GetShaderProgram();
-	for (unsigned int a = 0; a < 9; ++a)
-	{
-		u32 offset = VSVar_Loc[a].reg - const_number;
-		if(offset >= count) return;
-		u32 size = std::min(tmp.shader.UniformSize[a+10], count-offset);
-		if(size > 0)
-			glUniform4fv(tmp.shader.UniformLocations[a+10], size, f + 4*offset);
-	}
 }
 
 }  // namespace OGL
diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.h b/Source/Core/VideoCommon/Src/PixelShaderGen.h
index d9ce8c5874..b73b17168c 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.h
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.h
@@ -43,19 +43,6 @@ enum DSTALPHA_MODE
 	DSTALPHA_DUAL_SOURCE_BLEND // Use dual-source blending
 };
 
-// Annoying sure, can be removed once we get up to GLSL ~1.3
-const s_svar PSVar_Loc[] = { {C_COLORS, 4 },
-						{C_KCOLORS, 4 },
-						{C_ALPHA, 1 },
-						{C_TEXDIMS, 8 },
-						{C_ZBIAS, 2  },
-						{C_INDTEXSCALE, 2  },
-						{C_INDTEXMTX, 6 },
-						{C_FOG, 3 },
-						{C_PLIGHTS, 40 },
-						{C_PMATERIALS, 4 },
-						};
-
 #pragma pack(1)
 struct pixel_shader_uid_data
 {
diff --git a/Source/Core/VideoCommon/Src/PixelShaderManager.cpp b/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
index 3a13e287aa..970bdb6e58 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
+++ b/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
@@ -95,9 +95,6 @@ void PixelShaderManager::Shutdown()
 
 void PixelShaderManager::SetConstants(u32 components)
 {
-	if (g_ActiveConfig.backend_info.APIType == API_OPENGL && !g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-		dirty = true;
-	
 	for (int i = 0; i < 2; ++i)
 	{
 		if (s_nColorsChanged[i])
@@ -359,7 +356,7 @@ void PixelShaderManager::SetConstants(u32 components)
 		}
 	}
 	
-	if(dirty)
+	if(dirty && g_ActiveConfig.backend_info.APIType != API_OPENGL)
 	{
 		g_renderer->SetMultiPSConstant4fv(0, sizeof(constants)/16, (float*) &constants);
 		dirty = false;
diff --git a/Source/Core/VideoCommon/Src/VertexShaderGen.h b/Source/Core/VideoCommon/Src/VertexShaderGen.h
index c5f32c741d..667df311c2 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderGen.h
+++ b/Source/Core/VideoCommon/Src/VertexShaderGen.h
@@ -53,17 +53,6 @@
 #define C_DEPTHPARAMS           (C_POSTTRANSFORMMATRICES + 64)
 #define C_VENVCONST_END			(C_DEPTHPARAMS + 1)
 
-const s_svar VSVar_Loc[] = {  {C_POSNORMALMATRIX, 6 },
-						{C_PROJECTION, 4  },
-						{C_MATERIALS, 4 },
-						{C_LIGHTS, 40 },
-						{C_TEXMATRICES, 24 },
-						{C_TRANSFORMMATRICES, 64  },
-						{C_NORMALMATRICES, 32  },
-						{C_POSTTRANSFORMMATRICES, 64 },
-						{C_DEPTHPARAMS, 1 },
-						};
-
 #pragma pack(1)
 
 struct vertex_shader_uid_data
diff --git a/Source/Core/VideoCommon/Src/VertexShaderManager.cpp b/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
index 016fb517e4..20c4994cf8 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
+++ b/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
@@ -225,9 +225,6 @@ void VertexShaderManager::Dirty()
 // TODO: A cleaner way to control the matrices without making a mess in the parameters field
 void VertexShaderManager::SetConstants()
 {
-	if (g_ActiveConfig.backend_info.APIType == API_OPENGL && !g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-		dirty = true;
-
 	if (nTransformMatricesChanged[0] >= 0)
 	{
 		int startn = nTransformMatricesChanged[0] / 4;
@@ -521,7 +518,7 @@ void VertexShaderManager::SetConstants()
 		}
 	}
 	
-	if(dirty)
+	if(dirty && g_ActiveConfig.backend_info.APIType != API_OPENGL)
 	{
 		dirty = false;
 		g_renderer->SetMultiVSConstant4fv(0, sizeof(constants)/16, (float*) &constants);	
diff --git a/Source/Core/VideoCommon/Src/VideoCommon.h b/Source/Core/VideoCommon/Src/VideoCommon.h
index 54a2504631..ef36b1fed6 100644
--- a/Source/Core/VideoCommon/Src/VideoCommon.h
+++ b/Source/Core/VideoCommon/Src/VideoCommon.h
@@ -132,10 +132,5 @@ inline unsigned int GetPow2(unsigned int val)
 		++ret;
 	return ret;
 }
-struct s_svar
-{
-	const unsigned int reg;
-	const unsigned int size;
-};
 
 #endif  // _VIDEOCOMMON_H