Skip to content

Commit

Permalink
Merge pull request #14 from degasus/uboWorkaroundRemove
Browse files Browse the repository at this point in the history 
OGL: Remove non-UBO code path.
  • Loading branch information
@neobrain
neobrain committed Feb 4, 2014
2 parents 0008943 + 6089e44 commit af24ed7
Show file tree
Hide file tree
Showing 8 changed files with 67 additions and 180 deletions.
157 changes: 23 additions & 134 deletions Source/Core/VideoBackends/OGL/ProgramShaderCache.cpp
View file
Expand Up @@ -34,69 +34,13 @@ 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
I_COLORS,
I_KCOLORS,
I_ALPHA,
I_TEXDIMS,
I_ZBIAS ,
I_INDTEXSCALE ,
I_INDTEXMTX,
I_FOG,
I_PLIGHTS,
I_PMATERIALS,
// VERTEX SHADER UNIFORMS
I_POSNORMALMATRIX,
I_PROJECTION ,
I_MATERIALS,
I_LIGHTS,
I_TEXMATRICES,
I_TRANSFORMMATRICES ,
I_NORMALMATRICES ,
I_POSTTRANSFORMMATRICES,
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
Bind();

// Bind UBO
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO && !g_ActiveConfig.backend_info.bSupportShadingLanguage420pack)
if (!g_ActiveConfig.backend_info.bSupportShadingLanguage420pack)
{
GLint PSBlock_id = glGetUniformBlockIndex(glprogid, "PSBlock");
GLint VSBlock_id = glGetUniformBlockIndex(glprogid, "VSBlock");
Expand All @@ -107,34 +51,6 @@ void SHADER::SetProgramVariables()
glUniformBlockBinding(glprogid, VSBlock_id, 2);
}

// UBO workaround
if(!g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
for (int a = 0; a < NUM_UNIFORMS; ++a)
{
UniformLocations[a] = glGetUniformLocation(glprogid, UniformNames[a]);
UniformSize[a] = 0;
}

int max_uniforms = 0;
char name[50];
int size;

glGetProgramiv(glprogid, GL_ACTIVE_UNIFORMS, &max_uniforms);
for(int i=0; i<max_uniforms; i++)
{
glGetActiveUniform(glprogid, i, sizeof(name), NULL, &size, NULL, name);
for (int a = 0; a < NUM_UNIFORMS; ++a)
{
if(strstr(name, UniformNames[a]))
{
UniformSize[a] = size;
break;
}
}
}
}

// Bind Texture Sampler
for (int a = 0; a <= 9; ++a)
{
Expand Down Expand Up @@ -189,48 +105,27 @@ void SHADER::Bind()

void ProgramShaderCache::UploadConstants()
{
if(g_ActiveConfig.backend_info.bSupportsGLSLUBO)
if(PixelShaderManager::dirty || VertexShaderManager::dirty)
{
if(PixelShaderManager::dirty || VertexShaderManager::dirty)
{
auto buffer = s_buffer->Map(s_ubo_buffer_size, s_ubo_align);
auto buffer = s_buffer->Map(s_ubo_buffer_size, s_ubo_align);

memcpy(buffer.first,
&PixelShaderManager::constants, sizeof(PixelShaderConstants));
memcpy(buffer.first,
&PixelShaderManager::constants, sizeof(PixelShaderConstants));

memcpy(buffer.first + ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align),
&VertexShaderManager::constants, sizeof(VertexShaderConstants));
memcpy(buffer.first + ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align),
&VertexShaderManager::constants, sizeof(VertexShaderConstants));

s_buffer->Unmap(s_ubo_buffer_size);
glBindBufferRange(GL_UNIFORM_BUFFER, 1, s_buffer->m_buffer, buffer.second,
sizeof(PixelShaderConstants));
glBindBufferRange(GL_UNIFORM_BUFFER, 2, s_buffer->m_buffer, buffer.second + ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align),
sizeof(VertexShaderConstants));
s_buffer->Unmap(s_ubo_buffer_size);
glBindBufferRange(GL_UNIFORM_BUFFER, 1, s_buffer->m_buffer, buffer.second,
sizeof(PixelShaderConstants));
glBindBufferRange(GL_UNIFORM_BUFFER, 2, s_buffer->m_buffer, buffer.second + ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align),
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]);
}
PixelShaderManager::dirty = false;
VertexShaderManager::dirty = false;

ADDSTAT(stats.thisFrame.bytesUniformStreamed, s_ubo_buffer_size);
}

}

GLuint ProgramShaderCache::GetCurrentProgram(void)
Expand Down Expand Up @@ -462,17 +357,14 @@ void ProgramShaderCache::Init(void)
// We have to get the UBO alignment here because
// if we generate a buffer that isn't aligned
// then the UBO will fail.
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &s_ubo_align);
glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &s_ubo_align);

s_ubo_buffer_size = ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align) + ROUND_UP(sizeof(VertexShaderConstants), s_ubo_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 = StreamBuffer::Create(GL_UNIFORM_BUFFER, UBO_LENGTH);
}
// 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 = StreamBuffer::Create(GL_UNIFORM_BUFFER, UBO_LENGTH);

// Read our shader cache, only if supported
if (g_ogl_config.bSupportsGLSLCache && !g_Config.bEnableShaderDebugging)
Expand Down Expand Up @@ -542,11 +434,8 @@ void ProgramShaderCache::Shutdown(void)
pixel_uid_checker.Invalidate();
vertex_uid_checker.Invalidate();

if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
delete s_buffer;
s_buffer = 0;
}
delete s_buffer;
s_buffer = 0;
}

void ProgramShaderCache::CreateHeader ( void )
Expand Down Expand Up @@ -583,7 +472,7 @@ void ProgramShaderCache::CreateHeader ( void )
"%s\n"

, v==GLSLES3 ? "#version 300 es" : v==GLSL_130 ? "#version 130" : v==GLSL_140 ? "#version 140" : "#version 150"
, g_ActiveConfig.backend_info.bSupportsGLSLUBO && v<GLSL_140 ? "#extension GL_ARB_uniform_buffer_object : enable" : ""
, v<GLSL_140 ? "#extension GL_ARB_uniform_buffer_object : enable" : ""
, g_ActiveConfig.backend_info.bSupportsEarlyZ ? "#extension GL_ARB_shader_image_load_store : enable" : ""
, g_ActiveConfig.backend_info.bSupportShadingLanguage420pack ? "#extension GL_ARB_shading_language_420pack : enable" : ""

Expand Down
27 changes: 16 additions & 11 deletions Source/Core/VideoBackends/OGL/Render.cpp
View file
Expand Up @@ -420,6 +420,21 @@ Renderer::Renderer()
bSuccess = false;
}

if (!GLExtensions::Supports("GL_ARB_uniform_buffer_object"))
{
// ubo allow us to keep the current constants on shader switches
// we also can stream them much nicer and pack into it whatever we want to
PanicAlert("GPU: OGL ERROR: Need GL_ARB_uniform_buffer_object.\n"
"GPU: Does your video card support OpenGL 3.1?");
bSuccess = false;
}
else if (DriverDetails::HasBug(DriverDetails::BUG_BROKENUBO))
{
PanicAlert("Buggy GPU driver detected.\n"
"Please either install the closed-source GPU driver or update your Mesa 3D version.");
bSuccess = false;
}

if (!GLExtensions::Supports("GL_ARB_sampler_objects") && bSuccess)
{
// Our sampler cache uses this extension. It could easyly be workaround and it's by far the
Expand All @@ -440,7 +455,6 @@ Renderer::Renderer()
}

g_Config.backend_info.bSupportsDualSourceBlend = GLExtensions::Supports("GL_ARB_blend_func_extended");
g_Config.backend_info.bSupportsGLSLUBO = GLExtensions::Supports("GL_ARB_uniform_buffer_object") && !DriverDetails::HasBug(DriverDetails::BUG_ANNIHILATEDUBOS);
g_Config.backend_info.bSupportsPrimitiveRestart = !DriverDetails::HasBug(DriverDetails::BUG_PRIMITIVERESTART) &&
((GLExtensions::Version() >= 310) || GLExtensions::Supports("GL_NV_primitive_restart"));
g_Config.backend_info.bSupportsEarlyZ = GLExtensions::Supports("GL_ARB_shader_image_load_store");
Expand Down Expand Up @@ -520,24 +534,15 @@ Renderer::Renderer()
if(g_ogl_config.max_samples < 1)
g_ogl_config.max_samples = 1;

if(g_Config.backend_info.bSupportsGLSLUBO && DriverDetails::HasBug(DriverDetails::BUG_BROKENUBO))
{
g_Config.backend_info.bSupportsGLSLUBO = false;
ERROR_LOG(VIDEO, "Buggy driver detected. Disable UBO");
OSD::AddMessage("Major performance warning: Buggy GPU driver detected.", 20000);
OSD::AddMessage("Please either install the closed-source GPU driver or update your Mesa 3D version.", 20000);
}

UpdateActiveConfig();

OSD::AddMessage(StringFromFormat("Video Info: %s, %s, %s",
g_ogl_config.gl_vendor,
g_ogl_config.gl_renderer,
g_ogl_config.gl_version), 5000);

WARN_LOG(VIDEO,"Missing OGL Extensions: %s%s%s%s%s%s%s%s%s%s%s",
WARN_LOG(VIDEO,"Missing OGL Extensions: %s%s%s%s%s%s%s%s%s%s",
g_ActiveConfig.backend_info.bSupportsDualSourceBlend ? "" : "DualSourceBlend ",
g_ActiveConfig.backend_info.bSupportsGLSLUBO ? "" : "UniformBuffer ",
g_ActiveConfig.backend_info.bSupportsPrimitiveRestart ? "" : "PrimitiveRestart ",
g_ActiveConfig.backend_info.bSupportsEarlyZ ? "" : "EarlyZ ",
g_ogl_config.bSupportsGLPinnedMemory ? "" : "PinnedMemory ",
Expand Down
5 changes: 0 additions & 5 deletions Source/Core/VideoBackends/OGL/VertexManager.cpp
View file
Expand Up @@ -186,11 +186,6 @@ void VertexManager::vFlush()
if (useDstAlpha && !dualSourcePossible)
{
ProgramShaderCache::SetShader(DSTALPHA_ALPHA_PASS,g_nativeVertexFmt->m_components);
if (!g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
// Need to upload these again, if we don't support UBO
ProgramShaderCache::UploadConstants();
}

// only update alpha
glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_TRUE);
Expand Down
24 changes: 12 additions & 12 deletions Source/Core/VideoCommon/PixelShaderGen.cpp
View file
Expand Up @@ -283,23 +283,23 @@ static inline void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_T
}
out.Write("\n");

if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
if (ApiType == API_OPENGL)
out.Write("layout(std140%s) uniform PSBlock {\n", g_ActiveConfig.backend_info.bSupportShadingLanguage420pack ? ", binding = 1" : "");

DeclareUniform(out, ApiType, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_COLORS, "float4", I_COLORS"[4]");
DeclareUniform(out, ApiType, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_KCOLORS, "float4", I_KCOLORS"[4]");
DeclareUniform(out, ApiType, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_ALPHA, "float4", I_ALPHA"[1]"); // TODO: Why is this an array...-.-
DeclareUniform(out, ApiType, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_TEXDIMS, "float4", I_TEXDIMS"[8]");
DeclareUniform(out, ApiType, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_ZBIAS, "float4", I_ZBIAS"[2]");
DeclareUniform(out, ApiType, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_INDTEXSCALE, "float4", I_INDTEXSCALE"[2]");
DeclareUniform(out, ApiType, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_INDTEXMTX, "float4", I_INDTEXMTX"[6]");
DeclareUniform(out, ApiType, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_FOG, "float4", I_FOG"[3]");
DeclareUniform(out, ApiType, C_COLORS, "float4", I_COLORS"[4]");
DeclareUniform(out, ApiType, C_KCOLORS, "float4", I_KCOLORS"[4]");
DeclareUniform(out, ApiType, C_ALPHA, "float4", I_ALPHA"[1]"); // TODO: Why is this an array...-.-
DeclareUniform(out, ApiType, C_TEXDIMS, "float4", I_TEXDIMS"[8]");
DeclareUniform(out, ApiType, C_ZBIAS, "float4", I_ZBIAS"[2]");
DeclareUniform(out, ApiType, C_INDTEXSCALE, "float4", I_INDTEXSCALE"[2]");
DeclareUniform(out, ApiType, C_INDTEXMTX, "float4", I_INDTEXMTX"[6]");
DeclareUniform(out, ApiType, C_FOG, "float4", I_FOG"[3]");

// For pixel lighting - TODO: Should only be defined when per pixel lighting is enabled!
DeclareUniform(out, ApiType, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_PLIGHTS, "float4", I_PLIGHTS"[40]");
DeclareUniform(out, ApiType, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_PMATERIALS, "float4", I_PMATERIALS"[4]");
DeclareUniform(out, ApiType, C_PLIGHTS, "float4", I_PLIGHTS"[40]");
DeclareUniform(out, ApiType, C_PMATERIALS, "float4", I_PMATERIALS"[4]");

if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
if (ApiType == API_OPENGL)
out.Write("};\n");

if (ApiType == API_OPENGL)
Expand Down
8 changes: 4 additions & 4 deletions Source/Core/VideoCommon/ShaderGenCommon.h
View file
Expand Up @@ -174,18 +174,18 @@ static inline void WriteRegister(T& object, API_TYPE ApiType, const char *prefix
}

template<class T>
static inline void WriteLocation(T& object, API_TYPE ApiType, bool using_ubos)
static inline void WriteLocation(T& object, API_TYPE ApiType)
{
if (using_ubos)
if (ApiType == API_OPENGL)
return;

object.Write("uniform ");
}

template<class T>
static inline void DeclareUniform(T& object, API_TYPE api_type, bool using_ubos, const u32 num, const char* type, const char* name)
static inline void DeclareUniform(T& object, API_TYPE api_type, const u32 num, const char* type, const char* name)
{
WriteLocation(object, api_type, using_ubos);
WriteLocation(object, api_type);
object.Write("%s %s ", type, name);
WriteRegister(object, api_type, "c", num);
object.Write(";\n");
Expand Down
24 changes: 12 additions & 12 deletions Source/Core/VideoCommon/VertexShaderGen.cpp
View file
Expand Up @@ -83,20 +83,20 @@ static inline void GenerateVertexShader(T& out, u32 components, API_TYPE api_typ
_assert_(bpmem.genMode.numcolchans == xfregs.numChan.numColorChans);

// uniforms
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
if (api_type == API_OPENGL)
out.Write("layout(std140%s) uniform VSBlock {\n", g_ActiveConfig.backend_info.bSupportShadingLanguage420pack ? ", binding = 2" : "");

DeclareUniform(out, api_type, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_POSNORMALMATRIX, "float4", I_POSNORMALMATRIX"[6]");
DeclareUniform(out, api_type, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_PROJECTION, "float4", I_PROJECTION"[4]");
DeclareUniform(out, api_type, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_MATERIALS, "float4", I_MATERIALS"[4]");
DeclareUniform(out, api_type, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_LIGHTS, "float4", I_LIGHTS"[40]");
DeclareUniform(out, api_type, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_TEXMATRICES, "float4", I_TEXMATRICES"[24]");
DeclareUniform(out, api_type, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_TRANSFORMMATRICES, "float4", I_TRANSFORMMATRICES"[64]");
DeclareUniform(out, api_type, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_NORMALMATRICES, "float4", I_NORMALMATRICES"[32]");
DeclareUniform(out, api_type, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_POSTTRANSFORMMATRICES, "float4", I_POSTTRANSFORMMATRICES"[64]");
DeclareUniform(out, api_type, g_ActiveConfig.backend_info.bSupportsGLSLUBO, C_DEPTHPARAMS, "float4", I_DEPTHPARAMS);

if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
DeclareUniform(out, api_type, C_POSNORMALMATRIX, "float4", I_POSNORMALMATRIX"[6]");
DeclareUniform(out, api_type, C_PROJECTION, "float4", I_PROJECTION"[4]");
DeclareUniform(out, api_type, C_MATERIALS, "float4", I_MATERIALS"[4]");
DeclareUniform(out, api_type, C_LIGHTS, "float4", I_LIGHTS"[40]");
DeclareUniform(out, api_type, C_TEXMATRICES, "float4", I_TEXMATRICES"[24]");
DeclareUniform(out, api_type, C_TRANSFORMMATRICES, "float4", I_TRANSFORMMATRICES"[64]");
DeclareUniform(out, api_type, C_NORMALMATRICES, "float4", I_NORMALMATRICES"[32]");
DeclareUniform(out, api_type, C_POSTTRANSFORMMATRICES, "float4", I_POSTTRANSFORMMATRICES"[64]");
DeclareUniform(out, api_type, C_DEPTHPARAMS, "float4", I_DEPTHPARAMS);

if (api_type == API_OPENGL)
out.Write("};\n");

GenerateVSOutputStruct(out, api_type);
Expand Down
1 change: 0 additions & 1 deletion Source/Core/VideoCommon/VideoConfig.cpp
View file
Expand Up @@ -215,7 +215,6 @@ void VideoConfig::VerifyValidity()
if (!backend_info.bSupports3DVision) b3DVision = false;
if (!backend_info.bSupportsFormatReinterpretation) bEFBEmulateFormatChanges = false;
if (!backend_info.bSupportsPixelLighting) bEnablePixelLighting = false;
if (backend_info.APIType != API_OPENGL) backend_info.bSupportsGLSLUBO = false;
}

void VideoConfig::Save(const char *ini_file)
Expand Down
1 change: 0 additions & 1 deletion Source/Core/VideoCommon/VideoConfig.h
View file
Expand Up @@ -150,7 +150,6 @@ struct VideoConfig
bool bSupportsPrimitiveRestart;
bool bSupportsSeparateAlphaFunction;
bool bSupportsOversizedViewports;
bool bSupportsGLSLUBO; // needed by PixelShaderGen, so must stay in VideoCommon
bool bSupportsEarlyZ; // needed by PixelShaderGen, so must stay in VideoCommon
bool bSupportShadingLanguage420pack; // needed by ShaderGen, so must stay in VideoCommon
} backend_info;
Expand Down

0 comments on commit af24ed7

Please sign in to comment.