summaryrefslogtreecommitdiff
path: root/indra/llrender
diff options
context:
space:
mode:
Diffstat (limited to 'indra/llrender')
-rw-r--r--indra/llrender/llgl.cpp14
-rw-r--r--indra/llrender/llgl.h1
-rw-r--r--indra/llrender/llglheaders.h16
-rw-r--r--indra/llrender/llglslshader.cpp228
-rw-r--r--indra/llrender/llglslshader.h27
-rw-r--r--indra/llrender/llvertexbuffer.cpp14
6 files changed, 297 insertions, 3 deletions
diff --git a/indra/llrender/llgl.cpp b/indra/llrender/llgl.cpp
index 5429489075..81a97c7661 100644
--- a/indra/llrender/llgl.cpp
+++ b/indra/llrender/llgl.cpp
@@ -216,6 +216,11 @@ PFNGLGETQUERYIVARBPROC glGetQueryivARB = NULL;
PFNGLGETQUERYOBJECTIVARBPROC glGetQueryObjectivARB = NULL;
PFNGLGETQUERYOBJECTUIVARBPROC glGetQueryObjectuivARB = NULL;
+// GL_ARB_timer_query
+PFNGLQUERYCOUNTERPROC glQueryCounter = NULL;
+PFNGLGETQUERYOBJECTI64VPROC glGetQueryObjecti64v = NULL;
+PFNGLGETQUERYOBJECTUI64VPROC glGetQueryObjectui64v = NULL;
+
// GL_ARB_point_parameters
PFNGLPOINTPARAMETERFARBPROC glPointParameterfARB = NULL;
PFNGLPOINTPARAMETERFVARBPROC glPointParameterfvARB = NULL;
@@ -421,6 +426,7 @@ LLGLManager::LLGLManager() :
mHasFragmentShader(FALSE),
mNumTextureImageUnits(0),
mHasOcclusionQuery(FALSE),
+ mHasTimerQuery(FALSE),
mHasOcclusionQuery2(FALSE),
mHasPointParameters(FALSE),
mHasDrawBuffers(FALSE),
@@ -956,6 +962,7 @@ void LLGLManager::initExtensions()
mHasARBEnvCombine = ExtensionExists("GL_ARB_texture_env_combine", gGLHExts.mSysExts);
mHasCompressedTextures = glh_init_extensions("GL_ARB_texture_compression");
mHasOcclusionQuery = ExtensionExists("GL_ARB_occlusion_query", gGLHExts.mSysExts);
+ mHasTimerQuery = ExtensionExists("GL_ARB_timer_query", gGLHExts.mSysExts);
mHasOcclusionQuery2 = ExtensionExists("GL_ARB_occlusion_query2", gGLHExts.mSysExts);
mHasVertexBufferObject = ExtensionExists("GL_ARB_vertex_buffer_object", gGLHExts.mSysExts);
mHasVertexArrayObject = ExtensionExists("GL_ARB_vertex_array_object", gGLHExts.mSysExts);
@@ -1271,6 +1278,13 @@ void LLGLManager::initExtensions()
glGetQueryObjectivARB = (PFNGLGETQUERYOBJECTIVARBPROC)GLH_EXT_GET_PROC_ADDRESS("glGetQueryObjectivARB");
glGetQueryObjectuivARB = (PFNGLGETQUERYOBJECTUIVARBPROC)GLH_EXT_GET_PROC_ADDRESS("glGetQueryObjectuivARB");
}
+ if (mHasTimerQuery)
+ {
+ llinfos << "initExtensions() TimerQuery-related procs..." << llendl;
+ glQueryCounter = (PFNGLQUERYCOUNTERPROC) GLH_EXT_GET_PROC_ADDRESS("glQueryCounter");
+ glGetQueryObjecti64v = (PFNGLGETQUERYOBJECTI64VPROC) GLH_EXT_GET_PROC_ADDRESS("glGetQueryObjecti64v");
+ glGetQueryObjectui64v = (PFNGLGETQUERYOBJECTUI64VPROC) GLH_EXT_GET_PROC_ADDRESS("glGetQueryObjectui64v");
+ }
if (mHasPointParameters)
{
llinfos << "initExtensions() PointParameters-related procs..." << llendl;
diff --git a/indra/llrender/llgl.h b/indra/llrender/llgl.h
index 767d49cbf2..60597fd090 100644
--- a/indra/llrender/llgl.h
+++ b/indra/llrender/llgl.h
@@ -98,6 +98,7 @@ public:
BOOL mHasFragmentShader;
S32 mNumTextureImageUnits;
BOOL mHasOcclusionQuery;
+ BOOL mHasTimerQuery;
BOOL mHasOcclusionQuery2;
BOOL mHasPointParameters;
BOOL mHasDrawBuffers;
diff --git a/indra/llrender/llglheaders.h b/indra/llrender/llglheaders.h
index 509de51f4d..dace572953 100644
--- a/indra/llrender/llglheaders.h
+++ b/indra/llrender/llglheaders.h
@@ -116,6 +116,11 @@ extern PFNGLGETQUERYIVARBPROC glGetQueryivARB;
extern PFNGLGETQUERYOBJECTIVARBPROC glGetQueryObjectivARB;
extern PFNGLGETQUERYOBJECTUIVARBPROC glGetQueryObjectuivARB;
+// GL_ARB_timer_query
+extern PFNGLQUERYCOUNTERPROC glQueryCounter;
+extern PFNGLGETQUERYOBJECTI64VPROC glGetQueryObjecti64v;
+extern PFNGLGETQUERYOBJECTUI64VPROC glGetQueryObjectui64v;
+
// GL_ARB_point_parameters
extern PFNGLPOINTPARAMETERFARBPROC glPointParameterfARB;
extern PFNGLPOINTPARAMETERFVARBPROC glPointParameterfvARB;
@@ -378,6 +383,11 @@ extern PFNGLGETQUERYIVARBPROC glGetQueryivARB;
extern PFNGLGETQUERYOBJECTIVARBPROC glGetQueryObjectivARB;
extern PFNGLGETQUERYOBJECTUIVARBPROC glGetQueryObjectuivARB;
+// GL_ARB_timer_query
+extern PFNGLQUERYCOUNTERPROC glQueryCounter;
+extern PFNGLGETQUERYOBJECTI64VPROC glGetQueryObjecti64v;
+extern PFNGLGETQUERYOBJECTUI64VPROC glGetQueryObjectui64v;
+
// GL_ARB_point_parameters
extern PFNGLPOINTPARAMETERFARBPROC glPointParameterfARB;
extern PFNGLPOINTPARAMETERFVARBPROC glPointParameterfvARB;
@@ -619,6 +629,12 @@ extern PFNGLGETQUERYIVARBPROC glGetQueryivARB;
extern PFNGLGETQUERYOBJECTIVARBPROC glGetQueryObjectivARB;
extern PFNGLGETQUERYOBJECTUIVARBPROC glGetQueryObjectuivARB;
+// GL_ARB_timer_query
+extern PFNGLQUERYCOUNTERPROC glQueryCounter;
+extern PFNGLGETQUERYOBJECTI64VPROC glGetQueryObjecti64v;
+extern PFNGLGETQUERYOBJECTUI64VPROC glGetQueryObjectui64v;
+
+
// GL_ARB_point_parameters
extern PFNGLPOINTPARAMETERFARBPROC glPointParameterfARB;
extern PFNGLPOINTPARAMETERFVARBPROC glPointParameterfvARB;
diff --git a/indra/llrender/llglslshader.cpp b/indra/llrender/llglslshader.cpp
index 39ad9b9e9b..191c9862f0 100644
--- a/indra/llrender/llglslshader.cpp
+++ b/indra/llrender/llglslshader.cpp
@@ -53,6 +53,12 @@ GLhandleARB LLGLSLShader::sCurBoundShader = 0;
LLGLSLShader* LLGLSLShader::sCurBoundShaderPtr = NULL;
S32 LLGLSLShader::sIndexedTextureChannels = 0;
bool LLGLSLShader::sNoFixedFunction = false;
+bool LLGLSLShader::sProfileEnabled = false;
+std::set<LLGLSLShader*> LLGLSLShader::sInstances;
+U64 LLGLSLShader::sTotalTimeElapsed = 0;
+U32 LLGLSLShader::sTotalTrianglesDrawn = 0;
+U64 LLGLSLShader::sTotalSamplesDrawn = 0;
+U32 LLGLSLShader::sTotalDrawCalls = 0;
//UI shader -- declared here so llui_libtest will link properly
LLGLSLShader gUIProgram;
@@ -87,19 +93,188 @@ LLShaderFeatures::LLShaderFeatures()
//===============================
// LLGLSL Shader implementation
//===============================
+
+//static
+void LLGLSLShader::initProfile()
+{
+ sProfileEnabled = true;
+ sTotalTimeElapsed = 0;
+ sTotalTrianglesDrawn = 0;
+ sTotalSamplesDrawn = 0;
+ sTotalDrawCalls = 0;
+
+ for (std::set<LLGLSLShader*>::iterator iter = sInstances.begin(); iter != sInstances.end(); ++iter)
+ {
+ (*iter)->clearStats();
+ }
+}
+
+
+struct LLGLSLShaderCompareTimeElapsed
+{
+ bool operator()(const LLGLSLShader* const& lhs, const LLGLSLShader* const& rhs)
+ {
+ return lhs->mTimeElapsed < rhs->mTimeElapsed;
+ }
+};
+
+//static
+void LLGLSLShader::finishProfile()
+{
+ sProfileEnabled = false;
+
+ std::vector<LLGLSLShader*> sorted;
+
+ for (std::set<LLGLSLShader*>::iterator iter = sInstances.begin(); iter != sInstances.end(); ++iter)
+ {
+ sorted.push_back(*iter);
+ }
+
+ std::sort(sorted.begin(), sorted.end(), LLGLSLShaderCompareTimeElapsed());
+
+ for (std::vector<LLGLSLShader*>::iterator iter = sorted.begin(); iter != sorted.end(); ++iter)
+ {
+ (*iter)->dumpStats();
+ }
+
+ llinfos << "-----------------------------------" << llendl;
+ llinfos << "Total rendering time: " << llformat("%.4f ms", sTotalTimeElapsed/1000000.f) << llendl;
+ llinfos << "Total samples drawn: " << llformat("%.4f million", sTotalSamplesDrawn/1000000.f) << llendl;
+ llinfos << "Total triangles drawn: " << llformat("%.3f million", sTotalTrianglesDrawn/1000000.f) << llendl;
+}
+
+void LLGLSLShader::clearStats()
+{
+ mTrianglesDrawn = 0;
+ mTimeElapsed = 0;
+ mSamplesDrawn = 0;
+ mDrawCalls = 0;
+}
+
+void LLGLSLShader::dumpStats()
+{
+ if (mDrawCalls > 0)
+ {
+ llinfos << "=============================================" << llendl;
+ llinfos << mName << llendl;
+ for (U32 i = 0; i < mShaderFiles.size(); ++i)
+ {
+ llinfos << mShaderFiles[i].first << llendl;
+ }
+ llinfos << "=============================================" << llendl;
+
+ F32 ms = mTimeElapsed/1000000.f;
+ F32 seconds = ms/1000.f;
+
+ F32 pct_tris = (F32) mTrianglesDrawn/(F32)sTotalTrianglesDrawn*100.f;
+ F32 tris_sec = (F32) (mTrianglesDrawn/1000000.0);
+ tris_sec /= seconds;
+
+ F32 pct_samples = (F32) ((F64)mSamplesDrawn/(F64)sTotalSamplesDrawn)*100.f;
+ F32 samples_sec = (F32) mSamplesDrawn/1000000000.0;
+ samples_sec /= seconds;
+
+ F32 pct_calls = (F32) mDrawCalls/(F32)sTotalDrawCalls*100.f;
+ U32 avg_batch = mTrianglesDrawn/mDrawCalls;
+
+ llinfos << "Triangles Drawn: " << mTrianglesDrawn << " " << llformat("(%.2f pct of total, %.3f million/sec)", pct_tris, tris_sec ) << llendl;
+ llinfos << "Draw Calls: " << mDrawCalls << " " << llformat("(%.2f pct of total, avg %d tris/call)", pct_calls, avg_batch) << llendl;
+ llinfos << "SamplesDrawn: " << mSamplesDrawn << " " << llformat("(%.2f pct of total, %.3f billion/sec)", pct_samples, samples_sec) << llendl;
+ llinfos << "Time Elapsed: " << mTimeElapsed << " " << llformat("(%.2f pct of total, %.5f ms)\n", (F32) ((F64)mTimeElapsed/(F64)sTotalTimeElapsed)*100.f, ms) << llendl;
+ }
+}
+
+//static
+void LLGLSLShader::startProfile()
+{
+ if (sProfileEnabled && sCurBoundShaderPtr)
+ {
+ sCurBoundShaderPtr->placeProfileQuery();
+ }
+
+}
+
+//static
+void LLGLSLShader::stopProfile(U32 count, U32 mode)
+{
+ if (sProfileEnabled)
+ {
+ sCurBoundShaderPtr->readProfileQuery(count, mode);
+ }
+}
+
+void LLGLSLShader::placeProfileQuery()
+{
+#if !LL_DARWIN
+ if (mTimerQuery == 0)
+ {
+ glGenQueriesARB(1, &mTimerQuery);
+ }
+
+ glBeginQueryARB(GL_SAMPLES_PASSED, 1);
+ glBeginQueryARB(GL_TIME_ELAPSED, mTimerQuery);
+#endif
+}
+
+void LLGLSLShader::readProfileQuery(U32 count, U32 mode)
+{
+#if !LL_DARWIN
+ glEndQueryARB(GL_TIME_ELAPSED);
+ glEndQueryARB(GL_SAMPLES_PASSED);
+
+ U64 time_elapsed = 0;
+ glGetQueryObjectui64v(mTimerQuery, GL_QUERY_RESULT, &time_elapsed);
+
+ U64 samples_passed = 0;
+ glGetQueryObjectui64v(1, GL_QUERY_RESULT, &samples_passed);
+
+ sTotalTimeElapsed += time_elapsed;
+ mTimeElapsed += time_elapsed;
+
+ sTotalSamplesDrawn += samples_passed;
+ mSamplesDrawn += samples_passed;
+
+ U32 tri_count = 0;
+ switch (mode)
+ {
+ case LLRender::TRIANGLES: tri_count = count/3; break;
+ case LLRender::TRIANGLE_FAN: tri_count = count-2; break;
+ case LLRender::TRIANGLE_STRIP: tri_count = count-2; break;
+ default: tri_count = count; break; //points lines etc just use primitive count
+ }
+
+ mTrianglesDrawn += tri_count;
+ sTotalTrianglesDrawn += tri_count;
+
+ sTotalDrawCalls++;
+ mDrawCalls++;
+#endif
+}
+
+
+
LLGLSLShader::LLGLSLShader()
: mProgramObject(0),
mAttributeMask(0),
+ mTotalUniformSize(0),
mActiveTextureChannels(0),
mShaderLevel(0),
mShaderGroup(SG_DEFAULT),
- mUniformsDirty(FALSE)
+ mUniformsDirty(FALSE),
+ mTimerQuery(0)
{
+
+}
+LLGLSLShader::~LLGLSLShader()
+{
+
}
void LLGLSLShader::unload()
{
+ sInstances.erase(this);
+
stop_glerror();
mAttribute.clear();
mTexture.clear();
@@ -138,6 +313,8 @@ BOOL LLGLSLShader::createShader(vector<string> * attributes,
U32 varying_count,
const char** varyings)
{
+ sInstances.insert(this);
+
//reloading, reset matrix hash values
for (U32 i = 0; i < LLRender::NUM_MATRIX_MODES; ++i)
{
@@ -333,11 +510,56 @@ void LLGLSLShader::mapUniform(GLint index, const vector<string> * uniforms)
GLenum type;
GLsizei length;
- GLint size;
+ GLint size = -1;
char name[1024]; /* Flawfinder: ignore */
name[0] = 0;
+
glGetActiveUniformARB(mProgramObject, index, 1024, &length, &size, &type, (GLcharARB *)name);
+#if !LL_DARWIN
+ if (size > 0)
+ {
+ switch(type)
+ {
+ case GL_FLOAT_VEC2: size *= 2; break;
+ case GL_FLOAT_VEC3: size *= 3; break;
+ case GL_FLOAT_VEC4: size *= 4; break;
+ case GL_DOUBLE: size *= 2; break;
+ case GL_DOUBLE_VEC2: size *= 2; break;
+ case GL_DOUBLE_VEC3: size *= 6; break;
+ case GL_DOUBLE_VEC4: size *= 8; break;
+ case GL_INT_VEC2: size *= 2; break;
+ case GL_INT_VEC3: size *= 3; break;
+ case GL_INT_VEC4: size *= 4; break;
+ case GL_UNSIGNED_INT_VEC2: size *= 2; break;
+ case GL_UNSIGNED_INT_VEC3: size *= 3; break;
+ case GL_UNSIGNED_INT_VEC4: size *= 4; break;
+ case GL_BOOL_VEC2: size *= 2; break;
+ case GL_BOOL_VEC3: size *= 3; break;
+ case GL_BOOL_VEC4: size *= 4; break;
+ case GL_FLOAT_MAT2: size *= 4; break;
+ case GL_FLOAT_MAT3: size *= 9; break;
+ case GL_FLOAT_MAT4: size *= 16; break;
+ case GL_FLOAT_MAT2x3: size *= 6; break;
+ case GL_FLOAT_MAT2x4: size *= 8; break;
+ case GL_FLOAT_MAT3x2: size *= 6; break;
+ case GL_FLOAT_MAT3x4: size *= 12; break;
+ case GL_FLOAT_MAT4x2: size *= 8; break;
+ case GL_FLOAT_MAT4x3: size *= 12; break;
+ case GL_DOUBLE_MAT2: size *= 8; break;
+ case GL_DOUBLE_MAT3: size *= 18; break;
+ case GL_DOUBLE_MAT4: size *= 32; break;
+ case GL_DOUBLE_MAT2x3: size *= 12; break;
+ case GL_DOUBLE_MAT2x4: size *= 16; break;
+ case GL_DOUBLE_MAT3x2: size *= 12; break;
+ case GL_DOUBLE_MAT3x4: size *= 24; break;
+ case GL_DOUBLE_MAT4x2: size *= 16; break;
+ case GL_DOUBLE_MAT4x3: size *= 24; break;
+ }
+ mTotalUniformSize += size;
+ }
+#endif
+
S32 location = glGetUniformLocationARB(mProgramObject, name);
if (location != -1)
{
@@ -408,6 +630,7 @@ BOOL LLGLSLShader::mapUniforms(const vector<string> * uniforms)
{
BOOL res = TRUE;
+ mTotalUniformSize = 0;
mActiveTextureChannels = 0;
mUniform.clear();
mUniformMap.clear();
@@ -431,6 +654,7 @@ BOOL LLGLSLShader::mapUniforms(const vector<string> * uniforms)
unbind();
+ LL_DEBUGS("ShaderLoading") << "Total Uniform Size: " << mTotalUniformSize << llendl;
return res;
}
diff --git a/indra/llrender/llglslshader.h b/indra/llrender/llglslshader.h
index 725a7e2573..d67ece018f 100644
--- a/indra/llrender/llglslshader.h
+++ b/indra/llrender/llglslshader.h
@@ -67,14 +67,29 @@ public:
SG_WATER
};
+ static std::set<LLGLSLShader*> sInstances;
+ static bool sProfileEnabled;
+
LLGLSLShader();
+ ~LLGLSLShader();
static GLhandleARB sCurBoundShader;
static LLGLSLShader* sCurBoundShaderPtr;
static S32 sIndexedTextureChannels;
static bool sNoFixedFunction;
+ static void initProfile();
+ static void finishProfile();
+
+ static void startProfile();
+ static void stopProfile(U32 count, U32 mode);
+
void unload();
+ void clearStats();
+ void dumpStats();
+ void placeProfileQuery();
+ void readProfileQuery(U32 count, U32 mode);
+
BOOL createShader(std::vector<std::string> * attributes,
std::vector<std::string> * uniforms,
U32 varying_count = 0,
@@ -157,6 +172,7 @@ public:
std::map<std::string, GLint> mUniformMap; //lookup map of uniform name to uniform location
std::map<GLint, LLVector4> mValue; //lookup map of uniform location to last known value
std::vector<GLint> mTexture;
+ S32 mTotalUniformSize;
S32 mActiveTextureChannels;
S32 mShaderLevel;
S32 mShaderGroup;
@@ -165,6 +181,17 @@ public:
std::vector< std::pair< std::string, GLenum > > mShaderFiles;
std::string mName;
boost::unordered_map<std::string, std::string> mDefines;
+
+ //statistcis for profiling shader performance
+ U32 mTimerQuery;
+ U64 mTimeElapsed;
+ static U64 sTotalTimeElapsed;
+ U32 mTrianglesDrawn;
+ static U32 sTotalTrianglesDrawn;
+ U64 mSamplesDrawn;
+ static U64 sTotalSamplesDrawn;
+ U32 mDrawCalls;
+ static U32 sTotalDrawCalls;
};
//UI shader (declared here so llui_libtest will link properly)
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index 29fe5400a3..6108c6f821 100644
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -613,8 +613,9 @@ void LLVertexBuffer::drawArrays(U32 mode, const std::vector<LLVector3>& pos, con
glVertexPointer(3, GL_FLOAT, 0, pos[0].mV);
glNormalPointer(GL_FLOAT, 0, norm[0].mV);
}
-
+ LLGLSLShader::startProfile();
glDrawArrays(sGLMode[mode], 0, count);
+ LLGLSLShader::stopProfile(count, mode);
}
//static
@@ -651,7 +652,9 @@ void LLVertexBuffer::drawElements(U32 mode, const LLVector4a* pos, const LLVecto
glVertexPointer(3, GL_FLOAT, 16, pos);
}
+ LLGLSLShader::startProfile();
glDrawElements(sGLMode[mode], num_indices, GL_UNSIGNED_SHORT, indicesp);
+ LLGLSLShader::stopProfile(num_indices, mode);
}
void LLVertexBuffer::validateRange(U32 start, U32 end, U32 count, U32 indices_offset) const
@@ -751,9 +754,14 @@ void LLVertexBuffer::drawRange(U32 mode, U32 start, U32 end, U32 count, U32 indi
U16* idx = ((U16*) getIndicesPointer())+indices_offset;
stop_glerror();
+ LLGLSLShader::startProfile();
glDrawRangeElements(sGLMode[mode], start, end, count, GL_UNSIGNED_SHORT,
idx);
+ LLGLSLShader::stopProfile(count, mode);
stop_glerror();
+
+
+
placeFence();
}
@@ -797,8 +805,10 @@ void LLVertexBuffer::draw(U32 mode, U32 count, U32 indices_offset) const
}
stop_glerror();
+ LLGLSLShader::startProfile();
glDrawElements(sGLMode[mode], count, GL_UNSIGNED_SHORT,
((U16*) getIndicesPointer()) + indices_offset);
+ LLGLSLShader::stopProfile(count, mode);
stop_glerror();
placeFence();
}
@@ -838,7 +848,9 @@ void LLVertexBuffer::drawArrays(U32 mode, U32 first, U32 count) const
}
stop_glerror();
+ LLGLSLShader::startProfile();
glDrawArrays(sGLMode[mode], first, count);
+ LLGLSLShader::stopProfile(count, mode);
stop_glerror();
placeFence();
}
generated by cgit v1.2.3 (git 2.25.1) at 2024-12-03 21:55:45 +0000