From f67087b87277dc63096fbf4c42cdc17f84d3fbb8 Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Wed, 20 Feb 2013 19:30:46 -0600 Subject: MAINT-2370 Add debug setting "RenderAutoMuteRenderCostLimit" that automatically visually mutes avatars above a certain threshold (adjusted by Avatar Mesh Detail Slider) --- indra/newview/llface.cpp | 7 +++++++ 1 file changed, 7 insertions(+) (limited to 'indra/newview/llface.cpp') diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp index 28e4b32793..a7e225843c 100755 --- a/indra/newview/llface.cpp +++ b/indra/newview/llface.cpp @@ -51,6 +51,7 @@ #include "llviewerregion.h" #include "llviewerwindow.h" #include "llviewershadermgr.h" +#include "llvoavatar.h" #define LL_MAX_INDICES_COUNT 1000000 @@ -325,6 +326,12 @@ void LLFace::dirtyTexture() if (vobj) { vobj->mLODChanged = TRUE; + + LLVOAvatar* avatar = vobj->getAvatar(); + if (avatar) + { //avatar render cost may have changed + avatar->updateVisualComplexity(); + } } gPipeline.markRebuild(drawablep, LLDrawable::REBUILD_VOLUME, FALSE); } -- cgit v1.2.3 From f5e5396c3a17b6bcdc4eb49cda304a9047920fe1 Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Tue, 26 Feb 2013 15:15:08 -0600 Subject: MAINT-2371 First set of profile guided optimizations. Reviewed by Graham --- indra/newview/llface.cpp | 28 ++++++++++++++++++++-------- 1 file changed, 20 insertions(+), 8 deletions(-) (limited to 'indra/newview/llface.cpp') diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp index a7e225843c..6b3127decf 100755 --- a/indra/newview/llface.cpp +++ b/indra/newview/llface.cpp @@ -53,6 +53,7 @@ #include "llviewershadermgr.h" #include "llvoavatar.h" +extern BOOL gGLDebugLoggingEnabled; #define LL_MAX_INDICES_COUNT 1000000 @@ -1166,6 +1167,15 @@ static LLFastTimer::DeclareTimer FTM_FACE_GEOM_COLOR("Color"); static LLFastTimer::DeclareTimer FTM_FACE_GEOM_EMISSIVE("Emissive"); static LLFastTimer::DeclareTimer FTM_FACE_GEOM_WEIGHTS("Weights"); static LLFastTimer::DeclareTimer FTM_FACE_GEOM_BINORMAL("Binormal"); + +static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK("Face Feedback"); +static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_POSITION("Feedback Position"); +static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_NORMAL("Feedback Normal"); +static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_TEXTURE("Feedback Texture"); +static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_COLOR("Feedback Color"); +static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_EMISSIVE("Feedback Emissive"); +static LLFastTimer::DeclareTimer FTM_FACE_GEOM_FEEDBACK_BINORMAL("Feedback Binormal"); + static LLFastTimer::DeclareTimer FTM_FACE_GEOM_INDEX("Index"); static LLFastTimer::DeclareTimer FTM_FACE_GEOM_INDEX_TAIL("Tail"); static LLFastTimer::DeclareTimer FTM_FACE_POSITION_STORE("Pos"); @@ -1389,12 +1399,14 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, #ifdef GL_TRANSFORM_FEEDBACK_BUFFER if (use_transform_feedback && + mVertexBuffer->getUsage() == GL_DYNAMIC_COPY_ARB && gTransformPositionProgram.mProgramObject && //transform shaders are loaded mVertexBuffer->useVBOs() && //target buffer is in VRAM !rebuild_weights && //TODO: add support for weights !volume.isUnique()) //source volume is NOT flexi { //use transform feedback to pack vertex buffer - + //gGLDebugLoggingEnabled = TRUE; + LLFastTimer t(FTM_FACE_GEOM_FEEDBACK); LLVertexBuffer* buff = (LLVertexBuffer*) vf.mVertexBuffer.get(); if (vf.mVertexBuffer.isNull() || buff->getNumVerts() != vf.mNumVertices) @@ -1411,7 +1423,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, if (rebuild_pos) { - LLFastTimer t(FTM_FACE_GEOM_POSITION); + LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_POSITION); gTransformPositionProgram.bind(); mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_VERTEX, mGeomIndex, mGeomCount); @@ -1436,7 +1448,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, if (rebuild_color) { - LLFastTimer t(FTM_FACE_GEOM_COLOR); + LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_COLOR); gTransformColorProgram.bind(); mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_COLOR, mGeomIndex, mGeomCount); @@ -1452,7 +1464,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, if (rebuild_emissive) { - LLFastTimer t(FTM_FACE_GEOM_EMISSIVE); + LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_EMISSIVE); gTransformColorProgram.bind(); mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_EMISSIVE, mGeomIndex, mGeomCount); @@ -1473,7 +1485,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, if (rebuild_normal) { - LLFastTimer t(FTM_FACE_GEOM_NORMAL); + LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_NORMAL); gTransformNormalProgram.bind(); mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_NORMAL, mGeomIndex, mGeomCount); @@ -1486,7 +1498,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, if (rebuild_binormal) { - LLFastTimer t(FTM_FACE_GEOM_BINORMAL); + LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_BINORMAL); gTransformBinormalProgram.bind(); mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_BINORMAL, mGeomIndex, mGeomCount); @@ -1499,7 +1511,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, if (rebuild_tcoord) { - LLFastTimer t(FTM_FACE_GEOM_TEXTURE); + LLFastTimer t(FTM_FACE_GEOM_FEEDBACK_TEXTURE); gTransformTexCoordProgram.bind(); mVertexBuffer->bindForFeedback(0, LLVertexBuffer::TYPE_TEXCOORD0, mGeomIndex, mGeomCount); @@ -1522,13 +1534,13 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, } glBindBufferARB(GL_TRANSFORM_FEEDBACK_BUFFER, 0); - gGL.popMatrix(); if (cur_shader) { cur_shader->bind(); } + //gGLDebugLoggingEnabled = FALSE; } else #endif -- cgit v1.2.3 From df08808640031bd27a11177ea49a08f797d2d570 Mon Sep 17 00:00:00 2001 From: "Graham Madarasz (Graham)" Date: Thu, 28 Feb 2013 09:33:41 -0800 Subject: Improve perf of GLSL uniform lookups by name --- indra/newview/llface.cpp | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-) (limited to 'indra/newview/llface.cpp') diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp index 28e4b32793..3c7a49d5ce 100755 --- a/indra/newview/llface.cpp +++ b/indra/newview/llface.cpp @@ -55,6 +55,9 @@ #define LL_MAX_INDICES_COUNT 1000000 +static LLStaticHashedString sTextureIndexIn("texture_index_in"); +static LLStaticHashedString sColorIn("color_in"); + BOOL LLFace::sSafeRenderSelect = TRUE; // FALSE #define DOTVEC(a,b) (a.mV[0]*b.mV[0] + a.mV[1]*b.mV[1] + a.mV[2]*b.mV[2]) @@ -1417,8 +1420,8 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, vp[1] = 0; vp[2] = 0; vp[3] = 0; - - gTransformPositionProgram.uniform1i("texture_index_in", val); + + gTransformPositionProgram.uniform1i(sTextureIndexIn, val); glBeginTransformFeedback(GL_POINTS); buff->setBuffer(LLVertexBuffer::MAP_VERTEX); @@ -1436,7 +1439,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, S32 val = *((S32*) color.mV); - gTransformColorProgram.uniform1i("color_in", val); + gTransformColorProgram.uniform1i(sColorIn, val); glBeginTransformFeedback(GL_POINTS); buff->setBuffer(LLVertexBuffer::MAP_VERTEX); push_for_transform(buff, vf.mNumVertices, mGeomCount); @@ -1457,7 +1460,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, (glow << 16) | (glow << 24); - gTransformColorProgram.uniform1i("color_in", glow32); + gTransformColorProgram.uniform1i(sColorIn, glow32); glBeginTransformFeedback(GL_POINTS); buff->setBuffer(LLVertexBuffer::MAP_VERTEX); push_for_transform(buff, vf.mNumVertices, mGeomCount); -- cgit v1.2.3 From 609ed855e1160505238378a1be49e2b92e8496f5 Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Mon, 4 Mar 2013 18:01:42 -0600 Subject: MAINT-2371 More optimizations. Reviewed by Graham --- indra/newview/llface.cpp | 99 +++++++++++++++++++++++++++++++++--------------- 1 file changed, 69 insertions(+), 30 deletions(-) (limited to 'indra/newview/llface.cpp') diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp index 6b3127decf..ef91a459e7 100755 --- a/indra/newview/llface.cpp +++ b/indra/newview/llface.cpp @@ -1407,6 +1407,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, { //use transform feedback to pack vertex buffer //gGLDebugLoggingEnabled = TRUE; LLFastTimer t(FTM_FACE_GEOM_FEEDBACK); + LLGLEnable discard(GL_RASTERIZER_DISCARD); LLVertexBuffer* buff = (LLVertexBuffer*) vf.mVertexBuffer.get(); if (vf.mVertexBuffer.isNull() || buff->getNumVerts() != vf.mNumVertices) @@ -1955,21 +1956,31 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, if (rebuild_pos) { - LLFastTimer t(FTM_FACE_GEOM_POSITION); + LLVector4a* src = vf.mPositions; + + //_mm_prefetch((char*)src, _MM_HINT_T0); + + LLVector4a* end = src+num_vertices; + //LLVector4a* end_64 = end-4; + + //LLFastTimer t(FTM_FACE_GEOM_POSITION); llassert(num_vertices > 0); mVertexBuffer->getVertexStrider(vert, mGeomIndex, mGeomCount, map_range); - LLMatrix4a mat_vert; mat_vert.loadu(mat_vert_in); + + F32* dst = (F32*) vert.get(); + F32* end_f32 = dst+mGeomCount*4; - LLVector4a* src = vf.mPositions; - volatile F32* dst = (volatile F32*) vert.get(); - - volatile F32* end = dst+num_vertices*4; - LLVector4a res; + //_mm_prefetch((char*)dst, _MM_HINT_NTA); + //_mm_prefetch((char*)src, _MM_HINT_NTA); + + //_mm_prefetch((char*)dst, _MM_HINT_NTA); + LLVector4a res0; //,res1,res2,res3; + LLVector4a texIdx; S32 index = mTextureIndex < 255 ? mTextureIndex : 0; @@ -1986,29 +1997,53 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, texIdx.set(0,0,0,val); + LLVector4a tmp; + { - LLFastTimer t(FTM_FACE_POSITION_STORE); - LLVector4a tmp; + //LLFastTimer t2(FTM_FACE_POSITION_STORE); - do - { - mat_vert.affineTransform(*src++, res); - tmp.setSelectWithMask(mask, texIdx, res); + /*if (num_vertices > 4) + { //more than 64 bytes + while (src < end_64) + { + _mm_prefetch((char*)src + 64, _MM_HINT_T0); + _mm_prefetch((char*)dst + 64, _MM_HINT_T0); + + mat_vert.affineTransform(*src, res0); + tmp.setSelectWithMask(mask, texIdx, res0); + tmp.store4a((F32*) dst); + + mat_vert.affineTransform(*(src+1), res1); + tmp.setSelectWithMask(mask, texIdx, res1); + tmp.store4a((F32*) dst+4); + + mat_vert.affineTransform(*(src+2), res2); + tmp.setSelectWithMask(mask, texIdx, res2); + tmp.store4a((F32*) dst+8); + + mat_vert.affineTransform(*(src+3), res3); + tmp.setSelectWithMask(mask, texIdx, res3); + tmp.store4a((F32*) dst+12); + + dst += 16; + src += 4; + } + }*/ + + while (src < end) + { + mat_vert.affineTransform(*src++, res0); + tmp.setSelectWithMask(mask, texIdx, res0); tmp.store4a((F32*) dst); dst += 4; } - while(dst < end); } - + { - LLFastTimer t(FTM_FACE_POSITION_PAD); - S32 aligned_pad_vertices = mGeomCount - num_vertices; - res.set(res[0], res[1], res[2], 0.f); - - while (aligned_pad_vertices > 0) + //LLFastTimer t(FTM_FACE_POSITION_PAD); + while (dst < end_f32) { - --aligned_pad_vertices; - res.store4a((F32*) dst); + res0.store4a((F32*) dst); dst += 4; } } @@ -2022,15 +2057,17 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, if (rebuild_normal) { - LLFastTimer t(FTM_FACE_GEOM_NORMAL); + //LLFastTimer t(FTM_FACE_GEOM_NORMAL); mVertexBuffer->getNormalStrider(norm, mGeomIndex, mGeomCount, map_range); F32* normals = (F32*) norm.get(); - for (S32 i = 0; i < num_vertices; i++) - { + LLVector4a* src = vf.mNormals; + LLVector4a* end = src+num_vertices; + + while (src < end) + { LLVector4a normal; - mat_normal.rotate(vf.mNormals[i], normal); - normal.normalize3fast(); + mat_normal.rotate(*src++, normal); normal.store4a(normals); normals += 4; } @@ -2047,11 +2084,13 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, mVertexBuffer->getBinormalStrider(binorm, mGeomIndex, mGeomCount, map_range); F32* binormals = (F32*) binorm.get(); - for (S32 i = 0; i < num_vertices; i++) + LLVector4a* src = vf.mBinormals; + LLVector4a* end = vf.mBinormals+num_vertices; + + while (src < end) { LLVector4a binormal; - mat_normal.rotate(vf.mBinormals[i], binormal); - binormal.normalize3fast(); + mat_normal.rotate(*src++, binormal); binormal.store4a(binormals); binormals += 4; } -- cgit v1.2.3 From 6613d80d72931b13cc008c3dcc8ee90a39bec8f5 Mon Sep 17 00:00:00 2001 From: Graham Madarasz Date: Mon, 11 Mar 2013 14:19:05 -0700 Subject: Clean up moving llalignedarray and fast memcpy to llcommon --- indra/newview/llface.cpp | 3 +++ 1 file changed, 3 insertions(+) (limited to 'indra/newview/llface.cpp') diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp index 86e5f20812..5f86205175 100755 --- a/indra/newview/llface.cpp +++ b/indra/newview/llface.cpp @@ -64,6 +64,8 @@ BOOL LLFace::sSafeRenderSelect = TRUE; // FALSE #define DOTVEC(a,b) (a.mV[0]*b.mV[0] + a.mV[1]*b.mV[1] + a.mV[2]*b.mV[2]) +//#pragma GCC diagnostic ignored "-Wuninitialized" + /* For each vertex, given: B - binormal @@ -1982,6 +1984,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume, //_mm_prefetch((char*)dst, _MM_HINT_NTA); + LLVector4a res0; //,res1,res2,res3; LLVector4a texIdx; -- cgit v1.2.3 From 35383790fb250532fa7701638d39a8f1c6a0cf76 Mon Sep 17 00:00:00 2001 From: "Graham Madarasz (Graham Linden)" Date: Tue, 12 Mar 2013 04:40:51 -0700 Subject: Merge pragma warning hack to fix linux build --- indra/newview/llface.cpp | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) (limited to 'indra/newview/llface.cpp') diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp index 5f86205175..c3cb914120 100755 --- a/indra/newview/llface.cpp +++ b/indra/newview/llface.cpp @@ -53,6 +53,12 @@ #include "llviewershadermgr.h" #include "llvoavatar.h" +#if LL_LINUX +// Work-around spurious used before init warning on Vector4a +// +#pragma GCC diagnostic ignored "-Wuninitialized" +#endif + extern BOOL gGLDebugLoggingEnabled; #define LL_MAX_INDICES_COUNT 1000000 @@ -64,8 +70,6 @@ BOOL LLFace::sSafeRenderSelect = TRUE; // FALSE #define DOTVEC(a,b) (a.mV[0]*b.mV[0] + a.mV[1]*b.mV[1] + a.mV[2]*b.mV[2]) -//#pragma GCC diagnostic ignored "-Wuninitialized" - /* For each vertex, given: B - binormal -- cgit v1.2.3