diff options
-rw-r--r-- | indra/newview/llappviewer.cpp | 2 | ||||
-rw-r--r-- | indra/newview/llfloaterperformance.cpp | 11 | ||||
-rw-r--r-- | indra/newview/llperfstats.cpp | 67 | ||||
-rw-r--r-- | indra/newview/llperfstats.h | 68 | ||||
-rw-r--r-- | indra/newview/llviewerstats.cpp | 2 | ||||
-rw-r--r-- | indra/newview/llvoavatar.cpp | 146 | ||||
-rw-r--r-- | indra/newview/llvoavatar.h | 30 | ||||
-rw-r--r-- | indra/newview/llworld.cpp | 6 | ||||
-rw-r--r-- | indra/newview/pipeline.cpp | 42 |
9 files changed, 201 insertions, 173 deletions
diff --git a/indra/newview/llappviewer.cpp b/indra/newview/llappviewer.cpp index 6efbaeacf7..c9fc0e2cc8 100644 --- a/indra/newview/llappviewer.cpp +++ b/indra/newview/llappviewer.cpp @@ -4698,6 +4698,8 @@ void LLAppViewer::idle() LLFrameTimer::updateFrameTime(); LLFrameTimer::updateFrameCount(); LLEventTimer::updateClass(); + LLPerfStats::updateClass(); + // LLApp::stepFrame() performs the above three calls plus mRunner.run(). // Not sure why we don't call stepFrame() here, except that LLRunner seems // completely redundant with LLEventTimer. diff --git a/indra/newview/llfloaterperformance.cpp b/indra/newview/llfloaterperformance.cpp index 8c5745aa43..19fc3e673e 100644 --- a/indra/newview/llfloaterperformance.cpp +++ b/indra/newview/llfloaterperformance.cpp @@ -456,15 +456,8 @@ void LLFloaterPerformance::populateNearbyList() row[1]["column"] = "complex_value"; row[1]["type"] = "text"; - if (is_slow && !showTunedART) - { - row[1]["value"] = llformat( "%.f", LLPerfStats::raw_to_us( avatar->getLastART() ) ); - } - else - { - // use GPU time in us - row[1]["value"] = llformat( "%.f", render_av_gpu_ms * 1000.f); - } + // use GPU time in us + row[1]["value"] = llformat( "%.f", render_av_gpu_ms * 1000.f); row[1]["font"]["name"] = "SANSSERIF"; row[3]["column"] = "name"; diff --git a/indra/newview/llperfstats.cpp b/indra/newview/llperfstats.cpp index b680d8761b..2281475d4a 100644 --- a/indra/newview/llperfstats.cpp +++ b/indra/newview/llperfstats.cpp @@ -39,6 +39,11 @@ extern LLControlGroup gSavedSettings; namespace LLPerfStats { + // avatar timing metrics in ms (updated once per mainloop iteration) + std::atomic<F32> sTotalAvatarTime = 0.f; + std::atomic<F32> sAverageAvatarTime = 0.f; + std::atomic<F32> sMaxAvatarTime = 0.f; + std::atomic<int64_t> tunedAvatars{0}; std::atomic<U64> renderAvatarMaxART_ns{(U64)(ART_UNLIMITED_NANOS)}; // highest render time we'll allow without culling features bool belowTargetFPS{false}; @@ -141,14 +146,12 @@ namespace LLPerfStats resetChanges(); } - StatsRecorder::StatsRecorder():q(1024*16),t(&StatsRecorder::run) + StatsRecorder::StatsRecorder():q(1024*16) { // create a queue - // create a thread to consume from the queue tunables.initialiseFromSettings(); LLPerfStats::cpu_hertz = (F64)LLTrace::BlockTimer::countsPerSecond(); LLPerfStats::vsync_max_fps = gViewerWindow->getWindow()->getRefreshRate(); - t.detach(); } // static @@ -200,20 +203,6 @@ namespace LLPerfStats } } - auto& statsMapAv = statsDoubleBuffer[writeBuffer][static_cast<size_t>(ObjType_t::OT_AVATAR)]; - for(auto& stat_entry : statsMapAv) - { - for(auto& stat : avatarStatsToAvg) - { - auto val = stat_entry.second[static_cast<size_t>(stat)]; - if(val > SMOOTHING_PERIODS) - { - auto avg = statsDoubleBuffer[writeBuffer ^ 1][static_cast<size_t>(ObjType_t::OT_AVATAR)][stat_entry.first][static_cast<size_t>(stat)]; - stat_entry.second[static_cast<size_t>(stat)] = avg + (val / SMOOTHING_PERIODS) - (avg / SMOOTHING_PERIODS); - } - } - } - // swap the buffers if(enabled()) { @@ -293,6 +282,36 @@ namespace LLPerfStats } } + // called once per main loop iteration on main thread + void updateClass() + { + LL_PROFILE_ZONE_SCOPED_CATEGORY_STATS; + + sTotalAvatarTime = LLVOAvatar::getTotalGPURenderTime(); + sAverageAvatarTime = LLVOAvatar::getAverageGPURenderTime(); + sMaxAvatarTime = LLVOAvatar::getMaxGPURenderTime(); + + auto& general = LL::WorkQueue::getInstance("General"); + + if (general) + { + general->post([] { StatsRecorder::update(); }); + } + } + + // called once per main loop iteration on General thread + void StatsRecorder::update() + { + LL_PROFILE_ZONE_SCOPED_CATEGORY_STATS; + StatsRecord upd; + auto& instance{ StatsRecorder::getInstance() }; + + while (enabled() && !LLApp::isQuitting() && instance.q.tryPop(upd)) + { + instance.processUpdate(upd); + } + } + //static int StatsRecorder::countNearbyAvatars(S32 distance) { @@ -327,6 +346,8 @@ namespace LLPerfStats // static void StatsRecorder::updateAvatarParams() { + LL_PROFILE_ZONE_SCOPED_CATEGORY_STATS; + if(tunables.autoTuneTimeout) { LLPerfStats::lastSleepedFrame = gFrameCount; @@ -380,10 +401,10 @@ namespace LLPerfStats } } - auto av_render_max_raw = LLPerfStats::StatsRecorder::getMax(ObjType_t::OT_AVATAR, LLPerfStats::StatType_t::RENDER_COMBINED); + auto av_render_max_raw = ms_to_raw(sMaxAvatarTime); // Is our target frame time lower than current? If so we need to take action to reduce draw overheads. // cumulative avatar time (includes idle processing, attachments and base av) - auto tot_avatar_time_raw = LLPerfStats::StatsRecorder::getSum(ObjType_t::OT_AVATAR, LLPerfStats::StatType_t::RENDER_COMBINED); + auto tot_avatar_time_raw = ms_to_raw(sTotalAvatarTime); // The frametime budget we have based on the target FPS selected auto target_frame_time_raw = (U64)llround(LLPerfStats::cpu_hertz / (target_fps == 0 ? 1 : target_fps)); @@ -417,7 +438,7 @@ namespace LLPerfStats // if so we've got work to do // how much of the frame was spent on non avatar related work? - U64 non_avatar_time_raw = tot_frame_time_raw - tot_avatar_time_raw; + U64 non_avatar_time_raw = tot_frame_time_raw > tot_avatar_time_raw ? tot_frame_time_raw - tot_avatar_time_raw : 0; // If the target frame time < scene time (estimated as non_avatar time) U64 target_avatar_time_raw; @@ -475,7 +496,11 @@ namespace LLPerfStats { new_render_limit_ns = renderAvatarMaxART_ns; } - new_render_limit_ns -= LLPerfStats::ART_MIN_ADJUST_DOWN_NANOS; + + if (new_render_limit_ns > LLPerfStats::ART_MIN_ADJUST_DOWN_NANOS) + { + new_render_limit_ns -= LLPerfStats::ART_MIN_ADJUST_DOWN_NANOS; + } // bounce at the bottom to prevent "no limit" new_render_limit_ns = std::max((U64)new_render_limit_ns, (U64)LLPerfStats::ART_MINIMUM_NANOS); diff --git a/indra/newview/llperfstats.h b/indra/newview/llperfstats.h index a4768272b9..bb5677f237 100644 --- a/indra/newview/llperfstats.h +++ b/indra/newview/llperfstats.h @@ -42,6 +42,10 @@ extern U32 gFrameCount; extern LLUUID gAgentID; namespace LLPerfStats { + + // called once per main loop iteration + void updateClass(); + // Note if changing these, they should correspond with the log range of the correpsonding sliders static constexpr U64 ART_UNLIMITED_NANOS{50000000}; static constexpr U64 ART_MINIMUM_NANOS{100000}; @@ -68,7 +72,6 @@ namespace LLPerfStats enum class ObjType_t{ OT_GENERAL=0, // Also Unknown. Used for n/a type stats such as scenery - OT_AVATAR, OT_COUNT }; enum class StatType_t{ @@ -162,6 +165,9 @@ namespace LLPerfStats using Queue = LLThreadSafeQueue<StatsRecord>; public: + // called once per main loop iteration on General thread + static void update(); + static inline StatsRecorder& getInstance() { static StatsRecorder instance; @@ -241,7 +247,6 @@ namespace LLPerfStats auto ot{upd.objType}; auto& key{upd.objID}; - auto& avKey{upd.avID}; auto type {upd.statType}; auto val {upd.time}; @@ -251,13 +256,6 @@ namespace LLPerfStats doUpd(key, ot, type,val); return; } - - if (ot == ObjType_t::OT_AVATAR) - { - // LL_INFOS("perfstats") << "Avatar update:" << LL_ENDL; - doUpd(avKey, ot, type, val); - return; - } } static inline void doUpd(const LLUUID& key, ObjType_t ot, StatType_t type, uint64_t val) @@ -286,42 +284,7 @@ namespace LLPerfStats static void toggleBuffer(); static void clearStatsBuffers(); - // thread entry - static void run() - { - StatsRecord upd[10]; - auto & instance {StatsRecorder::getInstance()}; - LL_PROFILER_SET_THREAD_NAME("PerfStats"); - - while( enabled() && !LLApp::isExiting() ) - { - auto count = 0; - while (count < 10) - { - if (instance.q.tryPopFor(std::chrono::milliseconds(10), upd[count])) - { - count++; - } - else - { - break; - } - } - //LL_PROFILER_THREAD_BEGIN("PerfStats"); - if(count) - { - // LL_INFOS("perfstats") << "processing " << count << " updates." << LL_ENDL; - for(auto i =0; i < count; i++) - { - instance.processUpdate(upd[i]); - } - } - //LL_PROFILER_THREAD_END("PerfStats"); - } - } - Queue q; - std::thread t; ~StatsRecorder() = default; StatsRecorder(const StatsRecorder&) = delete; @@ -354,13 +317,6 @@ namespace LLPerfStats LL_PROFILE_ZONE_SCOPED_CATEGORY_STATS; }; - template < ObjType_t OD = ObjTypeDiscriminator, - std::enable_if_t<OD == ObjType_t::OT_AVATAR> * = nullptr> - RecordTime( const LLUUID & av, StatType_t type ):RecordTime<ObjTypeDiscriminator>(std::move(av), LLUUID::null, type) - { - //LL_PROFILE_ZONE_COLOR(tracy::Color::Purple); - }; - ~RecordTime() { if(!LLPerfStats::StatsRecorder::enabled()) @@ -375,13 +331,19 @@ namespace LLPerfStats }; }; - + inline double raw_to_ns(U64 raw) { return (static_cast<double>(raw) * 1000000000.0) / LLPerfStats::cpu_hertz; }; inline double raw_to_us(U64 raw) { return (static_cast<double>(raw) * 1000000.0) / LLPerfStats::cpu_hertz; }; inline double raw_to_ms(U64 raw) { return (static_cast<double>(raw) * 1000.0) / LLPerfStats::cpu_hertz; }; + inline U64 ns_to_raw(double ns) { return (U64)(LLPerfStats::cpu_hertz * (ns / 1000000000.0)); } + inline U64 us_to_raw(double us) { return (U64)(LLPerfStats::cpu_hertz * (us / 1000000.0)); } + inline U64 ms_to_raw(double ms) { return (U64)(LLPerfStats::cpu_hertz * (ms / 1000.0)); + + } + + using RecordSceneTime = RecordTime<ObjType_t::OT_GENERAL>; - using RecordAvatarTime = RecordTime<ObjType_t::OT_AVATAR>; };// namespace LLPerfStats diff --git a/indra/newview/llviewerstats.cpp b/indra/newview/llviewerstats.cpp index 609e8290da..37999e58ff 100644 --- a/indra/newview/llviewerstats.cpp +++ b/indra/newview/llviewerstats.cpp @@ -472,7 +472,7 @@ void update_statistics() auto tot_frame_time_raw = LLPerfStats::StatsRecorder::getSceneStat(LLPerfStats::StatType_t::RENDER_FRAME); // cumulative avatar time (includes idle processing, attachments and base av) - auto tot_avatar_time_raw = LLPerfStats::StatsRecorder::getSum(LLPerfStats::ObjType_t::OT_AVATAR, LLPerfStats::StatType_t::RENDER_COMBINED); + auto tot_avatar_time_raw = LLPerfStats::us_to_raw(LLVOAvatar::getTotalGPURenderTime()); // cumulative avatar render specific time (a bit arbitrary as the processing is too.) // auto tot_av_idle_time_raw = LLPerfStats::StatsRecorder::getSum(AvType, LLPerfStats::StatType_t::RENDER_IDLE); // auto tot_avatar_render_time_raw = tot_avatar_time_raw - tot_av_idle_time_raw; diff --git a/indra/newview/llvoavatar.cpp b/indra/newview/llvoavatar.cpp index 95e9321d6f..af65588709 100644 --- a/indra/newview/llvoavatar.cpp +++ b/indra/newview/llvoavatar.cpp @@ -2562,7 +2562,6 @@ void LLVOAvatar::idleUpdate(LLAgent &agent, const F64 &time) return; } // record time and refresh "tooSlow" status - LLPerfStats::RecordAvatarTime T(getID(), LLPerfStats::StatType_t::RENDER_IDLE); // per avatar "idle" time. updateTooSlow(); static LLCachedControl<bool> disable_all_render_types(gSavedSettings, "DisableAllRenderTypes"); @@ -8310,14 +8309,6 @@ bool LLVOAvatar::isTooSlow() const return mTooSlow; } -// use Avatar Render Time as complexity metric -// markARTStale - Mark stale and set the frameupdate to now so that we can wait at least one frame to get a revised number. -void LLVOAvatar::markARTStale() -{ - mARTStale=true; - mLastARTUpdateFrame = LLFrameTimer::getFrameCount(); -} - // Udpate Avatar state based on render time void LLVOAvatar::updateTooSlow() { @@ -8328,41 +8319,9 @@ void LLVOAvatar::updateTooSlow() // mTooSlow - Is the avatar flagged as being slow (includes shadow time) // mTooSlowWithoutShadows - Is the avatar flagged as being slow even with shadows removed. - // mARTStale - the rendertime we have is stale because of an update. We need to force a re-render to re-assess slowness - - if( mARTStale ) - { - if ( LLFrameTimer::getFrameCount() - mLastARTUpdateFrame < 5 ) - { - // LL_INFOS() << this->getFullname() << " marked stale " << LL_ENDL; - // we've not had a chance to update yet (allow a few to be certain a full frame has passed) - return; - } - - mARTStale = false; - mTooSlow = false; - mTooSlowWithoutShadows = false; - // LL_INFOS() << this->getFullname() << " refreshed ART combined = " << mRenderTime << " @ " << mLastARTUpdateFrame << LL_ENDL; - } - - // Either we're not stale or we've updated. - - U64 render_time_raw; - U64 render_geom_time_raw; - - if( !mTooSlow ) - { - // we are fully rendered, so we use the live values - std::lock_guard<std::mutex> lock{LLPerfStats::bufferToggleLock}; - render_time_raw = LLPerfStats::StatsRecorder::get(LLPerfStats::ObjType_t::OT_AVATAR, id, LLPerfStats::StatType_t::RENDER_COMBINED); - render_geom_time_raw = LLPerfStats::StatsRecorder::get(LLPerfStats::ObjType_t::OT_AVATAR, id, LLPerfStats::StatType_t::RENDER_GEOMETRY); - } - else - { - // use the cached values. - render_time_raw = mRenderTime; - render_geom_time_raw = mGeomTime; - } + + // get max render time in ms + F32 max_art_ms = (F32) (LLPerfStats::renderAvatarMaxART_ns / 1000000.0); bool autotune = LLPerfStats::tunables.userAutoTuneEnabled && !mIsControlAvatar && !isSelf(); @@ -8378,19 +8337,13 @@ void LLVOAvatar::updateTooSlow() } bool exceeds_max_ART = - ((LLPerfStats::renderAvatarMaxART_ns > 0) && (LLPerfStats::raw_to_ns(render_time_raw) >= LLPerfStats::renderAvatarMaxART_ns)); + ((LLPerfStats::renderAvatarMaxART_ns > 0) && + (mGPURenderTime >= max_art_ms)); // NOTE: don't use getGPURenderTime accessor here to avoid "isTooSlow" feedback loop if (exceeds_max_ART && !ignore_tune) { - if( !mTooSlow ) // if we were previously not slow (with or without shadows.) - { - // if we weren't capped, we are now - mLastARTUpdateFrame = LLFrameTimer::getFrameCount(); - mRenderTime = render_time_raw; - mGeomTime = render_geom_time_raw; - mARTStale = false; - mTooSlow = true; - } + mTooSlow = true; + if(!mTooSlowWithoutShadows) // if we were not previously above the full impostor cap { bool render_friend_or_exception = ( alwaysRenderFriends && LLAvatarTracker::instance().isBuddy( id ) ) || @@ -8398,13 +8351,12 @@ void LLVOAvatar::updateTooSlow() if( (!isSelf() || allowSelfImpostor) && !render_friend_or_exception ) { // Note: slow rendering Friends still get their shadows zapped. - mTooSlowWithoutShadows = (LLPerfStats::raw_to_ns(render_geom_time_raw) >= LLPerfStats::renderAvatarMaxART_ns); + mTooSlowWithoutShadows = getGPURenderTime()*2.f >= max_art_ms; // NOTE: assumes shadow rendering doubles render time } } } else { - // LL_INFOS() << this->getFullname() << " ("<< (combined?"combined":"geometry") << ") good render time = " << LLPerfStats::raw_to_ns(render_time_raw) << " vs ("<< LLVOAvatar::sRenderTimeCap_ns << " set @ " << mLastARTUpdateFrame << LL_ENDL; mTooSlow = false; mTooSlowWithoutShadows = false; @@ -11103,6 +11055,21 @@ void LLVOAvatar::calculateUpdateRenderComplexity() // HUD complexity LLHUDRenderNotifier::getInstance()->updateNotificationHUD(hud_complexity_list); } + + //schedule an update to ART next frame if needed + if (LLPerfStats::tunables.userAutoTuneEnabled && + LLPerfStats::tunables.userFPSTuningStrategy != LLPerfStats::TUNE_SCENE_ONLY && + !isVisuallyMuted()) + { + LLUUID id = getID(); // <== use id to make sure this avatar didn't get deleted between frames + LL::WorkQueue::getInstance("mainloop")->post([this, id]() + { + if (gObjectList.findObject(id) != nullptr) + { + gPipeline.profileAvatar(this); + } + }); + } } } @@ -11465,6 +11432,9 @@ void LLVOAvatar::readProfileQuery(S32 retries) glGetQueryObjectui64v(mGPUTimerQuery, GL_QUERY_RESULT, &time_elapsed); mGPURenderTime = time_elapsed / 1000000.f; mGPUProfilePending = false; + + setDebugText(llformat("%d", (S32)(mGPURenderTime * 1000.f))); + } else { // wait until next frame @@ -11477,3 +11447,67 @@ void LLVOAvatar::readProfileQuery(S32 retries) } } + +F32 LLVOAvatar::getGPURenderTime() +{ + return isVisuallyMuted() ? 0.f : mGPURenderTime; +} + +// static +F32 LLVOAvatar::getTotalGPURenderTime() +{ + LL_PROFILE_ZONE_SCOPED_CATEGORY_AVATAR; + + F32 ret = 0.f; + + for (LLCharacter* iter : LLCharacter::sInstances) + { + LLVOAvatar* inst = (LLVOAvatar*) iter; + ret += inst->getGPURenderTime(); + } + + return ret; +} + +F32 LLVOAvatar::getMaxGPURenderTime() +{ + LL_PROFILE_ZONE_SCOPED_CATEGORY_AVATAR; + + F32 ret = 0.f; + + for (LLCharacter* iter : LLCharacter::sInstances) + { + LLVOAvatar* inst = (LLVOAvatar*)iter; + ret = llmax(inst->getGPURenderTime(), ret); + } + + return ret; +} + +F32 LLVOAvatar::getAverageGPURenderTime() +{ + LL_PROFILE_ZONE_SCOPED_CATEGORY_AVATAR; + + F32 ret = 0.f; + + S32 count = 0; + + for (LLCharacter* iter : LLCharacter::sInstances) + { + LLVOAvatar* inst = (LLVOAvatar*)iter; + if (!inst->isTooSlow()) + { + ret += inst->getGPURenderTime(); + ++count; + } + } + + if (count > 0) + { + ret /= count; + } + + return ret; +} + + diff --git a/indra/newview/llvoavatar.h b/indra/newview/llvoavatar.h index 2ca44b041a..e8604464ae 100644 --- a/indra/newview/llvoavatar.h +++ b/indra/newview/llvoavatar.h @@ -309,11 +309,21 @@ public: void readProfileQuery(S32 retries); // get the GPU time in ms of rendering this avatar including all attachments - // returns -1 if this avatar has not been profiled using gPipeline.profileAvatar - F32 getGPURenderTime() { return mGPURenderTime; } + // returns 0.f if this avatar has not been profiled using gPipeline.profileAvatar + // or the avatar is visually muted + F32 getGPURenderTime(); + + // get the total GPU render time in ms of all avatars that have been benched + static F32 getTotalGPURenderTime(); + + // get the max GPU render time in ms of all avatars that have been benched + static F32 getMaxGPURenderTime(); + + // get the average GPU render time in ms of all avatars that have been benched + static F32 getAverageGPURenderTime(); // get the CPU time in ms of rendering this avatar including all attachments - // return -1 if this avatar has not been profiled using gPipeline.mProfileAvatar + // return 0.f if this avatar has not been profiled using gPipeline.mProfileAvatar F32 getCPURenderTime() { return mCPURenderTime; } @@ -401,8 +411,7 @@ public: void logMetricsTimerRecord(const std::string& phase_name, F32 elapsed, bool completed); void calcMutedAVColor(); - void markARTStale(); - + protected: LLViewerStats::PhaseMap& getPhases() { return mPhases; } BOOL updateIsFullyLoaded(); @@ -423,11 +432,6 @@ private: LLFrameTimer mFullyLoadedTimer; LLFrameTimer mRuthTimer; - U32 mLastARTUpdateFrame{0}; - U64 mRenderTime{0}; - U64 mGeomTime{0}; - bool mARTStale{true}; - bool mARTCapped{false}; // variables to hold "slowness" status bool mTooSlow{false}; bool mTooSlowWithoutShadows{false}; @@ -557,11 +561,11 @@ private: // profile results // GPU render time in ms - F32 mGPURenderTime = -1.f; + F32 mGPURenderTime = 0.f; bool mGPUProfilePending = false; // CPU render time in ms - F32 mCPURenderTime = -1.f; + F32 mCPURenderTime = 0.f; // the isTooComplex method uses these mutable values to avoid recalculating too frequently // DEPRECATED -- obsolete avatar render cost values @@ -1203,8 +1207,6 @@ public: // COF version of last appearance message received for this av. S32 mLastUpdateReceivedCOFVersion; - U64 getLastART() const { return mRenderTime; } - /** Diagnostics ** ** *******************************************************************************/ diff --git a/indra/newview/llworld.cpp b/indra/newview/llworld.cpp index 732ef1da6c..0e0dbdc071 100644 --- a/indra/newview/llworld.cpp +++ b/indra/newview/llworld.cpp @@ -1441,7 +1441,11 @@ F32 LLWorld::getNearbyAvatarsAndMaxGPUTime(std::vector<LLCharacter*> &valid_near char_iter++; continue; } - gPipeline.profileAvatar(avatar); + + if (!avatar->isTooSlow()) + { + gPipeline.profileAvatar(avatar); + } nearby_max_complexity = llmax(nearby_max_complexity, avatar->getGPURenderTime()); valid_nearby_avs.push_back(*char_iter); } diff --git a/indra/newview/pipeline.cpp b/indra/newview/pipeline.cpp index 4d9a8a594a..6ea86c4cbb 100644 --- a/indra/newview/pipeline.cpp +++ b/indra/newview/pipeline.cpp @@ -10073,6 +10073,9 @@ void LLPipeline::profileAvatar(LLVOAvatar* avatar, bool profile_attachments) LL_PROFILE_ZONE_SCOPED_CATEGORY_PIPELINE; + // don't continue to profile an avatar that is known to be too slow + llassert(!avatar->isTooSlow()); + LLGLSLShader* cur_shader = LLGLSLShader::sCurBoundShaderPtr; mRT->deferredScreen.bindTarget(); @@ -10345,25 +10348,28 @@ void LLPipeline::generateImpostor(LLVOAvatar* avatar, bool preview_avatar, bool resY = llmin(nhpo2((U32) (fov*pa)), (U32) 512); resX = llmin(nhpo2((U32) (atanf(tdim.mV[0]/distance)*2.f*RAD_TO_DEG*pa)), (U32) 512); - if (!avatar->mImpostor.isComplete()) - { - avatar->mImpostor.allocate(resX, resY, GL_RGBA, true); + if (!for_profile) + { + if (!avatar->mImpostor.isComplete()) + { + avatar->mImpostor.allocate(resX, resY, GL_RGBA, true); - if (LLPipeline::sRenderDeferred) - { - addDeferredAttachments(avatar->mImpostor, true); - } - - gGL.getTexUnit(0)->bind(&avatar->mImpostor); - gGL.getTexUnit(0)->setTextureFilteringOption(LLTexUnit::TFO_POINT); - gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE); - } - else if(resX != avatar->mImpostor.getWidth() || resY != avatar->mImpostor.getHeight()) - { - avatar->mImpostor.resize(resX,resY); - } + if (LLPipeline::sRenderDeferred) + { + addDeferredAttachments(avatar->mImpostor, true); + } + + gGL.getTexUnit(0)->bind(&avatar->mImpostor); + gGL.getTexUnit(0)->setTextureFilteringOption(LLTexUnit::TFO_POINT); + gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE); + } + else if (resX != avatar->mImpostor.getWidth() || resY != avatar->mImpostor.getHeight()) + { + avatar->mImpostor.resize(resX, resY); + } - avatar->mImpostor.bindTarget(); + avatar->mImpostor.bindTarget(); + } } F32 old_alpha = LLDrawPoolAvatar::sMinimumAlpha; @@ -10466,7 +10472,7 @@ void LLPipeline::generateImpostor(LLVOAvatar* avatar, bool preview_avatar, bool gGL.popMatrix(); } - if (!preview_avatar) + if (!preview_avatar && !for_profile) { avatar->mImpostor.flush(); avatar->setImpostorDim(tdim); |