From 24f8745914f7cec320d707f36895ac393575b861 Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Wed, 19 Mar 2014 17:57:00 -0500 Subject: MAINT-3131 Use benchmark to determine GPU class instead of GPU table. --- indra/newview/llfeaturemanager.cpp | 63 +++++++++++++++++++++++++++++++++++--- indra/newview/llglsandbox.cpp | 31 ++++++++++++++++--- indra/newview/llviewermenu.cpp | 2 +- 3 files changed, 86 insertions(+), 10 deletions(-) (limited to 'indra/newview') diff --git a/indra/newview/llfeaturemanager.cpp b/indra/newview/llfeaturemanager.cpp index ba6f26d3ef..938ea707f9 100755 --- a/indra/newview/llfeaturemanager.cpp +++ b/indra/newview/llfeaturemanager.cpp @@ -417,13 +417,67 @@ bool LLFeatureManager::parseFeatureTable(std::string filename) return parse_ok; } +F32 gpu_benchmark(); + bool LLFeatureManager::loadGPUClass() { + //get memory bandwidth from benchmark + F32 gbps = gpu_benchmark(); + + if (gbps < -1.f) + { //couldn't bench, use GLVersion + + if (gGLManager.mGLVersion < 2.f) + { + mGPUClass = GPU_CLASS_0; + } + else if (gGLManager.mGLVersion < 3.f) + { + mGPUClass = GPU_CLASS_1; + } + else if (gGLManager.mGLVersion < 3.3f) + { + mGPUClass = GPU_CLASS_2; + } + else if (gGLManager.mGLVersion < 4.f) + { + mGPUClass = GPU_CLASS_3; + } + else + { + mGPUClass = GPU_CLASS_4; + } + } + else if (gbps < 5.f) + { + mGPUClass = GPU_CLASS_0; + } + else if (gbps < 10.f) + { + mGPUClass = GPU_CLASS_1; + } + else if (gbps < 20.f) + { + mGPUClass = GPU_CLASS_2; + } + else if (gbps < 40.f) + { + mGPUClass = GPU_CLASS_3; + } + else if (gbps < 80.f) + { + mGPUClass = GPU_CLASS_4; + } + else + { + mGPUClass = GPU_CLASS_5; + } + // defaults - mGPUClass = GPU_CLASS_UNKNOWN; mGPUString = gGLManager.getRawGLString(); - mGPUSupported = FALSE; + mGPUSupported = TRUE; +#if 0 // first table is in the app dir std::string app_path = gDirUtilp->getAppRODataDir(); app_path += gDirUtilp->getDirDelimiter(); @@ -451,8 +505,8 @@ bool LLFeatureManager::loadGPUClass() { parse_ok = parseGPUTable(app_path); } - - return parse_ok; // indicates that the file parsed correctly, not that the gpu was recognized +#endif + return true; // indicates that the file parsed correctly, not that the gpu was recognized } @@ -719,6 +773,7 @@ void LLFeatureManager::init() void LLFeatureManager::applyRecommendedSettings() { + loadGPUClass(); // apply saved settings // cap the level at 2 (high) U32 level = llmax(GPU_CLASS_0, llmin(mGPUClass, GPU_CLASS_5)); diff --git a/indra/newview/llglsandbox.cpp b/indra/newview/llglsandbox.cpp index c4c1827266..92553e875d 100755 --- a/indra/newview/llglsandbox.cpp +++ b/indra/newview/llglsandbox.cpp @@ -878,13 +878,32 @@ void LLViewerObjectList::renderObjectBeacons() } -void gpu_benchmark() +F32 gpu_benchmark() { - if (!LLGLSLShader::sNoFixedFunction) - { //don't bother benchmarking the fixed function - return; + if (!gGLManager.mHasShaderObjects || !gGLManager.mHasTimerQuery) + { //don't bother benchmarking the fixed function or using CPU timers + return -1.f; } + + if (gBenchmarkProgram.mProgramObject == 0) + { + LLViewerShaderMgr::instance()->initAttribsAndUniforms(); + + gBenchmarkProgram.mName = "Benchmark Shader"; + gBenchmarkProgram.mFeatures.attachNothing = true; + gBenchmarkProgram.mShaderFiles.clear(); + gBenchmarkProgram.mShaderFiles.push_back(std::make_pair("interface/benchmarkV.glsl", GL_VERTEX_SHADER_ARB)); + gBenchmarkProgram.mShaderFiles.push_back(std::make_pair("interface/benchmarkF.glsl", GL_FRAGMENT_SHADER_ARB)); + gBenchmarkProgram.mShaderLevel = 1; + if (!gBenchmarkProgram.createShader(NULL, NULL)) + { + return -1.f; + } + } + + LLGLDisable blend(GL_BLEND); + //measure memory bandwidth by: // - allocating a batch of textures and render targets // - rendering those textures to those render targets @@ -981,7 +1000,7 @@ void gpu_benchmark() gBenchmarkProgram.unbind(); - LLGLSLShader::finishProfile(); + LLGLSLShader::finishProfile(false); LLImageGL::deleteTextures(count, source); @@ -1007,5 +1026,7 @@ void gpu_benchmark() { llinfos << "ARB_timer_query unavailable." << llendl; } + + return gbps; } diff --git a/indra/newview/llviewermenu.cpp b/indra/newview/llviewermenu.cpp index c7c8da27f3..d5acec7c9a 100755 --- a/indra/newview/llviewermenu.cpp +++ b/indra/newview/llviewermenu.cpp @@ -7197,7 +7197,7 @@ class LLAdvancedClickRenderProfile: public view_listener_t } }; -void gpu_benchmark(); +F32 gpu_benchmark(); class LLAdvancedClickRenderBenchmark: public view_listener_t { -- cgit v1.2.3 From 7defa95f6778778d850aa1aba3ea8cb16396d0dd Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Fri, 21 Mar 2014 14:30:14 -0500 Subject: MAINT-3131 fallback to CPU timers for return value of benchmark if ARB_timer_query is unavailable. --- indra/newview/llglsandbox.cpp | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) (limited to 'indra/newview') diff --git a/indra/newview/llglsandbox.cpp b/indra/newview/llglsandbox.cpp index 92553e875d..f620bf19fb 100755 --- a/indra/newview/llglsandbox.cpp +++ b/indra/newview/llglsandbox.cpp @@ -880,8 +880,8 @@ void LLViewerObjectList::renderObjectBeacons() F32 gpu_benchmark() { - if (!gGLManager.mHasShaderObjects || !gGLManager.mHasTimerQuery) - { //don't bother benchmarking the fixed function or using CPU timers + if (!gGLManager.mHasShaderObjects) + { //don't bother benchmarking the fixed function return -1.f; } @@ -1011,15 +1011,15 @@ F32 gpu_benchmark() llinfos << "Memory bandwidth is " << llformat("%.3f", gbps) << "GB/sec according to CPU timers" << llendl; - F32 ms = gBenchmarkProgram.mTimeElapsed/1000000.f; - F32 seconds = ms/1000.f; - - F64 samples_drawn = res*res*count*samples; - F32 samples_sec = (samples_drawn/1000000000.0)/seconds; - gbps = samples_sec*8; - if (gGLManager.mHasTimerQuery) { + F32 ms = gBenchmarkProgram.mTimeElapsed/1000000.f; + F32 seconds = ms/1000.f; + + F64 samples_drawn = res*res*count*samples; + F32 samples_sec = (samples_drawn/1000000000.0)/seconds; + gbps = samples_sec*8; + llinfos << "Memory bandwidth is " << llformat("%.3f", gbps) << "GB/sec according to ARB_timer_query" << llendl; } else -- cgit v1.2.3 From c142696f9dd18dcebee903444225026caf08e6e3 Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Fri, 21 Mar 2014 14:33:46 -0500 Subject: MAINT-3131 Actually fallback to GL version guess instead of treating error value from gpu_benchmark as actual memory bandwidth. --- indra/newview/llfeaturemanager.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'indra/newview') diff --git a/indra/newview/llfeaturemanager.cpp b/indra/newview/llfeaturemanager.cpp index 938ea707f9..5f37c0ea73 100755 --- a/indra/newview/llfeaturemanager.cpp +++ b/indra/newview/llfeaturemanager.cpp @@ -424,7 +424,7 @@ bool LLFeatureManager::loadGPUClass() //get memory bandwidth from benchmark F32 gbps = gpu_benchmark(); - if (gbps < -1.f) + if (gbps < 0.f) { //couldn't bench, use GLVersion if (gGLManager.mGLVersion < 2.f) -- cgit v1.2.3 From e43f43b4d9dbcaeb20884772b3fcba6df5e9a62c Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Mon, 24 Mar 2014 11:56:02 -0500 Subject: MAINT-3131 Attempt to work around drivers that lack timer query and ignore glFinish calls. --- indra/newview/llglsandbox.cpp | 24 ++++++++++++++++++++++-- 1 file changed, 22 insertions(+), 2 deletions(-) (limited to 'indra/newview') diff --git a/indra/newview/llglsandbox.cpp b/indra/newview/llglsandbox.cpp index f620bf19fb..0a423e6785 100755 --- a/indra/newview/llglsandbox.cpp +++ b/indra/newview/llglsandbox.cpp @@ -969,6 +969,8 @@ F32 gpu_benchmark() //wait for any previoius GL commands to finish glFinish(); + bool busted_finish = false; + for (S32 c = -1; c < samples; ++c) { LLTimer timer; @@ -983,7 +985,18 @@ F32 gpu_benchmark() } //wait for current batch of copies to finish - glFinish(); + if (busted_finish) + { + //read a pixel off the last target since some drivers seem to ignore glFinish + dest[count-1].bindTarget(); + U32 pixel = 0; + glReadPixels(0,0,1,1,GL_RGBA, GL_UNSIGNED_BYTE, &pixel); + dest[count-1].flush(); + } + else + { + glFinish(); + } F32 time = timer.getElapsedTimeF32(); @@ -994,7 +1007,14 @@ F32 gpu_benchmark() F32 gbps = gb/time; - results.push_back(gbps); + if (!gGLManager.mHasTimerQuery && !busted_finish && gbps > 128.f) + { //unrealistically high bandwidth for a card without timer queries, glFinish is probably ignored + busted_finish = true; + } + else + { + results.push_back(gbps); + } } } -- cgit v1.2.3