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/llglsandbox.cpp | 31 ++++++++++++++++++++++++++----- 1 file changed, 26 insertions(+), 5 deletions(-) (limited to 'indra/newview/llglsandbox.cpp') 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; } -- 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/llglsandbox.cpp') 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 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/llglsandbox.cpp') 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 From 4caa2746f3b1d336947dcc9e277af8b8c08b44e0 Mon Sep 17 00:00:00 2001 From: davep Date: Fri, 19 Sep 2014 13:01:41 -0500 Subject: MAINT-3131 Discard improbably high memory bandwidth measurements on OSX and default to Class 3 --- indra/newview/llglsandbox.cpp | 15 +++++++++++++-- 1 file changed, 13 insertions(+), 2 deletions(-) (limited to 'indra/newview/llglsandbox.cpp') diff --git a/indra/newview/llglsandbox.cpp b/indra/newview/llglsandbox.cpp index 0548de3895..4b8ac2b3cf 100755 --- a/indra/newview/llglsandbox.cpp +++ b/indra/newview/llglsandbox.cpp @@ -928,7 +928,7 @@ F32 gpu_benchmark() std::vector results; //build a random texture - U8 pixels[res*res*4]; + U8* pixels = new U8[res*res*4]; for (U32 i = 0; i < res*res*4; ++i) { @@ -950,6 +950,8 @@ F32 gpu_benchmark() LLImageGL::setManualImage(GL_TEXTURE_2D, 0, GL_RGBA, res,res,GL_RGBA, GL_UNSIGNED_BYTE, pixels); } + delete [] pixels; + //make a dummy triangle to draw with LLPointer buff = new LLVertexBuffer(LLVertexBuffer::MAP_VERTEX | LLVertexBuffer::MAP_TEXCOORD0, GL_STATIC_DRAW_ARB); buff->allocateBuffer(3, 0, true); @@ -1031,7 +1033,16 @@ F32 gpu_benchmark() F32 gbps = results[results.size()/2]; LL_INFOS() << "Memory bandwidth is " << llformat("%.3f", gbps) << "GB/sec according to CPU timers" << LL_ENDL; - + +#if LL_DARWIN + if (gbps > 512.f) + { + LL_INFOS() << "Memory bandwidth is improbably high and likely incorrect." << LL_ENDL; + //OSX is probably lying, discard result + gbps = -1.f; + } +#endif + if (gGLManager.mHasTimerQuery) { F32 ms = gBenchmarkProgram.mTimeElapsed/1000000.f; -- cgit v1.2.3