From b0d80552b3177269eb24ef9a1e0d58a98d95a96f Mon Sep 17 00:00:00 2001
From: Andrey Kleshchev <117672381+akleshchev@users.noreply.github.com>
Date: Sat, 15 Feb 2025 09:51:22 +0200
Subject: #2639 GPU test twice to fix misreported bandwidth (#3558)

A workaround, hopefully temporary.
First test returns half the throughput, do two tests. May be caused by
driver, may be some 'energy saving', will need to be checked later.
It was working the same way prior to ExtraFPS, but viewer was running
an extra CPU test.
---
 indra/newview/llglsandbox.cpp | 66 +++++++++++++++++++++++++------------------
 1 file changed, 39 insertions(+), 27 deletions(-)

(limited to 'indra/newview')

diff --git a/indra/newview/llglsandbox.cpp b/indra/newview/llglsandbox.cpp
index 112008172e..5484ce6276 100644
--- a/indra/newview/llglsandbox.cpp
+++ b/indra/newview/llglsandbox.cpp
@@ -903,6 +903,39 @@ private:
 };
 
 
+F32 shader_timer_benchmark(std::vector<LLRenderTarget> & dest, TextureHolder & texHolder, U32 textures_count, LLVertexBuffer * buff, F32 &seconds)
+{
+    // run GPU timer benchmark
+
+    //number of samples to take
+    const S32 samples = 64;
+
+    {
+        ShaderProfileHelper initProfile;
+        dest[0].bindTarget();
+        gBenchmarkProgram.bind();
+        for (S32 c = 0; c < samples; ++c)
+        {
+            for (U32 i = 0; i < textures_count; ++i)
+            {
+                texHolder.bind(i);
+                buff->setBuffer();
+                buff->drawArrays(LLRender::TRIANGLES, 0, 3);
+            }
+        }
+        gBenchmarkProgram.unbind();
+        dest[0].flush();
+    }
+
+    F32 ms = gBenchmarkProgram.mTimeElapsed / 1000000.f;
+    seconds = ms / 1000.f;
+
+    F64 samples_drawn = (F64)gBenchmarkProgram.mSamplesDrawn;
+    F64 gpixels_drawn = samples_drawn / 1000000000.0;
+    F32 samples_sec = (F32)(gpixels_drawn / seconds);
+    return samples_sec * 4;  // 4 bytes per sample
+}
+
 //-----------------------------------------------------------------------------
 // gpu_benchmark()
 //  returns measured memory bandwidth of GPU in gigabytes per second
@@ -944,9 +977,6 @@ F32 gpu_benchmark()
     //number of textures
     const U32 count = 32;
 
-    //number of samples to take
-    const S32 samples = 64;
-
     //time limit, allocation operations shouldn't take longer then 30 seconds, same for actual benchmark.
     const F32 time_limit = 30;
 
@@ -1036,33 +1066,15 @@ F32 gpu_benchmark()
 
     LLGLSLShader::unbind();
 
-    // run GPU timer benchmark
-    {
-        ShaderProfileHelper initProfile;
-        dest[0].bindTarget();
-        gBenchmarkProgram.bind();
-        for (S32 c = 0; c < samples; ++c)
-        {
-            for (U32 i = 0; i < count; ++i)
-            {
-                texHolder.bind(i);
-                buff->setBuffer();
-                buff->drawArrays(LLRender::TRIANGLES, 0, 3);
-            }
-        }
-        gBenchmarkProgram.unbind();
-        dest[0].flush();
-    }
+    // run GPU timer benchmark twice
+    F32 seconds = 0;
+    F32 gbps = shader_timer_benchmark(dest, texHolder, count, buff.get(), seconds);
 
-    F32 ms = gBenchmarkProgram.mTimeElapsed/1000000.f;
-    F32 seconds = ms/1000.f;
+    LL_INFOS("Benchmark") << "Memory bandwidth, 1st run is " << llformat("%.3f", gbps) << " GB/sec according to ARB_timer_query, total time " << seconds << " seconds" << LL_ENDL;
 
-    F64 samples_drawn = (F64)gBenchmarkProgram.mSamplesDrawn;
-    F64 gpixels_drawn = samples_drawn / 1000000000.0;
-    F32 samples_sec = (F32)(gpixels_drawn/seconds);
-    F32 gbps = samples_sec*4;  // 4 bytes per sample
+    gbps = shader_timer_benchmark(dest, texHolder, count, buff.get(), seconds);
 
-    LL_INFOS("Benchmark") << "Memory bandwidth is " << llformat("%.3f", gbps) << " GB/sec according to ARB_timer_query, total time " << seconds << " seconds" << LL_ENDL;
+    LL_INFOS("Benchmark") << "Memory bandwidth, final run is " << llformat("%.3f", gbps) << " GB/sec according to ARB_timer_query, total time " << seconds << " seconds" << LL_ENDL;
 
     return gbps;
 }
-- 
cgit v1.2.3