From 02f887ede0037c4ca30c84ccc5f5b58069da2f5f Mon Sep 17 00:00:00 2001
From: Stenzek <stenzek@gmail.com>
Date: Sun, 27 Nov 2016 18:15:02 +1000
Subject: [PATCH] OGL: Add GPUTimer class for measuring execution time of a
 draw/dispatch

---
 Source/Core/VideoBackends/OGL/GPUTimer.h      | 105 ++++++++++++++++++
 Source/Core/VideoBackends/OGL/OGL.vcxproj     |   3 +-
 .../VideoBackends/OGL/OGL.vcxproj.filters     |   5 +-
 .../Core/VideoBackends/OGL/TextureCache.cpp   |  12 ++
 4 files changed, 123 insertions(+), 2 deletions(-)
 create mode 100644 Source/Core/VideoBackends/OGL/GPUTimer.h

diff --git a/Source/Core/VideoBackends/OGL/GPUTimer.h b/Source/Core/VideoBackends/OGL/GPUTimer.h
new file mode 100644
index 0000000000..50724ab06f
--- /dev/null
+++ b/Source/Core/VideoBackends/OGL/GPUTimer.h
@@ -0,0 +1,105 @@
+// Copyright 2016 Dolphin Emulator Project
+// Licensed under GPLv2+
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "Common/GL/GLExtensions/GLExtensions.h"
+
+#ifndef GL_TIME_ELAPSED
+#define GL_TIME_ELAPSED 0x88BF
+#endif
+
+namespace OGL
+{
+/*
+ * This class can be used to measure the time it takes for the GPU to perform a draw call
+ * or compute dispatch. To use:
+ *
+ *   - Create an instance of GPUTimer before issuing the draw call.
+ *     (this can be before or after any binding that needs to be done)
+ *
+ *   - (optionally) call Begin(). This is not needed for a single draw call.
+ *
+ *   - Issue the draw call or compute dispatch as normal.
+ *
+ *   - (optionally) call End(). This is not necessary for a single draw call.
+ *
+ *   - Call GetTime{Seconds,Milliseconds,Nanoseconds} to determine how long the operation
+ *     took to execute on the GPU.
+ *
+ * NOTE: When the timer is read back, this will force a GL flush, so the more often a timer is used,
+ * the larger of a performance impact it will have. Only one timer can be active at any time, due to
+ * using GL_TIME_ELAPSED. This is not enforced by the class, however.
+ *
+ */
+class GPUTimer final
+{
+public:
+  GPUTimer()
+  {
+    glGenQueries(1, &m_query_id);
+    Begin();
+  }
+
+  ~GPUTimer()
+  {
+    End();
+    glDeleteQueries(1, &m_query_id);
+  }
+
+  void Begin()
+  {
+    if (m_started)
+      glEndQuery(GL_TIME_ELAPSED);
+
+    glBeginQuery(GL_TIME_ELAPSED, m_query_id);
+    m_started = true;
+  }
+
+  void End()
+  {
+    if (!m_started)
+      return;
+
+    glEndQuery(GL_TIME_ELAPSED);
+    m_started = false;
+  }
+
+  double GetTimeSeconds()
+  {
+    GetResult();
+    return static_cast<double>(m_result) / 1000000000.0;
+  }
+
+  double GetTimeMilliseconds()
+  {
+    GetResult();
+    return static_cast<double>(m_result) / 1000000.0;
+  }
+
+  u32 GetTimeNanoseconds()
+  {
+    GetResult();
+    return m_result;
+  }
+
+private:
+  void GetResult()
+  {
+    if (m_has_result)
+      return;
+
+    if (m_started)
+      End();
+
+    glGetQueryObjectuiv(m_query_id, GL_QUERY_RESULT, &m_result);
+    m_has_result = true;
+  }
+
+  GLuint m_query_id;
+  GLuint m_result = 0;
+  bool m_started = false;
+  bool m_has_result = false;
+};
+}  // namespace OGL
diff --git a/Source/Core/VideoBackends/OGL/OGL.vcxproj b/Source/Core/VideoBackends/OGL/OGL.vcxproj
index 0c234a8e71..3b945793a4 100644
--- a/Source/Core/VideoBackends/OGL/OGL.vcxproj
+++ b/Source/Core/VideoBackends/OGL/OGL.vcxproj
@@ -53,6 +53,7 @@
   <ItemGroup>
     <ClInclude Include="BoundingBox.h" />
     <ClInclude Include="FramebufferManager.h" />
+    <ClInclude Include="GPUTimer.h" />
     <ClInclude Include="PerfQuery.h" />
     <ClInclude Include="PostProcessing.h" />
     <ClInclude Include="ProgramShaderCache.h" />
@@ -79,4 +80,4 @@
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>
-</Project>
+</Project>
\ No newline at end of file
diff --git a/Source/Core/VideoBackends/OGL/OGL.vcxproj.filters b/Source/Core/VideoBackends/OGL/OGL.vcxproj.filters
index 282934d830..201a4045f8 100644
--- a/Source/Core/VideoBackends/OGL/OGL.vcxproj.filters
+++ b/Source/Core/VideoBackends/OGL/OGL.vcxproj.filters
@@ -90,8 +90,11 @@
     </ClInclude>
     <ClInclude Include="SamplerCache.h" />
     <ClInclude Include="VideoBackend.h" />
+    <ClInclude Include="GPUTimer.h">
+      <Filter>GLUtil</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <Text Include="CMakeLists.txt" />
   </ItemGroup>
-</Project>
+</Project>
\ No newline at end of file
diff --git a/Source/Core/VideoBackends/OGL/TextureCache.cpp b/Source/Core/VideoBackends/OGL/TextureCache.cpp
index 72b04fc5b1..bd6f99b2e8 100644
--- a/Source/Core/VideoBackends/OGL/TextureCache.cpp
+++ b/Source/Core/VideoBackends/OGL/TextureCache.cpp
@@ -16,6 +16,7 @@
 #include "Common/StringUtil.h"
 
 #include "VideoBackends/OGL/FramebufferManager.h"
+#include "VideoBackends/OGL/GPUTimer.h"
 #include "VideoBackends/OGL/ProgramShaderCache.h"
 #include "VideoBackends/OGL/Render.h"
 #include "VideoBackends/OGL/SamplerCache.h"
@@ -62,6 +63,8 @@ struct TextureDecodingProgramInfo
   bool valid = false;
 };
 
+//#define TIME_TEXTURE_DECODING 1
+
 static std::map<std::pair<u32, u32>, TextureDecodingProgramInfo> s_texture_decoding_program_info;
 static std::array<GLuint, TextureConversionShader::BUFFER_FORMAT_COUNT>
     s_texture_decoding_buffer_views;
@@ -713,6 +716,10 @@ void TextureCache::DecodeTextureOnGPU(TCacheEntryBase* entry, u32 dst_level, con
   if (iter == s_texture_decoding_program_info.end())
     return;
 
+#ifdef TIME_TEXTURE_DECODING
+  GPUTimer timer;
+#endif
+
   // Copy to GPU-visible buffer, aligned to the data type.
   auto info = iter->second;
   u32 bytes_per_buffer_elem =
@@ -775,5 +782,10 @@ void TextureCache::DecodeTextureOnGPU(TCacheEntryBase* entry, u32 dst_level, con
   glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT);
 
   TextureCache::SetStage();
+
+#ifdef TIME_TEXTURE_DECODING
+  WARN_LOG(VIDEO, "Decode texture format %u size %ux%u took %.4fms", static_cast<u32>(format),
+           width, height, timer.GetTimeMilliseconds());
+#endif
 }
 }