From 2b607e0f8f855a33bf653a69a1e8d5a85820f72f Mon Sep 17 00:00:00 2001
From: bkaradzic <branimirkaradzic@gmail.com>
Date: Tue, 26 Jun 2012 19:01:10 -0700
Subject: [PATCH] Fixed texture flags shifts. Added GL GPU timing.

---
 include/bgfx.h      |  8 ++++----
 src/renderer_gl.cpp | 37 +++++++++++++++++++++++++++++++++----
 src/renderer_gl.h   | 34 ++++++++++++++++++++++++++++++++++
 3 files changed, 71 insertions(+), 8 deletions(-)

diff --git a/include/bgfx.h b/include/bgfx.h
index 67efe4d4..02def501 100644
--- a/include/bgfx.h
+++ b/include/bgfx.h
@@ -112,16 +112,16 @@ namespace bgfx
 #define BGFX_TEXTURE_W_REPEAT           UINT32_C(0x00000000)
 #define BGFX_TEXTURE_W_MIRROR           UINT32_C(0x00000100)
 #define BGFX_TEXTURE_W_CLAMP            UINT32_C(0x00000200)
-#define BGFX_TEXTURE_W_SHIFT            4
+#define BGFX_TEXTURE_W_SHIFT            8
 #define BGFX_TEXTURE_W_MASK             UINT32_C(0x00000300)
 #define BGFX_TEXTURE_MIN_POINT          UINT32_C(0x00001000)
-#define BGFX_TEXTURE_MIN_SHIFT          8
+#define BGFX_TEXTURE_MIN_SHIFT          12
 #define BGFX_TEXTURE_MIN_MASK           UINT32_C(0x00001000)
 #define BGFX_TEXTURE_MAG_POINT          UINT32_C(0x00010000)
-#define BGFX_TEXTURE_MAG_SHIFT          12
+#define BGFX_TEXTURE_MAG_SHIFT          16
 #define BGFX_TEXTURE_MAG_MASK           UINT32_C(0x00010000)
 #define BGFX_TEXTURE_MIP_POINT          UINT32_C(0x00100000)
-#define BGFX_TEXTURE_MIP_SHIFT          16
+#define BGFX_TEXTURE_MIP_SHIFT          20
 #define BGFX_TEXTURE_MIP_MASK           UINT32_C(0x00100000)
 #define BGFX_TEXTURE_SRGB               UINT32_C(0x00200000)
 
diff --git a/src/renderer_gl.cpp b/src/renderer_gl.cpp
index 627bbbf2..7abfb65b 100644
--- a/src/renderer_gl.cpp
+++ b/src/renderer_gl.cpp
@@ -406,10 +406,17 @@ namespace bgfx
 		void init()
 		{
 			setRenderContextSize(BGFX_DEFAULT_WIDTH, BGFX_DEFAULT_HEIGHT);
+#if BGFX_CONFIG_RENDERER_OPENGL
+			m_queries.create();
+#endif // BGFX_CONFIG_RENDERER_OPENGL
 		}
 
 		void shutdown()
 		{
+#if BGFX_CONFIG_RENDERER_OPENGL
+			m_queries.destroy();
+#endif // BGFX_CONFIG_RENDERER_OPENGL
+
 #if BGFX_USE_WGL
 			if (NULL != m_hdc)
 			{
@@ -438,6 +445,9 @@ namespace bgfx
 		RenderTarget m_renderTargets[BGFX_CONFIG_MAX_RENDER_TARGETS];
 		UniformRegistry m_uniformReg;
 		void* m_uniforms[BGFX_CONFIG_MAX_UNIFORMS];
+#if BGFX_CONFIG_RENDERER_OPENGL
+		Queries m_queries;
+#endif // BGFX_CONFIG_RENDERER_OPENGL
 
 		TextVideoMem m_textVideoMem;
 
@@ -1841,6 +1851,14 @@ namespace bgfx
 	{
 		s_renderCtx.updateResolution(m_render->m_resolution);
 
+		int64_t elapsed = -bx::getHPCounter();
+#if BGFX_CONFIG_RENDERER_OPENGL
+		if (m_render->m_debug & (BGFX_DEBUG_IFH|BGFX_DEBUG_STATS) )
+		{
+			s_renderCtx.m_queries.begin(0, GL_TIME_ELAPSED);
+		}
+#endif // BGFX_CONFIG_RENDERER_OPENGL
+
 		if (0 < m_render->m_iboffset)
 		{
 			TransientIndexBuffer* ib = m_render->m_transientIb;
@@ -1880,8 +1898,6 @@ namespace bgfx
 		uint32_t statsNumPrims = 0;
 		uint32_t statsNumIndices = 0;
 
-		int64_t elapsed = -bx::getHPCounter();
-
 		if (0 == (m_render->m_debug&BGFX_DEBUG_IFH) )
 		{
 			for (uint32_t item = 0, numItems = m_render->m_num; item < numItems; ++item)
@@ -2373,6 +2389,13 @@ namespace bgfx
 
 		if (m_render->m_debug & (BGFX_DEBUG_IFH|BGFX_DEBUG_STATS) )
 		{
+			double elapsedGpuMs = 0.0;
+#if BGFX_CONFIG_RENDERER_OPENGL
+			s_renderCtx.m_queries.end(GL_TIME_ELAPSED);
+			uint64_t elapsedGl = s_renderCtx.m_queries.getResult(0);
+			elapsedGpuMs = double(elapsedGl)/1e6;
+#endif // BGFX_CONFIG_RENDERER_OPENGL
+
 			TextVideoMem& tvm = s_renderCtx.m_textVideoMem;
 
 			static int64_t next = now;
@@ -2382,11 +2405,17 @@ namespace bgfx
 				next = now + bx::getHPFrequency();
 				double freq = double(bx::getHPFrequency() );
 				double toMs = 1000.0/freq;
+				double elapsedCpuMs = double(elapsed)*toMs;
 
 				tvm.clear();
 				uint16_t pos = 10;
-				tvm.printf(10, pos++, 0x8e, "      Frame: %3.4f [ms] / %3.2f", double(frameTime)*toMs, freq/frameTime);
-				tvm.printf(10, pos++, 0x8e, " Draw calls: %4d / %3.4f [ms]", m_render->m_num, double(elapsed)*toMs);
+				tvm.printf(10, pos++, 0x8e, "  Frame CPU: %3.4f [ms] / %3.2f", double(frameTime)*toMs, freq/frameTime);
+				tvm.printf(10, pos++, 0x8e, " Draw calls: %4d / CPU %3.4f [ms] %c GPU %3.4f [ms]"
+					, m_render->m_num
+					, elapsedCpuMs
+					, elapsedCpuMs > elapsedGpuMs ? '>' : '<'
+					, elapsedGpuMs
+					);
 				tvm.printf(10, pos++, 0x8e, "      Prims: %7d", statsNumPrims);
 				tvm.printf(10, pos++, 0x8e, "    Indices: %7d", statsNumIndices);
 				tvm.printf(10, pos++, 0x8e, "   DVB size: %7d", m_render->m_vboffset);
diff --git a/src/renderer_gl.h b/src/renderer_gl.h
index fdc07daa..44afef13 100644
--- a/src/renderer_gl.h
+++ b/src/renderer_gl.h
@@ -350,6 +350,40 @@ namespace bgfx
 		uint8_t m_numPredefined;
 	};
 
+#if BGFX_CONFIG_RENDERER_OPENGL
+	struct Queries
+	{
+		void create()
+		{
+			glGenQueries(countof(m_queries), m_queries);
+		}
+
+		void destroy()
+		{
+			glDeleteQueries(countof(m_queries), m_queries);
+		}
+
+		void begin(uint16_t _id, GLenum _target) const
+		{
+			glBeginQuery(_target, m_queries[_id]);
+		}
+
+		void end(GLenum _target) const
+		{
+			glEndQuery(_target);
+		}
+
+		uint64_t getResult(uint16_t _id) const
+		{
+			uint64_t result;
+			glGetQueryObjectui64v(m_queries[_id], GL_QUERY_RESULT, &result);
+			return result;
+		}
+
+		GLuint m_queries[64];
+	};
+#endif // BGFX_CONFIG_RENDERER_OPENGL
+
 } // namespace bgfx
 
 #endif // __RENDERER_GL_H__