Merge pull request #573 from andr3wmac/extended_profiling

Extended Remotery profiling.

examples/common/entry/entry.cpp

@@ -4,12 +4,6 @@
  */
 
 #include <bx/bx.h>
-#if BX_PLATFORM_WINDOWS
-// BK - Remotery needs WinSock, but on VS2015/Win10 build
-//      fails if WinSock2 is included after Windows.h?!
-#	include <WinSock2.h>
-#endif // BX_PLATFORM_WINDOWS
-
 #include <bgfx/bgfx.h>
 #include <bx/string.h>
 #include <bx/readerwriter.h>
@@ -25,7 +19,7 @@
 #include "input.h"
 
 #define RMT_ENABLED ENTRY_CONFIG_PROFILER
-#include <remotery/lib/Remotery.c>
+#include <remotery/lib/Remotery.h>
 
 extern "C" int _main_(int _argc, char** _argv);
 

scripts/genie.lua

@@ -94,6 +94,7 @@ if _OPTIONS["with-profiler"] then
 	defines {
 		"ENTRY_CONFIG_PROFILER=1",
 		"BGFX_CONFIG_PROFILER_REMOTERY=1",
+        "_WINSOCKAPI_"
 	}
 end
 

src/bgfx.cpp

@@ -5,6 +5,12 @@
 
 #include "bgfx_p.h"
 
+#if BGFX_CONFIG_PROFILER_REMOTERY
+#	define RMT_USE_D3D11 BGFX_CONFIG_RENDERER_DIRECT3D11
+#	define RMT_USE_OPENGL 0
+#	include <remotery/lib/Remotery.c>
+#endif
+
 namespace bgfx
 {
 #define BGFX_MAIN_THREAD_MAGIC UINT32_C(0x78666762)

src/bgfx_p.h

@@ -53,14 +53,24 @@
 #	if BGFX_CONFIG_PROFILER_MICROPROFILE
 #		include <microprofile.h>
 #		define BGFX_PROFILER_SCOPE(_group, _name, _color) MICROPROFILE_SCOPEI(#_group, #_name, _color)
+#		define BGFX_PROFILER_BEGIN(_group, _name, _color) BX_NOOP()
+#		define BGFX_PROFILER_BEGIN_DYNAMIC(_namestr) BX_NOOP()
+#		define BGFX_PROFILER_END() BX_NOOP()
 #		define BGFX_PROFILER_SET_CURRENT_THREAD_NAME(_name) BX_NOOP()
 #	elif BGFX_CONFIG_PROFILER_REMOTERY
 #		define RMT_ENABLED BGFX_CONFIG_PROFILER_REMOTERY
+#		define RMT_USE_D3D11 BGFX_CONFIG_RENDERER_DIRECT3D11
 #		include <remotery/lib/Remotery.h>
 #		define BGFX_PROFILER_SCOPE(_group, _name, _color) rmt_ScopedCPUSample(_group##_##_name)
+#		define BGFX_PROFILER_BEGIN(_group, _name, _color) rmt_BeginCPUSample(_group##_##_name)
+#		define BGFX_PROFILER_BEGIN_DYNAMIC(_namestr) rmt_BeginCPUSampleDynamic(_namestr)
+#		define BGFX_PROFILER_END() rmt_EndCPUSample()
 #		define BGFX_PROFILER_SET_CURRENT_THREAD_NAME(_name) rmt_SetCurrentThreadName(_name)
 #	else
 #		define BGFX_PROFILER_SCOPE(_group, _name, _color) BX_NOOP()
+#		define BGFX_PROFILER_BEGIN(_group, _name, _color) BX_NOOP()
+#		define BGFX_PROFILER_BEGIN_DYNAMIC(_namestr) BX_NOOP()
+#		define BGFX_PROFILER_END() BX_NOOP()
 #		define BGFX_PROFILER_SET_CURRENT_THREAD_NAME(_name) BX_NOOP()
 #	endif // BGFX_CONFIG_PROFILER_*
 #endif // BGFX_PROFILER_SCOPE

src/renderer_d3d11.cpp

@@ -13,9 +13,24 @@
 #	include <windows.ui.xaml.media.dxinterop.h>
 #endif // BX_PLATFORM_WINRT
 
+#if BGFX_CONFIG_PROFILER_REMOTERY
+#	define BGFX_GPU_PROFILER_BIND(_device, _context) rmt_BindD3D11(_device, _context)
+#	define BGFX_GPU_PROFILER_UNBIND() rmt_UnbindD3D11()
+#	define BGFX_GPU_PROFILER_BEGIN(_group, _name, _color) rmt_BeginD3D11Sample(_group##_##_name)
+#	define BGFX_GPU_PROFILER_BEGIN_DYNAMIC(_namestr) rmt_BeginD3D11SampleDynamic(_namestr)
+#	define BGFX_GPU_PROFILER_END() rmt_EndD3D11Sample()
+#else
+#	define BGFX_GPU_PROFILER_BIND(_device, _context) BX_NOOP()
+#	define BGFX_GPU_PROFILER_UNBIND() BX_NOOP()
+#	define BGFX_GPU_PROFILER_BEGIN(_group, _name, _color) BX_NOOP()
+#	define BGFX_GPU_PROFILER_BEGIN_DYNAMIC(_namestr) BX_NOOP()
+#	define BGFX_GPU_PROFILER_END() BX_NOOP()
+#endif
+
 namespace bgfx { namespace d3d11
 {
 	static wchar_t s_viewNameW[BGFX_CONFIG_MAX_VIEWS][BGFX_CONFIG_MAX_VIEW_NAME];
+	static char s_viewName[BGFX_CONFIG_MAX_VIEWS][BGFX_CONFIG_MAX_VIEW_NAME];
 
 	struct PrimInfo
 	{
@@ -1405,9 +1420,8 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 				// Init reserved part of view name.
 				for (uint32_t ii = 0; ii < BGFX_CONFIG_MAX_VIEWS; ++ii)
 				{
-					char name[BGFX_CONFIG_MAX_VIEW_NAME_RESERVED+1];
-					bx::snprintf(name, sizeof(name), "%3d   ", ii);
-					mbstowcs(s_viewNameW[ii], name, BGFX_CONFIG_MAX_VIEW_NAME_RESERVED);
+					bx::snprintf(s_viewName[ii], BGFX_CONFIG_MAX_VIEW_NAME_RESERVED + 1, "%3d   ", ii);
+					mbstowcs(s_viewNameW[ii], s_viewName[ii], BGFX_CONFIG_MAX_VIEW_NAME_RESERVED);
 				}
 
 				if (BX_ENABLED(BGFX_CONFIG_DEBUG)
@@ -1449,6 +1463,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 				postReset();
 			}
 
+			BGFX_GPU_PROFILER_BIND(m_device, m_deviceCtx);
 			return true;
 
 		error:
@@ -1501,6 +1516,8 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 
 		void shutdown()
 		{
+			BGFX_GPU_PROFILER_UNBIND();
+
 			preReset();
 			m_ovr.shutdown();
 
@@ -1864,6 +1881,11 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 					, BX_COUNTOF(s_viewNameW[0])-BGFX_CONFIG_MAX_VIEW_NAME_RESERVED
 					);
 			}
+
+			bx::strlcpy(&s_viewName[_id][BGFX_CONFIG_MAX_VIEW_NAME_RESERVED]
+				, _name
+				, BX_COUNTOF(s_viewName[0]) - BGFX_CONFIG_MAX_VIEW_NAME_RESERVED
+				);
 		}
 
 		void updateUniform(uint16_t _loc, const void* _data, uint32_t _size) BX_OVERRIDE
@@ -4622,6 +4644,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 	void RendererContextD3D11::submit(Frame* _render, ClearQuad& _clearQuad, TextVideoMemBlitter& _textVideoMemBlitter)
 	{
 		PIX_BEGINEVENT(D3DCOLOR_RGBA(0xff, 0x00, 0x00, 0xff), L"rendererSubmit");
+		BGFX_GPU_PROFILER_BEGIN_DYNAMIC("rendererSubmit");
 
 		ID3D11DeviceContext* deviceCtx = m_deviceCtx;
 
@@ -4754,6 +4777,13 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 					}
 
 					PIX_ENDEVENT();
+					if (item > 1)
+					{
+						BGFX_GPU_PROFILER_END();
+						BGFX_PROFILER_END();
+					}
+					BGFX_PROFILER_BEGIN_DYNAMIC(s_viewName[view]);
+					BGFX_GPU_PROFILER_BEGIN_DYNAMIC(s_viewName[view]);
 
 					viewState.m_rect = _render->m_rect[view];
 					if (viewRestart)
@@ -5498,10 +5528,14 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 				captureElapsed = -bx::getHPCounter();
 				capture();
 				captureElapsed += bx::getHPCounter();
+
+				BGFX_GPU_PROFILER_END();
+				BGFX_PROFILER_END();
 			}
 		}
 
 		PIX_ENDEVENT();
+		BGFX_GPU_PROFILER_END();
 
 		int64_t now = bx::getHPCounter();
 		elapsed += now;

src/renderer_d3d9.cpp

@@ -11,6 +11,7 @@
 namespace bgfx { namespace d3d9
 {
 	static wchar_t s_viewNameW[BGFX_CONFIG_MAX_VIEWS][BGFX_CONFIG_MAX_VIEW_NAME];
+	static char s_viewName[BGFX_CONFIG_MAX_VIEWS][BGFX_CONFIG_MAX_VIEW_NAME];
 
 	struct PrimInfo
 	{
@@ -732,9 +733,8 @@ namespace bgfx { namespace d3d9
 			// Init reserved part of view name.
 			for (uint32_t ii = 0; ii < BGFX_CONFIG_MAX_VIEWS; ++ii)
 			{
-				char name[BGFX_CONFIG_MAX_VIEW_NAME_RESERVED+1];
-				bx::snprintf(name, sizeof(name), "%3d   ", ii);
-				mbstowcs(s_viewNameW[ii], name, BGFX_CONFIG_MAX_VIEW_NAME_RESERVED);
+				bx::snprintf(s_viewName[ii], BGFX_CONFIG_MAX_VIEW_NAME_RESERVED + 1, "%3d   ", ii);
+				mbstowcs(s_viewNameW[ii], s_viewName[ii], BGFX_CONFIG_MAX_VIEW_NAME_RESERVED);
 			}
 
 			if (NULL != m_deviceEx)
@@ -1106,6 +1106,11 @@ namespace bgfx { namespace d3d9
 					, BX_COUNTOF(s_viewNameW[0])-BGFX_CONFIG_MAX_VIEW_NAME_RESERVED
 					);
 			}
+
+			bx::strlcpy(&s_viewName[_id][BGFX_CONFIG_MAX_VIEW_NAME_RESERVED]
+				, _name
+				, BX_COUNTOF(s_viewName[0]) - BGFX_CONFIG_MAX_VIEW_NAME_RESERVED
+				);
 		}
 
 		void updateUniform(uint16_t _loc, const void* _data, uint32_t _size) BX_OVERRIDE
@@ -3564,6 +3569,11 @@ namespace bgfx { namespace d3d9
 
 					PIX_ENDEVENT();
 					PIX_BEGINEVENT(D3DCOLOR_RGBA(0xff, 0x00, 0x00, 0xff), s_viewNameW[key.m_view]);
+					if (item > 0)
+					{
+						BGFX_PROFILER_END();
+					}
+					BGFX_PROFILER_BEGIN_DYNAMIC(s_viewName[key.m_view]);
 
 					view = key.m_view;
 					programIdx = invalidHandle;
@@ -4077,6 +4087,8 @@ namespace bgfx { namespace d3d9
 				captureElapsed = -bx::getHPCounter();
 				capture();
 				captureElapsed += bx::getHPCounter();
+
+				BGFX_PROFILER_END();
 			}
 		}
 

src/renderer_gl.cpp

@@ -10,6 +10,20 @@
 #	include <bx/timer.h>
 #	include <bx/uint32_t.h>
 
+#if BGFX_CONFIG_PROFILER_REMOTERY
+#	define BGFX_GPU_PROFILER_BIND() rmt_BindOpenGL()
+#	define BGFX_GPU_PROFILER_UNBIND() rmt_UnbindOpenGL()
+#	define BGFX_GPU_PROFILER_BEGIN(_group, _name, _color) rmt_BeginOpenGLSample(_group##_##_name)
+#	define BGFX_GPU_PROFILER_BEGIN_DYNAMIC(_namestr) rmt_BeginOpenGLSampleDynamic(_namestr)
+#	define BGFX_GPU_PROFILER_END() rmt_EndOpenGLSample()
+#else
+#	define BGFX_GPU_PROFILER_BIND(_device, _context) BX_NOOP()
+#	define BGFX_GPU_PROFILER_UNBIND() BX_NOOP()
+#	define BGFX_GPU_PROFILER_BEGIN(_group, _name, _color) BX_NOOP()
+#	define BGFX_GPU_PROFILER_BEGIN_DYNAMIC(_namestr) BX_NOOP()
+#	define BGFX_GPU_PROFILER_END() BX_NOOP()
+#endif
+
 namespace bgfx { namespace gl
 {
 	static char s_viewName[BGFX_CONFIG_MAX_VIEWS][BGFX_CONFIG_MAX_VIEW_NAME];
@@ -1958,10 +1972,14 @@ namespace bgfx { namespace gl
 			}
 
 			ovrPostReset();
+
+			BGFX_GPU_PROFILER_BIND();
 		}
 
 		void shutdown()
 		{
+			BGFX_GPU_PROFILER_UNBIND();
+
 			ovrPreReset();
 			m_ovr.shutdown();
 
@@ -2272,13 +2290,10 @@ namespace bgfx { namespace gl
 
 		void updateViewName(uint8_t _id, const char* _name) BX_OVERRIDE
 		{
-			if (BX_ENABLED(BGFX_CONFIG_DEBUG_PIX) )
-			{
-				bx::strlcpy(&s_viewName[_id][BGFX_CONFIG_MAX_VIEW_NAME_RESERVED]
-					, _name
-					, BX_COUNTOF(s_viewName[0])-BGFX_CONFIG_MAX_VIEW_NAME_RESERVED
-					);
-			}
+			bx::strlcpy(&s_viewName[_id][BGFX_CONFIG_MAX_VIEW_NAME_RESERVED]
+				, _name
+				, BX_COUNTOF(s_viewName[0])-BGFX_CONFIG_MAX_VIEW_NAME_RESERVED
+				);
 		}
 
 		void updateUniform(uint16_t _loc, const void* _data, uint32_t _size) BX_OVERRIDE
@@ -5217,6 +5232,8 @@ namespace bgfx { namespace gl
 
 	void RendererContextGL::submit(Frame* _render, ClearQuad& _clearQuad, TextVideoMemBlitter& _textVideoMemBlitter)
 	{
+		BGFX_GPU_PROFILER_BEGIN_DYNAMIC("rendererSubmit");
+
 		if (1 < m_numWindows
 		&&  m_vaoSupport)
 		{
@@ -5383,6 +5400,14 @@ namespace bgfx { namespace gl
 						eye = 0;
 					}
 
+					if (item > 1)
+					{
+						BGFX_GPU_PROFILER_END();
+						BGFX_PROFILER_END();
+					}
+					BGFX_PROFILER_BEGIN_DYNAMIC(s_viewName[view]);
+					BGFX_GPU_PROFILER_BEGIN_DYNAMIC(s_viewName[view]);
+
 					viewState.m_rect = _render->m_rect[view];
 					if (viewRestart)
 					{
@@ -6266,9 +6291,14 @@ namespace bgfx { namespace gl
 				captureElapsed = -bx::getHPCounter();
 				capture();
 				captureElapsed += bx::getHPCounter();
+
+				BGFX_GPU_PROFILER_END();
+				BGFX_PROFILER_END();
 			}
 		}
 
+		BGFX_GPU_PROFILER_END();
+
 		m_glctx.makeCurrent(NULL);
 		int64_t now = bx::getHPCounter();
 		elapsed += now;