diff --git a/examples/common/aviwriter.h b/examples/common/aviwriter.h
index 6095b3c1..199e1800 100644
--- a/examples/common/aviwriter.h
+++ b/examples/common/aviwriter.h
@@ -91,8 +91,8 @@ struct AviWriter
 		bx::write(&m_writer, UINT16_C(0) );      // wPriority
 		bx::write(&m_writer, UINT16_C(0) );      // wLanguage
 		bx::write(&m_writer, UINT32_C(0) );      // dwInitialFrames
-		bx::write(&m_writer, UINT32_C(1000) );   // dwScale
-		bx::write(&m_writer, 1000*_fps);         // dwRate
+		bx::write(&m_writer, UINT32_C(1) );      // dwScale
+		bx::write(&m_writer, _fps);              // dwRate
 		bx::write(&m_writer, UINT32_C(0) );      // dwStart
 
 		m_lengthOffset = m_writer.seek();
@@ -137,19 +137,24 @@ struct AviWriter
 		{
 			int64_t pos = m_writer.seek();
 			m_writer.seek(m_moviListOffset, bx::Whence::Begin);
-			bx::write(&m_writer, uint32_t(pos-m_moviListOffset+4) );
+			bx::write(&m_writer, uint32_t(pos-m_moviListOffset-4) );
 			m_writer.seek(pos, bx::Whence::Begin);
 
 			bx::write(&m_writer, BX_MAKEFOURCC('i', 'd', 'x', '1') );
-			bx::write(&m_writer, UINT32_C(16) );
-			bx::write(&m_writer, BX_MAKEFOURCC('0', '0', 'd', 'b') );
-			bx::write(&m_writer, UINT32_C(16) );
-			bx::write(&m_writer, UINT32_C(4) );
-			bx::write(&m_writer, m_frameSize);
+			bx::write(&m_writer, m_numFrames*16);
+
+			for (uint32_t ii = 0, offset = 4; ii < m_numFrames; ++ii)
+			{
+				bx::write(&m_writer, BX_MAKEFOURCC('0', '0', 'd', 'b') );
+				bx::write(&m_writer, UINT32_C(16) );
+				bx::write(&m_writer, offset);
+				bx::write(&m_writer, m_frameSize);
+				offset += m_frameSize + 8;
+			}
 
 			pos = m_writer.seek();
 			m_writer.seek(m_riffSizeOffset, bx::Whence::Begin);
-			bx::write(&m_writer, uint32_t(pos-m_riffSizeOffset+4) );
+			bx::write(&m_writer, uint32_t(pos-m_riffSizeOffset-4) );
 
 			m_writer.seek(m_totalFramesOffset, bx::Whence::Begin);
 			bx::write(&m_writer, m_numFrames);
diff --git a/src/renderer_d3d11.cpp b/src/renderer_d3d11.cpp
index 576430da..d21e2c1b 100644
--- a/src/renderer_d3d11.cpp
+++ b/src/renderer_d3d11.cpp
@@ -1924,6 +1924,9 @@ namespace bgfx
 
 		s_renderCtx.updateResolution(m_render->m_resolution);
 
+		int64_t elapsed = -bx::getHPCounter();
+		int64_t captureElapsed = 0;
+
 		if (0 < m_render->m_iboffset)
 		{
 			TransientIndexBuffer* ib = m_render->m_transientIb;
@@ -1966,8 +1969,6 @@ namespace bgfx
 		uint32_t statsNumInstances = 0;
 		uint32_t statsNumPrimsRendered = 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)
@@ -2380,6 +2381,13 @@ namespace bgfx
 					statsNumPrimsRendered += numPrimsRendered;
 				}
 			}
+
+			if (0 < m_render->m_num)
+			{
+				captureElapsed = -bx::getHPCounter();
+				s_renderCtx.capture();
+				captureElapsed += bx::getHPCounter();
+			}
 		}
 
 		int64_t now = bx::getHPCounter();
@@ -2407,7 +2415,6 @@ 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;
@@ -2418,6 +2425,8 @@ namespace bgfx
 					, double(max)*toMs
 					, freq/frameTime
 					);
+
+				double elapsedCpuMs = double(elapsed)*toMs;
 				tvm.printf(10, pos++, 0x8e, " Draw calls: %4d / CPU %3.4f [ms]"
 					, m_render->m_num
 					, elapsedCpuMs
@@ -2427,6 +2436,9 @@ namespace bgfx
 					, statsNumInstances
 					, statsNumPrimsSubmitted
 					);
+
+				double captureMs = double(captureElapsed)*toMs;
+				tvm.printf(10, pos++, 0x8e, "    Capture: %3.4f [ms]", captureMs);
 				tvm.printf(10, pos++, 0x8e, "    Indices: %7d", statsNumIndices);
 				tvm.printf(10, pos++, 0x8e, "   DVB size: %7d", m_render->m_vboffset);
 				tvm.printf(10, pos++, 0x8e, "   DIB size: %7d", m_render->m_iboffset);
diff --git a/src/renderer_d3d9.cpp b/src/renderer_d3d9.cpp
index 4ad3ba11..adce237c 100644
--- a/src/renderer_d3d9.cpp
+++ b/src/renderer_d3d9.cpp
@@ -1997,6 +1997,9 @@ namespace bgfx
 
 		s_renderCtx.updateResolution(m_render->m_resolution);
 
+		int64_t elapsed = -bx::getHPCounter();
+		int64_t captureElapsed = 0;
+
 		device->BeginScene();
 
 		if (0 < m_render->m_iboffset)
@@ -2038,8 +2041,6 @@ namespace bgfx
 		uint32_t statsNumInstances = 0;
 		uint32_t statsNumPrimsRendered = 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)
@@ -2586,15 +2587,18 @@ namespace bgfx
 			}
 
 			PIX_ENDEVENT();
+
+			if (0 < m_render->m_num)
+			{
+				captureElapsed = -bx::getHPCounter();
+				s_renderCtx.capture();
+				captureElapsed += bx::getHPCounter();
+			}
 		}
 
 		int64_t now = bx::getHPCounter();
 		elapsed += now;
 
-		int64_t captureElapsed = -bx::getHPCounter();
-		s_renderCtx.capture();
-		captureElapsed += bx::getHPCounter();
-
 		static int64_t last = now;
 		int64_t frameTime = now - last;
 		last = now;
diff --git a/src/renderer_gl.cpp b/src/renderer_gl.cpp
index 56ab4ffb..2ba3912d 100644
--- a/src/renderer_gl.cpp
+++ b/src/renderer_gl.cpp
@@ -2352,6 +2352,8 @@ namespace bgfx
 		s_renderCtx.updateResolution(m_render->m_resolution);
 
 		int64_t elapsed = -bx::getHPCounter();
+		int64_t captureElapsed = 0;
+
 #if BGFX_CONFIG_RENDERER_OPENGL
 		if (m_render->m_debug & (BGFX_DEBUG_IFH|BGFX_DEBUG_STATS) )
 		{
@@ -2972,15 +2974,18 @@ namespace bgfx
 					}
 				}
 			}
+
+			if (0 < m_render->m_num)
+			{
+				captureElapsed = -bx::getHPCounter();
+				s_renderCtx.capture();
+				captureElapsed += bx::getHPCounter();
+			}
 		}
 
 		int64_t now = bx::getHPCounter();
 		elapsed += now;
 
-		int64_t captureElapsed = -bx::getHPCounter();
-		s_renderCtx.capture();
-		captureElapsed += bx::getHPCounter();
-
 		static int64_t last = now;
 		int64_t frameTime = now - last;
 		last = now;