diff --git a/src/renderer_d3d11.cpp b/src/renderer_d3d11.cpp
index 6dafafbf..a79a5145 100644
--- a/src/renderer_d3d11.cpp
+++ b/src/renderer_d3d11.cpp
@@ -3146,6 +3146,7 @@ namespace bgfx { namespace d3d11
 		uint32_t statsNumPrimsRendered[BX_COUNTOF(s_primInfo)] = {};
 		uint32_t statsNumInstances[BX_COUNTOF(s_primInfo)] = {};
 		uint32_t statsNumIndices = 0;
+		uint32_t statsKeyType[2] = {};
 
 		if (0 == (_render->m_debug&BGFX_DEBUG_IFH) )
 		{
@@ -3158,6 +3159,8 @@ namespace bgfx { namespace d3d11
 			for (int32_t item = 0, restartItem = numItems; item < numItems || restartItem < numItems;)
 			{
 				const bool isCompute = key.decode(_render->m_sortKeys[item], _render->m_viewRemap);
+				statsKeyType[isCompute]++;
+
 				const bool viewChanged = 0
 					|| key.m_view != view
 					|| item == numItems
@@ -3801,8 +3804,10 @@ namespace bgfx { namespace d3d11
 					);
 
 				double elapsedCpuMs = double(elapsed)*toMs;
-				tvm.printf(10, pos++, 0x8e, "  Draw calls: %4d / CPU %3.4f [ms]"
+				tvm.printf(10, pos++, 0x8e, "   Submitted: %4d (draw %4d, compute %4d) / CPU %3.4f [ms]"
 					, _render->m_num
+					, statsKeyType[0]
+					, statsKeyType[1]
 					, elapsedCpuMs
 					);
 				for (uint32_t ii = 0; ii < BX_COUNTOF(s_primName); ++ii)
diff --git a/src/renderer_d3d9.cpp b/src/renderer_d3d9.cpp
index 83b08b72..c034c071 100644
--- a/src/renderer_d3d9.cpp
+++ b/src/renderer_d3d9.cpp
@@ -2888,6 +2888,7 @@ namespace bgfx { namespace d3d9
 		uint32_t statsNumPrimsRendered[BX_COUNTOF(s_primInfo)] = {};
 		uint32_t statsNumInstances[BX_COUNTOF(s_primInfo)] = {};
 		uint32_t statsNumIndices = 0;
+		uint32_t statsKeyType[2] = {};
 
 		invalidateSamplerState();
 
@@ -2896,6 +2897,7 @@ namespace bgfx { namespace d3d9
 			for (uint32_t item = 0, numItems = _render->m_num; item < numItems; ++item)
 			{
 				const bool isCompute = key.decode(_render->m_sortKeys[item], _render->m_viewRemap);
+				statsKeyType[isCompute]++;
 
 				if (isCompute)
 				{
@@ -3426,8 +3428,10 @@ namespace bgfx { namespace d3d9
 					);
 
 				double elapsedCpuMs = double(elapsed)*toMs;
-				tvm.printf(10, pos++, 0x8e, "  Draw calls: %4d / CPU %3.4f [ms]"
+				tvm.printf(10, pos++, 0x8e, "   Submitted: %4d (draw %4d, compute %4d) / CPU %3.4f [ms]"
 					, _render->m_num
+					, statsKeyType[0]
+					, statsKeyType[1]
 					, elapsedCpuMs
 					);
 				for (uint32_t ii = 0; ii < BX_COUNTOF(s_primName); ++ii)
diff --git a/src/renderer_gl.cpp b/src/renderer_gl.cpp
index e1ed1543..3dec6c41 100644
--- a/src/renderer_gl.cpp
+++ b/src/renderer_gl.cpp
@@ -4421,6 +4421,7 @@ namespace bgfx { namespace gl
 		uint32_t statsNumPrimsRendered[BX_COUNTOF(s_primInfo)] = {};
 		uint32_t statsNumInstances[BX_COUNTOF(s_primInfo)] = {};
 		uint32_t statsNumIndices = 0;
+		uint32_t statsKeyType[2] = {};
 
 		if (0 == (_render->m_debug&BGFX_DEBUG_IFH) )
 		{
@@ -4435,6 +4436,8 @@ namespace bgfx { namespace gl
 			for (int32_t item = 0, restartItem = numItems; item < numItems || restartItem < numItems;)
 			{
 				const bool isCompute   = key.decode(_render->m_sortKeys[item], _render->m_viewRemap);
+				statsKeyType[isCompute]++;
+
 				const bool viewChanged = 0
 					|| key.m_view != view
 					|| item == numItems
@@ -5236,8 +5239,10 @@ namespace bgfx { namespace gl
 					);
 
 				double elapsedCpuMs = double(elapsed)*toMs;
-				tvm.printf(10, pos++, 0x8e, " Draw calls: %4d / CPU %3.4f [ms] %c GPU %3.4f [ms]"
+				tvm.printf(10, pos++, 0x8e, "  Submitted: %4d (draw %4d, compute %4d) / CPU %3.4f [ms] %c GPU %3.4f [ms]"
 					, _render->m_num
+					, statsKeyType[0]
+					, statsKeyType[1]
 					, elapsedCpuMs
 					, elapsedCpuMs > elapsedGpuMs ? '>' : '<'
 					, elapsedGpuMs