diff --git a/src/d_main.c b/src/d_main.c
index 28e134d65..2e1badc69 100644
--- a/src/d_main.c
+++ b/src/d_main.c
@@ -607,17 +607,25 @@ static void D_Display(void)
 			V_DrawThinString(30, 10, V_MONOSPACE | V_YELLOWMAP, s);
 			snprintf(s, sizeof s - 1, "rtot %d", rs_rendercalltime / divisor);
 			V_DrawThinString(30, 20, V_MONOSPACE | V_YELLOWMAP, s);
-			if (rendermode == render_opengl)// dont show unimplemented stats
+			snprintf(s, sizeof s - 1, "bsp  %d", rs_bsptime / divisor);
+			V_DrawThinString(30, 30, V_MONOSPACE | V_YELLOWMAP, s);
+			snprintf(s, sizeof s - 1, "nbsp %d", rs_numbspcalls);
+			V_DrawThinString(80, 10, V_MONOSPACE | V_BLUEMAP, s);
+			snprintf(s, sizeof s - 1, "nspr %d", rs_numsprites);
+			V_DrawThinString(80, 20, V_MONOSPACE | V_BLUEMAP, s);
+			snprintf(s, sizeof s - 1, "nnod %d", rs_numdrawnodes);
+			V_DrawThinString(80, 30, V_MONOSPACE | V_BLUEMAP, s);
+			snprintf(s, sizeof s - 1, "npob %d", rs_numpolyobjects);
+			V_DrawThinString(80, 40, V_MONOSPACE | V_BLUEMAP, s);
+			if (rendermode == render_opengl) // OpenGL specific stats
 			{
-				snprintf(s, sizeof s - 1, "bsp  %d", rs_bsptime / divisor);
-				V_DrawThinString(30, 30, V_MONOSPACE | V_YELLOWMAP, s);
-				snprintf(s, sizeof s - 1, "nsrt %d", rs_nodesorttime / divisor);
+				snprintf(s, sizeof s - 1, "nsrt %d", rs_hw_nodesorttime / divisor);
 				V_DrawThinString(30, 40, V_MONOSPACE | V_YELLOWMAP, s);
-				snprintf(s, sizeof s - 1, "ndrw %d", rs_nodedrawtime / divisor);
+				snprintf(s, sizeof s - 1, "ndrw %d", rs_hw_nodedrawtime / divisor);
 				V_DrawThinString(30, 50, V_MONOSPACE | V_YELLOWMAP, s);
-				snprintf(s, sizeof s - 1, "ssrt %d", rs_spritesorttime / divisor);
+				snprintf(s, sizeof s - 1, "ssrt %d", rs_hw_spritesorttime / divisor);
 				V_DrawThinString(30, 60, V_MONOSPACE | V_YELLOWMAP, s);
-				snprintf(s, sizeof s - 1, "sdrw %d", rs_spritedrawtime / divisor);
+				snprintf(s, sizeof s - 1, "sdrw %d", rs_hw_spritedrawtime / divisor);
 				V_DrawThinString(30, 70, V_MONOSPACE | V_YELLOWMAP, s);
 				/*snprintf(s, sizeof s - 1, "post %d", rs_posttime / divisor);
 				V_DrawThinString(30, 80, V_MONOSPACE | V_YELLOWMAP, s);*/
@@ -626,14 +634,6 @@ static void D_Display(void)
 				/*snprintf(s, sizeof s - 1, "test %d", rs_test / divisor);
 				V_DrawThinString(30, 100, V_MONOSPACE | V_YELLOWMAP, s);*/
 
-				snprintf(s, sizeof s - 1, "nbsp %d", rs_numbspcalls);
-				V_DrawThinString(80, 10, V_MONOSPACE | V_BLUEMAP, s);
-				snprintf(s, sizeof s - 1, "nnod %d", rs_numdrawnodes);
-				V_DrawThinString(80, 20, V_MONOSPACE | V_BLUEMAP, s);
-				snprintf(s, sizeof s - 1, "nspr %d", rs_numsprites);
-				V_DrawThinString(80, 30, V_MONOSPACE | V_BLUEMAP, s);
-				snprintf(s, sizeof s - 1, "npob %d", rs_numpolyobjects);
-				V_DrawThinString(80, 40, V_MONOSPACE | V_BLUEMAP, s);
 /*
 				if (cv_enable_batching.value)
 				{
@@ -658,10 +658,16 @@ static void D_Display(void)
 					V_DrawThinString(185, 30, V_MONOSPACE | V_PURPLEMAP, s);
 				}*/
 			}
-			else
+			else // software specific stats
 			{
+				snprintf(s, sizeof s - 1, "prtl %d", rs_sw_portaltime / divisor);
+				V_DrawThinString(30, 40, V_MONOSPACE | V_YELLOWMAP, s);
+				snprintf(s, sizeof s - 1, "plns %d", rs_sw_planetime / divisor);
+				V_DrawThinString(30, 50, V_MONOSPACE | V_YELLOWMAP, s);
+				snprintf(s, sizeof s - 1, "mskd %d", rs_sw_maskedtime / divisor);
+				V_DrawThinString(30, 60, V_MONOSPACE | V_YELLOWMAP, s);
 				snprintf(s, sizeof s - 1, "fin  %d", rs_swaptime / divisor);
-				V_DrawThinString(30, 30, V_MONOSPACE | V_YELLOWMAP, s);
+				V_DrawThinString(30, 70, V_MONOSPACE | V_YELLOWMAP, s);
 			}
 		}
 
diff --git a/src/hardware/hw_main.c b/src/hardware/hw_main.c
index 79cc83bba..bd2c5c9f4 100644
--- a/src/hardware/hw_main.c
+++ b/src/hardware/hw_main.c
@@ -147,17 +147,10 @@ static angle_t gr_aimingangle;
 static void HWR_SetTransformAiming(FTransform *trans, player_t *player, boolean skybox);
 
 // Render stats
-int rs_bsptime = 0;
-int rs_nodetime = 0;
-int rs_nodesorttime = 0;
-int rs_nodedrawtime = 0;
-int rs_spritesorttime = 0;
-int rs_spritedrawtime = 0;
-
-int rs_numdrawnodes = 0;
-int rs_numbspcalls = 0;
-int rs_numsprites = 0;
-int rs_numpolyobjects = 0;
+int rs_hw_nodesorttime = 0;
+int rs_hw_nodedrawtime = 0;
+int rs_hw_spritesorttime = 0;
+int rs_hw_spritedrawtime = 0;
 
 //int rs_posttime = 0;
 
@@ -4794,7 +4787,7 @@ static void HWR_CreateDrawNodes(void)
 	// If true, swap the draw order.
 	boolean shift = false;
 	
-	rs_nodesorttime = I_GetTimeMicros();
+	rs_hw_nodesorttime = I_GetTimeMicros();
 
 	for (i = 0; i < numplanes; i++, p++)
 	{
@@ -4917,9 +4910,9 @@ static void HWR_CreateDrawNodes(void)
 		} //i++
 	} // loop++
 
-	rs_nodesorttime = I_GetTimeMicros() - rs_nodesorttime;
+	rs_hw_nodesorttime = I_GetTimeMicros() - rs_hw_nodesorttime;
 	
-	rs_nodedrawtime = I_GetTimeMicros();
+	rs_hw_nodedrawtime = I_GetTimeMicros();
 
 	// Okay! Let's draw it all! Woo!
 	HWD.pfnSetTransform(&atransform);
@@ -4956,7 +4949,7 @@ static void HWR_CreateDrawNodes(void)
 		}
 	}
 	
-	rs_nodedrawtime = I_GetTimeMicros() - rs_nodedrawtime;
+	rs_hw_nodedrawtime = I_GetTimeMicros() - rs_hw_nodedrawtime;
 
 	numwalls = 0;
 	numplanes = 0;
@@ -6083,12 +6076,12 @@ void HWR_RenderPlayerView(INT32 viewnumber, player_t *player)
 
 	// Draw MD2 and sprites
 	rs_numsprites = gr_visspritecount;
-	rs_spritesorttime = I_GetTimeMicros();
+	rs_hw_spritesorttime = I_GetTimeMicros();
 	HWR_SortVisSprites();
-	rs_spritesorttime = I_GetTimeMicros() - rs_spritesorttime;
-	rs_spritedrawtime = I_GetTimeMicros();
+	rs_hw_spritesorttime = I_GetTimeMicros() - rs_hw_spritesorttime;
+	rs_hw_spritedrawtime = I_GetTimeMicros();
 	HWR_DrawSprites();
-	rs_spritedrawtime = I_GetTimeMicros() - rs_spritedrawtime;
+	rs_hw_spritedrawtime = I_GetTimeMicros() - rs_hw_spritedrawtime;
 
 #ifdef NEWCORONAS
 	//Hurdler: they must be drawn before translucent planes, what about gl fog?
@@ -6096,8 +6089,8 @@ void HWR_RenderPlayerView(INT32 viewnumber, player_t *player)
 #endif
 
 	rs_numdrawnodes = 0;
-	rs_nodesorttime = 0;
-	rs_nodedrawtime = 0;
+	rs_hw_nodesorttime = 0;
+	rs_hw_nodedrawtime = 0;
 	if (numplanes || numpolyplanes || numwalls) //Hurdler: render 3D water and transparent walls after everything
 	{
 		HWR_CreateDrawNodes();
diff --git a/src/hardware/hw_main.h b/src/hardware/hw_main.h
index f89809cc7..631593d6f 100644
--- a/src/hardware/hw_main.h
+++ b/src/hardware/hw_main.h
@@ -105,18 +105,12 @@ extern fixed_t *hwbbox;
 extern FTransform atransform;
 
 
-extern int rs_bsptime;
-extern int rs_nodetime;
-extern int rs_nodesorttime;
-extern int rs_nodedrawtime;
-extern int rs_spritesorttime;
-extern int rs_spritedrawtime;
+// Render stats
+extern int rs_hw_nodesorttime;
+extern int rs_hw_nodedrawtime;
+extern int rs_hw_spritesorttime;
+extern int rs_hw_spritedrawtime;
 
 //extern int rs_posttime;
 
-extern int rs_numdrawnodes;
-extern int rs_numbspcalls;
-extern int rs_numsprites;
-extern int rs_numpolyobjects;
-
 #endif
diff --git a/src/r_bsp.c b/src/r_bsp.c
index c0011f4b9..355d55bc6 100644
--- a/src/r_bsp.c
+++ b/src/r_bsp.c
@@ -812,6 +812,9 @@ static void R_AddPolyObjects(subsector_t *sub)
 		po = (polyobj_t *)(po->link.next);
 	}
 
+	// for render stats
+	rs_numpolyobjects += numpolys;
+
 	// sort polyobjects
 	R_SortPolyObjects(sub);
 
@@ -1363,6 +1366,9 @@ void R_RenderBSPNode(INT32 bspnum)
 {
 	node_t *bsp;
 	INT32 side;
+
+	rs_numbspcalls++;
+
 	while (!(bspnum & NF_SUBSECTOR))  // Found a subsector?
 	{
 		bsp = &nodes[bspnum];
diff --git a/src/r_main.c b/src/r_main.c
index 17eae9495..4d1be4b14 100644
--- a/src/r_main.c
+++ b/src/r_main.c
@@ -34,6 +34,7 @@
 #include "m_random.h" // quake camera shake
 #include "r_portal.h"
 #include "r_main.h"
+#include "i_system.h" // I_GetTimeMicros
 
 #ifdef HWRENDER
 #include "hardware/hw_main.h"
@@ -103,6 +104,17 @@ int rs_prevframetime = 0;
 int rs_rendercalltime = 0;
 int rs_swaptime = 0;
 
+int rs_bsptime = 0;
+
+int rs_sw_portaltime = 0;
+int rs_sw_planetime = 0;
+int rs_sw_maskedtime = 0;
+
+int rs_numbspcalls = 0;
+int rs_numsprites = 0;
+int rs_numdrawnodes = 0;
+int rs_numpolyobjects = 0;
+
 static CV_PossibleValue_t drawdist_cons_t[] = {
 	{256, "256"},	{512, "512"},	{768, "768"},
 	{1024, "1024"},	{1536, "1536"},	{2048, "2048"},
@@ -1452,7 +1464,11 @@ void R_RenderPlayerView(player_t *player)
 	mytotal = 0;
 	ProfZeroTimer();
 #endif
+	rs_numbspcalls = rs_numpolyobjects = rs_numdrawnodes = 0;
+	rs_bsptime = I_GetTimeMicros();
 	R_RenderBSPNode((INT32)numnodes - 1);
+	rs_bsptime = I_GetTimeMicros() - rs_bsptime;
+	rs_numsprites = visspritecount;
 #ifdef TIMING
 	RDMSR(0x10, &mycount);
 	mytotal += mycount; // 64bit add
@@ -1470,6 +1486,7 @@ void R_RenderPlayerView(player_t *player)
 		Portal_AddSkyboxPortals();
 
 	// Portal rendering. Hijacks the BSP traversal.
+	rs_sw_portaltime = I_GetTimeMicros();
 	if (portal_base)
 	{
 		portal_t *portal;
@@ -1509,15 +1526,20 @@ void R_RenderPlayerView(player_t *player)
 			Portal_Remove(portal);
 		}
 	}
+	rs_sw_portaltime = I_GetTimeMicros() - rs_sw_portaltime;
 
+	rs_sw_planetime = I_GetTimeMicros();
 	R_DrawPlanes();
 #ifdef FLOORSPLATS
 	R_DrawVisibleFloorSplats();
 #endif
+	rs_sw_planetime = I_GetTimeMicros() - rs_sw_planetime;
 
 	// draw mid texture and sprite
 	// And now 3D floors/sides!
+	rs_sw_maskedtime = I_GetTimeMicros();
 	R_DrawMasked(masks, nummasks);
+	rs_sw_maskedtime = I_GetTimeMicros() - rs_sw_maskedtime;
 
 	free(masks);
 }
diff --git a/src/r_main.h b/src/r_main.h
index 72d340bd9..99a25d86e 100644
--- a/src/r_main.h
+++ b/src/r_main.h
@@ -82,6 +82,17 @@ extern int rs_prevframetime;// time when previous frame was rendered
 extern int rs_rendercalltime;
 extern int rs_swaptime;
 
+extern int rs_bsptime;
+
+extern int rs_sw_portaltime;
+extern int rs_sw_planetime;
+extern int rs_sw_maskedtime;
+
+extern int rs_numbspcalls;
+extern int rs_numsprites;
+extern int rs_numdrawnodes;
+extern int rs_numpolyobjects;
+
 //
 // REFRESH - the actual rendering functions.
 //
diff --git a/src/r_things.c b/src/r_things.c
index d2f3b4902..02c347929 100644
--- a/src/r_things.c
+++ b/src/r_things.c
@@ -2552,6 +2552,8 @@ static drawnode_t *R_CreateDrawNode(drawnode_t *link)
 	node->thickseg = NULL;
 	node->ffloor = NULL;
 	node->sprite = NULL;
+
+	rs_numdrawnodes++;
 	return node;
 }