Announcing radeontop, a tool for viewing the GPU usage

[jarodcanal]

Thanks for the info, I will try to get the 3D acceleration working, and I will try radeontop again.
Regards

[jarodcanal]

glxinfo | grep render
direct rendering: Yes
OpenGL renderer string: Gallium 0.4 on AMD RS880

:-) got it working
(I just had to install package xserver-xorg-video-radeon from http://ppa.launchpad.net/xorg-edgers/ppa/ubuntu, and reboot)
Now, radeontop is showing some nice counters when running glxgears.
Thanks for your help.
Regards

[kwahoo]

Interesting tool. Does "Graphics pipe<<100%" mean CPU bottleneck? I've done some tests (AII 240+HD6670):
Amnesia: TDD (1280x720): Graphics pipe=100%
OilRush (1280x720): Graphics pipe=100%
Enemy Territory: Quake Wars Demo (1280x720): Graphics pipe=50-60%
Red Eclipse (1280x720): Graphics pipe=10-20% (but very high fps)
HL2: Lost Coast (1280x720/Wine): Graphics pipe=50-60%
Serious Sam 3 BFE (1024x768/Wine/D3D*): Graphics pipe=100%

*OpenGL backend works only with Catalyst

[curaga]

First tarball release up, 0.5.4.1. No real changes from .4, just more convenient packaging for people who don't like git.

Does "Graphics pipe<<100%" mean CPU bottleneck?

Yes, but you can't know whether that CPU bottleneck is in mesa or in the app just by these counters. Use CPU profilers to find that out.
Note that vsync can drop the pipe usage, for proper numbers disable vsync.

[frief]

Add overlayed bars (using '+' char) with exponential moving averages.
Used %5.1 format specifier for percentages to avoid jumping columns.

I added these overlayed bars for the first 4 percentages only, so it
can be seen what this would be about.

I would have liked to attach text file (the patch) to this mail but
the forum software doesn't let me (only videos and images).
Enclosing as code instead.

Greetings,
Frieder

From 2e52e85eef0a15177854bff38381331d972ad7f6 Mon Sep 17 00:00:00 2001
From: Frieder Ferlemann <Frieder.Ferlemann@NOSPAMweb.de>
Date: Tue, 10 Jul 2012 22:43:37 +0200
Subject: [PATCH] showing averaged and instantaneous bars

Add overlayed bars (using '+' char) with exponential moving averages.
Used %5.1 format specifier for percentages to avoid jumping columns.
---
 ui.c |   74 +++++++++++++++++++++++++++++++++++++++++++++++++----------------
 1 files changed, 56 insertions(+), 18 deletions(-)

diff --git a/ui.c b/ui.c
index 48b82cb..6e79f02 100644
--- a/ui.c
+++ b/ui.c
@@ -66,6 +66,30 @@ static void percentage(const unsigned int y, const unsigned int w, const float p
 	attroff(A_REVERSE);
 }
 
+static void percentage2(const unsigned int y, const unsigned int w, const float p, const float p2) {
+
+	const unsigned int x = (w/2) + 2;
+	unsigned int len = w - x - 1;
+	unsigned int len2 = len;
+	unsigned int minlen = 0;
+
+	len  = len  * (p  / 100.0);
+	len2 = len2 * (p2 / 100.0);
+	minlen = (len <= len2) ? len : len2;
+
+	attron(A_REVERSE);
+	mvhline(y, x, '+', minlen);
+
+	if(len <= len2) {
+		attroff(A_REVERSE);
+		mvhline(y, x+minlen, '+', len2-minlen);
+	}
+	else {
+		mvhline(y, x+minlen, ' ', len-minlen);
+		attroff(A_REVERSE);
+	}
+}
+
 void present(const unsigned int ticks, const char card[], const unsigned int color) {
 
 	// This does not need to be atomic. A delay here is acceptable.
@@ -122,39 +146,53 @@ void present(const unsigned int ticks, const char card[], const unsigned int col
 		float cr = 100.0 * (float) results->cr / ticks;
 		float cb = 100.0 * (float) results->cb / ticks;
 
+		// determins time constant for moving averages
+		const float alpha = 0.3;
+
+		static float ee_avg = 0;
+		static float vgt_avg = 0;
+		static float gui_avg = 0;
+		static float ta_avg = 0;
+
+		// calculate exponential moving averages
+		ee_avg  = ee_avg  * (1 - alpha) + ee  * alpha;
+		vgt_avg = vgt_avg * (1 - alpha) + vgt * alpha;
+		gui_avg = gui_avg * (1 - alpha) + gui * alpha;
+		ta_avg  = ta_avg  * (1 - alpha) + ta  * alpha;
+
 		mvhline(3, 0, ACS_HLINE, w);
 		mvvline(1, (w/2) + 1, ACS_VLINE, h);
 		mvaddch(3, (w/2) + 1, ACS_PLUS);
 
 		if (color) attron(COLOR_PAIR(1));
-		percentage(2, w, gui);
-		printright(2, hw, _("Graphics pipe %.2f%%"), gui);
+		percentage2(2, w, gui, gui_avg);
+		printright(2, hw, _("Graphics pipe %5.1f%% %5.1f%%"), gui, gui_avg);
 		if (color) attroff(COLOR_PAIR(1));
 
 		unsigned int start = 4;
 
-		percentage(start, w, ee);
-		printright(start++, hw, _("Event Engine %.2f%%"), ee);
+		percentage2(start, w, ee, ee_avg);
+		printright(start++, hw, _("Event Engine %5.1f%% %5.1f%%"), ee, ee_avg);
 
 		// Enough height?
 		if (h > bigh) start++;
 
 		if (color) attron(COLOR_PAIR(2));
-		percentage(start, w, vgt);
-		printright(start++, hw, _("Vertex Grouper + Tesselator %.2f%%"), vgt);
+		percentage2(start, w, vgt, vgt_avg);
+		printright(start++, hw, _("Vertex Grouper + Tesselator %5.1f%% %5.1f%%"), vgt, vgt_avg);
 		if (color) attroff(COLOR_PAIR(2));
 
 		// Enough height?
 		if (h > bigh) start++;
 
 		if (color) attron(COLOR_PAIR(3));
-		percentage(start, w, ta);
-		printright(start++, hw, _("Texture Addresser %.2f%%"), ta);
+		percentage2(start, w, ta, ta_avg);
+		printright(start++, hw, _("Texture Addresser %5.1f%% %5.1f%%"), ta, ta_avg);
 
 		// This is only present on R600
 		if (bits.tc) {
 			percentage(start, w, tc);
-			printright(start++, hw, _("Texture Cache %.2f%%"), tc);
+			printright(start++, hw, _("Texture Cache %5.1f%%"), tc);
 		}
 		if (color) attroff(COLOR_PAIR(3));
 
@@ -163,18 +201,18 @@ void present(const unsigned int ticks, const char card[], const unsigned int col
 
 		if (color) attron(COLOR_PAIR(4));
 		percentage(start, w, sx);
-		printright(start++, hw, _("Shader Export %.2f%%"), sx);
+		printright(start++, hw, _("Shader Export %5.1f%%"), sx);
 
 		percentage(start, w, sh);
-		printright(start++, hw, _("Sequencer Instruction Cache %.2f%%"), sh);
+		printright(start++, hw, _("Sequencer Instruction Cache %5.1f%%"), sh);
 
 		percentage(start, w, spi);
-		printright(start++, hw, _("Shader Interpolator %.2f%%"), spi);
+		printright(start++, hw, _("Shader Interpolator %5.1f%%"), spi);
 
 		// only on R600
 		if (bits.smx) {
 			percentage(start, w, smx);
-			printright(start++, hw, _("Shader Memory Exchange %.2f%%"), smx);
+			printright(start++, hw, _("Shader Memory Exchange %5.1f%%"), smx);
 		}
 		if (color) attroff(COLOR_PAIR(4));
 
@@ -182,25 +220,25 @@ void present(const unsigned int ticks, const char card[], const unsigned int col
 		if (h > bigh) start++;
 
 		percentage(start, w, sc);
-		printright(start++, hw, _("Scan Converter %.2f%%"), sc);
+		printright(start++, hw, _("Scan Converter %5.1f%%"), sc);
 
 		percentage(start, w, pa);
-		printright(start++, hw, _("Primitive Assembly %.2f%%"), pa);
+		printright(start++, hw, _("Primitive Assembly %5.1f%%"), pa);
 
 		// Enough height?
 		if (h > bigh) start++;
 
 		if (color) attron(COLOR_PAIR(5));
 		percentage(start, w, db);
-		printright(start++, hw, _("Depth Block %.2f%%"), db);
+		printright(start++, hw, _("Depth Block %5.1f%%"), db);
 
 		percentage(start, w, cb);
-		printright(start++, hw, _("Color Block %.2f%%"), cb);
+		printright(start++, hw, _("Color Block %5.1f%%"), cb);
 
 		// Only present on R600
 		if (bits.cr) {
 			percentage(start, w, cr);
-			printright(start++, hw, _("Clip Rectangle %.2f%%"), cr);
+			printright(start++, hw, _("Clip Rectangle %5.1f%%"), cr);
 		}
 		if (color) attroff(COLOR_PAIR(5));
 
-- 
1.7.7

[curaga]

Thanks. Please use the github interface in the future, forums tend to mangle patches.

Field width is changed, the lerp averages added on todo.

[curaga]

0.6 released with dumping support.

[TAXI]

Great tool!
Even if it just shows how much the CPU limits the driver. Example: I can run 2 instances of minecraft with max graphic settings and none of the values in radeontop goes near 100% while the GPU is in the low PM profile. Of course the FPS is bad and the CPU is at 100% (yes, I know, minecraft isn't the best here cause it eats up much CPU by default, but anyway...).

Michael should include radeontop measuring in his benchmarks.