[teeworlds] Re: Investigating high CPU usage
- From: Magnus Auvinen <magnus.auvinen@xxxxxxxxx>
- To: teeworlds@xxxxxxxxxxxxx
- Date: Mon, 2 Mar 2009 19:16:43 +0100
Hi,
Sorry for the very delayed response, long story, been busy, psu busted.
First, glFinish and glFlush are two separate functions. glFlush sends
everything to the gpu and continues. glFinish performs a flush and
then waits for the gpu to finish everything. If i remember correctly,
the glFinish call was added because some people experienced latency
issues because the cpu was several frames infront of the gpu. A good
graphics driver shouldn't let this happen. I've disabled the finish
call in trunk and we'll see what happens.
As for the micro optimizations of the rotation. You can see the
changelist here http://www.teeworlds.com/trac/teeworlds/changeset/1948
. I've did the rotate4 but with loop instead and added a check to see
if we need to rotate at all. Most of the calls to quads_draw comes
from the tilemap rendering and those doesn't have any rotation. You
could do a macro optimization and split up the tilemaps into chunks of
display lists that would be called instead of rendering each quad by
it self. This would reduce the cpu overhead quite much. I tend to do
this optimization when I get time for it but it's not very important
right now but if you are up for it, go ahead.
Cheers!
Magnus 'matricks' Auvinen
On Sat, Feb 14, 2009 at 12:37 AM, Bruno Tarquini <btarquini@xxxxxxxxx> wrote:
> Hi!
>
> When I play teeworlds, my CPU was always 100% EVEN with VSYNC set.
> CPU was even 100% in settings pages, which sounded weird.So I looked the
> code, did profiling... and found that 75% of the time was spend in
> libGLcore:
>
> 4925733 75.7125 libGLcore.so.173.14.15 (no symbols)
> 105697 1.6246 nvidia (no symbols)
> 93014 1.4297 libm-2.6.1.so cosf
> 82255 1.2643 teeworlds gfx_text_ex
> 74637 1.1472 zero (no symbols)
> 69720 1.0717 libm-2.6.1.so sinf
>
> Finally, i disabled gfx_finish and my CPU became happy :)
>
> here some test results:
>
> in video settings page:
> -----------------------
>
> vsync 0 finish 0 => fps 125 cpu 85%
> vsync 0 finish 1 => fps 100 cpu 90%
> vsync 1 finish 0 => fps 60 cpu 25% <-- here
> vsync 1 finish 1 => fps 60 cpu 95%
>
> in game:
> --------
>
> vsync 0 finish 0 => fps 150 cpu 85%
> vsync 0 finish 1 => fps 110 cpu 90% <-- finish reduce my fps !
> vsync 1 finish 0 => fps 60 cpu 30%
> vsync 1 finish 1 => fps 60 cpu 95%
>
> I googled a little and found that glFinish make CPU busy-waiting for
> GPU finish. In fact, calling glFinish after SDL_GL_SwapBuffers() in
> ec_gfx.c:gfx_swap() sounds weird 'cause SDL_GL_SwapBuffers() should
> already flush opengl commands buffers. Maybe this depends of openGL
> implementation.
> for me (linux):
> cpu: AMD Athlon(tm) XP 2800+
> gpu: NVIDIA-173.14.15 geforce FX5700
>
> So IMHO, glFinish() should be removed, or disabled by default or at
> least, when vsync is activated. Other point should be to reduce CPU
> usage when player isn't playing, maybe by increasing cpu_throttle.
>
>
> During my invetigations, I make some patches:
>
> -------------------------------------------------
> Make cpu throttle adjustable.
> user can make cl_cpu_throttle 10 (warn: reduce gaming performances!)
>
> diff -r c4f5cef1bc45 src/engine/client/ec_client.c
> --- a/src/engine/client/ec_client.c Sun Feb 08 18:15:45 2009 +0100
> +++ b/src/engine/client/ec_client.c Thu Feb 12 11:30:38 2009 +0100
> @@ -1816,8 +1816,10 @@
> /* be nice */
> if(config.dbg_stress)
> thread_sleep(5);
> - else if(config.cl_cpu_throttle || !gfx_window_active())
> - thread_sleep(1);
> + else if(config.cl_cpu_throttle)
> + thread_sleep(config.cl_cpu_throttle);
> + else if(!gfx_window_active())
> + thread_sleep(config.cl_cpu_throttle ?
> config.cl_cpu_throttle : 1);
>
> if(config.dbg_hitch)
> {
> diff -r c4f5cef1bc45 src/engine/e_config_variables.h
> --- a/src/engine/e_config_variables.h Sun Feb 08 18:15:45 2009 +0100
> +++ b/src/engine/e_config_variables.h Thu Feb 12 11:30:38 2009 +0100
> @@ -9,7 +9,7 @@
> MACRO_CONFIG_STR(password, 32, "", CFGFLAG_CLIENT, "Password to the
> server")
> MACRO_CONFIG_STR(logfile, 128, "", CFGFLAG_SAVE|CFGFLAG_CLIENT,
> "Filename to log all output to")
>
> -MACRO_CONFIG_INT(cl_cpu_throttle, 0, 0, 1, CFGFLAG_SAVE|CFGFLAG_CLIENT, "")
> +MACRO_CONFIG_INT(cl_cpu_throttle, 0, 0, 100,
> CFGFLAG_SAVE|CFGFLAG_CLIENT, "Adjust sleeping delay in event loop
> (milliseconds)")
> MACRO_CONFIG_INT(cl_editor, 0, 0, 1, CFGFLAG_CLIENT, "")
>
> MACRO_CONFIG_INT(cl_eventthread, 0, 0, 1, CFGFLAG_CLIENT, "Enables the
> usage of a thread to pump the events")
>
>
>
>
> -------------------------------------------------
> Optimize rotate function (oprofile showed it's 35% of cpu usage in
> settings page)
> factorize cos & sin calculation
>
>
> diff -r f43708f5c3f9 src/engine/client/ec_gfx.c
> --- a/src/engine/client/ec_gfx.c Thu Feb 12 11:30:38 2009 +0100
> +++ b/src/engine/client/ec_gfx.c Thu Feb 12 11:54:19 2009 +0100
> @@ -837,8 +837,10 @@
> {
> float x = point->x - center->x;
> float y = point->y - center->y;
> - point->x = x * cosf(rotation) - y * sinf(rotation) + center->x;
> - point->y = x * sinf(rotation) + y * cosf(rotation) + center->y;
> + float c = cosf(rotation);
> + float s = sinf(rotation);
> + point->x = x * c - y * s + center->x;
> + point->y = x * s + y * c + center->y;
> }
>
> void gfx_quads_draw(float x, float y, float w, float h)
> diff -r f43708f5c3f9 src/game/client/render_map.cpp
> --- a/src/game/client/render_map.cpp Thu Feb 12 11:30:38 2009 +0100
> +++ b/src/game/client/render_map.cpp Thu Feb 12 11:54:19 2009 +0100
> @@ -73,8 +73,10 @@
> {
> int x = point->x - center->x;
> int y = point->y - center->y;
> - point->x = (int)(x * cosf(rotation) - y * sinf(rotation) + center->x);
> - point->y = (int)(x * sinf(rotation) + y * cosf(rotation) + center->y);
> + float c = cosf(rotation);
> + float s = sinf(rotation);
> + point->x = (int)(x * c - y * s + center->x);
> + point->y = (int)(x * s + y * c + center->y);
> }
>
> void render_quads(QUAD *quads, int num_quads, void (*eval)(float
> time_offset, int env, float *channels), int renderflags)
>
>
>
> -------------------------------------------------
> Introduce rotate4 which rotate 4 vectors at the same time.
> save some trigo calculation.
>
>
> diff -r f7a250eb5301 src/engine/client/ec_gfx.c
> --- a/src/engine/client/ec_gfx.c Thu Feb 12 11:55:26 2009 +0100
> +++ b/src/engine/client/ec_gfx.c Thu Feb 12 12:29:25 2009 +0100
> @@ -843,6 +843,27 @@
> point->y = x * s + y * c + center->y;
> }
>
> +static void rotate4(VEC3 *center, VEC3 *point1, VEC3 *point2, VEC3
> *point3, VEC3 *point4)
> +{
> + float x, y;
> + float c = cosf(rotation);
> + float s = sinf(rotation);
> +
> +#define ROT_POINT(POINT) \
> + x = (POINT)->x - center->x; \
> + y = (POINT)->y - center->y; \
> + (POINT)->x = x * c - y * s + center->x; \
> + (POINT)->y = x * s + y * c + center->y
> +
> + ROT_POINT(point1);
> + ROT_POINT(point2);
> + ROT_POINT(point3);
> + ROT_POINT(point4);
> +
> +#undef ROT_POINT
> +}
> +
> +
> void gfx_quads_draw(float x, float y, float w, float h)
> {
> gfx_quads_drawTL(x-w/2, y-h/2,w,h);
> @@ -862,26 +883,27 @@
> vertices[num_vertices].pos.y = y;
> vertices[num_vertices].tex = texture[0];
> vertices[num_vertices].color = color[0];
> - rotate(¢er, &vertices[num_vertices].pos);
>
> vertices[num_vertices + 1].pos.x = x+width;
> vertices[num_vertices + 1].pos.y = y;
> vertices[num_vertices + 1].tex = texture[1];
> vertices[num_vertices + 1].color = color[1];
> - rotate(¢er, &vertices[num_vertices + 1].pos);
>
> vertices[num_vertices + 2].pos.x = x + width;
> vertices[num_vertices + 2].pos.y = y+height;
> vertices[num_vertices + 2].tex = texture[2];
> vertices[num_vertices + 2].color = color[2];
> - rotate(¢er, &vertices[num_vertices + 2].pos);
>
> vertices[num_vertices + 3].pos.x = x;
> vertices[num_vertices + 3].pos.y = y+height;
> vertices[num_vertices + 3].tex = texture[3];
> vertices[num_vertices + 3].color = color[3];
> - rotate(¢er, &vertices[num_vertices + 3].pos);
> -
> +
> + rotate4(¢er, &vertices[num_vertices].pos,
> + &vertices[num_vertices + 1].pos,
> + &vertices[num_vertices + 2].pos,
> + &vertices[num_vertices + 3].pos);
> +
> add_vertices(4);
> }
>
> diff -r f7a250eb5301 src/game/client/render_map.cpp
> --- a/src/game/client/render_map.cpp Thu Feb 12 11:55:26 2009 +0100
> +++ b/src/game/client/render_map.cpp Thu Feb 12 12:29:25 2009 +0100
> @@ -77,6 +77,26 @@
> float s = sinf(rotation);
> point->x = (int)(x * c - y * s + center->x);
> point->y = (int)(x * s + y * c + center->y);
> +}
> +
> +static void rotate4(POINT *center, POINT points[], float rotation)
> +{
> + int x, y;
> + float c = cosf(rotation);
> + float s = sinf(rotation);
> +
> +#define ROT_POINT(POINT) \
> + x = (POINT)->x - center->x; \
> + y = (POINT)->y - center->y; \
> + (POINT)->x = (int)(x * c - y * s + center->x); \
> + (POINT)->y = (int)(x * s + y * c + center->y)
> +
> + ROT_POINT(points);
> + ROT_POINT(points + 1);
> + ROT_POINT(points + 2);
> + ROT_POINT(points + 3);
> +
> +#undef ROT_POINT
> }
>
> void render_quads(QUAD *quads, int num_quads, void (*eval)(float
> time_offset, int env, float *channels), int renderflags)
> @@ -146,10 +166,7 @@
> rotated[3] = q->points[3];
> points = rotated;
>
> - rotate(&q->points[4], &rotated[0], rot);
> - rotate(&q->points[4], &rotated[1], rot);
> - rotate(&q->points[4], &rotated[2], rot);
> - rotate(&q->points[4], &rotated[3], rot);
> + rotate4(&q->points[4], rotated, rot);
> }
>
> gfx_quads_draw_freeform(
>
> Regards,
>
> Bruno
>
>
>
>
Other related posts:
