Jump to content

  • Log In with Google      Sign In   
  • Create Account

Opengl Multiple renderings, how to improve performance?


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
10 replies to this topic

#1 8mike   Members   -  Reputation: 127

Like
0Likes
Like

Posted 08 July 2014 - 11:49 AM

Hello everybody I am working on an application that needs to render the scene from multiple points of view. I notice that if I render once, even if the frag shader is long and complicated (writing to multiple 3D textures) it runs at 65 FPS. As soon as I add another rendering pass before that (simply rendering to 2 targets, colour and normals+depth) the framerate drops to 40. If I add a shadowmap pass it drops even further to 25-30 FPS. What is the best way to cope with multiple renderings and still retain a high framerate?

 

This is what I have right now:

1 - a shader that computes the normal+depth and colors, 2 targets on the same FBO

2 - the same shader gets called again for the shadowmap (different uniform parameters) but renders to a different FBO target

3 - a shader that writes values on a 3D texture using imageStore, new FBO

4 - a last shader that performs ray marching through the 3D texture and, at the same time, reads from the 2 maps created before

if i just run step 3 and 4 i get 65 FPS, if i add the 2 steps above the performance drops and the only explanation I have is that I am doing something useless in the middle cause the shaders in 3 and 4 are way more complicated (1000+ lines)

Thank you 



Sponsor:

#2 Rld_   Members   -  Reputation: 1521

Like
0Likes
Like

Posted 08 July 2014 - 01:46 PM

It's not strange that a severely un-optimized shader of 20 lines will get outperformed by a severely optimized shader of 100 lines (so to speak that is). Add the fact that you're using the same shader multiple times to render multiple viewpoints and your performance will go down the drain.

 

With that being said, we have no clue what your shaders look like so we can't really comment on anything if you're doing something wrong or not. It might be in the shaders, it might be in the way you are using the FBO, it might be something totally unrelated to either!

 

I'd suggest using a profiler to see where your bottleneck is, (double) check your shaders (1000+ lines for a single shader seems like a lot, but I don't know what you're aiming for) and check your code for silly mistakes.

 

But if you want help, post some code along with an explanation of what you're trying to achieve and I'm sure these lovely people are able to help you out!


Edited by Rld_, 08 July 2014 - 01:46 PM.


#3 8mike   Members   -  Reputation: 127

Like
0Likes
Like

Posted 08 July 2014 - 04:47 PM

I have Nsights installed in VS 13, would it work as a profiler for glsl? Btw the big shader is fine cause everything runs fast when i run it alone. The other one is slowing things down, Also, I have 5 FBOs with a total of 6 targets, maybe this is one of the things that is slowing down everything



#4 Rld_   Members   -  Reputation: 1521

Like
0Likes
Like

Posted 09 July 2014 - 01:44 AM

I never used Nsights before, but it seems like an option to rule some things out, but I'm not sure if it enables you to profile your shaders. A quick google already showed some options for you that might do the trick a bit better specifically for shaders.



#5 8mike   Members   -  Reputation: 127

Like
0Likes
Like

Posted 10 July 2014 - 06:08 AM

So I started doing some debugging with Nsight and I took a look at the events. This is what I saw:

 

I have numerous calls like this 

void glBindTexture(GLenum target = GL_TEXTURE_2D, GLuint texture = 15)	0x00010000	1539	0
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 1188, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 000354FC, GLint basevertex = 38051)	0x00010000	21038	22504
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 786, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0003678C, GLint basevertex = 38847)	0x00010000	17446	15872
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 786, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 000373D4, GLint basevertex = 39373)	0x00010000	16933	15292
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 786, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0003801C, GLint basevertex = 39899)	0x00010000	21552	14868
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 786, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 00038C64, GLint basevertex = 40425)	0x00010000	17960	14080
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 1188, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 000398AC, GLint basevertex = 40951)	0x00010000	16933	21740
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 1188, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0003AB3C, GLint basevertex = 41747)	0x00010000	16420	21856
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 786, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0003BDCC, GLint basevertex = 42543)	0x00010000	16933	15432
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 786, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0003CA14, GLint basevertex = 43069)	0x00010000	16933	15448
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 786, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0003D65C, GLint basevertex = 43595)	0x00010000	16420	14704
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 786, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0003E2A4, GLint basevertex = 44121)	0x00010000	16933	13788
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 1188, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0003EEEC, GLint basevertex = 44647)	0x00010000	16420	21668

and I guess I could turn those draw calls into a single call.

I then noticed this other thing:

void glBindTexture(GLenum target = GL_TEXTURE_2D, GLuint texture = 10)	0x00010000	1539	0
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 612, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0004017C, GLint basevertex = 45443)	0x00010000	21038	30748
void glBindTexture(GLenum target = GL_TEXTURE_2D, GLuint texture = 12)	0x00010000	1026	0
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 612, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 00040B0C, GLint basevertex = 45959)	0x00010000	20525	28388
void glBindTexture(GLenum target = GL_TEXTURE_2D, GLuint texture = 10)	0x00010000	1026	0
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 444, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0004149C, GLint basevertex = 46475)	0x00010000	20525	55416
void glBindTexture(GLenum target = GL_TEXTURE_2D, GLuint texture = 11)	0x00010000	1026	0
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 480, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 00041B8C, GLint basevertex = 46797)	0x00010000	20525	13716
void glBindTexture(GLenum target = GL_TEXTURE_2D, GLuint texture = 12)	0x00010000	1026	0
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 336, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0004230C, GLint basevertex = 47117)	0x00010000	20525	10060
void glBindTexture(GLenum target = GL_TEXTURE_2D, GLuint texture = 11)	0x00010000	1026	0
void glDrawElementsBaseVertex(GLenum mode = GL_TRIANGLES, GLsizei count = 480, GLenum type = GL_UNSIGNED_INT, GLvoid* indices = 0004284C, GLint basevertex = 47349)	0x00010000	20525	11676

Now in this one as you can notice I keep binding to the same 3 textures back and forth. Would it make sense to sort my elements based on texture object when I load the model, so that I can bind one texture and make all the draw calls and then go to the next?



#6 Rld_   Members   -  Reputation: 1521

Like
0Likes
Like

Posted 10 July 2014 - 06:18 AM

I'm not sure how expensive texture binding is, but it will probably increase performance by some if you keep it all together. Constantly (re)binding shaders is a costly thing, so you probably do want to have that sorted out.

 

Having less draw calls might also help.

 

You might also benefit from timing your instructions, see if something is taking up more time than it should. Most, if not all, profilers have an option to see your highest bottlenecks, but you can also time specific functions yourself if you want that.



#7 Hodgman   Moderators   -  Reputation: 31852

Like
0Likes
Like

Posted 10 July 2014 - 06:22 AM

I notice that if I render once, even if the frag shader is long and complicated (writing to multiple 3D textures) it runs at 65 FPS1. As soon as I add another rendering pass before that (simply rendering to 2 targets, colour and normals+depth) the framerate drops to 40 FPS2. If I add a shadowmap pass it drops even further to 25-30 FPS3

That's 1=15.4ms, 2=25ms, 3=40 to 33ms per frame maximum from the CPU and GPU.

So your original pass is 15.4ms, your added pass is (25-15.4=) 9.6ms, and your shadowmap pass is (33-25= / 40-25=) 8ms to 15 ms.
 
Those all sound like feasible numbers... I don't know why you expect to be able to do more work without spending more milliseconds of GPU time wink.png tongue.png
 
You can use gl_ext_timer_query to measure how long different passes are taking on the GPU. e.g. insert a query before/after your shadowmap pass, and then next frame, retrieve the results of the query to find out what the duration of those commands on the GPU actually was.



#8 8mike   Members   -  Reputation: 127

Like
0Likes
Like

Posted 10 July 2014 - 07:03 AM

Actually I turned off vsync so that I get the actual FPS. So these are the actual results (voxelization is the expensive shader)

 

voxelization alone: 94 FPS

 

normal+depth and color pass alone: 240 FPS (I could improve this by using a g-buffer instead of 2 textures)

 

shadowmap alone: 135 FPS

 

normal+depth+color & shadowmap: 160 FPS (those 2 passes use the same shader)

 

second pass + shadowmap + voxelisation (in this order) : 67 FPS

 

The weird thing is that if I run the shadowmap pass alone it's slower than if i run it after th other pass, so it means that I am probabily wasting something. 

I will go on and try to optimize the voxelisation part. It's the cyril crassin method of rendering everything while using the geometry shader to project to the dominant axis and use the fragment shader to write to a 3d texture. Right now I am writing to the 3d Tex by using imageStore even though that doesn't seem the bottleneck of the shader.

 

If I use gl_ext_timer_query i get those results

 

normal+depth & color : 10,336,416 ns

shadowmap: 5,412,064 ns

voxelise: 55,344,416 ns

which should run at 14 fps...

 

all I did was

glBeginQuery(GL_TIME_ELAPSED_EXT, queries[0]);
firstPass();
glEndQuery(GL_TIME_ELAPSED_EXT);
glBeginQuery(GL_TIME_ELAPSED_EXT, queries[1]);
ShadowMapPass();
glEndQuery(GL_TIME_ELAPSED_EXT);
glDisable(GL_DEPTH_TEST);
glDisable(GL_CULL_FACE);
glBeginQuery(GL_TIME_ELAPSED_EXT, queries[2]);
Voxelize(false);
glEndQuery(GL_TIME_ELAPSED_EXT);

while (!available) {
	glGetQueryObjectiv(queries[4 - 1], GL_QUERY_RESULT_AVAILABLE, &available);
}
for (int i = 0; i < 3; i++) {
  // See how much time the rendering of object i took in nanoseconds.
  glGetQueryObjectui64vEXT(queries[i], GL_QUERY_RESULT, &timeElapsed);
}

I assure that the fps count above is correct because it is the same one that is displayed by NSight



#9 haegarr   Crossbones+   -  Reputation: 4589

Like
0Likes
Like

Posted 10 July 2014 - 07:28 AM

As demonstrated by Hodgman above: Using a non-linear measure like FPS (which is in fact reciprocal) is meaningless for quantitative assessment of relative performance. 

 

However, in your query code, IMHO this line

glGetQueryObjectiv(queries[4 - 1], GL_QUERY_RESULT_AVAILABLE, &available);

should read so

glGetQueryObjectiv(queries[3 - 1], GL_QUERY_RESULT_AVAILABLE, &available);

because you want to wait for queries[2] to be ready, not for (the not existing) queries[3].


Edited by haegarr, 10 July 2014 - 07:30 AM.


#10 8mike   Members   -  Reputation: 127

Like
0Likes
Like

Posted 10 July 2014 - 08:15 AM

sorry, that was my typo, I had an extra query before. Still the results are the same I posted



#11 8mike   Members   -  Reputation: 127

Like
0Likes
Like

Posted 11 July 2014 - 01:09 PM

I also tried this: i put a glFinish() at the end so that all the commands in the queue would be executed at that point. This is what shows up in the debugger

 

1 void glFinish() 0x00010000 1772914 0
2 void glClear(GLbitfield mask = GL_DEPTH_BUFFER_BIT) 0x00010000 7749507 0

 

glClear is taking 7 ms of CPU time?

 

If i run other stuff before it it takes even 12ms. Also I have 2 more glClear and they both take this long. Is this normal?






Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS