VBOs extremely slow on real hardware (LG Optimus V 670) - Android 2.2.1

Graphics and GPU Programming Programming OpenGL

Started by douglasquaid84 August 09, 2011 05:14 PM

6 comments, last by capricorn 12 years, 8 months ago

137

Author

August 09, 2011 05:14 PM

[font="Arial,"][size="4"]I'm working on a 2d OpenGL graphics engine for my android game, so far I have implemented basic non textured quad rendering via VBOs.

[size="4"]To do this my graphics engine creates a VBO of a quad when initialized and upon rendering draws it using location/dimensions specified by a Polygon2D object.

[size="4"]When rendering 30 - 50 quads on actual hardware (LG Optimus V 670) the frame rate is around 5 - 10 and on the emulator it is 30 - 40.

[size="4"]Here's the code to give a better understanding

[size="4"]

public void CreateBuffers(GL10 gl)

{

    GL11 gl11 = (GL11)gl;



    mQuadBuffer = ByteBuffer.allocateDirect(QUAD2D.length * 4).order(ByteOrder.nativeOrder()).asFloatBuffer();



    mQuadBuffer.put(QUAD2D, 0, QUAD2D.length);





    mQuadBuffer.flip();



    int[] buffer = new int[1];



    gl11.glGenBuffers(1, buffer, 0);

    mQuadVBOId = buffer[0];



    gl11.glBindBuffer(GL11.GL_ARRAY_BUFFER, mQuadVBOId);

    gl11.glBufferData(GL11.GL_ARRAY_BUFFER, mQuadBuffer.capacity() * 4, mQuadBuffer, GL11.GL_STATIC_DRAW);

}







public void draw(GL10 gl) {



    GL11 gl11 = (GL11)gl;



    Polygon2D poly;

    int length = mPolygons.size();



    gl.glEnableClientState(GL11.GL_VERTEX_ARRAY);



    gl11.glBindBuffer(GL11.GL_ARRAY_BUFFER, mQuadVBOId);

    gl11.glVertexPointer(2, GL11.GL_FLOAT, 0, 0);



    for(int i = 0; i < length; i++)

    {



        poly = mPolygons.get(i);



        gl11.glPushMatrix();



        gl11.glTranslatef(poly.x, poly.y, 0);

        gl11.glScalef(poly.width, poly.height, 0);



        gl11.glDrawArrays(GL11.GL_LINE_LOOP, 0, 4);



        gl11.glPopMatrix(); 



    }



    gl11.glBindBuffer(GL11.GL_ARRAY_BUFFER, 0);

    gl.glDisableClientState(GL11.GL_VERTEX_ARRAY);



}

[/font][font="Arial,"][size="4"]Am I doing something wrong, other OpenGL applications seem to run fine such as Replica Island.

[size="4"]I doubt this is useful but here are the specs http://pdadb.net/ind...vm670_optimus_v

[/font]
[font="Arial,"]This is all regarding OpenGL- ES 1.1 for android operating system 2.2[/font]

kilah

514

August 09, 2011 05:44 PM

Hello Cairn!

You are not using VBO correctly. What you are doing is binding a buffer, using matrix operations , and then doing a drawcall. That is incorrect in pretty much every way. VBO are meant to be used to reduce draw calls, therefore what you should do is , create a buffer, fill the buffer and then on your draw function, bind and do the drawcall.

This means that you manually need to do every single operations you are doing in that loop, in another place, and filling an array of values that you will use in a single draw call later on.

Pseudocode





    // Somewhere in your update code

	for(int i = 0; i < length; i++)

    {



        poly = mPolygons.get(i);



		// Do vertex transform operations here, they way you prefer..

        gl11.glPushMatrix();



        gl11.glTranslatef(poly.x, poly.y, 0);

        gl11.glScalef(poly.width, poly.height, 0);



        gl11.glPopMatrix(); 



		// Save data.

        mPolygons.set( i, data );

	}

And later on, your draw code you should only call once glDrawArrays.

Hopefully this helps you.
Cheers.

douglasquaid84

137

Author

August 09, 2011 06:10 PM

I'm sorry if my understanding is incorrect but what your saying is:

initialization - create buffer

update - modify/fill buffer

draw - upload buffer and draw only once

Is that right or did I some how misconstrue what you said?

kilah

514

August 09, 2011 07:21 PM

exactly

douglasquaid84

137

Author

August 09, 2011 07:26 PM

I guess that makes sense if your trying to lessen your draw calls but that means you have to send vertex information to the gpu every update call right?

If that is so how are VBOs when using this method any better then array lists?

21st Century Moose

13,459

August 09, 2011 08:34 PM

Can you confirm whether or not you have a separate VBO for each quad you draw? It looks from your code as if you might, and if so this will kill performance.

Direct3D has need of instancing, but we do not. We have plenty of glVertexAttrib calls.

kilah

514

August 09, 2011 09:16 PM

@Cairn
VBO store data on whatever the driver decides that is the best option at that moment, generally speaking this means VRAM. So what you are doing, is uploading your data using a DMA petition through your driver, which is one of VBO main advantage. If you are uploading and array of data who is not changing, there is no need to upload again your information, and combined with a plausible VRAM storage + lower drawcalls, means your performance will increase a lot compared to direct draw calls (glBegin/glEnd). If you require to update information, VBO will also be faster than direct draw calls due to the fact that you are asking the driver for a DMA petition, and uploading it before doing the draw call. You are not uploading and drawing for each step, as you would with direct draw calls.

Cheers.

capricorn

139

August 09, 2011 09:34 PM

It's not VBOs that are slow, it's how you use them. On each loop iteration you perform two matrix multiplications and a draw call. Seems to me that your VBO contains a single quad (0,0,1,1), and you set it's position and dimensions with glTranslate and glScale calls. This is what impacts your performance (the draw call is negligible in this case). Try making a single VBO for all quads you draw. If you change the quads frequently then you'd want to use GL_DYNAMIC_DRAW for buffer usage. Then, you'll need a loop that updates all quad positions and sizes, and a draw call that draws all the quad (or you can make a second loop right away, because you'll need one anyway when you implement texturing).

Most people tend to think that adding another "long" loop would decrease performance, and they avoid it right away. In practice, it's putting all the calculations in a single loop is what decreases performance, and having several loops that are calculating, sorting and drawing is very often much faster.

VBOs are not "better" because they're in "video" memory and are supposedly not updated that often. They're "better" because the driver manages their memory; other factors like performance are merely consequences. Of course, updating VBO each frame impacts performance, but not as much as having one tiny VBO and spending a lot of CPU cycles to properly draw it.

WBW, capricorn

VBOs extremely slow on real hardware (LG Optimus V 670) - Android 2.2.1

This topic is closed to new replies.

Popular Topics

Recommended Tutorials

VBOs extremely slow on real hardware (LG Optimus V 670) - Android 2.2.1

This topic is closed to new replies.

Popular Topics

Recommended Tutorials

Reticulating splines