Started by Nov 24 2012 05:51 AM

,
7 replies to this topic

Posted 24 November 2012 - 05:51 AM

Hi,

I am trying to simulate a starfield of rotating cubes moving towards the user:

I am struggling to maintain a smooth animation on Android. I can't see why a Dual-core 1 GHz Cortex-A9 with a ULP GeForce GPU should not be able to do that. Here's what I do:

INITIALIZATION:

- Create a VBO containing 36 cube position vertices and 36 cube normal vertices, interleaved

- Set up a simple shader that takes positions, normals, color, MV matrix, MVP matrix, light position

- Connect the VBO to the shader's position and normal attribute respectively

EACH FRAME:

a] Movement (for each cube)

- calculate scaling matrix, rotation matrix A, translation matrix, rotation matrix B

- multiply all of the above matrices to obtain model matrix

- multiply model matrix and view matrix to obtain MV matrix

- multiply MV matrix and projection matrix to obtain MVP matrix

b] Draw (for each cube)

- hook up cube's MV matrix with shader (glUniformMatrix4fv)

- hook up cube's MVP matrix with shader (glUniformMatrix4fv)

- hook up cube's color constant with shader (glVertexAttrib4f)

- draw cube (glDrawArrays)

I have done a lot of profiling and timing analyzation, so I don't think I have any unintended performance leaks in my code. Rather, I believe there is a more general design flaw, and I hope someone can help me improve it.

I am trying to simulate a starfield of rotating cubes moving towards the user:

I am struggling to maintain a smooth animation on Android. I can't see why a Dual-core 1 GHz Cortex-A9 with a ULP GeForce GPU should not be able to do that. Here's what I do:

INITIALIZATION:

- Create a VBO containing 36 cube position vertices and 36 cube normal vertices, interleaved

- Set up a simple shader that takes positions, normals, color, MV matrix, MVP matrix, light position

- Connect the VBO to the shader's position and normal attribute respectively

EACH FRAME:

a] Movement (for each cube)

- calculate scaling matrix, rotation matrix A, translation matrix, rotation matrix B

- multiply all of the above matrices to obtain model matrix

- multiply model matrix and view matrix to obtain MV matrix

- multiply MV matrix and projection matrix to obtain MVP matrix

b] Draw (for each cube)

- hook up cube's MV matrix with shader (glUniformMatrix4fv)

- hook up cube's MVP matrix with shader (glUniformMatrix4fv)

- hook up cube's color constant with shader (glVertexAttrib4f)

- draw cube (glDrawArrays)

I have done a lot of profiling and timing analyzation, so I don't think I have any unintended performance leaks in my code. Rather, I believe there is a more general design flaw, and I hope someone can help me improve it.

Posted 24 November 2012 - 06:06 AM

Do you have any profiling tools that allow you to measure CPU and GPU timings independently? The first step will be determining which processor is the bottleneck, so you can focus your optimisations usefully.

That said, a solution will probably involve call glDraw less often than once-per-cube.

On the CPU side, every "batch" of geometry has a certain amount of overhead, so you want to reduce the total amount of gl* calls, particularly glDraw* calls.

On the GPU side, you want every glDraw* "batch" to contain as many vertices and as many pixels as possible. It seems like many of your cubes cover hardly any pixels, which may greatly exaggerate the per-draw overheads.

That said, a solution will probably involve call glDraw less often than once-per-cube.

On the CPU side, every "batch" of geometry has a certain amount of overhead, so you want to reduce the total amount of gl* calls, particularly glDraw* calls.

On the GPU side, you want every glDraw* "batch" to contain as many vertices and as many pixels as possible. It seems like many of your cubes cover hardly any pixels, which may greatly exaggerate the per-draw overheads.

Posted 24 November 2012 - 06:24 AM

That said, a solution will probably involve call glDraw less often than once-per-cube.

True, but

Posted 24 November 2012 - 06:40 AM

That said, a solution will probably involve call glDraw less often than once-per-cube.

True, buthow? Since every cube needs its own scaling, rotation and translation, how can I combine glDrawArrays() calls?

Option 1: Software transform the vertices. That is, do the matrix multiply on the CPU, then you can send all the cubes in one go.

Option 2: 'Hardware skinning' style solution. Instead of doing one cube at a time, put 16* cubes into your VBO. The vertices for each cube have indices, which you use in your vertex shader to look up into an array of matrix uniforms.

From experience doing similar things on iOS, I'd expect option 1 to be the better choice.

*Probably you'll want this number to be as large as possible within the constaints of the max amount of uniform space your GPU supports.

**Edited by C0lumbo, 24 November 2012 - 06:40 AM.**

Posted 24 November 2012 - 07:27 AM

you don't need to do matrix mult to translate for example

1D, 2D or 3D translation is 3, 6 and 9 flops respectively

rotation is harder to optimize, but a 2D rotation is alot cheaper

1D, 2D or 3D translation is 3, 6 and 9 flops respectively

rotation is harder to optimize, but a 2D rotation is alot cheaper

Posted 22 December 2012 - 07:52 AM

I know this thread is a month old, but I have invested serious thought and work into the project, and I'd like to post an update to show that I aprecciate your answers, and to help other people with similar issues.

**PROFILING**

Do you have any profiling tools that allow you to measure CPU and GPU timings independently? The first step will be determining which processor is the bottleneck, so you can focus your optimisations usefully.

I believe that you have a very good point here; I have mostly been "optimizing blindly", which is considered bad practice. I have tried / considered the following options:

**1. Android SDK Tools:** the profiler shipped with the SDK shows CPU time, but not GPU time. Also, it profiles my own application, but I would like to see what's going on in the whole system. Google has recognized this and published a tool for system-wide profiling called *systrace*, but it's only available for Jelly Bean. The same goes for *dumpsys gfxinfo* which in combination with the Jelly Bean developer option "Profile GPU Rendering" outputs a stat about how much time is spent processing, drawing and swapping UI. See Chet Haase and Guy Romain's excellent presentation about Android UI performance for more information about those tools. For me, they are not an option, I am stuck with Honeycomb for various reasons.

**2. Log output:** yes I know it is stone age, but I thought it would be interesting to see how much time my application spends in my *move()* (CPU) and *draw()* (GPU) methods. The results are not very conclusive; I guess this has to to with multithreading and the way Android handles vsync.

**3. NVIDIA Tools:** there is a tool for Tegra 2 platforms called PerfHUD ES that looks very promising: detailed information about draw calls and lots of other GPU-related information. I am currently trying to get this running on honeycomb. Any help aprecciated.

**OPTIMIZING**

Option 1: Software transform the vertices. That is, do the matrix multiply on the CPU, then you can send all the cubes in one go.

Option 2: 'Hardware skinning' style solution. Instead of doing one cube at a time, put 16* cubes into your VBO. The vertices for each cube have indices, which you use in your vertex shader to look up into an array of matrix uniforms.

Both of your options seem very reasonable approaches to take. I decided to implement option 2 first. The OpenGL ES specification calls this method "Dynamic Indexing". It took me an hour to rearrange my application accordingly, and a whole day to find out how to get hold of the vertex index inside the shader and use it to address my transformation matrices. It's not straightforward on OpenGL ES 2.0 because for some fucking reason they decided to leave out the crucial gl_VertexID variable there. Sorry, but this tiny little detail really drove me mad. Anyways, the solution is quite simple, once you know how to do it. Anton Holmquist has a short, to-the-point blog post about it, which I wish I'd found earlier.

The one big drawback about this method is, like you pointed out, the limited uniform space. For those who don't have a clue what that is (like I did): it is the space available for declaring uniform variables in the shader. I read somewhere that this space relates to the amount / size of registers the GPU has - correct me if I'm wrong here. For anyone interested, calling *glGetIntegerv(GL_MAX_VERTEX_UNIFORM_VECTORS)* will return a number that says how much uniform space you have on your system. The specification for OpenGL ES 2.0 says it has to be at least 128. The number is expressed in vectors, and since each matrix has 4 vectors, that means you could declare a maximum of 32 matrix variables or, in my case, two arrays of 16 matrices. In other words, I can batch a maximum of 16 cubes now.

**NEXT**

Dynamic indexing has certainly improved performance, but I am not entirely happy yet. I will implement the abovementioned option 1, hoping to improve performance by shifting work from the GPU to the CPU and, as a next step, parallize the *move()* and *draw()* operations.

Posted 22 December 2012 - 11:05 PM

Doing the the matrix calculations in software shouldn't be too bad. But if you are CPU limited (maybe the case on Android) then here is an alternative that does the matrix calculation on the GPU:

Pass a vec4 Rotation (rotX, rotY, rotZ, angle) to your vertex shader. Then calculate the rotation matrix in your shader:

mat4 transform = mat4(1.0); //identity ... make rotation matrix from axis and angle ... gl_Position = projection * view * transform * a_position;

Theoretically you could put the Position and Normal attributes into a separate static VBO since they are unchanged (and update the view matrix instead). If you fill all your instance data in one go, you could end up with a single draw call per frame.

Posted 23 December 2012 - 02:42 AM

I do think the software approach will beat the vertex shader approach, so you should probably go ahead and give that a shot.

But - I think you should be able to manage more than 16 cubes per batch. Firstly, you don't need a full 4x4 matrix - you can switch to 4x3 matrices and implicitly assume that the last row is 0, 0, 0, 1. You might need to transpose your matrices to achieve this.

Also, I don't think you need two arrays of matrices. I assume one of the matrices is for your normals, and one for the positions. But you can usually use the same matrix for your positions and your normals and simply ignore the position for the normals.

So, by my reckoning you should be able to manage 128/3=42 cubes each batch. Your GLSL code might end up looking a bit like this (untested, not compiled):

uniform float4 g_vMatrices[42*3];

...

vWorldPos.x = dot(g_vMatrices[iCubeIndexAttribute*3], float4(vPositionAttribute, 1.0));

vWorldPos.y = dot(g_vMatrices[iCubeIndexAttribute*3+1], float4(vPositionAttribute, 1.0));

vWorldPos.z = dot(g_vMatrices[iCubeIndexAttribute*3+2], float4(vPositionAttribute, 1.0));

vWorldNormal.x = dot(g_vMatrices[iCubeIndexAttribute*3], float4(vNormalAttribute, 0.0));

vWorldNormal.y = dot(g_vMatrices[iCubeIndexAttribute*3+1], float4(vNormalAttribute, 0.0));

vWorldNormal.z = dot(g_vMatrices[iCubeIndexAttribute*3+2], float4(vNormalAttribute, 0.0));

Oops - just realised you'd then need to transform the positions by the viewproj matrix which will take up a few more uniforms, so you'll end up with only 41 cubes per batch.

But - I think you should be able to manage more than 16 cubes per batch. Firstly, you don't need a full 4x4 matrix - you can switch to 4x3 matrices and implicitly assume that the last row is 0, 0, 0, 1. You might need to transpose your matrices to achieve this.

Also, I don't think you need two arrays of matrices. I assume one of the matrices is for your normals, and one for the positions. But you can usually use the same matrix for your positions and your normals and simply ignore the position for the normals.

So, by my reckoning you should be able to manage 128/3=42 cubes each batch. Your GLSL code might end up looking a bit like this (untested, not compiled):

uniform float4 g_vMatrices[42*3];

...

vWorldPos.x = dot(g_vMatrices[iCubeIndexAttribute*3], float4(vPositionAttribute, 1.0));

vWorldPos.y = dot(g_vMatrices[iCubeIndexAttribute*3+1], float4(vPositionAttribute, 1.0));

vWorldPos.z = dot(g_vMatrices[iCubeIndexAttribute*3+2], float4(vPositionAttribute, 1.0));

vWorldNormal.x = dot(g_vMatrices[iCubeIndexAttribute*3], float4(vNormalAttribute, 0.0));

vWorldNormal.y = dot(g_vMatrices[iCubeIndexAttribute*3+1], float4(vNormalAttribute, 0.0));

vWorldNormal.z = dot(g_vMatrices[iCubeIndexAttribute*3+2], float4(vNormalAttribute, 0.0));

Oops - just realised you'd then need to transform the positions by the viewproj matrix which will take up a few more uniforms, so you'll end up with only 41 cubes per batch.

**Edited by C0lumbo, 23 December 2012 - 02:45 AM.**