Showing posts with label vertex array. Show all posts
Showing posts with label vertex array. Show all posts

2009-10-31

Vertex Buffer Objects - speed testing

For drawing 32,768 meshes of 12*16*2 triangles each (1 triangle strip):

OpenGL 2.1 hardware:
VA/IA: ~230ms
VBO: ~30ms

OpenGL 1.4 hardware:
VA/IA: ~230ms
VBO: ~240ms

The note is, although your card may say it supports the VBO extension, or support VBO as per the OpenGL 1.5 specs, it doesn't.

VBO's only work in 1.5 GL's or later, due to hardware restrictions.

Remember; If you worry about optimization, you'd better have timing tests to verify that it even matters. Saving about 10% or more is useful.
Also, small frame optimizations are useless; Anyone can do those, and they generally introduce unmaintainable hacks. Worry about your high level optimizations, like using "most common failed test first" and algorithmic and design changes to maximize your logical speed.

Anyways. Here's a buncha UV spheres, as mentioned (generated using xcore + mingl):



Currently working on the texture functions in mingl. Having to write compressors/decompressors for all the DXT formats. It'll be quite useful though.

Z out. Happy Halloween.

2009-10-10

KDTrees and blender gui synthetics

Remember kids, if you use a synthetic oil, it's not a good idea to switch back to regular.



As far as I understand it, "KDTrees" are simply a data structure that says:

def kdtree:
def node:
x,y,w,h
split
low_node *
high_node *
parent *
direction

And, each node occupies a region of space defined by a x,y,w,h box (or x,y,z) and "splits" that box into TWO boxes, along only one of the axes (either x or y). This way, you can more efficiently partition static objects into a scene; this is like all BSP / octree scene management techniques; However, updating a KD tree is rather expensive, it really is good for storing static scenes though as it requires a far lesser node density and can expand into any other volume as needed.

So, they're used in blender's GUI; I'm getting my mockup GUI system to work slowly, and here are many panels defined by glViewports with text in them. The text system is nothing more than a badass texture font I used some free program to generate with 2 px padding and as a alpha only texture, so the memory required for those letters is quite minimal; although fonts are not conducive to mipmapping, so you must manually provide other LOD fonts for your program. Don't forget that always mipmapping is better for your hardware.



I'm also perfecting my cMesh class, which will provide a full next generation mesh including armature animation and animux system, basically equaling and exceeding the capabilities of all existing high definition games; And doing so in a manner consistent with next gen technology (VBO/Shader) while still retaining the ability to be processed via older cards, at great CPU expense.

It looks something like this:



The concept is present in all games I have hacked, and works like this:

A mesh consists of "tiles", which are separate VA/IA's that contain some amount of triangle strips/triangles/quadstrips with a consistent vertex format and ideally contiguous data. This means each tile can store different vertex parameters, the usual culprits are:

current vertex position (3)
next vertex position + weight (4) //For mesh keyframes only
texcoord 0 (2) //1st texture coordinate
texcoord 1 (2) //2nd texture coordinate (if you have seperate UV mappings, not efficient)
normal (3) //Required for any lighting
binormal (3) //can be shader generated, but space may need to exist in VA array)
matrix indicies (4) //Required to skin a mesh that has more than 1 matrix deforming it
matrix weights (4) //Required to skin a mesh properly

So, each tile can behave as a independent mesh. Above a tile is a tile state, which contains the RC (RenderCommand) array for actually issuing the stripping indicies and material changes and matrix palette updates for this mesh.

Above that is the TileTree, so you can swap/zsort/prioritize and use lod tiles as needed.

the concept of a tile is to load all mesh data into your graphic card, and then merely tell the card what to use to render via a few array binding commands. This is extremely efficient, as it supports mesh instancing to a large degree, and prevents transferring data to the card. The only drawback is you still have to transmit matrix updates to the card, but even a 270 bone character (like Valgirt) this is still far less information that a small VA.

So, once you have tiles setup, we have a generic bone/matrix animation system, with a matrix class (yes, I like quat's too, but matrices require less conversions in a game setting; these matricies can be changed to be only state driven mind you.).

The heart of this system is the base animation classes, which consist of "Keys" which are blocks of static data to interpolate between, "Channels" which are a list of keys and times, and "Animations" which combine channels together to form complete animations.

The Animux is a animation multiplexer, which combines any number of animations together, so that you can use multiple types of animations together, for instance "Run_legs" and "Shoot_Torso" like Quake, or "Face_Phoneme_ma" and "Run". This allows you to make characters that can walk, talk, and run + shoot at the same time, and even use IK calculations as you want.

I've probably done this thrice before, but THIS time, I've got it nailed, and have massive amounts of evidence and experience with the new GLSL to support this type of design. Hopefully, I can get AniStar up and running so I can actually have a program that can make animations for given characters.

Z out.

2008-11-15

I hate GM, I love Blender

So I hate GM again.

Usually doesn't take long. ;P


However, now I have developed a much more robust triangle stripping algorithm that directly translates into huge vertex/VA savings in rendering. This is a really good thing, as instancing a model requires copying the base meshes's vertex VA, which means each instance takes a good chunk of memory (for a 1300 vertex red dragon, that translates to ~1800 vertices for strips, or 3900 for raw triangles. So, each instance requires 4*3*1300 bytes, which is about 16 kB, lol)

Either way, now it's onto making good things like a Memory Manager so I can dump ram as I see fit, and do tracked allocations and all that low level shit that I shouldn't have to do again. Also, I now have to use VBO's (if you card supports them) to do the rendering, and this is a very interesting thing:

VERTEX ARRAYS ARE SLOWER THAN IMMEDIATE MODE

Why, you ask? Draw a picture for yourself. Immediate mode is a 1 direction pipeline. The GPU pipeline just gets pushed data. The end. Vertex Arrays require synchronization of the GPU and system memory, which is about like trying to unflush a toiet. My FPS drops from ~230 to near 20 with immediate mode versus VA's. With VBO's, I can get into the 780 FPS range.

Therefore, use VBO's. Make sure your array data is SEPARATE (do not stride) so you can easily plug in instance VA data.

More testing is required, but here is a toolchain picture that makes me happy:


Note that this is a goofy looking stupid lizard monster. Also, note that it has collision boxes (defined each by two verticies inside the keyframe). Also note, that that collision information is exported correctly into SDL. Does this look anything like Mortal Kombat or Mugen to you? Exactly. The problem is, this lizard has bones. And, I had to write MORE scripts to convert it's animations into mesh keyframes. And MORE scripts to read and export the animation file that defines the animations, as well as game information like attack boxes and physics and so forth.

However, it's really nice to be able to quickly make something, and see it work in the game. No doubting that at all.

Until next update.

-Z