这篇相当于是对前三篇的总结,基本效果如下:
在初试PyOpenGL一 (Python+OpenGL)讲解Pyopengl环境搭建,网格,球体,第一与第三人称摄像机的实现。在初试PyOpenGL二 (Python+OpenGL)基本地形生成与高度检测 里以用高程图生成地形以及以球体做三人称漫游。初试PyOpenGL三 (Python+OpenGL)GPGPU基本运算与乒乓技术 里实现了基本的GPGPU运算。
我认为比较完善的GPU粒子系统应该如下,粒子初始化可以放在CPU里,但是相关数据运算首先要放在GPU里,并且运算后的数据也应该放在显存里,而不是内存里。故用第三篇实现GPU粒子系统不满足,因为他数据是存放在纹理中,要放入VBO里,必需先读取经过内存,然后存放入显存里,这里虽然运算是放入GPU了,但是数据要经过显存-内存-显存的过程,产生不必要的消耗,并且,因为数据是存放在纹理的像素里,故限定在片断着色器中,这二个限制导致第三篇里的内容不能用来实现GPU粒子系统,而是用来实现一些需要结合CPU与GPU结合处理的运算。
在这里,我们采用OpenGL 里的Transform Feedback,和第三篇采用FBO结合浮点纹理不同,Transform Feedback简单来说,传入一个VBO,经过GPU运算后,放入另一个VBO中,注意二点,操作都是针对VBO,也就是针对显存,故不需要经过CPU与内存,还有一点就是在Transform Feedback里,一个缓存不能同时作为输入和输出。
首先来看一下简单的例子介绍Transform Feedback的基本应用,首先指出一点,GLSL3.0与GLSL4.0的Transform Feedback写法有些区别,手上分别有支持3.0与4.0的显示,但是为了更好的兼容性,选择3.0的写法,相应代码和着色器代码如下:
1 tf_v = """
2 #version 330
3 in float inValue;
4 out float outValue;
5 out float out2;
6 void main() {
7 outValue = inValue+3.0;
8 out2 = 1.0;
9 }"""
简单变换反馈的着色器
1 this.tfProgram = glCreateProgram()
2 this.tfProgram = ShaderProgram(this.tfProgram)
3 tfvshader = shaders.compileShader(tf_v,GL_VERTEX_SHADER)
4 glAttachShader(this.tfProgram,tfvshader)
5 LP_LP_c_char = POINTER(POINTER(c_char))
6 ptrs = (c_char_p * 2)('outValue', 'out2')
7 print ptrs,len(ptrs)
8 c_array = cast(ptrs, LP_LP_c_char)
9 glTransformFeedbackVaryings(this.tfProgram, len(ptrs), c_array, GL_INTERLEAVED_ATTRIBS)
10 glLinkProgram(this.tfProgram)
11 this.tfProgram.invalue = glGetAttribLocation(this.tfProgram,"inValue")
着色器基本参数设置
1 class transformFeedback(common):
2 def __init__(this,pro):
3 data = [1.0, 2.0, 3.0, 4.0, 5.0]
4 data1 = [1.0] * 5
5 this.vbo = vbo.VBO(ny.array(data,'f'))
6 this.tbo = vbo.VBO(ny.array(data1,'f'))
7 glUseProgram(pro)
8 pi = pro.invalue
9 #this.vbo = glGenBuffers(1)
10 #glBindBuffer(GL_ARRAY_BUFFER, this.vbo)
11 #output data
12 this.tbo = glGenBuffers(1)
13 glBindBuffer(GL_ARRAY_BUFFER, this.tbo)
14 glBufferData(GL_ARRAY_BUFFER, 40, None, GL_STATIC_DRAW)
15 #input data
16 this.vbo.bind()
17 glEnableVertexAttribArray(pi)
18 #in pyopengl,the glVertexAttribPointer last two params must not be 0,0
19 glVertexAttribPointer(pi,1,GL_FLOAT,False,4*1,this.vbo)
20 glEnable(GL_RASTERIZER_DISCARD)
21 glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER, 0, this.tbo)
22 glBeginTransformFeedback(GL_POINTS)
23 glDrawArrays(GL_POINTS, 0, 5)
24 glEndTransformFeedback()
25 glDisable(GL_RASTERIZER_DISCARD)
26 glDisableVertexAttribArray(pi)
27 glFlush()
28
29 glBindBuffer(GL_ARRAY_BUFFER, this.tbo)
30 buffer = (ctypes.c_float * 10)()
31 #get buffer pointer
32 point = ctypes.cast(buffer, ctypes.POINTER(ctypes.c_float))
33 glGetBufferSubData(GL_ARRAY_BUFFER, 0, 10 * 4,point)
34 #convert pointer to array
35 array = ny.ctypeslib.as_array(point,(10,))
36 print "tf",array
37
38 bf = glMapBuffer(GL_TRANSFORM_FEEDBACK_BUFFER,GL_READ_WRITE)
39 pointv = ctypes.cast(bf, ctypes.POINTER(ctypes.c_float))
40 arrayv = ny.ctypeslib.as_array(pointv,(5,))
41 print "tfv",arrayv
42 glUnmapBuffer(GL_ARRAY_BUFFER)
Transform Feedback基本流程
着色器里代码很简单,传入一个float数据,返回二个float数据,上面我们传入一个数组,[1.0, 2.0, 3.0, 4.0, 5.0],经过着色器里简单运算,分别返回这个数据加3值,与一个固定值1.0.然后在transformFeedback我们为了验证正确与否,需要读取VBO里的数据。在这里,pyopengl可以使用glGetBufferSubData与glMapBuffer来得到VBO里的数据,需要注意的是,python与c之间的一些指针,数据的转换,引入ctype,声明ctype类型的数组,然后转换成对应的指针,填充这个数组后,然后转换把指针转化成numpy里的数组.得到的数据如下:
可以看到,传出的数据是4,1,5,1,6,1,7,1,8,1,对比传入的是1.0, 2.0, 3.0, 4.0, 5.0。验证正确。
下面我们以上面的例子来实现我们的粒子系统,这里先入相关Python代码。
1 class particleSystem(object):
2 def __init__(this,len=1):
3 this.length = len
4 this.cparticles = [0.0] * 7 * len
5 this.nparticles = [0.0] * 7 * len
6 this.index = 0
7 this.center = 0.0,0.0
8 this.currenttime = 0.0
9 this.height = 2.0
10 this.init1()
11 this.createVAO()
12 def init1(this):
13 #pos(x,y,z),vel(x,y,z),time
14 for i in range(this.length):
15 ind = i * 7
16 px,py,pz,tt = ind,ind + 1,ind + 2,ind + 6
17 vx,vy,vz = ind + 3,ind + 4,ind + 5
18 this.cparticles[px] = 0.0
19 this.cparticles[py] = 3.0
20 this.cparticles[pz] = random.uniform(0,5)
21 this.cparticles[vx] = random.random()
22 this.cparticles[vy] = 0.0
23 this.cparticles[vz] = 0.0
24 this.cparticles[tt] = random.uniform(1.0,40.0)#random.uniform(0, 3 * this.height)
25 def createVAO(this):
26 this.currvbo = vbo.VBO(ny.array(this.cparticles,'f'))
27 this.nextvbo = vbo.VBO(ny.array(this.nparticles,'f'))
28 def render(this,program):
29 ind = this.index % 2
30 span = time.time() - this.currenttime if this.currenttime != 0.0 else 0.0
31 invbo,outvbo = (this.currvbo,this.nextvbo) if ind == 0 else (this.nextvbo,this.currvbo)
32 #gpu compute.
33 print span
34 glUseProgram(program)
35 glUniform1f(program.span, span)
36 glUniform1f(program.live, 40)
37 this.update(invbo,outvbo)
38 glUseProgram(0)
39 #draw particle.
40 glColor(0.5,0.8,0.9)
41 glPointSize(3.0)
42 outvbo.bind()
43 glVertexPointer(3,GL_FLOAT,28,outvbo)
44 glDrawArrays(GL_POINTS, 0, this.length)
45 outvbo.unbind()
46 this.index = this.index + 1
47 this.currenttime = time.time()
48 def update(this,fvbo,svbo):
49 #fvbo->shader(GPU)->svbo,should svbo and fvbo both bind.
50 svbo.bind()
51 fvbo.bind()
52 glEnableVertexAttribArray(0)
53 glEnableVertexAttribArray(1)
54 glEnableVertexAttribArray(2)
55 glVertexAttribPointer(0,3,GL_FLOAT,False,4 * 7,fvbo)
56 glVertexAttribPointer(1,3,GL_FLOAT,False,4 * 7,fvbo + 12)
57 glVertexAttribPointer(2,1,GL_FLOAT,False,4 * 7,fvbo + 24)
58 glEnable(GL_RASTERIZER_DISCARD)
59 glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER,0,svbo)
60 glBeginTransformFeedback(GL_POINTS)
61 glDrawArrays(GL_POINTS, 0, this.length)
62 glEndTransformFeedback()
63 glDisable(GL_RASTERIZER_DISCARD)
64 glDisableVertexAttribArray(0)
65 glDisableVertexAttribArray(1)
66 glDisableVertexAttribArray(2)
67 fvbo.unbind()
68 #query gpu data is chage?
69 #svbo.bind()
70 #bf = glMapBuffer(GL_ARRAY_BUFFER,GL_READ_WRITE)
71 #pointv = ctypes.cast(bf, ctypes.POINTER(ctypes.c_float))
72 #arrayv = ny.ctypeslib.as_array(pointv,(70,))
73 #print "tfv",arrayv
74 #glUnmapBuffer(GL_ARRAY_BUFFER)
粒子系统乒乓
结合前面的例子和上文中的乒乓来看,粒子在这里我们每个定义七个数据,前三个用来表示他的位置,后三个用来表示他的速度,最后一个用来表示他在显存里的存活时间。在update就是把数据从一个缓存经过GPU运算放入另一个缓存的过程,例如第一桢,我们传入fvbo,然后数据输出到svbo.在第二桢里,数据就从svbo经过GPU传入到fvbo,第三,第四分别如第一,第二。这样就能实现如第三篇中的乒乓技术。然后在显示render里,我们就用当前输出的缓存里的数据简单的输出显示,本文只是介绍用法,实现如雪花,雨滴,瀑布等特效需要对相关初始化粒子,着色器代码,添加纹理做更改,但是基本处理还是如上。
下面是着色器代码,实现粒子与球的碰撞,也有与地面的交互。代码如下:
1 particle_v = """
2 #version 330
3 in vec3 pos;
4 in vec3 vel;
5 in float time;
6 uniform float span;
7 uniform vec2 planeSacle;
8 uniform sampler2D plane;
9 uniform vec3 sphere;
10 uniform float live;
11 out vec3 outpos;
12 out vec3 outvel;
13 out float outtime;
14 void main() {
15 outpos = pos + vel*span;
16 vec2 uv = vec2(pos.xz/planeSacle + vec2(0.5,0.5));
17 uv.y = 1.0 - uv.y;
18 float hight = texture2D(plane, uv).r;
19 vec3 tvel = vel;
20 //sphere collision
21 float radius = sphere.y;
22 vec3 sphereh = sphere + vec3(0.0,hight,0.0);
23 if(distance(outpos,sphereh) <= radius)
24 {
25 tvel = reflect(vel,normalize(outpos-sphereh))/2.0;
26 }
27 tvel = tvel + vec3(0.0,-0.5,0.0)*span;
28
29 //ground collision
30 if(hight > outpos.y)
31 {
32 outpos.y = hight;
33 tvel = vec3(max(vel.x-span*1.1,0.0),0.0,max(vel.z - span*1.1,0.0));
34 }
35 //update particle live
36 outtime = time + span;
37 if(outtime>=live)
38 {
39 outpos = vec3(0.0,3.0,hight*5.0);
40 outtime = 0.0;
41 tvel = vec3(hight,0.0,0.0);
42 }
43 outvel = tvel;
44 }"""
粒子系统着色器代码
整个过程比较简单,也只考虑一些基本的碰撞,比如球的速度也应该影响碰撞后粒子的方向,但是这里只考虑粒子碰撞球后反射的方向,与地面的碰撞后,不会反弹,会慢慢停止向前移动。
最后一些相关着色器的参数设置代码。
1 this.particleProgram = glCreateProgram()
2 this.particleProgram = ShaderProgram(this.particleProgram)
3 particleshader = shaders.compileShader(particle_v,GL_VERTEX_SHADER)
4 glAttachShader(this.particleProgram,particleshader)
5 LP_LP_c_char = POINTER(POINTER(c_char))
6 ptrs = (c_char_p * 3)('outpos', 'outvel','outtime')
7 c_array = cast(ptrs, LP_LP_c_char)
8 glTransformFeedbackVaryings(this.particleProgram, len(ptrs), c_array, GL_INTERLEAVED_ATTRIBS)
9 glLinkProgram(this.particleProgram)
10 this.particleProgram.pos = glGetAttribLocation(this.particleProgram,"pos")
11 this.particleProgram.vel = glGetAttribLocation(this.particleProgram,"vel")
12 this.particleProgram.time = glGetAttribLocation(this.particleProgram,"time")
13 this.particleProgram.span = glGetUniformLocation(this.particleProgram,"span")
14 this.particleProgram.live = glGetUniformLocation(this.particleProgram,"live")
15 this.particleProgram.plane = glGetUniformLocation(this.particleProgram,"plane")
16 this.particleProgram.planeSacle = glGetUniformLocation(this.particleProgram,"planeSacle")
17 this.particleProgram.sphere = glGetUniformLocation(this.particleProgram,"sphere")
粒子系统参数设置
在本文中,试着用了5千W个粒子,发现初始化很慢,花了十几秒,但是桢数和5000个粒子基本没有差别,从这里可以看出,GPU并行处理的强大之处。
完整代码:PythonGPU粒子系统.zip 操作方式EDSF前后左右移动,WR分别向上与向下,鼠标右键加移动鼠标控制方向,V切换第一人称与第三人称。UP与DOWN切换前面操作的移动幅度。