Centre Blaise Pascal » Bench4GPU

#!/usr/bin/env python

#

# TrouNoir model using PyOpenCL

#

# CC BY-NC-SA 2011 : <emmanuel.quemener@ens-lyon.fr>

#

# Thanks to Andreas Klockner for PyOpenCL:

# http://mathema.tician.de/software/pyopencl

#

# Interesting links:

# http://viennacl.sourceforge.net/viennacl-documentation.html

# http://enja.org/2011/02/22/adventures-in-pyopencl-part-1-getting-started-with-python/

import pyopencl as cl

import numpy

from PIL import Image

import time,string

from numpy.random import randint as nprnd

import sys

import getopt

import matplotlib.pyplot as plt

KERNEL_CODE=string.Template("""

#define PI (float)3.14159265359

#define nbr 200

float atanp(float x,float y)

{

  float angle;

  angle=atan2(y,x);

  if (angle<0)

    {

      angle+=(float)2.e0f*PI;

    }

  return angle;

}

float f(float v)

{

  return v;

}

float g(float u,float m,float b)

{

  return (3.*m/b*pow(u,2)-u);

}

void calcul(float *us,float *vs,float up,float vp,

            float h,float m,float b)

{

  float c0,c1,c2,c3,d0,d1,d2,d3;

  c0=h*f(vp);

  c1=h*f(vp+c0/2.);

  c2=h*f(vp+c1/2.);

  c3=h*f(vp+c2);

  d0=h*g(up,m,b);

  d1=h*g(up+d0/2.,m,b);

  d2=h*g(up+d1/2.,m,b);

  d3=h*g(up+d2,m,b);

  *us=up+(c0+2.*c1+2.*c2+c3)/6.;

  *vs=vp+(d0+2.*d1+2.*d2+d3)/6.;

}

void rungekutta(float *ps,float *us,float *vs,

                float pp,float up,float vp,

                float h,float m,float b)

{

  calcul(us,vs,up,vp,h,m,b);

  *ps=pp+h;

}

float decalage_spectral(float r,float b,float phi,

                         float tho,float m)

{

  return (sqrt(1-3*m/r)/(1+sqrt(m/pow(r,3))*b*sin(tho)*sin(phi)));

}

float spectre(float rf,int q,float b,float db,

             float h,float r,float m,float bss)

{

  float flx;

  int fi;

  fi=(int)(rf*nbr/bss);

  flx=pow(r/m,q)*pow(rf,4)*b*db*h;

  return(flx);

}

float spectre_cn(float rf,float b,float db,

                float h,float r,float m,float bss)

{

  float flx;

  float nu_rec,nu_em,qu,v,temp,temp_em,flux_int,m_point,planck,c,k;

  int fi,posfreq;

  planck=6.62e-34;

  k=1.38e-23;

  temp=3.e7;

  m_point=1.e14;

  v=1.-3./r;

  qu=1/sqrt(1-3./r)/sqrt(r)*(sqrt(r)-sqrt(6.)+sqrt(3.)/2.*log((sqrt(r)+sqrt(3.))/(sqrt(r)-sqrt(3.))*(sqrt(6.)-sqrt(3.))/(sqrt(6.)+sqrt(3.))));

  temp_em=temp*sqrt(m)*exp(0.25*log(m_point))*exp(-0.75*log(r))*

    exp(-0.125*log(v))*exp(0.25*log(qu));

  flux_int=0;

  flx=0;

  for (fi=1;fi<nbr;fi++)

    {

      nu_em=bss*fi/nbr;

      nu_rec=nu_em*rf;

      posfreq=1./bss*nu_rec*nbr;

      if ((posfreq>0)&&(posfreq<nbr))

        {

          flux_int=2*planck/9.e16f*pow(nu_em,3)/(exp(planck*nu_em/k/temp_em)-1.)*pow(rf,3)*b*db*h;

          flx+=flux_int;

        }

    }

  return(flx);

}

void impact(float phi,float r,float b,float tho,float m,

            float *zp,float *fp,

            int q,float db,

            float h,float bss,int raie)

{

  float flx,rf;

  rf=decalage_spectral(r,b,phi,tho,m);

  if (raie==0)

    {

      flx=spectre(rf,q,b,db,h,r,m,bss);

    }

  else

    {

      flx=spectre_cn(rf,b,db,h,r,m,bss);

    }

  *zp=1./rf;

  *fp=flx;

}

__kernel void EachPixel(__global float *MyImage)

{

   uint xi=(uint)get_global_id(0);

   uint yi=(uint)get_global_id(1);

   uint sizex=(uint)get_global_size(0);

   uint sizey=(uint)get_global_size(1);

   // Perform trajectory for each pixel

  float m,rs,ri,re,tho,ro;

  int q;

  float bss,stp;

  int nmx,dim;

  float d,bmx,db,b,h;

  float up,vp,pp;

  float us,vs,ps;

  float rp[2000];

  float zmx,fmx,zen,fen;

  float flux_tot,impc_tot;

  float phi,thi,thx,phd,php,nr,r;

  int ni,ii,i,imx,j,n,tst,dist,raie,pc,fcl,zcl;

  int nh;

  int ExitOnImpact;

  float zp,fp;

  m=1.;

  rs=2.*m;

  ri=3.*rs;

  re=12.;

  ro=100.;

  //tho=PI/180.*80;

  tho=PI/180.*0.;

  q=-2;

  dist=1;

  raie=0;

  nmx=(float)sizex/2.;

  stp=0.5;

  // Autosize for image

  bmx=1.25*re;

  b=0.;

  thx=asin(bmx/ro);

  pc=0;

  h=PI/5.e2f;

  // set origin as center of image

  float x=(float)xi-(float)(sizex/2)+(float)5e-1f;

  float y=(float)(sizey/2)-(float)yi-(float)5e-1f;

  // angle extracted from cylindric symmetry

  phi=atanp(x,y);

  phd=atanp(cos(phi)*sin(tho),cos(tho));

  if (dist==1)

  {

     db=bmx/nmx;

     // impact parameter

     b=sqrt(x*x+y*y)*(float)2.e0f/(float)sizex*bmx;

     up=0.;

     vp=1.;

  }

  else

  {

     b=sqrt(x*x+y*y)*(float)2.e0f/(float)sizex*bmx;

     thi=asin(b/ro);

     db=bmx/nmx;

     up=sin(thi);

     vp=cos(thi);

  }

  pp=0.;

  nh=0;

  rungekutta(&ps,&us,&vs,pp,up,vp,h,m,b);

  rp[nh]=fabs(b/us);

  ExitOnImpact=0;

  do

  {

     nh++;

     pp=ps;

     up=us;

     vp=vs;

     rungekutta(&ps,&us,&vs,pp,up,vp,h,m,b);

     rp[nh]=fabs(b/us);

     ExitOnImpact = (((fmod(pp,PI)<fmod(phd,PI))&&(fmod(ps,PI)>fmod(phd,PI)))||((fmod(ps,PI)<fmod(phd,PI))&&(fmod(pp,PI)>fmod(phd,PI))))&&(rp[nh]>ri)&&(rp[nh]<re)?1:0;

  } while ((rp[nh]>=rs)&&(rp[nh]<=rp[0])&&(ExitOnImpact==0));

  if (ExitOnImpact==1) {

     impact(phi,rp[nh-1],b,tho,m,&zp,&fp,q,db,h,bss,raie);

  }

  else

  {

     zp=0.;

  }

//

  MyImage[yi+sizex*xi]=(float)zp;

  barrier(CLK_LOCAL_MEM_FENCE);

}

""")

def ImageOutput(sigma,prefix):

    Max=sigma.max()

    Min=sigma.min()

    # Normalize value as 8bits Integer

    SigmaInt=(255*(sigma-Min)/(Max-Min)).astype('uint8')

    image = Image.fromarray(SigmaInt)

    image.save("%s.jpg" % prefix)

def BlackHole(MyImage,Device):

    kernel_params = {}

    # Je detecte un peripherique GPU dans la liste des peripheriques

    Id=0

    HasXPU=False

    for platform in cl.get_platforms():

        for device in platform.get_devices():

            if Id==Device:

                XPU=device

                print "CPU/GPU selected: ",device.name.lstrip()

                HasXPU=True

            Id+=1

    if HasXPU==False:

        print "No XPU #%i found in all of %i devices, sorry..." % (Device,Id-1)

        sys.exit()

    ctx = cl.Context([XPU])

    queue = cl.CommandQueue(ctx,

                            properties=cl.command_queue_properties.PROFILING_ENABLE)

    BlackHoleCL = cl.Program(ctx,KERNEL_CODE.substitute(kernel_params)).build()

    # Je recupere les flag possibles pour les buffers

    mf = cl.mem_flags

    print(MyImage)

    # Program based on Kernel2

    MyImageCL = cl.Buffer(ctx, mf.WRITE_ONLY | mf.COPY_HOST_PTR, hostbuf=MyImage)

    start_time=time.time()

    CLLaunch=BlackHoleCL.EachPixel(queue,(MyImage.shape[0],MyImage.shape[1]),None,MyImageCL)

    CLLaunch.wait()

    elapsed = time.time()-start_time

    print("Elapsed %f: " % elapsed)

    cl.enqueue_copy(queue,MyImage,MyImageCL).wait()

    print(MyImage)

    MyImageCL.release()

    ImageOutput(MyImage,"TrouNoir")

    return(elapsed)

if __name__=='__main__':

    GpuStyle = 'OpenCL'

    Mass = 1.

    # Internal Radius 3 times de Schwarzschild Radius

    InternalRadius=6.*Mass

    #

    ExternalRadius=12.

    #

    ObserverDistance=100.

    # Angle with normal to disc 10 degrees

    Angle = numpy.pi/180.*(90.-10.)

    # Radiation of disc : BlackBody or Monochromatic

    BlackBody = True

    # Camera : Angular Camera or plate with the size of camera

    AngularCamera = True

    # Size of image

    Size=256

    # Variable Type

    VariableType='FP32'

    # ?

    q=-2

    HowToUse='%s -h [Help] -b [BlackBodyEmission] -c [AngularCamera] -s <SizeInPixels> -m <Mass> -i <DiscInternalRadius> -x <DiscExternalRadius> -o <ObservatorDistance> -a <AngleAboveDisc> -d <DeviceId> -g <CUDA/OpenCL> -t <FP32/FP64>'

    try:

        opts, args = getopt.getopt(sys.argv[1:],"hbcs:m:i:x:o:a:d:g:t:",["blackbody","camera","size=","mass=","internal=","external=","observer=","angle=","device=","gpustyle=","variabletype="])

    except getopt.GetoptError:

        print(HowToUse % sys.argv[0])

        sys.exit(2)

    # List of Devices

    Devices=[]

    Alu={}

    for opt, arg in opts:

        if opt == '-h':

            print(HowToUse % sys.argv[0])

            print("\nInformations about devices detected under OpenCL API:")

            # For PyOpenCL import

            try:

                import pyopencl as cl

                Id=0

                for platform in cl.get_platforms():

                    for device in platform.get_devices():

                        #deviceType=cl.device_type.to_string(device.type)

                        deviceType="xPU"

                        print("Device #%i from %s of type %s : %s" % (Id,platform.vendor.lstrip(),deviceType,device.name.lstrip()))

                        Id=Id+1

            except:

                print("Your platform does not seem to support OpenCL")

            print("\nInformations about devices detected under CUDA API:")

                # For PyCUDA import

            try:

                import pycuda.driver as cuda

                cuda.init()

                for Id in range(cuda.Device.count()):

                    device=cuda.Device(Id)

                    print("Device #%i of type GPU : %s" % (Id,device.name()))

                print

            except:

                print("Your platform does not seem to support CUDA")

            sys.exit()

        elif opt in ("-d", "--device"):

#            Devices.append(int(arg))

            Device=int(arg)

        elif opt in ("-g", "--gpustyle"):

            GpuStyle = arg

        elif opt in ("-t", "--variabletype"):

            VariableType = arg

        elif opt in ("-s", "--size"):

            Size = int(arg)

        elif opt in ("-m", "--mass"):

            Mass = float(arg)

        elif opt in ("-i", "--internal"):

            InternalRadius = float(arg)

        elif opt in ("-e", "--external"):

            ExternalRadius = float(arg)

        elif opt in ("-o", "--observer"):

            ObserverDistance = float(arg)

        elif opt in ("-a", "--angle"):

            Angle = numpy.pi/180.*(90.-float(arg))

        elif opt in ("-b", "--blackbody"):

            BlackBody = True

        elif opt in ("-c", "--camera"):

            AngularCamera = True

    print("Device Identification selected : %s" % Device)

    print("GpuStyle used : %s" % GpuStyle)

    print("VariableType : %s" % VariableType)

    print("Size : %i" % Size)

    print("Mass : %f" % Mass)

    print("Internal Radius : %f" % InternalRadius)

    print("External Radius : %f" % ExternalRadius)

    print("Observer Distance : %f" % ObserverDistance)

    print("Angle with normal of (in radians) : %f" % Angle)

    print("Black Body Disc Emission (monochromatic instead) : %s" % BlackBody)

    print("Angular Camera (dimension of object instead) : %s" % AngularCamera)

    if GpuStyle=='CUDA':

        print("\nSelection of CUDA device")

        try:

            # For PyCUDA import

            import pycuda.driver as cuda

            cuda.init()

            for Id in range(cuda.Device.count()):

                device=cuda.Device(Id)

                print("Device #%i of type GPU : %s" % (Id,device.name()))

                if Id in Devices:

                    Alu[Id]='GPU'

        except ImportError:

            print("Platform does not seem to support CUDA")

    if GpuStyle=='OpenCL':

        print("\nSelection of OpenCL device")

        try:

            # For PyOpenCL import

            import pyopencl as cl

            Id=0

            for platform in cl.get_platforms():

                for device in platform.get_devices():

                    #deviceType=cl.device_type.to_string(device.type)

                    deviceType="xPU"

                    print("Device #%i from %s of type %s : %s" % (Id,platform.vendor.lstrip().rstrip(),deviceType,device.name.lstrip().rstrip()))

                    if Id in Devices:

                    # Set the Alu as detected Device Type

                        Alu[Id]=deviceType

                    Id=Id+1

        except ImportError:

            print("Platform does not seem to support OpenCL")

#    print(Devices,Alu)

#    MyImage=numpy.where(numpy.random.zeros(Size,Size)>0,1,-1).astype(numpy.float32)

    MyImage=numpy.zeros((Size,Size),dtype=numpy.float32)

    print(MyImage)

    duration=BlackHole(MyImage,Device)