1 # Mandelbrot calculate using GPU, Serial numpy and faster numpy
   2 # Use to show the speed difference between CPU and GPU calculations
   3 # ian@ianozsvald.com July 2010
   4 
   5 # Based on vegaseat's TKinter/numpy example code from 2006
   6 # http://www.daniweb.com/code/snippet216851.html#
   7 # with minor changes to move to numpy from the obsolete Numeric
   8 
   9 import sys
  10 import numpy as nm
  11 
  12 import Tkinter as tk
  13 import Image          # PIL
  14 import ImageTk        # PIL
  15 
  16 import pycuda.driver as drv
  17 import pycuda.tools
  18 import pycuda.autoinit
  19 from pycuda.compiler import SourceModule
  20 import pycuda.gpuarray as gpuarray
  21 
  22 # set width and height of window, more pixels take longer to calculate
  23 w = 1000
  24 h = 1000
  25 
  26 from pycuda.elementwise import ElementwiseKernel
  27 complex_gpu = ElementwiseKernel(
  28         "pycuda::complex<float> *z, pycuda::complex<float> *q, int *iteration, int maxiter",
  29             "for (int n=0; n < maxiter; n++) {z[i] = (z[i]*z[i])+q[i]; if (abs(z[i]) > 2.0f) {iteration[i]=n; z[i] = pycuda::complex<float>(); q[i] = pycuda::complex<float>();};}",
  30         "complex5",
  31         preamble="#include <pycuda-complex.hpp>",)
  32 
  33 def calculate_z_gpu(q, maxiter, z):
  34     output = nm.resize(nm.array(0,), q.shape)
  35     q_gpu = gpuarray.to_gpu(q.astype(nm.complex64))
  36     z_gpu = gpuarray.to_gpu(z.astype(nm.complex64))
  37     iterations_gpu = gpuarray.to_gpu(output)
  38     # the for loop and complex calculations are all done on the GPU
  39     # we bring the iterations_gpu array back to determine pixel colours later
  40     complex_gpu(z_gpu, q_gpu, iterations_gpu, maxiter)
  41 
  42     iterations = iterations_gpu.get()
  43     return iterations
  44 
  45 
  46 def calculate_z_numpy_gpu(q, maxiter, z):
  47     """Calculate z using numpy on the GPU via gpuarray"""
  48     outputg = gpuarray.to_gpu(nm.resize(nm.array(0,), q.shape).astype(nm.int32))
  49     zg = gpuarray.to_gpu(z.astype(nm.complex64))
  50     qg = gpuarray.to_gpu(q.astype(nm.complex64))
  51     # 2.0 as an array
  52     twosg = gpuarray.to_gpu(nm.array([2.0]*zg.size).astype(nm.float32))
  53     # 0+0j as an array
  54     cmplx0sg = gpuarray.to_gpu(nm.array([0+0j]*zg.size).astype(nm.complex64))
  55     # for abs_zg > twosg result
  56     comparison_result = gpuarray.to_gpu(nm.array([False]*zg.size).astype(nm.bool))
  57     # we'll add 1 to iterg after each iteration
  58     iterg = gpuarray.to_gpu(nm.array([0]*zg.size).astype(nm.int32))
  59 
  60     for iter in range(maxiter):
  61         zg = zg*zg + qg
  62 
  63         # abs returns a complex (rather than a float) from the complex
  64         # input where the real component is the absolute value (which
  65         # looks like a bug) so I take the .real after abs()
  66         abs_zg = abs(zg).real
  67 
  68         comparison_result = abs_zg > twosg
  69         qg = gpuarray.if_positive(comparison_result, cmplx0sg, qg)
  70         zg = gpuarray.if_positive(comparison_result, cmplx0sg, zg)
  71         outputg = gpuarray.if_positive(comparison_result, iterg, outputg)
  72         iterg = iterg + 1
  73     output = outputg.get()
  74     return output
  75 
  76 
  77 def calculate_z_numpy(q, maxiter, z):
  78     # calculate z using numpy, this is the original
  79     # routine from vegaseat's URL
  80     # NOTE this routine was faster using a default of double-precision complex nbrs
  81     # rather than the current single precision
  82     output = nm.resize(nm.array(0,), q.shape).astype(nm.int32)
  83     for iter in range(maxiter):
  84         z = z*z + q
  85         done = nm.greater(abs(z), 2.0)
  86         q = nm.where(done,0+0j, q)
  87         z = nm.where(done,0+0j, z)
  88         output = nm.where(done, iter, output)
  89     return output
  90 
  91 def calculate_z_serial(q, maxiter, z):
  92     # calculate z using pure python with numpy arrays
  93     # this routine unrolls calculate_z_numpy as an intermediate
  94     # step to the creation of calculate_z_gpu
  95     # it runs slower than calculate_z_numpy
  96     output = nm.resize(nm.array(0,), q.shape).astype(nm.int32)
  97     for i in range(len(q)):
  98         if i % 100 == 0:
  99             # print out some progress info since it is so slow...
 100             print "%0.2f%% complete" % (1.0/len(q) * i * 100)
 101         for iter in range(maxiter):
 102             z[i] = z[i]*z[i] + q[i]
 103             if abs(z[i]) > 2.0:
 104                 q[i] = 0+0j
 105                 z[i] = 0+0j
 106                 output[i] = iter
 107     return output
 108 
 109 
 110 show_instructions = False
 111 if len(sys.argv) == 1:
 112     show_instructions = True
 113 
 114 if len(sys.argv) > 1:
 115     if sys.argv[1] not in ['gpu', 'gpuarray', 'numpy', 'python']:
 116         show_instructions = True
 117 
 118 if show_instructions:
 119     print "Usage: python mandelbrot.py [gpu|gpuarray|numpy|python]"
 120     print "Where:"
 121     print " gpu is a pure CUDA solution on the GPU"
 122     print " gpuarray uses a numpy-like CUDA wrapper in Python on the GPU"
 123     print " numpy is a pure Numpy (C-based) solution on the CPU"
 124     print " python is a pure Python solution on the CPU with numpy arrays"
 125     sys.exit(0)
 126 
 127 routine = {'gpuarray':calculate_z_numpy_gpu,
 128             'gpu':calculate_z_gpu,
 129             'numpy':calculate_z_numpy,
 130             'python':calculate_z_serial}
 131 
 132 calculate_z = routine[sys.argv[1]]
 133 ##if sys.argv[1] == 'python':
 134 #    import psyco
 135 #    psyco.full()
 136 
 137 # Using a WinXP Intel Core2 Duo 2.66GHz CPU (1 CPU used)
 138 # with a 9800GT GPU I get the following timings (smaller is better).
 139 # With 200x200 problem with max iterations set at 300:
 140 # calculate_z_gpu: 0.03s
 141 # calculate_z_serial: 8.7s
 142 # calculate_z_numpy: 0.3s
 143 #
 144 # Using WinXP Intel 2.9GHz CPU (1 CPU used)
 145 # with a GTX 480 GPU I get the following using 1000x1000 plot with 1000 max iterations:
 146 # gpu: 0.07s
 147 # gpuarray: 3.4s
 148 # numpy: 43.4s
 149 # python (serial): 1605.6s
 150 
 151 class Mandelbrot(object):
 152     def __init__(self):
 153         # create window
 154         self.root = tk.Tk()
 155         self.root.title("Mandelbrot Set")
 156         self.create_image()
 157         self.create_label()
 158         # start event loop
 159         self.root.mainloop()
 160 
 161 
 162     def draw(self, x1, x2, y1, y2, maxiter=300):
 163         # draw the Mandelbrot set, from numpy example
 164         xx = nm.arange(x1, x2, (x2-x1)/w*2)
 165         yy = nm.arange(y2, y1, (y1-y2)/h*2) * 1j
 166         # force yy, q and z to use 32 bit floats rather than
 167         # the default 64 doubles for nm.complex for consistency with CUDA
 168         yy = yy.astype(nm.complex64)
 169         q = nm.ravel(xx+yy[:, nm.newaxis]).astype(nm.complex64)
 170         z = nm.zeros(q.shape, nm.complex64)
 171 
 172         start_main = drv.Event()
 173         end_main  = drv.Event()
 174         start_main.record()
 175         output = calculate_z(q, maxiter, z)
 176 
 177         end_main.record()
 178         end_main.synchronize()
 179         secs = start_main.time_till(end_main)*1e-3
 180         print "Main took", secs
 181 
 182         output = (output + (256*output) + (256**2)*output) * 8
 183         # convert output to a string
 184         self.mandel = output.tostring()
 185 
 186     def create_image(self):
 187         """"
 188         create the image from the draw() string
 189         """
 190         self.im = Image.new("RGB", (w/2, h/2))
 191         # you can experiment with these x and y ranges
 192         self.draw(-2.13, 0.77, -1.3, 1.3, 1000)
 193         self.im.fromstring(self.mandel, "raw", "RGBX", 0, -1)
 194 
 195     def create_label(self):
 196         # put the image on a label widget
 197         self.image = ImageTk.PhotoImage(self.im)
 198         self.label = tk.Label(self.root, image=self.image)
 199         self.label.pack()
 200 
 201 # test the class
 202 if __name__ == '__main__':
 203     test = Mandelbrot()