root / Ising / GPU / Ising2D-GPU.py @ 94
Historique | Voir | Annoter | Télécharger (23,63 ko)
1 |
#!/usr/bin/env python
|
---|---|
2 |
#
|
3 |
# Ising2D model in serial mode
|
4 |
#
|
5 |
# CC BY-NC-SA 2011 : <emmanuel.quemener@ens-lyon.fr>
|
6 |
|
7 |
import sys |
8 |
import numpy |
9 |
import math |
10 |
from PIL import Image |
11 |
from math import exp |
12 |
from random import random |
13 |
import time |
14 |
import getopt |
15 |
import matplotlib.pyplot as plt |
16 |
from numpy.random import randint as nprnd |
17 |
|
18 |
KERNEL_CODE_OPENCL="""
|
19 |
|
20 |
// Marsaglia RNG very simple implementation
|
21 |
#define znew ((z=36969*(z&65535)+(z>>16))<<16)
|
22 |
#define wnew ((w=18000*(w&65535)+(w>>16))&65535)
|
23 |
#define MWC (znew+wnew)
|
24 |
#define SHR3 (jsr=(jsr=(jsr=jsr^(jsr<<17))^(jsr>>13))^(jsr<<5))
|
25 |
#define CONG (jcong=69069*jcong+1234567)
|
26 |
#define KISS ((MWC^CONG)+SHR3)
|
27 |
|
28 |
#define MWCfp MWC * 2.328306435454494e-10f
|
29 |
#define KISSfp KISS * 2.328306435454494e-10f
|
30 |
|
31 |
__kernel void MainLoopOne(__global char *s,float T,float J,float B,
|
32 |
uint sizex,uint sizey,
|
33 |
uint iterations,uint seed_w,uint seed_z)
|
34 |
|
35 |
{
|
36 |
uint z=seed_z;
|
37 |
uint w=seed_w;
|
38 |
|
39 |
for (uint i=0;i<iterations;i++) {
|
40 |
|
41 |
uint x=(uint)(MWC%sizex) ;
|
42 |
uint y=(uint)(MWC%sizey) ;
|
43 |
|
44 |
int p=s[x+sizex*y];
|
45 |
|
46 |
int d=s[x+sizex*((y+1)%sizey)];
|
47 |
int u=s[x+sizex*((y-1)%sizey)];
|
48 |
int l=s[((x-1)%sizex)+sizex*y];
|
49 |
int r=s[((x+1)%sizex)+sizex*y];
|
50 |
|
51 |
float DeltaE=2.0f*p*(J*(u+d+l+r)+B);
|
52 |
|
53 |
int factor=((DeltaE < 0.0f) || (MWCfp < exp(-DeltaE/T))) ? -1:1;
|
54 |
s[x%sizex+sizex*(y%sizey)] = (char)factor*p;
|
55 |
}
|
56 |
barrier(CLK_GLOBAL_MEM_FENCE);
|
57 |
}
|
58 |
|
59 |
__kernel void MainLoopGlobal(__global char *s,__global float *T,float J,float B,
|
60 |
uint sizex,uint sizey,
|
61 |
uint iterations,uint seed_w,uint seed_z)
|
62 |
|
63 |
{
|
64 |
uint z=seed_z/(get_global_id(0)+1);
|
65 |
uint w=seed_w/(get_global_id(0)+1);
|
66 |
float t=T[get_global_id(0)];
|
67 |
uint ind=get_global_id(0);
|
68 |
|
69 |
for (uint i=0;i<iterations;i++) {
|
70 |
|
71 |
uint x=(uint)(MWC%sizex) ;
|
72 |
uint y=(uint)(MWC%sizey) ;
|
73 |
|
74 |
int p=s[x+sizex*(y+sizey*ind)];
|
75 |
|
76 |
int d=s[x+sizex*((y+1)%sizey+sizey*ind)];
|
77 |
int u=s[x+sizex*((y-1)%sizey+sizey*ind)];
|
78 |
int l=s[((x-1)%sizex)+sizex*(y+sizey*ind)];
|
79 |
int r=s[((x+1)%sizex)+sizex*(y+sizey*ind)];
|
80 |
|
81 |
float DeltaE=2.0f*p*(J*(u+d+l+r)+B);
|
82 |
|
83 |
int factor=((DeltaE < 0.0f) || (MWCfp < exp(-DeltaE/t))) ? -1:1;
|
84 |
s[x%sizex+sizex*(y%sizey+sizey*ind)] = (char)factor*p;
|
85 |
|
86 |
}
|
87 |
|
88 |
barrier(CLK_GLOBAL_MEM_FENCE);
|
89 |
|
90 |
}
|
91 |
|
92 |
__kernel void MainLoopHybrid(__global char *s,__global float *T,float J,float B,
|
93 |
uint sizex,uint sizey,
|
94 |
uint iterations,uint seed_w,uint seed_z)
|
95 |
|
96 |
{
|
97 |
uint z=seed_z/(get_group_id(0)*get_num_groups(0)+get_local_id(0)+1);
|
98 |
uint w=seed_w/(get_group_id(0)*get_num_groups(0)+get_local_id(0)+1);
|
99 |
float t=T[get_group_id(0)*get_num_groups(0)+get_local_id(0)];
|
100 |
uint ind=get_group_id(0)*get_num_groups(0)+get_local_id(0);
|
101 |
|
102 |
for (uint i=0;i<iterations;i++) {
|
103 |
|
104 |
uint x=(uint)(MWC%sizex) ;
|
105 |
uint y=(uint)(MWC%sizey) ;
|
106 |
|
107 |
int p=s[x+sizex*(y+sizey*ind)];
|
108 |
|
109 |
int d=s[x+sizex*((y+1)%sizey+sizey*ind)];
|
110 |
int u=s[x+sizex*((y-1)%sizey+sizey*ind)];
|
111 |
int l=s[((x-1)%sizex)+sizex*(y+sizey*ind)];
|
112 |
int r=s[((x+1)%sizex)+sizex*(y+sizey*ind)];
|
113 |
|
114 |
float DeltaE=2.0f*p*(J*(u+d+l+r)+B);
|
115 |
|
116 |
int factor=((DeltaE < 0.0f) || (MWCfp < exp(-DeltaE/t))) ? -1:1;
|
117 |
s[x%sizex+sizex*(y%sizey+sizey*ind)] = (char)factor*p;
|
118 |
|
119 |
}
|
120 |
|
121 |
barrier(CLK_GLOBAL_MEM_FENCE);
|
122 |
|
123 |
}
|
124 |
|
125 |
__kernel void MainLoopLocal(__global char *s,__global float *T,float J,float B,
|
126 |
uint sizex,uint sizey,
|
127 |
uint iterations,uint seed_w,uint seed_z)
|
128 |
{
|
129 |
uint z=seed_z/(get_local_id(0)+1);
|
130 |
uint w=seed_w/(get_local_id(0)+1);
|
131 |
float t=T[get_local_id(0)];
|
132 |
uint ind=get_local_id(0);
|
133 |
|
134 |
for (uint i=0;i<iterations;i++) {
|
135 |
|
136 |
uint x=(uint)(MWC%sizex) ;
|
137 |
uint y=(uint)(MWC%sizey) ;
|
138 |
|
139 |
int p=s[x+sizex*(y+sizey*ind)];
|
140 |
|
141 |
int d=s[x+sizex*((y+1)%sizey+sizey*ind)];
|
142 |
int u=s[x+sizex*((y-1)%sizey+sizey*ind)];
|
143 |
int l=s[((x-1)%sizex)+sizex*(y+sizey*ind)];
|
144 |
int r=s[((x+1)%sizex)+sizex*(y+sizey*ind)];
|
145 |
|
146 |
float DeltaE=2.0f*p*(J*(u+d+l+r)+B);
|
147 |
|
148 |
int factor=((DeltaE < 0.0f) || (MWCfp < exp(-DeltaE/t))) ? -1:1;
|
149 |
s[x%sizex+sizex*(y%sizey+sizey*ind)] = (char)factor*p;
|
150 |
}
|
151 |
|
152 |
barrier(CLK_LOCAL_MEM_FENCE);
|
153 |
barrier(CLK_GLOBAL_MEM_FENCE);
|
154 |
|
155 |
}
|
156 |
"""
|
157 |
|
158 |
KERNEL_CODE_CUDA="""
|
159 |
|
160 |
// Marsaglia RNG very simple implementation
|
161 |
#define znew ((z=36969*(z&65535)+(z>>16))<<16)
|
162 |
#define wnew ((w=18000*(w&65535)+(w>>16))&65535)
|
163 |
#define MWC (znew+wnew)
|
164 |
#define SHR3 (jsr=(jsr=(jsr=jsr^(jsr<<17))^(jsr>>13))^(jsr<<5))
|
165 |
#define CONG (jcong=69069*jcong+1234567)
|
166 |
#define KISS ((MWC^CONG)+SHR3)
|
167 |
|
168 |
#define MWCfp MWC * 2.328306435454494e-10f
|
169 |
#define KISSfp KISS * 2.328306435454494e-10f
|
170 |
|
171 |
__global__ void MainLoopOne(char *s,float T,float J,float B,
|
172 |
uint sizex,uint sizey,
|
173 |
uint iterations,uint seed_w,uint seed_z)
|
174 |
|
175 |
{
|
176 |
uint z=seed_z;
|
177 |
uint w=seed_w;
|
178 |
|
179 |
for (uint i=0;i<iterations;i++) {
|
180 |
|
181 |
uint x=(uint)(MWC%sizex) ;
|
182 |
uint y=(uint)(MWC%sizey) ;
|
183 |
|
184 |
int p=s[x+sizex*y];
|
185 |
|
186 |
int d=s[x+sizex*((y+1)%sizey)];
|
187 |
int u=s[x+sizex*((y-1)%sizey)];
|
188 |
int l=s[((x-1)%sizex)+sizex*y];
|
189 |
int r=s[((x+1)%sizex)+sizex*y];
|
190 |
|
191 |
float DeltaE=2.0f*p*(J*(u+d+l+r)+B);
|
192 |
|
193 |
int factor=((DeltaE < 0.0f) || (MWCfp < exp(-DeltaE/T))) ? -1:1;
|
194 |
s[x%sizex+sizex*(y%sizey)] = (char)factor*p;
|
195 |
}
|
196 |
__syncthreads();
|
197 |
|
198 |
}
|
199 |
|
200 |
__global__ void MainLoopGlobal(char *s,float *T,float J,float B,
|
201 |
uint sizex,uint sizey,
|
202 |
uint iterations,uint seed_w,uint seed_z)
|
203 |
|
204 |
{
|
205 |
uint z=seed_z/(blockIdx.x+1);
|
206 |
uint w=seed_w/(blockIdx.x+1);
|
207 |
float t=T[blockIdx.x];
|
208 |
uint ind=blockIdx.x;
|
209 |
|
210 |
for (uint i=0;i<iterations;i++) {
|
211 |
|
212 |
uint x=(uint)(MWC%sizex) ;
|
213 |
uint y=(uint)(MWC%sizey) ;
|
214 |
|
215 |
int p=s[x+sizex*(y+sizey*ind)];
|
216 |
|
217 |
int d=s[x+sizex*((y+1)%sizey+sizey*ind)];
|
218 |
int u=s[x+sizex*((y-1)%sizey+sizey*ind)];
|
219 |
int l=s[((x-1)%sizex)+sizex*(y+sizey*ind)];
|
220 |
int r=s[((x+1)%sizex)+sizex*(y+sizey*ind)];
|
221 |
|
222 |
float DeltaE=2.0f*p*(J*(u+d+l+r)+B);
|
223 |
|
224 |
int factor=((DeltaE < 0.0f) || (MWCfp < exp(-DeltaE/t))) ? -1:1;
|
225 |
s[x%sizex+sizex*(y%sizey+sizey*ind)] = (char)factor*p;
|
226 |
|
227 |
}
|
228 |
__syncthreads();
|
229 |
|
230 |
}
|
231 |
|
232 |
__global__ void MainLoopHybrid(char *s,float *T,float J,float B,
|
233 |
uint sizex,uint sizey,
|
234 |
uint iterations,uint seed_w,uint seed_z)
|
235 |
|
236 |
{
|
237 |
uint z=seed_z/(blockDim.x*blockIdx.x+threadIdx.x+1);
|
238 |
uint w=seed_w/(blockDim.x*blockIdx.x+threadIdx.x+1);
|
239 |
float t=T[blockDim.x*blockIdx.x+threadIdx.x];
|
240 |
uint ind=blockDim.x*blockIdx.x+threadIdx.x;
|
241 |
|
242 |
for (uint i=0;i<iterations;i++) {
|
243 |
|
244 |
uint x=(uint)(MWC%sizex) ;
|
245 |
uint y=(uint)(MWC%sizey) ;
|
246 |
|
247 |
int p=s[x+sizex*(y+sizey*ind)];
|
248 |
|
249 |
int d=s[x+sizex*((y+1)%sizey+sizey*ind)];
|
250 |
int u=s[x+sizex*((y-1)%sizey+sizey*ind)];
|
251 |
int l=s[((x-1)%sizex)+sizex*(y+sizey*ind)];
|
252 |
int r=s[((x+1)%sizex)+sizex*(y+sizey*ind)];
|
253 |
|
254 |
float DeltaE=2.0f*p*(J*(u+d+l+r)+B);
|
255 |
|
256 |
int factor=((DeltaE < 0.0f) || (MWCfp < exp(-DeltaE/t))) ? -1:1;
|
257 |
s[x%sizex+sizex*(y%sizey+sizey*ind)] = (char)factor*p;
|
258 |
|
259 |
}
|
260 |
__syncthreads();
|
261 |
|
262 |
}
|
263 |
|
264 |
__global__ void MainLoopLocal(char *s,float *T,float J,float B,
|
265 |
uint sizex,uint sizey,
|
266 |
uint iterations,uint seed_w,uint seed_z)
|
267 |
{
|
268 |
uint z=seed_z/(threadIdx.x+1);
|
269 |
uint w=seed_w/(threadIdx.x+1);
|
270 |
float t=T[threadIdx.x];
|
271 |
uint ind=threadIdx.x;
|
272 |
|
273 |
for (uint i=0;i<iterations;i++) {
|
274 |
|
275 |
uint x=(uint)(MWC%sizex) ;
|
276 |
uint y=(uint)(MWC%sizey) ;
|
277 |
|
278 |
int p=s[x+sizex*(y+sizey*ind)];
|
279 |
|
280 |
int d=s[x+sizex*((y+1)%sizey+sizey*ind)];
|
281 |
int u=s[x+sizex*((y-1)%sizey+sizey*ind)];
|
282 |
int l=s[((x-1)%sizex)+sizex*(y+sizey*ind)];
|
283 |
int r=s[((x+1)%sizex)+sizex*(y+sizey*ind)];
|
284 |
|
285 |
float DeltaE=2.0f*p*(J*(u+d+l+r)+B);
|
286 |
|
287 |
int factor=((DeltaE < 0.0f) || (MWCfp < exp(-DeltaE/t))) ? -1:1;
|
288 |
s[x%sizex+sizex*(y%sizey+sizey*ind)] = (char)factor*p;
|
289 |
}
|
290 |
__syncthreads();
|
291 |
|
292 |
}
|
293 |
"""
|
294 |
|
295 |
# find prime factors of a number
|
296 |
# Get for WWW :
|
297 |
# http://pythonism.wordpress.com/2008/05/17/looking-at-factorisation-in-python/
|
298 |
def PrimeFactors(x): |
299 |
factorlist=numpy.array([]).astype('uint32')
|
300 |
loop=2
|
301 |
while loop<=x:
|
302 |
if x%loop==0: |
303 |
x/=loop |
304 |
factorlist=numpy.append(factorlist,[loop]) |
305 |
else:
|
306 |
loop+=1
|
307 |
return factorlist
|
308 |
|
309 |
# Try to find the best thread number in Hybrid approach (Blocks&Threads)
|
310 |
# output is thread number
|
311 |
def BestThreadsNumber(jobs): |
312 |
factors=PrimeFactors(jobs) |
313 |
matrix=numpy.append([factors],[factors[::-1]],axis=0) |
314 |
threads=1
|
315 |
for factor in matrix.transpose().ravel(): |
316 |
threads=threads*factor |
317 |
if threads*threads>jobs:
|
318 |
break
|
319 |
return(long(threads)) |
320 |
|
321 |
def ImageOutput(sigma,prefix): |
322 |
Max=sigma.max() |
323 |
Min=sigma.min() |
324 |
|
325 |
# Normalize value as 8bits Integer
|
326 |
SigmaInt=(255*(sigma-Min)/(Max-Min)).astype('uint8') |
327 |
image = Image.fromarray(SigmaInt) |
328 |
image.save("%s.jpg" % prefix)
|
329 |
|
330 |
def Metropolis(sigma,T,J,B,iterations): |
331 |
start=time.time() |
332 |
|
333 |
SizeX,SizeY=sigma.shape |
334 |
|
335 |
for p in xrange(0,iterations): |
336 |
# Random access coordonate
|
337 |
X,Y=numpy.random.randint(SizeX),numpy.random.randint(SizeY) |
338 |
|
339 |
DeltaE=J*sigma[X,Y]*(2*(sigma[X,(Y+1)%SizeY]+ |
340 |
sigma[X,(Y-1)%SizeY]+
|
341 |
sigma[(X-1)%SizeX,Y]+
|
342 |
sigma[(X+1)%SizeX,Y])+B)
|
343 |
|
344 |
if DeltaE < 0. or random() < exp(-DeltaE/T): |
345 |
sigma[X,Y]=-sigma[X,Y] |
346 |
duration=time.time()-start |
347 |
return(duration)
|
348 |
|
349 |
def MetropolisAllOpenCL(sigmaDict,TList,J,B,iterations,jobs,ParaStyle,Alu,Device): |
350 |
|
351 |
# sigmaDict & Tlist are NOT respectively array & float
|
352 |
# sigmaDict : dict of array for each temperatoire
|
353 |
# TList : list of temperatures
|
354 |
|
355 |
# Initialisation des variables en les CASTant correctement
|
356 |
|
357 |
# Je detecte un peripherique GPU dans la liste des peripheriques
|
358 |
|
359 |
HasGPU=False
|
360 |
Id=1
|
361 |
# Primary Device selection based on Device Id
|
362 |
for platform in cl.get_platforms(): |
363 |
for device in platform.get_devices(): |
364 |
deviceType=cl.device_type.to_string(device.type) |
365 |
if Id==Device and not HasGPU: |
366 |
GPU=device |
367 |
print "CPU/GPU selected: ",device.name |
368 |
HasGPU=True
|
369 |
Id=Id+1
|
370 |
# Default Device selection based on ALU Type
|
371 |
for platform in cl.get_platforms(): |
372 |
for device in platform.get_devices(): |
373 |
deviceType=cl.device_type.to_string(device.type) |
374 |
if deviceType=="GPU" and Alu=="GPU" and not HasGPU: |
375 |
GPU=device |
376 |
print "GPU selected: ",device.name |
377 |
HasGPU=True
|
378 |
if deviceType=="CPU" and Alu=="CPU" and not HasGPU: |
379 |
GPU=device |
380 |
print "CPU selected: ",device.name |
381 |
HasGPU=True
|
382 |
|
383 |
# Je cree le contexte et la queue pour son execution
|
384 |
# ctx = cl.create_some_context()
|
385 |
ctx = cl.Context([GPU]) |
386 |
queue = cl.CommandQueue(ctx, |
387 |
properties=cl.command_queue_properties.PROFILING_ENABLE) |
388 |
|
389 |
# Je recupere les flag possibles pour les buffers
|
390 |
mf = cl.mem_flags |
391 |
|
392 |
# Concatenate all sigma in single array
|
393 |
sigma=numpy.copy(sigmaDict[TList[0]])
|
394 |
for T in TList[1:]: |
395 |
sigma=numpy.concatenate((sigma,sigmaDict[T]),axis=1)
|
396 |
|
397 |
sigmaCL = cl.Buffer(ctx, mf.WRITE_ONLY|mf.COPY_HOST_PTR,hostbuf=sigma) |
398 |
TCL = cl.Buffer(ctx, mf.WRITE_ONLY|mf.COPY_HOST_PTR,hostbuf=TList) |
399 |
|
400 |
MetropolisCL = cl.Program(ctx,KERNEL_CODE_OPENCL).build( \ |
401 |
options = "-cl-mad-enable -cl-fast-relaxed-math")
|
402 |
|
403 |
SizeX,SizeY=sigmaDict[TList[0]].shape
|
404 |
|
405 |
if ParaStyle=='Blocks': |
406 |
# Call OpenCL kernel
|
407 |
# (1,) is Global work size (only 1 work size)
|
408 |
# (1,) is local work size
|
409 |
# SeedZCL is lattice translated in CL format
|
410 |
# SeedWCL is lattice translated in CL format
|
411 |
# step is number of iterations
|
412 |
CLLaunch=MetropolisCL.MainLoopGlobal(queue,(jobs,),None,
|
413 |
sigmaCL, |
414 |
TCL, |
415 |
numpy.float32(J), |
416 |
numpy.float32(B), |
417 |
numpy.uint32(SizeX), |
418 |
numpy.uint32(SizeY), |
419 |
numpy.uint32(iterations), |
420 |
numpy.uint32(nprnd(2**31-1)), |
421 |
numpy.uint32(nprnd(2**31-1))) |
422 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \ |
423 |
(Alu,jobs,1,ParaStyle)
|
424 |
elif ParaStyle=='Threads': |
425 |
# It's necessary to put a Local_ID equal to a Global_ID
|
426 |
# Jobs are to be considerated as global number of jobs to do
|
427 |
# And to be distributed to entities
|
428 |
# For example :
|
429 |
# G_ID=10 & L_ID=10 : 10 Threads on 1 UC
|
430 |
# G_ID=10 & L_ID=1 : 10 Threads on 1 UC
|
431 |
|
432 |
CLLaunch=MetropolisCL.MainLoopLocal(queue,(jobs,),(jobs,), |
433 |
sigmaCL, |
434 |
TCL, |
435 |
numpy.float32(J), |
436 |
numpy.float32(B), |
437 |
numpy.uint32(SizeX), |
438 |
numpy.uint32(SizeY), |
439 |
numpy.uint32(iterations), |
440 |
numpy.uint32(nprnd(2**31-1)), |
441 |
numpy.uint32(nprnd(2**31-1))) |
442 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \ |
443 |
(Alu,1,jobs,ParaStyle)
|
444 |
else:
|
445 |
threads=BestThreadsNumber(jobs) |
446 |
# en OpenCL, necessaire de mettre un Global_id identique au local_id
|
447 |
CLLaunch=MetropolisCL.MainLoopHybrid(queue,(jobs,),(threads,), |
448 |
sigmaCL, |
449 |
TCL, |
450 |
numpy.float32(J), |
451 |
numpy.float32(B), |
452 |
numpy.uint32(SizeX), |
453 |
numpy.uint32(SizeY), |
454 |
numpy.uint32(iterations), |
455 |
numpy.uint32(nprnd(2**31-1)), |
456 |
numpy.uint32(nprnd(2**31-1))) |
457 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \ |
458 |
(Alu,jobs/threads,threads,ParaStyle) |
459 |
|
460 |
CLLaunch.wait() |
461 |
cl.enqueue_copy(queue, sigma, sigmaCL).wait() |
462 |
elapsed = 1e-9*(CLLaunch.profile.end - CLLaunch.profile.start)
|
463 |
sigmaCL.release() |
464 |
|
465 |
results=numpy.split(sigma,len(TList),axis=1) |
466 |
for T in TList: |
467 |
sigmaDict[T]=numpy.copy(results[numpy.nonzero(TList == T)[0][0]]) |
468 |
|
469 |
return(elapsed)
|
470 |
|
471 |
def MetropolisAllCuda(sigmaDict,TList,J,B,iterations,jobs,ParaStyle,Alu,Device): |
472 |
|
473 |
# sigmaDict & Tlist are NOT respectively array & float
|
474 |
# sigmaDict : dict of array for each temperatoire
|
475 |
# TList : list of temperatures
|
476 |
|
477 |
# Avec PyCUDA autoinit, rien a faire !
|
478 |
|
479 |
mod = SourceModule(KERNEL_CODE_CUDA) |
480 |
|
481 |
MetropolisBlocksCuda=mod.get_function("MainLoopGlobal")
|
482 |
MetropolisThreadsCuda=mod.get_function("MainLoopLocal")
|
483 |
MetropolisHybridCuda=mod.get_function("MainLoopHybrid")
|
484 |
|
485 |
# Concatenate all sigma in single array
|
486 |
sigma=numpy.copy(sigmaDict[TList[0]])
|
487 |
for T in TList[1:]: |
488 |
sigma=numpy.concatenate((sigma,sigmaDict[T]),axis=1)
|
489 |
|
490 |
sigmaCU=cuda.InOut(sigma) |
491 |
TCU=cuda.InOut(TList) |
492 |
|
493 |
SizeX,SizeY=sigmaDict[TList[0]].shape
|
494 |
|
495 |
start = pycuda.driver.Event() |
496 |
stop = pycuda.driver.Event() |
497 |
|
498 |
start.record() |
499 |
start.synchronize() |
500 |
if ParaStyle=='Blocks': |
501 |
# Call CUDA kernel
|
502 |
# (1,) is Global work size (only 1 work size)
|
503 |
# (1,) is local work size
|
504 |
# SeedZCL is lattice translated in CL format
|
505 |
# SeedWCL is lattice translated in CL format
|
506 |
# step is number of iterations
|
507 |
MetropolisBlocksCuda(sigmaCU, |
508 |
TCU, |
509 |
numpy.float32(J), |
510 |
numpy.float32(B), |
511 |
numpy.uint32(SizeX), |
512 |
numpy.uint32(SizeY), |
513 |
numpy.uint32(iterations), |
514 |
numpy.uint32(nprnd(2**31-1)), |
515 |
numpy.uint32(nprnd(2**31-1)), |
516 |
grid=(jobs,1),block=(1,1,1)) |
517 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \ |
518 |
(Alu,jobs,1,ParaStyle)
|
519 |
elif ParaStyle=='Threads': |
520 |
MetropolisThreadsCuda(sigmaCU, |
521 |
TCU, |
522 |
numpy.float32(J), |
523 |
numpy.float32(B), |
524 |
numpy.uint32(SizeX), |
525 |
numpy.uint32(SizeY), |
526 |
numpy.uint32(iterations), |
527 |
numpy.uint32(nprnd(2**31-1)), |
528 |
numpy.uint32(nprnd(2**31-1)), |
529 |
grid=(1,1),block=(jobs,1,1)) |
530 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \ |
531 |
(Alu,1,jobs,ParaStyle)
|
532 |
else:
|
533 |
threads=BestThreadsNumber(jobs) |
534 |
MetropolisHybridCuda(sigmaCU, |
535 |
TCU, |
536 |
numpy.float32(J), |
537 |
numpy.float32(B), |
538 |
numpy.uint32(SizeX), |
539 |
numpy.uint32(SizeY), |
540 |
numpy.uint32(iterations), |
541 |
numpy.uint32(nprnd(2**31-1)), |
542 |
numpy.uint32(nprnd(2**31-1)), |
543 |
grid=(jobs/threads,1),block=(threads,1,1)) |
544 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \ |
545 |
(Alu,jobs/threads,threads,ParaStyle) |
546 |
|
547 |
stop.record() |
548 |
stop.synchronize() |
549 |
elapsed = start.time_till(stop)*1e-3
|
550 |
|
551 |
results=numpy.split(sigma,len(TList),axis=1) |
552 |
for T in TList: |
553 |
sigmaDict[T]=numpy.copy(results[numpy.nonzero(TList == T)[0][0]]) |
554 |
|
555 |
return(elapsed)
|
556 |
|
557 |
|
558 |
def Magnetization(sigma,M): |
559 |
return(numpy.sum(sigma)/(sigma.shape[0]*sigma.shape[1]*1.0)) |
560 |
|
561 |
def Energy(sigma,J): |
562 |
# Copier et caster
|
563 |
E=numpy.copy(sigma).astype(numpy.float32) |
564 |
|
565 |
# Appel par slice
|
566 |
E[1:-1,1:-1]=-J*E[1:-1,1:-1]*(E[:-2,1:-1]+E[2:,1:-1]+ |
567 |
E[1:-1,:-2]+E[1:-1,2:]) |
568 |
|
569 |
# Bien nettoyer la peripherie
|
570 |
E[:,0]=0 |
571 |
E[:,-1]=0 |
572 |
E[0,:]=0 |
573 |
E[-1,:]=0 |
574 |
|
575 |
Energy=numpy.sum(E) |
576 |
|
577 |
return(Energy/(E.shape[0]*E.shape[1]*1.0)) |
578 |
|
579 |
def DisplayCurves(T,E,M,J,B): |
580 |
|
581 |
plt.xlabel("Temperature")
|
582 |
plt.ylabel("Energy")
|
583 |
|
584 |
Experience,=plt.plot(T,E,label="Energy")
|
585 |
|
586 |
plt.legend() |
587 |
plt.show() |
588 |
|
589 |
|
590 |
if __name__=='__main__': |
591 |
|
592 |
# Set defaults values
|
593 |
# Alu can be CPU or GPU
|
594 |
Alu='CPU'
|
595 |
# Id of GPU : 0
|
596 |
Device=0
|
597 |
# GPU style can be Cuda (Nvidia implementation) or OpenCL
|
598 |
GpuStyle='OpenCL'
|
599 |
# Parallel distribution can be on Threads or Blocks
|
600 |
ParaStyle='Blocks'
|
601 |
# Coupling factor
|
602 |
J=1.
|
603 |
# Magnetic Field
|
604 |
B=0.
|
605 |
# Size of Lattice
|
606 |
Size=256
|
607 |
# Default Temperatures (start, end, step)
|
608 |
Tmin=0.1
|
609 |
Tmax=5
|
610 |
Tstep=0.1
|
611 |
# Default Number of Iterations
|
612 |
Iterations=Size*Size |
613 |
# Curves is True to print the curves
|
614 |
Curves=False
|
615 |
|
616 |
try:
|
617 |
opts, args = getopt.getopt(sys.argv[1:],"hcj:b:z:i:s:e:p:a:d:g:t:",["coupling=","magneticfield=","size=","iterations=","tempstart=","tempend=","tempstep=","alu=","gpustyle=","parastyle="]) |
618 |
except getopt.GetoptError:
|
619 |
print '%s -j <Coupling Factor> -b <Magnetic Field> -z <Size of Lattice> -i <Iterations> -s <Minimum Temperature> -e <Maximum Temperature> -p <steP Temperature> -c (Print Curves) -a <CPU/GPU> -d <DeviceId> -g <CUDA/OpenCL> -t <Threads/Blocks>' % sys.argv[0] |
620 |
sys.exit(2)
|
621 |
|
622 |
|
623 |
for opt, arg in opts: |
624 |
if opt == '-h': |
625 |
print '%s -j <Coupling Factor> -b <Magnetic Field> -z <Size of Lattice> -i <Iterations> -s <Minimum Temperature> -e <Maximum Temperature> -p <steP Temperature> -c (Print Curves) -a <CPU/GPU> -d <DeviceId> -g <CUDA/OpenCL> -t <Threads/Blocks/Hybrid>' % sys.argv[0] |
626 |
sys.exit() |
627 |
elif opt == '-c': |
628 |
Curves=True
|
629 |
elif opt in ("-j", "--coupling"): |
630 |
J = float(arg)
|
631 |
elif opt in ("-b", "--magneticfield"): |
632 |
B = float(arg)
|
633 |
elif opt in ("-s", "--tempmin"): |
634 |
Tmin = float(arg)
|
635 |
elif opt in ("-e", "--tempmax"): |
636 |
Tmax = float(arg)
|
637 |
elif opt in ("-p", "--tempstep"): |
638 |
Tstep = float(arg)
|
639 |
elif opt in ("-i", "--iterations"): |
640 |
Iterations = int(arg)
|
641 |
elif opt in ("-z", "--size"): |
642 |
Size = int(arg)
|
643 |
elif opt in ("-a", "--alu"): |
644 |
Alu = arg |
645 |
elif opt in ("-d", "--device"): |
646 |
Device = int(arg)
|
647 |
elif opt in ("-g", "--gpustyle"): |
648 |
GpuStyle = arg |
649 |
elif opt in ("-t", "--parastyle"): |
650 |
ParaStyle = arg |
651 |
|
652 |
if Alu=='CPU' and GpuStyle=='CUDA': |
653 |
print "Alu can't be CPU for CUDA, set Alu to GPU" |
654 |
Alu='GPU'
|
655 |
|
656 |
if ParaStyle not in ('Blocks','Threads','Hybrid'): |
657 |
print "%s not exists, ParaStyle set as Threads !" % ParaStyle |
658 |
ParaStyle='Blocks'
|
659 |
|
660 |
print "Compute unit : %s" % Alu |
661 |
print "Device Identification : %s" % Device |
662 |
print "GpuStyle used : %s" % GpuStyle |
663 |
print "Parallel Style used : %s" % ParaStyle |
664 |
print "Coupling Factor : %s" % J |
665 |
print "Magnetic Field : %s" % B |
666 |
print "Size of lattice : %s" % Size |
667 |
print "Iterations : %s" % Iterations |
668 |
print "Temperature on start : %s" % Tmin |
669 |
print "Temperature on end : %s" % Tmax |
670 |
print "Temperature step : %s" % Tstep |
671 |
|
672 |
if GpuStyle=='CUDA': |
673 |
# For PyCUDA import
|
674 |
import pycuda.driver as cuda |
675 |
import pycuda.gpuarray as gpuarray |
676 |
import pycuda.autoinit |
677 |
from pycuda.compiler import SourceModule |
678 |
|
679 |
if GpuStyle=='OpenCL': |
680 |
# For PyOpenCL import
|
681 |
import pyopencl as cl |
682 |
Id=1
|
683 |
for platform in cl.get_platforms(): |
684 |
for device in platform.get_devices(): |
685 |
deviceType=cl.device_type.to_string(device.type) |
686 |
print "Device #%i of type %s : %s" % (Id,deviceType,device.name) |
687 |
Id=Id+1
|
688 |
|
689 |
LAPIMAGE=False
|
690 |
|
691 |
sigmaIn=numpy.where(numpy.random.randn(Size,Size)>0,1,-1).astype(numpy.int8) |
692 |
|
693 |
ImageOutput(sigmaIn,"Ising2D_Serial_%i_Initial" % (Size))
|
694 |
|
695 |
# La temperature est passee comme parametre, attention au CAST !
|
696 |
Trange=numpy.arange(Tmin,Tmax+Tstep,Tstep).astype(numpy.float32) |
697 |
|
698 |
E=[] |
699 |
M=[] |
700 |
|
701 |
sigma={} |
702 |
for T in Trange: |
703 |
sigma[T]=numpy.copy(sigmaIn) |
704 |
|
705 |
jobs=len(Trange)
|
706 |
|
707 |
# For GPU, all process are launched
|
708 |
if GpuStyle=='CUDA': |
709 |
duration=MetropolisAllCuda(sigma,Trange,J,B,Iterations,jobs,ParaStyle,Alu,Device) |
710 |
else:
|
711 |
duration=MetropolisAllOpenCL(sigma,Trange,J,B,Iterations,jobs,ParaStyle,Alu,Device) |
712 |
|
713 |
print BestThreadsNumber(len(Trange)) |
714 |
|
715 |
for T in Trange: |
716 |
E=numpy.append(E,Energy(sigma[T],J)) |
717 |
M=numpy.append(M,Magnetization(sigma[T],B)) |
718 |
print "CPU Time for each : %f" % (duration/len(Trange)) |
719 |
print "Total Energy at Temperature %f : %f" % (T,E[-1]) |
720 |
print "Total Magnetization at Temperature %f : %f" % (T,M[-1]) |
721 |
ImageOutput(sigma[T],"Ising2D_Serial_%i_%1.1f_Final" % (Size,T))
|
722 |
|
723 |
if Curves:
|
724 |
DisplayCurves(Trange,E,M,J,B) |
725 |
|
726 |
|