root / Pi / GPU / Pi-GPU.py-20130213-2 @ 124
Historique | Voir | Annoter | Télécharger (18,28 ko)
1 |
#!/usr/bin/env python |
---|---|
2 |
|
3 |
# |
4 |
# Pi-by-MC using PyCUDA/PyOpenCL |
5 |
# |
6 |
# CC BY-NC-SA 2011 : <emmanuel.quemener@ens-lyon.fr> |
7 |
# |
8 |
# Thanks to Andreas Klockner for PyCUDA: |
9 |
# http://mathema.tician.de/software/pycuda |
10 |
# |
11 |
|
12 |
# 2013-01-01 : problems with launch timeout |
13 |
# http://stackoverflow.com/questions/497685/how-do-you-get-around-the-maximum-cuda-run-time |
14 |
# Option "Interactive" "0" in /etc/X11/xorg.conf |
15 |
|
16 |
# Common tools |
17 |
import numpy |
18 |
from numpy.random import randint as nprnd |
19 |
import sys |
20 |
import getopt |
21 |
import time |
22 |
import matplotlib.pyplot as plt |
23 |
import math |
24 |
from scipy.optimize import curve_fit |
25 |
from socket import gethostname |
26 |
|
27 |
# Predicted Amdahl Law (Reduced with s=1-p) |
28 |
def AmdahlR(N, T1, p): |
29 |
return (T1*(1-p+p/N)) |
30 |
|
31 |
# Predicted Amdahl Law |
32 |
def Amdahl(N, T1, s, p): |
33 |
return (T1*(s+p/N)) |
34 |
|
35 |
# Predicted Mylq Law with first order |
36 |
def Mylq(N, T1,s,c,p): |
37 |
return (T1*(s+c*N+p/N)) |
38 |
|
39 |
# Predicted Mylq Law with second order |
40 |
def Mylq2(N, T1,s,c1,c2,p): |
41 |
return (T1*(s+c1*N+c2*N*N+p/N)) |
42 |
|
43 |
KERNEL_CODE_CUDA=""" |
44 |
|
45 |
// Marsaglia RNG very simple implementation |
46 |
|
47 |
#define znew ((z=36969*(z&65535)+(z>>16))<<16) |
48 |
#define wnew ((w=18000*(w&65535)+(w>>16))&65535) |
49 |
#define MWC (znew+wnew) |
50 |
#define SHR3 (jsr=(jsr=(jsr=jsr^(jsr<<17))^(jsr>>13))^(jsr<<5)) |
51 |
#define CONG (jcong=69069*jcong+1234567) |
52 |
#define KISS ((MWC^CONG)+SHR3) |
53 |
|
54 |
#define MWCfp MWC * 2.328306435454494e-10f |
55 |
#define KISSfp KISS * 2.328306435454494e-10f |
56 |
|
57 |
__global__ void MainLoopBlocks(uint *s,uint iterations,uint seed_w,uint seed_z) |
58 |
{ |
59 |
uint z=seed_z/(blockIdx.x+1); |
60 |
uint w=seed_w/(blockIdx.x+1); |
61 |
|
62 |
int total=0; |
63 |
|
64 |
for (uint i=0;i<iterations;i++) { |
65 |
|
66 |
float x=MWCfp ; |
67 |
float y=MWCfp ; |
68 |
|
69 |
// Matching test |
70 |
int inside=((x*x+y*y) < 1.0f) ? 1:0; |
71 |
total+=inside; |
72 |
|
73 |
} |
74 |
|
75 |
s[blockIdx.x]=total; |
76 |
__syncthreads(); |
77 |
|
78 |
} |
79 |
|
80 |
__global__ void MainLoopThreads(uint *s,uint iterations,uint seed_w,uint seed_z) |
81 |
{ |
82 |
uint z=seed_z/(threadIdx.x+1); |
83 |
uint w=seed_w/(threadIdx.x+1); |
84 |
|
85 |
int total=0; |
86 |
|
87 |
for (uint i=0;i<iterations;i++) { |
88 |
|
89 |
float x=MWCfp ; |
90 |
float y=MWCfp ; |
91 |
|
92 |
// Matching test |
93 |
int inside=((x*x+y*y) < 1.0f) ? 1:0; |
94 |
total+=inside; |
95 |
|
96 |
} |
97 |
|
98 |
s[threadIdx.x]=total; |
99 |
__syncthreads(); |
100 |
|
101 |
} |
102 |
|
103 |
__global__ void MainLoopHybrid(uint *s,uint iterations,uint seed_w,uint seed_z) |
104 |
{ |
105 |
uint z=seed_z/(blockDim.x*blockIdx.x+threadIdx.x+1); |
106 |
uint w=seed_w/(blockDim.x*blockIdx.x+threadIdx.x+1); |
107 |
|
108 |
int total=0; |
109 |
|
110 |
for (uint i=0;i<iterations;i++) { |
111 |
|
112 |
float x=MWCfp ; |
113 |
float y=MWCfp ; |
114 |
|
115 |
// Matching test |
116 |
int inside=((x*x+y*y) < 1.0f) ? 1:0; |
117 |
total+=inside; |
118 |
|
119 |
} |
120 |
|
121 |
s[blockDim.x*blockIdx.x+threadIdx.x]=total; |
122 |
__syncthreads(); |
123 |
|
124 |
} |
125 |
""" |
126 |
|
127 |
KERNEL_CODE_OPENCL=""" |
128 |
|
129 |
// Marsaglia RNG very simple implementation |
130 |
#define znew ((z=36969*(z&65535)+(z>>16))<<16) |
131 |
#define wnew ((w=18000*(w&65535)+(w>>16))&65535) |
132 |
#define MWC (znew+wnew) |
133 |
#define SHR3 (jsr=(jsr=(jsr=jsr^(jsr<<17))^(jsr>>13))^(jsr<<5)) |
134 |
#define CONG (jcong=69069*jcong+1234567) |
135 |
#define KISS ((MWC^CONG)+SHR3) |
136 |
|
137 |
#define MWCfp MWC * 2.328306435454494e-10f |
138 |
#define KISSfp KISS * 2.328306435454494e-10f |
139 |
|
140 |
__kernel void MainLoopGlobal(__global uint *s,uint iterations,uint seed_w,uint seed_z) |
141 |
{ |
142 |
uint z=seed_z/(get_global_id(0)+1); |
143 |
uint w=seed_w/(get_global_id(0)+1); |
144 |
|
145 |
int total=0; |
146 |
|
147 |
for (uint i=0;i<iterations;i++) { |
148 |
|
149 |
float x=MWCfp ; |
150 |
float y=MWCfp ; |
151 |
|
152 |
// Matching test |
153 |
int inside=((x*x+y*y) < 1.0f) ? 1:0; |
154 |
total+=inside; |
155 |
} |
156 |
s[get_global_id(0)]=total; |
157 |
barrier(CLK_GLOBAL_MEM_FENCE); |
158 |
|
159 |
} |
160 |
|
161 |
__kernel void MainLoopLocal(__global uint *s,uint iterations,uint seed_w,uint seed_z) |
162 |
{ |
163 |
uint z=seed_z/(get_local_id(0)+1); |
164 |
uint w=seed_w/(get_local_id(0)+1); |
165 |
|
166 |
int total=0; |
167 |
|
168 |
for (uint i=0;i<iterations;i++) { |
169 |
|
170 |
float x=MWCfp ; |
171 |
float y=MWCfp ; |
172 |
|
173 |
// Matching test |
174 |
int inside=((x*x+y*y) < 1.0f) ? 1:0; |
175 |
total+=inside; |
176 |
} |
177 |
s[get_local_id(0)]=total; |
178 |
barrier(CLK_LOCAL_MEM_FENCE); |
179 |
|
180 |
} |
181 |
|
182 |
__kernel void MainLoopHybrid(__global uint *s,uint iterations,uint seed_w,uint seed_z) |
183 |
{ |
184 |
uint z=seed_z/(get_group_id(0)*get_num_groups(0)+get_local_id(0)+1); |
185 |
uint w=seed_w/(get_group_id(0)*get_num_groups(0)+get_local_id(0)+1); |
186 |
|
187 |
// uint jsr=123456789; |
188 |
// uint jcong=380116160; |
189 |
|
190 |
int total=0; |
191 |
|
192 |
for (uint i=0;i<iterations;i++) { |
193 |
|
194 |
float x=MWCfp ; |
195 |
float y=MWCfp ; |
196 |
|
197 |
// Matching test |
198 |
int inside=((x*x+y*y) < 1.0f) ? 1:0; |
199 |
total+=inside; |
200 |
} |
201 |
barrier(CLK_LOCAL_MEM_FENCE); |
202 |
s[get_group_id(0)*get_num_groups(0)+get_local_id(0)]=total; |
203 |
|
204 |
} |
205 |
""" |
206 |
|
207 |
def MetropolisCuda(circle,iterations,steps,jobs,ParaStyle): |
208 |
|
209 |
# Avec PyCUDA autoinit, rien a faire ! |
210 |
|
211 |
circleCU = cuda.InOut(circle) |
212 |
|
213 |
mod = SourceModule(KERNEL_CODE_CUDA) |
214 |
|
215 |
MetropolisBlocksCU=mod.get_function("MainLoopBlocks") |
216 |
MetropolisJobsCU=mod.get_function("MainLoopThreads") |
217 |
MetropolisHybridCU=mod.get_function("MainLoopHybrid") |
218 |
|
219 |
start = pycuda.driver.Event() |
220 |
stop = pycuda.driver.Event() |
221 |
|
222 |
MyPi=numpy.zeros(steps) |
223 |
MyDuration=numpy.zeros(steps) |
224 |
|
225 |
if iterations%jobs==0: |
226 |
iterationsCL=numpy.uint32(iterations/jobs+1) |
227 |
iterationsNew=iterationsCL*jobs |
228 |
else: |
229 |
iterationsCL=numpy.uint32(iterations/jobs) |
230 |
iterationsNew=iterations |
231 |
|
232 |
for i in range(steps): |
233 |
start.record() |
234 |
start.synchronize() |
235 |
if ParaStyle=='Blocks': |
236 |
MetropolisBlocksCU(circleCU, |
237 |
numpy.uint32(iterationsCL), |
238 |
numpy.uint32(nprnd(2**32/jobs)), |
239 |
numpy.uint32(nprnd(2**32/jobs)), |
240 |
grid=(jobs,1), |
241 |
block=(1,1,1)) |
242 |
print "GPU with %i %s done" % (jobs,ParaStyle) |
243 |
elif ParaStyle=='Hybrid': |
244 |
blocks=jobs/int(math.sqrt(float(jobs))) |
245 |
MetropolisHybridCU(circleCU, |
246 |
numpy.uint32(iterationsCL), |
247 |
numpy.uint32(nprnd(2**32/jobs)), |
248 |
numpy.uint32(nprnd(2**32/jobs)), |
249 |
grid=(blocks,1), |
250 |
block=(jobs/blocks,1,1)) |
251 |
print "GPU with (blocks,jobs)=(%i,%i) %s done" % (blocks,jobs/blocks,ParaStyle) |
252 |
else: |
253 |
MetropolisJobsCU(circleCU, |
254 |
numpy.uint32(iterationsCL), |
255 |
numpy.uint32(nprnd(2**32/jobs)), |
256 |
numpy.uint32(nprnd(2**32/jobs)), |
257 |
grid=(1,1), |
258 |
block=(jobs,1,1)) |
259 |
print "GPU with %i %s done" % (jobs,ParaStyle) |
260 |
stop.record() |
261 |
stop.synchronize() |
262 |
|
263 |
#elapsed = stop.time_since(start)*1e-3 |
264 |
elapsed = start.time_till(stop)*1e-3 |
265 |
|
266 |
#print circle,float(numpy.sum(circle)) |
267 |
MyPi[i]=4.*float(numpy.sum(circle))/float(iterationsCL) |
268 |
MyDuration[i]=elapsed |
269 |
#print MyPi[i],MyDuration[i] |
270 |
#time.sleep(1) |
271 |
|
272 |
print jobs,numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration) |
273 |
|
274 |
return(numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration)) |
275 |
|
276 |
|
277 |
def MetropolisOpenCL(circle,iterations,steps,jobs,ParaStyle,Alu,Device): |
278 |
|
279 |
# Initialisation des variables en les CASTant correctement |
280 |
|
281 |
# Je detecte un peripherique GPU dans la liste des peripheriques |
282 |
# for platform in cl.get_platforms(): |
283 |
# for device in platform.get_devices(): |
284 |
# if cl.device_type.to_string(device.type)=='GPU': |
285 |
# GPU=device |
286 |
#print "GPU detected: ",device.name |
287 |
|
288 |
HasGPU=False |
289 |
Id=1 |
290 |
# Device selection based on choice (default is GPU) |
291 |
for platform in cl.get_platforms(): |
292 |
for device in platform.get_devices(): |
293 |
if not HasGPU: |
294 |
deviceType=cl.device_type.to_string(device.type) |
295 |
if deviceType=="GPU" and Alu=="GPU" and Id==Device: |
296 |
GPU=device |
297 |
print "GPU selected: ",device.name |
298 |
HasGPU=True |
299 |
if deviceType=="CPU" and Alu=="CPU": |
300 |
GPU=device |
301 |
print "CPU selected: ",device.name |
302 |
HasGPU=True |
303 |
Id=Id+1 |
304 |
|
305 |
# Je cree le contexte et la queue pour son execution |
306 |
#ctx = cl.create_some_context() |
307 |
ctx = cl.Context([GPU]) |
308 |
queue = cl.CommandQueue(ctx, |
309 |
properties=cl.command_queue_properties.PROFILING_ENABLE) |
310 |
|
311 |
# Je recupere les flag possibles pour les buffers |
312 |
mf = cl.mem_flags |
313 |
|
314 |
circleCL = cl.Buffer(ctx, mf.WRITE_ONLY|mf.COPY_HOST_PTR,hostbuf=circle) |
315 |
|
316 |
MetropolisCL = cl.Program(ctx,KERNEL_CODE_OPENCL).build( \ |
317 |
options = "-cl-mad-enable -cl-fast-relaxed-math") |
318 |
|
319 |
#MetropolisCL = cl.Program(ctx,KERNEL_CODE_OPENCL).build() |
320 |
|
321 |
i=0 |
322 |
|
323 |
MyPi=numpy.zeros(steps) |
324 |
MyDuration=numpy.zeros(steps) |
325 |
|
326 |
if iterations%jobs==0: |
327 |
iterationsCL=numpy.uint32(iterations/jobs+1) |
328 |
iterationsNew=iterationsCL*jobs |
329 |
else: |
330 |
iterationsCL=numpy.uint32(iterations/jobs) |
331 |
iterationsNew=iterations |
332 |
|
333 |
blocks=int(math.sqrt(jobs)) |
334 |
|
335 |
for i in range(steps): |
336 |
|
337 |
if ParaStyle=='Blocks': |
338 |
# Call OpenCL kernel |
339 |
# (1,) is Global work size (only 1 work size) |
340 |
# (1,) is local work size |
341 |
# circleCL is lattice translated in CL format |
342 |
# SeedZCL is lattice translated in CL format |
343 |
# SeedWCL is lattice translated in CL format |
344 |
# step is number of iterations |
345 |
CLLaunch=MetropolisCL.MainLoopGlobal(queue,(jobs,),None, |
346 |
circleCL, |
347 |
numpy.uint32(iterationsCL), |
348 |
numpy.uint32(nprnd(2**32/jobs)), |
349 |
numpy.uint32(nprnd(2**32/jobs))) |
350 |
print "%s with %i %s done" % (Alu,jobs,ParaStyle) |
351 |
elif ParaStyle=='Hybrid': |
352 |
# en OpenCL, necessaire de mettre un Global_id identique au local_id |
353 |
CLLaunch=MetropolisCL.MainLoopHybrid(queue,(blocks*blocks,),(blocks,), |
354 |
circleCL, |
355 |
numpy.uint32(iterationsCL), |
356 |
numpy.uint32(nprnd(2**32/jobs)), |
357 |
numpy.uint32(nprnd(2**32/jobs))) |
358 |
print "%s with (Blocks,Threads)=(%i,%i) %s done" % (Alu,blocks,blocks,ParaStyle) |
359 |
else: |
360 |
# en OpenCL, necessaire de mettre un Global_id identique au local_id |
361 |
CLLaunch=MetropolisCL.MainLoopLocal(queue,(jobs,),(jobs,), |
362 |
circleCL, |
363 |
numpy.uint32(iterationsCL), |
364 |
numpy.uint32(nprnd(2**32/jobs)), |
365 |
numpy.uint32(nprnd(2**32/jobs))) |
366 |
print "%s with %i %s done" % (Alu,jobs,ParaStyle) |
367 |
|
368 |
CLLaunch.wait() |
369 |
cl.enqueue_copy(queue, circle, circleCL).wait() |
370 |
|
371 |
elapsed = 1e-9*(CLLaunch.profile.end - CLLaunch.profile.start) |
372 |
|
373 |
#print circle,float(numpy.sum(circle)) |
374 |
MyPi[i]=4.*float(numpy.sum(circle))/float(iterationsNew) |
375 |
MyDuration[i]=elapsed |
376 |
#print MyPi[i],MyDuration[i] |
377 |
|
378 |
circleCL.release() |
379 |
|
380 |
#print jobs,numpy.mean(MyPi),numpy.median(MyPi),numpy.std(MyPi) |
381 |
print jobs,numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration) |
382 |
|
383 |
return(numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration)) |
384 |
|
385 |
|
386 |
def FitAndPrint(N,D,Curves): |
387 |
|
388 |
try: |
389 |
coeffs_Amdahl, matcov_Amdahl = curve_fit(Amdahl, N, D) |
390 |
|
391 |
D_Amdahl=Amdahl(N,coeffs_Amdahl[0],coeffs_Amdahl[1],coeffs_Amdahl[2]) |
392 |
coeffs_Amdahl[1]=coeffs_Amdahl[1]*coeffs_Amdahl[0]/D[0] |
393 |
coeffs_Amdahl[2]=coeffs_Amdahl[2]*coeffs_Amdahl[0]/D[0] |
394 |
coeffs_Amdahl[0]=D[0] |
395 |
print "Amdahl Normalized: T=%.2f(%.6f+%.6f/N)" % \ |
396 |
(coeffs_Amdahl[0],coeffs_Amdahl[1],coeffs_Amdahl[2]) |
397 |
except: |
398 |
print "Impossible to fit for Amdahl law : only %i elements" % len(D) |
399 |
|
400 |
try: |
401 |
coeffs_AmdahlR, matcov_AmdahlR = curve_fit(AmdahlR, N, D) |
402 |
|
403 |
D_AmdahlR=AmdahlR(N,coeffs_AmdahlR[0],coeffs_AmdahlR[1]) |
404 |
coeffs_AmdahlR[1]=coeffs_AmdahlR[1]*coeffs_AmdahlR[0]/D[0] |
405 |
coeffs_AmdahlR[0]=D[0] |
406 |
print "Amdahl Reduced Normalized: T=%.2f(%.6f+%.6f/N)" % \ |
407 |
(coeffs_AmdahlR[0],1-coeffs_AmdahlR[1],coeffs_AmdahlR[1]) |
408 |
|
409 |
except: |
410 |
print "Impossible to fit for Reduced Amdahl law : only %i elements" % len(D) |
411 |
|
412 |
try: |
413 |
coeffs_Mylq, matcov_Mylq = curve_fit(Mylq, N, D) |
414 |
|
415 |
coeffs_Mylq[1]=coeffs_Mylq[1]*coeffs_Mylq[0]/D[0] |
416 |
coeffs_Mylq[2]=coeffs_Mylq[2]*coeffs_Mylq[0]/D[0] |
417 |
coeffs_Mylq[3]=coeffs_Mylq[3]*coeffs_Mylq[0]/D[0] |
418 |
coeffs_Mylq[0]=D[0] |
419 |
print "Mylq Normalized : T=%.2f(%.6f+%.6f*N+%.6f/N)" % (coeffs_Mylq[0], |
420 |
coeffs_Mylq[1], |
421 |
coeffs_Mylq[2], |
422 |
coeffs_Mylq[3]) |
423 |
D_Mylq=Mylq(N,coeffs_Mylq[0],coeffs_Mylq[1],coeffs_Mylq[2], |
424 |
coeffs_Mylq[3]) |
425 |
except: |
426 |
print "Impossible to fit for Mylq law : only %i elements" % len(D) |
427 |
|
428 |
try: |
429 |
coeffs_Mylq2, matcov_Mylq2 = curve_fit(Mylq2, N, D) |
430 |
|
431 |
coeffs_Mylq2[1]=coeffs_Mylq2[1]*coeffs_Mylq2[0]/D[0] |
432 |
coeffs_Mylq2[2]=coeffs_Mylq2[2]*coeffs_Mylq2[0]/D[0] |
433 |
coeffs_Mylq2[3]=coeffs_Mylq2[3]*coeffs_Mylq2[0]/D[0] |
434 |
coeffs_Mylq2[4]=coeffs_Mylq2[4]*coeffs_Mylq2[0]/D[0] |
435 |
coeffs_Mylq2[0]=D[0] |
436 |
print "Mylq 2nd order Normalized: T=%.2f(%.6f+%.6f*N+%.6f*N^2+%.6f/N)" % \ |
437 |
(coeffs_Mylq2[0],coeffs_Mylq2[1],coeffs_Mylq2[2],coeffs_Mylq2[3], |
438 |
coeffs_Mylq2[4]) |
439 |
|
440 |
except: |
441 |
print "Impossible to fit for 2nd order Mylq law : only %i elements" % len(D) |
442 |
|
443 |
if Curves: |
444 |
plt.xlabel("Number of Threads/work Items") |
445 |
plt.ylabel("Total Elapsed Time") |
446 |
|
447 |
Experience,=plt.plot(N,D,'ro') |
448 |
try: |
449 |
pAmdahl,=plt.plot(N,D_Amdahl,label="Loi de Amdahl") |
450 |
pMylq,=plt.plot(N,D_Mylq,label="Loi de Mylq") |
451 |
except: |
452 |
print "Fit curves seem not to be available" |
453 |
|
454 |
plt.legend() |
455 |
plt.show() |
456 |
|
457 |
if __name__=='__main__': |
458 |
|
459 |
# Set defaults values |
460 |
# Alu can be CPU or GPU |
461 |
Alu='CPU' |
462 |
# Id of GPU |
463 |
Device=1 |
464 |
# GPU style can be Cuda (Nvidia implementation) or OpenCL |
465 |
GpuStyle='OpenCL' |
466 |
# Parallel distribution can be on Threads or Blocks |
467 |
ParaStyle='Threads' |
468 |
# Iterations is integer |
469 |
Iterations=1000000 |
470 |
# JobStart in first number of Jobs to explore |
471 |
JobStart=1 |
472 |
# JobEnd is last number of Jobs to explore |
473 |
JobEnd=512 |
474 |
# Redo is the times to redo the test to improve metrology |
475 |
Redo=1 |
476 |
# OutMetrology is method for duration estimation : False is GPU inside |
477 |
OutMetrology=False |
478 |
# Curves is True to print the curves |
479 |
Curves=False |
480 |
|
481 |
try: |
482 |
opts, args = getopt.getopt(sys.argv[1:],"hoca:g:p:i:s:e:r:d:",["alu=","gpustyle=","parastyle=","iterations=","jobstart=","jobend=","redo=","device="]) |
483 |
except getopt.GetoptError: |
484 |
print '%s -o (Out of Core Metrology) -c (Print Curves) -a <CPU/GPU> -d <DeviceId> -g <CUDA/OpenCL> -p <Threads/Hybrid/Blocks> -i <Iterations> -s <JobStart> -e <JobEnd> -r <RedoToImproveStats>' % sys.argv[0] |
485 |
sys.exit(2) |
486 |
|
487 |
for opt, arg in opts: |
488 |
if opt == '-h': |
489 |
print '%s -o (Out of Core Metrology) -c (Print Curves) -a <CPU/GPU> -d <DeviceId> -g <CUDA/OpenCL> -p <Threads/Hybrid/Blocks> -i <Iterations> -s <JobStart> -e <JobEnd> -r <RedoToImproveStats>' % sys.argv[0] |
490 |
sys.exit() |
491 |
elif opt == '-o': |
492 |
OutMetrology=True |
493 |
elif opt == '-c': |
494 |
Curves=True |
495 |
elif opt in ("-a", "--alu"): |
496 |
Alu = arg |
497 |
elif opt in ("-d", "--device"): |
498 |
Device = int(arg) |
499 |
elif opt in ("-g", "--gpustyle"): |
500 |
GpuStyle = arg |
501 |
elif opt in ("-p", "--parastyle"): |
502 |
ParaStyle = arg |
503 |
elif opt in ("-i", "--iterations"): |
504 |
Iterations = numpy.uint32(arg) |
505 |
elif opt in ("-s", "--jobstart"): |
506 |
JobStart = int(arg) |
507 |
elif opt in ("-e", "--jobend"): |
508 |
JobEnd = int(arg) |
509 |
elif opt in ("-r", "--redo"): |
510 |
Redo = int(arg) |
511 |
|
512 |
if Alu=='CPU' and GpuStyle=='CUDA': |
513 |
print "Alu can't be CPU for CUDA, set Alu to GPU" |
514 |
Alu='GPU' |
515 |
|
516 |
if ParaStyle not in ('Blocks','Threads','Hybrid'): |
517 |
print "%s not exists, ParaStyle set as Threads !" % ParaStyle |
518 |
ParaStyle='Threads' |
519 |
|
520 |
print "Compute unit : %s" % Alu |
521 |
print "Device Identification : %s" % Device |
522 |
print "GpuStyle used : %s" % GpuStyle |
523 |
print "Parallel Style used : %s" % ParaStyle |
524 |
print "Iterations : %s" % Iterations |
525 |
print "Number of threads on start : %s" % JobStart |
526 |
print "Number of threads on end : %s" % JobEnd |
527 |
print "Number of redo : %s" % Redo |
528 |
print "Metrology done out of CPU/GPU : %r" % OutMetrology |
529 |
|
530 |
if GpuStyle=='CUDA': |
531 |
# For PyCUDA import |
532 |
import pycuda.driver as cuda |
533 |
import pycuda.gpuarray as gpuarray |
534 |
import pycuda.autoinit |
535 |
from pycuda.compiler import SourceModule |
536 |
|
537 |
if GpuStyle=='OpenCL': |
538 |
# For PyOpenCL import |
539 |
import pyopencl as cl |
540 |
Id=1 |
541 |
for platform in cl.get_platforms(): |
542 |
for device in platform.get_devices(): |
543 |
deviceType=cl.device_type.to_string(device.type) |
544 |
print "Device #%i of type %s : %s" % (Id,deviceType,device.name) |
545 |
Id=Id+1 |
546 |
|
547 |
average=numpy.array([]).astype(numpy.float32) |
548 |
median=numpy.array([]).astype(numpy.float32) |
549 |
stddev=numpy.array([]).astype(numpy.float32) |
550 |
|
551 |
ExploredJobs=numpy.array([]).astype(numpy.uint32) |
552 |
|
553 |
Jobs=JobStart |
554 |
|
555 |
while Jobs <= JobEnd: |
556 |
avg,med,std=0,0,0 |
557 |
ExploredJobs=numpy.append(ExploredJobs,Jobs) |
558 |
circle=numpy.zeros(Jobs).astype(numpy.uint32) |
559 |
|
560 |
if OutMetrology: |
561 |
duration=numpy.array([]).astype(numpy.float32) |
562 |
for i in range(Redo): |
563 |
start=time.time() |
564 |
if GpuStyle=='CUDA': |
565 |
try: |
566 |
MetropolisCuda(circle,Iterations,1,Jobs,ParaStyle) |
567 |
except: |
568 |
print "Problem with %i // computations on Cuda" % Jobs |
569 |
elif GpuStyle=='OpenCL': |
570 |
try: |
571 |
MetropolisOpenCL(circle,Iterations,1,Jobs,ParaStyle,Alu,Device) |
572 |
except: |
573 |
print "Problem with %i // computations on OpenCL" % Jobs |
574 |
duration=numpy.append(duration,time.time()-start) |
575 |
avg=numpy.mean(duration) |
576 |
med=numpy.median(duration) |
577 |
std=numpy.std(duration) |
578 |
else: |
579 |
if GpuStyle=='CUDA': |
580 |
try: |
581 |
avg,med,std=MetropolisCuda(circle,Iterations,Redo,Jobs,ParaStyle) |
582 |
except: |
583 |
print "Problem with %i // computations on Cuda" % Jobs |
584 |
elif GpuStyle=='OpenCL': |
585 |
try: |
586 |
avg,med,std=MetropolisOpenCL(circle,Iterations,Redo,Jobs,ParaStyle,Alu,Device) |
587 |
except: |
588 |
print "Problem with %i // computations on OpenCL" % Jobs |
589 |
|
590 |
if (avg,med,std) != (0,0,0): |
591 |
print "avg,med,std",avg,med,std |
592 |
average=numpy.append(average,avg) |
593 |
median=numpy.append(median,med) |
594 |
stddev=numpy.append(stddev,std) |
595 |
else: |
596 |
print "Values seem to be wrong..." |
597 |
#THREADS*=2 |
598 |
numpy.savez("Pi_%s_%s_%s_%s_%i_%.8i_%s" % (Alu,GpuStyle,ParaStyle,JobStart,JobEnd,Iterations,gethostname()),(ExploredJobs,average,median,stddev)) |
599 |
Jobs+=1 |
600 |
|
601 |
FitAndPrint(ExploredJobs,median,Curves) |
602 |
|