root / Pi / GPU / Pi-GPU.py-20130205 @ 49
Historique | Voir | Annoter | Télécharger (15,79 ko)
1 |
#!/usr/bin/env python |
---|---|
2 |
# |
3 |
# Pi-by-MC using PyCUDA |
4 |
# |
5 |
# CC BY-NC-SA 2011 : <emmanuel.quemener@ens-lyon.fr> |
6 |
# |
7 |
# Thanks to Andreas Klockner for PyCUDA: |
8 |
# http://mathema.tician.de/software/pycuda |
9 |
# |
10 |
|
11 |
# 2013-01-01 : problems with launch timeout |
12 |
# nvidia-smi -c 1 : ko |
13 |
# nvidia-smi -c 3 : ko |
14 |
# nvidia-smi --gom=1 : not supported |
15 |
# http://stackoverflow.com/questions/497685/how-do-you-get-around-the-maximum-cuda-run-time |
16 |
# Option "Interactive" "0" in /etc/X11/xorg.conf |
17 |
|
18 |
# Common tools |
19 |
import numpy |
20 |
from numpy.random import randint as nprnd |
21 |
import sys |
22 |
import getopt |
23 |
import time |
24 |
import matplotlib.pyplot as plt |
25 |
from scipy.optimize import curve_fit |
26 |
from socket import gethostname |
27 |
|
28 |
# Predicted Amdahl Law (Reduced with s=1-p) |
29 |
def AmdahlR(N, T1, p): |
30 |
return (T1*(1-p+p/N)) |
31 |
|
32 |
# Predicted Amdahl Law |
33 |
def Amdahl(N, T1, s, p): |
34 |
return (T1*(s+p/N)) |
35 |
|
36 |
# Predicted Mylq Law with first order |
37 |
def Mylq(N, T1,s,c,p): |
38 |
return (T1*(s+c*N+p/N)) |
39 |
|
40 |
# Predicted Mylq Law with second order |
41 |
def Mylq2(N, T1,s,c1,c2,p): |
42 |
return (T1*(s+c1*N+c2*N*N+p/N)) |
43 |
|
44 |
KERNEL_CODE_CUDA=""" |
45 |
|
46 |
// Marsaglia RNG very simple implementation |
47 |
|
48 |
#define znew ((z=36969*(z&65535)+(z>>16))<<16) |
49 |
#define wnew ((w=18000*(w&65535)+(w>>16))&65535) |
50 |
#define MWC (znew+wnew) |
51 |
#define SHR3 (jsr=(jsr=(jsr=jsr^(jsr<<17))^(jsr>>13))^(jsr<<5)) |
52 |
#define CONG (jcong=69069*jcong+1234567) |
53 |
#define KISS ((MWC^CONG)+SHR3) |
54 |
|
55 |
//#define MWCfp (MWC + 2147483648.0f) * 2.328306435454494e-10f |
56 |
//#define KISSfp (KISS + 2147483648.0f) * 2.328306435454494e-10f |
57 |
#define MWCfp MWC * 2.328306435454494e-10f |
58 |
#define KISSfp KISS * 2.328306435454494e-10f |
59 |
|
60 |
__global__ void MainLoopBlocks(uint *s,uint iterations,uint seed_w,uint seed_z) |
61 |
{ |
62 |
uint z=seed_z/(blockIdx.x+1); |
63 |
uint w=seed_w/(blockIdx.x+1); |
64 |
// uint jsr=123456789; |
65 |
// uint jcong=380116160; |
66 |
|
67 |
int total=0; |
68 |
|
69 |
for (uint i=0;i<iterations;i++) { |
70 |
|
71 |
float x=MWCfp ; |
72 |
float y=MWCfp ; |
73 |
|
74 |
// Matching test |
75 |
int inside=((x*x+y*y) < 1.0f) ? 1:0; |
76 |
total+=inside; |
77 |
|
78 |
} |
79 |
|
80 |
s[blockIdx.x]=total; |
81 |
__syncthreads(); |
82 |
|
83 |
} |
84 |
|
85 |
__global__ void MainLoopThreads(uint *s,uint iterations,uint seed_w,uint seed_z) |
86 |
{ |
87 |
uint z=seed_z/(threadIdx.x+1); |
88 |
uint w=seed_w/(threadIdx.x+1); |
89 |
// uint jsr=123456789; |
90 |
// uint jcong=380116160; |
91 |
|
92 |
int total=0; |
93 |
|
94 |
for (uint i=0;i<iterations;i++) { |
95 |
|
96 |
float x=MWCfp ; |
97 |
float y=MWCfp ; |
98 |
|
99 |
// Matching test |
100 |
int inside=((x*x+y*y) < 1.0f) ? 1:0; |
101 |
total+=inside; |
102 |
|
103 |
} |
104 |
|
105 |
s[threadIdx.x]=total; |
106 |
__syncthreads(); |
107 |
|
108 |
} |
109 |
""" |
110 |
|
111 |
KERNEL_CODE_OPENCL=""" |
112 |
|
113 |
// Marsaglia RNG very simple implementation |
114 |
#define znew ((z=36969*(z&65535)+(z>>16))<<16) |
115 |
#define wnew ((w=18000*(w&65535)+(w>>16))&65535) |
116 |
#define MWC (znew+wnew) |
117 |
#define SHR3 (jsr=(jsr=(jsr=jsr^(jsr<<17))^(jsr>>13))^(jsr<<5)) |
118 |
#define CONG (jcong=69069*jcong+1234567) |
119 |
#define KISS ((MWC^CONG)+SHR3) |
120 |
|
121 |
//#define MWCfp (MWC + 2147483648.0f) * 2.328306435454494e-10f |
122 |
//#define KISSfp (KISS + 2147483648.0f) * 2.328306435454494e-10f |
123 |
#define MWCfp MWC * 2.328306435454494e-10f |
124 |
#define KISSfp KISS * 2.328306435454494e-10f |
125 |
|
126 |
__kernel void MainLoopGlobal(__global int *s,uint iterations,uint seed_w,uint seed_z) |
127 |
{ |
128 |
uint z=seed_z/(get_global_id(0)+1); |
129 |
uint w=seed_w/(get_global_id(0)+1); |
130 |
// uint jsr=123456789; |
131 |
// uint jcong=380116160; |
132 |
|
133 |
int total=0; |
134 |
|
135 |
for (uint i=0;i<iterations;i++) { |
136 |
|
137 |
float x=MWCfp ; |
138 |
float y=MWCfp ; |
139 |
|
140 |
// Matching test |
141 |
int inside=((x*x+y*y) < 1.0f) ? 1:0; |
142 |
total+=inside; |
143 |
} |
144 |
s[get_global_id(0)]=total; |
145 |
barrier(CLK_GLOBAL_MEM_FENCE); |
146 |
|
147 |
} |
148 |
|
149 |
__kernel void MainLoopLocal(__global int *s,uint iterations,uint seed_w,uint seed_z) |
150 |
{ |
151 |
uint z=seed_z/(get_local_id(0)+1); |
152 |
uint w=seed_w/(get_local_id(0)+1); |
153 |
// uint jsr=123456789; |
154 |
// uint jcong=380116160; |
155 |
|
156 |
int total=0; |
157 |
|
158 |
for (uint i=0;i<iterations;i++) { |
159 |
|
160 |
float x=MWCfp ; |
161 |
float y=MWCfp ; |
162 |
|
163 |
// Matching test |
164 |
int inside=((x*x+y*y) < 1.0f) ? 1:0; |
165 |
total+=inside; |
166 |
} |
167 |
s[get_local_id(0)]=total; |
168 |
barrier(CLK_LOCAL_MEM_FENCE); |
169 |
|
170 |
} |
171 |
""" |
172 |
|
173 |
def MetropolisCuda(circle,iterations,steps,threads,ParaStyle): |
174 |
|
175 |
# Avec PyCUDA autoinit, rien a faire ! |
176 |
|
177 |
circleCU = cuda.InOut(circle) |
178 |
|
179 |
mod = SourceModule(KERNEL_CODE_CUDA) |
180 |
|
181 |
MetropolisBlocksCU=mod.get_function("MainLoopBlocks") |
182 |
MetropolisThreadsCU=mod.get_function("MainLoopThreads") |
183 |
|
184 |
start = pycuda.driver.Event() |
185 |
stop = pycuda.driver.Event() |
186 |
|
187 |
MyPi=numpy.zeros(steps) |
188 |
MyDuration=numpy.zeros(steps) |
189 |
|
190 |
if iterations%threads==0: |
191 |
iterationsCL=numpy.uint32(iterations/threads+1) |
192 |
iterationsNew=iterationsCL*threads |
193 |
else: |
194 |
iterationsCL=numpy.uint32(iterations/threads) |
195 |
iterationsNew=iterations |
196 |
|
197 |
for i in range(steps): |
198 |
start.record() |
199 |
start.synchronize() |
200 |
if ParaStyle=='Blocks': |
201 |
MetropolisBlocksCU(circleCU, |
202 |
numpy.uint32(iterationsCL), |
203 |
numpy.uint32(nprnd(2**32/threads)), |
204 |
numpy.uint32(nprnd(2**32/threads)), |
205 |
grid=(threads,1), |
206 |
block=(1,1,1)) |
207 |
else: |
208 |
MetropolisThreadsCU(circleCU, |
209 |
numpy.uint32(iterationsCL), |
210 |
numpy.uint32(nprnd(2**32/threads)), |
211 |
numpy.uint32(nprnd(2**32/threads)), |
212 |
grid=(1,1), |
213 |
block=(threads,1,1)) |
214 |
print "GPU done" |
215 |
stop.record() |
216 |
stop.synchronize() |
217 |
|
218 |
#elapsed = stop.time_since(start)*1e-3 |
219 |
elapsed = start.time_till(stop)*1e-3 |
220 |
|
221 |
#print circle,float(numpy.sum(circle)) |
222 |
MyPi[i]=4.*float(numpy.sum(circle))/float(iterationsCL) |
223 |
MyDuration[i]=elapsed |
224 |
#print MyPi[i],MyDuration[i] |
225 |
#time.sleep(1) |
226 |
|
227 |
print threads,numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration) |
228 |
|
229 |
return(numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration)) |
230 |
|
231 |
|
232 |
def MetropolisOpenCL(circle,iterations,steps,threads,ParaStyle,Alu): |
233 |
|
234 |
# Initialisation des variables en les CASTant correctement |
235 |
|
236 |
# Je detecte un peripherique GPU dans la liste des peripheriques |
237 |
# for platform in cl.get_platforms(): |
238 |
# for device in platform.get_devices(): |
239 |
# if cl.device_type.to_string(device.type)=='GPU': |
240 |
# GPU=device |
241 |
#print "GPU detected: ",device.name |
242 |
|
243 |
HasGPU=False |
244 |
# Device selection based on choice (default is GPU) |
245 |
for platform in cl.get_platforms(): |
246 |
for device in platform.get_devices(): |
247 |
if not HasGPU: |
248 |
deviceType=cl.device_type.to_string(device.type) |
249 |
if deviceType=="GPU" and Alu=="GPU": |
250 |
GPU=device |
251 |
print "GPU selected: ",device.name |
252 |
HasGPU=True |
253 |
if deviceType=="CPU" and Alu=="CPU": |
254 |
GPU=device |
255 |
print "CPU selected: ",device.name |
256 |
HasGPU=True |
257 |
|
258 |
# Je cree le contexte et la queue pour son execution |
259 |
#ctx = cl.create_some_context() |
260 |
ctx = cl.Context([GPU]) |
261 |
queue = cl.CommandQueue(ctx, |
262 |
properties=cl.command_queue_properties.PROFILING_ENABLE) |
263 |
|
264 |
# Je recupere les flag possibles pour les buffers |
265 |
mf = cl.mem_flags |
266 |
|
267 |
circleCL = cl.Buffer(ctx, mf.WRITE_ONLY|mf.COPY_HOST_PTR,hostbuf=circle) |
268 |
|
269 |
MetropolisCL = cl.Program(ctx,KERNEL_CODE_OPENCL).build( \ |
270 |
options = "-cl-mad-enable -cl-fast-relaxed-math") |
271 |
|
272 |
#MetropolisCL = cl.Program(ctx,KERNEL_CODE_OPENCL).build() |
273 |
|
274 |
i=0 |
275 |
|
276 |
MyPi=numpy.zeros(steps) |
277 |
MyDuration=numpy.zeros(steps) |
278 |
|
279 |
if iterations%threads==0: |
280 |
iterationsCL=numpy.uint32(iterations/threads+1) |
281 |
iterationsNew=iterationsCL*threads |
282 |
else: |
283 |
iterationsCL=numpy.uint32(iterations/threads) |
284 |
iterationsNew=iterations |
285 |
|
286 |
for i in range(steps): |
287 |
|
288 |
if ParaStyle=='Blocks': |
289 |
# Call OpenCL kernel |
290 |
# (1,) is Global work size (only 1 work size) |
291 |
# (1,) is local work size |
292 |
# circleCL is lattice translated in CL format |
293 |
# SeedZCL is lattice translated in CL format |
294 |
# SeedWCL is lattice translated in CL format |
295 |
# step is number of iterations |
296 |
CLLaunch=MetropolisCL.MainLoopGlobal(queue,(threads,),(1,), |
297 |
circleCL, |
298 |
numpy.uint32(iterationsCL), |
299 |
numpy.uint32(nprnd(2**32/threads)), |
300 |
numpy.uint32(nprnd(2**32/threads))) |
301 |
print "%s with %s done" % (Alu,ParaStyle) |
302 |
else: |
303 |
# en OpenCL, necessaire de mettre un Global_id identique au local_id |
304 |
CLLaunch=MetropolisCL.MainLoopLocal(queue,(threads,),(threads,), |
305 |
circleCL, |
306 |
numpy.uint32(iterationsCL), |
307 |
numpy.uint32(nprnd(2**32/threads)), |
308 |
numpy.uint32(nprnd(2**32/threads))) |
309 |
print "%s with %s done" % (Alu,ParaStyle) |
310 |
|
311 |
CLLaunch.wait() |
312 |
cl.enqueue_copy(queue, circle, circleCL).wait() |
313 |
|
314 |
elapsed = 1e-9*(CLLaunch.profile.end - CLLaunch.profile.start) |
315 |
|
316 |
#print circle,float(numpy.sum(circle)) |
317 |
MyPi[i]=4.*float(numpy.sum(circle))/float(iterationsNew) |
318 |
MyDuration[i]=elapsed |
319 |
#print MyPi[i],MyDuration[i] |
320 |
|
321 |
circleCL.release() |
322 |
|
323 |
#print threads,numpy.mean(MyPi),numpy.median(MyPi),numpy.std(MyPi) |
324 |
print threads,numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration) |
325 |
|
326 |
return(numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration)) |
327 |
|
328 |
|
329 |
def FitAndPrint(N,D,Curves): |
330 |
|
331 |
try: |
332 |
coeffs_Amdahl, matcov_Amdahl = curve_fit(Amdahl, N, D) |
333 |
|
334 |
D_Amdahl=Amdahl(N,coeffs_Amdahl[0],coeffs_Amdahl[1],coeffs_Amdahl[2]) |
335 |
coeffs_Amdahl[1]=coeffs_Amdahl[1]*coeffs_Amdahl[0]/D[0] |
336 |
coeffs_Amdahl[2]=coeffs_Amdahl[2]*coeffs_Amdahl[0]/D[0] |
337 |
coeffs_Amdahl[0]=D[0] |
338 |
print "Amdahl Normalized: T=%.2f(%.5f+%.5f/N)" % \ |
339 |
(coeffs_Amdahl[0],coeffs_Amdahl[1],coeffs_Amdahl[2]) |
340 |
except: |
341 |
print "Impossible to fit for Amdahl law : only %i elements" % len(D) |
342 |
|
343 |
try: |
344 |
coeffs_AmdahlR, matcov_AmdahlR = curve_fit(AmdahlR, N, D) |
345 |
|
346 |
D_AmdahlR=AmdahlR(N,coeffs_AmdahlR[0],coeffs_AmdahlR[1]) |
347 |
coeffs_AmdahlR[1]=coeffs_AmdahlR[1]*coeffs_AmdahlR[0]/D[0] |
348 |
coeffs_AmdahlR[0]=D[0] |
349 |
print "Amdahl Reduced Normalized: T=%.2f(%.5f+%.5f/N)" % \ |
350 |
(coeffs_AmdahlR[0],1-coeffs_AmdahlR[1],coeffs_AmdahlR[1]) |
351 |
|
352 |
except: |
353 |
print "Impossible to fit for Reduced Amdahl law : only %i elements" % len(D) |
354 |
|
355 |
try: |
356 |
coeffs_Mylq, matcov_Mylq = curve_fit(Mylq, N, D) |
357 |
|
358 |
coeffs_Mylq[1]=coeffs_Mylq[1]*coeffs_Mylq[0]/D[0] |
359 |
coeffs_Mylq[2]=coeffs_Mylq[2]*coeffs_Mylq[0]/D[0] |
360 |
coeffs_Mylq[3]=coeffs_Mylq[3]*coeffs_Mylq[0]/D[0] |
361 |
coeffs_Mylq[0]=D[0] |
362 |
print "Mylq Normalized : T=%.2f(%.5f+%.5f*N+%.5f/N)" % (coeffs_Mylq[0], |
363 |
coeffs_Mylq[1], |
364 |
coeffs_Mylq[2], |
365 |
coeffs_Mylq[3]) |
366 |
D_Mylq=Mylq(N,coeffs_Mylq[0],coeffs_Mylq[1],coeffs_Mylq[2], |
367 |
coeffs_Mylq[3]) |
368 |
except: |
369 |
print "Impossible to fit for Mylq law : only %i elements" % len(D) |
370 |
|
371 |
try: |
372 |
coeffs_Mylq2, matcov_Mylq2 = curve_fit(Mylq2, N, D) |
373 |
|
374 |
coeffs_Mylq2[1]=coeffs_Mylq2[1]*coeffs_Mylq2[0]/D[0] |
375 |
coeffs_Mylq2[2]=coeffs_Mylq2[2]*coeffs_Mylq2[0]/D[0] |
376 |
coeffs_Mylq2[3]=coeffs_Mylq2[3]*coeffs_Mylq2[0]/D[0] |
377 |
coeffs_Mylq2[4]=coeffs_Mylq2[4]*coeffs_Mylq2[0]/D[0] |
378 |
coeffs_Mylq2[0]=D[0] |
379 |
print "Mylq 2nd order Normalized: T=%.2f(%.5f+%.5f*N+%.5f*N^2+%.5f/N)" % \ |
380 |
(coeffs_Mylq2[0],coeffs_Mylq2[1],coeffs_Mylq2[2],coeffs_Mylq2[3], |
381 |
coeffs_Mylq2[4]) |
382 |
|
383 |
except: |
384 |
print "Impossible to fit for 2nd order Mylq law : only %i elements" % len(D) |
385 |
|
386 |
if Curves: |
387 |
plt.xlabel("Number of Threads/work Items") |
388 |
plt.ylabel("Total Elapsed Time") |
389 |
|
390 |
Experience,=plt.plot(N,D,'ro') |
391 |
try: |
392 |
pAmdahl,=plt.plot(N,D_Amdahl,label="Loi de Amdahl") |
393 |
pMylq,=plt.plot(N,D_Mylq,label="Loi de Mylq") |
394 |
except: |
395 |
print "Fit curves seem not to be available" |
396 |
|
397 |
plt.legend() |
398 |
plt.show() |
399 |
|
400 |
if __name__=='__main__': |
401 |
|
402 |
# Set defaults values |
403 |
# Alu can be CPU or GPU |
404 |
Alu='CPU' |
405 |
# GPU style can be Cuda (Nvidia implementation) or OpenCL |
406 |
GpuStyle='OpenCL' |
407 |
# Parallel distribution can be on Threads or Blocks |
408 |
ParaStyle='Threads' |
409 |
# Iterations is integer |
410 |
Iterations=1000000000 |
411 |
# ThreadStart in first number of Threads to explore |
412 |
ThreadStart=1 |
413 |
# ThreadEnd is last number of Threads to explore |
414 |
ThreadEnd=512 |
415 |
# Redo is the times to redo the test to improve metrology |
416 |
Redo=1 |
417 |
# OutMetrology is method for duration estimation : False is GPU inside |
418 |
OutMetrology=False |
419 |
# Curves is True to print the curves |
420 |
Curves=False |
421 |
|
422 |
try: |
423 |
opts, args = getopt.getopt(sys.argv[1:],"hoca:g:p:i:s:e:r:",["alu=","gpustyle=","parastyle=","iterations=","threadstart=","threadend=","redo="]) |
424 |
except getopt.GetoptError: |
425 |
print '%s -o -a <CPU/GPU> -g <CUDA/OpenCL> -p <ParaStyle> -i <Iterations> -s <ThreadStart> -e <ThreadEnd> -r <RedoToImproveStats>' % sys.argv[0] |
426 |
sys.exit(2) |
427 |
|
428 |
for opt, arg in opts: |
429 |
if opt == '-h': |
430 |
print '%s -o (Out of Core Metrology) -c (Print Curves) -a <CPU/GPU> -g <CUDA/OpenCL> -p <Threads/Blocks> -i <Iterations> -s <ThreadStart> -e <ThreadEnd> -r <RedoToImproveStats>' % sys.argv[0] |
431 |
sys.exit() |
432 |
elif opt == '-o': |
433 |
OutMetrology=True |
434 |
elif opt == '-c': |
435 |
Curves=True |
436 |
elif opt in ("-a", "--alu"): |
437 |
Alu = arg |
438 |
elif opt in ("-g", "--gpustyle"): |
439 |
GpuStyle = arg |
440 |
elif opt in ("-p", "--parastyle"): |
441 |
ParaStyle = arg |
442 |
elif opt in ("-i", "--iterations"): |
443 |
Iterations = numpy.uint32(arg) |
444 |
elif opt in ("-s", "--threadstart"): |
445 |
ThreadStart = int(arg) |
446 |
elif opt in ("-e", "--threadend"): |
447 |
ThreadEnd = int(arg) |
448 |
elif opt in ("-r", "--redo"): |
449 |
Redo = int(arg) |
450 |
|
451 |
if Alu=='CPU' and GpuStyle=='CUDA': |
452 |
print "Alu can't be CPU for CUDA, set Alu to GPU" |
453 |
Alu='GPU' |
454 |
|
455 |
if ParaStyle not in ('Blocks','Threads'): |
456 |
print "%s not exists, ParaStyle set as Threads !" % ParaStyle |
457 |
ParaStyle='Threads' |
458 |
|
459 |
print "Compute unit : %s" % Alu |
460 |
print "GpuStyle used : %s" % GpuStyle |
461 |
print "Parallel Style used : %s" % ParaStyle |
462 |
print "Iterations : %s" % Iterations |
463 |
print "Number of threads on start : %s" % ThreadStart |
464 |
print "Number of threads on end : %s" % ThreadEnd |
465 |
print "Number of redo : %s" % Redo |
466 |
print "Metrology done out of CPU/GPU : %r" % OutMetrology |
467 |
|
468 |
if GpuStyle=='CUDA': |
469 |
# For PyCUDA import |
470 |
import pycuda.driver as cuda |
471 |
import pycuda.gpuarray as gpuarray |
472 |
import pycuda.autoinit |
473 |
from pycuda.compiler import SourceModule |
474 |
|
475 |
if GpuStyle=='OpenCL': |
476 |
# For PyOpenCL import |
477 |
import pyopencl as cl |
478 |
|
479 |
average=numpy.array([]).astype(numpy.float32) |
480 |
median=numpy.array([]).astype(numpy.float32) |
481 |
stddev=numpy.array([]).astype(numpy.float32) |
482 |
|
483 |
ExploredThreads=numpy.array([]).astype(numpy.uint32) |
484 |
|
485 |
Threads=ThreadStart |
486 |
|
487 |
while Threads <= ThreadEnd: |
488 |
ExploredThreads=numpy.append(ExploredThreads,Threads) |
489 |
circle=numpy.zeros(Threads).astype(numpy.uint32) |
490 |
|
491 |
if OutMetrology: |
492 |
duration=numpy.array([]).astype(numpy.float32) |
493 |
for i in range(Redo): |
494 |
start=time.time() |
495 |
if GpuStyle=='CUDA': |
496 |
MetropolisCuda(circle,Iterations,1,Threads,ParaStyle) |
497 |
elif GpuStyle=='OpenCL': |
498 |
MetropolisOpenCL(circle,Iterations,1,Threads,ParaStyle,Alu) |
499 |
duration=numpy.append(duration,time.time()-start) |
500 |
avg=numpy.mean(duration) |
501 |
med=numpy.median(duration) |
502 |
std=numpy.std(duration) |
503 |
else: |
504 |
if GpuStyle=='CUDA': |
505 |
avg,med,std=MetropolisCuda(circle,Iterations,Redo,Threads,ParaStyle) |
506 |
elif GpuStyle=='OpenCL': |
507 |
avg,med,std=MetropolisOpenCL(circle,Iterations,Redo,Threads, |
508 |
ParaStyle,Alu) |
509 |
print "avg,med,std",avg,med,std |
510 |
average=numpy.append(average,avg) |
511 |
median=numpy.append(median,med) |
512 |
stddev=numpy.append(stddev,std) |
513 |
#THREADS*=2 |
514 |
Threads+=1 |
515 |
|
516 |
numpy.savez("Pi_%s_%s_%s_%s_%i_%.8i_%s" % (Alu,GpuStyle,ParaStyle,ThreadStart,ThreadEnd,Iterations,gethostname()),(ExploredThreads,average,median,stddev)) |
517 |
|
518 |
FitAndPrint(ExploredThreads,median,Curves) |
519 |
|