|
1 |
#!/usr/bin/env python
|
|
2 |
|
|
3 |
#
|
|
4 |
# Splutter-by-MonteCarlo using PyCUDA/PyOpenCL
|
|
5 |
#
|
|
6 |
# CC BY-NC-SA 2014 : <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 math
|
|
23 |
from socket import gethostname
|
|
24 |
|
|
25 |
# find prime factors of a number
|
|
26 |
# Get for WWW :
|
|
27 |
# http://pythonism.wordpress.com/2008/05/17/looking-at-factorisation-in-python/
|
|
28 |
def PrimeFactors(x):
|
|
29 |
factorlist=numpy.array([]).astype('uint32')
|
|
30 |
loop=2
|
|
31 |
while loop<=x:
|
|
32 |
if x%loop==0:
|
|
33 |
x/=loop
|
|
34 |
factorlist=numpy.append(factorlist,[loop])
|
|
35 |
else:
|
|
36 |
loop+=1
|
|
37 |
return factorlist
|
|
38 |
|
|
39 |
# Try to find the best thread number in Hybrid approach (Blocks&Threads)
|
|
40 |
# output is thread number
|
|
41 |
def BestThreadsNumber(jobs):
|
|
42 |
factors=PrimeFactors(jobs)
|
|
43 |
matrix=numpy.append([factors],[factors[::-1]],axis=0)
|
|
44 |
threads=1
|
|
45 |
for factor in matrix.transpose().ravel():
|
|
46 |
threads=threads*factor
|
|
47 |
if threads*threads>jobs:
|
|
48 |
break
|
|
49 |
return(long(threads))
|
|
50 |
|
|
51 |
# Predicted Amdahl Law (Reduced with s=1-p)
|
|
52 |
def AmdahlR(N, T1, p):
|
|
53 |
return (T1*(1-p+p/N))
|
|
54 |
|
|
55 |
# Predicted Amdahl Law
|
|
56 |
def Amdahl(N, T1, s, p):
|
|
57 |
return (T1*(s+p/N))
|
|
58 |
|
|
59 |
# Predicted Mylq Law with first order
|
|
60 |
def Mylq(N, T1,s,c,p):
|
|
61 |
return (T1*(s+p/N)+c*N)
|
|
62 |
|
|
63 |
# Predicted Mylq Law with second order
|
|
64 |
def Mylq2(N, T1,s,c1,c2,p):
|
|
65 |
return (T1*(s+p/N)+c1*N+c2*N*N)
|
|
66 |
|
|
67 |
KERNEL_CODE_CUDA="""
|
|
68 |
|
|
69 |
// Marsaglia RNG very simple implementation
|
|
70 |
|
|
71 |
#define znew ((z=36969*(z&65535)+(z>>16))<<16)
|
|
72 |
#define wnew ((w=18000*(w&65535)+(w>>16))&65535)
|
|
73 |
#define MWC (znew+wnew)
|
|
74 |
#define SHR3 (jsr=(jsr=(jsr=jsr^(jsr<<17))^(jsr>>13))^(jsr<<5))
|
|
75 |
#define CONG (jcong=69069*jcong+1234567)
|
|
76 |
#define KISS ((MWC^CONG)+SHR3)
|
|
77 |
|
|
78 |
#define MWCfp MWC * 2.328306435454494e-10f
|
|
79 |
#define KISSfp KISS * 2.328306435454494e-10f
|
|
80 |
|
|
81 |
#define MAX 4294967296
|
|
82 |
|
|
83 |
uint rotl(uint value, int shift) {
|
|
84 |
return (value << shift) | (value >> (sizeof(value) * 8 - shift));
|
|
85 |
}
|
|
86 |
|
|
87 |
uint rotr(uint value, int shift) {
|
|
88 |
return (value >> shift) | (value << (sizeof(value) * 8 - shift));
|
|
89 |
}
|
|
90 |
|
|
91 |
__global__ void MainLoopBlocks(uint *s,uint size,ulong iterations,uint seed_w,uint seed_z)
|
|
92 |
{
|
|
93 |
// uint z=rotl(seed_z,blockIdx.x);
|
|
94 |
// uint w=rotr(seed_w,blockIdx.x);
|
|
95 |
|
|
96 |
// uint jsr=rotl(seed_z,blockIdx.x);
|
|
97 |
// uint jcong=rotr(seed_w,blockIdx.x);
|
|
98 |
|
|
99 |
uint z=seed_z/(blockIdx.x+1);
|
|
100 |
uint w=seed_w%(blockIdx.x+1);
|
|
101 |
|
|
102 |
uint jsr=seed_z/(blockIdx.x+1);
|
|
103 |
uint jcong=seed_w%(blockIdx.x+1);
|
|
104 |
|
|
105 |
for (ulong i=0;i<iterations;i++) {
|
|
106 |
|
|
107 |
s[(uint)(((ulong)size*(ulong)CONG)/(ulong)MAX)]+=1;
|
|
108 |
|
|
109 |
}
|
|
110 |
__threadfence_block();
|
|
111 |
}
|
|
112 |
|
|
113 |
__global__ void MainLoopThreads(uint *s,uint size,ulong iterations,uint seed_w,uint seed_z)
|
|
114 |
{
|
|
115 |
// uint z=rotl(seed_z,threadIdx.x);
|
|
116 |
// uint w=rotr(seed_w,threadIdx.x);
|
|
117 |
|
|
118 |
// uint jsr=rotl(seed_z,threadIdx.x);
|
|
119 |
// uint jcong=rotr(seed_w,threadIdx.x);
|
|
120 |
|
|
121 |
uint z=seed_z;
|
|
122 |
uint w=seed_w;
|
|
123 |
|
|
124 |
uint jsr=seed_z;
|
|
125 |
uint jcong=seed_w;
|
|
126 |
|
|
127 |
for (ulong i=0;i<iterations;i++) {
|
|
128 |
|
|
129 |
s[(uint)(((ulong)size*(ulong)CONG)/(ulong)MAX)]+=1;
|
|
130 |
}
|
|
131 |
|
|
132 |
__syncthreads();
|
|
133 |
}
|
|
134 |
|
|
135 |
__global__ void MainLoopHybrid(uint *s,uint size,ulong iterations,uint seed_w,uint seed_z)
|
|
136 |
{
|
|
137 |
// uint z=rotl(seed_z,blockDim.x*blockIdx.x+threadIdx.x threadIdx.x);
|
|
138 |
// uint w=rotr(seed_w,blockDim.x*blockIdx.x+threadIdx.x);
|
|
139 |
|
|
140 |
// uint jsr=rotl(seed_z,blockDim.x*blockIdx.x+threadIdx.x);
|
|
141 |
// uint jcong=rotr(seed_w,blockDim.x*blockIdx.x+threadIdx.x);
|
|
142 |
|
|
143 |
uint z=seed_z;
|
|
144 |
uint w=seed_w;
|
|
145 |
|
|
146 |
uint jsr=seed_z;
|
|
147 |
uint jcong=seed_w;
|
|
148 |
|
|
149 |
for (ulong i=0;i<iterations;i++) {
|
|
150 |
|
|
151 |
s[(uint)(((ulong)size*(ulong)CONG)/(ulong)MAX)]+=1;
|
|
152 |
}
|
|
153 |
|
|
154 |
__syncthreads();
|
|
155 |
|
|
156 |
}
|
|
157 |
"""
|
|
158 |
|
|
159 |
KERNEL_CODE_OPENCL="""
|
|
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 |
|
|
164 |
#define MWC (znew+wnew)
|
|
165 |
#define SHR3 (jsr=(jsr=(jsr=jsr^(jsr<<17))^(jsr>>13))^(jsr<<5))
|
|
166 |
#define CONG (jcong=69069*jcong+1234567)
|
|
167 |
#define KISS ((MWC^CONG)+SHR3)
|
|
168 |
|
|
169 |
#define CONGfp CONG * 2.328306435454494e-10f
|
|
170 |
#define SHR3fp SHR3 * 2.328306435454494e-10f
|
|
171 |
#define MWCfp MWC * 2.328306435454494e-10f
|
|
172 |
#define KISSfp KISS * 2.328306435454494e-10f
|
|
173 |
|
|
174 |
#define MAX 4294967296
|
|
175 |
|
|
176 |
uint rotl(uint value, int shift) {
|
|
177 |
return (value << shift) | (value >> (sizeof(value) * CHAR_BIT - shift));
|
|
178 |
}
|
|
179 |
|
|
180 |
uint rotr(uint value, int shift) {
|
|
181 |
return (value >> shift) | (value << (sizeof(value) * CHAR_BIT - shift));
|
|
182 |
}
|
|
183 |
|
|
184 |
__kernel void MainLoopGlobal(__global uint *s,const uint space,const ulong iterations,const uint seed_w,const uint seed_z)
|
|
185 |
{
|
|
186 |
__private const float id=(float)get_global_id(0);
|
|
187 |
__private const float size=(float)get_global_size(0);
|
|
188 |
__private const float block=space/size;
|
|
189 |
|
|
190 |
__private uint z=seed_z;
|
|
191 |
__private uint w=seed_w;
|
|
192 |
|
|
193 |
__private uint jsr=seed_z;
|
|
194 |
__private uint jcong=seed_w;
|
|
195 |
|
|
196 |
for (__private ulong i=0;i<iterations;i++) {
|
|
197 |
|
|
198 |
// Standard version does not work for several processes (some lost!)
|
|
199 |
//__private uint position=(uint)(((ulong)space*(ulong)MWC)/(ulong)MAX);
|
|
200 |
// Dense version
|
|
201 |
__private uint position=(uint)( (ulong)space*((ulong)CONG+MAX*(ulong)id)/(ulong)size/(ulong)MAX );
|
|
202 |
// Float version seems to be the best...
|
|
203 |
//__private uint position=(uint)( block*(CONGfp+id) );
|
|
204 |
|
|
205 |
s[position]++;
|
|
206 |
}
|
|
207 |
|
|
208 |
barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);
|
|
209 |
|
|
210 |
}
|
|
211 |
|
|
212 |
__kernel void MainLoopLocal(__global uint *s,uint size,ulong iterations,uint seed_w,uint seed_z)
|
|
213 |
{
|
|
214 |
uint z=rotl(seed_z,get_local_id(0));
|
|
215 |
uint w=rotr(seed_w,get_local_id(0));
|
|
216 |
|
|
217 |
uint jsr=rotl(seed_z,get_local_id(0));
|
|
218 |
uint jcong=rotr(seed_w,get_local_id(0));
|
|
219 |
|
|
220 |
for (ulong i=0;i<iterations;i++) {
|
|
221 |
|
|
222 |
s[(int)(((ulong)size*(ulong)CONG)/(ulong)MAX)]+=(uint)1;
|
|
223 |
}
|
|
224 |
|
|
225 |
|
|
226 |
barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);
|
|
227 |
}
|
|
228 |
|
|
229 |
__kernel void MainLoopHybrid(__global uint *s,uint size,ulong iterations,uint seed_w,uint seed_z)
|
|
230 |
{
|
|
231 |
uint z=rotl(seed_z,get_group_id(0)*get_num_groups(0)+get_local_id(0));
|
|
232 |
uint w=rotr(seed_w,get_group_id(0)*get_num_groups(0)+get_local_id(0));
|
|
233 |
|
|
234 |
uint jsr=rotl(seed_z,get_group_id(0)*get_num_groups(0)+get_local_id(0));
|
|
235 |
uint jcong=rotr(seed_w,get_group_id(0)*get_num_groups(0)+get_local_id(0));
|
|
236 |
|
|
237 |
for (ulong i=0;i<iterations;i++) {
|
|
238 |
|
|
239 |
s[(int)(((ulong)size*(ulong)CONG)/(ulong)MAX)]+=1;
|
|
240 |
|
|
241 |
barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);
|
|
242 |
}
|
|
243 |
|
|
244 |
|
|
245 |
}
|
|
246 |
"""
|
|
247 |
|
|
248 |
def MetropolisCuda(circle,iterations,steps,jobs,ParaStyle):
|
|
249 |
|
|
250 |
# Avec PyCUDA autoinit, rien a faire !
|
|
251 |
|
|
252 |
circleCU = cuda.InOut(circle)
|
|
253 |
|
|
254 |
mod = SourceModule(KERNEL_CODE_CUDA)
|
|
255 |
|
|
256 |
MetropolisBlocksCU=mod.get_function("MainLoopBlocks")
|
|
257 |
MetropolisJobsCU=mod.get_function("MainLoopThreads")
|
|
258 |
MetropolisHybridCU=mod.get_function("MainLoopHybrid")
|
|
259 |
|
|
260 |
start = pycuda.driver.Event()
|
|
261 |
stop = pycuda.driver.Event()
|
|
262 |
|
|
263 |
MySplutter=numpy.zeros(steps)
|
|
264 |
MyDuration=numpy.zeros(steps)
|
|
265 |
|
|
266 |
if iterations%jobs==0:
|
|
267 |
iterationsCL=numpy.uint64(iterations/jobs)
|
|
268 |
else:
|
|
269 |
iterationsCL=numpy.uint64(iterations/jobs+1)
|
|
270 |
|
|
271 |
iterationsNew=iterationsCL*jobs
|
|
272 |
|
|
273 |
for i in range(steps):
|
|
274 |
|
|
275 |
Splutter=numpy.zeros(1024).astype(numpy.uint32)
|
|
276 |
|
|
277 |
print Splutter
|
|
278 |
|
|
279 |
SplutterCU = cuda.InOut(Splutter)
|
|
280 |
|
|
281 |
start.record()
|
|
282 |
start.synchronize()
|
|
283 |
if ParaStyle=='Blocks':
|
|
284 |
MetropolisBlocksCU(SplutterCU,
|
|
285 |
numpy.uint32(len(Splutter)),
|
|
286 |
numpy.uint64(iterationsCL),
|
|
287 |
numpy.uint32(nprnd(2**30/jobs)),
|
|
288 |
numpy.uint32(nprnd(2**30/jobs)),
|
|
289 |
grid=(jobs,1),
|
|
290 |
block=(1,1,1))
|
|
291 |
|
|
292 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \
|
|
293 |
(Alu,jobs,1,ParaStyle)
|
|
294 |
elif ParaStyle=='Hybrid':
|
|
295 |
threads=BestThreadsNumber(jobs)
|
|
296 |
MetropolisHybridCU(SplutterCU,
|
|
297 |
numpy.uint32(len(Splutter)),
|
|
298 |
numpy.uint64(iterationsCL),
|
|
299 |
numpy.uint32(nprnd(2**30/jobs)),
|
|
300 |
numpy.uint32(nprnd(2**30/jobs)),
|
|
301 |
grid=(jobs,1),
|
|
302 |
block=(threads,1,1))
|
|
303 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \
|
|
304 |
(Alu,jobs/threads,threads,ParaStyle)
|
|
305 |
else:
|
|
306 |
MetropolisJobsCU(SplutterCU,
|
|
307 |
numpy.uint32(len(Splutter)),
|
|
308 |
numpy.uint64(iterationsCL),
|
|
309 |
numpy.uint32(nprnd(2**30/jobs)),
|
|
310 |
numpy.uint32(nprnd(2**30/jobs)),
|
|
311 |
grid=(1,1),
|
|
312 |
block=(jobs,1,1))
|
|
313 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \
|
|
314 |
(Alu,jobs,1,ParaStyle)
|
|
315 |
stop.record()
|
|
316 |
stop.synchronize()
|
|
317 |
|
|
318 |
elapsed = start.time_till(stop)*1e-3
|
|
319 |
|
|
320 |
print Splutter,sum(Splutter)
|
|
321 |
MySplutter[i]=numpy.median(Splutter)
|
|
322 |
print numpy.mean(Splutter),MySplutter[i],numpy.std(Splutter)
|
|
323 |
|
|
324 |
MyDuration[i]=elapsed
|
|
325 |
|
|
326 |
#AllPi=4./numpy.float32(iterationsCL)*circle.astype(numpy.float32)
|
|
327 |
#MyPi[i]=numpy.median(AllPi)
|
|
328 |
#print MyPi[i],numpy.std(AllPi),MyDuration[i]
|
|
329 |
|
|
330 |
|
|
331 |
print jobs,numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration)
|
|
332 |
|
|
333 |
return(numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration))
|
|
334 |
|
|
335 |
|
|
336 |
def MetropolisOpenCL(circle,iterations,steps,jobs,ParaStyle,Alu,Device):
|
|
337 |
|
|
338 |
# Initialisation des variables en les CASTant correctement
|
|
339 |
|
|
340 |
if Device==0:
|
|
341 |
print "Enter XPU selector based on ALU type: first selected"
|
|
342 |
HasXPU=False
|
|
343 |
# Default Device selection based on ALU Type
|
|
344 |
for platform in cl.get_platforms():
|
|
345 |
for device in platform.get_devices():
|
|
346 |
deviceType=cl.device_type.to_string(device.type)
|
|
347 |
deviceMemory=device.max_mem_alloc_size
|
|
348 |
if deviceType=="GPU" and Alu=="GPU" and not HasXPU:
|
|
349 |
XPU=device
|
|
350 |
print "GPU selected: ",device.name
|
|
351 |
HasXPU=True
|
|
352 |
MemoryXPU=deviceMemory
|
|
353 |
if deviceType=="CPU" and Alu=="CPU" and not HasXPU:
|
|
354 |
XPU=device
|
|
355 |
print "CPU selected: ",device.name
|
|
356 |
HasXPU=True
|
|
357 |
MemoryXPU=deviceMemory
|
|
358 |
else:
|
|
359 |
print "Enter XPU selector based on device number & ALU type"
|
|
360 |
Id=1
|
|
361 |
HasXPU=False
|
|
362 |
# Primary Device selection based on Device Id
|
|
363 |
for platform in cl.get_platforms():
|
|
364 |
for device in platform.get_devices():
|
|
365 |
deviceType=cl.device_type.to_string(device.type)
|
|
366 |
deviceMemory=device.max_mem_alloc_size
|
|
367 |
if Id==Device and Alu==deviceType and HasXPU==False:
|
|
368 |
XPU=device
|
|
369 |
print "CPU/GPU selected: ",device.name
|
|
370 |
HasXPU=True
|
|
371 |
MemoryXPU=deviceMemory
|
|
372 |
Id=Id+1
|
|
373 |
if HasXPU==False:
|
|
374 |
print "No XPU #%i of type %s found in all of %i devices, sorry..." % \
|
|
375 |
(Device,Alu,Id-1)
|
|
376 |
return(0,0,0)
|
|
377 |
|
|
378 |
# Je cree le contexte et la queue pour son execution
|
|
379 |
ctx = cl.Context([XPU])
|
|
380 |
queue = cl.CommandQueue(ctx,
|
|
381 |
properties=cl.command_queue_properties.PROFILING_ENABLE)
|
|
382 |
|
|
383 |
# Je recupere les flag possibles pour les buffers
|
|
384 |
mf = cl.mem_flags
|
|
385 |
|
|
386 |
MetropolisCL = cl.Program(ctx,KERNEL_CODE_OPENCL).build(options = "-cl-mad-enable -cl-fast-relaxed-math")
|
|
387 |
#MetropolisCL = cl.Program(ctx,KERNEL_CODE_OPENCL).build()
|
|
388 |
|
|
389 |
MyDuration=numpy.zeros(steps)
|
|
390 |
|
|
391 |
if iterations%jobs==0:
|
|
392 |
iterationsCL=numpy.uint64(iterations/jobs)
|
|
393 |
else:
|
|
394 |
iterationsCL=numpy.uint64(iterations/jobs+1)
|
|
395 |
|
|
396 |
iterationsNew=numpy.uint64(iterationsCL*jobs)
|
|
397 |
|
|
398 |
print iterations,iterationsCL,iterationsNew
|
|
399 |
|
|
400 |
MySplutter=numpy.zeros(steps)
|
|
401 |
|
|
402 |
print 'toto ',2**(int)(numpy.log2(MemoryXPU/4)),(MemoryXPU/jobs/4)*jobs
|
|
403 |
|
|
404 |
for i in range(steps):
|
|
405 |
|
|
406 |
#Splutter=numpy.zeros(2**(int)(numpy.log2(MemoryXPU/4))).astype(numpy.uint32)
|
|
407 |
#Splutter=numpy.zeros(1024).astype(numpy.uint32)
|
|
408 |
|
|
409 |
#Splutter=numpy.zeros(jobs).astype(numpy.uint32)
|
|
410 |
Splutter=numpy.zeros(jobs*16).astype(numpy.uint32)
|
|
411 |
|
|
412 |
print Splutter,len(Splutter)
|
|
413 |
|
|
414 |
SplutterCL = cl.Buffer(ctx, mf.WRITE_ONLY|mf.COPY_HOST_PTR,hostbuf=Splutter)
|
|
415 |
|
|
416 |
if ParaStyle=='Blocks':
|
|
417 |
# Call OpenCL kernel
|
|
418 |
# (1,) is Global work size (only 1 work size)
|
|
419 |
# (1,) is local work size
|
|
420 |
# circleCL is lattice translated in CL format
|
|
421 |
# SeedZCL is lattice translated in CL format
|
|
422 |
# SeedWCL is lattice translated in CL format
|
|
423 |
# step is number of iterations
|
|
424 |
# CLLaunch=MetropolisCL.MainLoopGlobal(queue,(jobs,),None,
|
|
425 |
# SplutterCL,
|
|
426 |
# numpy.uint32(len(Splutter)),
|
|
427 |
# numpy.uint64(iterationsCL),
|
|
428 |
# numpy.uint32(nprnd(2**30/jobs)),
|
|
429 |
# numpy.uint32(nprnd(2**30/jobs)))
|
|
430 |
CLLaunch=MetropolisCL.MainLoopGlobal(queue,(jobs,),None,
|
|
431 |
SplutterCL,
|
|
432 |
numpy.uint32(len(Splutter)),
|
|
433 |
numpy.uint64(iterationsCL),
|
|
434 |
numpy.uint32(521288629),
|
|
435 |
numpy.uint32(362436069))
|
|
436 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \
|
|
437 |
(Alu,jobs,1,ParaStyle)
|
|
438 |
elif ParaStyle=='Hybrid':
|
|
439 |
threads=BestThreadsNumber(jobs)
|
|
440 |
# en OpenCL, necessaire de mettre un Global_id identique au local_id
|
|
441 |
CLLaunch=MetropolisCL.MainLoopHybrid(queue,(jobs,),(threads,),
|
|
442 |
SplutterCL,
|
|
443 |
numpy.uint32(len(Splutter)),
|
|
444 |
numpy.uint64(iterationsCL),
|
|
445 |
numpy.uint32(nprnd(2**30/jobs)),
|
|
446 |
numpy.uint32(nprnd(2**30/jobs)))
|
|
447 |
|
|
448 |
print "%s with (WorkItems/Threads)=(%i,%i) %s method done" % \
|
|
449 |
(Alu,jobs/threads,threads,ParaStyle)
|
|
450 |
else:
|
|
451 |
# en OpenCL, necessaire de mettre un Global_id identique au local_id
|
|
452 |
CLLaunch=MetropolisCL.MainLoopLocal(queue,(jobs,),(jobs,),
|
|
453 |
SplutterCL,
|
|
454 |
numpy.uint32(len(Splutter)),
|
|
455 |
numpy.uint64(iterationsCL),
|
|
456 |
numpy.uint32(nprnd(2**30/jobs)),
|
|
457 |
numpy.uint32(nprnd(2**30/jobs)))
|
|
458 |
print "%s with %i %s done" % (Alu,jobs,ParaStyle)
|
|
459 |
|
|
460 |
CLLaunch.wait()
|
|
461 |
cl.enqueue_copy(queue, Splutter, SplutterCL).wait()
|
|
462 |
|
|
463 |
elapsed = 1e-9*(CLLaunch.profile.end - CLLaunch.profile.start)
|
|
464 |
|
|
465 |
MyDuration[i]=elapsed
|
|
466 |
print Splutter,sum(Splutter)
|
|
467 |
MySplutter[i]=numpy.median(Splutter)
|
|
468 |
print numpy.mean(Splutter)*len(Splutter),MySplutter[i]*len(Splutter),numpy.std(Splutter)
|
|
469 |
|
|
470 |
SplutterCL.release()
|
|
471 |
|
|
472 |
print jobs,numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration)
|
|
473 |
|
|
474 |
return(numpy.mean(MyDuration),numpy.median(MyDuration),numpy.std(MyDuration))
|
|
475 |
|
|
476 |
|
|
477 |
def FitAndPrint(N,D,Curves):
|
|
478 |
|
|
479 |
from scipy.optimize import curve_fit
|
|
480 |
import matplotlib.pyplot as plt
|
|
481 |
|
|
482 |
try:
|
|
483 |
coeffs_Amdahl, matcov_Amdahl = curve_fit(Amdahl, N, D)
|
|
484 |
|
|
485 |
D_Amdahl=Amdahl(N,coeffs_Amdahl[0],coeffs_Amdahl[1],coeffs_Amdahl[2])
|
|
486 |
coeffs_Amdahl[1]=coeffs_Amdahl[1]*coeffs_Amdahl[0]/D[0]
|
|
487 |
coeffs_Amdahl[2]=coeffs_Amdahl[2]*coeffs_Amdahl[0]/D[0]
|
|
488 |
coeffs_Amdahl[0]=D[0]
|
|
489 |
print "Amdahl Normalized: T=%.2f(%.6f+%.6f/N)" % \
|
|
490 |
(coeffs_Amdahl[0],coeffs_Amdahl[1],coeffs_Amdahl[2])
|
|
491 |
except:
|
|
492 |
print "Impossible to fit for Amdahl law : only %i elements" % len(D)
|
|
493 |
|
|
494 |
try:
|
|
495 |
coeffs_AmdahlR, matcov_AmdahlR = curve_fit(AmdahlR, N, D)
|
|
496 |
|
|
497 |
D_AmdahlR=AmdahlR(N,coeffs_AmdahlR[0],coeffs_AmdahlR[1])
|
|
498 |
coeffs_AmdahlR[1]=coeffs_AmdahlR[1]*coeffs_AmdahlR[0]/D[0]
|
|
499 |
coeffs_AmdahlR[0]=D[0]
|
|
500 |
print "Amdahl Reduced Normalized: T=%.2f(%.6f+%.6f/N)" % \
|
|
501 |
(coeffs_AmdahlR[0],1-coeffs_AmdahlR[1],coeffs_AmdahlR[1])
|
|
502 |
|
|
503 |
except:
|
|
504 |
print "Impossible to fit for Reduced Amdahl law : only %i elements" % len(D)
|
|
505 |
|
|
506 |
try:
|
|
507 |
coeffs_Mylq, matcov_Mylq = curve_fit(Mylq, N, D)
|
|
508 |
|
|
509 |
coeffs_Mylq[1]=coeffs_Mylq[1]*coeffs_Mylq[0]/D[0]
|
|
510 |
# coeffs_Mylq[2]=coeffs_Mylq[2]*coeffs_Mylq[0]/D[0]
|
|
511 |
coeffs_Mylq[3]=coeffs_Mylq[3]*coeffs_Mylq[0]/D[0]
|
|
512 |
coeffs_Mylq[0]=D[0]
|
|
513 |
print "Mylq Normalized : T=%.2f(%.6f+%.6f/N)+%.6f*N" % (coeffs_Mylq[0],
|
|
514 |
coeffs_Mylq[1],
|
|
515 |
coeffs_Mylq[3],
|
|
516 |
coeffs_Mylq[2])
|
|
517 |
D_Mylq=Mylq(N,coeffs_Mylq[0],coeffs_Mylq[1],coeffs_Mylq[2],
|
|
518 |
coeffs_Mylq[3])
|
|
519 |
except:
|
|
520 |
print "Impossible to fit for Mylq law : only %i elements" % len(D)
|
|
521 |
|
|
522 |
try:
|
|
523 |
coeffs_Mylq2, matcov_Mylq2 = curve_fit(Mylq2, N, D)
|
|
524 |
|
|
525 |
coeffs_Mylq2[1]=coeffs_Mylq2[1]*coeffs_Mylq2[0]/D[0]
|
|
526 |
# coeffs_Mylq2[2]=coeffs_Mylq2[2]*coeffs_Mylq2[0]/D[0]
|
|
527 |
# coeffs_Mylq2[3]=coeffs_Mylq2[3]*coeffs_Mylq2[0]/D[0]
|
|
528 |
coeffs_Mylq2[4]=coeffs_Mylq2[4]*coeffs_Mylq2[0]/D[0]
|
|
529 |
coeffs_Mylq2[0]=D[0]
|
|
530 |
print "Mylq 2nd order Normalized: T=%.2f(%.6f+%.6f/N)+%.6f*N+%.6f*N^2" % \
|
|
531 |
(coeffs_Mylq2[0],coeffs_Mylq2[1],
|
|
532 |
coeffs_Mylq2[4],coeffs_Mylq2[2],coeffs_Mylq2[3])
|
|
533 |
|
|
534 |
except:
|
|
535 |
print "Impossible to fit for 2nd order Mylq law : only %i elements" % len(D)
|
|
536 |
|
|
537 |
if Curves:
|
|
538 |
plt.xlabel("Number of Threads/work Items")
|
|
539 |
plt.ylabel("Total Elapsed Time")
|
|
540 |
|
|
541 |
Experience,=plt.plot(N,D,'ro')
|
|
542 |
try:
|
|
543 |
pAmdahl,=plt.plot(N,D_Amdahl,label="Loi de Amdahl")
|
|
544 |
pMylq,=plt.plot(N,D_Mylq,label="Loi de Mylq")
|
|
545 |
except:
|
|
546 |
print "Fit curves seem not to be available"
|
|
547 |
|
|
548 |
plt.legend()
|
|
549 |
plt.show()
|
|
550 |
|
|
551 |
if __name__=='__main__':
|
|
552 |
|
|
553 |
# Set defaults values
|
|
554 |
|
|
555 |
# Alu can be CPU, GPU or ACCELERATOR
|
|
556 |
Alu='CPU'
|
|
557 |
# Id of GPU : 1 is for first find !
|
|
558 |
Device=0
|
|
559 |
# GPU style can be Cuda (Nvidia implementation) or OpenCL
|
|
560 |
GpuStyle='OpenCL'
|
|
561 |
# Parallel distribution can be on Threads or Blocks
|
|
562 |
ParaStyle='Blocks'
|
|
563 |
# Iterations is integer
|
|
564 |
Iterations=100000000
|
|
565 |
# JobStart in first number of Jobs to explore
|
|
566 |
JobStart=1
|
|
567 |
# JobEnd is last number of Jobs to explore
|
|
568 |
JobEnd=16
|
|
569 |
# JobStep is the step of Jobs to explore
|
|
570 |
JobStep=1
|
|
571 |
# Redo is the times to redo the test to improve metrology
|
|
572 |
Redo=1
|
|
573 |
# OutMetrology is method for duration estimation : False is GPU inside
|
|
574 |
OutMetrology=False
|
|
575 |
Metrology='InMetro'
|
|
576 |
# Curves is True to print the curves
|
|
577 |
Curves=False
|
|
578 |
# Fit is True to print the curves
|
|
579 |
Fit=False
|
|
580 |
|
|
581 |
try:
|
|
582 |
opts, args = getopt.getopt(sys.argv[1:],"hoclfa:g:p:i:s:e:t:r:d:",["alu=","gpustyle=","parastyle=","iterations=","jobstart=","jobend=","jobstep=","redo=","device="])
|
|
583 |
except getopt.GetoptError:
|
|
584 |
print '%s -o (Out of Core Metrology) -c (Print Curves) -f (Fit to Amdahl Law) -a <CPU/GPU/ACCELERATOR> -d <DeviceId> -g <CUDA/OpenCL> -p <Threads/Hybrid/Blocks> -i <Iterations> -s <JobStart> -e <JobEnd> -t <JobStep> -r <RedoToImproveStats> ' % sys.argv[0]
|
|
585 |
sys.exit(2)
|
|
586 |
|
|
587 |
for opt, arg in opts:
|
|
588 |
if opt == '-h':
|
|
589 |
print '%s -o (Out of Core Metrology) -c (Print Curves) -f (Fit to Amdahl Law) -a <CPU/GPU/ACCELERATOR> -d <DeviceId> -g <CUDA/OpenCL> -p <Threads/Hybrid/Blocks> -i <Iterations> -s <JobStart> -e <JobEnd> -t <JobStep> -r <RedoToImproveStats>' % sys.argv[0]
|
|
590 |
|
|
591 |
print "\nInformations about devices detected under OpenCL:"
|
|
592 |
# For PyOpenCL import
|
|
593 |
try:
|
|
594 |
import pyopencl as cl
|
|
595 |
Id=1
|
|
596 |
for platform in cl.get_platforms():
|
|
597 |
for device in platform.get_devices():
|
|
598 |
deviceType=cl.device_type.to_string(device.type)
|
|
599 |
deviceMemory=device.max_mem_alloc_size
|
|
600 |
print "Device #%i of type %s with memory %i : %s" % (Id,deviceType,deviceMemory,device.name)
|
|
601 |
Id=Id+1
|
|
602 |
|
|
603 |
print
|
|
604 |
sys.exit()
|
|
605 |
except ImportError:
|
|
606 |
print "Your platform does not seem to support OpenCL"
|
|
607 |
|
|
608 |
elif opt == '-o':
|
|
609 |
OutMetrology=True
|
|
610 |
Metrology='OutMetro'
|
|
611 |
elif opt == '-c':
|
|
612 |
Curves=True
|
|
613 |
elif opt == '-f':
|
|
614 |
Fit=True
|
|
615 |
elif opt in ("-a", "--alu"):
|
|
616 |
Alu = arg
|
|
617 |
elif opt in ("-d", "--device"):
|
|
618 |
Device = int(arg)
|
|
619 |
elif opt in ("-g", "--gpustyle"):
|
|
620 |
GpuStyle = arg
|
|
621 |
elif opt in ("-p", "--parastyle"):
|
|
622 |
ParaStyle = arg
|
|
623 |
elif opt in ("-i", "--iterations"):
|
|
624 |
Iterations = numpy.uint64(arg)
|
|
625 |
elif opt in ("-s", "--jobstart"):
|
|
626 |
JobStart = int(arg)
|
|
627 |
elif opt in ("-e", "--jobend"):
|
|
628 |
JobEnd = int(arg)
|
|
629 |
elif opt in ("-t", "--jobstep"):
|
|
630 |
JobStep = int(arg)
|
|
631 |
elif opt in ("-r", "--redo"):
|
|
632 |
Redo = int(arg)
|
|
633 |
|
|
634 |
if Alu=='CPU' and GpuStyle=='CUDA':
|
|
635 |
print "Alu can't be CPU for CUDA, set Alu to GPU"
|
|
636 |
Alu='GPU'
|
|
637 |
|
|
638 |
if ParaStyle not in ('Blocks','Threads','Hybrid'):
|
|
639 |
print "%s not exists, ParaStyle set as Threads !" % ParaStyle
|
|
640 |
ParaStyle='Threads'
|
|
641 |
|
|
642 |
print "Compute unit : %s" % Alu
|
|
643 |
print "Device Identification : %s" % Device
|
|
644 |
print "GpuStyle used : %s" % GpuStyle
|
|
645 |
print "Parallel Style used : %s" % ParaStyle
|
|
646 |
print "Iterations : %s" % Iterations
|
|
647 |
print "Number of threads on start : %s" % JobStart
|
|
648 |
print "Number of threads on end : %s" % JobEnd
|
|
649 |
print "Number of redo : %s" % Redo
|
|
650 |
print "Metrology done out of CPU/GPU : %r" % OutMetrology
|
|
651 |
|
|
652 |
if GpuStyle=='CUDA':
|
|
653 |
try:
|
|
654 |
# For PyCUDA import
|
|
655 |
import pycuda.driver as cuda
|
|
656 |
import pycuda.gpuarray as gpuarray
|
|
657 |
import pycuda.autoinit
|
|
658 |
from pycuda.compiler import SourceModule
|
|
659 |
except ImportError:
|
|
660 |
print "Platform does not seem to support CUDA"
|
|
661 |
|
|
662 |
if GpuStyle=='OpenCL':
|
|
663 |
try:
|
|
664 |
# For PyOpenCL import
|
|
665 |
import pyopencl as cl
|
|
666 |
Id=1
|
|
667 |
for platform in cl.get_platforms():
|
|
668 |
for device in platform.get_devices():
|
|
669 |
deviceType=cl.device_type.to_string(device.type)
|
|
670 |
print "Device #%i of type %s : %s" % (Id,deviceType,device.name)
|
|
671 |
if Id == Device:
|
|
672 |
# Set the Alu as detected Device Type
|
|
673 |
Alu=deviceType
|
|
674 |
Id=Id+1
|
|
675 |
except ImportError:
|
|
676 |
print "Platform does not seem to support CUDA"
|
|
677 |
|
|
678 |
average=numpy.array([]).astype(numpy.float32)
|
|
679 |
median=numpy.array([]).astype(numpy.float32)
|
|
680 |
stddev=numpy.array([]).astype(numpy.float32)
|
|
681 |
|
|
682 |
ExploredJobs=numpy.array([]).astype(numpy.uint32)
|
|
683 |
|
|
684 |
Jobs=JobStart
|
|
685 |
|
|
686 |
while Jobs <= JobEnd:
|
|
687 |
avg,med,std=0,0,0
|
|
688 |
ExploredJobs=numpy.append(ExploredJobs,Jobs)
|
|
689 |
circle=numpy.zeros(Jobs).astype(numpy.uint64)
|
|
690 |
|
|
691 |
if OutMetrology:
|
|
692 |
duration=numpy.array([]).astype(numpy.float32)
|
|
693 |
for i in range(Redo):
|
|
694 |
start=time.time()
|
|
695 |
if GpuStyle=='CUDA':
|
|
696 |
try:
|
|
697 |
a,m,s=MetropolisCuda(circle,Iterations,1,Jobs,ParaStyle)
|
|
698 |
except:
|
|
699 |
print "Problem with %i // computations on Cuda" % Jobs
|
|
700 |
elif GpuStyle=='OpenCL':
|
|
701 |
try:
|
|
702 |
a,m,s=MetropolisOpenCL(circle,Iterations,1,Jobs,ParaStyle,
|
|
703 |
Alu,Device)
|
|
704 |
except:
|
|
705 |
print "Problem with %i // computations on OpenCL" % Jobs
|
|
706 |
duration=numpy.append(duration,time.time()-start)
|
|
707 |
if (a,m,s) != (0,0,0):
|
|
708 |
avg=numpy.mean(duration)
|
|
709 |
med=numpy.median(duration)
|
|
710 |
std=numpy.std(duration)
|
|
711 |
else:
|
|
712 |
print "Values seem to be wrong..."
|
|
713 |
else:
|
|
714 |
if GpuStyle=='CUDA':
|
|
715 |
try:
|
|
716 |
avg,med,std=MetropolisCuda(circle,Iterations,Redo,Jobs,ParaStyle)
|
|
717 |
except:
|
|
718 |
print "Problem with %i // computations on Cuda" % Jobs
|
|
719 |
elif GpuStyle=='OpenCL':
|
|
720 |
try:
|
|
721 |
avg,med,std=MetropolisOpenCL(circle,Iterations,Redo,Jobs,ParaStyle,Alu,Device)
|
|
722 |
except:
|
|
723 |
print "Problem with %i // computations on OpenCL" % Jobs
|
|
724 |
|
|
725 |
if (avg,med,std) != (0,0,0):
|
|
726 |
print "jobs,avg,med,std",Jobs,avg,med,std
|
|
727 |
average=numpy.append(average,avg)
|
|
728 |
median=numpy.append(median,med)
|
|
729 |
stddev=numpy.append(stddev,std)
|
|
730 |
else:
|
|
731 |
print "Values seem to be wrong..."
|
|
732 |
#THREADS*=2
|
|
733 |
if len(average)!=0:
|
|
734 |
numpy.savez("Splutter_%s_%s_%s_%s_%i_%.8i_Device%i_%s_%s" % (Alu,GpuStyle,ParaStyle,JobStart,JobEnd,Iterations,Device,Metrology,gethostname()),(ExploredJobs,average,median,stddev))
|
|
735 |
ToSave=[ ExploredJobs,average,median,stddev ]
|
|
736 |
numpy.savetxt("Splutter_%s_%s_%s_%s_%i_%.8i_Device%i_%s_%s" % (Alu,GpuStyle,ParaStyle,JobStart,JobEnd,Iterations,Device,Metrology,gethostname()),numpy.transpose(ToSave))
|
|
737 |
Jobs+=JobStep
|
|
738 |
|
|
739 |
if Fit:
|
|
740 |
FitAndPrint(ExploredJobs,median,Curves)
|
|
741 |
|
| 0 |
742 |
|