gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Input-dependent models now activated for all the operations

Browse Source
This commit is contained in:
Philippe Tillet
2014-10-04 08:58:11 +02:00
parent 044419f9f0
commit fc8b450a7c
5 changed files with 89 additions and 90 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
autotune/config.ini
View File

@@ -4,16 +4,16 @@ tmp-folder = /tmp/
[vector-axpy]
devices = 0
precision = single
size = 10000000
#~ size = 10000000
#~
#~ [matrix-axpy]
#~ devices = 0
#~ precision = all
#~ precision = single
#~ size = 3072, 3072
#~
#~ [row-wise-reduction]
#~ devices = 0
#~ precision = all
#~ precision = single
#~ layout = N, T
#~ size = 3968, 3968
@@ -21,4 +21,4 @@ size = 10000000
devices = 0
precision = single
layout = NT
size = 1536, 1536, 1536
#size = 1536, 1536, 1536

40
autotune/python/autotune.py
View File

@@ -8,6 +8,7 @@ from external.configobj import ConfigObj
import pyopencl as cl
import pyviennacl as vcl
import numpy as np
from pyviennacl import backend
from pyviennacl import opencl
from pyviennacl import atidlas
@@ -73,32 +74,45 @@ def do_tuning(config_fname, spec_fname, viennacl_root):
with open(fname, "w+") as archive:
return optimize.genetic(statement, device, TYPES[operation]['template'], lambda p: TYPES[operation]['template'](p, *other_params),
lambda t: TYPES[operation]['perf-index']([datatype().itemsize, sizes, t]), TYPES[operation]['perf-measure'], archive)
#Helper
def tune(execution_handler, nTuning, nDataPoints, draw):
if 'size' in p:
profile = execution_handler(map_to_list(int, p['size']))
else:
def compute_perf(x, t):
return TYPES[operation]['perf-index']([datatype().itemsize, x, t])
X, Y, profiles = generate_dataset(TYPES[operation]['template'], execution_handler, nTuning, nDataPoints, compute_perf, draw)
train_model(X, Y, profiles, TYPES[operation]['perf-measure'])
#Vector AXPY
if operation=='vector-axpy':
def execution_handler(sizes, fname=os.devnull, parameters=None):
x = vcl.Vector(sizes[0], context=ctx, dtype=datatype)
y = vcl.Vector(sizes[0], context=ctx, dtype=datatype)
return execute(device, vcl.Statement(vcl.ElementProd(vcl.exp(x + y),vcl.cos(x + y))), (), sizes, fname, parameters)
if 'size' in p:
profile = execution_handler(map_to_list(int, p['size']))
tune(execution_handler, 50, 10000, lambda : 64*np.random.randint(low=10, high=100000, size=1))
#Matrix AXPY
if operation=='matrix-axpy':
A = vcl.Matrix(s, context=ctx, dtype=datatype)
B = vcl.Matrix(s, context=ctx, dtype=datatype)
execute(A+B, ())
def execution_handler(sizes, fname=os.devnull, parameters=None):
A = vcl.Matrix(sizes, context=ctx, dtype=datatype)
B = vcl.Matrix(sizes, context=ctx, dtype=datatype)
return execute(device, vcl.Statement(A+B), (), sizes, fname, parameters)
tune(execution_handler, 50, 10000, lambda : 64*np.random.randint(low=5, high=100, size=2))
#Row-wise reduction
if operation=='row-wise-reduction':
layouts = map_to_list((str,p['layout']))
layouts = map_to_list(str,p['layout'])
if 'all' in layouts:
layouts = ['N', 'T']
for A_trans in layouts:
A = vcl.Matrix(s if A_trans=='N' else s[::-1], context=ctx, dtype=datatype, layout=vcl.COL_MAJOR)
x = vcl.Vector(s[1] if A_trans=='N' else s[0], context=ctx, dtype=datatype)
def execution_handler(sizes, fname=os.devnull, parameters=None):
A = vcl.Matrix(sizes if A_trans=='N' else sizes[::-1], context=ctx, dtype=datatype, layout=vcl.COL_MAJOR)
x = vcl.Vector(sizes[1] if A_trans=='N' else sizes[0], context=ctx, dtype=datatype)
LHS = A if A_trans=='N' else A.T
execute(LHS*x, ())
execute(device, vcl.Statement(LHS*x), (), sizes, fname, parameters)
tune(execution_handler, 50, 10000, lambda : 64*np.random.randint(low=5, high=100, size=2))
#Matrix Product
if operation=='matrix-product':
layouts = map_to_list((str,p['layout']))
layouts = map_to_list(str,p['layout'])
if 'all' in layouts:
layouts = ['NN', 'NT', 'TN', 'TT']
for layout in layouts:
@@ -114,11 +128,7 @@ def do_tuning(config_fname, spec_fname, viennacl_root):
C = vcl.Matrix((sizes[0], sizes[2]), context=ctx, dtype = datatype, layout=vcl.COL_MAJOR)
statement = vcl.Statement(vcl.Assign(C,LHS*RHS*alpha + C*beta))
return execute(device, statement,(A_trans, B_trans), sizes, fname, parameters)
if 'size' in p:
profile = execution_handler(map(int, p['size']))
else:
X, Y, profiles = generate_dataset(TYPES[operation]['template'], execution_handler)
train_model(X, Y, profiles)
tune(execution_handler, 50, 10000, lambda : 64*np.random.randint(low=1, high=40, size=3))

102
autotune/python/dataset.py
View File

@@ -6,71 +6,59 @@ import numpy as np
from sklearn.neighbors.kde import KernelDensity
from pyviennacl.atidlas import FetchingPolicy
def decode(y):
fetch = [FetchingPolicy.FETCH_FROM_LOCAL, FetchingPolicy.FETCH_FROM_GLOBAL_CONTIGUOUS, FetchingPolicy.FETCH_FROM_GLOBAL_STRIDED]
y[7] = fetch[y[7]]
y[8] = fetch[y[8]]
return y
def resample(X, tbincount, densities, step):
def resample(X, draw):
Xtuples = [tuple(x) for x in X]
r = random.random()
while(True):
if(len(tbincount)==0 or len(densities)==0 or r<=1.0/len(densities)):
x = np.array([step*random.randint(1,40), step*random.randint(1,40), step*random.randint(1,40)])
else:
probs = [1.0/x if x>0 else 0 for x in tbincount]
distr = np.random.choice(range(tbincount.size), p = probs/np.sum(probs))
x = densities[distr].sample()[0]
x = np.maximum(np.ones(x.shape),(x - step/2).astype(int)/step + 1)*step
x = draw()
if tuple(x) not in Xtuples:
break
return x.astype(int)
def generate_dataset(TemplateType, execution_handler):
I = 50
step = 64
path = "./data"
def generate_dataset(TemplateType, execution_handler, nTuning, nDataPoints, compute_perf, draw):
# print "Getting some good profiles..."
# X = np.empty((I, 3))
# t = np.empty(I)
# profiles = []
# for i in range(I):
# x = resample(X, [], [], step)
# y = execution_handler(x)
# if y not in profiles:
# profiles.append(y)
# idx = profiles.index(y)
# X[i,:] = x
# t[i] = idx
# densities = [KernelDensity(kernel='gaussian', bandwidth=2*step).fit(X[t==i,:]) for i in range(int(max(t))+1)];
#
# print "Generating the dataset..."
# N = 10000
# Y = np.empty((N, len(profiles)))
# X = np.empty((N,3))
# t = []
#
# for i in range(N):
# x = resample(X, [], [], step)
# for j,y in enumerate(profiles):
# T = execution_handler(x, os.devnull, decode(map(int, y)))
# Y[i,j] = 2*1e-9*x[0]*x[1]*x[2]/T
# idx = np.argmax(Y[i,:])
# X[i,:] = x
# t = np.argmax(Y[:i+1,], axis=1)
# densities[idx].fit(X[t==idx,:])
# if i%10==0:
# sys.stdout.write('%d data points generated\r'%i)
# sys.stdout.flush()
#
# np.savetxt(os.path.join(path,"profiles.csv"), profiles)
# np.savetxt(os.path.join(path,"X.csv"), X)
# np.savetxt(os.path.join(path,"Y.csv"), Y)
print "Getting some good profiles..."
nDim = draw().size
X = np.empty((nTuning, nDim))
t = np.empty(nTuning)
profiles = []
for i in range(nTuning):
x = resample(X, draw)
y = execution_handler(x)
if y not in profiles:
profiles.append(y)
idx = profiles.index(y)
X[i,:] = x
t[i] = idx
profiles = np.loadtxt(os.path.join(path,"profiles.csv"))
X = np.loadtxt(os.path.join(path,"X.csv"))
Y = np.loadtxt(os.path.join(path,"Y.csv"))
print "Generating the dataset..."
Y = np.empty((nDataPoints, len(profiles)))
X = np.empty((nDataPoints, nDim))
t = []
for i in range(nDataPoints):
x = resample(X, draw)
for j,y in enumerate(profiles):
T = execution_handler(x, os.devnull, y)
Y[i,j] = compute_perf(x, T)
idx = np.argmax(Y[i,:])
X[i,:] = x
t = np.argmax(Y[:i+1,], axis=1)
if i%10==0:
sys.stdout.write('%d data points generated\r'%i)
sys.stdout.flush()
template_name = TemplateType.__name__
dir = os.path.join("data", template_name)
if not os.path.exists(dir):
os.makedirs(dir)
np.savetxt(os.path.join(dir,"profiles.csv"), profiles)
np.savetxt(os.path.join(dir,"X.csv"), X)
np.savetxt(os.path.join(dir,"Y.csv"), Y)
profiles = np.loadtxt(os.path.join(dir, "profiles.csv"))
X = np.loadtxt(os.path.join(dir, "X.csv"),ndmin=2)
Y = np.loadtxt(os.path.join(dir, "Y.csv"),ndmin=2)
return X, Y, profiles

15
autotune/python/genetic.py
View File

@@ -40,13 +40,15 @@ class GeneticOperators(object):
self.ParameterType = TemplateType.Parameters
self.build_template = build_template
self.cache = {}
self.indpb = 0.05
self.out = out
self.genome_info = {
vcl.atidlas.VectorAxpyTemplate: [3,4,4,vcl.atidlas.FetchingPolicy],
vcl.atidlas.MatrixAxpyTemplate: [3,3,3,3,3,vcl.atidlas.FetchingPolicy],
vcl.atidlas.RowWiseReductionTemplate: [3,3,3,4,vcl.atidlas.FetchingPolicy],
vcl.atidlas.MatrixProductTemplate: [3,3,3,3,3,3,3,vcl.atidlas.FetchingPolicy,vcl.atidlas.FetchingPolicy,3]
}[TemplateType]
self.indpb = 1.0/sum([1 if x==vcl.atidlas.FetchingPolicy else x for x in self.genome_info])
creator.create("FitnessMin", base.Fitness, weights=(-1.0,))
creator.create("Individual", list, fitness=creator.FitnessMin)
@@ -149,7 +151,7 @@ class GeneticOperators(object):
ind.fitness.values = fit
hof.update(population)
while time.time() - start_time < maxtime:
while time.time() - start_time < maxtime and gen < maxgen:
# Vary the population
offspring = []
for _ in xrange(mu):
@@ -166,9 +168,8 @@ class GeneticOperators(object):
offspring.append(ind)
else: # Apply reproduction
offspring.append(random.choice(population))
#~ for x in offspring:
#~ print self.decode(x)
#for x in offspring:
#print self.decode(x)
# Evaluate the individuals with an invalid fitness
invalid_ind = [ind for ind in offspring if not ind.fitness.valid]
fitnesses = self.toolbox.map(self.evaluate, invalid_ind)
@@ -180,9 +181,9 @@ class GeneticOperators(object):
population[:] = self.toolbox.select(population + offspring, mu)
#Update
gen = gen + 1
best_profile = '(%s)'%','.join(map(str,self.decode(hof[0])));
best_profile = '(%s)'%','.join(map(str,self.decode(hof[0])))
best_performance = compute_perf(hof[0].fitness.values[0])
sys.stdout.write('Time %d | Best %d %s [ for %s ]\r'%(time.time() - start_time, best_performance, perf_metric, best_profile))
sys.stdout.write('Generation %d | Time %d | Best %d %s [ for %s ]\r'%(gen, time.time() - start_time, best_performance, perf_metric, best_profile))
sys.stdout.flush()
sys.stdout.write('\n')
return self.decode(hof[0])

8
autotune/python/model.py
View File

@@ -8,7 +8,7 @@ from pybrain.supervised.trainers import BackpropTrainer
from pybrain.structure import LinearLayer, TanhLayer, SigmoidLayer, SoftmaxLayer, FeedForwardNetwork, BiasUnit
from pybrain.tools.neuralnets import NNregression, Trainer
def train_model(X, Y, profiles):
def train_model(X, Y, profiles, metric):
#Preprocessing
Xmean = np.mean(X, axis=0)
Xstd = np.std(X, axis=0)
@@ -43,7 +43,7 @@ def train_model(X, Y, profiles):
np.set_printoptions(precision=2)
print("-----------------")
print("Average testing speedup : %f (Optimal : %f)"%(sp.stats.gmean(speedups), sp.stats.gmean(optspeedups)))
print("Average GFLOP/s : %f (Default %f, Optimal %f)"%(np.mean(np.multiply(GFlops,speedups)), np.mean(GFlops), np.mean(np.multiply(GFlops,optspeedups))))
print("Minimum speedup is %f wrt %i GFlops"%(np.min(speedups), GFlops[np.argmin(speedups)]))
print("Maximum speedup is %f wrt %i GFlops for %s"%(np.max(speedups), GFlops[np.argmax(speedups)], X[np.argmax(speedups)]*Xstd+Xmean))
print("Average %s: %f (Default %f, Optimal %f)"%(metric, np.mean(np.multiply(GFlops,speedups)), np.mean(GFlops), np.mean(np.multiply(GFlops,optspeedups))))
print("Minimum speedup is %f wrt %i %s"%(np.min(speedups), GFlops[np.argmin(speedups)], metric))
print("Maximum speedup is %f wrt %i %s"%(np.max(speedups), GFlops[np.argmax(speedups)], metric))
print("--------")