def __init__(self, ic=None, nthr=256):
super(CudaSim, self).__init__()
# def gpu(gsc, exc, vel
# cat=concatenate):
self.nthr = nthr
self.idel = (self.dataset.distances / self.vel / self.dt).astype(int32)
self.horizon = self.idel.max() + 1
self.hist = zeros((self.horizon, self.n, self.nthr)) + 0.1
self.conn = self.dataset.weights
self.state = ic or 0.1 * ones((self.n, self.nsv, self.nthr))
self.exc = self.exc * ones((self.nthr, ))
self.gsc = self.gsc * ones((self.nthr, )) / self.n
genc.RPointer.gpu = True
src = genc.module(genc.model(dt=self.dt,
noise=False,
gpu=True,
**genc.fhn),
genc.step(self.n,
len(genc.fhn['eqns']),
model=genc.fhn['name'],
noise=False,
gpu=True),
genc.wrap(self.horizon, gpu=True),
gpu=True)
self.mod = cuda.srcmod(src, ['step'], debug=False)
def __init__(self, ic=None):
super(CeeSim, self).__init__()
C = ascontiguousarray
idel = (self.dataset.distances / self.vel / self.dt).astype(int64)
self.idel = C(idel)
self.horizon = self.idel.max() + 1
self.conn = C(self.dataset.weights)
self.gsc = C(array([self.gsc])) / self.n
self.exc = C(array([self.exc]))
self.hist = C(zeros((self.horizon, self.n)) + 0.1)
self.state = C(ic or 0.1 * ones((self.n, 2)))
src = genc.module(
genc.model(dt=self.dt, **genc.fhn),
genc.step(self.n, len(genc.fhn['eqns']), model=genc.fhn['name']),
genc.wrap(self.horizon))
self.mod = ccpp.srcmod(src, ['step'], debug=False)
for arr_name in ['idel', 'conn', 'hist', 'state', 'gsc', 'exc']:
arr = getattr(self, arr_name)
ptr = arr.ctypes.data_as(ctypes.c_void_p)
setattr(self, 'p_%s' % (arr_name, ), ptr)
def gpu(gsc, exc, vel, dt, dataset, tf=1500, ds=80, model="fhn_euler", cvar=0,
kblock=128, ublock=1024, cat=concatenate):
ts = r_[0:tf:dt]
n = dataset.weights.shape[0]
nsv = model_nsvs[model]
nthr = len(gsc)
npad = nthr % ublock if nthr%kblock or nthr%ublock and nthr > ublock else 0
nthr += npad
idel = (dataset.distances/vel/dt).astype(int32)
hist = random.uniform(low=-1., high=1., size=(idel.max()+1, n, nthr))
conn = dataset.weights
X = random.uniform(low=-2, high=2, size=(n, nsv, nthr))
Xs = empty((1+len(ts)/ds, n, nsv, nthr), dtype=float32)
hist[-1, ...] = X[:, cvar, :]
"""
mod = srcmod('parsweep.cu', ['kernel', 'update'],
horizon=idel.max()+1, dt=dt, ds=ds, n=n, cvar=cvar, model=model, nsv=nsv)
"""
# todo, delay eval, spec exec context ONCE
gen.RPointer.gpu = True
mod = srcmod(
gen.module(gen.model(dt=dt, noise=False, gpu=True, **gen.fhn),
gen.step(n, len(gen.fhn['eqns']), model=gen.fhn['name'],
noise=False, gpu=True),
gen.wrap(idel.max() + 1, gpu=True), gpu=True),
['step'],
debug=False)
with arrays_on_gpu(_timed=False, _memdebug=True, idel=idel.astype(int32),
hist=hist.astype(float32), conn=conn.astype(float32),
X=X.astype(float32), exc=cat((exc, zeros((npad,)))).astype(float32),
gsc=cat((gsc, zeros((npad,)))).astype(float32)) as g:
Xs[0, ...] = g.X.get()
for step, t in enumerate(ts):
mod.step(int32(step), g.idel, g.hist, g.conn, g.X, g.gsc, g.exc,
block=(kblock, 1, 1), grid=(nthr/kblock, 1))
# maybe separate again for perf & to be sure no thread mixing...
"""
mod.update(int32(step), g.hist, g.X,
block=(ublock if nthr>=ublock else nthr, 1, 1),
grid=(nthr/ublock if nthr/ublock > 0 else 1, 1))
"""
if step%ds == 0 and not (1+step/ds)>=len(Xs):
Xs[1+step/ds, ...] = g.X.get()
Xs = rollaxis(Xs, 3)
return Xs[:-npad] if npad else Xs
def gpu(gsc,
exc,
vel,
dt,
dataset,
tf=1500,
ds=80,
model="fhn_euler",
cvar=0,
kblock=128,
ublock=1024,
cat=concatenate):
ts = r_[0:tf:dt]
n = dataset.weights.shape[0]
nsv = model_nsvs[model]
nthr = len(gsc)
npad = nthr % ublock if nthr % kblock or nthr % ublock and nthr > ublock else 0
nthr += npad
idel = (dataset.distances / vel / dt).astype(int32)
hist = random.uniform(low=-1., high=1., size=(idel.max() + 1, n, nthr))
conn = dataset.weights
X = random.uniform(low=-2, high=2, size=(n, nsv, nthr))
Xs = empty((1 + len(ts) / ds, n, nsv, nthr), dtype=float32)
hist[-1, ...] = X[:, cvar, :]
"""
mod = srcmod('parsweep.cu', ['kernel', 'update'],
horizon=idel.max()+1, dt=dt, ds=ds, n=n, cvar=cvar, model=model, nsv=nsv)
"""
# todo, delay eval, spec exec context ONCE
gen.RPointer.gpu = True
mod = srcmod(gen.module(gen.model(dt=dt, noise=False, gpu=True, **gen.fhn),
gen.step(n,
len(gen.fhn['eqns']),
model=gen.fhn['name'],
noise=False,
gpu=True),
gen.wrap(idel.max() + 1, gpu=True),
gpu=True), ['step'],
debug=False)
with arrays_on_gpu(_timed=False,
_memdebug=True,
idel=idel.astype(int32),
hist=hist.astype(float32),
conn=conn.astype(float32),
X=X.astype(float32),
exc=cat((exc, zeros((npad, )))).astype(float32),
gsc=cat((gsc, zeros((npad, )))).astype(float32)) as g:
Xs[0, ...] = g.X.get()
for step, t in enumerate(ts):
mod.step(int32(step),
g.idel,
g.hist,
g.conn,
g.X,
g.gsc,
g.exc,
block=(kblock, 1, 1),
grid=(nthr / kblock, 1))
# maybe separate again for perf & to be sure no thread mixing...
"""
mod.update(int32(step), g.hist, g.X,
block=(ublock if nthr>=ublock else nthr, 1, 1),
grid=(nthr/ublock if nthr/ublock > 0 else 1, 1))
"""
if step % ds == 0 and not (1 + step / ds) >= len(Xs):
Xs[1 + step / ds, ...] = g.X.get()
Xs = rollaxis(Xs, 3)
return Xs[:-npad] if npad else Xs