def get_divergence_error(vector):
for mu in range(3):
fft.idft(vector[mu], vector_x[mu])
derivs.divergence(queue, vector_x, div)
derivs(queue, fx=vector_x[0], pdx=pdx[0])
derivs(queue, fx=vector_x[1], pdy=pdx[1])
derivs(queue, fx=vector_x[2], pdz=pdx[2])
norm = sum([clm.fabs(pdx[mu]) for mu in range(3)])
max_err = cla.max(clm.fabs(div)) / cla.max(norm)
avg_err = cla.sum(clm.fabs(div)) / cla.sum(norm)
return max_err, avg_err
def get_divergence_errors(hij):
max_errors = []
avg_errors = []
for i in range(1, 4):
for mu in range(3):
fft.idft(hij[tensor_id(i, mu + 1)], vector_x[mu])
derivs.divergence(queue, vector_x, div)
derivs(queue, fx=vector_x[0], pdx=pdx[0])
derivs(queue, fx=vector_x[1], pdy=pdx[1])
derivs(queue, fx=vector_x[2], pdz=pdx[2])
norm = sum([clm.fabs(pdx[mu]) for mu in range(3)])
max_errors.append(cla.max(clm.fabs(div)) / cla.max(norm))
avg_errors.append(cla.sum(clm.fabs(div)) / cla.sum(norm))
return np.array(max_errors), np.array(avg_errors)
def test_outoforderqueue_clmath(ctx_factory):
context = ctx_factory()
try:
queue = cl.CommandQueue(context,
properties=cl.command_queue_properties.OUT_OF_ORDER_EXEC_MODE_ENABLE)
except Exception:
pytest.skip("out-of-order queue not available")
a = np.random.rand(10**6).astype(np.dtype('float32'))
a_gpu = cl_array.to_device(queue, a)
# testing that clmath functions wait for and create events
b_gpu = clmath.fabs(clmath.sin(a_gpu * 5))
queue.finish()
b1 = b_gpu.get()
b = np.abs(np.sin(a * 5))
assert np.abs(b1 - b).mean() < 1e-5
def test_outoforderqueue_clmath(ctx_factory):
context = ctx_factory()
try:
queue = cl.CommandQueue(context,
properties=cl.command_queue_properties.
OUT_OF_ORDER_EXEC_MODE_ENABLE)
except Exception:
pytest.skip("out-of-order queue not available")
a = np.random.rand(10**6).astype(np.dtype('float32'))
a_gpu = cl_array.to_device(queue, a)
# testing that clmath functions wait for and create events
b_gpu = clmath.fabs(clmath.sin(a_gpu * 5))
queue.finish()
b1 = b_gpu.get()
b = np.abs(np.sin(a * 5))
assert np.abs(b1 - b).mean() < 1e-5
def test_spectral_poisson(ctx_factory, grid_shape, proc_shape, h, dtype,
timing=False):
if ctx_factory:
ctx = ctx_factory()
else:
ctx = ps.choose_device_and_make_context()
queue = cl.CommandQueue(ctx)
mpi = ps.DomainDecomposition(proc_shape, h, grid_shape=grid_shape)
rank_shape, _ = mpi.get_rank_shape_start(grid_shape)
fft = ps.DFT(mpi, ctx, queue, grid_shape, dtype)
L = (3, 5, 7)
dx = tuple(Li / Ni for Li, Ni in zip(L, grid_shape))
dk = tuple(2 * np.pi / Li for Li in L)
if h == 0:
def get_evals_2(k, dx):
return - k**2
derivs = ps.SpectralCollocator(fft, dk)
else:
from pystella.derivs import SecondCenteredDifference
get_evals_2 = SecondCenteredDifference(h).get_eigenvalues
derivs = ps.FiniteDifferencer(mpi, h, dx, stream=False)
solver = ps.SpectralPoissonSolver(fft, dk, dx, get_evals_2)
pencil_shape = tuple(ni + 2*h for ni in rank_shape)
statistics = ps.FieldStatistics(mpi, 0, rank_shape=rank_shape,
grid_size=np.product(grid_shape))
fx = cla.empty(queue, pencil_shape, dtype)
rho = clr.rand(queue, rank_shape, dtype)
rho -= statistics(rho)["mean"]
lap = cla.empty(queue, rank_shape, dtype)
rho_h = rho.get()
for m_squared in (0, 1.2, 19.2):
solver(queue, fx, rho, m_squared=m_squared)
fx_h = fx.get()
if h > 0:
fx_h = fx_h[h:-h, h:-h, h:-h]
derivs(queue, fx=fx, lap=lap)
diff = np.fabs(lap.get() - rho_h - m_squared * fx_h)
max_err = np.max(diff) / cla.max(clm.fabs(rho))
avg_err = np.sum(diff) / cla.sum(clm.fabs(rho))
max_rtol = 1e-12 if dtype == np.float64 else 1e-4
avg_rtol = 1e-13 if dtype == np.float64 else 1e-5
assert max_err < max_rtol and avg_err < avg_rtol, \
f"solution inaccurate for {h=}, {grid_shape=}, {proc_shape=}"
if timing:
from common import timer
time = timer(lambda: solver(queue, fx, rho, m_squared=m_squared), ntime=10)
if mpi.rank == 0:
print(f"poisson took {time:.3f} ms for {grid_shape=}, {proc_shape=}")
def test_step(ctx_factory, proc_shape, dtype, Stepper):
if proc_shape != (1, 1, 1):
pytest.skip("test step only on one rank")
if ctx_factory:
ctx = ctx_factory()
else:
ctx = ps.choose_device_and_make_context()
queue = cl.CommandQueue(ctx)
from pystella.step import LowStorageRKStepper
is_low_storage = LowStorageRKStepper in Stepper.__bases__
rank_shape = (1, 1, 8)
init_vals = np.linspace(1, 3, 8)
if is_low_storage:
y = cla.zeros(queue, rank_shape, dtype)
y[0, 0, :] = init_vals
y0 = y.copy()
else:
num_copies = Stepper.num_copies
y = cla.zeros(queue, (num_copies, ) + rank_shape, dtype)
y[0, 0, 0, :] = init_vals
y0 = y[0].copy()
dtlist = [.1, .05, .025]
for n in [-1., -2., -3., -4.]:
max_errs = {}
for dt in dtlist:
def sol(y0, t):
return ((-1 + n) * (-t + y0**(1 - n) / (-1 + n)))**(1 /
(1 - n))
_y = ps.Field("y")
rhs_dict = {_y: _y**n}
stepper = Stepper(rhs_dict,
dt=dt,
halo_shape=0,
rank_shape=rank_shape)
if is_low_storage:
y[0, 0, :] = init_vals
else:
y[0, 0, 0, :] = init_vals
t = 0
errs = []
while t < .1:
for s in range(stepper.num_stages):
stepper(s, queue=queue, y=y, filter_args=True)
t += dt
if is_low_storage:
errs.append(cla.max(clm.fabs(1. - sol(y0, t) / y)).get())
else:
errs.append(
cla.max(clm.fabs(1. - sol(y0, t) / y[0])).get())
max_errs[dt] = np.max(errs)
order = stepper.expected_order
print(f"{order=}, {n=}")
print(max_errs)
print([
max_errs[a] / max_errs[b] for a, b in zip(dtlist[:-1], dtlist[1:])
])
order = stepper.expected_order
rtol = dtlist[-1]**order if dtype == np.float64 else 1e-1
assert list(max_errs.values())[-1] < rtol, \
f"Stepper solution inaccurate for {n=}"
for a, b in zip(dtlist[:-1], dtlist[1:]):
assert max_errs[a] / max_errs[b] > .9 * (a/b)**order, \
f"Stepper convergence failing for {n=}"
def U_to_P(params, G, U, P, Pout=None, iter_max=None):
s = G.slices
sh = G.shapes
# Set some default parameters, but allow overrides
if iter_max is None:
if 'invert_iter_max' in params:
iter_max = params['invert_iter_max']
else:
iter_max = 8
if 'invert_err_tol' in params:
err_tol = params['invert_err_tol']
else:
err_tol = 1.e-8
if 'invert_iter_delta' in params:
delta = params['invert_iter_delta']
else:
delta = 1.e-5
if 'gamma_max' not in params:
params['gamma_max'] = 25
# Don't overwrite old memory by default, just allow using old values
# Caller can pass new/old memory but is responsible that it contain logical values if we don't update it
if Pout is None:
Pout = P.copy()
# Update the primitive B-fields
G.vecdivbygeom(params['queue'],
u=U[s.B3VEC],
g=G.gdet_d[Loci.CENT.value],
out=Pout[s.B3VEC])
# Cached constant quantities
global ncov, ncon, lgdet
if ncov is None:
# For later on
ncov = cl_array.zeros(params['queue'],
sh.grid_vector,
dtype=np.float64)
G.timesgeom(params['queue'],
u=cl_array.empty_like(ncov[0]).fill(1.0),
g=-G.lapse_d[Loci.CENT.value],
out=ncov[0])
ncon = G.raise_grid(ncov)
lgdet = G.lapse_d[Loci.CENT.value] / G.gdet_d[Loci.CENT.value]
# Eflag will indicate inversion failures
# Define a generic kernel so we can split out flagging in the future
# Use it to catch negative density early on
eflag = cl_array.zeros(params['queue'], sh.grid_scalar, dtype=np.int32)
global knl_set_eflag
if knl_set_eflag is None:
code = add_ghosts(
"""eflag[i,j,k] = if(var[i,j,k] < 0, flag, eflag[i,j,k])""")
knl_set_eflag = lp.make_kernel(
sh.isl_grid_scalar,
code, [
*scalarArrayArgs("eflag", dtype=np.int32),
*scalarArrayArgs("var"),
lp.ValueArg("flag", dtype=np.int32), ...
],
default_offset=lp.auto)
knl_set_eflag = tune_grid_kernel(knl_set_eflag,
sh.bulk_scalar,
ng=G.NG)
evt, _ = knl_set_eflag(params['queue'],
var=U[s.RHO],
eflag=eflag,
flag=-100)
# Convert from conserved variables to four-vectors
Bcon = cl_array.zeros(params['queue'], sh.grid_vector, dtype=np.float64)
G.vectimesgeom(params['queue'], u=U[s.B3VEC], g=lgdet, out=Bcon[1:])
Qcov = cl_array.empty_like(Bcon)
G.timesgeom(params['queue'], u=(U[s.UU] - U[s.RHO]), g=lgdet, out=Qcov[0])
G.vectimesgeom(params['queue'], u=U[s.U3VEC], g=lgdet, out=Qcov[1:])
Bcov = G.lower_grid(Bcon)
Qcon = G.raise_grid(Qcov)
# This will have fringes of zeros still!
Bsq = G.dot(Bcon, Bcov)
QdB = G.dot(Bcon, Qcov)
Qdotn = G.dot(Qcon, ncov)
Qsq = G.dot(Qcon, Qcov)
Qtsq = Qsq + Qdotn**2
Qtcon = cl_array.empty_like(Qcon)
for i in range(4):
Qtcon[i] = Qcon[i] + ncon[i] * Qdotn
# Set up eqn for W', the energy density
D = cl_array.zeros_like(Qsq)
G.timesgeom(params['queue'], u=U[s.RHO], g=lgdet, out=D)
Ep = -Qdotn - D
del Bcov, Qcon, Qcov
# Numerical rootfinding
# Take guesses from primitives
Wp = Wp_func(params, G, P, Loci.CENT, eflag)
# Trap on any failures so far if debugging. They're very rare.
if 'debug' in params and params['debug']:
if np.any(eflag.get()[s.bulk] != 0):
raise ValueError("Unexpected flag set!")
# Step around the guess & evaluate errors
h = delta * Wp # TODO stable enough? Need fancy subtraction from iharm3d?
errp = err_eqn(params, G, Bsq, D, Ep, QdB, Qtsq, Wp + h, eflag)
err = err_eqn(params, G, Bsq, D, Ep, QdB, Qtsq, Wp, eflag)
errm = err_eqn(params, G, Bsq, D, Ep, QdB, Qtsq, Wp - h, eflag)
# Preserve Wp/err before updating them below
Wp1 = Wp.copy()
err1 = err.copy()
global knl_utop_prep
if knl_utop_prep is None:
# TODO keep an accumulator here to avoid that costly any() call?
code = add_ghosts("""
# TODO put error/prep in here, not calling below
# Attempt a Halley/Muller/Bailey/Press step
dedW := (errp[i,j,k] - errm[i,j,k]) / (2 * h[i,j,k])
dedW2 := (errp[i,j,k] - 2.*err[i,j,k] + errm[i,j,k]) / (h[i,j,k]**2)
# Limit size of 2nd derivative correction
# Loopy trick common in HARM: define intermediate variables (xt, xt2, xt3...) to impose clipping
# This allows assignments or substitutions while keeping dependencies straight
ft := 0.5*err[i,j,k]*dedW2/(dedW**2)
ft2 := if(ft > 0.3, 0.3, ft)
f := if(ft2 < -0.3, -0.3, ft2)
# Limit size of step
dWt := -err[i,j,k] / dedW / (1. - f)
dWt2 := if(dWt < -0.5*Wp[i,j,k], -0.5*Wp[i,j,k], dWt)
dW := if(dWt2 > 2.0*Wp[i,j,k], 2.0*Wp[i,j,k], dWt2)
Wp[i,j,k] = Wp[i,j,k] + dW {id=wp}
# Guarantee we take one step in every bulk zone
stop_flag[i,j,k] = 0
# This would avoid the step where there's convergence, but would require taking 2*dW above or similar
#stop_flag[i,j,k] = (fabs(dW / Wp[i,j,k]) < err_tol) + \
# (fabs(err[i,j,k] / Wp[i,j,k]) < err_tol) {dep=wp,nosync=wp}
""")
knl_utop_prep = lp.make_kernel(
sh.isl_grid_scalar,
code, [
*scalarArrayArgs("err", "errp", "errm", "h", "Wp"),
*scalarArrayArgs("stop_flag", dtype=np.int8), ...
],
assumptions=sh.assume_grid,
seq_dependencies=True)
knl_utop_prep = lp.fix_parameters(knl_utop_prep, err_tol=err_tol)
knl_utop_prep = tune_grid_kernel(knl_utop_prep,
sh.bulk_scalar,
ng=G.NG)
print("Compiled utop_prep")
# Fill stop_flag with 1 so we don't have to worry about ghost zones taking steps
stop_flag = cl_array.empty(params['queue'], sh.grid_scalar,
dtype=np.int8).fill(1)
evt, _ = knl_utop_prep(params['queue'],
err=err,
errp=errp,
errm=errm,
h=h,
Wp=Wp,
stop_flag=stop_flag)
evt.wait()
err_eqn(params, G, Bsq, D, Ep, QdB, Qtsq, Wp, eflag, out=err)
# Iteration kernel for 1Dw solver
global knl_utop_iter
if knl_utop_iter is None:
code = add_ghosts("""
# Evaluate whether we need to do any of this
<> go = not(stop_flag[i,j,k]) {id=insn_go}
# Normal secant increment is dW. Limit guess to between 0.5 and 2 times current value
dWt := (Wp1[i,j,k] - Wp[i,j,k]) * err[i,j,k] / (err[i,j,k] - err1[i,j,k])
dWt2 := if(dWt < -0.5*Wp[i,j,k], -0.5*Wp[i,j,k], dWt)
<> dW = if(dWt2 > 2.0*Wp[i,j,k], 2.0*Wp[i,j,k], dWt2) {id=dw,if=go}
# Preserve last values, after use but before any changes
Wp1[i,j,k] = Wp[i,j,k] {id=wp1,nosync=dw,if=go}
err1[i,j,k] = err[i,j,k] {nosync=dw,if=go}
# Update Wp. Err will be updated outside kernel
Wp[i,j,k] = Wp[i,j,k] + dW {id=wp,nosync=dw:wp1,if=go}
# Set flag not to continue in zones that have converged
stop_flag[i,j,k] = if(fabs(dW / Wp[i,j,k]) < err_tol, 1, stop_flag[i,j,k]) {nosync=dw:wp:insn_go,if=go}
# For the future, when we've defined err_eqn for loopy kernels
# err = err_eqn(Bsq, D, Ep, QdB, Qtsq, Wp, gam, gamma_max, eflag) {if=go}
# stop_flag[i,j,k] = stop_flag[i,j,k] + (fabs(err[] / Wp[]) < err_tol) {if=go}
""")
knl_utop_iter = lp.make_kernel(
sh.isl_grid_scalar,
code, [
*scalarArrayArgs("Wp", "Wp1", "err", "err1"),
*scalarArrayArgs("stop_flag", dtype=np.int8), ...
],
assumptions=sh.assume_grid,
default_offset=lp.auto,
seq_dependencies=True)
knl_utop_iter = lp.fix_parameters(knl_utop_iter, err_tol=err_tol)
knl_utop_iter = tune_grid_kernel(knl_utop_iter,
sh.bulk_scalar,
ng=G.NG)
print("Compiled utop_iter")
# Iterate at least once to set new values from first step
# TODO Needed now we set Wp, err1 right?
for niter in range(iter_max):
#print("U_to_P iter")
evt, _ = knl_utop_iter(params['queue'],
Wp=Wp,
Wp1=Wp1,
err=err,
err1=err1,
stop_flag=stop_flag)
err = err_eqn(params, G, Bsq, D, Ep, QdB, Qtsq, Wp, eflag)
# TODO there may be better/faster if/reduction statements here...
stop_flag |= (clm.fabs(err / Wp) < err_tol)
if cl_array.min(stop_flag) >= 1:
break
# If secant method failed to converge, do not set primitives other than B
eflag += (stop_flag == 0)
del Wp1, err, err1, stop_flag
# Find utsq, gamma, rho0 from Wp
gamma = gamma_func(params, G, Bsq, D, QdB, Qtsq, Wp, eflag)
if 'debug' in params and params['debug']:
if np.any(gamma.get()[s.bulk] < 1.):
raise ValueError("gamma < 1 failure!")
# Find the scalars
global knl_utop_set
if knl_utop_set is None:
code = add_ghosts(
replace_prim_names("""
rho0 := D[i,j,k] / gamma[i,j,k]
W := Wp[i,j,k] + D[i,j,k]
w := W / (gamma[i,j,k]**2)
pres := (w - rho0) * (gam - 1.) / gam
u := w - (rho0 + pres)
# Set flag if prims are < 0
eflag[i,j,k] = if((u < 0)*(rho0 < 0), 8, if(u < 0, 7, if(rho0 < 0, 6, eflag[i,j,k]))) {id=ef}
# Don't update flagged primitives (necessary? Could skip the branch if fixup does ok)
<> set = not(eflag[i,j,k]) {dep=ef,nosync=ef}
P[RHO,i,j,k] = rho0 {if=set}
P[UU,i,j,k] = u {if=set}
P[U1,i,j,k] = (gamma[i,j,k] / (W + Bsq[i,j,k])) * (Qtcon[1,i,j,k] + QdB[i,j,k] * Bcon[1,i,j,k] / W) {if=set}
P[U2,i,j,k] = (gamma[i,j,k] / (W + Bsq[i,j,k])) * (Qtcon[2,i,j,k] + QdB[i,j,k] * Bcon[2,i,j,k] / W) {if=set}
P[U3,i,j,k] = (gamma[i,j,k] / (W + Bsq[i,j,k])) * (Qtcon[3,i,j,k] + QdB[i,j,k] * Bcon[3,i,j,k] / W) {if=set}
"""))
if 'electrons' in params and params['electrons']:
code += add_ghosts("""
P[KEL,i,j,k] = U[KEL,i,j,k]/U[RHO,i,j,k]
P[KTOT,i,j,k] = U[KTOT,i,j,k]/U[RHO,i,j,k]
""")
knl_utop_set = lp.make_kernel(
sh.isl_grid_scalar,
code, [
*primsArrayArgs("P"), *vecArrayArgs("Qtcon", "Bcon"),
*scalarArrayArgs("D", "gamma", "Wp", "Bsq", "QdB"),
*scalarArrayArgs("eflag", dtype=np.int32), ...
],
assumptions=sh.assume_grid,
default_offset=lp.auto)
knl_utop_set = lp.fix_parameters(knl_utop_set,
gam=params['gam'],
nprim=params['n_prim'],
ndim=4)
knl_utop_set = tune_grid_kernel(knl_utop_set, sh.bulk_scalar, ng=G.NG)
print("Compiled utop_set")
evt, _ = knl_utop_set(params['queue'],
P=Pout,
Qtcon=Qtcon,
Bcon=Bcon,
D=D,
gamma=gamma,
Wp=Wp,
Bsq=Bsq,
QdB=QdB,
eflag=eflag)
evt.wait()
del Qtcon, Bcon, D, gamma, Wp, Bsq, QdB
# Trap on flags early in test problems
if 'debug' in params and params['debug']:
n_nonzero = np.count_nonzero(eflag.get())
if n_nonzero > 0:
print("Nonzero eflag in bulk: {}\nFlags: {}".format(
n_nonzero, np.argwhere(eflag.get() != 0)))
return Pout, eflag
