a_mic = gd.empty(nbands, usemic=True)
b_mic = gd.empty(nbands, usemic=True)
c_mic = stream.bind(c)
np.random.seed(10)
a_mic.array[:] = np.random.random(a_mic.shape)
b_mic.array[:] = np.random.random(b_mic.shape)
# a_mic.update_device()
# b_mic.update_device()
# warm-up
for i in range(3):
a_mic.update_device()
b_mic.update_device()
gd.integrate(a_mic, b_mic, hermitian=False, _transposed_result=c)
c_mic.update_device()
mic_gemm(1.0, a_mic, c_mic, 0.0, b_mic)
b_mic.update_host()
t0 = time()
# equal(np.sum(c), 3600.89536641, 1e-6)
for i in range(repeats):
a_mic.update_device()
b_mic.update_device()
gd.integrate(a_mic, b_mic, hermitian=False, _transposed_result=c)
c_mic.update_device()
mic_gemm(1.0, a_mic, c_mic, 0.0, b_mic)
b_mic.update_host()
t1 = time()
if rank == 0:
print "Check", np.sum(b_mic.array), "Time", (t1 - t0) / repeats
def matrix_multiply(self, C_NN, psit_nG, P_ani=None, out_nG=None):
"""Calculate new linear combinations of wave functions.
Results will be put in the *P_ani* dict and a new psit_nG returned::
__ __
~ \ ~ ~a ~ \ ~a ~
psi <-- ) C psi and <p |psi > <-- ) C <p |psi >
n /__ nn' n' i n /__ nn' i n'
n' n'
Parameters:
C_NN: ndarray
Matrix representation of the requested linear combinations. Even
with a hermitian operator, this matrix need not be self-adjoint.
However, unlike the results from calculate_matrix_elements, it is
assumed that all matrix elements are filled in (use e.g. tri2full).
psit_nG: ndarray
Set of vectors in which the matrix elements are evaluated.
P_ani: dict
Dictionary of projector overlap integrals P_ni = <p_i | psit_nG>.
"""
if self.A_nn is None:
self.allocate_arrays()
band_comm = self.bd.comm
B = band_comm.size
J = self.nblocks
N = self.bd.mynbands
C_NN = self.bmd.redistribute_input(C_NN)
if B == 1 and J == 1:
# Simple case:
if use_mic:
work_nG = self.work1_xG_mic
else:
work_nG = reshape(self.work1_xG, psit_nG.shape)
if out_nG is None:
out_nG = work_nG
# out_nG[:] = 117 # gemm may not like nan's
elif out_nG is psit_nG:
work_nG[:] = psit_nG
psit_nG = work_nG
if use_mic:
if self.gd.comm.rank == 0:
offload_report(1)
C_NN_mic = self.A_nn_mic
C_NN_mic.array[:] = C_NN[:]
C_NN_mic.update_device()
stream.sync()
mic_gemm(1.0, psit_nG, C_NN_mic, 0.0, out_nG)
if self.gd.comm.rank == 0:
offload_report(0)
else:
self.gd.gemm(1.0, psit_nG, C_NN, 0.0, out_nG)
if P_ani:
for P_ni in P_ani.values():
gemm(1.0, P_ni.copy(), C_NN, 0.0, P_ni)
return out_nG
# Now it gets nasty! We parallelize over B groups of bands and
# each grid chunk is divided in J smaller slices (less memory).
Q = B # always non-hermitian XXX
rank = band_comm.rank
shape = psit_nG.shape
psit_nG = psit_nG.reshape(N, -1)
G = psit_nG.shape[1] # number of grid-points
g = int(np.ceil(G / float(J)))
# Buffers for send/receive of pre-multiplication versions of P_ani's.
sbuf_nI = rbuf_nI = None
if P_ani:
sbuf_nI = np.hstack([P_ni for P_ni in P_ani.values()])
sbuf_nI = np.ascontiguousarray(sbuf_nI)
if B > 1:
rbuf_nI = np.empty_like(sbuf_nI)
# Because of the amount of communication involved, we need to
# be syncronized up to this point but only on the 1D band_comm
# communication ring
band_comm.barrier()
while g * J >= G + g: # remove extra slice(s)
J -= 1
assert 0 < g * J < G + g
work1_xG = reshape(self.work1_xG, (self.X, ) + psit_nG.shape[1:])
work2_xG = reshape(self.work2_xG, (self.X, ) + psit_nG.shape[1:])
for j in range(J):
G1 = j * g
G2 = G1 + g
if G2 > G:
G2 = G
g = G2 - G1
sbuf_ng = reshape(work1_xG, (N, g))
rbuf_ng = reshape(work2_xG, (N, g))
sbuf_ng[:] = psit_nG[:, G1:G2]
beta = 0.0
cycle_P_ani = (j == J - 1 and P_ani)
for q in range(Q):
# Start sending currently buffered kets to rank below
# and receiving next set of kets from rank above us.
# If we're at the last slice, start cycling P_ani too.
if q < Q - 1:
self._initialize_cycle(sbuf_ng, rbuf_ng, sbuf_nI, rbuf_nI,
cycle_P_ani)
# Calculate wave-function contributions from the current slice
# of grid data by the current mynbands x mynbands matrix block.
C_nn = self.bmd.extract_block(C_NN, (rank + q) % B, rank)
self.gd.gemm(1.0, sbuf_ng, C_nn, beta, psit_nG[:, G1:G2])
# If we're at the last slice, add contributions to P_ani's.
if cycle_P_ani:
I1 = 0
for P_ni in P_ani.values():
I2 = I1 + P_ni.shape[1]
gemm(1.0, sbuf_nI[:, I1:I2], C_nn, beta, P_ni)
I1 = I2
# Wait for all send/receives to finish before next iteration.
# Swap send and receive buffer such that next becomes current.
# If we're at the last slice, also finishes the P_ani cycle.
if q < Q - 1:
sbuf_ng, rbuf_ng, sbuf_nI, rbuf_nI = self._finish_cycle(
sbuf_ng, rbuf_ng, sbuf_nI, rbuf_nI, cycle_P_ani)
# First iteration was special because we initialized the kets
if q == 0:
beta = 1.0
psit_nG.shape = shape
return psit_nG
def matrix_multiply(self, C_NN, psit_nG, P_ani=None, out_nG=None):
"""Calculate new linear combinations of wave functions.
Results will be put in the *P_ani* dict and a new psit_nG returned::
__ __
~ \ ~ ~a ~ \ ~a ~
psi <-- ) C psi and <p |psi > <-- ) C <p |psi >
n /__ nn' n' i n /__ nn' i n'
n' n'
Parameters:
C_NN: ndarray
Matrix representation of the requested linear combinations. Even
with a hermitian operator, this matrix need not be self-adjoint.
However, unlike the results from calculate_matrix_elements, it is
assumed that all matrix elements are filled in (use e.g. tri2full).
psit_nG: ndarray
Set of vectors in which the matrix elements are evaluated.
P_ani: dict
Dictionary of projector overlap integrals P_ni = <p_i | psit_nG>.
"""
if self.A_nn is None:
self.allocate_arrays()
band_comm = self.bd.comm
B = band_comm.size
J = self.nblocks
N = self.bd.mynbands
C_NN = self.bmd.redistribute_input(C_NN)
if B == 1 and J == 1:
# Simple case:
if use_mic:
work_nG = self.work1_xG_mic
else:
work_nG = reshape(self.work1_xG, psit_nG.shape)
if out_nG is None:
out_nG = work_nG
# out_nG[:] = 117 # gemm may not like nan's
elif out_nG is psit_nG:
work_nG[:] = psit_nG
psit_nG = work_nG
if use_mic:
if self.gd.comm.rank == 0:
offload_report(1)
C_NN_mic = self.A_nn_mic
C_NN_mic.array[:] = C_NN[:]
C_NN_mic.update_device()
stream.sync()
mic_gemm(1.0, psit_nG, C_NN_mic, 0.0, out_nG)
if self.gd.comm.rank == 0:
offload_report(0)
else:
self.gd.gemm(1.0, psit_nG, C_NN, 0.0, out_nG)
if P_ani:
for P_ni in P_ani.values():
gemm(1.0, P_ni.copy(), C_NN, 0.0, P_ni)
return out_nG
# Now it gets nasty! We parallelize over B groups of bands and
# each grid chunk is divided in J smaller slices (less memory).
Q = B # always non-hermitian XXX
rank = band_comm.rank
shape = psit_nG.shape
psit_nG = psit_nG.reshape(N, -1)
G = psit_nG.shape[1] # number of grid-points
g = int(np.ceil(G / float(J)))
# Buffers for send/receive of pre-multiplication versions of P_ani's.
sbuf_nI = rbuf_nI = None
if P_ani:
sbuf_nI = np.hstack([P_ni for P_ni in P_ani.values()])
sbuf_nI = np.ascontiguousarray(sbuf_nI)
if B > 1:
rbuf_nI = np.empty_like(sbuf_nI)
# Because of the amount of communication involved, we need to
# be syncronized up to this point but only on the 1D band_comm
# communication ring
band_comm.barrier()
while g * J >= G + g: # remove extra slice(s)
J -= 1
assert 0 < g * J < G + g
work1_xG = reshape(self.work1_xG, (self.X,) + psit_nG.shape[1:])
work2_xG = reshape(self.work2_xG, (self.X,) + psit_nG.shape[1:])
for j in range(J):
G1 = j * g
G2 = G1 + g
if G2 > G:
G2 = G
g = G2 - G1
sbuf_ng = reshape(work1_xG, (N, g))
rbuf_ng = reshape(work2_xG, (N, g))
sbuf_ng[:] = psit_nG[:, G1:G2]
beta = 0.0
cycle_P_ani = (j == J - 1 and P_ani)
for q in range(Q):
# Start sending currently buffered kets to rank below
# and receiving next set of kets from rank above us.
# If we're at the last slice, start cycling P_ani too.
if q < Q - 1:
self._initialize_cycle(sbuf_ng, rbuf_ng,
sbuf_nI, rbuf_nI, cycle_P_ani)
# Calculate wave-function contributions from the current slice
# of grid data by the current mynbands x mynbands matrix block.
C_nn = self.bmd.extract_block(C_NN, (rank + q) % B, rank)
self.gd.gemm(1.0, sbuf_ng, C_nn, beta, psit_nG[:, G1:G2])
# If we're at the last slice, add contributions to P_ani's.
if cycle_P_ani:
I1 = 0
for P_ni in P_ani.values():
I2 = I1 + P_ni.shape[1]
gemm(1.0, sbuf_nI[:, I1:I2], C_nn, beta, P_ni)
I1 = I2
# Wait for all send/receives to finish before next iteration.
# Swap send and receive buffer such that next becomes current.
# If we're at the last slice, also finishes the P_ani cycle.
if q < Q - 1:
sbuf_ng, rbuf_ng, sbuf_nI, rbuf_nI = self._finish_cycle(
sbuf_ng, rbuf_ng, sbuf_nI, rbuf_nI, cycle_P_ani)
# First iteration was special because we initialized the kets
if q == 0:
beta = 1.0
psit_nG.shape = shape
return psit_nG
def integrate(self, a_xg, b_yg=None,
global_integral=True, hermitian=False,
_transposed_result=None):
"""Integrate function(s) over domain.
a_xg: ndarray
Function(s) to be integrated.
b_yg: ndarray
If present, integrate a_xg.conj() * b_yg.
global_integral: bool
If the array(s) are distributed over several domains, then the
total sum will be returned. To get the local contribution
only, use global_integral=False.
hermitian: bool
Result is hermitian.
_transposed_result: ndarray
Long story. Don't use this unless you are a method of the
MatrixOperator class ..."""
xshape = a_xg.shape[:-3]
if b_yg is None:
# Only one array:
result = a_xg.reshape(xshape + (-1,)).sum(axis=-1) * self.dv
if global_integral:
if result.ndim == 0:
result = self.comm.sum(result)
else:
self.comm.sum(result)
return result
if isinstance(a_xg, mic.OffloadArray):
# offload arrays have to be contiguous in any case
A_xg = a_xg
B_yg = b_yg
else:
A_xg = np.ascontiguousarray(a_xg.reshape((-1,) + a_xg.shape[-3:]))
B_yg = np.ascontiguousarray(b_yg.reshape((-1,) + b_yg.shape[-3:]))
if _transposed_result is None:
result_yx = np.zeros((len(B_yg), len(A_xg)), A_xg.dtype)
else:
result_yx = _transposed_result
global_integral = False
if isinstance(a_xg, mic.OffloadArray):
result_yx_mic = stream.bind(result_yx)
stream.sync()
# result_yx_mic.fillfrom(result_yx)
# result_yx_mic.array[:] = result_yx[:]
# result_yx_mic.update_device()
if a_xg is b_yg:
if isinstance(a_xg, mic.OffloadArray):
# dsyrk performs badly in MIC so use dgemm here
# mic_rk(self.dv, A_xg, 0.0, result_yx_mic)
mic_gemm(self.dv, A_xg, A_xg, 0.0, result_yx_mic, 'c')
else:
rk(self.dv, A_xg, 0.0, result_yx)
elif hermitian:
if isinstance(a_xg, mic.OffloadArray):
mic_r2k(self.dv, A_xg, B_yg, 0.0, result_yx_mic)
else:
r2k(0.5 * self.dv, A_xg, B_yg, 0.0, result_yx)
else:
if isinstance(a_xg, mic.OffloadArray):
mic_gemm(self.dv, A_xg, B_yg, 0.0, result_yx_mic, 'c')
else:
gemm(self.dv, A_xg, B_yg, 0.0, result_yx, 'c')
if isinstance(a_xg, mic.OffloadArray):
result_yx_mic.update_host()
stream.sync()
if global_integral:
self.comm.sum(result_yx)
yshape = b_yg.shape[:-3]
result = result_yx.T.reshape(xshape + yshape)
if result.ndim == 0:
return result.item()
else:
return result
# a_mic.array = a_mic.array.reshape(a_mic.shape[:-3] + (-1,))
a_mic.update_host()
a_mic_sum = np.sum(a_mic.array)
print " sum(a_mic)=" + str(a_mic_sum)
b_mic = mic.offload_array(b.shape, dtype=float)
b_mic.fillfrom(b)
# b_mic.update_device()
# b_mic.array = b_mic.array.reshape(b_mic.shape[:-3] + (-1,))
b_mic.update_host()
b_mic_sum = np.sum(b_mic.array)
print " sum(b_mic)=" + str(b_mic_sum)
# c_mic = offload_array(c.shape, dtype=float)
# c_mic.fill(0.0);
c_mic = device.associate(c)
c_mic.update_device()
t0 = time()
for r in range(repeats):
mic_gemm(alpha, a_mic, b_mic, beta, c_mic, "c")
#mic_r2k(alpha, a_mic, b_mic, beta, c_mic)
c_mic.update_host()
c[:] = c_mic.array[:]
t1 = time()
print "MIC time", t1 - t0
print "MIC checks"
c_mic_sum = np.sum(c_mic.array)
print " sum(c_mic)=" + str(c_mic_sum)
print " sum(c)=" + str(np.sum(c))
