def t3():
out = clone_here(a[i,
j]) # Move data to the current device
rhs1 = clone_here(a[i, k])
rhs2 = clone_here(a[j, k])
out = update(rhs1, rhs2, out)
copy(a[i, j], out) # Move the result to the global array
def t3():
out = clone_here(a[i,
j]) # Move data to the current device
rhs1 = clone_here(a[i, k])
rhs2 = clone_here(a[j, k])
out -= rhs1 @ rhs2.T
copy(a[i, j], out) # Move the result to the global array
def matmul_task():
old_device = cp.cuda.Device()
#local_start = time.perf_counter()
b_block_local = clone_here(b_block)
#cp.cuda.get_current_stream().synchronize()
#communication_stop = time.perf_counter()
# cupy doesn't support the out argument for matmul yet so we have to copy.
# cp.matmul(a_block, b_block_local.T, out = c_block)
c_block[:] = a_block @ b_block_local.T
def __getitem__(self, index: IndexType): # -> Union[Array, List[Array]]
"""
Read partitions and make sure they are on the current device.
:param index: index of the target partition(s).
.. todo:
Multiple partitions are currently returned as a Python list of partitions (ndarrays).
"""
if not isinstance(index, tuple):
index = (index, )
ret = []
parse_index(
self._latest_view,
index,
step=lambda I, i: I[i],
stop=lambda x: ret.append(clone_here(x) if is_array(x) else x))
if len(ret) == 1:
if ret[0] is None: warn("Partition has been freed!")
return ret[0]
warn(
"Multiple partitions are currently returned as a Python list of partitions (ndarrays)."
)
return ret
def t2():
dblock = clone_here(a[j, j])
dblock = cholesky(dblock)
copy(a[j, j], dblock)
def t1():
out = clone_here(a[j, j]) # Move data to the current device
rhs = clone_here(a[j, k])
out = update(rhs, rhs, out)
copy(a[j, j], out) # Move the result to the global array
def get_gpu_memory(i:int, j:int, num_gpus:int): dev_id = i % num_gpus local_id = i // num_gpus src = gpu_arrs[dev_id][local_id][j] dst = clone_here(src) return dst
def t3():
out = clone_here(a[i, j])
rhs1 = clone_here(a[i, k])
rhs2 = clone_here(a[j, k])
out -= rhs1 @ rhs2.T
copy(a[i, j], out)
def t1():
out = clone_here(a[j, j])
rhs = clone_here(a[j, k])
out -= rhs @ rhs.T
copy(a[j, j], out)
def cholesky_inplace(a):
if a.shape[0] != a.shape[1]:
raise ValueError("A square array is required.")
ca = clone_here(a)
ca[:] = cupy.linalg.cholesky(ca)
copy(a, ca)
def t4():
factor = clone_here(a[j, j])
panel = clone_here(a[i, j])
panel = ltriang_solve(factor, panel)
copy(a[i, j], panel)
def t1():
#print("t1[", i, "] start", sep='')
#copy_start = time.time()
A_block_local = clone_here(A_block)
#copy_end = time.time()
Q1_blocked[i], R1[R1_lower:R1_upper] = qr_block(A_block_local)
def t3():
#print("t3[", i, "] start", sep='')
Q1_block_local = clone_here(Q1_blocked[i])
Q2_block_local = clone_here(Q2_block)
Q[Q_lower:Q_upper] = matmul_block(Q1_block_local, Q2_block_local)