def __init__(self, path):
self.path = path
self.fdimensions = []
self.bytes_header = -1
self.verbose = gpt.default.is_verbose("io")
self.size = 1
gpt.barrier()
def __init__(self, root, params, write):
self.root = root
self.params = params
if not "grids" in self.params:
self.params["grids"] = {}
else:
if type(self.params["grids"]) == gpt.grid:
self.params["grids"] = [self.params["grids"]]
self.params["grids"] = dict([(g.describe(), g)
for g in self.params["grids"]])
self.verbose = gpt.default.is_verbose("io")
if gpt.rank() == 0:
os.makedirs(self.root, exist_ok=True)
if write:
self.glb = gpt.FILE(root + "/global", "wb")
for f in glob.glob("%s/??/*.field" % self.root):
os.unlink(f)
else:
self.glb = gpt.FILE(root + "/global", "r+b")
else:
self.glb = None
self.loc = {}
self.pos = {}
self.loc_desc = ""
# now sync since only root has created directory
gpt.barrier()
def __init__(self, root, write, params):
self.root = root
self.params = params
self.params["grids"] = {}
self.verbose = gpt.default.is_verbose("io")
if gpt.rank() == 0:
os.makedirs(self.root, exist_ok=True)
if write:
self.glb = gpt.FILE(root + "/global", "wb")
for f in glob.glob("%s/??/*.field" % self.root):
os.unlink(f)
else:
self.glb = gpt.FILE(root + "/global", "rb")
else:
self.glb = None
self.loc = {}
self.pos = {}
self.loc_desc = ""
self.cache = {}
# If we write, keep an index buffer
self.index_file = io.StringIO("") if write else None
# now sync since only root has created directory
gpt.barrier()
def perform(self, root):
for target in self.targets:
dst = f"{root}/{target}/propagators"
g.message(dst, "Delete", os.path.exists(dst))
if os.path.exists(dst) and g.rank() == 0:
shutil.rmtree(dst)
g.barrier()
def write(self, t, cc):
if self.f is not None:
self.f.write(struct.pack("i", len(t) + 1))
self.f.write((t + "\0").encode("utf-8"))
ln = len(cc)
ccr = [
fff for sublist in ((c.real, c.imag) for c in cc)
for fff in sublist
]
bindata = struct.pack("d" * 2 * ln, *ccr)
crc32comp = binascii.crc32(bindata) & 0xFFFFFFFF
self.f.write(struct.pack("II", crc32comp, ln))
self.f.write(bindata)
self.f.flush()
gpt.barrier()
def download(dst, src):
src_a = src.split("/")
assert len(src_a) >= 3
host = src_a[0]
assert host[-1] == ":"
assert src_a[1] == ""
path = "/".join(src_a[2:])
assert host in base
baseurl = base[host]
if gpt.rank() == 0:
verbose = gpt.default.is_verbose("repository")
t0 = gpt.time()
if "GPT_REPOSITORY" in os.environ:
root = os.environ["GPT_REPOSITORY"]
shutil.copy2(f"{root}/{path}", dst)
mode = "copy"
else:
filename, header = request.urlretrieve(f"{baseurl}/{path}", filename=dst)
mode = "download"
t1 = gpt.time()
filesize = os.path.getsize(dst)
speedMBs = filesize / 1024.0 ** 2.0 / (t1 - t0)
if verbose:
gpt.message(f"Repository {mode} {src} in {t1-t0:g} s at {speedMBs:g} MB/s")
# add a barrier so that all nodes have file after download
gpt.barrier()
# os.scandir to trigger network filesystem synchronization
os.scandir(os.path.dirname(dst))
def close(self):
if self.f is not None:
self.f.close()
self.f = None
gpt.barrier()
def save(filename, objs, params):
# split data to save
assert len(objs) == 3
basis = objs[0]
cevec = objs[1]
ev = objs[2]
# verbosity
verbose = gpt.default.is_verbose("io")
if verbose:
gpt.message(
"Saving %d basis vectors, %d coarse-grid vectors, %d eigenvalues to %s"
% (len(basis), len(cevec), len(ev), filename))
# create directory
if gpt.rank() == 0:
os.makedirs(filename, exist_ok=True)
# now sync since only root has created directory
gpt.barrier()
# write eigenvalues
if gpt.rank() == 0:
f = open("%s/eigen-values.txt" % filename, "wt")
f.write("%d\n" % len(ev))
for v in ev:
f.write("%.15E\n" % v)
f.close()
# site checkerboard
# only odd is used in this file format but
# would be easy to generalize here
site_cb = gpt.odd
# grids
assert len(basis) > 0
assert len(cevec) > 0
fgrid = basis[0].grid
cgrid = cevec[0].grid
# mpi layout
if params["mpi"] is not None:
mpi = params["mpi"]
else:
mpi = fgrid.mpi
assert mpi[0] == 1 # assert no mpi in 5th direction
# params
assert basis[0].checkerboard() == site_cb
nd = 5
assert len(fgrid.ldimensions) == nd
fdimensions = fgrid.fdimensions
ldimensions = [conformDiv(fdimensions[i], mpi[i]) for i in range(nd)]
assert fgrid.precision == gpt.single
s = ldimensions
b = [
conformDiv(fgrid.fdimensions[i], cgrid.fdimensions[i])
for i in range(nd)
]
nb = [conformDiv(s[i], b[i]) for i in range(nd)]
neigen = len(cevec)
nbasis = len(basis)
if "nsingle" in params:
nsingle = params["nsingle"]
assert nsingle <= nbasis
else:
nsingle = nbasis
nsingleCap = min([nsingle, nbasis])
blocks = numpy.prod(nb)
FP16_COEF_EXP_SHARE_FLOATS = 10
# write metadata
if gpt.rank() == 0:
fmeta = open("%s/metadata.txt" % filename, "wt")
for i in range(nd):
fmeta.write("s[%d] = %d\n" % (i, s[(i + 1) % nd]))
for i in range(nd):
fmeta.write("b[%d] = %d\n" % (i, b[(i + 1) % nd]))
for i in range(nd):
fmeta.write("nb[%d] = %d\n" % (i, nb[(i + 1) % nd]))
fmeta.write("neig = %d\n" % neigen)
fmeta.write("nkeep = %d\n" % nbasis)
fmeta.write("nkeep_single = %d\n" % nsingle)
fmeta.write("blocks = %d\n" % blocks)
fmeta.write("FP16_COEF_EXP_SHARE_FLOATS = %d\n" %
FP16_COEF_EXP_SHARE_FLOATS)
fmeta.flush() # write crc32 later
# create cartesian view on fine grid
cv0 = gpt.cartesian_view(-1, mpi, fdimensions, fgrid.cb, site_cb)
views = cv0.views_for_node(fgrid)
crc32 = numpy.array([0] * cv0.ranks, dtype=numpy.uint64)
# timing
t0 = gpt.time()
totalSizeGB = 0
dt_fp16 = 1e-30
dt_distr = 1e-30
dt_munge = 1e-30
dt_crc = 1e-30
dt_fwrite = 1e-30
t0 = gpt.time()
# load all views
if verbose:
gpt.message("Saving %s with %d views per node" %
(filename, len(views)))
for i, v in enumerate(views):
cv = gpt.cartesian_view(v if v is not None else -1, mpi, fdimensions,
fgrid.cb, site_cb)
cvc = gpt.cartesian_view(v if v is not None else -1, mpi,
cgrid.fdimensions, gpt.full, gpt.none)
pos_coarse = gpt.coordinates(cvc, "canonical")
dn, fn = get_local_name(filename, cv)
if fn is not None:
os.makedirs(dn, exist_ok=True)
# sizes
slot_lsites = numpy.prod(cv.view_dimensions)
assert slot_lsites % blocks == 0
block_data_size_single = slot_lsites * 12 // 2 // blocks * 2 * 4
block_data_size_fp16 = FP_16_SIZE(slot_lsites * 12 // 2 // blocks * 2,
24)
coarse_block_size_part_fp32 = 2 * (4 * nsingleCap)
coarse_block_size_part_fp16 = 2 * (FP_16_SIZE(
nbasis - nsingleCap, FP16_COEF_EXP_SHARE_FLOATS))
coarse_vector_size = (coarse_block_size_part_fp32 +
coarse_block_size_part_fp16) * blocks
totalSize = (
blocks *
(block_data_size_single * nsingleCap + block_data_size_fp16 *
(nbasis - nsingleCap)) + neigen * coarse_vector_size)
totalSizeGB += totalSize / 1024.0**3.0 if v is not None else 0.0
# checksum
crc32_comp = 0
# file
f = gpt.FILE(fn, "wb") if fn is not None else None
# block positions
pos = [
cgpt.coordinates_from_block(cv.top, cv.bottom, b, nb,
"canonicalOdd") for b in range(blocks)
]
# group blocks
read_blocks = blocks
block_reduce = 1
max_read_blocks = get_param(params, "max_read_blocks", 8)
while read_blocks > max_read_blocks and read_blocks % 2 == 0:
pos = [
numpy.concatenate((pos[2 * i + 0], pos[2 * i + 1]))
for i in range(read_blocks // 2)
]
block_data_size_single *= 2
block_data_size_fp16 *= 2
read_blocks //= 2
block_reduce *= 2
# make read-only to enable caching
for x in pos:
x.setflags(write=0)
# single-precision data
data = memoryview(bytearray(block_data_size_single * nsingleCap))
data_munged = memoryview(bytearray(block_data_size_single *
nsingleCap))
for b in range(read_blocks):
fgrid.barrier()
dt_distr -= gpt.time()
lhs_size = basis[0].otype.nfloats * 4 * len(pos[b])
lhs = data_munged[0:lhs_size]
distribute_plan = gpt.copy_plan(lhs, basis[0])
distribute_plan.destination += gpt.global_memory_view(
fgrid, [[fgrid.processor, lhs, 0, lhs.nbytes]])
distribute_plan.source += basis[0].view[pos[b]]
distribute_plan = distribute_plan()
lhs = None
for i in range(nsingleCap):
distribute_plan(
data_munged[block_data_size_single *
i:block_data_size_single * (i + 1)],
basis[i],
)
dt_distr += gpt.time()
if f is not None:
dt_munge -= gpt.time()
cgpt.munge_inner_outer(
data,
data_munged,
block_reduce,
nsingleCap,
)
dt_munge += gpt.time()
dt_crc -= gpt.time()
crc32_comp = gpt.crc32(data, crc32_comp)
dt_crc += gpt.time()
fgrid.barrier()
dt_fwrite -= gpt.time()
if f is not None:
f.write(data)
globalWriteGB = len(data) / 1024.0**3.0
else:
globalWriteGB = 0.0
globalWriteGB = fgrid.globalsum(globalWriteGB)
dt_fwrite += gpt.time()
totalSizeGB += globalWriteGB
if verbose:
gpt.message(
"* write %g GB: fwrite at %g GB/s, crc32 at %g GB/s, munge at %g GB/s, distribute at %g GB/s"
% (
totalSizeGB,
totalSizeGB / dt_fwrite,
totalSizeGB / dt_crc,
totalSizeGB / dt_munge,
totalSizeGB / dt_distr,
))
# fp16 data
if nbasis != nsingleCap:
# allocate data buffer
data_fp32 = memoryview(
bytearray(block_data_size_single * (nbasis - nsingleCap)))
data_munged = memoryview(
bytearray(block_data_size_single * (nbasis - nsingleCap)))
data = memoryview(
bytearray(block_data_size_fp16 * (nbasis - nsingleCap)))
for b in range(read_blocks):
fgrid.barrier()
dt_distr -= gpt.time()
lhs_size = basis[0].otype.nfloats * 4 * len(pos[b])
lhs = data_munged[0:lhs_size]
distribute_plan = gpt.copy_plan(lhs, basis[0])
distribute_plan.destination += gpt.global_memory_view(
fgrid, [[fgrid.processor, lhs, 0, lhs.nbytes]])
distribute_plan.source += basis[0].view[pos[b]]
distribute_plan = distribute_plan()
lhs = None
for i in range(nsingleCap, nbasis):
j = i - nsingleCap
distribute_plan(
data_munged[j * block_data_size_single:(j + 1) *
block_data_size_single],
basis[i],
)
dt_distr += gpt.time()
if f is not None:
dt_munge -= gpt.time()
cgpt.munge_inner_outer(
data_fp32,
data_munged,
block_reduce,
nbasis - nsingleCap,
)
dt_munge += gpt.time()
dt_fp16 -= gpt.time()
cgpt.fp32_to_fp16(data, data_fp32, 24)
dt_fp16 += gpt.time()
dt_crc -= gpt.time()
crc32_comp = gpt.crc32(data, crc32_comp)
dt_crc += gpt.time()
fgrid.barrier()
dt_fwrite -= gpt.time()
if f is not None:
f.write(data)
globalWriteGB = len(data) / 1024.0**3.0
else:
globalWriteGB = 0.0
globalWriteGB = fgrid.globalsum(globalWriteGB)
dt_fwrite += gpt.time()
totalSizeGB += globalWriteGB
if verbose:
gpt.message(
"* write %g GB: fwrite at %g GB/s, crc32 at %g GB/s, munge at %g GB/s, distribute at %g GB/s, fp16 at %g GB/s"
% (
totalSizeGB,
totalSizeGB / dt_fwrite,
totalSizeGB / dt_crc,
totalSizeGB / dt_munge,
totalSizeGB / dt_distr,
totalSizeGB / dt_fp16,
))
# coarse grid data
data = memoryview(bytearray(coarse_vector_size))
data_fp32 = memoryview(
bytearray(cevec[0].otype.nfloats * 4 * len(pos_coarse)))
distribute_plan = gpt.copy_plan(data_fp32, cevec[0])
distribute_plan.destination += gpt.global_memory_view(
cgrid, [[cgrid.processor, data_fp32, 0, data_fp32.nbytes]])
distribute_plan.source += cevec[0].view[pos_coarse]
distribute_plan = distribute_plan()
for j in range(neigen):
fgrid.barrier()
dt_distr -= gpt.time()
distribute_plan(data_fp32, cevec[j])
dt_distr += gpt.time()
if f is not None:
dt_fp16 -= gpt.time()
cgpt.fp32_to_mixed_fp32fp16(
data,
data_fp32,
coarse_block_size_part_fp32,
coarse_block_size_part_fp16,
FP16_COEF_EXP_SHARE_FLOATS,
)
dt_fp16 += gpt.time()
dt_crc -= gpt.time()
crc32_comp = gpt.crc32(data, crc32_comp)
dt_crc += gpt.time()
fgrid.barrier()
dt_fwrite -= gpt.time()
if f is not None:
f.write(data)
globalWriteGB = len(data) / 1024.0**3.0
else:
globalWriteGB = 0.0
globalWriteGB = fgrid.globalsum(globalWriteGB)
dt_fwrite += gpt.time()
totalSizeGB += globalWriteGB
if verbose and j % (neigen // 10) == 0:
gpt.message(
"* write %g GB: fwrite at %g GB/s, crc32 at %g GB/s, munge at %g GB/s, distribute at %g GB/s, fp16 at %g GB/s"
% (
totalSizeGB,
totalSizeGB / dt_fwrite,
totalSizeGB / dt_crc,
totalSizeGB / dt_munge,
totalSizeGB / dt_distr,
totalSizeGB / dt_fp16,
))
# save crc
crc32[cv.rank] = crc32_comp
# synchronize crc32
fgrid.globalsum(crc32)
# timing
t1 = gpt.time()
# write crc to metadata
if gpt.rank() == 0:
for i in range(len(crc32)):
fmeta.write("crc32[%d] = %X\n" % (i, crc32[i]))
fmeta.close()
# verbosity
if verbose:
gpt.message("* save %g GB at %g GB/s" % (totalSizeGB, totalSizeGB /
(t1 - t0)))
self.fn = fn
super().__init__("file_" + fn, ["file_" + f for f in needfn])
def perform(self, root):
f = open(f"{root}/{self.name}/{self.fn}", "wt")
f.write(self.fn + "\n")
f.close()
def check(self, root):
return os.path.exists(f"{root}/{self.name}/{self.fn}")
if os.path.exists(root) and g.rank() == 0:
shutil.rmtree(root)
g.barrier()
fail_B = True
for jid in range(5):
j = g.jobs.next(
root,
[
job_create_file("A", []),
job_create_file("C", ["B", "A"]),
job_create_file("B", ["A"]),
],
stale_seconds=0.0,
)
if j is not None and j.name == "file_B" and fail_B:
