def _emit_build_profiling(self):
if not is_log_enabled_for('PERF'):
return
# Rounder to K decimal places
fround = lambda i, n=100: ceil(i * n) / n
timings = self._profiler.py_timers.copy()
tot = timings.pop('op-compile')
perf("Operator `%s` generated in %.2f s" % (self.name, fround(tot)))
max_hotspots = 3
threshold = 20.
def _emit_timings(timings, indent=''):
timings.pop('total', None)
entries = sorted(timings,
key=lambda i: timings[i]['total'],
reverse=True)
for i in entries[:max_hotspots]:
v = fround(timings[i]['total'])
perc = fround(v / tot * 100, n=10)
if perc > threshold:
perf("%s%s: %.2f s (%.1f %%)" %
(indent, i.lstrip('_'), v, perc))
_emit_timings(timings[i], ' ' * len(indent) + ' * ')
_emit_timings(timings, ' * ')
if self._profiler._ops:
ops = ['%d --> %d' % i for i in self._profiler._ops]
perf("Flops reduction after symbolic optimization: [%s]" %
' ; '.join(ops))
def _emit_timings(timings, indent=''):
timings.pop('total', None)
entries = sorted(timings, key=lambda i: timings[i]['total'], reverse=True)
for i in entries[:max_hotspots]:
v = fround(timings[i]['total'])
perc = fround(v/tot*100, n=10)
if perc > threshold:
perf("%s%s: %.2f s (%.1f %%)" % (indent, i.lstrip('_'), v, perc))
_emit_timings(timings[i], ' '*len(indent) + ' * ')
def _profile_output(self, args):
"""Produce a performance summary of the profiled sections."""
summary = self._profiler.summary(args, self._dtype)
info("Operator `%s` run in %.2f s" % (self.name, sum(summary.timings.values())))
for k, v in summary.items():
itershapes = [",".join(str(i) for i in its) for its in v.itershapes]
if len(itershapes) > 1:
name = "%s<%s>" % (k, ",".join("<%s>" % i for i in itershapes))
elif len(itershapes) == 1:
name = "%s<%s>" % (k, itershapes[0])
else:
name = None
gpointss = ", %.2f GPts/s" % v.gpointss if v.gpointss else ''
perf("* %s with OI=%.2f computed in %.3f s [%.2f GFlops/s%s]" %
(name, v.oi, v.time, v.gflopss, gpointss))
return summary
def _jit_compile(self):
"""
JIT-compile the C code generated by the Operator.
It is ensured that JIT compilation will only be performed once per
Operator, reagardless of how many times this method is invoked.
"""
if self._lib is None:
with self._profiler.timer_on('jit-compile'):
recompiled, src_file = self._compiler.jit_compile(
self._soname, str(self.ccode))
elapsed = self._profiler.py_timers['jit-compile']
if recompiled:
perf("Operator `%s` jit-compiled `%s` in %.2f s with `%s`" %
(self.name, src_file, elapsed, self._compiler))
else:
perf("Operator `%s` fetched `%s` in %.2f s from jit-cache" %
(self.name, src_file, elapsed))
def _emit_build_profiling(self):
if not is_log_enabled_for('PERF'):
return
# Rounder to K decimal places
fround = lambda i, n=100: ceil(i * n) / n
timings = self._profiler.py_timers.copy()
tot = timings.pop('op-compile')
perf("Operator `%s` generated in %.2f s" % (self.name, fround(tot)))
max_hotspots = 3
for i in sorted(timings, key=timings.get, reverse=True)[:max_hotspots]:
v = fround(timings[i])
perc = fround(v / tot * 100, n=10)
if perc > 20.:
perf("- [Hotspot] %s: %.2f s (%.1f %%)" %
(i.lstrip('_'), v, perc))
def autotune(operator, arguments, parameters, tunable):
"""
Acting as a high-order function, take as input an operator and a list of
operator arguments to perform empirical autotuning. Some of the operator
arguments are marked as tunable.
"""
# We get passed all the arguments, but the cfunction only requires a subset
at_arguments = OrderedDict([(p.name, arguments[p.name])
for p in parameters])
# User-provided output data must not be altered
output = [i.name for i in operator.output]
for k, v in arguments.items():
if k in output:
at_arguments[k] = v.copy()
iterations = FindNodes(Iteration).visit(operator.body)
dim_mapper = {i.dim.name: i.dim for i in iterations}
# Shrink the iteration space of time-stepping dimension so that auto-tuner
# runs will finish quickly
steppers = [i for i in iterations if i.dim.is_Time]
if len(steppers) == 0:
timesteps = 1
elif len(steppers) == 1:
stepper = steppers[0]
start = at_arguments[stepper.dim.min_name]
timesteps = stepper.extent(start=start,
finish=options['at_squeezer']) - 1
if timesteps < 0:
timesteps = options['at_squeezer'] - timesteps
perf("AutoTuner: Number of timesteps adjusted to %d" % timesteps)
at_arguments[stepper.dim.min_name] = start
at_arguments[stepper.dim.max_name] = timesteps
if stepper.dim.is_Stepping:
at_arguments[stepper.dim.parent.min_name] = start
at_arguments[stepper.dim.parent.max_name] = timesteps
else:
warning("AutoTuner: Couldn't understand loop structure; giving up")
return arguments
# Attempted block sizes ...
mapper = OrderedDict([(i.argument.symbolic_size.name, i) for i in tunable])
# ... Defaults (basic mode)
blocksizes = [
OrderedDict([(i, v) for i in mapper]) for v in options['at_blocksize']
]
# ... Always try the entire iteration space (degenerate block)
itershape = [
mapper[i].iteration.symbolic_extent.subs(arguments) for i in mapper
]
blocksizes.append(
OrderedDict([(i, mapper[i].iteration.extent(0, j - 1))
for i, j in zip(mapper, itershape)]))
# ... More attempts if auto-tuning in aggressive mode
if configuration['autotuning'].level == 'aggressive':
blocksizes = more_heuristic_attempts(blocksizes)
# How many temporaries are allocated on the stack?
# Will drop block sizes that might lead to a stack overflow
functions = FindSymbols('symbolics').visit(operator.body +
operator.elemental_functions)
stack_shapes = [
i.symbolic_shape for i in functions if i.is_Array and i._mem_stack
]
stack_space = sum(reduce(mul, i, 1)
for i in stack_shapes) * operator._dtype().itemsize
# Note: there is only a single loop over 'blocksize' because only
# square blocks are tested
timings = OrderedDict()
for bs in blocksizes:
illegal = False
for k, v in at_arguments.items():
if k in bs:
val = bs[k]
start = mapper[k].original_dim.symbolic_start.subs(arguments)
end = mapper[k].original_dim.symbolic_end.subs(arguments)
if val <= mapper[k].iteration.extent(start, end):
at_arguments[k] = val
else:
# Block size cannot be larger than actual dimension
illegal = True
break
if illegal:
continue
# Make sure we remain within stack bounds, otherwise skip block size
dim_sizes = {}
for k, v in at_arguments.items():
if k in bs:
dim_sizes[mapper[k].argument.symbolic_size] = bs[k]
elif k in dim_mapper:
dim_sizes[dim_mapper[k].symbolic_size] = v
try:
bs_stack_space = stack_space.xreplace(dim_sizes)
except AttributeError:
bs_stack_space = stack_space
try:
if int(bs_stack_space) > options['at_stack_limit']:
continue
except TypeError:
# We should never get here
warning(
"AutoTuner: Couldn't determine stack size; skipping block shape %s"
% str(bs))
continue
# Use AutoTuner-specific profiler structs
timer = operator.profiler.timer.reset()
at_arguments[operator.profiler.name] = timer
operator.cfunction(*list(at_arguments.values()))
elapsed = sum(getattr(timer._obj, i) for i, _ in timer._obj._fields_)
timings[tuple(bs.items())] = elapsed
perf("AutoTuner: Block shape <%s> took %f (s) in %d timesteps" %
(','.join('%d' % i for i in bs.values()), elapsed, timesteps))
try:
best = dict(min(timings, key=timings.get))
perf("AutoTuner: Selected block shape %s" % best)
except ValueError:
warning("AutoTuner: Couldn't find legal block shapes")
return arguments
# Build the new argument list
tuned = OrderedDict()
for k, v in arguments.items():
tuned[k] = best[k] if k in mapper else v
# Reset the profiling struct
assert operator.profiler.name in tuned
tuned[operator.profiler.name] = operator.profiler.timer.reset()
return tuned
def log(msg):
perf("AutoTuner: %s" % msg)
def _emit_apply_profiling(self, args):
"""Produce a performance summary of the profiled sections."""
# Rounder to 2 decimal places
fround = lambda i: ceil(i * 100) / 100
info("Operator `%s` ran in %.2f s" %
(self.name, fround(self._profiler.py_timers['apply'])))
summary = self._profiler.summary(args,
self._dtype,
reduce_over='apply')
if not is_log_enabled_for('PERF'):
# Do not waste time
return summary
if summary.globals:
# Note that with MPI enabled, the global performance indicators
# represent "cross-rank" performance data
metrics = []
v = summary.globals.get('vanilla')
if v is not None:
metrics.append("OI=%.2f" % fround(v.oi))
metrics.append("%.2f GFlops/s" % fround(v.gflopss))
v = summary.globals.get('fdlike')
if v is not None:
metrics.append("%.2f GPts/s" % fround(v.gpointss))
if metrics:
perf("Global performance: [%s]" % ', '.join(metrics))
perf("Local performance:")
indent = " " * 2
else:
indent = ""
# Emit local, i.e. "per-rank" performance. Without MPI, this is the only
# thing that will be emitted
for k, v in summary.items():
rank = "[rank%d]" % k.rank if k.rank is not None else ""
oi = "OI=%.2f" % fround(v.oi)
gflopss = "%.2f GFlops/s" % fround(v.gflopss)
gpointss = "%.2f GPts/s" % fround(
v.gpointss) if v.gpointss else None
metrics = ", ".join(i for i in [oi, gflopss, gpointss]
if i is not None)
itershapes = [
",".join(str(i) for i in its) for its in v.itershapes
]
if len(itershapes) > 1:
itershapes = ",".join("<%s>" % i for i in itershapes)
elif len(itershapes) == 1:
itershapes = itershapes[0]
else:
itershapes = ""
name = "%s%s<%s>" % (k.name, rank, itershapes)
perf("%s* %s ran in %.2f s [%s]" %
(indent, name, fround(v.time), metrics))
for n, time in summary.subsections.get(k.name, {}).items():
perf("%s+ %s ran in %.2f s [%.2f%%]" %
(indent * 2, n, time, fround(time / v.time * 100)))
# Emit performance mode and arguments
perf_args = {}
for i in self.input + self.dimensions:
if not i.is_PerfKnob:
continue
try:
perf_args[i.name] = args[i.name]
except KeyError:
# Try with the aliases
for a in i._arg_names:
if a in args:
perf_args[a] = args[a]
break
perf("Performance[mode=%s] arguments: %s" %
(self._state['optimizations'], perf_args))
return summary
def _emit_apply_profiling(self, args):
"""Produce a performance summary of the profiled sections."""
# Rounder to 2 decimal places
fround = lambda i: ceil(i * 100) / 100
info("Operator `%s` run in %.2f s" %
(self.name, fround(self._profiler.py_timers['apply'])))
summary = self._profiler.summary(args,
self._dtype,
reduce_over='apply')
if not is_log_enabled_for('PERF'):
# Do not waste time
return summary
if summary.globals:
indent = " " * 2
perf("Global performance indicators")
# With MPI enabled, the 'vanilla' entry contains "cross-rank" performance data
v = summary.globals.get('vanilla')
if v is not None:
gflopss = "%.2f GFlops/s" % fround(v.gflopss)
gpointss = "%.2f GPts/s" % fround(
v.gpointss) if v.gpointss else None
metrics = ", ".join(i for i in [gflopss, gpointss]
if i is not None)
perf(
"%s* Operator `%s` with OI=%.2f computed in %.2f s [%s]" %
(indent, self.name, fround(v.oi), fround(v.time), metrics))
v = summary.globals.get('fdlike')
if v is not None:
perf("%s* Achieved %.2f FD-GPts/s" % (indent, v.gpointss))
perf("Local performance indicators")
else:
indent = ""
# Emit local, i.e. "per-rank" performance. Without MPI, this is the only
# thing that will be emitted
for k, v in summary.items():
rank = "[rank%d]" % k.rank if k.rank is not None else ""
gflopss = "%.2f GFlops/s" % fround(v.gflopss)
gpointss = "%.2f GPts/s" % fround(
v.gpointss) if v.gpointss else None
metrics = ", ".join(i for i in [gflopss, gpointss]
if i is not None)
itershapes = [
",".join(str(i) for i in its) for its in v.itershapes
]
if len(itershapes) > 1:
name = "%s%s<%s>" % (k.name, rank, ",".join(
"<%s>" % i for i in itershapes))
perf("%s* %s with OI=%.2f computed in %.2f s [%s]" %
(indent, name, fround(v.oi), fround(v.time), metrics))
elif len(itershapes) == 1:
name = "%s%s<%s>" % (k.name, rank, itershapes[0])
perf("%s* %s with OI=%.2f computed in %.2f s [%s]" %
(indent, name, fround(v.oi), fround(v.time), metrics))
else:
name = k.name
perf("%s* %s%s computed in %.2f s" %
(indent, name, rank, fround(v.time)))
# Emit relevant configuration values
perf("Configuration: %s" % self._state['optimizations'])
# Emit relevant performance arguments
perf_args = {}
for i in self.input + self.dimensions:
if not i.is_PerfKnob:
continue
try:
perf_args[i.name] = args[i.name]
except KeyError:
# Try with the aliases
for a in i._arg_names:
if a in args:
perf_args[a] = args[a]
break
perf("Performance arguments: %s" % perf_args)
return summary
def log(msg):
perf("AutoTuner: %s" % msg)
def _profile_output(self, args):
"""Produce a performance summary of the profiled sections."""
# Rounder to 2 decimal places
fround = lambda i: ceil(i * 100) / 100
info("Operator `%s` run in %.2f s" %
(self.name, fround(self._profiler.py_timers['apply'])))
summary = self._profiler.summary(args,
self._dtype,
reduce_over='apply')
if summary.globals:
indent = " " * 2
perf("Global performance indicators")
# With MPI enabled, the 'vanilla' entry contains "cross-rank" performance data
v = summary.globals.get('vanilla')
if v is not None:
gflopss = "%.2f GFlops/s" % fround(v.gflopss)
gpointss = "%.2f GPts/s" % fround(
v.gpointss) if v.gpointss else None
metrics = ", ".join(i for i in [gflopss, gpointss]
if i is not None)
perf(
"%s* Operator `%s` with OI=%.2f computed in %.2f s [%s]" %
(indent, self.name, fround(v.oi), fround(v.time), metrics))
v = summary.globals.get('fdlike')
if v is not None:
perf("%s* Achieved %.2f FD-GPts/s" % (indent, v.gpointss))
perf("Local performance indicators")
else:
indent = ""
# Emit local, i.e. "per-rank" performance. Without MPI, this is the only
# thing that will be emitted
for k, v in summary.items():
rank = "[rank%d]" % k.rank if k.rank is not None else ""
gflopss = "%.2f GFlops/s" % fround(v.gflopss)
gpointss = "%.2f GPts/s" % fround(
v.gpointss) if v.gpointss else None
metrics = ", ".join(i for i in [gflopss, gpointss]
if i is not None)
itershapes = [
",".join(str(i) for i in its) for its in v.itershapes
]
if len(itershapes) > 1:
name = "%s%s<%s>" % (k.name, rank, ",".join(
"<%s>" % i for i in itershapes))
perf("%s* %s with OI=%.2f computed in %.2f s [%s]" %
(indent, name, fround(v.oi), fround(v.time), metrics))
elif len(itershapes) == 1:
name = "%s%s<%s>" % (k.name, rank, itershapes[0])
perf("%s* %s with OI=%.2f computed in %.2f s [%s]" %
(indent, name, fround(v.oi), fround(v.time), metrics))
else:
name = k.name
perf("%s* %s%s computed in %.2f s" %
(indent, name, rank, fround(v.time)))
perf("Configuration: %s" % self._state['optimizations'])
return summary
