def __init__(self, kernel, machine, cores=1):
"""Initialize cache simulation based predictor from kernel and machine object."""
CachePredictor.__init__(self, kernel, machine, cores)
# Get the machine's cache model and simulator
csim = self.machine.get_cachesim(self.cores)
# FIXME handle multiple datatypes
element_size = self.kernel.datatypes_size[self.kernel.datatype]
cacheline_size = self.machine['cacheline size']
elements_per_cacheline = int(cacheline_size // element_size)
# Gathering some loop information:
inner_loop = list(self.kernel.get_loop_stack(subs_consts=True))[-1]
inner_index = symbol_pos_int(inner_loop['index'])
inner_increment = inner_loop[
'increment'] # Calculate the number of iterations for warm-up
max_cache_size = max(
map(lambda c: c.size(), csim.levels(with_mem=False)))
max_array_size = max(
self.kernel.array_sizes(in_bytes=True, subs_consts=True).values())
# Regular Initialization
warmup_indices = {
symbol_pos_int(l['index']):
((l['stop'] - l['start']) // l['increment']) // 3
for l in self.kernel.get_loop_stack(subs_consts=True)
}
warmup_iteration_count = self.kernel.indices_to_global_iterator(
warmup_indices)
# Make sure we are not handeling gigabytes of data, but 1.5x the maximum cache size
while warmup_iteration_count * element_size > max_cache_size * 1.5:
# Decreasing indices (starting from outer), until total size is small enough
for l in self.kernel.get_loop_stack():
index = symbol_pos_int(l['index'])
if warmup_indices[index] > l['start']:
warmup_indices[index] -= 1
# get offset in iterations and size that a change on this level provokes
diff_iterations = (
warmup_iteration_count -
self.kernel.indices_to_global_iterator(warmup_indices))
diff_size = diff_iterations * element_size
warmup_indices[index] = max(
l['start'],
(warmup_iteration_count -
(max_cache_size * 1.5) // element_size) // diff_size)
break
warmup_iteration_count = self.kernel.indices_to_global_iterator(
warmup_indices)
offsets = []
if max_array_size < 2 * max_cache_size:
# Full caching possible, go through all itreration before actual initialization
offsets = list(
self.kernel.compile_global_offsets(
iteration=range(0, self.kernel.iteration_length())))
# Align iteration count with cachelines
# do this by aligning either writes (preferred) or reads
# Assumption: writes (and reads) increase linearly
o = list(
self.kernel.compile_global_offsets(
iteration=warmup_iteration_count))[0]
if o[1]:
# we have a write to work with:
first_offset = min(o[1])
else:
# we use reads
first_offset = min(o[0])
# Distance from cacheline boundary (in bytes)
diff = first_offset - \
(int(first_offset) >> csim.first_level.cl_bits << csim.first_level.cl_bits)
warmup_iteration_count -= (diff // element_size) // inner_increment
warmup_indices = self.kernel.global_iterator_to_indices(
warmup_iteration_count)
offsets += list(
self.kernel.compile_global_offsets(
iteration=range(0, warmup_iteration_count)))
# Do the warm-up
csim.loadstore(offsets, length=element_size)
# FIXME compile_global_offsets should already expand to element_size
# Reset stats to conclude warm-up phase
csim.reset_stats()
# Benchmark iterations:
# Starting point is one past the last warmup element
bench_iteration_start = warmup_iteration_count
# End point is the end of the current dimension (cacheline alligned)
first_dim_factor = int(
(inner_loop['stop'] - warmup_indices[inner_index] - 1) //
(elements_per_cacheline // inner_increment))
# If end point is less than one cacheline away, go beyond for 100 cachelines and
# warn user of potentially inaccurate results
if first_dim_factor == 0:
# TODO a nicer solution would be to do less warmup iterations to select a
# cacheline within a first dimension, if possible
print(
'Warning: (automatic) warmup vs benchmark iteration choice was not perfect '
'and may lead to inaccurate cache miss predictions. This is most likely the '
'result of too few inner loop iterations ({} from {} to {}).'.
format(inner_loop['index'], inner_loop['start'],
inner_loop['stop']))
first_dim_factor = 100
bench_iteration_end = (
bench_iteration_start +
elements_per_cacheline * inner_increment * first_dim_factor)
# compile access needed for one cache-line
offsets = list(
self.kernel.compile_global_offsets(
iteration=range(bench_iteration_start, bench_iteration_end)))
# simulate
csim.loadstore(offsets, length=element_size)
# FIXME compile_global_offsets should already expand to element_size
# use stats to build results
self.stats = list(csim.stats())
self.first_dim_factor = first_dim_factor
def __init__(self, kernel, machine, cores=1):
"""Initialize layer condition based predictor from kernel and machine object."""
CachePredictor.__init__(self, kernel, machine, cores=cores)
# check that layer conditions can be applied on this kernel:
# 1. All iterations may only have a step width of 1
loop_stack = list(self.kernel.get_loop_stack())
if any([l['increment'] != 1 for l in loop_stack]):
raise ValueError(
"Can not apply layer condition, since not all loops are of step "
"length 1.")
# 2. The order of iterations must be reflected in the order of indices in all array
# references containing the inner loop index. If the inner loop index is not part of the
# reference, the reference is simply ignored
index_order = [symbol_pos_int(l['index']) for l in loop_stack]
for var_name, arefs in chain(self.kernel.sources.items(),
self.kernel.destinations.items()):
if next(iter(arefs)) is None:
# Anything that is a scalar may be ignored
continue
for a in [self.kernel.access_to_sympy(var_name, a) for a in arefs]:
for t in a.expand().as_ordered_terms():
# Check each and every term if they are valid according to loop order and array
# initialization
idx = t.free_symbols.intersection(index_order)
# Terms without any indices can be treat as constant offsets and are acceptable
if not idx:
continue
if len(idx) != 1:
raise ValueError(
"Only one loop counter may appear per term. "
"Problematic term: {}.".format(t))
else: # len(idx) == 1
idx = idx.pop()
# Check that number of multiplication match access order of iterator
pow_dict = {
k: v
for k, v in t.as_powers_dict().items() if k != idx
}
stride_dim = sum(pow_dict.values())
error = False
try:
if loop_stack[-stride_dim -
1]['index'] != idx.name:
error = True
except IndexError:
error = True
if error:
raise ValueError(
"Number of multiplications in index term does not "
"match loop counter order. "
"Problematic term: {}.".format(t))
# 3. Indices may only increase with one
inner_index = symbol_pos_int(loop_stack[-1]['index'])
inner_increment = loop_stack[-1]['increment']
# TODO use a public interface, not self.kernel._*
for arefs in chain(chain(*self.kernel.sources.values()),
chain(*self.kernel.destinations.values())):
if arefs is None:
continue
for expr in arefs:
diff = expr.subs(inner_index, 1 + inner_increment) - expr.subs(
inner_index, 1)
if diff != 0 and diff != 1:
# TODO support -1 aswell
raise ValueError(
"Can not apply layer condition, array references may not "
"increment more then one per iteration.")
# FIXME handle multiple datatypes
element_size = self.kernel.datatypes_size[self.kernel.datatype]
accesses = {}
destinations = set()
distances = []
for var_name in self.kernel.variables:
# Gather all access to current variable/array
accesses[var_name] = self.kernel.sources.get(
var_name,
set()).union(self.kernel.destinations.get(var_name, set()))
# Skip non-variable offsets, where acs is [None, None, None] (or similar) or only made
# up from constant offsets
if not any(accesses[var_name]) or not any([
a == inner_index or a.coeff(inner_index) != 0
for a in chain.from_iterable(accesses[var_name])
]):
continue
destinations.update([
(var_name, tuple(r))
for r in self.kernel.destinations.get(var_name, [])
])
acs = accesses[var_name]
# Transform them into sympy expressions
acs = [self.kernel.access_to_sympy(var_name, r) for r in acs]
# Replace constants with their integer counter parts, to make the entries sortable
acs = [self.kernel.subs_consts(e) for e in acs]
# Sort accesses by decreasing order
acs.sort(reverse=True)
# Create reuse distances by substracting accesses pairwise in decreasing order
distances += [(acs[i - 1] - acs[i]).simplify()
for i in range(1, len(acs))]
# Add infinity for each array
distances.append(sympy.oo)
# Sort distances by decreasing order
distances.sort(reverse=True)
# Create copy of distances in bytes:
distances_bytes = [d * element_size for d in distances]
# CAREFUL! From here on we are working in byte offsets and not in indices anymore.
# converting access sets to lists, otherwise pprint will fail during obligatory sorting step
results = {
'accesses': {k: list(v)
for k, v in accesses.items()},
'distances': distances,
'destinations': destinations,
'distances_bytes': distances_bytes,
'cache': []
}
sum_array_sizes = sum(
self.kernel.array_sizes(in_bytes=True, subs_consts=True).values())
for c in self.machine.get_cachesim(self.cores).levels(with_mem=False):
# Assuming increasing order of cache sizes
hits = 0
misses = len(distances_bytes)
cache_requirement = 0
tail = 0
# Test for full caching
if c.size() > sum_array_sizes:
hits = misses
misses = 0
cache_requirement = sum_array_sizes
else:
for tail in sorted(set(distances_bytes), reverse=True):
# Assuming decreasing order of tails
# Ignoring infinity tail:
if tail is sympy.oo:
continue
cache_requirement = (
# Sum of inter-access caches
sum([d for d in distances_bytes if d <= tail]) +
# Tails
tail * len([d for d in distances_bytes if d > tail]))
if cache_requirement <= c.size():
# If we found a tail that fits into our available cache size
# note hits and misses and break
hits = len([d for d in distances_bytes if d <= tail])
misses = len([d for d in distances_bytes if d > tail])
break
# Resulting analysis for current cache level
# TODO include loads and stores
results['cache'].append({
'name':
c.name,
'hits':
hits,
'misses':
misses,
'evicts':
len(destinations) if c.size() < sum_array_sizes else 0,
'requirement':
cache_requirement,
'tail':
tail
})
self.results = results
def __init__(self, kernel, machine, cores=1):
"""Initialize layer condition based predictor from kernel and machine object."""
CachePredictor.__init__(self, kernel, machine, cores=cores)
# check that layer conditions can be applied on this kernel:
# 1. All iterations may only have a step width of 1
loop_stack = list(self.kernel.get_loop_stack())
if any([l['increment'] != 1 for l in loop_stack]):
raise ValueError("Can not apply layer condition, since not all loops are of step "
"length 1.")
# 2. The order of iterations must be reflected in the order of indices in all array
# references containing the inner loop index. If the inner loop index is not part of the
# reference, the reference is simply ignored
index_order = [symbol_pos_int(l['index']) for l in loop_stack]
for var_name, arefs in chain(self.kernel.sources.items(), self.kernel.destinations.items()):
if next(iter(arefs)) is None:
# Anything that is a scalar may be ignored
continue
for a in [self.kernel.access_to_sympy(var_name, a) for a in arefs]:
for t in a.expand().as_ordered_terms():
# Check each and every term if they are valid according to loop order and array
# initialization
idx = t.free_symbols.intersection(index_order)
# Terms without any indices can be treat as constant offsets and are acceptable
if not idx:
continue
if len(idx) != 1:
raise ValueError("Only one loop counter may appear per term. "
"Problematic term: {}.".format(t))
else: # len(idx) == 1
idx = idx.pop()
# Check that number of multiplication match access order of iterator
pow_dict = {k: v for k, v in t.as_powers_dict().items()
if k != idx}
stride_dim = sum(pow_dict.values())
error = False
try:
if loop_stack[-stride_dim-1]['index'] != idx.name:
error = True
except IndexError:
error = True
if error:
raise ValueError("Number of multiplications in index term does not "
"match loop counter order. "
"Problematic term: {}.".format(t))
# 3. Indices may only increase with one
inner_index = symbol_pos_int(loop_stack[-1]['index'])
inner_increment = loop_stack[-1]['increment']
# TODO use a public interface, not self.kernel._*
for arefs in chain(chain(*self.kernel.sources.values()),
chain(*self.kernel.destinations.values())):
if arefs is None:
continue
for expr in arefs:
diff = expr.subs(inner_index, 1+inner_increment) - expr.subs(inner_index, 1)
if diff != 0 and diff != 1:
# TODO support -1 aswell
raise ValueError("Can not apply layer condition, array references may not "
"increment more then one per iteration.")
# FIXME handle multiple datatypes
element_size = self.kernel.datatypes_size[self.kernel.datatype]
accesses = {}
destinations = set()
distances = []
for var_name in self.kernel.variables:
# Gather all access to current variable/array
accesses[var_name] = self.kernel.sources.get(var_name, set()).union(
self.kernel.destinations.get(var_name, set()))
# Skip non-variable offsets, where acs is [None, None, None] (or similar) or only made
# up from constant offsets
if not any(accesses[var_name]) or not any(
[a == inner_index or a.coeff(inner_index) != 0
for a in chain.from_iterable(accesses[var_name])]):
continue
destinations.update(
[(var_name, tuple(r)) for r in self.kernel.destinations.get(var_name, [])])
acs = accesses[var_name]
# Transform them into sympy expressions
acs = [self.kernel.access_to_sympy(var_name, r) for r in acs]
# Replace constants with their integer counter parts, to make the entries sortable
acs = [self.kernel.subs_consts(e) for e in acs]
# Sort accesses by decreasing order
acs.sort(reverse=True)
# Create reuse distances by substracting accesses pairwise in decreasing order
distances += [(acs[i-1]-acs[i]).simplify() for i in range(1, len(acs))]
# Add infinity for each array
distances.append(sympy.oo)
# Sort distances by decreasing order
distances.sort(reverse=True)
# Create copy of distances in bytes:
distances_bytes = [d*element_size for d in distances]
# CAREFUL! From here on we are working in byte offsets and not in indices anymore.
# converting access sets to lists, otherwise pprint will fail during obligatory sorting step
results = {'accesses': {k: list(v) for k,v in accesses.items()},
'distances': distances,
'destinations': destinations,
'distances_bytes': distances_bytes,
'cache': []}
sum_array_sizes = sum(self.kernel.array_sizes(in_bytes=True, subs_consts=True).values())
for c in self.machine.get_cachesim(self.cores).levels(with_mem=False):
# Assuming increasing order of cache sizes
hits = 0
misses = len(distances_bytes)
cache_requirement = 0
tail = 0
# Test for full caching
if c.size() > sum_array_sizes:
hits = misses
misses = 0
cache_requirement = sum_array_sizes
else:
for tail in sorted(set(distances_bytes), reverse=True):
# Assuming decreasing order of tails
# Ignoring infinity tail:
if tail is sympy.oo:
continue
cache_requirement = (
# Sum of inter-access caches
sum([d for d in distances_bytes if d <= tail]) +
# Tails
tail*len([d for d in distances_bytes if d > tail]))
if cache_requirement <= c.size():
# If we found a tail that fits into our available cache size
# note hits and misses and break
hits = len([d for d in distances_bytes if d <= tail])
misses = len([d for d in distances_bytes if d > tail])
break
# Resulting analysis for current cache level
# TODO include loads and stores
results['cache'].append({
'name': c.name,
'hits': hits,
'misses': misses,
'evicts': len(destinations) if c.size() < sum_array_sizes else 0,
'requirement': cache_requirement,
'tail': tail})
self.results = results
def __init__old(self, kernel, machine, cores=1):
"""Initialize cache simulation based predictor from kernel and machine object."""
CachePredictor.__init__(self, kernel, machine, cores)
# Get the machine's cache model and simulator
csim = self.machine.get_cachesim(self.cores)
# FIXME handle multiple datatypes
element_size = self.kernel.datatypes_size[self.kernel.datatype]
cacheline_size = self.machine['cacheline size']
elements_per_cacheline = int(cacheline_size // element_size)
# Gathering some loop information:
inner_loop = list(self.kernel.get_loop_stack(subs_consts=True))[-1]
inner_index = symbol_pos_int(inner_loop['index'])
inner_increment = inner_loop['increment'] # Calculate the number of iterations for warm-up
max_cache_size = max(map(lambda c: c.size(), csim.levels(with_mem=False)))
max_array_size = max(self.kernel.array_sizes(in_bytes=True, subs_consts=True).values())
# Regular Initialization
warmup_indices = {
symbol_pos_int(l['index']): ((l['stop']-l['start'])//l['increment'])//3
for l in self.kernel.get_loop_stack(subs_consts=True)}
warmup_iteration_count = self.kernel.indices_to_global_iterator(warmup_indices)
# Make sure we are not handeling gigabytes of data, but 1.5x the maximum cache size
while warmup_iteration_count * element_size > max_cache_size*1.5:
# Decreasing indices (starting from outer), until total size is small enough
for l in self.kernel.get_loop_stack():
index = symbol_pos_int(l['index'])
if warmup_indices[index] > l['start']:
warmup_indices[index] -= 1
# get offset in iterations and size that a change on this level provokes
diff_iterations = (warmup_iteration_count -
self.kernel.indices_to_global_iterator(warmup_indices))
diff_size = diff_iterations*element_size
warmup_indices[index] = max(
l['start'],
(warmup_iteration_count - (max_cache_size*1.5) // element_size) // diff_size
)
break
warmup_iteration_count = self.kernel.indices_to_global_iterator(warmup_indices)
# Align iteration count with cachelines
# do this by aligning either writes (preferred) or reads
# Assumption: writes (and reads) increase linearly
o = self.kernel.compile_global_offsets(iteration=warmup_iteration_count)[0]
if len(o[1]):
# we have a write to work with:
first_offset = min(o[1])
else:
# we use reads
first_offset = min(o[0])
# Distance from cacheline boundary (in bytes)
diff = first_offset - \
(int(first_offset) >> csim.first_level.cl_bits << csim.first_level.cl_bits)
warmup_iteration_count -= (diff//element_size)//inner_increment
warmup_indices = self.kernel.global_iterator_to_indices(warmup_iteration_count)
offsets = self.kernel.compile_global_offsets(
iteration=range(0, warmup_iteration_count))
if max_array_size < 2*max_cache_size:
# Full caching possible, go through all itreration before actual initialization
offsets = np.concatenate((self.kernel.compile_global_offsets(
iteration=range(0, self.kernel.iteration_length())),
offsets))
# Do the warm-up
csim.loadstore(offsets, length=element_size)
# FIXME compile_global_offsets should already expand to element_size
# Reset stats to conclude warm-up phase
csim.reset_stats()
# Benchmark iterations:
# Starting point is one past the last warmup element
bench_iteration_start = warmup_iteration_count
# End point is the end of the current dimension (cacheline alligned)
first_dim_factor = int((inner_loop['stop'] - warmup_indices[inner_index] - 1)
// (elements_per_cacheline // inner_increment))
# If end point is less than one cacheline away, go beyond for 100 cachelines and
# warn user of potentially inaccurate results
if first_dim_factor == 0:
# TODO a nicer solution would be to do less warmup iterations to select a
# cacheline within a first dimension, if possible
print('Warning: (automatic) warmup vs benchmark iteration choice was not perfect '
'and may lead to inaccurate cache miss predictions. This is most likely the '
'result of too few inner loop iterations ({} from {} to {}).'.format(
inner_loop['index'], inner_loop['start'], inner_loop['stop']
))
first_dim_factor = 100
bench_iteration_end = (bench_iteration_start +
elements_per_cacheline * inner_increment * first_dim_factor)
# compile access needed for one cache-line
offsets = self.kernel.compile_global_offsets(
iteration=range(bench_iteration_start, bench_iteration_end))
# simulate
csim.loadstore(offsets, length=element_size)
# FIXME compile_global_offsets should already expand to element_size
# use stats to build results
self.stats = list(csim.stats())
self.first_dim_factor = first_dim_factor
def __init__(self, kernel, machine, cores=1):
"""Initialize cache simulation based predictor from kernel and machine object."""
CachePredictor.__init__(self, kernel, machine, cores)
# Get the machine's cache model and simulator
self.csim = self.machine.get_cachesim(self.cores)
# FIXME handle multiple datatypes
element_size = self.kernel.datatypes_size[self.kernel.datatype]
cacheline_size = self.machine['cacheline size']
elements_per_cacheline = int(cacheline_size // element_size)
iterations_per_cacheline = (sympy.Integer(self.machine['cacheline size']) /
sympy.Integer(self.kernel.bytes_per_iteration))
# Gathering some loop information:
inner_loop = list(self.kernel.get_loop_stack(subs_consts=True))[-1]
inner_index = symbol_pos_int(inner_loop['index'])
inner_increment = inner_loop['increment'] # Calculate the number of iterations for warm-up
total_length = self.kernel.iteration_length()
max_iterations = self.kernel.subs_consts(total_length)
max_cache_size = sum([c.size() for c in self.csim.levels(with_mem=False)])
# Warmup
# Phase 1:
# define warmup interval boundaries
max_steps = 100
warmup_increment = ceildiv(max_cache_size // element_size, max_steps // 2)
invalid_entries = self.csim.count_invalid_entries()
step = 0
warmup_iteration = 0
complete_sweep = False
while invalid_entries > 0 and step < max_steps and not complete_sweep:
prev_warmup_iteration = warmup_iteration
warmup_iteration = warmup_iteration + warmup_increment
if warmup_iteration > max_iterations:
warmup_iteration = max_iterations
complete_sweep = True
# print("warmup_iteration1", warmup_iteration)
offsets = self.kernel.compile_global_offsets(
iteration=range(prev_warmup_iteration, warmup_iteration))
self.csim.loadstore(offsets, length=element_size)
invalid_entries = self.csim.count_invalid_entries()
# TODO more intelligent break criteria based on change of invalid entries might be
# useful for early termination.
# print("invalid_entries", invalid_entries)
step += 1
# Phase 2:
# Check that there is enough space left for benchmarking
if complete_sweep:
warmup_iteration = 0
else:
# Are we far away from max_iterations?
# print("max_iterations", max_iterations)
if warmup_iteration > max_iterations - 100000:
# To close to end, need to complete sweep
complete_sweep = True
prev_warmup_iteration = warmup_iteration
warmup_iteration = max_iterations
# print("warmup_iteration2", warmup_iteration, end="; ")
offsets = self.kernel.compile_global_offsets(
iteration=range(prev_warmup_iteration, warmup_iteration))
self.csim.loadstore(offsets, length=element_size)
warmup_iteration = 0
if not complete_sweep and invalid_entries > 0:
print("Warning: Unable to perform complete sweep nor initialize cache completely. "
"This might introduce inaccuracies (additional cache misses) in the cache "
"prediction.")
# Phase 3:
# Iterate to safe handover point
prev_warmup_iteration = warmup_iteration
warmup_iteration = self._align_iteration_with_cl_boundary(warmup_iteration, subtract=False)
if warmup_iteration != prev_warmup_iteration:
# print("warmup_iteration3", warmup_iteration)
offsets = self.kernel.compile_global_offsets(
iteration=range(prev_warmup_iteration, warmup_iteration))
self.csim.loadstore(offsets, length=element_size)
# Reset stats to conclude warm-up phase
self.csim.reset_stats()
# Benchmark
bench_iteration = self._align_iteration_with_cl_boundary(min(
warmup_iteration + 100000, max_iterations - 1))
# print("bench_iteration", bench_iteration)
first_dim_factor = float((bench_iteration - warmup_iteration) / iterations_per_cacheline)
# If end point is less than 100 cacheline away, warn user of inaccuracy
if first_dim_factor < 1000:
print("Warning: benchmark iterations are very low ({} CL). This may lead to inaccurate "
"cache predictions.".format(first_dim_factor))
# Compile access needed for one cache-line
offsets = self.kernel.compile_global_offsets(
iteration=range(warmup_iteration, bench_iteration))
# Run cache simulation
self.csim.loadstore(offsets, length=element_size)
# FIXME compile_global_offsets should already expand to element_size
# use stats to build results
self.stats = list(self.csim.stats())
self.first_dim_factor = first_dim_factor
def __init__(self, kernel, machine, cores=1):
"""Initialize cache simulation based predictor from kernel and machine object."""
CachePredictor.__init__(self, kernel, machine, cores)
# Get the machine's cache model and simulator
self.csim = self.machine.get_cachesim(self.cores)
# FIXME handle multiple datatypes
element_size = self.kernel.datatypes_size[self.kernel.datatype]
cacheline_size = self.machine['cacheline size']
elements_per_cacheline = int(cacheline_size // element_size)
# Gathering some loop information:
inner_loop = list(self.kernel.get_loop_stack(subs_consts=True))[-1]
inner_index = symbol_pos_int(inner_loop['index'])
inner_increment = inner_loop[
'increment'] # Calculate the number of iterations for warm-up
total_length = self.kernel.iteration_length()
max_iterations = self.kernel.subs_consts(total_length)
max_cache_size = sum(
[c.size() for c in self.csim.levels(with_mem=False)])
# Warmup
# Phase 1:
# define warmup interval boundaries
max_steps = 100
warmup_increment = ceildiv(max_cache_size // element_size, max_steps)
invalid_entries = self.csim.count_invalid_entries()
step = 0
warmup_iteration = 0
complete_sweep = False
while invalid_entries > 0 and step < max_steps and not complete_sweep:
prev_warmup_iteration = warmup_iteration
warmup_iteration = warmup_iteration + warmup_increment
if warmup_iteration > max_iterations:
warmup_iteration = max_iterations
complete_sweep = True
# print("warmup_iteration1", warmup_iteration)
offsets = self.kernel.compile_global_offsets(
iteration=range(prev_warmup_iteration, warmup_iteration))
self.csim.loadstore(offsets, length=element_size)
invalid_entries = self.csim.count_invalid_entries()
# TODO more intelligent break criteria based on change of invalid entries might be
# useful for early termination.
# print("invalid_entries", invalid_entries)
step += 1
# Phase 2:
# Check that there is enough space left for benchmarking
if complete_sweep:
warmup_iteration = 0
else:
# Are we far away from max_iterations?
# print("max_iterations", max_iterations)
if warmup_iteration > max_iterations - 100000:
# To close to end, need to complete sweep
complete_sweep = True
prev_warmup_iteration = warmup_iteration
warmup_iteration = max_iterations
# print("warmup_iteration2", warmup_iteration, end="; ")
offsets = self.kernel.compile_global_offsets(
iteration=range(prev_warmup_iteration, warmup_iteration))
self.csim.loadstore(offsets, length=element_size)
warmup_iteration = 0
if not complete_sweep and invalid_entries > 0:
print(
"Warning: Unable to perform complete sweep nor initialize cache completely. "
"This might introduce inaccuracies (additional cache misses) in the cache "
"prediction.")
# Phase 3:
# Iterate to safe handover point
prev_warmup_iteration = warmup_iteration
warmup_iteration = self._align_iteration_with_cl_boundary(
warmup_iteration, subtract=False)
if warmup_iteration != prev_warmup_iteration:
# print("warmup_iteration3", warmup_iteration)
offsets = self.kernel.compile_global_offsets(
iteration=range(prev_warmup_iteration, warmup_iteration))
self.csim.loadstore(offsets, length=element_size)
# Reset stats to conclude warm-up phase
self.csim.reset_stats()
# Benchmark
bench_iteration = self._align_iteration_with_cl_boundary(
min(warmup_iteration + 100000, max_iterations - 1))
# print("bench_iteration", bench_iteration)
first_dim_factor = float(
(bench_iteration - warmup_iteration) / elements_per_cacheline)
# If end point is less than 100 cacheline away, warn user of inaccuracy
if first_dim_factor < 1000:
print(
"Warning: benchmark iterations are very low ({} CL). This may lead to inaccurate "
"cache predictions.".format(first_dim_factor))
# Compile access needed for one cache-line
offsets = self.kernel.compile_global_offsets(
iteration=range(warmup_iteration, bench_iteration))
# Run cache simulation
self.csim.loadstore(offsets, length=element_size)
# FIXME compile_global_offsets should already expand to element_size
# use stats to build results
self.stats = list(self.csim.stats())
self.first_dim_factor = first_dim_factor
def __init__(self, kernel, machine, cores=1, symbolic=False):
"""Initialize layer condition based predictor from kernel and machine object."""
CachePredictor.__init__(self, kernel, machine, cores=cores)
# check that layer conditions can be applied on this kernel:
# 1. All iterations may only have a step width of 1
loop_stack = list(self.kernel.get_loop_stack())
if any([l['increment'] != 1 for l in loop_stack]):
raise ValueError("Can not apply layer condition, since not all loops are of step "
"length 1.")
# 2. The order of iterations must be reflected in the order of indices in all array
# references containing the inner loop index. If the inner loop index is not part of the
# reference, the reference is simply ignored
index_order = [symbol_pos_int(l['index']) for l in loop_stack]
for var_name, arefs in chain(self.kernel.sources.items(), self.kernel.destinations.items()):
if next(iter(arefs)) is None:
# Anything that is a scalar may be ignored
continue
for a in [self.kernel.access_to_sympy(var_name, a) for a in arefs]:
for t in a.expand().as_ordered_terms():
# Check each and every term if they are valid according to loop order and array
# initialization
idx = t.free_symbols.intersection(index_order)
# Terms without any indices can be treat as constant offsets and are acceptable
if not idx:
continue
if len(idx) != 1:
raise ValueError("Only one loop counter may appear per term. "
"Problematic term: {}.".format(t))
else: # len(idx) == 1
idx = idx.pop()
# Check that number of multiplication match access order of iterator
pow_dict = {k: v for k, v in t.as_powers_dict().items()
if k != idx}
stride_dim = sum(pow_dict.values())
error = False
try:
if loop_stack[-stride_dim-1]['index'] != idx.name:
error = True
except IndexError:
error = True
if error:
raise ValueError("Number of multiplications in index term does not "
"match loop counter order. "
"Problematic term: {}.".format(t))
# 3. Indices may only increase with one
inner_index = symbol_pos_int(loop_stack[-1]['index'])
inner_increment = loop_stack[-1]['increment']
# TODO use a public interface, not self.kernel._*
for aref in chain(chain(*self.kernel.sources.values()),
chain(*self.kernel.destinations.values())):
if aref is None:
continue
for expr in aref:
diff = expr.subs(inner_index, 1+inner_increment) - expr.subs(inner_index, 1)
if diff != 0 and diff != 1:
# TODO support -1 aswell
raise ValueError("Can not apply layer condition, array references may not "
"increment more then one per iteration.")
# FIXME handle multiple datatypes
element_size = self.kernel.datatypes_size[self.kernel.datatype]
indices = list([symbol_pos_int(l[0]) for l in self.kernel._loop_stack])
sympy_accesses = self.kernel.compile_sympy_accesses()
accesses = {}
destinations = set()
distances = []
for var_name in sorted(self.kernel.variables):
# Gather all access to current variable/array and delinearize accesses
accesses[var_name] = []
dimension_factors = None
for a in sympy_accesses[var_name]:
o, df = split_sympy_access_in_dim_offset_and_factor(a, indices)
accesses[var_name].append(o)
if dimension_factors is None:
dimension_factors = df
elif dimension_factors[-len(indices):] != df[-len(indices):]:
raise ValueError("Extracted dimension factors are different within one "
"variable. Must be a bug. {!r} != {!r}".format(
df, dimension_factors))
elif len(dimension_factors) < len(df):
dimension_factors = df
# Skip non-variable offsets, where acs is [None, None, None] (or similar) or only made
# up from constant offsets
if not any(accesses[var_name]) or not any(
[a == inner_index or a.coeff(inner_index) != 0
for a in chain.from_iterable(accesses[var_name])]):
continue
destinations.update(
[(var_name, tuple(r)) for r in self.kernel.destinations.get(var_name, [])])
acs = list(accesses[var_name])
# If accesses are of unequal length, pad with leading zero elements
max_dims = max(map(len, acs))
for i in range(len(acs)):
if len(acs[i]) < max_dims:
acs[i] = (sympy.Integer(0),)*(max_dims-len(acs[i])) + acs[i]
# Sort accesses by decreasing order
acs.sort(reverse=True)
# Transform back into sympy expressions
for i in range(len(acs)):
acs[i] = reduce(sympy.Add, [f*df for f, df in zip(acs[i], dimension_factors)])
# Create reuse distances by substracting accesses pairwise in decreasing order
distances += [(acs[i-1]-acs[i]).simplify() for i in range(1, len(acs))]
# Add infinity for each array
distances.append(sympy.oo)
# Sort distances by decreasing order
distances.sort(reverse=True, key=sympy_expr_abs_distance_key)
# Create copy of distances in bytes:
distances_bytes = [d*element_size for d in distances]
# CAREFUL! From here on we are working in byte offsets and not in indices anymore.
# converting access sets to lists, otherwise pprint will fail during obligatory sorting step
results = {'accesses': {k: sorted(list(v), key=cmp_to_key(uneven_tuple_cmp))
for k,v in accesses.items()},
'distances': distances,
'destinations': destinations,
'distances_bytes': distances_bytes,
'cache': []}
sum_array_sizes = sum(self.kernel.array_sizes(in_bytes=True, subs_consts=False).values())
for c in self.machine.get_cachesim(self.cores).levels(with_mem=False):
# Assuming increasing order of cache sizes
options = []
# Full caching
options.append({
'condition': (c.size() > sum_array_sizes).simplify().expand(),
'hits': len(distances),
'misses': 0,
'evicts': 0,
'tail': sympy.oo,
})
for tail in sorted(set([d.simplify().expand() for d in distances_bytes]), reverse=True,
key=sympy_expr_abs_distance_key):
# Assuming decreasing order of tails
# Ignoring infinity tail:
if tail is sympy.oo:
continue
cache_requirement = (
# Sum of inter-access caches
sum([d for d in distances_bytes
if sympy_expr_abs_distance_key(d) <= sympy_expr_abs_distance_key(tail)]
) +
# Tails
tail*len([d for d in distances_bytes
if sympy_expr_abs_distance_key(d) >
sympy_expr_abs_distance_key(tail)]))
condition = (cache_requirement <= c.size()).simplify().expand()
hits = len(
[d for d in distances_bytes
if sympy_expr_abs_distance_key(d) <= sympy_expr_abs_distance_key(tail)])
misses = len(
[d for d in distances_bytes
if sympy_expr_abs_distance_key(d) > sympy_expr_abs_distance_key(tail)])
# Resulting analysis
options.append({
'condition': condition,
'hits': hits,
'misses': misses,
'evicts': len(destinations),
'tail': tail})
# If we encountered a True condition, break to not include multiple such.
if isinstance(condition, BooleanTrue):
break
if not isinstance(options[-1]['condition'], BooleanTrue):
# Fallback: no condition matched
options.append({
'condition': True,
'hits': 0,
'misses': len(distances),
'evicts': len(destinations),
'tail': 0
})
if symbolic:
results['cache'].append(options)
else:
for o in options:
if self.kernel.subs_consts(o['condition']):
results['cache'].append(o)
break
self.results = results