def test_bufshr_wide_fetch_example(self):
''' Example scheme using bufshr with wide fetch. '''
# Make a PartitionScheme that allows bufshr for IFM.
part = PartitionScheme(order=range(pe.NUM),
pdims=((2, 2), (1, 1), (2, 1), (1, 1)))
bufshr = BufShrScheme(self.par_proc_region, part)
self.assertEqual(
bufshr.size(de.IFM), 4, 'test_bufshr_wide_fetch_example: '
'made-up PartitionScheme is not expected: '
'{}, bufshr size for {} {}.'.format(part, de.IFM,
bufshr.size(de.IFM)))
for t1, t2 in [((3, 3, 1), (1, 1, 2)), ((1, 3, 2), (3, 1, 1))]:
# Make a LoopBlockingScheme that has wide fetch for IFM.
p_nld = self._part_nld(part)
bl_ts = (tuple(
util.idivc(p_nld.loopcnt[lpe], t1[lpe] * t2[lpe])
for lpe in range(le.NUM)), t1, t2)
# At GBUF level, from inner to outer: le.BAT, le.IFM, le.OFM.
bl_ords = (tuple(range(le.NUM)), (1, 2, 0))
lbs = LoopBlockingScheme(p_nld, bl_ts, bl_ords,
self.resource['PAR'], bufshr,
self.options['BUFSHR'])
self.assertTrue(lbs.is_valid())
self.assertGreater(sum(lbs.get_noc_access()), 0)
self.assertEqual(
lbs.bufshr_subgrp_size[de.IFM], 4,
'test_bufshr_wide_fetch_example: '
'made-up LoopBlockingScheme is not expected: '
'{}, bufshr subgrp size for {} {}.'.format(
(bl_ts, bl_ords), de.IFM, lbs.bufshr_subgrp_size[de.IFM]))
self.assertGreater(
lbs.bufshr_wide_fetch_width[de.IFM], 1,
'test_bufshr_wide_fetch_example: '
'made-up LoopBlockingScheme is not expected: '
'{}, bufshr wide fetch width for {} {}.'.format(
(bl_ts, bl_ords), de.IFM,
lbs.bufshr_wide_fetch_width[de.IFM]))
self.assertGreater(
lbs.bufshr_rot_round_cnt[de.IFM], 0,
'test_bufshr_wide_fetch_example: '
'made-up LoopBlockingScheme is not expected: '
'{}, bufshr rotation rounds for {} {}'.format(
(bl_ts, bl_ords), de.IFM,
lbs.bufshr_rot_round_cnt[de.IFM]))
# Sim.
dram_access, gbuf_access, bufshr_stats = \
self._sim_access_conv(lbs, get_bufshr=True)
self._verify_bufshr_stats(dram_access, gbuf_access, bufshr_stats,
lbs, bufshr,
'test_bufshr_wide_fetch_example')
def test_bufshr_multisubgrp_example(self):
''' Example scheme using bufshr with multiple subgroups in a group. '''
# Make a PartitionScheme that allows bufshr for IFM.
part = PartitionScheme(order=list(reversed(range(pe.NUM))),
pdims=((2, 2), (1, 1), (2, 1), (1, 1)))
bufshr = BufShrScheme(self.par_proc_region, part)
self.assertEqual(
bufshr.size(de.IFM), 4, 'test_bufshr_multisubgrp_example: '
'made-up PartitionScheme is not expected: '
'{}, bufshr size for {} {}.'.format(part, de.IFM,
bufshr.size(de.IFM)))
# Make a LoopBlockingScheme that has multi subgroups per group for IFM.
p_nld = self._part_nld(part)
bl_ts = ((util.idivc(p_nld.loopcnt[le.IFM],
1), util.idivc(p_nld.loopcnt[le.OFM], 3),
util.idivc(p_nld.loopcnt[le.BAT], 2)), (1, 3, 2), (1, 1, 1))
# At GBUF level, from inner to outer: le.BAT, le.OFM, le.IFM.
bl_ords = (tuple(range(le.NUM)), (2, 1, 0))
lbs = LoopBlockingScheme(p_nld, bl_ts, bl_ords, self.resource['PAR'],
bufshr, self.options['BUFSHR'])
self.assertTrue(lbs.is_valid())
self.assertGreater(sum(lbs.get_noc_access()), 0)
self.assertGreater(
lbs.bufshr_grp_size[de.IFM], lbs.bufshr_subgrp_size[de.IFM],
'test_bufshr_multisubgrp_example: '
'made-up LoopBlockingScheme is not expected: '
'{}, bufshr grp size {}, bufshr subgrp size {}'.format(
(bl_ts, bl_ords), lbs.bufshr_grp_size, lbs.bufshr_subgrp_size))
self.assertGreater(
lbs.bufshr_rot_round_cnt[de.IFM], 0,
'test_bufshr_multisubgrp_example: '
'made-up LoopBlockingScheme is not expected: '
'{}, bufshr rotation rounds for {} {}'.format(
(bl_ts, bl_ords), de.IFM, lbs.bufshr_rot_round_cnt[de.IFM]))
# Sim.
dram_access, gbuf_access, bufshr_stats = \
self._sim_access_conv(lbs, get_bufshr=True)
self._verify_bufshr_stats(dram_access, gbuf_access, bufshr_stats, lbs,
bufshr, 'test_bufshr_multisubgrp_example')
def _lbs(self,
bl_ts,
bl_ords=None,
wlkey='BASE',
rsrckey='BASE',
optkey='BASE'):
''' Make a LoopBlockingScheme instance. '''
bl_ords = (tuple(range(le.NUM)), tuple(range(le.NUM))) \
if not bl_ords else bl_ords
return LoopBlockingScheme(self.nld[wlkey], bl_ts, bl_ords,
self.resource[rsrckey], self.bufshr,
self.options[optkey])
def test_bufshr_rotation_example(self):
''' Example scheme using bufshr with rotation. '''
# Make a PartitionScheme that allows bufshr for all data categories.
part = PartitionScheme(order=range(pe.NUM),
pdims=((2, 1), (1, 2), (1, 1), (2, 1)))
bufshr = BufShrScheme(self.par_proc_region, part)
self.assertTrue(
all(bufshr.size(dce) > 1 for dce in range(de.NUM)),
'test_bufshr_rotation_example: '
'made-up PartitionScheme is not expected: '
'{}, bufshr size {}'.format(
part, [bufshr.size(dce) for dce in range(de.NUM)]))
# Make a LoopBlockingScheme that uses bufshr for all data categories.
p_nld = self._part_nld(part)
bl_ts = ((util.idivc(p_nld.loopcnt[le.IFM],
6), util.idivc(p_nld.loopcnt[le.OFM], 9),
util.idivc(p_nld.loopcnt[le.BAT], 2)), (3, 3, 2), (2, 3, 1))
bl_ords = (tuple(range(le.NUM)), tuple(range(le.NUM)))
lbs = LoopBlockingScheme(p_nld, bl_ts, bl_ords, self.resource['PAR'],
bufshr, self.options['BUFSHR'])
self.assertTrue(lbs.is_valid())
self.assertGreater(sum(lbs.get_noc_access()), 0)
self.assertTrue(
all(sgs > 1 for sgs in lbs.bufshr_subgrp_size)
and all(t > 1 for t in bl_ts[0]), 'test_bufshr_rotation_example: '
'made-up LoopBlockingScheme is not expected: '
'{}, top factors {}, bufshr subgrp size {}'.format(
(bl_ts, bl_ords), bl_ts[0], lbs.bufshr_subgrp_size))
# Sim.
dram_access, gbuf_access, bufshr_stats = \
self._sim_access_conv(lbs, get_bufshr=True)
self._verify_bufshr_stats(dram_access, gbuf_access, bufshr_stats, lbs,
bufshr, 'test_bufshr_rotation_example')
def test_ordered_loops(self):
''' Get ordered_loops. '''
assert list(range(le.NUM)) == [le.IFM, le.OFM, le.BAT]
self.assertListEqual(
LoopBlockingScheme.ordered_loops((3, 5, 2), (2, 0, 1)),
[(le.IFM, 3), (le.BAT, 2), (le.OFM, 5)])
# Trivial loops at different positions.
self.assertListEqual(
LoopBlockingScheme.ordered_loops((3, 5, 1), (0, 1, 2)),
[(le.OFM, 5), (le.IFM, 3)])
self.assertListEqual(
LoopBlockingScheme.ordered_loops((3, 5, 1), (1, 2, 0)),
[(le.OFM, 5), (le.IFM, 3)])
self.assertListEqual(
LoopBlockingScheme.ordered_loops((3, 5, 1), (0, 2, 1)),
[(le.OFM, 5), (le.IFM, 3)])
# Different loops are trivial.
self.assertListEqual(
LoopBlockingScheme.ordered_loops((1, 5, 2), (0, 2, 1)),
[(le.OFM, 5), (le.BAT, 2)])
self.assertListEqual(
LoopBlockingScheme.ordered_loops((3, 1, 2), (0, 2, 1)),
[(le.BAT, 2), (le.IFM, 3)])
# Multiple trivial loops.
self.assertListEqual(
LoopBlockingScheme.ordered_loops((1, 5, 1), (0, 1, 2)),
[(le.OFM, 5)])
self.assertListEqual(
LoopBlockingScheme.ordered_loops((1, 1, 1), (0, 1, 2)),
[])
for bl_t, bl_ord in itertools.product(
itertools.product(*[range(1, 8)] * 3),
itertools.permutations(range(le.NUM))):
ord_loops = LoopBlockingScheme.ordered_loops(bl_t, bl_ord)
self.assertTrue(all(len(tpl) == 2 for tpl in ord_loops))
self.assertFalse(any(tpl[1] <= 1 for tpl in ord_loops))
self.assertEqual(len(ord_loops), le.NUM - bl_t.count(1))
self.assertTrue(all(tpl[1] == bl_t[tpl[0]] for tpl in ord_loops))
rev_loops = LoopBlockingScheme.ordered_loops(bl_t, bl_ord,
reverse=True)
ord_lpes = LoopBlockingScheme.ordered_loops(bl_t, bl_ord,
lpe_only=True)
self.assertEqual(len(rev_loops), len(ord_loops))
self.assertEqual(len(ord_lpes), len(ord_loops))
self.assertListEqual(list(reversed(rev_loops)), ord_loops)
self.assertListEqual([tpl[0] for tpl in ord_loops], ord_lpes)