def test_gather_access_footprint_2():
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"c[2*i] = a[i]",
name="matmul", assumptions="n >= 1")
knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32))
from loopy.statistics import gather_access_footprints, count
fp = gather_access_footprints(knl)
params = {"n": 200}
for key, footprint in six.iteritems(fp):
assert count(knl, footprint).eval_with_dict(params) == 200
print(key, count(knl, footprint))
def test_gather_access_footprint_2():
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"c[2*i] = a[i]",
name="matmul", assumptions="n >= 1")
knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32))
from loopy.statistics import gather_access_footprints, count
fp = gather_access_footprints(knl)
params = {"n": 200}
for key, footprint in six.iteritems(fp):
assert count(knl, footprint).eval_with_dict(params) == 200
print(key, count(knl, footprint))
def test_strided_footprint():
param_dict = dict(n=2**20)
knl = lp.make_kernel("[n] -> {[i]: 0<=i<n}", ["z[i] = x[3*i]"], name="s3")
knl = lp.add_and_infer_dtypes(knl, dict(x=np.float32))
unr = 4
bx = 256
knl = lp.split_iname(knl, "i", bx * unr, outer_tag="g.0", slabs=(0, 1))
knl = lp.split_iname(knl, "i_inner", bx, outer_tag="unr", inner_tag="l.0")
footprints = lp.gather_access_footprints(knl)
x_l_foot = footprints[('x', 'read')]
from loopy.statistics import count
num = count(knl, x_l_foot).eval_with_dict(param_dict)
denom = count(knl, x_l_foot.remove_divs()).eval_with_dict(param_dict)
assert 2 * num < denom
def test_gather_access_footprint():
knl = lp.make_kernel("{[i,k,j]: 0<=i,j,k<n}",
["c[i, j] = sum(k, a[i, k]*b[k, j]) + a[i,j]"],
name="matmul",
assumptions="n >= 1")
knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32, b=np.float32))
from loopy.statistics import gather_access_footprints, count
fp = gather_access_footprints(knl)
for key, footprint in six.iteritems(fp):
print(key, count(knl, footprint))
def test_gather_access_footprint():
knl = lp.make_kernel(
"{[i,k,j]: 0<=i,j,k<n}",
[
"c[i, j] = sum(k, a[i, k]*b[k, j]) + a[i,j]"
],
name="matmul", assumptions="n >= 1")
knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32, b=np.float32))
from loopy.statistics import gather_access_footprints, count
fp = gather_access_footprints(knl)
for key, footprint in six.iteritems(fp):
print(key, count(knl, footprint))