def test_fp_prec_comparison():
# FIXME: This test should succeed even when the number is exactly
# representable in single precision.
#
# https://gitlab.tiker.net/inducer/loopy/issues/187
fortran_src_dp = """
subroutine assign_scalar(a)
real*8 a(1)
a(1) = 1.1d0
end
"""
prg_dp = lp.parse_fortran(fortran_src_dp)
fortran_src_sp = """
subroutine assign_scalar(a)
real*8 a(1)
a(1) = 1.1
end
"""
prg_sp = lp.parse_fortran(fortran_src_sp)
assert prg_sp != prg_dp
def test_fuse_kernels(ctx_factory):
fortran_template = """
subroutine {name}(nelements, ndofs, result, d, q)
implicit none
integer e, i, j, k
integer nelements, ndofs
real*8 result(nelements, ndofs, ndofs)
real*8 q(nelements, ndofs, ndofs)
real*8 d(ndofs, ndofs)
real*8 prev
do e = 1,nelements
do i = 1,ndofs
do j = 1,ndofs
do k = 1,ndofs
{inner}
end do
end do
end do
end do
end subroutine
"""
xd_line = """
prev = result(e,i,j)
result(e,i,j) = prev + d(i,k)*q(e,i,k)
"""
yd_line = """
prev = result(e,i,j)
result(e,i,j) = prev + d(i,k)*q(e,k,j)
"""
xderiv, = lp.parse_fortran(
fortran_template.format(inner=xd_line, name="xderiv"))
yderiv, = lp.parse_fortran(
fortran_template.format(inner=yd_line, name="yderiv"))
xyderiv, = lp.parse_fortran(
fortran_template.format(inner=(xd_line + "\n" + yd_line),
name="xyderiv"))
knl = lp.fuse_kernels((xderiv, yderiv))
knl = lp.prioritize_loops(knl, "e,i,j,k")
assert len(knl.temporary_variables) == 2
# This is needed for correctness, otherwise ordering could foul things up.
knl = lp.assignment_to_subst(knl, "prev")
knl = lp.assignment_to_subst(knl, "prev_0")
ctx = ctx_factory()
lp.auto_test_vs_ref(xyderiv,
ctx,
knl,
parameters=dict(nelements=20, ndofs=4))
def test_fuse_kernels(ctx_factory):
fortran_template = """
subroutine {name}(nelements, ndofs, result, d, q)
implicit none
integer e, i, j, k
integer nelements, ndofs
real*8 result(nelements, ndofs, ndofs)
real*8 q(nelements, ndofs, ndofs)
real*8 d(ndofs, ndofs)
real*8 prev
do e = 1,nelements
do i = 1,ndofs
do j = 1,ndofs
do k = 1,ndofs
{inner}
end do
end do
end do
end do
end subroutine
"""
xd_line = """
prev = result(e,i,j)
result(e,i,j) = prev + d(i,k)*q(e,i,k)
"""
yd_line = """
prev = result(e,i,j)
result(e,i,j) = prev + d(i,k)*q(e,k,j)
"""
xderiv = lp.parse_fortran(
fortran_template.format(inner=xd_line, name="xderiv"))
yderiv = lp.parse_fortran(
fortran_template.format(inner=yd_line, name="yderiv"))
xyderiv = lp.parse_fortran(
fortran_template.format(inner=(xd_line + "\n" + yd_line),
name="xyderiv"))
knl = lp.fuse_kernels((xderiv["xderiv"], yderiv["yderiv"]),
data_flow=[("result", 0, 1)])
knl = knl.with_kernel(
lp.prioritize_loops(knl["xderiv_and_yderiv"], "e,i,j,k"))
assert len(knl["xderiv_and_yderiv"].temporary_variables) == 2
ctx = ctx_factory()
lp.auto_test_vs_ref(xyderiv,
ctx,
knl,
parameters=dict(nelements=20, ndofs=4))
def test_fuse_kernels(ctx_factory):
fortran_template = """
subroutine {name}(nelements, ndofs, result, d, q)
implicit none
integer e, i, j, k
integer nelements, ndofs
real*8 result(nelements, ndofs, ndofs)
real*8 q(nelements, ndofs, ndofs)
real*8 d(ndofs, ndofs)
real*8 prev
do e = 1,nelements
do i = 1,ndofs
do j = 1,ndofs
do k = 1,ndofs
{inner}
end do
end do
end do
end do
end subroutine
"""
xd_line = """
prev = result(e,i,j)
result(e,i,j) = prev + d(i,k)*q(e,i,k)
"""
yd_line = """
prev = result(e,i,j)
result(e,i,j) = prev + d(i,k)*q(e,k,j)
"""
xderiv, = lp.parse_fortran(
fortran_template.format(inner=xd_line, name="xderiv"))
yderiv, = lp.parse_fortran(
fortran_template.format(inner=yd_line, name="yderiv"))
xyderiv, = lp.parse_fortran(
fortran_template.format(
inner=(xd_line + "\n" + yd_line), name="xyderiv"))
knl = lp.fuse_kernels((xderiv, yderiv))
knl = lp.set_loop_priority(knl, "e,i,j,k")
assert len(knl.temporary_variables) == 2
# This is needed for correctness, otherwise ordering could foul things up.
knl = lp.assignment_to_subst(knl, "prev")
knl = lp.assignment_to_subst(knl, "prev_0")
ctx = ctx_factory()
lp.auto_test_vs_ref(xyderiv, ctx, knl, parameters=dict(nelements=20, ndofs=4))
def test_assignment_to_subst_two_defs(ctx_factory):
fortran_src = """
subroutine fill(out, out2, inp, n)
implicit none
real*8 a, out(n), out2(n), inp(n)
integer n, i
do i = 1, n
a = inp(i)
out(i) = 5*a
a = 3*inp(n)
out2(i) = 6*a
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
ref_knl = knl
knl = lp.assignment_to_subst(knl, "a")
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=5))
def test_tagged(ctx_factory):
fortran_src = """
subroutine rot_norm(out, alpha, out2, inp, inp2, n)
implicit none
real*8 a, b, r, out(n), out2(n), inp(n), inp2(n)
real*8 alpha
integer n, i
do i = 1, n
!$loopy begin tagged: input
a = cos(alpha)*inp(i) + sin(alpha)*inp2(i)
b = -sin(alpha)*inp(i) + cos(alpha)*inp2(i)
!$loopy end tagged: input
r = sqrt(a**2 + b**2)
a = a/r
b = b/r
out(i) = a
out2(i) = b
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
assert sum(1 for insn in lp.find_instructions(knl, "tag:input")) == 2
def test_assignment_to_subst_two_defs(ctx_factory):
fortran_src = """
subroutine fill(out, out2, inp, n)
implicit none
real*8 a, out(n), out2(n), inp(n)
integer n, i
do i = 1, n
a = inp(i)
out(i) = 5*a
a = 3*inp(n)
out2(i) = 6*a
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
ref_knl = knl
knl = lp.assignment_to_subst(knl, "a")
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=5))
def test_assignment_to_subst_indices(ctx_factory):
fortran_src = """
subroutine fill(out, out2, inp, n)
implicit none
real*8 a(n), out(n), out2(n), inp(n)
integer n, i
do i = 1, n
a(i) = 6*inp(i)
enddo
do i = 1, n
out(i) = 5*a(i)
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
knl = lp.fix_parameters(knl, n=5)
ref_knl = knl
assert "a" in knl.temporary_variables
knl = lp.assignment_to_subst(knl, "a")
assert "a" not in knl.temporary_variables
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_knl, ctx, knl)
def test_if(ctx_factory):
fortran_src = """
subroutine fill(out, out2, inp, n)
implicit none
real*8 a, b, out(n), out2(n), inp(n)
integer n, i, j
do i = 1, n
a = inp(i)
if (a.ge.3) then
b = 2*a
do j = 1,3
b = 3 * b
end do
out(i) = 5*b
else
out(i) = 4*a
endif
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
ref_knl = knl
knl = lp.assignment_to_subst(knl, "a")
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=5))
def test_tagged(ctx_factory):
fortran_src = """
subroutine rot_norm(out, alpha, out2, inp, inp2, n)
implicit none
real*8 a, b, r, out(n), out2(n), inp(n), inp2(n)
real*8 alpha
integer n, i
do i = 1, n
!$loopy begin tagged: input
a = cos(alpha)*inp(i) + sin(alpha)*inp2(i)
b = -sin(alpha)*inp(i) + cos(alpha)*inp2(i)
!$loopy end tagged: input
r = sqrt(a**2 + b**2)
a = a/r
b = b/r
out(i) = a
out2(i) = b
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
assert sum(1 for insn in lp.find_instructions(knl, "tag:input")) == 2
def test_assignment_to_subst_indices(ctx_factory):
fortran_src = """
subroutine fill(out, out2, inp, n)
implicit none
real*8 a(n), out(n), out2(n), inp(n)
integer n, i
do i = 1, n
a(i) = 6*inp(i)
enddo
do i = 1, n
out(i) = 5*a(i)
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
knl = lp.fix_parameters(knl, n=5)
ref_knl = knl
assert "a" in knl.temporary_variables
knl = lp.assignment_to_subst(knl, "a")
assert "a" not in knl.temporary_variables
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_knl, ctx, knl)
def test_if(ctx_factory):
fortran_src = """
subroutine fill(out, out2, inp, n)
implicit none
real*8 a, b, out(n), out2(n), inp(n)
integer n, i, j
do i = 1, n
a = inp(i)
if (a.ge.3) then
b = 2*a
do j = 1,3
b = 3 * b
end do
out(i) = 5*b
else
out(i) = 4*a
endif
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
ref_knl = knl
knl = lp.assignment_to_subst(knl, "a")
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=5))
def test_fuse_kernels(ctx_factory):
fortran_template = """
subroutine {name}(nelements, ndofs, result, d, q)
implicit none
integer e, i, j, k
integer nelements, ndofs
real*8 result(nelements, ndofs, ndofs)
real*8 q(nelements, ndofs, ndofs)
real*8 d(ndofs, ndofs)
real*8 prev
do e = 1,nelements
do i = 1,ndofs
do j = 1,ndofs
do k = 1,ndofs
{inner}
end do
end do
end do
end do
end subroutine
"""
xd_line = """
prev = result(e,i,j)
result(e,i,j) = prev + d(i,k)*q(e,i,k)
"""
yd_line = """
prev = result(e,i,j)
result(e,i,j) = prev + d(i,k)*q(e,k,j)
"""
xderiv, = lp.parse_fortran(
fortran_template.format(inner=xd_line, name="xderiv"))
yderiv, = lp.parse_fortran(
fortran_template.format(inner=yd_line, name="yderiv"))
xyderiv, = lp.parse_fortran(
fortran_template.format(
inner=(xd_line + "\n" + yd_line), name="xyderiv"))
knl = lp.fuse_kernels((xderiv, yderiv), data_flow=[("result", 0, 1)])
knl = lp.prioritize_loops(knl, "e,i,j,k")
assert len(knl.temporary_variables) == 2
ctx = ctx_factory()
lp.auto_test_vs_ref(xyderiv, ctx, knl, parameters=dict(nelements=20, ndofs=4))
def test_matmul(ctx_factory, buffer_inames):
ctx = ctx_factory()
if (buffer_inames and
ctx.devices[0].platform.name == "Portable Computing Language"):
pytest.skip("crashes on pocl")
logging.basicConfig(level=logging.INFO)
fortran_src = """
subroutine dgemm(m,n,ell,a,b,c)
implicit none
real*8 a(m,ell),b(ell,n),c(m,n)
integer m,n,k,i,j,ell
do j = 1,n
do i = 1,m
do k = 1,ell
c(i,j) = c(i,j) + b(k,j)*a(i,k)
end do
end do
end do
end subroutine
"""
knl, = lp.parse_fortran(fortran_src)
assert len(knl.domains) == 1
ref_knl = knl
knl = lp.split_iname(knl, "i", 16,
outer_tag="g.0", inner_tag="l.1")
knl = lp.split_iname(knl, "j", 8,
outer_tag="g.1", inner_tag="l.0")
knl = lp.split_iname(knl, "k", 32)
knl = lp.assume(knl, "n mod 32 = 0")
knl = lp.assume(knl, "m mod 32 = 0")
knl = lp.assume(knl, "ell mod 16 = 0")
knl = lp.extract_subst(knl, "a_acc", "a[i1,i2]", parameters="i1, i2")
knl = lp.extract_subst(knl, "b_acc", "b[i1,i2]", parameters="i1, i2")
knl = lp.precompute(knl, "a_acc", "k_inner,i_inner",
precompute_outer_inames='i_outer, j_outer, k_outer',
default_tag="l.auto")
knl = lp.precompute(knl, "b_acc", "j_inner,k_inner",
precompute_outer_inames='i_outer, j_outer, k_outer',
default_tag="l.auto")
knl = lp.buffer_array(knl, "c", buffer_inames=buffer_inames,
init_expression="0", store_expression="base+buffer")
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=128, m=128, ell=128))
def test_precompute_some_exist(ctx_factory):
fortran_src = """
subroutine dgemm(m,n,ell,a,b,c)
implicit none
real*8 a(m,ell),b(ell,n),c(m,n)
integer m,n,k,i,j,ell
do j = 1,n
do i = 1,m
do k = 1,ell
c(i,j) = c(i,j) + b(k,j)*a(i,k)
end do
end do
end do
end subroutine
"""
knl = lp.parse_fortran(fortran_src)
assert len(knl["dgemm"].domains) == 1
knl = lp.split_iname(knl, "i", 8, outer_tag="g.0", inner_tag="l.1")
knl = lp.split_iname(knl, "j", 8, outer_tag="g.1", inner_tag="l.0")
knl = lp.split_iname(knl, "k", 8)
knl = lp.assume(knl, "n mod 8 = 0")
knl = lp.assume(knl, "m mod 8 = 0")
knl = lp.assume(knl, "ell mod 8 = 0")
knl = lp.extract_subst(knl, "a_acc", "a[i1,i2]", parameters="i1, i2")
knl = lp.extract_subst(knl, "b_acc", "b[i1,i2]", parameters="i1, i2")
knl = lp.precompute(knl,
"a_acc",
"k_inner,i_inner",
precompute_inames="ktemp,itemp",
precompute_outer_inames="i_outer, j_outer, k_outer",
default_tag="l.auto")
knl = lp.precompute(knl,
"b_acc",
"j_inner,k_inner",
precompute_inames="itemp,k2temp",
precompute_outer_inames="i_outer, j_outer, k_outer",
default_tag="l.auto")
ref_knl = knl
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_knl,
ctx,
knl,
parameters=dict(n=128, m=128, ell=128))
def test_matmul(ctx_factory, buffer_inames):
logging.basicConfig(level=logging.INFO)
fortran_src = """
subroutine dgemm(m,n,l,a,b,c)
implicit none
real*8 a(m,l),b(l,n),c(m,n)
integer m,n,k,i,j,l
do j = 1,n
do i = 1,m
do k = 1,l
c(i,j) = c(i,j) + b(k,j)*a(i,k)
end do
end do
end do
end subroutine
"""
knl, = lp.parse_fortran(fortran_src)
assert len(knl.domains) == 1
ref_knl = knl
knl = lp.split_iname(knl, "i", 16, outer_tag="g.0", inner_tag="l.1")
knl = lp.split_iname(knl, "j", 8, outer_tag="g.1", inner_tag="l.0")
knl = lp.split_iname(knl, "k", 32)
knl = lp.assume(knl, "n mod 32 = 0")
knl = lp.assume(knl, "m mod 32 = 0")
knl = lp.assume(knl, "l mod 16 = 0")
knl = lp.extract_subst(knl, "a_acc", "a[i1,i2]", parameters="i1, i2")
knl = lp.extract_subst(knl, "b_acc", "b[i1,i2]", parameters="i1, i2")
knl = lp.precompute(knl, "a_acc", "k_inner,i_inner")
knl = lp.precompute(knl, "b_acc", "j_inner,k_inner")
knl = lp.buffer_array(knl,
"c",
buffer_inames=buffer_inames,
init_expression="0",
store_expression="base+buffer")
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_knl,
ctx,
knl,
parameters=dict(n=128, m=128, l=128))
def test_batched_sparse():
fortran_src = """
subroutine sparse(rowstarts, colindices, values, m, n, nvecs, nvals, x, y)
implicit none
integer rowstarts(m+1), colindices(nvals)
real*8 values(nvals)
real*8 x(n, nvecs), y(n, nvecs), rowsum(nvecs)
integer m, n, rowstart, rowend, length, nvals, nvecs
integer i, j, k
do i = 1, m
rowstart = rowstarts(i)
rowend = rowstarts(i+1)
length = rowend - rowstart
do k = 1, nvecs
rowsum(k) = 0
enddo
do k = 1, nvecs
do j = 1, length
rowsum(k) = rowsum(k) + &
x(colindices(rowstart+j-1),k)*values(rowstart+j-1)
end do
end do
do k = 1, nvecs
y(i,k) = rowsum(k)
end do
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
knl = lp.split_iname(knl, "i", 128)
knl = lp.tag_inames(knl, {"i_outer": "g.0"})
knl = lp.tag_inames(knl, {"i_inner": "l.0"})
knl = lp.add_prefetch(knl, "values",
default_tag="l.auto")
knl = lp.add_prefetch(knl, "colindices",
default_tag="l.auto")
knl = lp.fix_parameters(knl, nvecs=4)
def test_batched_sparse():
fortran_src = """
subroutine sparse(rowstarts, colindices, values, m, n, nvecs, nvals, x, y)
implicit none
integer rowstarts(m+1), colindices(nvals)
real*8 values(nvals)
real*8 x(n, nvecs), y(n, nvecs), rowsum(nvecs)
integer m, n, rowstart, rowend, length, nvals, nvecs
integer i, j, k
do i = 1, m
rowstart = rowstarts(i)
rowend = rowstarts(i+1)
length = rowend - rowstart
do k = 1, nvecs
rowsum(k) = 0
enddo
do k = 1, nvecs
do j = 1, length
rowsum(k) = rowsum(k) + &
x(colindices(rowstart+j-1),k)*values(rowstart+j-1)
end do
end do
do k = 1, nvecs
y(i,k) = rowsum(k)
end do
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
knl = lp.split_iname(knl, "i", 128)
knl = lp.tag_inames(knl, {"i_outer": "g.0"})
knl = lp.tag_inames(knl, {"i_inner": "l.0"})
knl = lp.add_prefetch(knl, "values",
default_tag="l.auto")
knl = lp.add_prefetch(knl, "colindices",
default_tag="l.auto")
knl = lp.fix_parameters(knl, nvecs=4)
def test_matmul(ctx_factory, buffer_inames):
logging.basicConfig(level=logging.INFO)
fortran_src = """
subroutine dgemm(m,n,l,a,b,c)
implicit none
real*8 a(m,l),b(l,n),c(m,n)
integer m,n,k,i,j,l
do j = 1,n
do i = 1,m
do k = 1,l
c(i,j) = c(i,j) + b(k,j)*a(i,k)
end do
end do
end do
end subroutine
"""
knl, = lp.parse_fortran(fortran_src)
assert len(knl.domains) == 1
ref_knl = knl
knl = lp.split_iname(knl, "i", 16,
outer_tag="g.0", inner_tag="l.1")
knl = lp.split_iname(knl, "j", 8,
outer_tag="g.1", inner_tag="l.0")
knl = lp.split_iname(knl, "k", 32)
knl = lp.assume(knl, "n mod 32 = 0")
knl = lp.assume(knl, "m mod 32 = 0")
knl = lp.assume(knl, "l mod 16 = 0")
knl = lp.extract_subst(knl, "a_acc", "a[i1,i2]", parameters="i1, i2")
knl = lp.extract_subst(knl, "b_acc", "b[i1,i2]", parameters="i1, i2")
knl = lp.precompute(knl, "a_acc", "k_inner,i_inner")
knl = lp.precompute(knl, "b_acc", "j_inner,k_inner")
knl = lp.buffer_array(knl, "c", buffer_inames=buffer_inames,
init_expression="0", store_expression="base+buffer")
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=128, m=128, l=128))
def test_precompute_some_exist(ctx_factory):
fortran_src = """
subroutine dgemm(m,n,ell,a,b,c)
implicit none
real*8 a(m,ell),b(ell,n),c(m,n)
integer m,n,k,i,j,ell
do j = 1,n
do i = 1,m
do k = 1,ell
c(i,j) = c(i,j) + b(k,j)*a(i,k)
end do
end do
end do
end subroutine
"""
knl, = lp.parse_fortran(fortran_src)
assert len(knl.domains) == 1
knl = lp.split_iname(knl, "i", 8,
outer_tag="g.0", inner_tag="l.1")
knl = lp.split_iname(knl, "j", 8,
outer_tag="g.1", inner_tag="l.0")
knl = lp.split_iname(knl, "k", 8)
knl = lp.assume(knl, "n mod 8 = 0")
knl = lp.assume(knl, "m mod 8 = 0")
knl = lp.assume(knl, "ell mod 8 = 0")
knl = lp.extract_subst(knl, "a_acc", "a[i1,i2]", parameters="i1, i2")
knl = lp.extract_subst(knl, "b_acc", "b[i1,i2]", parameters="i1, i2")
knl = lp.precompute(knl, "a_acc", "k_inner,i_inner",
precompute_inames="ktemp,itemp",
default_tag="l.auto")
knl = lp.precompute(knl, "b_acc", "j_inner,k_inner",
precompute_inames="itemp,k2temp",
default_tag="l.auto")
ref_knl = knl
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=128, m=128, ell=128))
def test_fill_const(ctx_factory):
fortran_src = """
subroutine fill(out, a, n)
implicit none
real*8 a, out(n)
integer n, i
do i = 1, n
out(i) = 3.45
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
ctx = ctx_factory()
lp.auto_test_vs_ref(knl, ctx, knl, parameters=dict(n=5, a=5))
def test_fill_const(ctx_factory):
fortran_src = """
subroutine fill(out, a, n)
implicit none
real*8 a, out(n)
integer n, i
do i = 1, n
out(i) = 3.45
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
ctx = ctx_factory()
lp.auto_test_vs_ref(knl, ctx, knl, parameters=dict(n=5, a=5))
def test_division_in_shapes(ctx_factory):
fortran_src = """
subroutine halve(m, a)
implicit none
integer m, i, j
real*8 a(m/2,m/2)
do i = 1,m/2
do j = 1,m/2
a(i, j) = 2*a(i, j)
end do
end do
end subroutine
"""
t_unit = lp.parse_fortran(fortran_src)
ref_t_unit = t_unit
print(t_unit)
ctx = ctx_factory()
lp.auto_test_vs_ref(ref_t_unit, ctx, t_unit, parameters=dict(m=128))
def test_assign_double_precision_scalar_as_rational(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
fortran_src = """
subroutine assign_scalar(a)
real*8 a(1)
a(1) = 11
a(1) = a(1) / 10
end
"""
t_unit = lp.parse_fortran(fortran_src)
a_dev = cl.array.empty(queue, 1, dtype=np.float64, order="F")
t_unit(queue, a=a_dev)
abs_err = abs(a_dev.get()[0] - 1.1)
assert abs_err < 1e-15
def test_domain_fusion_imperfectly_nested():
fortran_src = """
subroutine imperfect(n, m, a, b)
implicit none
integer i, j, n, m
real a(n), b(n,n)
do i=1, n
a(i) = i
do j=1, m
b(i,j) = i*j
end do
end do
end subroutine
"""
t_unit = lp.parse_fortran(fortran_src)
# If n > 0 and m == 0, a single domain would be empty,
# leading (incorrectly) to no assignments to 'a'.
assert len(t_unit["imperfect"].domains) > 1
def test_assign_double_precision_scalar(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
fortran_src = """
subroutine assign_scalar(a)
real*8 a(1)
a(1) = 1.1d0
end
"""
t_unit = lp.parse_fortran(fortran_src)
print(lp.generate_code_v2(t_unit).device_code())
assert "1.1;" in lp.generate_code_v2(t_unit).device_code()
a_dev = cl.array.empty(queue, 1, dtype=np.float64, order="F")
t_unit(queue, a=a_dev)
abs_err = abs(a_dev.get()[0] - 1.1)
assert abs_err < 1e-15
def test_asterisk_in_shape(ctx_factory):
fortran_src = """
subroutine fill(out, out2, inp, n)
implicit none
real*8 a, out(n), out2(n), inp(*)
integer n, i
do i = 1, n
a = inp(n)
out(i) = 5*a
out2(i) = 6*a
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl(queue, inp=np.array([1, 2, 3.]), n=3)
def test_asterisk_in_shape(ctx_factory):
fortran_src = """
subroutine fill(out, out2, inp, n)
implicit none
real*8 a, out(n), out2(n), inp(*)
integer n, i
do i = 1, n
a = inp(n)
out(i) = 5*a
out2(i) = 6*a
end do
end
"""
knl, = lp.parse_fortran(fortran_src)
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl(queue, inp=np.array([1, 2, 3.]), n=3)
def test_fortran_subroutines():
fortran_src = """
subroutine twice(n, a)
implicit none
real*8 a(n)
integer i,n
do i=1,n
a(i) = a(i) * 2
end do
end subroutine
subroutine twice_cross(n, a, i)
implicit none
integer i, n
real*8 a(n,n)
call twice(n, a(1:n, i))
call twice(n, a(i, 1:n))
end subroutine
"""
t_unit = lp.parse_fortran(fortran_src).with_entrypoints("twice_cross")
print(lp.generate_code_v2(t_unit).device_code())
def fortran_kernel(self, line, cell):
result = lp.parse_fortran(cell)
for knl in result:
self.shell.user_ns[knl.name] = knl
def test_gnuma_horiz_kernel(ctx_factory, ilp_multiple, Nq, opt_level):
ctx = ctx_factory()
filename = "strongVolumeKernels.f90"
with open(filename, "r") as sourcef:
source = sourcef.read()
source = source.replace("datafloat", "real*4")
hsv_r, hsv_s = [
knl
for knl in lp.parse_fortran(source, filename, auto_dependencies=False)
if "KernelR" in knl.name or "KernelS" in knl.name
]
hsv_r = lp.tag_instructions(hsv_r, "rknl")
hsv_s = lp.tag_instructions(hsv_s, "sknl")
hsv = lp.fuse_kernels([hsv_r, hsv_s], ["_r", "_s"])
#hsv = hsv_s
from gnuma_loopy_transforms import (fix_euler_parameters,
set_q_storage_format,
set_D_storage_format)
hsv = lp.fix_parameters(hsv, Nq=Nq)
hsv = lp.set_loop_priority(hsv, "e,k,j,i")
hsv = lp.tag_inames(hsv, dict(e="g.0", j="l.1", i="l.0"))
hsv = lp.assume(hsv, "elements >= 1")
hsv = fix_euler_parameters(hsv, p_p0=1, p_Gamma=1.4, p_R=1)
for name in ["Q", "rhsQ"]:
hsv = set_q_storage_format(hsv, name)
hsv = set_D_storage_format(hsv)
#hsv = lp.add_prefetch(hsv, "volumeGeometricFactors")
ref_hsv = hsv
if opt_level == 0:
tap_hsv = hsv
hsv = lp.add_prefetch(hsv, "D[:,:]")
if opt_level == 1:
tap_hsv = hsv
# turn the first reads into subst rules
local_prep_var_names = set()
for insn in lp.find_instructions(hsv, "tag:local_prep"):
assignee, = insn.assignee_var_names()
local_prep_var_names.add(assignee)
hsv = lp.assignment_to_subst(hsv, assignee)
# precompute fluxes
hsv = lp.assignment_to_subst(hsv, "JinvD_r")
hsv = lp.assignment_to_subst(hsv, "JinvD_s")
r_fluxes = lp.find_instructions(hsv, "tag:compute_fluxes and tag:rknl")
s_fluxes = lp.find_instructions(hsv, "tag:compute_fluxes and tag:sknl")
if ilp_multiple > 1:
hsv = lp.split_iname(hsv, "k", 2, inner_tag="ilp")
ilp_inames = ("k_inner", )
flux_ilp_inames = ("kk", )
else:
ilp_inames = ()
flux_ilp_inames = ()
rtmps = []
stmps = []
flux_store_idx = 0
for rflux_insn, sflux_insn in zip(r_fluxes, s_fluxes):
for knl_tag, insn, flux_inames, tmps, flux_precomp_inames in [
("rknl", rflux_insn, (
"j",
"n",
), rtmps, (
"jj",
"ii",
)),
("sknl", sflux_insn, (
"i",
"n",
), stmps, (
"ii",
"jj",
)),
]:
flux_var, = insn.assignee_var_names()
print(insn)
reader, = lp.find_instructions(
hsv,
"tag:{knl_tag} and reads:{flux_var}".format(knl_tag=knl_tag,
flux_var=flux_var))
hsv = lp.assignment_to_subst(hsv, flux_var)
flux_store_name = "flux_store_%d" % flux_store_idx
flux_store_idx += 1
tmps.append(flux_store_name)
hsv = lp.precompute(hsv,
flux_var + "_subst",
flux_inames + ilp_inames,
temporary_name=flux_store_name,
precompute_inames=flux_precomp_inames +
flux_ilp_inames,
default_tag=None)
if flux_var.endswith("_s"):
hsv = lp.tag_array_axes(hsv, flux_store_name, "N0,N1,N2?")
else:
hsv = lp.tag_array_axes(hsv, flux_store_name, "N1,N0,N2?")
n_iname = "n_" + flux_var.replace("_r", "").replace("_s", "")
if n_iname.endswith("_0"):
n_iname = n_iname[:-2]
hsv = lp.rename_iname(hsv,
"n",
n_iname,
within="id:" + reader.id,
existing_ok=True)
hsv = lp.tag_inames(hsv, dict(ii="l.0", jj="l.1"))
for iname in flux_ilp_inames:
hsv = lp.tag_inames(hsv, {iname: "ilp"})
hsv = lp.alias_temporaries(hsv, rtmps)
hsv = lp.alias_temporaries(hsv, stmps)
if opt_level == 2:
tap_hsv = hsv
for prep_var_name in local_prep_var_names:
if prep_var_name.startswith("Jinv") or "_s" in prep_var_name:
continue
hsv = lp.precompute(
hsv, lp.find_one_rule_matching(hsv, prep_var_name + "_*subst*"))
if opt_level == 3:
tap_hsv = hsv
hsv = lp.add_prefetch(hsv, "Q[ii,jj,k,:,:,e]", sweep_inames=ilp_inames)
if opt_level == 4:
tap_hsv = hsv
tap_hsv = lp.tag_inames(
tap_hsv, dict(Q_dim_field_inner="unr", Q_dim_field_outer="unr"))
hsv = lp.buffer_array(hsv,
"rhsQ",
ilp_inames,
fetch_bounding_box=True,
default_tag="for",
init_expression="0",
store_expression="base + buffer")
if opt_level == 5:
tap_hsv = hsv
tap_hsv = lp.tag_inames(
tap_hsv,
dict(rhsQ_init_field_inner="unr",
rhsQ_store_field_inner="unr",
rhsQ_init_field_outer="unr",
rhsQ_store_field_outer="unr",
Q_dim_field_inner="unr",
Q_dim_field_outer="unr"))
# buffer axes need to be vectorized in order for this to work
hsv = lp.tag_array_axes(hsv, "rhsQ_buf", "c?,vec,c")
hsv = lp.tag_array_axes(hsv, "Q_fetch", "c?,vec,c")
hsv = lp.tag_array_axes(hsv, "D_fetch", "f,f")
hsv = lp.tag_inames(hsv, {
"Q_dim_k": "unr",
"rhsQ_init_k": "unr",
"rhsQ_store_k": "unr"
},
ignore_nonexistent=True)
if opt_level == 6:
tap_hsv = hsv
tap_hsv = lp.tag_inames(
tap_hsv,
dict(rhsQ_init_field_inner="unr",
rhsQ_store_field_inner="unr",
rhsQ_init_field_outer="unr",
rhsQ_store_field_outer="unr",
Q_dim_field_inner="unr",
Q_dim_field_outer="unr"))
hsv = lp.tag_inames(
hsv,
dict(rhsQ_init_field_inner="vec",
rhsQ_store_field_inner="vec",
rhsQ_init_field_outer="unr",
rhsQ_store_field_outer="unr",
Q_dim_field_inner="vec",
Q_dim_field_outer="unr"))
if opt_level == 7:
tap_hsv = hsv
hsv = lp.collect_common_factors_on_increment(
hsv, "rhsQ_buf", vary_by_axes=(0, ) if ilp_multiple > 1 else ())
if opt_level >= 8:
tap_hsv = hsv
hsv = tap_hsv
if 1:
print("OPS")
op_poly = lp.get_op_poly(hsv)
print(lp.stringify_stats_mapping(op_poly))
print("MEM")
gmem_poly = lp.sum_mem_access_to_bytes(lp.get_gmem_access_poly(hsv))
print(lp.stringify_stats_mapping(gmem_poly))
hsv = lp.set_options(hsv,
cl_build_options=[
"-cl-denorms-are-zero",
"-cl-fast-relaxed-math",
"-cl-finite-math-only",
"-cl-mad-enable",
"-cl-no-signed-zeros",
])
hsv = hsv.copy(name="horizontalStrongVolumeKernel")
results = lp.auto_test_vs_ref(ref_hsv,
ctx,
hsv,
parameters=dict(elements=300),
quiet=True)
elapsed = results["elapsed_wall"]
print("elapsed", elapsed)
def fortran_kernel(self, line, cell):
result = lp.parse_fortran(cell)
for knl in result:
self.shell.user_ns[knl.name] = knl
def fortran_kernel(self, line, cell):
result = lp.parse_fortran(cell)
self.shell.user_ns["prog"] = result
def test_gnuma_horiz_kernel(ctx_factory, ilp_multiple, Nq, opt_level):
ctx = ctx_factory()
filename = "strongVolumeKernels.f90"
with open(filename, "r") as sourcef:
source = sourcef.read()
source = source.replace("datafloat", "real*4")
hsv_r, hsv_s = [
knl for knl in lp.parse_fortran(source, filename, auto_dependencies=False)
if "KernelR" in knl.name or "KernelS" in knl.name
]
hsv_r = lp.tag_instructions(hsv_r, "rknl")
hsv_s = lp.tag_instructions(hsv_s, "sknl")
hsv = lp.fuse_kernels([hsv_r, hsv_s], ["_r", "_s"])
#hsv = hsv_s
from gnuma_loopy_transforms import (
fix_euler_parameters,
set_q_storage_format, set_D_storage_format)
hsv = lp.fix_parameters(hsv, Nq=Nq)
hsv = lp.set_loop_priority(hsv, "e,k,j,i")
hsv = lp.tag_inames(hsv, dict(e="g.0", j="l.1", i="l.0"))
hsv = lp.assume(hsv, "elements >= 1")
hsv = fix_euler_parameters(hsv, p_p0=1, p_Gamma=1.4, p_R=1)
for name in ["Q", "rhsQ"]:
hsv = set_q_storage_format(hsv, name)
hsv = set_D_storage_format(hsv)
#hsv = lp.add_prefetch(hsv, "volumeGeometricFactors")
ref_hsv = hsv
if opt_level == 0:
tap_hsv = hsv
hsv = lp.add_prefetch(hsv, "D[:,:]")
if opt_level == 1:
tap_hsv = hsv
# turn the first reads into subst rules
local_prep_var_names = set()
for insn in lp.find_instructions(hsv, "tag:local_prep"):
assignee, = insn.assignee_var_names()
local_prep_var_names.add(assignee)
hsv = lp.assignment_to_subst(hsv, assignee)
# precompute fluxes
hsv = lp.assignment_to_subst(hsv, "JinvD_r")
hsv = lp.assignment_to_subst(hsv, "JinvD_s")
r_fluxes = lp.find_instructions(hsv, "tag:compute_fluxes and tag:rknl")
s_fluxes = lp.find_instructions(hsv, "tag:compute_fluxes and tag:sknl")
if ilp_multiple > 1:
hsv = lp.split_iname(hsv, "k", 2, inner_tag="ilp")
ilp_inames = ("k_inner",)
flux_ilp_inames = ("kk",)
else:
ilp_inames = ()
flux_ilp_inames = ()
rtmps = []
stmps = []
flux_store_idx = 0
for rflux_insn, sflux_insn in zip(r_fluxes, s_fluxes):
for knl_tag, insn, flux_inames, tmps, flux_precomp_inames in [
("rknl", rflux_insn, ("j", "n",), rtmps, ("jj", "ii",)),
("sknl", sflux_insn, ("i", "n",), stmps, ("ii", "jj",)),
]:
flux_var, = insn.assignee_var_names()
print(insn)
reader, = lp.find_instructions(hsv,
"tag:{knl_tag} and reads:{flux_var}"
.format(knl_tag=knl_tag, flux_var=flux_var))
hsv = lp.assignment_to_subst(hsv, flux_var)
flux_store_name = "flux_store_%d" % flux_store_idx
flux_store_idx += 1
tmps.append(flux_store_name)
hsv = lp.precompute(hsv, flux_var+"_subst", flux_inames + ilp_inames,
temporary_name=flux_store_name,
precompute_inames=flux_precomp_inames + flux_ilp_inames,
default_tag=None)
if flux_var.endswith("_s"):
hsv = lp.tag_data_axes(hsv, flux_store_name, "N0,N1,N2?")
else:
hsv = lp.tag_data_axes(hsv, flux_store_name, "N1,N0,N2?")
n_iname = "n_"+flux_var.replace("_r", "").replace("_s", "")
if n_iname.endswith("_0"):
n_iname = n_iname[:-2]
hsv = lp.rename_iname(hsv, "n", n_iname, within="id:"+reader.id,
existing_ok=True)
hsv = lp.tag_inames(hsv, dict(ii="l.0", jj="l.1"))
for iname in flux_ilp_inames:
hsv = lp.tag_inames(hsv, {iname: "ilp"})
hsv = lp.alias_temporaries(hsv, rtmps)
hsv = lp.alias_temporaries(hsv, stmps)
if opt_level == 2:
tap_hsv = hsv
for prep_var_name in local_prep_var_names:
if prep_var_name.startswith("Jinv") or "_s" in prep_var_name:
continue
hsv = lp.precompute(hsv,
lp.find_one_rule_matching(hsv, prep_var_name+"_*subst*"))
if opt_level == 3:
tap_hsv = hsv
hsv = lp.add_prefetch(hsv, "Q[ii,jj,k,:,:,e]", sweep_inames=ilp_inames)
if opt_level == 4:
tap_hsv = hsv
tap_hsv = lp.tag_inames(tap_hsv, dict(
Q_dim_field_inner="unr",
Q_dim_field_outer="unr"))
hsv = lp.buffer_array(hsv, "rhsQ", ilp_inames,
fetch_bounding_box=True, default_tag="for",
init_expression="0", store_expression="base + buffer")
if opt_level == 5:
tap_hsv = hsv
tap_hsv = lp.tag_inames(tap_hsv, dict(
rhsQ_init_field_inner="unr", rhsQ_store_field_inner="unr",
rhsQ_init_field_outer="unr", rhsQ_store_field_outer="unr",
Q_dim_field_inner="unr",
Q_dim_field_outer="unr"))
# buffer axes need to be vectorized in order for this to work
hsv = lp.tag_data_axes(hsv, "rhsQ_buf", "c?,vec,c")
hsv = lp.tag_data_axes(hsv, "Q_fetch", "c?,vec,c")
hsv = lp.tag_data_axes(hsv, "D_fetch", "f,f")
hsv = lp.tag_inames(hsv,
{"Q_dim_k": "unr", "rhsQ_init_k": "unr", "rhsQ_store_k": "unr"},
ignore_nonexistent=True)
if opt_level == 6:
tap_hsv = hsv
tap_hsv = lp.tag_inames(tap_hsv, dict(
rhsQ_init_field_inner="unr", rhsQ_store_field_inner="unr",
rhsQ_init_field_outer="unr", rhsQ_store_field_outer="unr",
Q_dim_field_inner="unr",
Q_dim_field_outer="unr"))
hsv = lp.tag_inames(hsv, dict(
rhsQ_init_field_inner="vec", rhsQ_store_field_inner="vec",
rhsQ_init_field_outer="unr", rhsQ_store_field_outer="unr",
Q_dim_field_inner="vec",
Q_dim_field_outer="unr"))
if opt_level == 7:
tap_hsv = hsv
hsv = lp.collect_common_factors_on_increment(hsv, "rhsQ_buf",
vary_by_axes=(0,) if ilp_multiple > 1 else ())
if opt_level >= 8:
tap_hsv = hsv
hsv = tap_hsv
if 1:
print("OPS")
op_poly = lp.get_op_poly(hsv)
print(lp.stringify_stats_mapping(op_poly))
print("MEM")
gmem_poly = lp.sum_mem_access_to_bytes(lp.get_gmem_access_poly(hsv))
print(lp.stringify_stats_mapping(gmem_poly))
hsv = lp.set_options(hsv, cl_build_options=[
"-cl-denorms-are-zero",
"-cl-fast-relaxed-math",
"-cl-finite-math-only",
"-cl-mad-enable",
"-cl-no-signed-zeros",
])
hsv = hsv.copy(name="horizontalStrongVolumeKernel")
results = lp.auto_test_vs_ref(ref_hsv, ctx, hsv, parameters=dict(elements=300),
quiet=True)
elapsed = results["elapsed_wall"]
print("elapsed", elapsed)