def variant_2(knl):
knl = lp.split_iname(knl, "i", 16, outer_tag="g.1", inner_tag="l.1")
knl = lp.split_iname(knl, "j", 16, outer_tag="g.0", inner_tag="l.0")
knl = lp.add_prefetch(knl, "a", ["i_inner", "j_inner"],
fetch_bounding_box=True)
knl = lp.set_loop_priority(knl, ["a_dim_0_outer", "a_dim_1_outer"])
return knl
def variant_2(knl):
knl = lp.split_iname(knl, "i", 16, outer_tag="g.1", inner_tag="l.1")
knl = lp.split_iname(knl, "j", 16, outer_tag="g.0", inner_tag="l.0")
knl = lp.add_prefetch(knl, "a", ["i_inner", "j_inner"],
fetch_bounding_box=True)
knl = lp.set_loop_priority(knl, ["a_dim_0_outer", "a_dim_1_outer"])
return knl
def test_precompute_with_preexisting_inames(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[e,i,j,k]: 0<=e<E and 0<=i,j,k<n}",
"""
result[e,i] = sum(j, D1[i,j]*u[e,j])
result2[e,i] = sum(k, D2[i,k]*u[e,k])
""")
knl = lp.add_and_infer_dtypes(knl, {
"u": np.float32,
"D1": np.float32,
"D2": np.float32,
})
knl = lp.fix_parameters(knl, n=13)
ref_knl = knl
knl = lp.extract_subst(knl, "D1_subst", "D1[ii,jj]", parameters="ii,jj")
knl = lp.extract_subst(knl, "D2_subst", "D2[ii,jj]", parameters="ii,jj")
knl = lp.precompute(knl, "D1_subst", "i,j", default_tag="for",
precompute_inames="ii,jj")
knl = lp.precompute(knl, "D2_subst", "i,k", default_tag="for",
precompute_inames="ii,jj")
knl = lp.set_loop_priority(knl, "ii,jj,e,j,k")
lp.auto_test_vs_ref(
ref_knl, ctx, knl,
parameters=dict(E=200))
def test_slab_decomposition_does_not_double_execute(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"a[i] = 2*a[i]",
assumptions="n>=1")
ref_knl = knl
for outer_tag in ["for", "g.0"]:
knl = ref_knl
knl = lp.split_iname(knl, "i", 4, slabs=(0, 1), inner_tag="unr",
outer_tag=outer_tag)
knl = lp.set_loop_priority(knl, "i_outer")
a = cl.array.empty(queue, 20, np.float32)
a.fill(17)
a_ref = a.copy()
a_knl = a.copy()
knl = lp.set_options(knl, write_cl=True)
print("TEST-----------------------------------------")
knl(queue, a=a_knl)
print("REF-----------------------------------------")
ref_knl(queue, a=a_ref)
print("DONE-----------------------------------------")
print("REF", a_ref)
print("KNL", a_knl)
assert (a_ref == a_knl).get().all()
print("_________________________________")
def tune_prims_kernel(knl, shape=None, ng=None, prefetch_args=()):
"""Parameters for 3D"""
knl = spec_prims_kernel(knl, shape, ng)
# This assumes linear, see above
knl = _lp.split_iname(knl,
"k",
lsize[0],
outer_tag=otags[0],
inner_tag=itags[0])
knl = _lp.split_iname(knl,
"j",
lsize[1],
outer_tag=otags[1],
inner_tag=itags[1])
knl = _lp.split_iname(knl,
"i",
lsize[2],
outer_tag=otags[2],
inner_tag=itags[2])
knl = _lp.set_loop_priority(
knl, "p,i_outer,j_outer,k_outer,i_inner,j_inner,k_inner")
for arg in prefetch_args:
knl = _lp.add_prefetch(knl,
arg,
"i_inner,j_inner,k_inner",
default_tag="l.auto")
#knl = lp.set_options(knl, no_numpy=True)
return knl
def test_ilp_loop_bound(ctx_factory):
# The salient bit of this test is that a joint bound on (outer, inner)
# from a split occurs in a setting where the inner loop has been ilp'ed.
# In 'normal' parallel loops, the inner index is available for conditionals
# throughout. In ILP'd loops, not so much.
ctx = ctx_factory()
knl = lp.make_kernel(
"{ [i,j,k]: 0<=i,j,k<n }",
"""
out[i,k] = sum(j, a[i,j]*b[j,k])
""",
[
lp.GlobalArg("a,b", np.float32, shape=lp.auto),
"...",
],
assumptions="n>=1")
ref_knl = knl
knl = lp.set_loop_priority(knl, "j,i,k")
knl = lp.split_iname(knl, "k", 4, inner_tag="ilp")
lp.auto_test_vs_ref(ref_knl, ctx, knl,
parameters=dict(
n=200
))
def test_poisson_fem(ctx_factory):
# Stolen from Peter Coogan and Rob Kirby for FEM assembly
ctx = ctx_factory()
nbf = 5
nqp = 5
sdim = 3
knl = lp.make_kernel(
"{ [c,i,j,k,ell,ell2,ell3]: \
0 <= c < nels and \
0 <= i < nbf and \
0 <= j < nbf and \
0 <= k < nqp and \
0 <= ell,ell2 < sdim}",
"""
dpsi(bf,k0,dir) := \
simul_reduce(sum, ell2, DFinv[c,ell2,dir] * DPsi[bf,k0,ell2] )
Ael[c,i,j] = \
J[c] * w[k] * sum(ell, dpsi(i,k,ell) * dpsi(j,k,ell))
""",
assumptions="nels>=1 and nbf >= 1 and nels mod 4 = 0")
print(knl)
knl = lp.fix_parameters(knl, nbf=nbf, sdim=sdim, nqp=nqp)
ref_knl = knl
knl = lp.set_loop_priority(knl, ["c", "j", "i", "k"])
def variant_1(knl):
knl = lp.precompute(knl, "dpsi", "i,k,ell", default_tag='for')
knl = lp.set_loop_priority(knl, "c,i,j")
return knl
def variant_2(knl):
knl = lp.precompute(knl, "dpsi", "i,ell", default_tag='for')
knl = lp.set_loop_priority(knl, "c,i,j")
return knl
def add_types(knl):
return lp.add_and_infer_dtypes(knl, dict(
w=np.float32,
J=np.float32,
DPsi=np.float32,
DFinv=np.float32,
))
for variant in [
#variant_1,
variant_2
]:
knl = variant(knl)
lp.auto_test_vs_ref(
add_types(ref_knl), ctx, add_types(knl),
parameters=dict(n=5, nels=15, nbf=5, sdim=2, nqp=7))
def transform_surface_kernel(knl):
#print knl
knl = lp.tag_inames(knl, dict(k="g.0", n="l.0", m="l.0"))
knl = lp.split_iname(knl, "mp", 4, inner_tag="unr")
knl = lp.add_prefetch(knl, "LIFT")
for name in ["nx", "ny", "nz", "Fscale", "bc"]:
knl = lp.add_prefetch(knl, name)
knl = lp.set_loop_priority(knl, "mp_outer,mp_inner")
return knl
def transform_surface_kernel(knl):
#print knl
knl = lp.tag_inames(knl, dict(k="g.0", n="l.0", m="l.0"))
knl = lp.split_iname(knl, "mp", 4, inner_tag="unr")
knl = lp.add_prefetch(knl, "LIFT")
for name in ["nx", "ny", "nz", "Fscale", "bc"]:
knl = lp.add_prefetch(knl, name)
knl = lp.set_loop_priority(knl, "mp_outer,mp_inner")
return knl
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 variant_gpu(knl):
knl = lp.expand_subst(knl)
knl = lp.split_iname(knl, "i", 256,
outer_tag="g.0", inner_tag="l.0")
knl = lp.split_iname(knl, "j", 256)
knl = lp.add_prefetch(knl, "x[j,k]", ["j_inner", "k"],
["x_fetch_j", "x_fetch_k"], default_tag=None)
knl = lp.tag_inames(knl, dict(x_fetch_k="unr", x_fetch_j="l.0"))
knl = lp.add_prefetch(knl, "x[i,k]", ["k"], default_tag=None)
knl = lp.set_loop_priority(knl, ["j_outer", "j_inner"])
return knl
def variant_gpu(knl):
knl = lp.expand_subst(knl)
knl = lp.split_iname(knl, "i", 256, outer_tag="g.0", inner_tag="l.0")
knl = lp.split_iname(knl, "j", 256)
knl = lp.add_prefetch(knl,
"x[j,k]", ["j_inner", "k"],
["x_fetch_j", "x_fetch_k"],
default_tag=None)
knl = lp.tag_inames(knl, dict(x_fetch_k="unr", x_fetch_j="l.0"))
knl = lp.add_prefetch(knl, "x[i,k]", ["k"], default_tag=None)
knl = lp.set_loop_priority(knl, ["j_outer", "j_inner"])
return knl
def test_index_cse(ctx_factory):
knl = lp.make_kernel(["{[i,j,k,l,m]:0<=i,j,k,l,m<n}"], """
for i
for j
c[i,j,m] = sum((k,l), a[i,j,l]*b[i,j,k,l])
end
end
""")
knl = lp.tag_inames(knl, "l:unr")
knl = lp.set_loop_priority(knl, "i,j,k,l")
knl = lp.add_and_infer_dtypes(knl, {"a": np.float32, "b": np.float32})
knl = lp.fix_parameters(knl, n=5)
print(lp.generate_code_v2(knl).device_code())
def test_chunk_iname(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i] = 2*a[i]",
[lp.GlobalArg("out,a", np.float32, shape=lp.auto), "..."],
assumptions="n>0")
ref_knl = knl
knl = lp.chunk_iname(knl, "i", 3, inner_tag="l.0")
knl = lp.set_loop_priority(knl, "i_outer, i_inner")
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=130))
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_chunk_iname(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{ [i]: 0<=i<n }",
"out[i] = 2*a[i]",
[
lp.GlobalArg("out,a", np.float32, shape=lp.auto),
"..."
],
assumptions="n>0")
ref_knl = knl
knl = lp.chunk_iname(knl, "i", 3, inner_tag="l.0")
knl = lp.set_loop_priority(knl, "i_outer, i_inner")
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=130))
def test_numba_cuda_target():
knl = lp.make_kernel("{[i,j,k]: 0<=i,j<M and 0<=k<N}",
"D[i,j] = sqrt(sum(k, (X[i, k]-X[j, k])**2))",
target=lp.NumbaCudaTarget())
knl = lp.assume(knl, "M>0")
knl = lp.split_iname(knl, "i", 16, outer_tag='g.0')
knl = lp.split_iname(knl, "j", 128, inner_tag='l.0', slabs=(0, 1))
knl = lp.add_prefetch(knl, "X[i,:]")
knl = lp.fix_parameters(knl, N=3)
knl = lp.set_loop_priority(knl, "i_inner,j_outer")
knl = lp.tag_inames(knl, "k:unr")
knl = lp.tag_array_axes(knl, "X", "N0,N1")
knl = lp.add_and_infer_dtypes(knl, {"X": np.float32})
print(lp.generate_code_v2(knl).all_code())
def test_assume(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel("{[i]: 0<=i<n}", "a[i] = a[i] + 1",
[lp.GlobalArg("a", np.float32, shape="n"), "..."])
knl = lp.split_iname(knl, "i", 16)
knl = lp.set_loop_priority(knl, "i_outer,i_inner")
knl = lp.assume(knl, "n mod 16 = 0")
knl = lp.assume(knl, "n > 10")
knl = lp.preprocess_kernel(knl, ctx.devices[0])
kernel_gen = lp.generate_loop_schedules(knl)
for gen_knl in kernel_gen:
print(gen_knl)
compiled = lp.CompiledKernel(ctx, gen_knl)
print(compiled.get_code())
assert "if" not in compiled.get_code()
def test_numba_cuda_target():
knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j<M and 0<=k<N}",
"D[i,j] = sqrt(sum(k, (X[i, k]-X[j, k])**2))",
target=lp.NumbaCudaTarget())
knl = lp.assume(knl, "M>0")
knl = lp.split_iname(knl, "i", 16, outer_tag='g.0')
knl = lp.split_iname(knl, "j", 128, inner_tag='l.0', slabs=(0, 1))
knl = lp.add_prefetch(knl, "X[i,:]")
knl = lp.fix_parameters(knl, N=3)
knl = lp.set_loop_priority(knl, "i_inner,j_outer")
knl = lp.tag_inames(knl, "k:unr")
knl = lp.tag_array_axes(knl, "X", "N0,N1")
knl = lp.add_and_infer_dtypes(knl, {"X": np.float32})
print(lp.generate_code_v2(knl).all_code())
def test_generate_c_snippet():
from loopy.target.c import CTarget
from pymbolic import var
I = var("I") # noqa
f = var("f")
df = var("df")
q_v = var("q_v")
eN = var("eN") # noqa
k = var("k")
u = var("u")
from functools import partial
l_sum = partial(lp.Reduction, "sum", allow_simultaneous=True)
Instr = lp.Assignment # noqa
knl = lp.make_kernel(
"{[I, k]: 0<=I<nSpace and 0<=k<nQuad}",
[
Instr(f[I], l_sum(k, q_v[k, I]*u)),
Instr(df[I], l_sum(k, q_v[k, I])),
],
[
lp.GlobalArg("q_v", np.float64, shape="nQuad, nSpace"),
lp.GlobalArg("f,df", np.float64, shape="nSpace"),
lp.ValueArg("u", np.float64),
"...",
],
target=CTarget(),
assumptions="nQuad>=1")
if 0: # enable to play with prefetching
# (prefetch currently requires constant sizes)
knl = lp.fix_parameters(knl, nQuad=5, nSpace=3)
knl = lp.add_prefetch(knl, "q_v", "k,I", default_tag=None)
knl = lp.split_iname(knl, "k", 4, inner_tag="unr", slabs=(0, 1))
knl = lp.set_loop_priority(knl, "I,k_outer,k_inner")
knl = lp.preprocess_kernel(knl)
knl = lp.get_one_scheduled_kernel(knl)
print(lp.generate_body(knl))
def test_assume(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[i]: 0<=i<n}",
"a[i] = a[i] + 1",
[lp.GlobalArg("a", np.float32, shape="n"), "..."])
knl = lp.split_iname(knl, "i", 16)
knl = lp.set_loop_priority(knl, "i_outer,i_inner")
knl = lp.assume(knl, "n mod 16 = 0")
knl = lp.assume(knl, "n > 10")
knl = lp.preprocess_kernel(knl, ctx.devices[0])
kernel_gen = lp.generate_loop_schedules(knl)
for gen_knl in kernel_gen:
print(gen_knl)
compiled = lp.CompiledKernel(ctx, gen_knl)
print(compiled.get_code())
assert "if" not in compiled.get_code()
def test_generate_c_snippet():
from loopy.target.c import CTarget
from pymbolic import var
I = var("I") # noqa
f = var("f")
df = var("df")
q_v = var("q_v")
eN = var("eN") # noqa
k = var("k")
u = var("u")
from functools import partial
l_sum = partial(lp.Reduction, "sum", allow_simultaneous=True)
Instr = lp.Assignment # noqa
knl = lp.make_kernel("{[I, k]: 0<=I<nSpace and 0<=k<nQuad}", [
Instr(f[I], l_sum(k, q_v[k, I] * u)),
Instr(df[I], l_sum(k, q_v[k, I])),
], [
lp.GlobalArg("q_v", np.float64, shape="nQuad, nSpace"),
lp.GlobalArg("f,df", np.float64, shape="nSpace"),
lp.ValueArg("u", np.float64),
"...",
],
target=CTarget(),
assumptions="nQuad>=1")
if 0: # enable to play with prefetching
# (prefetch currently requires constant sizes)
knl = lp.fix_parameters(knl, nQuad=5, nSpace=3)
knl = lp.add_prefetch(knl, "q_v", "k,I", default_tag=None)
knl = lp.split_iname(knl, "k", 4, inner_tag="unr", slabs=(0, 1))
knl = lp.set_loop_priority(knl, "I,k_outer,k_inner")
knl = lp.preprocess_kernel(knl)
knl = lp.get_one_scheduled_kernel(knl)
print(lp.generate_body(knl))
def test_ilp_loop_bound(ctx_factory):
# The salient bit of this test is that a joint bound on (outer, inner)
# from a split occurs in a setting where the inner loop has been ilp'ed.
# In 'normal' parallel loops, the inner index is available for conditionals
# throughout. In ILP'd loops, not so much.
ctx = ctx_factory()
knl = lp.make_kernel("{ [i,j,k]: 0<=i,j,k<n }",
"""
out[i,k] = sum(j, a[i,j]*b[j,k])
""", [
lp.GlobalArg("a,b", np.float32, shape=lp.auto),
"...",
],
assumptions="n>=1")
ref_knl = knl
knl = lp.set_loop_priority(knl, "j,i,k")
knl = lp.split_iname(knl, "k", 4, inner_tag="ilp")
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(n=200))
def test_slab_decomposition_does_not_double_execute(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
knl = lp.make_kernel("{ [i]: 0<=i<n }",
"a[i] = 2*a[i]",
assumptions="n>=1")
ref_knl = knl
for outer_tag in ["for", "g.0"]:
knl = ref_knl
knl = lp.split_iname(knl,
"i",
4,
slabs=(0, 1),
inner_tag="unr",
outer_tag=outer_tag)
knl = lp.set_loop_priority(knl, "i_outer")
a = cl.array.empty(queue, 20, np.float32)
a.fill(17)
a_ref = a.copy()
a_knl = a.copy()
knl = lp.set_options(knl, write_cl=True)
print("TEST-----------------------------------------")
knl(queue, a=a_knl)
print("REF-----------------------------------------")
ref_knl(queue, a=a_ref)
print("DONE-----------------------------------------")
print("REF", a_ref)
print("KNL", a_knl)
assert (a_ref == a_knl).get().all()
print("_________________________________")
def test_precompute_with_preexisting_inames(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[e,i,j,k]: 0<=e<E and 0<=i,j,k<n}", """
result[e,i] = sum(j, D1[i,j]*u[e,j])
result2[e,i] = sum(k, D2[i,k]*u[e,k])
""")
knl = lp.add_and_infer_dtypes(knl, {
"u": np.float32,
"D1": np.float32,
"D2": np.float32,
})
knl = lp.fix_parameters(knl, n=13)
ref_knl = knl
knl = lp.extract_subst(knl, "D1_subst", "D1[ii,jj]", parameters="ii,jj")
knl = lp.extract_subst(knl, "D2_subst", "D2[ii,jj]", parameters="ii,jj")
knl = lp.precompute(knl,
"D1_subst",
"i,j",
default_tag="for",
precompute_inames="ii,jj")
knl = lp.precompute(knl,
"D2_subst",
"i,k",
default_tag="for",
precompute_inames="ii,jj")
knl = lp.set_loop_priority(knl, "ii,jj,e,j,k")
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters=dict(E=200))
nimg=nimg,
f_w=f_w))
ref_knl = knl
##############################################################
# Play with things in here to optimize for speed
lp.split_reduction_outward(knl, "color")
## Order of loop variables
knl = lp.set_loop_priority(knl, [
"im_x_outer",
"im_y_outer",
"im_x_inner",
"im_y_inner",
"n",
"feat",
"f_x",
"f_y"])
## Split loop into innner/outer. 2nd arg is size of inner loop
knl = lp.split_iname(knl, "im_x", 16)
knl = lp.split_iname(knl, "im_y", 16)
#knl = lp.split_iname(knl, "color", 3)
## specifying block/tread ordering of each varible: g.N, l.N, unr, [seq], ilp
knl = lp.tag_inames(knl, dict(im_x_inner="l.0",
im_x_outer="g.0",
im_y_inner="l.1",
def variant_1(knl):
knl = lp.split_iname(knl, "im_x", 16, inner_tag="l.0")
knl = lp.set_loop_priority(knl, "iimg,im_x_outer,im_y,ifeat,f_x,f_y")
return knl
def variant_1(knl):
knl = lp.add_prefetch(knl, "a")
knl = lp.add_prefetch(knl, "b")
knl = lp.set_loop_priority(knl, ["i", "j"])
return knl
def set_up_volume_loop(kernel, Nq):
kernel = lp.fix_parameters(kernel, Nq=Nq)
kernel = lp.set_loop_priority(kernel, "e,k,j,i")
kernel = lp.tag_inames(kernel, dict(e="g.0", j="l.1", i="l.0"))
kernel = lp.assume(kernel, "elements >= 1")
return kernel
def variant_1(knl):
knl = lp.split_iname(knl, "im_x", 16, inner_tag="l.0")
knl = lp.set_loop_priority(knl, "iimg,im_x_outer,im_y,ifeat,f_x,f_y")
return knl
def set_up_volume_loop(kernel, Nq):
kernel = lp.fix_parameters(kernel, Nq=Nq)
kernel = lp.set_loop_priority(kernel, "e,k,j,i")
kernel = lp.tag_inames(kernel, dict(e="g.0", j="l.1", i="l.0"))
kernel = lp.assume(kernel, "elements >= 1")
return kernel
def variant_1(knl):
knl = lp.add_prefetch(knl, "a")
knl = lp.add_prefetch(knl, "b")
knl = lp.set_loop_priority(knl, ["i", "j"])
return knl
def variant_2(knl):
knl = lp.precompute(knl, "dpsi", "i,ell", default_tag='for')
knl = lp.set_loop_priority(knl, "c,i,j")
return knl
def variant_1(knl):
knl = lp.add_prefetch(knl, "a")
knl = lp.add_prefetch(knl, "b")
knl = lp.set_loop_priority(knl, ["i", "j"])
knl = lp.add_inames_to_insn(knl, "i", "writes:b_fetch")
return 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 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)