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, 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_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_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_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_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_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_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_finite_difference_expr_subst(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
grid = np.linspace(0, 2*np.pi, 2048, endpoint=False)
h = grid[1] - grid[0]
u = cl.clmath.sin(cl.array.to_device(queue, grid))
fin_diff_knl = lp.make_kernel(
"{[i]: 1<=i<=n}",
"out[i] = -(f[i+1] - f[i-1])/h",
[lp.GlobalArg("out", shape="n+2"), "..."])
flux_knl = lp.make_kernel(
"{[j]: 1<=j<=n}",
"f[j] = u[j]**2/2",
[
lp.GlobalArg("f", shape="n+2"),
lp.GlobalArg("u", shape="n+2"),
])
fused_knl = lp.fuse_kernels([fin_diff_knl, flux_knl],
data_flow=[
("f", 1, 0)
])
fused_knl = lp.set_options(fused_knl, write_cl=True)
evt, _ = fused_knl(queue, u=u, h=np.float32(1e-1))
fused_knl = lp.assignment_to_subst(fused_knl, "f")
fused_knl = lp.set_options(fused_knl, write_cl=True)
# This is the real test here: The automatically generated
# shape expressions are '2+n' and the ones above are 'n+2'.
# Is loopy smart enough to understand that these are equal?
evt, _ = fused_knl(queue, u=u, h=np.float32(1e-1))
fused0_knl = lp.affine_map_inames(fused_knl, "i", "inew", "inew+1=i")
gpu_knl = lp.split_iname(
fused0_knl, "inew", 128, outer_tag="g.0", inner_tag="l.0")
precomp_knl = lp.precompute(
gpu_knl, "f_subst", "inew_inner", fetch_bounding_box=True)
precomp_knl = lp.tag_inames(precomp_knl, {"j_0_outer": "unr"})
precomp_knl = lp.set_options(precomp_knl, return_dict=True)
evt, _ = precomp_knl(queue, u=u, h=h)
def test_finite_difference_expr_subst(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
grid = np.linspace(0, 2 * np.pi, 2048, endpoint=False)
h = grid[1] - grid[0]
u = cl.clmath.sin(cl.array.to_device(queue, grid))
fin_diff_knl = lp.make_kernel("{[i]: 1<=i<=n}",
"out[i] = -(f[i+1] - f[i-1])/h",
[lp.GlobalArg("out", shape="n+2"), "..."])
flux_knl = lp.make_kernel("{[j]: 1<=j<=n}", "f[j] = u[j]**2/2", [
lp.GlobalArg("f", shape="n+2"),
lp.GlobalArg("u", shape="n+2"),
])
fused_knl = lp.fuse_kernels([fin_diff_knl, flux_knl],
data_flow=[("f", 1, 0)])
fused_knl = lp.set_options(fused_knl, write_cl=True)
evt, _ = fused_knl(queue, u=u, h=np.float32(1e-1))
fused_knl = lp.assignment_to_subst(fused_knl, "f")
fused_knl = lp.set_options(fused_knl, write_cl=True)
# This is the real test here: The automatically generated
# shape expressions are '2+n' and the ones above are 'n+2'.
# Is loopy smart enough to understand that these are equal?
evt, _ = fused_knl(queue, u=u, h=np.float32(1e-1))
fused0_knl = lp.affine_map_inames(fused_knl, "i", "inew", "inew+1=i")
gpu_knl = lp.split_iname(fused0_knl,
"inew",
128,
outer_tag="g.0",
inner_tag="l.0")
precomp_knl = lp.precompute(gpu_knl,
"f_subst",
"inew_inner",
fetch_bounding_box=True)
precomp_knl = lp.tag_inames(precomp_knl, {"j_0_outer": "unr"})
precomp_knl = lp.set_options(precomp_knl, return_dict=True)
evt, _ = precomp_knl(queue, u=u, h=h)
def test_precompute_does_not_lead_to_dep_cycle(ctx_factory):
# See https://github.com/inducer/loopy/issues/498
ctx = ctx_factory()
knl = lp.make_kernel(
"{[i]: 0<=i<10}", """
<> tmp0[i] = 2 * i
<> tmp1[i] = 2 * tmp0[i]
<> tmp2[i] = 3 * tmp1[i]
out[i] = 2*tmp1[i] + 3*tmp2[i]
""")
ref_knl = knl
knl = lp.assignment_to_subst(knl, "tmp1")
knl = lp.precompute(knl, "tmp1_subst")
lp.auto_test_vs_ref(knl, ctx, ref_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)
