def test_global_parallel_reduction_simpler(ctx_factory, size):
ctx = ctx_factory()
pytest.xfail("very sensitive to kernel ordering, fails unused hw-axis check")
knl = lp.make_kernel(
"{[l,g,j]: 0 <= l < nl and 0 <= g,j < ng}",
"""
<> key = make_uint2(l+nl*g, 1234) {inames=l:g}
<> ctr = make_uint4(0, 1, 2, 3) {inames=l:g,id=init_ctr}
<> vals, ctr = philox4x32_f32(ctr, key) {dep=init_ctr}
<> tmp[g] = sum(l, vals.s0 + 1j*vals.s1 + vals.s2 + 1j*vals.s3)
result = sum(j, tmp[j])
""")
ng = 50
knl = lp.fix_parameters(knl, ng=ng)
knl = lp.set_options(knl, write_cl=True)
ref_knl = knl
knl = lp.split_iname(knl, "l", 128, inner_tag="l.0")
knl = lp.split_reduction_outward(knl, "l_inner")
knl = lp.tag_inames(knl, "g:g.0,j:l.0")
lp.auto_test_vs_ref(ref_knl, ctx, knl, parameters={"nl": size})
def test_split_reduction(ctx_factory):
knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<n}", """
b = sum((i,j,k), a[i,j,k])
""", [
lp.GlobalArg("box_source_starts,box_source_counts_nonchild,a",
None,
shape=None), "..."
])
knl = lp.split_reduction_outward(knl, "j,k")
def test_split_reduction(ctx_factory):
knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<n}",
"""
b = sum((i,j,k), a[i,j,k])
""",
[
lp.GlobalArg("box_source_starts,box_source_counts_nonchild,a",
None, shape=None),
"..."])
knl = lp.split_reduction_outward(knl, "j,k")
def test_global_parallel_reduction(ctx_factory, size):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[i]: 0 <= i < n }", """
# Using z[0] instead of z works around a bug in ancient PyOpenCL.
z[0] = sum(i, a[i])
""")
knl = lp.add_and_infer_dtypes(knl, {"a": np.float32})
ref_knl = knl
gsize = 128
knl = lp.split_iname(knl, "i", gsize * 20)
knl = lp.split_iname(knl, "i_inner", gsize, inner_tag="l.0")
knl = lp.split_reduction_outward(knl, "i_outer")
knl = lp.split_reduction_inward(knl, "i_inner_outer")
from loopy.transform.data import reduction_arg_to_subst_rule
knl = reduction_arg_to_subst_rule(knl, "i_outer")
knl = lp.precompute(knl,
"red_i_outer_arg",
"i_outer",
temporary_address_space=lp.AddressSpace.GLOBAL,
default_tag="l.auto")
knl = lp.realize_reduction(knl)
knl = lp.tag_inames(knl, "i_outer_0:g.0")
# Keep the i_outer accumulator on the correct (lower) side of the barrier,
# otherwise there will be useless save/reload code generated.
knl = lp.add_dependency(knl, "writes:acc_i_outer",
"id:red_i_outer_arg_barrier")
lp.auto_test_vs_ref(ref_knl,
ctx,
knl,
parameters={"n": size},
print_ref_code=True)
def test_global_parallel_reduction(ctx_factory, size):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[i]: 0 <= i < n }",
"""
# Using z[0] instead of z works around a bug in ancient PyOpenCL.
z[0] = sum(i, a[i])
""")
knl = lp.add_and_infer_dtypes(knl, {"a": np.float32})
ref_knl = knl
gsize = 128
knl = lp.split_iname(knl, "i", gsize * 20)
knl = lp.split_iname(knl, "i_inner", gsize, inner_tag="l.0")
knl = lp.split_reduction_outward(knl, "i_outer")
knl = lp.split_reduction_inward(knl, "i_inner_outer")
from loopy.transform.data import reduction_arg_to_subst_rule
knl = reduction_arg_to_subst_rule(knl, "i_outer")
knl = lp.precompute(knl, "red_i_outer_arg", "i_outer",
temporary_scope=lp.temp_var_scope.GLOBAL,
default_tag="l.auto")
knl = lp.realize_reduction(knl)
knl = lp.tag_inames(knl, "i_outer_0:g.0")
# Keep the i_outer accumulator on the correct (lower) side of the barrier,
# otherwise there will be useless save/reload code generated.
knl = lp.add_dependency(
knl, "writes:acc_i_outer",
"id:red_i_outer_arg_barrier")
lp.auto_test_vs_ref(
ref_knl, ctx, knl, parameters={"n": size},
print_ref_code=True)
lp.GlobalArg("img,f", np.float32, shape=lp.auto),
"..."
],
defines=dict(ncolors=n_input_channels,
nfeature_maps=n_output_channels,
im_w=im_w, im_h=im_h,
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)