def test_join_inames(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[i,j]: 0<=i,j<16}",
["b[i,j] = 2*a[i,j]"],
[
lp.GlobalArg("a", np.float32, shape=(
16,
16,
)),
lp.GlobalArg("b", np.float32, shape=(
16,
16,
))
],
)
ref_knl = knl
knl = lp.add_prefetch(knl,
"a",
sweep_inames=["i", "j"],
default_tag="l.auto")
knl = lp.join_inames(knl, ["a_dim_0", "a_dim_1"])
lp.auto_test_vs_ref(ref_knl, ctx, knl, print_ref_code=True)
def test_join_inames(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
"{[i,j]: 0<=i,j<16}",
[
"b[i,j] = 2*a[i,j]"
],
[
lp.GlobalArg("a", np.float32, shape=(16, 16,)),
lp.GlobalArg("b", np.float32, shape=(16, 16,))
],
)
ref_knl = knl
knl = lp.add_prefetch(knl, "a", sweep_inames=["i", "j"])
knl = lp.join_inames(knl, ["a_dim_0", "a_dim_1"])
lp.auto_test_vs_ref(ref_knl, ctx, knl, print_ref_code=True)