def test_magma_fermi_matrix_mul(ctx_factory):
dtype = np.float32
ctx = ctx_factory()
order = "C"
n = get_suitable_size(ctx)
if (not ctx.devices[0].image_support
or ctx.devices[0].platform.name == "Portable Computing Language"):
pytest.skip("crashes on pocl")
image_format = cl.ImageFormat(cl.channel_order.R, cl.channel_type.FLOAT)
if image_format not in cl.get_supported_image_formats(
ctx, cl.mem_flags.READ_ONLY, cl.mem_object_type.IMAGE2D):
pytest.skip("image format not supported")
knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<%d}" % n, ["c[i, j] = sum(k, a[i, k]*b[k, j])"], [
lp.ImageArg("a", dtype, shape=(n, n)),
lp.ImageArg("b", dtype, shape=(n, n)),
lp.GlobalArg("c", dtype, shape=(n, n), order=order),
],
name="matmul")
seq_knl = knl
i_reg = 4
j_reg = 4
i_chunks = 16
j_chunks = 16
knl = lp.split_iname(knl, "i", i_reg * i_chunks, outer_tag="g.0")
knl = lp.split_iname(knl,
"i_inner",
i_reg,
outer_tag="l.0",
inner_tag="ilp")
knl = lp.split_iname(knl, "j", j_reg * j_chunks, outer_tag="g.1")
knl = lp.split_iname(knl,
"j_inner",
j_reg,
outer_tag="l.1",
inner_tag="ilp")
knl = lp.split_iname(knl, "k", 16)
knl = lp.split_iname(knl, "k_inner", 8, outer_tag="unr")
# FIXME
#knl = lp.add_prefetch(knl, 'a', ["k_inner", "i_inner_inner", "i_inner_outer"],
# default_tag="l.auto")
#knl = lp.add_prefetch(knl, 'b',
# ["k_inner", ("j_inner_inner", "j_inner_outer"),], default_tag="l.auto")
lp.auto_test_vs_ref(seq_knl,
ctx,
knl,
op_count=[2 * n**3 / 1e9],
op_label=["GFlops"],
parameters={},
blacklist_ref_vendors="pocl")
def no_test_image_matrix_mul_ilp(ctx_factory):
dtype = np.float32
ctx = ctx_factory()
order = "C"
if (not ctx.devices[0].image_support
or ctx.devices[0].platform.name == "Portable Computing Language"):
pytest.skip("crashes on pocl")
image_format = cl.ImageFormat(cl.channel_order.R, cl.channel_type.FLOAT)
if image_format not in cl.get_supported_image_formats(
ctx, cl.mem_flags.READ_ONLY, cl.mem_object_type.IMAGE2D):
pytest.skip("image format not supported")
n = get_suitable_size(ctx)
knl = lp.make_kernel(
"{[i,j,k]: 0<=i,j,k<%d}" % n, ["c[i, j] = sum(k, a[i, k]*b[k, j])"], [
lp.ImageArg("a", dtype, shape=(n, n)),
lp.ImageArg("b", dtype, shape=(n, n)),
lp.GlobalArg("c", dtype, shape=(n, n), order=order),
],
name="matmul")
seq_knl = knl
ilp = 4
knl = lp.split_iname(knl, "i", 2, outer_tag="g.0", inner_tag="l.1")
j_inner_split = 4
knl = lp.split_iname(knl, "j", ilp * j_inner_split, outer_tag="g.1")
knl = lp.split_iname(knl,
"j_inner",
j_inner_split,
outer_tag="ilp",
inner_tag="l.0")
knl = lp.split_iname(knl, "k", 2)
# conflict-free?
knl = lp.add_prefetch(knl,
'a', ["i_inner", "k_inner"],
default_tag="l.auto")
knl = lp.add_prefetch(knl,
'b', ["j_inner_outer", "j_inner_inner", "k_inner"],
default_tag="l.auto")
lp.auto_test_vs_ref(seq_knl,
ctx,
knl,
op_count=[2 * n**3 / 1e9],
op_label=["GFlops"],
parameters={})