def test_workload_padding(
self,
target,
input_shape,
weight_shape,
stride,
padding,
dilation,
dtype,
ref_data,
):
a_np, w_np, b_np, c_np = ref_data
_, _, out_height, out_width = c_np.shape
A = te.placeholder(input_shape, name="A", dtype=dtype)
W = te.placeholder(weight_shape, name="W", dtype=dtype)
with tvm.target.Target(target):
wkl = _get_workload(A, W, (stride, stride), padding, dilation, dtype)
# check if tile_ow candidates are the factors of the right output weight.
cfg = autotvm.get_config()
_fallback_schedule(cfg, wkl)
ow_tile = np.prod(cfg["tile_ow"].size)
tvm.testing.assert_allclose(ow_tile, out_width)
def verify_workload_padding():
_, _, out_height, out_width = get_const_tuple(c_np.shape)
wkl = _get_workload(A, W, (stride, stride), padding, dilation, dtype)
# check if tile_ow candidates are the factors of the right output weight.
cfg = autotvm.get_config()
_fallback_schedule(cfg, wkl)
ow_tile = np.prod(cfg["tile_ow"].size)
tvm.testing.assert_allclose(ow_tile, out_width)
def verify_workload_padding():
_, _, out_height, out_width = get_const_tuple(c_np.shape)
wkl = _get_workload(A, W, (stride, stride), padding, dilation, dtype)
# for testing functionality,
# we choose arbitrary int32_lanes and num_int8_elements can divide the channel,
# regardless of the performance.
int32_lanes, num_int8_elements = num_filter, in_channel
# check if tile_ow candidates are the factors of the right output weight.
cfg = autotvm.get_config()
fallback_schedule_cpu_common_int8(cfg, wkl, int32_lanes, num_int8_elements)
ow_tile = np.prod(cfg["tile_ow"].size)
tvm.testing.assert_allclose(ow_tile, out_width)
