def _make_matrices(broadcast, ifm_layout, ifm2_layout, ofm_layout,
ofm_channels):
broadcast_h, broadcast_w, broadcast_c = broadcast
nhwc_to_nhcwb16, nhcwb16_to_nhwc = get_layout_transform_matrices(
ofm_channels)
ifm_matrix = [
[1, 0, 0, 0, 0],
[0, 1, 0, 0, 0],
[0, 0, 1, 0, 0],
[0, 0, 0, 1, 0],
[0, 0, 0, 0, 1],
]
ifm2_matrix = [
[1, 0, 0, 0, 0],
[0, (1 - broadcast_h), 0, 0, broadcast_h],
[0, 0, (1 - broadcast_w), 0, broadcast_w],
[0, 0, 0, (1 - broadcast_c), broadcast_c],
[0, 0, 0, 0, 1],
]
if ofm_layout == "NHCWB16":
ifm_matrix = np.matmul(ifm_matrix, nhcwb16_to_nhwc).tolist()
ifm2_matrix = np.matmul(ifm2_matrix, nhcwb16_to_nhwc).tolist()
if ifm_layout == "NHCWB16":
ifm_matrix = np.matmul(nhwc_to_nhcwb16, ifm_matrix).tolist()
if ifm2_layout == "NHCWB16":
ifm2_matrix = np.matmul(nhwc_to_nhcwb16, ifm2_matrix).tolist()
return (ifm_matrix, ifm2_matrix)
def _make_matrices(ifm_layout, ofm_layout, ofm_channels):
nhwc_to_nhcwb16, nhcwb16_to_nhwc = get_layout_transform_matrices(ofm_channels)
ifm_matrix = [
[1, 0, 0, 0, 0],
[0, 1, 0, 0, 0],
[0, 0, 1, 0, 0],
[0, 0, 0, 1, 0],
[0, 0, 0, 0, 1],
]
if ofm_layout == "NHCWB16":
ifm_matrix = np.matmul(ifm_matrix, nhcwb16_to_nhwc).tolist()
if ifm_layout == "NHCWB16":
ifm_matrix = np.matmul(nhwc_to_nhcwb16, ifm_matrix).tolist()
return ifm_matrix
def test_ethosu_unary_elementwise_matcher(ofm_shape, ifm_layout, ofm_layout, op_type):
ifm_shape = ofm_shape.copy()
ofm_channels = ofm_shape[3]
nhwc_to_nhcwb16, _ = get_layout_transform_matrices(ofm_channels)
if ifm_layout == "NHCWB16":
ifm_shape = [
int(math.ceil(n))
for n in np.matmul(
nhwc_to_nhcwb16,
ifm_shape
+ [
1,
],
).tolist()[:-1]
]
if ofm_layout == "NHCWB16":
ofm_shape = [
int(math.ceil(n))
for n in np.matmul(
nhwc_to_nhcwb16,
ofm_shape
+ [
1,
],
).tolist()[:-1]
]
order = [1, 2, 4, 3, 0]
else:
order = [1, 2, 3, 4]
ifm = te.placeholder(ifm_shape, dtype="int8")
lut = te.placeholder((), dtype="uint8")
out = unary_elementwise_compute(
ifm=ifm,
lut=lut,
operator_type=op_type,
ifm_scale=1,
ifm_zero_point=0,
ofm_scale=1,
ofm_zero_point=0,
ofm_channels=ofm_channels,
activation="NONE",
clip_min=0,
clip_max=0,
rounding_mode="TFL",
ifm_layout=ifm_layout,
ofm_layout=ofm_layout,
)
ifm_propagator = out.op.attrs["ifm_propagator"]
offset = [0] * len(ofm_shape)
stripes = [0] * len(ofm_shape)
output_stripe_config = cs.StripeConfig(ofm_shape, ofm_shape, ofm_shape, order, stripes, offset)
ifm_transform = _make_matrices(ifm_layout, ofm_layout, ofm_channels)
device_config = cs.EthosuDeviceConfig("ethos-u55-256")
part = match_ethosu_unary_elementwise(out, device_config)
assert isinstance(part, cs.EthosuPart)
assert len(part.propagators) == 1
assert part.propagators[0].transform == ifm_transform
propagated_ifm = ifm_propagator.propagate(output_stripe_config).shape
# The layout transforms that have the exact number of output channels in them
# will lose no information about the number of channels
assert ifm_shape == propagated_ifm
def make_matrices(
op_type,
kernel,
stride,
padding,
ifm_layout,
ofm_layout,
dilation=(1, 1),
ifm_channels=1,
ofm_channels=1,
):
kernel_h, kernel_w = kernel
stride_h, stride_w = stride
dilation_h, dilation_w = dilation
dilated_kernel_h = (kernel_h - 1) * dilation_h + 1
dilated_kernel_w = (kernel_w - 1) * dilation_w + 1
nhwc_to_nhcwb16, nhcwb16_to_nhwc = get_layout_transform_matrices(ofm_channels)
if op_type == "ethosu_conv2d":
ifm_matrix = [
[1, 0, 0, 0, 0],
[0, stride_h, 0, 0, (dilated_kernel_h - stride_h)],
[0, 0, stride_w, 0, (dilated_kernel_w - stride_w)],
[0, 0, 0, 0, ifm_channels],
[0, 0, 0, 0, 1],
]
weight_matrix = [
[0, 0, 0, 1, 0],
[0, 0, 0, 0, kernel_h],
[0, 0, 0, 0, kernel_w],
[0, 0, 0, 0, ifm_channels],
[0, 0, 0, 0, 1],
]
elif op_type == "ethosu_depthwise_conv2d":
ifm_matrix = [
[1, 0, 0, 0, 0],
[0, stride_h, 0, 0, (dilated_kernel_h - stride_h)],
[0, 0, stride_w, 0, (dilated_kernel_w - stride_w)],
[0, 0, 0, 1, 0],
[0, 0, 0, 0, 1],
]
weight_matrix = [
[0, 0, 0, 1, 0],
[0, 0, 0, 0, kernel_h],
[0, 0, 0, 0, kernel_w],
[0, 0, 0, 0, 1],
[0, 0, 0, 0, 1],
]
elif op_type == "ethosu_pooling":
ifm_matrix = [
[1, 0, 0, 0, 0],
[0, stride_h, 0, 0, (dilated_kernel_h - stride_h)],
[0, 0, stride_w, 0, (dilated_kernel_w - stride_w)],
[0, 0, 0, 1, 0],
[0, 0, 0, 0, 1],
]
weight_matrix = [
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
]
scale_bias_matrix = [
[0, 0, 0, 1, 0],
[0, 0, 0, 0, 10],
[0, 0, 0, 0, 1],
]
if ofm_layout == "NHCWB16":
ifm_matrix = np.matmul(ifm_matrix, nhcwb16_to_nhwc).tolist()
weight_matrix = np.matmul(weight_matrix, nhcwb16_to_nhwc).tolist()
scale_bias_matrix = np.matmul(scale_bias_matrix, nhcwb16_to_nhwc).tolist()
if ifm_layout == "NHCWB16":
ifm_matrix = np.matmul(nhwc_to_nhcwb16, ifm_matrix).tolist()
ifm_offset = (
[0, -padding[0], -padding[1], 0]
if ifm_layout == "NHWC"
else [0, -padding[0], 0, -padding[1], 0]
)
weight_offset = [0, 0, 0, 0]
scale_bias_offset = [0, 0]
return (
ifm_matrix,
ifm_offset,
weight_matrix,
weight_offset,
scale_bias_matrix,
scale_bias_offset,
)
def test_best_block_config(
test_id,
op_type,
activation,
kernel,
stride,
dilation,
padding,
in_shape,
out_shape,
layouts,
acc_config,
expected_block_configs,
):
ofm_channels = out_shape[3]
ifm_channels = in_shape[3]
nhwc_to_nhcwb16, _ = get_layout_transform_matrices(ofm_channels)
ifm_matrix, ifm_offset, weight_matrix, weight_offset, _, _ = make_matrices(
op_type,
kernel,
stride,
padding,
layouts[0],
layouts[1],
dilation,
ifm_channels,
ofm_channels,
)
if layouts[0] == "NHCWB16":
in_shape = [
int(math.ceil(n))
for n in np.matmul(nhwc_to_nhcwb16, in_shape + (1, )).tolist()[:-1]
]
if layouts[1] == "NHCWB16":
out_shape = [
int(math.ceil(n)) for n in np.matmul(nhwc_to_nhcwb16, out_shape +
(1, )).tolist()[:-1]
]
propagator = cs.Propagator(ifm_matrix, ifm_offset)
weight_propagator = cs.Propagator(weight_matrix, weight_offset)
subkernels = ((kernel[0] + 7) // 8) * ((kernel[1] + 7) // 8)
op_attrs = {
"op": op_type,
"activation": activation,
"stride_h": stride[0],
"stride_w": stride[1],
"dilation_h": dilation[0],
"dilation_w": dilation[1],
}
device_config = cs.EthosuDeviceConfig(acc_config)
block_configs = device_config.get_valid_block_configs(
propagator,
op_attrs,
out_shape,
ofm_channels,
ifm_channels,
layouts[1],
layouts[0],
"int8",
"int8",
kernel[0],
kernel[1],
)
output_quantum = [1, 1, 2, 8]
if layouts[1] == "NHCWB16":
output_quantum = [1, 1, 1, 2, 8]
# Create EthosUPart
te_subgraph = cs.TESubgraph([], None)
part = cs.EthosuPart(
te_subgraph,
[propagator, weight_propagator],
output_quantum,
subkernels,
block_configs,
1,
)
# Add tensors
input_tensor = cs.Tensor(in_shape, "int8")
part.set_input(0, input_tensor)
if op_type == "ethosu_conv2d":
weight_tensor = cs.Tensor(
[ofm_channels, kernel[0], kernel[1], ifm_channels], "int8")
part.set_input(1, weight_tensor)
elif op_type == "ethosu_depthwise_conv2d":
weight_tensor = cs.Tensor([ofm_channels, kernel[0], kernel[1], 1],
"int8")
part.set_input(1, weight_tensor)
output_tensor = cs.Tensor(out_shape, "int8")
part.set_output(output_tensor)
order = [1, 2, 3, 4] if layouts[1] == "NHCWB16" else [1, 2, 4, 3, 0]
stripes = [1] * len(output_quantum)
offset = [0] * len(output_quantum)
stripe_config = cs.StripeConfig(out_shape, out_shape, out_shape, order,
stripes, offset)
block = part.get_block_config(stripe_config)
block_shape = tuple(int(a) for a in block.output_shape)
assert block_shape in expected_block_configs[test_id]