def test_generate_output_stripe_configs_single_dimension():
stripe_factors = 3
subgraph = cs.TESubgraph([], None)
part_1 = cs.InlinePart(
subgraph,
[
cs.Propagator(
[[2, 0, 0], [0, 2, 0], [0, 0, 1]],
[0, 0],
),
],
)
tensor_1 = cs.Tensor([800, 800], "uint8")
tensor_2 = cs.Tensor([400, 400], "uint8")
part_1.set_input(0, tensor_1)
part_1.set_output(tensor_2)
tensor_1.add_consumer(part_1)
tensor_2.add_producer(part_1)
expected_stripe_configs = {
cs.StripeConfig([400, 1], [400, 400], [400, 1], [2, 1], [1, 400], [0, 0]),
cs.StripeConfig([400, 200], [400, 400], [400, 200], [2, 1], [1, 2], [0, 0]),
cs.StripeConfig([1, 400], [400, 400], [1, 400], [1, 2], [400, 1], [0, 0]),
cs.StripeConfig([200, 400], [400, 400], [200, 400], [1, 2], [2, 1], [0, 0]),
cs.StripeConfig([400, 400], [400, 400], [400, 400], [1, 2], [1, 1], [0, 0]),
}
output_stripe_configs = _generate_output_stripe_configs(
part=part_1, stripe_factors=stripe_factors, enable_striping=True, multi_dimensional=False
)
assert len(output_stripe_configs) == len(expected_stripe_configs)
assert set(output_stripe_configs) == expected_stripe_configs
def test_inline_part():
subgraph = cs.TESubgraph([], None)
part = cs.InlinePart(
subgraph,
[
cs.Propagator(
[[0, 1, 0], [1, 0, 0], [0, 0, 1]],
[0, 0],
),
],
)
output_stripe_config = cs.StripeConfig([2, 4], [8, 8], [2, 4], [1, 2],
[4, 2], [0, 0])
input_stripe_config = cs.StripeConfig([4, 2], [8, 8], [4, 2], [2, 1],
[2, 4], [0, 0])
assert part.input_tensors == [None]
assert part.output_tensor == None
assert len(part.propagators) == 1
assert part.in_line == True
assert part.get_stripe_align_hint() == [1, 1]
performance_info = part.get_performance_info(output_stripe_config,
is_rolling=False)
assert performance_info.compute_cycles == 0
assert performance_info.read_bytes == [0]
assert performance_info.write_bytes == 0
input_stripe_configs = part.calculate_input_stripe_configs(
output_stripe_config)
assert len(input_stripe_configs) == 1
assert input_stripe_configs[0] == input_stripe_config
def test_ethosu_part():
te_subgraph = cs.TESubgraph([], None)
output_quantum = [1, 2, 2, 8]
propagator = cs.Propagator(
[[1, 0, 0, 0, 2], [0, 1, 0, 0, 2], [0, 0, 1, 0, 0], [0, 0, 0, 1, 0], [0, 0, 0, 0, 1]],
[0, 0, 0, 0],
)
stripe_config = cs.StripeConfig(
[1, 4, 4, 16], [1, 64, 72, 96], [1, 4, 4, 16], [1, 2, 3, 4], [1, 16, 13, 6], [0, 0, 0, 0]
)
subkernels = 3
valid_block_configs = [cs.BlockConfig([1, 2, 4, 16], 15000, 7500)]
part = EthosuPart(
te_subgraph,
[propagator],
output_quantum,
subkernels,
valid_block_configs,
1,
)
input_tensor = cs.Tensor(shape=[1, 66, 74, 16], dtype="int8")
part.set_input(0, input_tensor)
assert part.get_stripe_align_hint() == output_quantum
# Check that the performance model runs, don't verify output
part.get_performance_info(stripe_config, BufferMode.ROLLING)
part.get_performance_info(stripe_config, BufferMode.RECOMPUTE)
def test_ethosu_part():
te_subgraph = pl.TESubgraph([], None)
output_quantum = [1, 2, 2, 8]
quantum_cycles = 32
propagator = pl.Propagator(
[[1, 0, 0, 0, 2], [0, 1, 0, 0, 2], [0, 0, 1, 0, 0], [0, 0, 0, 1, 0],
[0, 0, 0, 0, 1]],
[0, 0, 0, 0],
)
stripe_config = pl.StripeConfig([1, 4, 4, 16], [1, 64, 72, 96],
[1, 4, 4, 16], [1, 2, 3, 4],
[1, 16, 13, 6], [0, 0, 0, 0])
part = EthosuPart(te_subgraph, [propagator], output_quantum,
quantum_cycles)
assert part.get_stripe_align_hint() == output_quantum
# Check that the performance model runs, don't verify output
part.get_performance_info(stripe_config, False)
part.get_performance_info(stripe_config, True)
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 test_best_block_config(
test_id,
op_type,
activation,
kernel,
stride,
dilation,
padding,
in_shape,
out_shape,
layouts,
acc_config,
expected_block_configs,
):
nhwc_to_nhcwb16 = [
[1, 0, 0, 0, 0],
[0, 1, 0, 0, 0],
[0, 0, 0, 1 / 16, 0],
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 16],
[0, 0, 0, 0, 1],
]
nhcwb16_to_nhwc = [
[1, 0, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0],
[0, 0, 16, 0, 1, -16],
[0, 0, 0, 0, 0, 1],
]
ifm_matrix, ifm_offset, weight_matrix, weight_offset, _, _ = make_matrices(
op_type, kernel, stride, padding, layouts[0], layouts[1], dilation,
in_shape[3])
ofm_channels = out_shape[3]
ifm_channels = in_shape[3]
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,
)
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]
def test_conv_performance(
accelerator,
op_type,
activation,
kernel,
stride,
dilation,
padding,
in_shape,
out_shape,
block_shape,
input_block_shape,
expected,
):
ifm_channels = in_shape[3]
ifm_matrix, ifm_offset, weight_matrix, weight_offset, _, _ = make_matrices(
op_type,
kernel,
stride,
padding,
"NHWC",
"NHWC",
dilation,
ifm_channels,
)
propagator = cs.Propagator(ifm_matrix, ifm_offset)
weight_propagator = cs.Propagator(weight_matrix, weight_offset)
subkernels = ((kernel[0] + 7) // 8) * ((kernel[1] + 7) // 8)
device_config = cs.EthosuDeviceConfig(accelerator)
output_cycles = device_config._get_output_cycles(op_type, "", "int8", "int8", activation)
output_cycles *= reduce(lambda a, b: a * b, block_shape, 1)
is_partkernel = device_config.is_partkernel(
op_type, ifm_channels, "int8", kernel[0] * kernel[1]
)
compute_cycles = device_config._estimate_compute_cycles_per_block(
op_type,
_Shape(block_shape),
_Shape(input_block_shape),
kernel[0],
kernel[1],
ifm_channels,
"int8",
is_partkernel,
)
block_configs = [
cs.BlockConfig(input_block_shape, block_shape, compute_cycles, int(output_cycles))
]
output_quantum = [1, 1, 2, 8]
te_subgraph = cs.TESubgraph([], None)
part = cs.EthosuPart(
te_subgraph,
[propagator, weight_propagator],
output_quantum,
subkernels,
block_configs,
1,
)
part.set_input(0, cs.Tensor(in_shape, "int8"))
part.set_input(1, cs.Tensor([ifm_channels, kernel[0], kernel[1], out_shape[-1]], "int8"))
part.set_output(cs.Tensor(out_shape, "int8"))
stripes = [1] * len(output_quantum)
offset = [0] * len(output_quantum)
order = [1, 2, 3, 4]
stripe_config = cs.StripeConfig(out_shape, out_shape, out_shape, order, stripes, offset)
compute_cycles = part.get_performance_info(stripe_config, cs.BufferMode.ROLLING).compute_cycles
tolerance = expected * 0.1
assert expected - tolerance <= compute_cycles <= expected + tolerance
def test_ethosu_conv2d_block_config_from_matcher(ifm_layout, ofm_layout,
ifm_channels,
expected_cycles):
ofm_channels = 10
ifm_height = 123
ifm_width = 155
ifm_shape = ((1, ifm_height, ifm_width,
ifm_channels) if ifm_layout == "NHWC" else
(1, ifm_height, 1 + ((ifm_channels - 1) // 16), ifm_width,
16))
weight_shape = (ofm_channels, 3, 3, ifm_channels)
scale_bias_shape = (ofm_channels, 10)
ifm = te.placeholder(ifm_shape, dtype="int8")
weight = te.placeholder(weight_shape, dtype="int8")
scale_bias = te.placeholder(scale_bias_shape, dtype="uint8")
lut = te.placeholder((), dtype="uint8")
out = conv2d_compute(
ifm=ifm,
weight=weight,
scale_bias=scale_bias,
lut=lut,
ifm_scale=1,
ifm_zero_point=0,
ofm_scale=1,
ofm_zero_point=0,
weight_zero_point=0,
strides=(1, 1),
padding=(0, 0, 0, 0),
dilation=(1, 1),
activation="NONE",
clip_min=0,
clip_max=0,
upscale="NONE",
rounding_mode="TFL",
ifm_layout=ifm_layout,
ofm_layout=ofm_layout,
)
device_config = cs.EthosuDeviceConfig("ethos-u55-256")
part = match_ethosu_conv2d(out, device_config)
ofm_shape = [int(i) for i in part.subgraph.output_tensor.shape]
# Add inputs and outputs to the part
input_tensor = cs.Tensor(ifm_shape, "int8")
part.set_input(0, input_tensor)
weight_tensor = cs.Tensor(weight_shape, "int8")
part.set_input(1, weight_tensor)
scale_bias_tensor = cs.Tensor(scale_bias_shape, "int8")
part.set_input(2, scale_bias_tensor)
output_tensor = cs.Tensor(ofm_shape, "int8")
part.set_output(output_tensor)
# Create a stripe of a size of the output tensor
order = [1, 2, 3, 4] if ofm_layout == "NHWC" else [1, 2, 4, 3, 0]
stripes = [1] * len(order)
offset = [0] * len(order)
stripe_config = cs.StripeConfig(ofm_shape, ofm_shape, ofm_shape, order,
stripes, offset)
block = part.get_block_config(stripe_config)
# Since we dont know the values of the variables we passed to the get_valid_block_configs in
# the matcher, best we can do is to verify the compute cycle count since the channels have a
# significant effect on it
assert block.compute_cycles == expected_cycles
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 = [
[1, 0, 0, 0, 0],
[0, 1, 0, 0, 0],
[0, 0, 0, 1 / 16, 0],
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 16],
[0, 0, 0, 0, 1],
]
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)
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
# Layout conversions will align the propagated IFMs to the brick, i.e. 16
# so the expected ifm_shape needs to be rounded up to 16
if ifm_layout != ofm_layout:
assert ifm_shape[:-1] == propagated_ifm[:-1]
assert ((ifm_shape[-1] + 16 - 1) // 16) * 16 == propagated_ifm[-1]
else:
assert ifm_shape == propagated_ifm