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