def get_optimal_block_config(
npu_op: vapi.NpuOperation,
accel_config: vapi.NpuAccelerator) -> vapi.NpuShape3D:
"""
"The NPU's unit of work is known as a block. It will fetch block(s) from Input
Feature Map (IFM) and a compute block for Output Feature Map (OFM).
Therefore, we need to pick an optimal block configuration considering bandwidth
to bring IFM blocks and the number of OFM block computes need to happen
to cover the OFM as indicated by the npu op.
Parameters
----------
npu_op : ethosu.vela.api.NpuOperation
The NPU operation and its params
accel_config : ethosu.vela.api.NpuAccelerator
The NPU accelerator config
Returns
-------
ethosu.vela.api.NpuShape3D :
The optimal block config for the operator
"""
options = tvm.transform.PassContext.current().config.get(
"relay.ext.ethos-u.options", None)
if options and options.dev_force_block_config:
block_config = [
int(v) for v in options.dev_force_block_config.split("x")
]
return vapi.NpuShape3D(height=block_config[0],
width=block_config[1],
depth=block_config[2])
all_valid_block_configs = vapi.npu_find_block_configs(npu_op, accel_config)
return _get_optimal_block_config(all_valid_block_configs)
def get_optimal_block_config(
npu_op: vapi.NpuOperation,
accel_config: vapi.NpuAccelerator) -> vapi.NpuShape3D:
"""
"The NPU's unit of work is known as a block. It will fetch block(s) from Input
Feature Map (IFM) and a compute block for Output Feature Map (OFM).
Therefore, we need to pick an optimal block configuration considering bandwidth
to bring IFM blocks and the number of OFM block computes need to happen
to cover the OFM as indicated by the npu op.
Parameters
----------
npu_op : ethosu.vela.api.NpuOperation
The NPU operation and its params
accel_config : ethosu.vela.api.NpuAccelerator
The NPU accelerator config
Returns
-------
ethosu.vela.api.NpuShape3D :
The optimal block config for the operator
"""
all_valid_block_configs = vapi.npu_find_block_configs(npu_op, accel_config)
return _get_optimal_block_config(all_valid_block_configs)