diff options
Diffstat (limited to 'ethosu/vela/npu_performance.py')
| -rw-r--r-- | ethosu/vela/npu_performance.py | 33 |
1 files changed, 31 insertions, 2 deletions
diff --git a/ethosu/vela/npu_performance.py b/ethosu/vela/npu_performance.py index 21b420bf..08967f49 100644 --- a/ethosu/vela/npu_performance.py +++ b/ethosu/vela/npu_performance.py @@ -22,6 +22,8 @@ import copy from enum import auto from enum import IntEnum +from typing import Set +from uuid import UUID import numpy as np @@ -31,11 +33,13 @@ from .architecture_features import Accelerator from .architecture_features import NpuBlockType from .architecture_features import SHRAMElements from .architecture_features import TensorFormat +from .nn_graph import Graph from .numeric_util import round_up from .operation import Kernel from .operation import Op from .scheduler import Schedule from .scheduler import SchedulerOperation +from .scheduler import SchedulerOpInfo from .shape4d import Shape4D from .tensor import BandwidthDirection from .tensor import MemArea @@ -725,16 +729,39 @@ def estimate_full_op_performance( return bws, macs, cycles_a -def calc_new_performance_for_network(nng, arch): +def calc_new_performance_for_network(nng: Graph, arch): total_bws = make_bandwidth_array() total_macs = 0 total_cycles = np.zeros(PassCycles.Size) + total_weight_size = 0 + total_encoded_weight_size = 0 + + # Store unique instances of original/encoded weight tensor uuids to prevent double counting of weights + original_weight_uuids: Set[UUID] = set() + encoded_npu_weight_uuids: Set[UUID] = set() for sg in nng.subgraphs: prev_op = None for sched_op in sg.sched_ops: - op_info = sg.schedule.cost_map[sched_op] + op_info: SchedulerOpInfo = sg.schedule.cost_map[sched_op] bws, macs, cycles = estimate_full_op_performance(arch, sg.schedule, sched_op, prev_op, op_info.block_config) + + # Tensors for calculating weight sizes + original_weight = sched_op.parent_op.weights + encoded_npu_weight = op_info.npu_weights_tensor + + # Save UUIDs of original_weight so only unique instances of tensors are used to calculate weights + if original_weight and (original_weight.equivalence_id not in original_weight_uuids): + + original_weight_uuids.add(original_weight.equivalence_id) + total_weight_size += original_weight.values.itemsize * original_weight.values.size + + # Save UUIDs of encoded_npu_weight so only unique instances of tensors are used to calculate weights + if encoded_npu_weight and (encoded_npu_weight.equivalence_id not in encoded_npu_weight_uuids): + + encoded_npu_weight_uuids.add(encoded_npu_weight) + total_encoded_weight_size += len(encoded_npu_weight.buffer) + total_bws += bws total_macs += macs total_cycles += cycles @@ -743,3 +770,5 @@ def calc_new_performance_for_network(nng, arch): nng.bandwidths = total_bws nng.macs = total_macs nng.cycles = total_cycles + nng.total_original_weights = total_weight_size + nng.total_npu_encoded_weights = total_encoded_weight_size |