diff options
| author | Johan Alfven <johan.alfven@arm.com> | 2024-01-25 12:32:13 +0100 |
|---|---|---|
| committer | Johan Alfven <johan.alfven@arm.com> | 2024-01-30 12:25:05 +0000 |
| commit | 014bc28324014dc82e3b352d7a81df9ed0338c56 (patch) | |
| tree | b0b5e765319531da441874e03c4c81980bd6f800 | |
| parent | cbec599c97b8620580ed5f80ea253770df872fa1 (diff) | |
| download | ethos-u-vela-014bc28324014dc82e3b352d7a81df9ed0338c56.tar.gz | |
MLBEDSW-8569: MLCE: Reported number of CPU ops are wrong
- A Pack op is implemented by several AvgPool ops. Depending
on number of CPU ops and graph topology this could result in that
the AvgPool ops ended up in different nodes. One of these node
had the Pack output referenced to it but the other node did not.
As a result the full graph was not traversed when calculating CPU
ops.
- The compiled network works as intended but the number of
reported CPU was wrong.
- Added new method that extracts the ops using the passes in
the sub graphs which fix the problem.
Change-Id: Ie88ebd4669783559258ae763737a4c7f86c905f8
Signed-off-by: Johan Alfven <johan.alfven@arm.com>
| -rw-r--r-- | ethosu/vela/nn_graph.py | 17 | ||||
| -rw-r--r-- | ethosu/vela/stats_writer.py | 6 |
2 files changed, 19 insertions, 4 deletions
diff --git a/ethosu/vela/nn_graph.py b/ethosu/vela/nn_graph.py index 50266d5..92c7e1b 100644 --- a/ethosu/vela/nn_graph.py +++ b/ethosu/vela/nn_graph.py @@ -1,4 +1,4 @@ -# SPDX-FileCopyrightText: Copyright 2020-2023 Arm Limited and/or its affiliates <open-source-office@arm.com> +# SPDX-FileCopyrightText: Copyright 2020-2024 Arm Limited and/or its affiliates <open-source-office@arm.com> # # SPDX-License-Identifier: Apache-2.0 # @@ -318,6 +318,7 @@ class Subgraph: ps.outputs = [out_tens for out_tens in ps.outputs if len(out_tens.consumers()) > 0] ps.ops = [op for op in ps.ops if op.outputs[0] in ps.outputs] + # get_all_ops is used when traversing the original graph def get_all_ops(self): all_ops = [] visit_op_set = set() @@ -344,6 +345,20 @@ class Subgraph: return all_ops + # get_all_ops_from_passes is used by stats writer to calculate the number of + # CPU and NPU ops + # Due to a side effect get_all_ops might not be traversing the full graph + # after extract_npu_subgraph have been called and should not be used by stats writer. + # The reason is that the main graph might have NPU nodes with no visible outputs + # and therefore the nodes will be missed. + def get_all_ops_from_passes(self): + all_ops = [] + for idx, ps in enumerate(self.passes): + for op in ps.ops: + all_ops.append(op) + + return all_ops + def print_operators(self, ignore_placeholder_const=True, show_attributes=True): print(f"Operators of Subgraph {self.name}") diff --git a/ethosu/vela/stats_writer.py b/ethosu/vela/stats_writer.py index 25c9030..b743a5f 100644 --- a/ethosu/vela/stats_writer.py +++ b/ethosu/vela/stats_writer.py @@ -1,4 +1,4 @@ -# SPDX-FileCopyrightText: Copyright 2020-2022 Arm Limited and/or its affiliates <open-source-office@arm.com> +# SPDX-FileCopyrightText: Copyright 2020-2022, 2024 Arm Limited and/or its affiliates <open-source-office@arm.com> # # SPDX-License-Identifier: Apache-2.0 # @@ -365,11 +365,11 @@ def print_performance_metrics(nng, arch, show_cpu_operations=False, verbose_weig for sg in nng.subgraphs: if sg.placement == PassPlacement.Cpu: - for op in sg.get_all_ops(): + for op in sg.get_all_ops_from_passes(): if op.type not in ir_only_ops: cpu_operations.append(op) elif sg.placement == PassPlacement.Npu: - for op in sg.get_all_ops(): + for op in sg.get_all_ops_from_passes(): if op.type not in ir_only_ops: npu_operations.append(op) |