diff options
| author | Patrik Gustavsson <patrik.gustavsson@arm.com> | 2020-08-13 13:41:05 +0200 |
|---|---|---|
| committer | Fredrik Knutsson <fredrik.knutsson.hunnebo@gmail.com> | 2020-08-14 10:49:15 +0000 |
| commit | 458a208c44f70a9848f1e8e2e91f28ce3641c48f (patch) | |
| tree | 37f23561f75d61746383dafc987b411646baaed8 /ethosu/vela/scheduler.py | |
| parent | be733cf04bb262d4eee791d76f01cecd64ff9255 (diff) | |
| download | ethos-u-vela-458a208c44f70a9848f1e8e2e91f28ce3641c48f.tar.gz | |
MLBEDSW-2570 Avoid usage of NHCWB16 for some cases
Avoid usage of NHCWB16 when Stack/Pack/Concat is performed in axis 3,
and the "concat start" of each slice to be combined is not a multiple
of 16.
Signed-off-by: Patrik Gustavsson <patrik.gustavsson@arm.com>
Change-Id: If3f7b4a3424be3c86fc2dc48e8649ce4c4f49485
Diffstat (limited to 'ethosu/vela/scheduler.py')
| -rw-r--r-- | ethosu/vela/scheduler.py | 33 |
1 files changed, 19 insertions, 14 deletions
diff --git a/ethosu/vela/scheduler.py b/ethosu/vela/scheduler.py index cc9278fd..f3b3a79c 100644 --- a/ethosu/vela/scheduler.py +++ b/ethosu/vela/scheduler.py @@ -670,14 +670,16 @@ class DynamicProgrammingScheduler: for pred_candidate in ps.dag_predecessors: if len(pred_candidate.outputs) == 1 and pred_candidate.outputs[0] == ifm_tensor: # we found a predecessor that produces this IFM tensor - if len(pred_candidate.successors) == 1 and pred_candidate.successors[0] == ps: - # and it only has one successor, namely us - if pred_candidate.placement == PassPlacement.Npu: - if pred_candidate.npu_block_type in self.ifm_stream_npu_blocks: - # and it is on the Npu - if not self.avoid_for_spilling(pred_candidate): - # and fusable - it's a candidate - pred_pass_list.append(pred_candidate) + if not ifm_tensor.avoid_NHCWB16: + # and NHCWB16 format is not to be avoided + if len(pred_candidate.successors) == 1 and pred_candidate.successors[0] == ps: + # and it only has one successor, namely us + if pred_candidate.placement == PassPlacement.Npu: + if pred_candidate.npu_block_type in self.ifm_stream_npu_blocks: + # and it is on the Npu + if not self.avoid_for_spilling(pred_candidate): + # and fusable - it's a candidate + pred_pass_list.append(pred_candidate) if not pred_pass_list: return ABORT_SEARCH @@ -953,12 +955,15 @@ class DynamicProgrammingScheduler: if output.purpose != TensorPurpose.FeatureMap: continue - use_NHCWB16 = True - for op in output.consumer_list: - if op is None or op.type == "Reshape": - use_NHCWB16 = False - else: - use_NHCWB16 &= op.run_on_npu + use_NHCWB16 = not output.avoid_NHCWB16 + + if use_NHCWB16: + # Check consumers, to see if NHCWB16 can be used in the output + for op in output.consumer_list: + if op is None or op.type == "Reshape": + use_NHCWB16 = False + else: + use_NHCWB16 &= op.run_on_npu if use_NHCWB16: output.set_format(TensorFormat.NHCWB16, arch) |