diff options
| author | Patrik Gustavsson <patrik.gustavsson@arm.com> | 2020-12-01 16:02:29 +0100 |
|---|---|---|
| committer | Patrik Gustavsson <patrik.gustavsson@arm.com> | 2020-12-18 16:33:32 +0100 |
| commit | 2349d429d926e258e9a61d34c7fd97660ab9fb98 (patch) | |
| tree | b5151d0f12428e47d64b1fb2ce4f2f8c19304a0d /ethosu/vela/pass_packing.py | |
| parent | 528a56df829b65f7a2c61953650b123c461095f7 (diff) | |
| download | ethos-u-vela-2349d429d926e258e9a61d34c7fd97660ab9fb98.tar.gz | |
MLBEDSW-3654 Add/use op ifm/ofm shapes
Add ifm/ofm shapes to op
Changed to rely on these shapes
Signed-off-by: Patrik Gustavsson <patrik.gustavsson@arm.com>
Change-Id: I571535a1dcadc2bdb04a3c727a8e1c49703b174d
Diffstat (limited to 'ethosu/vela/pass_packing.py')
| -rw-r--r-- | ethosu/vela/pass_packing.py | 19 |
1 files changed, 19 insertions, 0 deletions
diff --git a/ethosu/vela/pass_packing.py b/ethosu/vela/pass_packing.py index 9bc04f29..095a78d4 100644 --- a/ethosu/vela/pass_packing.py +++ b/ethosu/vela/pass_packing.py @@ -397,11 +397,28 @@ def pack_into_passes(nng, arch, verbose_packing=False): if len(ps.inputs) > 2: ps.ifm_tensor = ps.inputs[-2] + + # Get the corresponding ifm_shapes + for op in input_ops_list + [primary_op]: + if ps.ifm_tensor == op.ifm: + ps.ifm_shapes.append(op.ifm_shapes[0]) + elif ps.ifm_tensor == op.ifm2: + ps.ifm_shapes.append(op.ifm_shapes[1]) + for op in input_ops_list + [primary_op]: + if ps.ifm2_tensor == op.ifm: + ps.ifm_shapes.append(op.ifm_shapes[0]) + elif ps.ifm2_tensor == op.ifm2: + ps.ifm_shapes.append(op.ifm_shapes[1]) else: ps.ifm_tensor = ifm_tensor ps.ifm2_tensor = None + if ps.primary_op is not None: + ps.ifm_shapes.append(ps.primary_op.ifm_shapes[0]) ps.ofm_tensor = ofm_tensor + if ps.primary_op is not None: + ps.ofm_shapes.append(ps.primary_op.ofm_shapes[0]) + assert ps.placement != PassPlacement.Npu or ps.ofm_tensor is not None ps.weight_tensor = ps.get_primary_op_ifm_weights()[1] ps.scale_tensor = ps.get_primary_op_ifm_weights_biases_ofm()[2] @@ -436,6 +453,8 @@ def pack_into_passes(nng, arch, verbose_packing=False): avgpool_out = inp.clone("_avgpooled") avgpool_out.consumer_list.append(op) avgpool_op.set_output_tensor(avgpool_out) + avgpool_op.ifm_shapes = op.ifm_shapes + avgpool_op.ofm_shapes = op.ofm_shapes op.inputs[0] = avgpool_out op_list.insert(0, avgpool_op) |