diff options
| author | Fredrik Svedberg <fredrik.svedberg@arm.com> | 2020-06-03 15:43:31 +0200 |
|---|---|---|
| committer | Louis Verhaard <louis.verhaard@arm.com> | 2020-08-05 16:26:04 +0200 |
| commit | a0c3624899edc601525a589643c802469003f89d (patch) | |
| tree | 4fc52db04cd29901b3e5d4a7425a7a641e9647fb /ethosu/vela/pass_packing.py | |
| parent | 9a03fdff316662be69a1adc4e391e43bc6519b08 (diff) | |
| download | ethos-u-vela-a0c3624899edc601525a589643c802469003f89d.tar.gz | |
[MLBEDSW-2335] SoftMax int16
Added graph rewrite of Softmax for int16.
Change-Id: Id7885af6056a23e8b8362fb61ae94283251eb398
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>
Diffstat (limited to 'ethosu/vela/pass_packing.py')
| -rw-r--r-- | ethosu/vela/pass_packing.py | 15 |
1 files changed, 9 insertions, 6 deletions
diff --git a/ethosu/vela/pass_packing.py b/ethosu/vela/pass_packing.py index c14a70be..8fb95f03 100644 --- a/ethosu/vela/pass_packing.py +++ b/ethosu/vela/pass_packing.py @@ -66,6 +66,7 @@ mac_main_ops = set( "MaxPool", "AvgPoolAct", "MaxPoolAct", + "ReduceSum", # deconvolution "ResizeBilinear", ) @@ -85,10 +86,12 @@ binary_elem_wise_main_ops = set( "Sub", "Minimum", "Maximum", + "SHL", + "SHR", ) ) -unary_elem_wise_main_ops = set(("LeakyRelu", "Abs")) # Unary element-wise operations +unary_elem_wise_main_ops = set(("LeakyRelu", "Abs", "CLZ",)) # Unary element-wise operations elem_wise_main_ops = binary_elem_wise_main_ops | unary_elem_wise_main_ops @@ -417,13 +420,12 @@ def pack_into_passes(nng, arch, verbose_packing=False): # Swap broadcast input if applicable broadcast_input_check(ps) + # If only 1 input, IFM and IFM2 will be the same tensor ps.ifm_tensor = ps.inputs[0] + ps.ifm2_tensor = ps.inputs[-1] - if len(ps.inputs) == 1: - # Only 1 input, IFM and IFM2 are the same tensor - ps.ifm2_tensor = ps.inputs[0] - else: - ps.ifm2_tensor = ps.inputs[1] + if len(ps.inputs) > 2: + ps.ifm_tensor = ps.inputs[-2] else: ps.ifm_tensor = ifm_tensor ps.ifm2_tensor = None @@ -432,6 +434,7 @@ def pack_into_passes(nng, arch, verbose_packing=False): 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] + ps.lut_tensor = ps.get_primary_op_lut() for op in ps.ops: op.scheduled_pass = ps |