diff options
author | Louis Verhaard <louis.verhaard@arm.com> | 2021-02-09 16:08:26 +0100 |
---|---|---|
committer | Louis Verhaard <louis.verhaard@arm.com> | 2021-02-17 16:29:20 +0100 |
commit | 1a92f78e14f31f1423824228deb0628b7a9a9071 (patch) | |
tree | 4e0de9a2e6b5b7d9159b25cbfead4a625c134a3c /ethosu/vela/softmax.py | |
parent | 8d0f4890aa0ceae92a33ebb789701ff644a6fcaa (diff) | |
download | ethos-u-vela-1a92f78e14f31f1423824228deb0628b7a9a9071.tar.gz |
MLBEDSW-4022: support PAD followed by pool operator
PAD followed by max/average pool is run on NPU if NPU
padding can be used. Average pool is converted to depthwise.
Change-Id: Icc3652e6d9ecff5ac3dc7d92080313d90c245404
Signed-off-by: Louis Verhaard <louis.verhaard@arm.com>
Diffstat (limited to 'ethosu/vela/softmax.py')
-rw-r--r-- | ethosu/vela/softmax.py | 12 |
1 files changed, 5 insertions, 7 deletions
diff --git a/ethosu/vela/softmax.py b/ethosu/vela/softmax.py index 4418f011..520ec23e 100644 --- a/ethosu/vela/softmax.py +++ b/ethosu/vela/softmax.py @@ -1,4 +1,4 @@ -# Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved. +# Copyright (C) 2020-2021 Arm Limited or its affiliates. All rights reserved. # # Copyright 2017 The TensorFlow Authors. All Rights Reserved. # @@ -287,11 +287,9 @@ class SoftMax: shift = create_const_tensor( f"{name}_const", [1, 1, 1, 1], DataType.int32, [12], np.int32, quantization=no_scale_quant ) - rescaled_exp = add_op_get_ofm( - create_shr( - name, ifm_exp, shift, no_scale_quant, activation, attrs={"rounding_mode": NpuRoundingMode.NATURAL}, - ) - ) + shr_op = create_shr(name, ifm_exp, shift, no_scale_quant, activation) + shr_op.rounding_mode = NpuRoundingMode.NATURAL + rescaled_exp = add_op_get_ofm(shr_op) # PASS 3 - Reduce sum sum_of_exp = add_op_get_ofm( @@ -421,7 +419,7 @@ class SoftMax: # PASS 30 - SHR shr30_op = Operation(Op.SHR, f"{self.op.name}_shr{pass_number}") - shr30_op.attrs["rounding_mode"] = NpuRoundingMode.NATURAL + shr30_op.rounding_mode = NpuRoundingMode.NATURAL shr30_op.add_input_tensor(scaled_exp) shr30_op.add_input_tensor(right_shift) shr30_op.set_output_tensor(ofm) |