diff options
author | Johan Gunnarsson <johan.gunnarsson@arm.com> | 2023-08-10 13:10:44 +0200 |
---|---|---|
committer | Johan Gunnarsson <johan.gunnarsson@arm.com> | 2023-08-29 16:54:41 +0200 |
commit | 985563791a811e1ea3b5137f97e5a5fc4dafd4b1 (patch) | |
tree | cd1d9a41c9e194e9fcd1fe9ee090d8ef07a640a9 /ethosu/vela/high_level_command_to_npu_op.py | |
parent | c02eaa3e25840aee4ff909df263d4d0673227c5d (diff) | |
download | ethos-u-vela-985563791a811e1ea3b5137f97e5a5fc4dafd4b1.tar.gz |
MLBEDSW-7881: Convert Quantize op to Avgpool op in graph optimiser
This convert is already done in the pass packing stage, but doing it
in the graph optimiser stage is better.
Change-Id: Ib9baa98d115cf88491ce39936972a93467a378ce
Signed-off-by: Johan Gunnarsson <johan.gunnarsson@arm.com>
Diffstat (limited to 'ethosu/vela/high_level_command_to_npu_op.py')
-rw-r--r-- | ethosu/vela/high_level_command_to_npu_op.py | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/ethosu/vela/high_level_command_to_npu_op.py b/ethosu/vela/high_level_command_to_npu_op.py index 79ac3929..4384f2c1 100644 --- a/ethosu/vela/high_level_command_to_npu_op.py +++ b/ethosu/vela/high_level_command_to_npu_op.py @@ -143,7 +143,7 @@ def get_rounding_mode(op: Operation, fused_quantize: bool) -> NpuRoundingMode: if op.type.is_resize_op(): rounding_mode = NpuRoundingMode.NATURAL elif ( - op._original_type.npu_block_type in (NpuBlockType.ConvolutionMxN, NpuBlockType.ConvolutionDepthWise) + op.original_type.npu_block_type in (NpuBlockType.ConvolutionMxN, NpuBlockType.ConvolutionDepthWise) and op.ifm.dtype == DataType.int16 ): rounding_mode = NpuRoundingMode.NATURAL @@ -334,7 +334,7 @@ def use_zero_point_0(ps, tens: Tensor, is_ifm_tensor: bool) -> bool: return False if ps.primary_op.type == Op.AvgPool and ps.primary_op.explicit_scaling: return False - fused_quantize = any(op.type == Op.Quantize for op in ps.ops) + fused_quantize = any(op.type == Op.Quantize or op.original_type == Op.Quantize for op in ps.ops) forced_ofm_quantization = ps.primary_op.forced_output_quantization use_0 = ( ( @@ -521,7 +521,7 @@ def set_common_op_fields(npu_op: NpuBlockOperation, cmd: NpuStripe, arch: Archit if cmd.weight_tensor is not None: npu_op.weights, npu_op.biases = create_weights(cmd.weight_tensor, cmd.weight_box, cmd.scale_tensor, arch) npu_op.activation = create_npu_activation(op) - npu_op.fused_quantize = any(op.type == Op.Quantize for op in ps.ops) + npu_op.fused_quantize = any(op.type == Op.Quantize or op.original_type == Op.Quantize for op in ps.ops) npu_op.rounding_mode = get_rounding_mode(op, npu_op.fused_quantize) npu_op.block_config = NpuShape3D(height=ps.block_config[0], width=ps.block_config[1], depth=ps.block_config[3]) |