diff options
| author | Patrik Gustavsson <patrik.gustavsson@arm.com> | 2020-11-18 07:59:09 +0100 |
|---|---|---|
| committer | Patrik Gustavsson <patrik.gustavsson@arm.com> | 2020-11-18 10:16:20 +0100 |
| commit | b0ca274e904ca9080f2baa22fdb538c58b20aab2 (patch) | |
| tree | a19dd0aa7f1e7c60a224c8d4ff6aae749712e96a | |
| parent | 837dc1bc42323fa723b72fe51919bc2f013e5a26 (diff) | |
| download | ethos-u-vela-b0ca274e904ca9080f2baa22fdb538c58b20aab2.tar.gz | |
MLBEDSW-3494 Fix rounding of fused Quantized
Fix for setting rounding to TFL for fused Quantized
Signed-off-by: Patrik Gustavsson <patrik.gustavsson@arm.com>
Change-Id: Ic203f95f8916e330bcbf5792b52661b6f3e99bfc
| -rw-r--r-- | ethosu/vela/high_level_command_to_npu_op.py | 13 |
1 files changed, 9 insertions, 4 deletions
diff --git a/ethosu/vela/high_level_command_to_npu_op.py b/ethosu/vela/high_level_command_to_npu_op.py index 812e8e9..61ce1c9 100644 --- a/ethosu/vela/high_level_command_to_npu_op.py +++ b/ethosu/vela/high_level_command_to_npu_op.py @@ -130,7 +130,7 @@ def ifm_ifm2_correct_order(ifm_shape: List[int], ifm2_shape: List[int]) -> bool: return True -def get_rounding_mode(op: Operation) -> NpuRoundingMode: +def get_rounding_mode(op: Operation, fused_quantize: bool) -> NpuRoundingMode: """Specifies type of rounding to be used""" rounding_mode = NpuRoundingMode.TFL if op.type == Op.ResizeBilinear: @@ -140,7 +140,12 @@ def get_rounding_mode(op: Operation) -> NpuRoundingMode: and op.ifm.dtype == DataType.int16 ): rounding_mode = NpuRoundingMode.NATURAL - elif op.type.is_avgpool_op() and op.memory_function == Op.ConcatSliceWrite and op.kernel.elements_wh() == 1: + elif ( + not fused_quantize + and op.type.is_avgpool_op() + and op.memory_function == Op.ConcatSliceWrite + and op.kernel.elements_wh() == 1 + ): rounding_mode = NpuRoundingMode.NATURAL rounding_mode = op.attrs.get("rounding_mode", rounding_mode) return rounding_mode @@ -353,14 +358,14 @@ def set_common_op_fields(npu_op: NpuBlockOperation, cmd: NpuStripe, arch: Archit if cmd.scale_tensor is not None: npu_op.biases = create_biases(cmd.weight_tensor, cmd.scale_tensor, cmd.weight_box, arch) npu_op.activation = create_npu_activation(op) - npu_op.rounding_mode = get_rounding_mode(op) + npu_op.fused_quantize = any(op.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]) if not op.type.is_elementwise_op(): npu_op.padding = create_padding(cmd, op) npu_op.kernel = to_npu_kernel(op.kernel) npu_op.ifm_upscale = get_upscale(op) - npu_op.fused_quantize = any(op.type == Op.Quantize for op in ps.ops) return npu_op |