Skip to content

Use single rounding as default for TOSA lowering #4591

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Aug 14, 2024
+14 −33
Merged

Use single rounding as default for TOSA lowering #4591

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
37 changes: 11 additions & 26 deletions backends/arm/operators/op_addmm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,10 +12,7 @@
register_node_visitor,
)
from executorch.backends.arm.tosa_mapping import TosaArg
from executorch.backends.arm.tosa_quant_utils import (
compute_multiplier_and_shift,
get_quant_node_args,
)
from executorch.backends.arm.tosa_quant_utils import build_rescale, get_quant_node_args

from executorch.backends.arm.tosa_utils import build_reshape
from executorch.exir.dialects._ops import ops as exir_ops
Expand Down Expand Up @@ -128,32 +125,20 @@ def define_node(
weight_scale = get_quant_node_args(weight_node_q_node).scale

output_rescale_scale = (input_scale * weight_scale) / consumer_node_scale
(
multiplier_output,
shift_output,
) = compute_multiplier_and_shift(output_rescale_scale)

attr_rescale_output = ts.TosaSerializerAttribute()
attr_rescale_output.RescaleAttribute(
input_zp=0,
output_zp=consumer_node_node_zp,
multiplier=[multiplier_output],
shift=[shift_output],
scale32=True,
double_round=True,
per_channel=False,
input_unsigned=False,
output_unsigned=False,
)

reshaped_res = tosa_graph.addIntermediate(result_shape, ts.DType.INT32)
build_reshape(tosa_graph, conv2d_res.name, result_shape, reshaped_res.name)

tosa_graph.addOperator(
TosaOp.Op().RESCALE,
[reshaped_res.name],
[output.name],
attr_rescale_output,
build_rescale(
tosa_fb=tosa_graph,
scale=output_rescale_scale,
input_node=reshaped_res,
output_name=output.name,
output_type=ts.DType.INT8,
output_shape=reshaped_res.shape,
input_zp=0,
output_zp=consumer_node_node_zp,
is_double_round=False,
)

else:
Expand Down
10 changes: 3 additions & 7 deletions backends/arm/tosa_quant_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,7 +171,7 @@ def build_rescale(
output_shape,
input_zp,
output_zp,
is_double_round,
is_double_round=False,
):
scale_width = 32 if is_scale32(output_type) else 16
multiplier, shift = compute_multiplier_and_shift(scale, scale_width)
Expand All @@ -197,7 +197,7 @@ def build_rescale(


def build_rescale_to_int32(
tosa_fb, input, input_zp, rescale_scale, is_scale32=True, is_double_round=True
tosa_fb, input, input_zp, rescale_scale, is_scale32=True, is_double_round=False
) -> TosaSerializerTensor:
multiplier, shift = compute_multiplier_and_shift(rescale_scale)
attr_rescale = ts.TosaSerializerAttribute()
Expand Down Expand Up @@ -230,7 +230,7 @@ def build_rescale_from_int32(
output_zp,
rescale_scale,
is_scale32=True,
is_double_round=True,
is_double_round=False,
) -> TosaSerializerTensor:
multiplier, shift = compute_multiplier_and_shift(rescale_scale)
attr_rescale_output = ts.TosaSerializerAttribute()
Expand Down Expand Up @@ -329,9 +329,6 @@ def build_rescale_conv_output(
output_scale,
output_zp,
):
# Only use double round if we are doing 32 bit scaling
double_round = is_scale32(output_type)

# TODO add check to verify if this is a Per-channel quantization.
post_conv2d_scale = (input_scale.number * weight_scale.number) / output_scale.number

Expand All @@ -345,6 +342,5 @@ def build_rescale_conv_output(
op.shape,
0,
output_zp.number,
double_round,
)
return
Loading