Skip to content

Arm backend: Add int32 support to aten.mul.Tensor with BI/INT profile #11964

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
zingo merged 2 commits into from
Jun 25, 2025
Merged

Arm backend: Add int32 support to aten.mul.Tensor with BI/INT profile #11964

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7048351
Arm backend: Add int32 support to aten.mul.Tensor with BI/INT profile
YufengShi-dudu Jun 19, 2025
f705749
Merge branch 'main' into add-int32-support-to-mul-node-visitor-bi-int
YufengShi-dudu Jun 25, 2025
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
158 changes: 96 additions & 62 deletions backends/arm/operators/op_mul.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,38 +50,50 @@ def define_node(
validate_num_inputs(self.target, inputs, 2)
validate_same_dtype(self.target, [*inputs, output], ts)
validate_valid_dtype(
self.target, [*inputs, output], ts.DType.INT8, output.tosa_spec
self.target,
[*inputs, output],
[ts.DType.INT8, ts.DType.INT32],
output.tosa_spec,
)

dim_order = (
inputs[0].dim_order
if len(inputs[0].shape) > len(inputs[1].shape)
else inputs[1].dim_order
)
input_A = inputs[0]
input_B = inputs[1]
input_qparams = get_input_qparams(node)
input_A_qargs = input_qparams[0]
input_B_qargs = input_qparams[1]
input_A.shape = tutils.tosa_shape(input_A.shape, input_A.dim_order)
input_B.shape = tutils.tosa_shape(input_B.shape, input_B.dim_order)

# Rescale inputs to INT32 with zp=0
input_A_rescaled = tqutils.build_rescale_to_int32(
tosa_graph,
input_A,
input_A_qargs.get_zp_per_tensor(),
1.0,
)
input_B_rescaled = tqutils.build_rescale_to_int32(
tosa_graph,
input_B,
input_B_qargs.get_zp_per_tensor(),
1.0,
)

output_shape = tutils.tosa_shape(output.shape, output.dim_order)
mul_output = tosa_graph.addIntermediate(output_shape, ts.DType.INT32)
if inputs[0].dtype == ts.DType.INT8:
input_A = inputs[0]
input_B = inputs[1]
input_qparams = get_input_qparams(node)
input_A_qargs = input_qparams[0]
input_B_qargs = input_qparams[1]
input_A.shape = tutils.tosa_shape(input_A.shape, input_A.dim_order)
input_B.shape = tutils.tosa_shape(input_B.shape, input_B.dim_order)

# Rescale inputs to INT32 with zp=0
input_A_rescaled = tqutils.build_rescale_to_int32(
tosa_graph,
input_A,
input_A_qargs.get_zp_per_tensor(),
1.0,
)
input_B_rescaled = tqutils.build_rescale_to_int32(
tosa_graph,
input_B,
input_B_qargs.get_zp_per_tensor(),
1.0,
)
else:
# input[0].dtype == ts.DType.INT32
# Non quantized input, natively support by TOSA.MUL
input_A_rescaled, input_B_rescaled = inputs[0], inputs[1]

if output.dtype == ts.DType.INT8:
output_shape = tutils.tosa_shape(output.shape, output.dim_order)
mul_output = tosa_graph.addIntermediate(output_shape, ts.DType.INT32)
else:
# output.dtype == ts.DType.INT32
mul_output = output

input1, input2 = tutils.reshape_for_broadcast(
tosa_graph,
Expand All @@ -101,10 +113,16 @@ def define_node(
[mul_output.name],
attr,
)
output_scale = (
input_A_qargs.get_scale_per_tensor() * input_B_qargs.get_scale_per_tensor()
)
tqutils.insert_rescale_op_to_int8(tosa_graph, mul_output, output_scale, node)

if output.dtype == ts.DType.INT8:
# Scale output back to 8 bit
output_scale = (
input_A_qargs.get_scale_per_tensor() # type: ignore[possibly-undefined]
* input_B_qargs.get_scale_per_tensor() # type: ignore[possibly-undefined]
)
tqutils.insert_rescale_op_to_int8(
tosa_graph, mul_output, output_scale, node
)


@register_node_visitor
Expand Down Expand Up @@ -161,35 +179,47 @@ def define_node(
validate_num_inputs(self.target, inputs, 2)
validate_same_dtype(self.target, [*inputs, output], ts)
validate_valid_dtype(
self.target, [*inputs, output], ts.DType.INT8, output.tosa_spec
)

input_A = inputs[0]
input_B = inputs[1]
input_qparams = get_input_qparams(node)
input_A_qargs = input_qparams[0]
input_B_qargs = input_qparams[1]
input_A.shape = tutils.tosa_shape(input_A.shape, input_A.dim_order)
input_B.shape = tutils.tosa_shape(input_B.shape, input_B.dim_order)

# Rescale inputs to INT32 with zp=0
input_A_rescaled = tqutils.build_rescale_to_int32(
tosa_graph,
input_A,
input_A_qargs.get_zp_per_tensor(),
1.0,
tosa_spec=self.tosa_spec,
)
input_B_rescaled = tqutils.build_rescale_to_int32(
tosa_graph,
input_B,
input_B_qargs.get_zp_per_tensor(),
1.0,
tosa_spec=self.tosa_spec,
self.target,
[*inputs, output],
[ts.DType.INT8, ts.DType.INT32],
output.tosa_spec,
)

output_shape = tutils.tosa_shape(output.shape, output.dim_order)
mul_output = tosa_graph.addIntermediate(output_shape, ts.DType.INT32)
if inputs[0].dtype == ts.DType.INT8:
input_A = inputs[0]
input_B = inputs[1]
input_qparams = get_input_qparams(node)
input_A_qargs = input_qparams[0]
input_B_qargs = input_qparams[1]
input_A.shape = tutils.tosa_shape(input_A.shape, input_A.dim_order)
input_B.shape = tutils.tosa_shape(input_B.shape, input_B.dim_order)

# Rescale inputs to INT32 with zp=0
input_A_rescaled = tqutils.build_rescale_to_int32(
tosa_graph,
input_A,
input_A_qargs.get_zp_per_tensor(),
1.0,
tosa_spec=self.tosa_spec,
)
input_B_rescaled = tqutils.build_rescale_to_int32(
tosa_graph,
input_B,
input_B_qargs.get_zp_per_tensor(),
1.0,
tosa_spec=self.tosa_spec,
)
else:
# input[0].dtype == ts.DType.INT32
# Non quantized input, natively support by TOSA.MUL
input_A_rescaled, input_B_rescaled = inputs[0], inputs[1]

if output.dtype == ts.DType.INT8:
output_shape = tutils.tosa_shape(output.shape, output.dim_order)
mul_output = tosa_graph.addIntermediate(output_shape, ts.DType.INT32)
else:
# output.dtype == ts.DType.INT32
mul_output = output

# Do the INT32 Mul
tosa_graph.addConst([1], ts.DType.INT8, 0, name=f"{node.name}_shift")
Expand All @@ -198,12 +228,16 @@ def define_node(
[input_A_rescaled.name, input_B_rescaled.name, f"{node.name}_shift"],
[mul_output.name],
)
output_scale = (
input_A_qargs.get_scale_per_tensor() * input_B_qargs.get_scale_per_tensor()
)
tqutils.insert_rescale_op_to_int8(
tosa_graph, mul_output, output_scale, node, self.tosa_spec
)

if output.dtype == ts.DType.INT8:
# Scale output back to 8 bit
output_scale = (
input_A_qargs.get_scale_per_tensor() # type: ignore[possibly-undefined]
* input_B_qargs.get_scale_per_tensor() # type: ignore[possibly-undefined]
)
tqutils.insert_rescale_op_to_int8(
tosa_graph, mul_output, output_scale, node, self.tosa_spec
)


@register_node_visitor
Expand Down
84 changes: 84 additions & 0 deletions backends/arm/test/ops/test_mul.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,23 @@
}


test_data_suite_int32 = {
# (test_name, input, other,) See torch.mul() for info
"op_mul_rank4_randn_int32": lambda: (
torch.randint(0, 10, (1, 10, 25, 20), dtype=torch.int32),
torch.randint(0, 10, (1, 10, 25, 20), dtype=torch.int32),
),
"op_mul_rank4_randn_mutltiple_broadcasts_int32": lambda: (
torch.randint(0, 10, (1, 4, 4, 1), dtype=torch.int32),
torch.randint(0, 10, (1, 1, 4, 4), dtype=torch.int32),
),
"op_mul_rank4_randn_broadcast_int32": lambda: (
torch.randint(0, 10, (1, 10, 25, 20), dtype=torch.int32),
torch.randint(0, 10, (1, 25, 20), dtype=torch.int32),
),
}


class Mul(torch.nn.Module):

def forward(
Expand Down Expand Up @@ -111,6 +128,17 @@ def test_mul_tensor_tosa_MI_diff_input_ranks(test_data: torch.Tensor):
pipeline.run()


@common.parametrize("test_data", test_data_suite_int32)
def test_mul_tensor_tosa_MI_int32(test_data: torch.Tensor):
pipeline = TosaPipelineMI[input_t1](
Mul(),
test_data(),
aten_op,
exir_op=[],
)
pipeline.run()


@common.parametrize("test_data", test_data_suite_2)
def test_mul_tensor_tosa_BI_diff_input_ranks(test_data: torch.Tensor):
pipeline = TosaPipelineBI[input_t1](
Expand All @@ -133,6 +161,18 @@ def test_mul_tensor_tosa_BI(test_data: torch.Tensor):
pipeline.run()


@common.parametrize("test_data", test_data_suite_int32)
def test_mul_tensor_tosa_BI_int32(test_data: torch.Tensor):
pipeline = TosaPipelineBI[input_t1](
Mul(),
test_data(),
aten_op,
exir_op=[],
)
pipeline.pop_stage("check.quant_nodes")
pipeline.run()


@common.parametrize("test_data", test_data_suite)
@common.XfailIfNoCorstone300
def test_mul_tensor_u55_BI(test_data: torch.Tensor):
Expand All @@ -157,3 +197,47 @@ def test_mul_tensor_u85_BI(test_data: torch.Tensor):
run_on_fvp=True,
)
pipeline.run()


@common.parametrize(
"test_data",
test_data_suite_int32,
xfails={
# TODO: MLETORCH-1132 Investigate why tests with inputs that require broadcasting fail on u55/u85
"op_mul_rank4_randn_mutltiple_broadcasts_int32": "RuntimeError: mean(): could not infer output dtype. Input dtype must be either a floating point or complex dtype. Got: Int",
"op_mul_rank4_randn_broadcast_int32": "RuntimeError: mean(): could not infer output dtype. Input dtype must be either a floating point or complex dtype. Got: Int",
},
)
@common.XfailIfNoCorstone300
def test_mul_tensor_u55_BI_int32(test_data: torch.Tensor):
pipeline = EthosU55PipelineBI[input_t1](
Mul(),
test_data(),
aten_op,
exir_ops=[],
run_on_fvp=True,
)
pipeline.pop_stage("check.quant_nodes")
pipeline.run()


@common.parametrize(
"test_data",
test_data_suite_int32,
xfails={
# TODO: MLETORCH-1132 Investigate why tests with inputs that require broadcasting fail on u55/u85
"op_mul_rank4_randn_mutltiple_broadcasts_int32": "RuntimeError: mean(): could not infer output dtype. Input dtype must be either a floating point or complex dtype. Got: Int",
"op_mul_rank4_randn_broadcast_int32": "RuntimeError: mean(): could not infer output dtype. Input dtype must be either a floating point or complex dtype. Got: Int",
},
)
@common.XfailIfNoCorstone320
def test_mul_tensor_u85_BI_int32(test_data: torch.Tensor):
pipeline = EthosU85PipelineBI[input_t1](
Mul(),
test_data(),
aten_op,
exir_ops=[],
run_on_fvp=True,
)
pipeline.pop_stage("check.quant_nodes")
pipeline.run()
Loading