This is an automated email from the ASF dual-hosted git repository.
mshr pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/main by this push:
new b3b6024027 [Relax] Add FRelaxInferLayout for gather_elements operator
(#18642)
b3b6024027 is described below
commit b3b6024027c9b83471880dfb7af892c618274131
Author: Guan-Ming (Wesley) Chiu <[email protected]>
AuthorDate: Wed Jan 7 19:33:01 2026 +0800
[Relax] Add FRelaxInferLayout for gather_elements operator (#18642)
## Why
The gather_elements operator lacked layout inference support, preventing
it from participating in layout transformations during the ConvertLayout
pass.
## How
- Add InferLayoutGatherElements function that transforms the axis
attribute according to the input layout
- Register FRelaxInferLayout attribute
---------
Co-authored-by: gemini-code-assist[bot]
<176961590+gemini-code-assist[bot]@users.noreply.github.com>
---
src/relax/op/tensor/manipulate.cc | 34 ++++++++++++++
.../python/relax/test_transform_convert_layout.py | 53 ++++++++++++++++++++++
2 files changed, 87 insertions(+)
diff --git a/src/relax/op/tensor/manipulate.cc
b/src/relax/op/tensor/manipulate.cc
index 3170b28eeb..afb749a297 100644
--- a/src/relax/op/tensor/manipulate.cc
+++ b/src/relax/op/tensor/manipulate.cc
@@ -2150,12 +2150,46 @@ StructInfo InferStructInfoGatherElements(const Call&
call, const BlockBuilder& c
return TensorStructInfo(data_sinfo->dtype, indices_sinfo->ndim,
data_sinfo->vdevice);
}
+InferLayoutOutput InferLayoutGatherElements(
+ const Call& call, const ffi::Map<ffi::String, ffi::Array<ffi::String>>&
desired_layouts,
+ const VarLayoutMap& var_layout_map) {
+ ICHECK(NoDesiredLayout(call, desired_layouts));
+ const auto* attrs = call->attrs.as<GatherElementsAttrs>();
+ ICHECK(attrs) << "Invalid Call";
+
+ LayoutDecision data_layout = GetLayoutDecision(var_layout_map,
call->args[0]);
+ LayoutDecision indices_layout = GetLayoutDecision(var_layout_map,
call->args[1]);
+
+ LayoutDecision layout = data_layout;
+ // If data_layout is initial and indices_layout is not, prefer
indices_layout.
+ bool data_is_initial =
+ data_layout->layout.name() ==
InitialLayout(data_layout->layout.ndim()).name();
+ bool indices_is_initial =
+ indices_layout->layout.name() ==
InitialLayout(indices_layout->layout.ndim()).name();
+ if (data_is_initial && !indices_is_initial) {
+ layout = indices_layout;
+ }
+
+ if (layout->layout.ndim() != layout->layout.ndim_primal()) {
+ const auto* tensor_sinfo =
GetStructInfoAs<TensorStructInfoNode>(call->args[0]);
+ ICHECK(tensor_sinfo != nullptr) << "Invalid Call";
+ ICHECK(!tensor_sinfo->IsUnknownNdim()) << "Only support static ndim for
now";
+ int ndim = tensor_sinfo->ndim;
+ layout = LayoutDecision(InitialLayout(ndim));
+ }
+
+ ObjectPtr<GatherElementsAttrs> new_attrs =
ffi::make_object<GatherElementsAttrs>(*attrs);
+ new_attrs->axis = FindAxis(layout->layout, attrs->axis->value);
+ return InferLayoutOutput({layout, layout}, {layout}, Attrs(new_attrs));
+}
+
TVM_REGISTER_OP("relax.gather_elements")
.set_attrs_type<GatherElementsAttrs>()
.set_num_inputs(2)
.add_argument("data", "Tensor", "The input tensor.")
.add_argument("indices", "Tensor", "The indices tensor.")
.set_attr<FInferStructInfo>("FInferStructInfo",
InferStructInfoGatherElements)
+ .set_attr<FRelaxInferLayout>("FRelaxInferLayout",
InferLayoutGatherElements)
.set_attr<Bool>("FPurity", Bool(true));
/* relax.gather_nd */
diff --git a/tests/python/relax/test_transform_convert_layout.py
b/tests/python/relax/test_transform_convert_layout.py
index 221d680ebc..84fa9e70c7 100644
--- a/tests/python/relax/test_transform_convert_layout.py
+++ b/tests/python/relax/test_transform_convert_layout.py
@@ -5434,5 +5434,58 @@ def test_conv2d_scatter_nd():
verify(Input, Expected)
+def test_conv2d_gather_elements():
+ @I.ir_module
+ class Input:
+ @R.function
+ def main(
+ x: R.Tensor((2, 3, 28, 28), "float32"),
+ w: R.Tensor((4, 3, 3, 3), "float32"),
+ indices: R.Tensor((2, 4, 26, 26), "int64"),
+ ) -> R.Tensor(None, "float32", ndim=4):
+ with R.dataflow():
+ data: R.Tensor((2, 4, 26, 26), "float32") = R.nn.conv2d(x, w,
out_dtype="float32")
+ gv = R.gather_elements(data, indices, axis=1)
+ R.output(gv)
+ return gv
+
+ @I.ir_module
+ class Expected:
+ @R.function
+ def main(
+ x: R.Tensor((2, 3, 28, 28), dtype="float32"),
+ w: R.Tensor((4, 3, 3, 3), dtype="float32"),
+ indices: R.Tensor((2, 4, 26, 26), dtype="int64"),
+ ) -> R.Tensor(None, dtype="float32", ndim=4):
+ with R.dataflow():
+ lv: R.Tensor((2, 28, 28, 3), dtype="float32") =
R.permute_dims(x, axes=[0, 2, 3, 1])
+ lv1: R.Tensor((4, 3, 3, 3), dtype="float32") =
R.permute_dims(w, axes=[0, 2, 3, 1])
+ data: R.Tensor((2, 26, 26, 4), dtype="float32") = R.nn.conv2d(
+ lv,
+ lv1,
+ strides=[1, 1],
+ padding=[0, 0, 0, 0],
+ dilation=[1, 1],
+ groups=1,
+ data_layout="NHWC",
+ kernel_layout="OHWI",
+ out_layout="NHWC",
+ out_dtype="float32",
+ )
+ lv2: R.Tensor((2, 26, 26, 4), dtype="int64") = R.permute_dims(
+ indices, axes=[0, 2, 3, 1]
+ )
+ lv3: R.Tensor((2, 26, 26, 4), dtype="float32") =
R.gather_elements(
+ data, lv2, axis=3
+ )
+ gv: R.Tensor((2, 4, 26, 26), dtype="float32") = R.permute_dims(
+ lv3, axes=[0, 3, 1, 2]
+ )
+ R.output(gv)
+ return gv
+
+ verify(Input, Expected)
+
+
if __name__ == "__main__":
tvm.testing.main()
