eliminate slice shape #82
Conversation
HSQ79815
force-pushed
the
eliminate_nop_slice_shape
branch
from
557c48f to
ab3b293
Compare
| const auto &dims_of_shape_node_input = shape_node->input()->sizes(); | ||
| std::vector<Dimension> result_of_shape_op; | ||
|
|
||
| auto add_y_if_negative = [](int64_t x, int64_t y) -> int64_t { |
There was a problem hiding this comment.
please extract this function into some common file for example creating a new file named onnxoptimizer/passes/utils.h
There was a problem hiding this comment.
Other passes also use this helper function, so I will handle with it by new PR.
| namespace ONNX_NAMESPACE { | ||
| namespace optimization { | ||
|
|
||
| struct EliminateSliceShape final : public PredicateBasedPass { |
There was a problem hiding this comment.
EliminateShapeSlice is better because the order of eliminated nodes in graph is Shape -> Slice
There was a problem hiding this comment.
what do you think of EliminateSliceAfterShape ? In this pass, Pattern Shape -> Slice be matched, we want to eliminate Slice op , but Shape will reserve.
Signed-off-by: haoshengqiang <[email protected]>
Signed-off-by: haoshengqiang <[email protected]>
HSQ79815
force-pushed
the
eliminate_nop_slice_shape
branch
from
ab3b293 to
2d3a205
Compare
Signed-off-by: haoshengqiang <[email protected]>
Signed-off-by: haoshengqiang <[email protected]>
| if (!fetch_first_value_of_tensor(node->inputs()[1], slice_start) || | ||
| !fetch_first_value_of_tensor(node->inputs()[2], slice_end) || | ||
| (node->inputs().size() == 5 && | ||
| !fetch_first_value_of_tensor(node->inputs()[4], slice_step)) || |
There was a problem hiding this comment.
Is fetch_sole_value_of_tensor a better name? And then an ONNX_ASSERT should be added to check the tensor has only one element.
Signed-off-by: haoshengqiang <[email protected]>
Signed-off-by: haoshengqiang [email protected]