graph: utils: pm: support commutative select #3166
Conversation
| @@ -252,7 +254,8 @@ bool match_node_inputs(const binding_t &b, match_context_t *ctx, | |||
| if (node_inputs_matcher.get_node()->get_inputs().size() | |||
| == VARIADIC_INPUT_NUM) { | |||
| matching_status = node_inputs_matcher.match_variadic_inputs(); | |||
| } else if (!has_commutative_inputs(node_inputs_matcher.get_op())) { | |||
| } else if (!has_commutative_inputs(node_inputs_matcher.get_op(), | |||
| {b.bind_op_port, b.bind_port})) { | |||
There was a problem hiding this comment.
I'm wondering, for select, whose first input is non-commutative, and second & third inputs are commutative, how can we make sure all the inputs are correctly visited and matched?
There was a problem hiding this comment.
When bind_node is select, bind_port is 2 and bind_op_port is 1 which conform to commutative inputs = {1, 2}.
There was a problem hiding this comment.
I understand that. I mean how is the 0-th input matched? As it's non-commutative.
There was a problem hiding this comment.
Do you mean the first or the second input will swap with the first one? I doubt whether that's a valid graph, since the data type of the 0-th input is boolean after all. As for other cases, like when all three input data types are the same but only two are commutative, I don't think we have such a requirement.
There was a problem hiding this comment.
I have added gtest for verifying match_node_inputs()
ShanSimu
force-pushed
the
shaojiec/select_commutative
branch
from
7ffd645 to
b5fc799
Compare
|
please format the commit message with correct component, graph: interface: ..., graph: utils: pm: ... |
Thanks for letting me know. I split the commit into three parts based on file paths, using the directory structure as the distinction: |
ShanSimu
force-pushed
the
shaojiec/select_commutative
branch
from
b5fc799 to
26cd4b3
Compare
ShanSimu
changed the title
ShanSimu
changed the title
ShanSimu
force-pushed
the
shaojiec/select_commutative
branch
from
26cd4b3 to
7e9349a
Compare
ShanSimu
force-pushed
the
shaojiec/select_commutative
branch
from
7e9349a to
37c9cd8
Compare
| // MatMul Add | ||
| // \ / | ||
| // Select | ||
| auto pmatmul = graphp->append_op(MatMul); |
There was a problem hiding this comment.
Please add another input for select's 0-th input (e.g. GreaterEqual), we want to know if the three inputs (both commutative and non-commutative) can all be matched
| agraph.finalize(); | ||
|
|
||
| std::vector<op_t *> fusion_ops; | ||
| EXPECT_TRUE(match_pattern(agraph.get_ops()[0].get(), graphp, fusion_ops)); |
There was a problem hiding this comment.
And here we need to match from the select op, to verify our assumption.
Description
The original setting for "commutative" was limited to operators with only two inputs. However, the select operator has three inputs, and only two of them are commutative.
A new member variable,
std::set<size_t> commutative_inputs, has been added toop_schemato store the indices of commutative inputs. The functionset_commutative_inputs(std::set<size_t> inputs = {0, 1})uses default parameters to maintain support for the original commutative operators.In
nested_matcher.cpp, the functionshas_commutative_inputs,match_commutative_inputs, andmatch_node_inputshave been updated to include checks for whether a port belongs to the commutative inputs.This PR is for fix MFDNN-13383. Another method is also implemented here, which is achieved by swapping the mapping method of ports inside and outside the subgraph.