[draft][luci/pass] Introduce FuseGRU Pass #14252
Conversation
This PR introduces FuseGRUPass for fusing decomposed gru pattern into single CircleGRU. ONE-DCO-1.0-Signed-off-by: Artem Balyshev <[email protected]> ONE-DCO-1.0-Signed-off-by: Chunseok Lee <[email protected]>
| if (_while_node == nullptr) | ||
| return false; | ||
|
|
||
| // 1 - check condition graph: only one Less operation |
There was a problem hiding this comment.
only one Less operation and below condition doesn't match.
- 1/ fix comment like
Less operation should exist - 2/ fix implementation to check only one
Lessexist
| if (fc_nodes.size() != 2 or mul_nodes.size() != 3 or logistic_nodes.size() != 2 or | ||
| split_nodes.size() != 2 or add_nodes.size() != 6 or gather_nodes.size() != 1 or | ||
| reshape_nodes.size() != 1 or sub_nodes.size() != 1 or tanh_nodes.size() != 1 or | ||
| split_out_nodes.size() != 6) | ||
| return false; |
There was a problem hiding this comment.
As I commented in the draft code, I don't agree on only check number of Ops to check.
| split_out_nodes.size() != 6) | ||
| return false; | ||
|
|
||
| // Check structure |
There was a problem hiding this comment.
I don't understand the algorithm of below check codes.
Please explain what is happening.
| luci::CircleGRU *create_circle_gru(loco::Graph *graph); | ||
|
|
||
| private: | ||
| const GRUPatternBase *_p; |
There was a problem hiding this comment.
| const GRUPatternBase *_p; | |
| // initialized at ctor | |
| const GRUPatternBase *_p; |
| break; | ||
|
|
||
| default: | ||
| assert(false); |
There was a problem hiding this comment.
let's throw instead of assert
| } | ||
| else | ||
| { | ||
| bias_ih_cloned = _p->_pattern_last_node->graph()->nodes()->create<luci::CircleOutputExclude>(); |
There was a problem hiding this comment.
| bias_ih_cloned = _p->_pattern_last_node->graph()->nodes()->create<luci::CircleOutputExclude>(); | |
| bias_ih_cloned = graph->nodes()->create<luci::CircleOutputExclude>(); |
?
|
|
||
| luci::CircleGRU *FuseGRU::create_circle_gru(loco::Graph *graph) | ||
| { | ||
| assert(graph); |
There was a problem hiding this comment.
If this graph and _p->_pattern_last_node->graph() are different, please add a not about this.
As I understand, we're looking with multiple sub graph objects so graph ptr can be different.
| * | | ||
| * [Out_1] | ||
| */ | ||
| class GRUPattern1 final : public GRUPatternBase |
There was a problem hiding this comment.
I assume 1 suffix in GRUPattern1 is to prepare more patterns.
Please leave a note about thus.
Or please use just GRUPattern here and later we can rename this when we add more patterns.
|
|
||
| g.init(); | ||
|
|
||
| EXPECT_FALSE(pass.run(g.g())); |
There was a problem hiding this comment.
adding a note about what negative would help understanding FuseGRUTestNegGraph.
It's a bunch of nodes connected and I can't catch what it's doing.
Co-authored-by: SaeHie Park <[email protected]>
Co-authored-by: SaeHie Park <[email protected]>
|
BTW, what happens to the dead subgraphs after fusion? We have dead node elimination passes, but AFAIK we don't have such a pass for dead subgraphs. If dead subgraphs are not removed in the model, model size may be increased. |
the draft has deal subgraph elimination pass :) I will submit it soon |
Co-authored-by: Hyukjin Jeong <[email protected]>
|
@chunseoklee , how about splitting this PR to several small ones, so that review takes shorter. |
I will try, but I am not sure that how to split FUseGRUPass.cpp into smaller ones. Then, I will change this PR to draft state. |
chunseoklee
changed the title
Co-authored-by: Hyukjin Jeong <[email protected]>
This PR introduces FuseGRUPass for fusing decomposed gru pattern into single CircleGRU.
draft : #14237
issue : #12263
ONE-DCO-1.0-Signed-off-by: Artem Balyshev [email protected]
ONE-DCO-1.0-Signed-off-by: Chunseok Lee [email protected]