Byte threshold for ConstantFolding transformation #25239
Conversation
github-actions
bot
added
category: Core
| @@ -125,6 +125,7 @@ bool ov::pass::MOCTransformations::run_on_model(const std::shared_ptr<ov::Model> | |||
| f->validate_nodes_and_infer_types(); | |||
| } | |||
|
|
|||
| const int64_t cf_threshold = 1024; | |||
There was a problem hiding this comment.
can you give explanation what is threshold? How is this value selected?
Also, I think it should be a macro, sort of configuration for ConstantFolding
There was a problem hiding this comment.
1024 is just a preliminary estimation: usually ConstantFolding folds 3-4 operations (~250 bytes each) from xml file during one step, so to make the sum of IR xml + bin sizes more or less stable, we allow ConstantFolding to increase the bin size to this value.
ok, I will add the macro
| // do not apply ConstantFolding for this particular node because | ||
| // the created Constant might increase the bin size beyond the required threshold. | ||
| continue; | ||
| } |
There was a problem hiding this comment.
Could you please clearify. CF_BYTE_THRESHOLD - is difference for entire model or for particular node? I see current calculation is doing for each node.
| // So ~1000 bytes will be deleted from xml file. | ||
| // To make the sum of xml + bin more or less stable, | ||
| // we allow ConstantFolding to increase the bin size to this value. | ||
| #define CF_BYTE_THRESHOLD 1024 |
There was a problem hiding this comment.
| #define CF_BYTE_THRESHOLD 1024 | |
| constexpr size_t CF_BYTE_THRESHOLD 1024 |
?
| private: | ||
| // this variable controls the threshold by which it is allowed to increase | ||
| // the size of the binary file during one step of the ConstantFolding execution. | ||
| int64_t m_byte_threshold; |
There was a problem hiding this comment.
| int64_t m_byte_threshold; | |
| const size_t m_byte_threshold; |
Should we allow change after creation (remove const is in most cases easier).
Set it to max as default and use it if less then max.
| for (const auto& input_value : node->input_values()) { | ||
| auto constant = ov::as_type_ptr<ov::op::v0::Constant>(input_value.get_node_shared_ptr()); | ||
| if (constant && is_threshold_applicable) { | ||
| byte_size_before += static_cast<int64_t>(constant->get_byte_size()); |
There was a problem hiding this comment.
If threshold would be unsigned then this byte size also can be and cast will be not required.
| if (constant && is_threshold_applicable) { | ||
| byte_size_before += static_cast<int64_t>(constant->get_byte_size()); | ||
| } else { | ||
| is_threshold_applicable = false; |
There was a problem hiding this comment.
| is_threshold_applicable = false; | |
| is_threshold_applicable = false; | |
| break; |
?
| byte_size_after += static_cast<int64_t>( | ||
| element::get_memory_size(output.get_element_type(), ov::shape_size(output.get_shape()))); | ||
| } else { | ||
| is_threshold_applicable = false; |
There was a problem hiding this comment.
| is_threshold_applicable = false; | |
| is_threshold_applicable = false; | |
| break; |
?
| broadcast_v1->set_friendly_name("test"); | ||
| auto f = make_shared<Model>(broadcast_v1, ParameterVector{}); | ||
|
|
||
| run_constant_folding_with_threshold(f, 1024); |
There was a problem hiding this comment.
| run_constant_folding_with_threshold(f, 1024); | |
| run_constant_folding_with_threshold(f, CF_BYTE_THRESHOLD ); |
|
This PR will be closed in a week because of 2 weeks of no activity. |
|
This PR was closed because it has been stalled for 2 week with no activity. |
Details:
Adding a byte threshold for ConstantFolding transformation.
If byte size of the folded constant exceed this threshold, we do not fold this operation.
Tickets: