### Performance
 - synthetic test, elementwise add `a = b + 1`. Improvement around 2%

 - models (together with #452 )

|model|latency|prev_latency|delta|
|--------|--------|--------|--------|
|bert-base-uncased|19.6886|20.0891|2.034
|densenet121|35.0828|35.2025|0.341
|efficientnet_b0|18.8482|18.9655|0.622
|mobilenet_v2|11.5672|11.5901|0.198
|resnet50|29.0077|29.6469|2.204
|vit_b_16|123.344|126.705|2.725
|**GMEAN**| | |**1.349**

  
|model|latency|prev_latency|delta|
|--------|--------|--------|--------|
|bert-base-uncased|19.6886|19.9814|1.487
|densenet121|35.0828|35.2032|0.343
|efficientnet_b0|18.8482|18.9711|0.652
|mobilenet_v2|11.5672|11.5913|0.208
|resnet50|29.0077|29.101|0.322
|vit_b_16|123.344|124.837|1.21
|**GMEAN**| | |**0.702**

The above comparisons were done against exactly the same `main` branch.
Yes, we have a big fluctuation in perf results :(


### TODO
- doesn't work with dynamic shapes yet
- only `spatial` mapping is supported right now. Another mapping support
should be investigating

Сontinue indexes optimizations started in #458.

- We compare the difference between analyzing `expr` and flatten index.
Before we transformed when difference was a constant. Now when
difference don't contain loop variables.
 - Added new pass that expands leading `repeat()` mapping.


Bug fix:
 - In `Poli` class save and compare `var` itself but not `var.hint`

Other:
- always convert `Constant` to `Poli`. It avoid a bit confusing code
when we have to make 2 checks `isinstance(p, int) or p.is_constant()`

### Performance
 - synthetic test, elementwise add `a = b + 1`. Improvement around 2%

 - models (together with #452 )

|model|latency|prev_latency|delta|
|--------|--------|--------|--------|
|bert-base-uncased|19.6886|20.0891|2.034
|densenet121|35.0828|35.2025|0.341
|efficientnet_b0|18.8482|18.9655|0.622
|mobilenet_v2|11.5672|11.5901|0.198
|resnet50|29.0077|29.6469|2.204
|vit_b_16|123.344|126.705|2.725
|**GMEAN**| | |**1.349**

  
|model|latency|prev_latency|delta|
|--------|--------|--------|--------|
|bert-base-uncased|19.6886|19.9814|1.487
|densenet121|35.0828|35.2032|0.343
|efficientnet_b0|18.8482|18.9711|0.652
|mobilenet_v2|11.5672|11.5913|0.208
|resnet50|29.0077|29.101|0.322
|vit_b_16|123.344|124.837|1.21
|**GMEAN**| | |**0.702**

The above comparisons were done against exactly the same `main` branch.
Yes, we have a big fluctuation in perf results :(


### TODO
- doesn't work with dynamic shapes yet
- only `spatial` mapping is supported right now. Another mapping support
should be investigating

Сontinue indexes optimizations started in #458.

- We compare the difference between analyzing `expr` and flatten index.
Before we transformed when difference was a constant. Now when
difference don't contain loop variables.
 - Added new pass that expands leading `repeat()` mapping.


Bug fix:
 - In `Poli` class save and compare `var` itself but not `var.hint`

Other:
- always convert `Constant` to `Poli`. It avoid a bit confusing code
when we have to make 2 checks `isinstance(p, int) or p.is_constant()`

### Performance
 - synthetic test, elementwise add `a = b + 1`. Improvement around 2%

 - models (together with #452 )

|model|latency|prev_latency|delta|
|--------|--------|--------|--------|
|bert-base-uncased|19.6886|20.0891|2.034
|densenet121|35.0828|35.2025|0.341
|efficientnet_b0|18.8482|18.9655|0.622
|mobilenet_v2|11.5672|11.5901|0.198
|resnet50|29.0077|29.6469|2.204
|vit_b_16|123.344|126.705|2.725
|**GMEAN**| | |**1.349**

  
|model|latency|prev_latency|delta|
|--------|--------|--------|--------|
|bert-base-uncased|19.6886|19.9814|1.487
|densenet121|35.0828|35.2032|0.343
|efficientnet_b0|18.8482|18.9711|0.652
|mobilenet_v2|11.5672|11.5913|0.208
|resnet50|29.0077|29.101|0.322
|vit_b_16|123.344|124.837|1.21
|**GMEAN**| | |**0.702**

The above comparisons were done against exactly the same `main` branch.
Yes, we have a big fluctuation in perf results :(


### TODO
- doesn't work with dynamic shapes yet
- only `spatial` mapping is supported right now. Another mapping support
should be investigating

Сontinue indexes optimizations started in #458.

- We compare the difference between analyzing `expr` and flatten index.
Before we transformed when difference was a constant. Now when
difference don't contain loop variables.
 - Added new pass that expands leading `repeat()` mapping.


Bug fix:
 - In `Poli` class save and compare `var` itself but not `var.hint`

Other:
- always convert `Constant` to `Poli`. It avoid a bit confusing code
when we have to make 2 checks `isinstance(p, int) or p.is_constant()`