!test

Review updated until commit b7778bd

Description

• Add new utility function resetContiguityFromTensor in ir_utils namespace

• Function infers and resets TensorView contiguity based on actual at::Tensor sizes/strides

• Handles complex dimension types: iteration, broadcast, and reduction domains

• Includes comprehensive test coverage with 8 test cases for different contiguity patterns

!test

!test

!test

Greptile Summary

Added utility function ir_utils::resetContiguityFromTensor that infers and sets TensorView contiguity based on actual tensor strides. The implementation uses a two-pointer algorithm traversing right-to-left through the allocation domain, correctly handling broadcast dimensions (set to nullopt, consume sizes/strides entry) and reduction dimensions (set to nullopt, don't consume entry). Contiguity is determined by comparing consecutive strides: innermost dimension is contiguous if stride==1, others are contiguous if stride equals the product of the next dimension's stride and size.

Key changes:

• csrc/ir/utils.cpp: Implemented resetContiguityFromTensor with proper index management and validation
• csrc/ir/utils.h: Added function declaration with clear documentation
• tests/cpp/test_utils.cpp: Added comprehensive test covering all 8 contiguity combinations for 3 iteration dimensions mixed with broadcast/reduction dimensions

Testing coverage: The test validates all 2^3=8 combinations of contiguity patterns for a 5-dimension allocation domain [Iteration, Broadcast, Iteration, Reduction, Iteration] mapped to tensor shape [3, 1, 4, 6], ensuring correct handling of non-leftmost broadcast dimensions and reduction dimensions with no sizes/strides entries.

Confidence Score: 4/5

• This PR is safe to merge with minimal risk - adds well-tested utility function with no changes to existing code paths
• Score reflects solid implementation with comprehensive testing. The algorithm correctly handles all tested cases with proper index management and validation. Minor deduction due to potential edge case with size-0 dimensions (rare in practice but worth noting). The utility function is self-contained and will be used in a future PR, so no immediate production impact.
• No files require special attention - all changes are additive and well-tested

Important Files Changed

Sequence Diagram

sequenceDiagram
    participant Client
    participant resetContiguityFromTensor
    participant inferAllocationSizesAndStrides
    participant TensorView
    participant TensorDomain

    Client->>resetContiguityFromTensor: Call with TensorView & at::Tensor
    
    alt tensor not defined
        resetContiguityFromTensor->>Client: Early return
    end
    
    resetContiguityFromTensor->>inferAllocationSizesAndStrides: Get sizes & strides
    Note over inferAllocationSizesAndStrides: Returns vectors excluding<br/>reduction dimensions
    inferAllocationSizesAndStrides-->>resetContiguityFromTensor: (sizes, strides)
    
    resetContiguityFromTensor->>TensorView: getMaybeAllocationDomain()
    TensorView-->>resetContiguityFromTensor: alloc domain (all dims)
    
    Note over resetContiguityFromTensor: Initialize contiguity vector<br/>(all nullopt)
    
    loop Right-to-left traversal (alloc_idx)
        alt Reduction dimension
            Note over resetContiguityFromTensor: Keep nullopt<br/>Don't consume sizes/strides
        else Broadcast dimension
            Note over resetContiguityFromTensor: Keep nullopt<br/>Consume sizes/strides entry<br/>(decrement sizes_idx)
        else Regular iteration dimension
            alt Rightmost non-skipped dim
                Note over resetContiguityFromTensor: Check: stride == 1
            else Other dimensions
                Note over resetContiguityFromTensor: Check: stride ==<br/>prev_stride * prev_size
            end
            Note over resetContiguityFromTensor: Set contiguity[alloc_idx]<br/>Update prev_non_skipped_sizes_idx
        end
    end
    
    Note over resetContiguityFromTensor: Validate all sizes/strides<br/>consumed (sizes_idx == -1)
    
    resetContiguityFromTensor->>TensorView: setContiguity(contiguity)
    TensorView->>TensorDomain: setContiguity(contiguity)
    
    resetContiguityFromTensor->>Client: Return

!test

!test