The goal of this project is to implement a Value-Domain Analysis pass for tensor compilers using the MLIR TOSA dialect. This analysis maps tensor slices—subsets of a tensor with one fewer dimension—to their specific min-max ranges.

For example, a 3D tensor is analyzed as a series of 2D slices. If a slice contains only zeros, as seen in the "zero-table" results, the analysis identifies this sparsity precisely.

Optimization Use Cases

By understanding the value range of each slice, a compiler can perform several high-impact optimizations:

• Per-Channel Quantization: Automatically determine the scale and zero-point for INT8 quantization based on the dynamic range of individual filters (slices).
• Layer Elimination: If a slice's range is strictly positive (e.g., ), a subsequent ReLU layer is redundant and can be eliminated for that slice.
• Sparsity-Aware Execution: Identify "zero-slices" (where Min=0 and Max=0) to skip computation or use sparse-tensor storage formats.
• Bit-Width Optimization: Determine if a slice can fit into a smaller data type (e.g., FP16 or INT16) without loss of precision.
• Dead Code Elimination: If the value range of a tensor is constant across its entire domain, downstream operations that branch on those values can be simplified.

Build

To build the VISHAP optimizer, ensure you have MLIR 22.0 configured and run:

mkdir build
cd build
cmake ..
make

Run

You can run the value profiler on any TOSA-based MLIR file. For example, to analyze the 3D tensor provided in the test suite:

./bin/vishap-opt ../tests/value_profiler/t0.mlir

Example Output

The tool analyzes slices along the first dimension. For a hypothetical 3 * 2 * 2 tensor:

• Slice 0: Min=0.0, Max=2.0 (Contains values like 1 and 2)
• Slice 1: Min=0.0, Max=0.0 (Zero-slice)
• Slice 2: Min=0.0, Max=4.0 (Contains values like 3 and 4)