Migration guide for Tensor update

In SlangPy version 0.41 we undertook a major rewrite of the Tensor API to bring it inline with existing standards and simplify its use going forwards. The primary changes made are:

Python API

  • NDBuffer is fully deprecated

  • Tensor is the sole ND container type, and supports both differentiable and non-differentiable data

Slang API

  • NDBuffer is removed

  • Tensor and RWTensor types now only store non-differentiable data

  • WTensor introduced to store write-only non-differentiable data

  • DiffTensor, WDiffTensor and RWDiffTensor introduced to store differentiable data

  • get and set methods replaced with load and store methods respectively

  • getv and setv methods replaced with load and store methods respectively

  • subscript operators correctly implemented for all Tensors

Migrating

Most migration steps can be automated using search-and-replace, however doing them in the correct order is important to avoid naming issues.

1. Remove use of NDBuffer in Python code

  • Replace all instances of NDBuffer with Tensor

  • Fix any use of the NDBuffer constructor to use Tensor.empty

Note: Use of NDBuffer constructor with positional arguments may not be easily fixable with search-and-replace. It is probably worth reviewing these instances manually.

2. Replace differentiable Tensors parameters in Slang

If you have exclusively used the Tensor type as function parameters, simply search and replacing as below should be sufficient.

  • Replace all instances of Tensor with IDiffTensor

  • Replace all instances of RWTensor with IRWDiffTensor

  • Replace all instances of GradInTensor with IWDiffTensor

  • Replace all instances of GradOutTensor with IDiffTensor

  • Replace all instances of GradInOutTensor with IRWDiffTensor

Where tensors have been used as variables, it can depend on the use case. Typically:

  • Tensor/RWTensor will typically be either:

    • Keep the same if using purely as ND storage

    • Replace with PrimalTensor/ RWPrimalTensor if you need it to be compatible with IDiffTensor interfaces

  • GradInTensor should be changed to WDiffTensor

  • GradOutTensor should be changed to DiffTensor

  • GradInOutTensor should be changed to RWDiffTensor

3. Replace NDBuffers with non-differentiable Tensors in Slang code

For function arguments:

  • Replace all instances of NDBuffer with ITensor

  • Replace all instances of RWWBuffer with IRWTensor

Where NDBuffers have been used as variables:

  • Replace all instances of NDBuffer with Tensor

  • Replace all instances of RWBuffer with RWTensor

4. Fix get/set and getv/setv method calls

  • Replace all instances of .get( with .load(

  • Replace all instances of .set( with .store(

  • Replace all instances of .getv( with .load(

  • Replace all instances of .setv( with .store(

5. Fix any errors attempting to access gradient buffers in Slang

If Slang has compile errors attempting to access the ‘grad’ properties of tensors, switch them from using IDiffTensor to the concrete DiffTensor (or corresponding) types.