search

RubixML / Tensor

A library and extension that provides objects for scientific computing in PHP.
https://rubixml.com
MIT License
232 stars 28 forks source link

Fixes for PHP 8.2 crashes #41

Closed 27pchrisl closed 1 year ago

27pchrisl commented 1 year ago

This PR changes how arrays are accessed in all the operations. It appears that for some reason the hashtables are becoming non-packed, but I am not sure why. Changing directly accessing arData to using zend_hash_index_find fixes this issue. Under the hood zend_hash_index_find evaluates whether the array is properly packed, and delegates to directly accessing arData. So I think this change should be as performant as it was before, if your arrays are properly packed.

All the tests run successfully with these changes.

There are still deprecation warnings about self and static calls, but it looks like Zephir has not yet developed a fix for this. https://github.com/zephir-lang/zephir/issues/2405

There are also warnings like "Implicit conversion from float 3.3 to int loses precision", which appear to be coming from calls into Zend mod_function. I was not sure of the mathematically correct way to resolve this one. Minor changes like int to float warnings are also fixed, and a return type of Vector is fixed to match the non-extension code.

(Potentially) fixes #33 and #38

github-actions[bot] commented 1 year ago

CLA Assistant Lite bot All contributors have signed the CLA ✍️ ✅

27pchrisl commented 1 year ago

I have read the CLA Document and I hereby sign the CLA

andrewdalpino commented 1 year ago

Awesome @27pchrisl thanks for looking into this for us! Your help is greatly appreciated.

Did you run the benchmarks with these changes yet by any chance? I'm curious to know if we take a performance hit. If there is a performance hit, do you think we could try to find the point at which these arrays become "unpacked" and "re-pack" them as a way to mitigate the performance loss?

andrewdalpino commented 1 year ago

There doesn't appear to be much (if any) of a performance impact.

Baseline ...

|---|---|---|---|---|---|---|---| | MatrixMatrixMultiplyBench | multiply | | 1 | 5 | 112.114mb | 0.015s | ±9.78% | | MatrixVectorMultiplyBench | multiply | | 1 | 5 | 75.193mb | 0.011s | ±9.72% | | MatrixScalarMultiplyBench | multiply | | 1 | 5 | 75.119mb | 0.011s | ±1.85% | | VectorVectorMultiplyBench | multiply | | 1 | 5 | 3.737mb | 0.385ms | ±2.14% |

After your changes ...

|---|---|---|---|---|---|---|---| | MatrixMatrixMultiplyBench | multiply | | 1 | 5 | 112.114mb | 0.015s | ±12.67% | | MatrixVectorMultiplyBench | multiply | | 1 | 5 | 75.193mb | 0.012s | ±9.72% | | MatrixScalarMultiplyBench | multiply | | 1 | 5 | 75.119mb | 0.011s | ±0.19% | | VectorVectorMultiplyBench | multiply | | 1 | 5 | 3.737mb | 0.393ms | ±58.90% |