mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 12:54:11 +08:00
31681bcacc
Summary: ARM has provided with an SVE128 box-cox implementation. It uses the same underlying algorithm as the previous version, but it has better log and exp implementations. These supplied mathematical functions have switches to adjust the precision/speed trade-off. We've noted a slight precision improvement, while also about a 5% peroformance increase Before: ZeroLambda1 61.66ns 16.22M NonZeroLambda1 125.73ns 7.95M NonZeroLambdaManyColumns 1.84ms 542.11 NonZeroLambdaEigenColumnar 262.31us 3.81K NonZeroLambdaEigenRowMajor 275.17us 3.63K NonZeroLambdaWithPyTorchColumnar 97.43us 10.26K NonZeroLambdaWithPyTorchRowMajor 90.82us 11.01K NonZeroLambdaWithPyTorchRowMajorFullBatch 96.96us 10.31K NonZeroLambdaBatch 151.84us 6.59K After: ZeroLambda1 57.85ns 17.29M NonZeroLambda1 118.85ns 8.41M NonZeroLambdaManyColumns 1.82ms 548.16 NonZeroLambdaEigenColumnar 261.67us 3.82K NonZeroLambdaEigenRowMajor 274.53us 3.64K NonZeroLambdaWithPyTorchColumnar 89.12us 11.22K NonZeroLambdaWithPyTorchRowMajor 83.49us 11.98K NonZeroLambdaWithPyTorchRowMajorFullBatch 88.79us 11.26K NonZeroLambdaBatch 144.74us 6.91K Test Plan: Correctness: buck2 test @//mode/opt //koski/functions_contrib/df4ai/tests:batch_box_cox_test Performance: buck2 run @//mode/opt //koski/functions_contrib/df4ai/benchmark:boxcox_benchmark Differential Revision: D83485704 Privacy Context Container: L1196524 Pull Request resolved: https://github.com/pytorch/pytorch/pull/164152 Approved by: https://github.com/ezyang