search

ion-g-ion / torchTT

Tensor-Train decomposition in pytorch
MIT License
40 stars 10 forks source link

Anyway to improve the speed of .round() function? #21

Open hanfluid opened 6 months ago

ion-g-ion commented 6 months ago

Hi. The round function performs 2 sweeps one for the orthogonalization (QR decompositions) and one for the actual compression (SVDs). The asymptotic complexity cannot be reduced. Possible issues for a slow round can be:

hanfluid commented 6 months ago

Thanks a lot. Could you please elaborate the first point a little more? Is it like doing .round() everytime after a multiplication?

On Mon, Mar 4, 2024 at 1:21 PM ion-g-ion @.***> wrote:

Hi. The round function performs 2 sweeps one for the orthogonalization (QR decompositions) and one for the actual compression (SVDs). The asymptotic complexity cannot be reduced. Possible issues for a slow round can be:

  • ranks arising from successive multiplications. This can be solved by AMEn multiplications.
  • not using the right linear algebra backend. Intel MKL is usually the good choice in terms of performance.

— Reply to this email directly, view it on GitHub https://github.com/ion-g-ion/torchTT/issues/21#issuecomment-1977294092, or unsubscribe https://github.com/notifications/unsubscribe-auth/AKGVYA4XJX6EOVGZ47CLYKTYWTCTFAVCNFSM6AAAAABECWGV2OVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNZXGI4TIMBZGI . You are receiving this because you authored the thread.Message ID: @.***>