A question about fusion of TVM and MXNet/mshadow

[Tao Sun <su...@gmail.com>]

After some reading and learning of MXNet, I tentatively made the conclusion
that the migration of MXNet to TVM, if it exists, has yet finished.
Currently, mshadow, on which most of current operator code is still based,
seems to carry out the most functionality that TVM can carry ultimately.
Will anyone correct/confirm this conclusion? Will the community have a
plan/calendar to migrate to TVM? Thanks.


Tao Sun

Re: A question about fusion of TVM and MXNet/mshadow

[Tao Sun <su...@gmail.com>]

My suggestion is that there be a formal doc explaining the relationship
between TVM/NNVM and MXNet/mshadow and the next step of their roles in the
MXNet as a whole. I think this would ease the learning curve of potential
developers.

On Wed, Aug 8, 2018 at 3:59 PM, Tianqi Chen <tq...@cs.washington.edu>
wrote:

> It is true that most of the current GPU code depends on mshadow. Porting
> the operator code entirely over to TVM will take quite a huge effort. So a
> more gradual path forward is to could be drop-in TVM to support cases that
> it optimizes well(ARM, AMDGPU, accelerators) while keeping the old
> infrastructure around for a while.
>
> This is my part of the technical assessment. There is not yet a proposal
> for a complete migration over TVM in the community.
>
> Tianqi
>
> On Wed, Aug 8, 2018 at 12:52 PM, Tao Sun <su...@gmail.com> wrote:
>
> > After some reading and learning of MXNet, I tentatively made the
> conclusion
> > that the migration of MXNet to TVM, if it exists, has yet finished.
> > Currently, mshadow, on which most of current operator code is still
> based,
> > seems to carry out the most functionality that TVM can carry ultimately.
> > Will anyone correct/confirm this conclusion? Will the community have a
> > plan/calendar to migrate to TVM? Thanks.
> >
> >
> > Tao Sun
> >
>

Re: A question about fusion of TVM and MXNet/mshadow

[Tianqi Chen <tq...@cs.washington.edu>]

It is true that most of the current GPU code depends on mshadow. Porting
the operator code entirely over to TVM will take quite a huge effort. So a
more gradual path forward is to could be drop-in TVM to support cases that
it optimizes well(ARM, AMDGPU, accelerators) while keeping the old
infrastructure around for a while.

This is my part of the technical assessment. There is not yet a proposal
for a complete migration over TVM in the community.

Tianqi

On Wed, Aug 8, 2018 at 12:52 PM, Tao Sun <su...@gmail.com> wrote:

> After some reading and learning of MXNet, I tentatively made the conclusion
> that the migration of MXNet to TVM, if it exists, has yet finished.
> Currently, mshadow, on which most of current operator code is still based,
> seems to carry out the most functionality that TVM can carry ultimately.
> Will anyone correct/confirm this conclusion? Will the community have a
> plan/calendar to migrate to TVM? Thanks.
>
>
> Tao Sun
>