Hi All,
I was experimenting to decompose a relay expression (operator or a function)
into the same kind of operator. Mainly, if I have to do vector addition, I just
want to replicate the number of additions but each addition operates on part of
the data.
For example, to add two vectors of s
@comaniac - are you assuming that user needs to extend from the ExprMutator
class?
I have been mostly user of TVM, and now, I'd like to spend some time to
understand relay.
How does this method differs from the post_order_visit function provided by
TVM?
[quote="comaniac, post:3, topic:6
Hi All,
I am wondering if someone (or if there already exists) functionality that
converts neural network module from PyTorch to Relay IR?
I have seen PyTorhc/TVM project, but I am not sure if this project is
converting the pytorch to Relay IR or the old TVM IR?
Thanks,
S.
---
[Visit
Dear All,
I am wondering how can I write a Relay pass that tiles conv2d by the output
channels (data partitioning) in Relay graph level?
For example, let us assume that I have some relay program like below, and I
want to able to traverse the relay graph that contains this conv2d, and able t
Thank you @masahi
This is very helpful. But, I am more than puzzled.
Let us say you have two HW units of capable of running (the original example
add1 and add2). So according to your answer, the add1 and add2 CAN NOT run in
parallel? Could you provide some insights and on this? Also provide
Thank you @hht
This is very useful. I have two follow up questions.
1) what is the purpose of external_mods in the LoweredOutput structure?
2) Ia m wondering if I can get more details about how the CodeGen in TVM works?
I mean what is the sequence. I know it starts from Relay, and I am
@hht - thank you again. Now it makes a kind of sense. Could you please clarify
what do you mean by "Parallelism only exists in the module."? My understanding
is that there is only one Module, and module contains multiple graph nodes that
can run in parallel.
[quote="hht, post:5, topic:657
@hht -- this is definitely interesting. In my given example, add1 and add2 are
Op types, and thus, I'd expect them to be run in parallel in a HW that is
capable of running two adders ("+") in parallel.
[quote="hht, post:5, topic:6572"]
There is no strategy to enforce parallelism to the op_e
@hht -- thank you very much.
So is this mean, we can not enforce parallelism to GraphRuntime? If I
understand correctly, it looks like GraphRuntime does not run add1 and add2 in
parallel?
Basically, I am wondering if there is a mechanism to enforce parallelism to
GraphRuntime from the hi
Dear All,
I am wondering how the execution order of operators is defined at runtime in
TVM?
For example, in the following example, add1 and add2 are parallel, and how the
TVM runtime schedules these on hardware? (Surely, it depends on target HW, but
assuming we have a HW that its capable of
Thank you very much.
How operators are running (or being scheduled) is one thing that TVM needs
documentation.
I think it is important for people to understand how the graph operators are
being executed because there is more parallelism in the graph level than
operator level in some network
Dear All,
I am new to TVM and having trouble understanding the way TVM selects and
executes operators.
Question: How TVM decides which operator (add1 and add2 are parellel) to
execute as follows? and where is this information (and how this information) is
being figured out by TVM?
```
x1
12 matches
Mail list logo
