Printf mcs mutex#574
Conversation
mrutt92
force-pushed
the
printf_mcs_mutex
branch
from
9c0a69d to
b5faa01
Compare
|
is this one ready to go, max? |
| extern int __bsg_pod_x; //The X Cord of the tile's pod (lives in DRAM) | ||
| extern int __bsg_pod_y; //The Y cord of the tile's pod (lives in DRAM) |
There was a problem hiding this comment.
You could grab these values from cfg_pod CSR at the beginning. There is no need to have the host set these.
There was a problem hiding this comment.
I guess that's true. It save's a trivial few instruction words in the initialization to have the host set it. That CSR resets to the current pod but it can be set by the core to something else (like if it wants to read another pod's DRAM).
There was a problem hiding this comment.
I'm honestly not 100% sure which is better. I think it's better to have the host set it? It's just a couple extra packets in the nbf and we don't need to entangle it with the semantics of the pod csr. But I'm not sure, I'm open to be talked down from that.
There was a problem hiding this comment.
I don't think it makes sense to keep these variable in DRAM, because if you change the pod x/y CSR then now you are reading those values from different pods' DRAM space. I think it's better off to make global x/y readable by CSR instructions (read-only), then it doesn't take up space in neither DRAM or DMEM.
There was a problem hiding this comment.
I think that makes a lot of sense. That'll also decouple these values from the semantics of the pod address register.
There was a problem hiding this comment.
Why is it necessary to access the X and Y coordinate of the pod? As a general principle we would like to avoid code that depends on this, since it is a virtualization hazard...
There was a problem hiding this comment.
It's a global lock.
There was a problem hiding this comment.
Yes why do we need a global lock across pods..
There was a problem hiding this comment.
Maybe we don't? If each pod has its own IO node then there's no need.
There was a problem hiding this comment.
At the moment, there's only one IO node shared across pods. A global lock prevents the IO node from being slammed with packets from all pods.
mrutt92
force-pushed
the
printf_mcs_mutex
branch
from
b5faa01 to
2cb035b
Compare
|
Seems okay though? With individual pod locks, the throughput is already
throttled to some extent?
…On Wed, Sep 29, 2021 at 11:56 AM mrutt92 ***@***.***> wrote:
***@***.**** commented on this pull request.
------------------------------
In software/bsg_manycore_lib/bsg_tile_config_vars.h
<#574 (comment)>
:
> +extern int __bsg_pod_x; //The X Cord of the tile's pod (lives in DRAM)
+extern int __bsg_pod_y; //The Y cord of the tile's pod (lives in DRAM)
At the moment, there's only one IO node shared across pods. A global lock
prevents the IO node from being slammed with packets from all pods.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#574 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AEFG5AFMCSU3ZEU5CKRUTITUENONDANCNFSM5DSUZDRQ>
.
|
|
My thought is that the lock is presumably there because of lack of
re-entrancy in some library; which would not apply for multipod, since they
have their own DRAM spaces. Or is the lock there to reduce interspersing of
characters or ... ?
On Wed, Sep 29, 2021 at 12:05 PM Michael Nguyen Taylor <
***@***.***> wrote:
… Seems okay though? With individual pod locks, the throughput is already
throttled to some extent?
On Wed, Sep 29, 2021 at 11:56 AM mrutt92 ***@***.***> wrote:
> ***@***.**** commented on this pull request.
> ------------------------------
>
> In software/bsg_manycore_lib/bsg_tile_config_vars.h
> <#574 (comment)>
> :
>
> > +extern int __bsg_pod_x; //The X Cord of the tile's pod (lives in DRAM)
> +extern int __bsg_pod_y; //The Y cord of the tile's pod (lives in DRAM)
>
> At the moment, there's only one IO node shared across pods. A global lock
> prevents the IO node from being slammed with packets from all pods.
>
> —
> You are receiving this because you commented.
> Reply to this email directly, view it on GitHub
> <#574 (comment)>,
> or unsubscribe
> <https://github.com/notifications/unsubscribe-auth/AEFG5AFMCSU3ZEU5CKRUTITUENONDANCNFSM5DSUZDRQ>
> .
>
|
|
It prevents one IO node from being slammed and it also serves as keeping the messages from being garbled. |
3 participants
No description provided.