Fix/kernel termination return by guosran · Pull Request #297 · tancheng/VectorCGRA

guosran · 2026-06-19T19:10:51Z

Summary

Addresses #292 with a minimized dynamic-control termination fix. This PR only updates CtrlMemDynamicRTL.py and its focused test.

Changes

Mark the final real dynamic control word with is_last_ctrl.
Avoid relying on a synthetic timeout COMPLETE packet after the final step.
Add a concrete example as an inline RTL comment.

No new ports or interfaces are introduced.

guosran · 2026-06-26T00:32:09Z

Concrete example:
Completion should come from the explicit RET path, not from a control-memory timeout.

ctrl steps:
ADD
MUL
RET value=42

Before:
ctrl memory could send COMPLETE when the step counter reached the end.

After:
CtrlMemDynamicRTL waits for the RET response from the element path and forwards that result to the controller.

yyan7223 · 2026-06-27T04:59:07Z

+         (s.recv_pkt_from_controller_queue.send.msg.payload.cmd == CMD_CONFIG_GEP_STRIDE) | \
         (s.recv_pkt_from_controller_queue.send.msg.payload.cmd == CMD_UPDATE_COUNTER_SHADOW_VALUE) | \
-         (s.recv_pkt_from_controller_queue.send.msg.payload.cmd == CMD_RESET_LEAF_COUNTER):
+         (s.recv_pkt_from_controller_queue.send.msg.payload.cmd == CMD_RESET_LEAF_COUNTER) | \


Those changes are unrelated to this PR, right?

yyan7223 · 2026-06-27T05:43:08Z

          s.send_pkt_to_controller.msg @= \
-              IntraCgraPktType(zext(s.tile_id, IntraPktTileIdType), num_tiles, 0, 0, 0, 0, 0, 0, 0, 0,
+              IntraCgraPktType(s.tile_id, num_tiles, 0, 0, 0, 0, 0, 0, 0, 0,
                               s.recv_from_element_queue.send.msg)


How do you make sure the result of OPT_RET is sent back to CPU? recv_from_element_queue may contain the results of any operations.

yyan7223 · 2026-06-27T05:48:33Z

-      # Sets each FU's op code as NAH when prologue execution has not completed.
-      # As FU is supposed to do nothing during prologue.
-      if s.prologue_count_outport_fu != 0:
-        s.send_ctrl.msg.operation @= OPT_NAH


Why it is removed?

yyan7223 · 2026-06-27T05:59:41Z

+      # FlexibleFuRTL so the real control word remains visible to the
+      # crossbars and debug signals.
+      if ~s.send_ctrl.val:
+        s.send_ctrl.msg.operation @= OPT_START


Why OPT_START?

yyan7223 · 2026-06-27T06:03:50Z

+      s.send_ctrl.msg.is_last_ctrl        @= \
+          s.reg_file.rdata[0].is_last_ctrl | \
+          ((s.total_ctrl_steps_val > 0) &
+           (s.times == s.total_ctrl_steps_val - TimeType(1)))


Can you explain why we need ((s.total_ctrl_steps_val > 0) & (s.times == s.total_ctrl_steps_val - TimeType(1))) ?

yyan7223 · 2026-06-27T06:04:13Z


    @update
-    def update_send_ctrl_val():
+    def update_send_ctrl():


yyan7223

Are you using AI to help you with the PR?

tancheng · 2026-06-27T07:07:49Z

Are you using AI to help you with the PR?

Of coarse. And using AI is totally fine. And we should encourage that. @yyan7223 do you have any concern~?

But for this PR, AI might not aware of our original motivation, we indeed provided the total_steps to make the kernel be able to return COMPLETE for better debugging or for the systolic array case (where there is no RETURN instruction mapped).

yyan7223 · 2026-06-27T07:16:19Z

Concrete example: Completion should come from the explicit RET path, not from a control-memory timeout.

ctrl steps: ADD MUL RET value=42

Before: ctrl memory could send COMPLETE when the step counter reached the end.

After: CtrlMemDynamicRTL waits for the RET response from the element path and forwards that result to the controller.

I agree with your proposal.
For implementation, I think we should simultaneously check whether the opcode of current send_ctrl is OPT_RET and the predicate of its execution result is 1.

yyan7223 · 2026-06-27T07:24:14Z

Are you using AI to help you with the PR?

Of coarse. And using AI is totally fine. And we should encourage that. @yyan7223 do you have any concern~?

But for this PR, AI might not aware of our original motivation, we indeed provided the total_steps to make the kernel be able to return COMPLETE for better debugging or for the systolic array case (where there is no RETURN instruction mapped).

I'm totally fine with AI, I also use it for coding, and I check every line of code it generated carefully to make sure it is reasonable.

tancheng · 2026-06-27T18:36:11Z

Concrete example: Completion should come from the explicit RET path, not from a control-memory timeout.
ctrl steps: ADD MUL RET value=42
Before: ctrl memory could send COMPLETE when the step counter reached the end.
After: CtrlMemDynamicRTL waits for the RET response from the element path and forwards that result to the controller.

I agree with your proposal. For implementation, I think we should simultaneously check whether the opcode of current send_ctrl is OPT_RET and the predicate of its execution result is 1.

hmm, I suggest we keep the original solution as well. i.e., it can also send COMPLETE if total_ctrl_steps_val is reached. It already check whether it is non-zero:

elif ((s.total_ctrl_steps_val > 0

I thought including this total_ctrl_steps_val won't introduce any bug..

tancheng requested review from tancheng and yyan7223 June 22, 2026 06:43

tancheng reviewed Jun 22, 2026

View reviewed changes

Comment thread fu/single/CompRTL.py Outdated

tancheng requested a review from ShangkunLi June 22, 2026 06:52

guosran force-pushed the fix/kernel-termination-return branch from 7726455 to 83e1c87 Compare June 23, 2026 20:41

This was referenced Jun 24, 2026

[P0] Conv test numerical value mismatch #292

Open

Fix/comp terminal predicate rtl #305

Open

guosran force-pushed the fix/kernel-termination-return branch from 83e1c87 to 6358200 Compare June 25, 2026 14:43

guosran changed the base branch from kernel-submit to master June 25, 2026 14:52

guosran marked this pull request as draft June 25, 2026 14:53

Fix dynamic ctrl termination

5904e1a

guosran force-pushed the fix/kernel-termination-return branch from 6358200 to 5904e1a Compare June 25, 2026 19:52

guosran marked this pull request as ready for review June 26, 2026 00:32

yyan7223 reviewed Jun 27, 2026

View reviewed changes

Comment thread mem/ctrl/CtrlMemDynamicRTL.py

@update

def update_send_ctrl_val():

def update_send_ctrl():

yyan7223 Jun 27, 2026

Copy link
Copy Markdown

Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why?

yyan7223 reviewed Jun 27, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix/kernel termination return#297

Fix/kernel termination return#297
guosran wants to merge 1 commit into
tancheng:masterfrom
guosran:fix/kernel-termination-return

guosran commented Jun 19, 2026 •

edited

Loading

Uh oh!

Uh oh!

guosran commented Jun 26, 2026

Uh oh!

yyan7223 Jun 27, 2026

Uh oh!

yyan7223 Jun 27, 2026

Uh oh!

yyan7223 Jun 27, 2026

Uh oh!

yyan7223 Jun 27, 2026

Uh oh!

yyan7223 Jun 27, 2026

Uh oh!

yyan7223 Jun 27, 2026

Uh oh!

yyan7223 left a comment

Uh oh!

tancheng commented Jun 27, 2026

Uh oh!

yyan7223 commented Jun 27, 2026

Uh oh!

yyan7223 commented Jun 27, 2026

Uh oh!

tancheng commented Jun 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

guosran commented Jun 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Changes

Uh oh!

Uh oh!

guosran commented Jun 26, 2026

Uh oh!

yyan7223 Jun 27, 2026

Choose a reason for hiding this comment

Uh oh!

yyan7223 Jun 27, 2026

Choose a reason for hiding this comment

Uh oh!

yyan7223 Jun 27, 2026

Choose a reason for hiding this comment

Uh oh!

yyan7223 Jun 27, 2026

Choose a reason for hiding this comment

Uh oh!

yyan7223 Jun 27, 2026

Choose a reason for hiding this comment

Uh oh!

yyan7223 Jun 27, 2026

Choose a reason for hiding this comment

Uh oh!

yyan7223 left a comment

Choose a reason for hiding this comment

Uh oh!

tancheng commented Jun 27, 2026

Uh oh!

yyan7223 commented Jun 27, 2026

Uh oh!

yyan7223 commented Jun 27, 2026

Uh oh!

tancheng commented Jun 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

guosran commented Jun 19, 2026 •

edited

Loading