InsertSync: ping/pong multibuffer support#196
Open
TaoTao-real wants to merge 7 commits intozhangstevenunity:mainfrom
Open
InsertSync: ping/pong multibuffer support#196TaoTao-real wants to merge 7 commits intozhangstevenunity:mainfrom
TaoTao-real wants to merge 7 commits intozhangstevenunity:mainfrom
Conversation
TaoTao-real
changed the title
Contributor
Author
|
Remote NPU validation (Ascend910, CANN 8.5.0, pto-isa 08a50c5):
Generated C++ now selects between ping/pong addresses (int64) and builds a loop-local tile via |
Contributor
Author
|
Checked and reconciled local multibuffer work with this PR.\n\n- PR head branch is already at local latest commit (no extra local multibuffer commits pending integration).\n- The subset-based handwritten multibuffer regression is already included ().\n- Local sample check passed for the two multibuffer Sync cases:\n - \n - \n\nNo additional code update is required for PR #196 at this point. |
Contributor
Author
|
Follow-up (fixed formatting): Checked and reconciled local multibuffer work with this PR.
No additional code update is required for PR #196 at this point. |
TaoTao-real
changed the title
- Add pto.multi_buffer=2 attr plumbing into PlanMemory (alloc_tile -> memref.alloc). - Detect ping/pong via planned address overlap-matrix and emit dynamic event-id set/wait. - Add EnableMultiBuffer pass to materialize loop-local ping/pong selection. - Add Sync sample + runop guard; fix PTOViewToMemref typed accessor crash for bitcast/treshape.
emitc.call_opaque requires an IntegerAttr placeholder to print SSA operands. Add the operand placeholder for event_id so set_flag/wait_flag receive the dynamic event argument, and extend the Sync multibuf runop guard to catch missing 3rd argument.
- Track view-like alias closures (bind_tile/subview/casts) from multi-address pointer_cast. - Build ping/pong selector in the LCA loop and rematerialize loop-local alias ops so tile allocations also switch ping/pong. - Update multibuf pingpong sample to use alloc_tile + TLOAD/TSTORE so the generated C++ builds on A2/A3 pto-isa.
Materialize ping/pong by selecting between i64 addresses and building a loop-local PointerCastOp. This keeps bind_tile lowering able to trace the defining PointerCastOp and avoids generating C++ that casts Tile<> to __ubuf__ pointers (which breaks A2/A3 compilation). Update the multibuf runop guard accordingly.
The ping/pong base addresses are an implementation detail of PlanMemory. Check for >=2 distinct int64 constants + ternary address selection, rather than hard-coding 0/512.
TaoTao-real
force-pushed
the
codex/multibuffer-pingpong
branch
from
a15da46 to
4183bf4
Compare
FangRui0
reviewed
Mar 9, 2026
| `[` $src_pipe `,` $dst_pipe `,` $event_id `]` attr-dict | ||
| }]; | ||
| } | ||
|
|
Contributor
There was a problem hiding this comment.
新增op务必同时更新文档 docs/PTO_IR_manual.md
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This change adds first-class double-buffer (ping/pong) support to PTOAS so that:
What’s included
pto.multi_buffer = 2 : i32.pto.alloc_tile -> memref.alloc + pto.bind_tilepropagatespto.multi_bufferonto the createdmemref.allocso PlanMemory can observe it.pto.multi_bufferonmemref.alloc(currently only supports value2; other values error out to avoid silent miscompile) and emits multi-addresspto.pointer_cast(addrs=[ping, pong]).eventIdNum=2is enabled only when the planned addresses satisfy a ping/pong overlap-matrix proof (i==joverlaps,i!=jdoes not overlap). Otherwise it safely degrades to single-buffer.pto.set_flag_dyn/pto.wait_flag_dyn(event-id is an SSAindex) + EmitC lowering.pto.pointer_cast(so runtime actually alternates ping/pong). This also avoids emitting Tile->pointer casts that are not guaranteed to compile on A2/A3 toolchains.test/samples/Sync/test_inject_sync_multibuf_pingpong.py+runop.shregression guard locking in:pto.bitcast/pto.treshapeafteralloc_tilelowering inPTOViewToMemref.Testing
PYTHONPATH=... bash test/samples/runop.sh -t SyncPYTHONPATH=... bash test/samples/runop.sh allNotes
pto.multi_buffer = 2for the first version.