Repository navigation
Fix LDG MMIO GPU semantics with an Omega-owned oracle control - #172
Merged
Merged
Conversation
Signed-off-by: Drake Stapleton <[email protected]>
aien-spark
force-pushed
the
fix/sm121-ldg-mmio
branch
from
October 28, 2026 18:13
0f96bf4 to
3c16328
Compare
aien-spark
marked this pull request as ready for review
October 29, 2026 17:49
# Conflicts: # src/cudapp_om_units.c # src/cudapp_om_units.ident
drakestapleton
left a comment
drakestapleton
left a comment
Contributor
There was a problem hiding this comment.
Classification: ready-to-merge. All checks green on cccf95a, no conflicts, not draft. The encoding fix is host-verified against nvdisasm 13.0.85 (773/773 SASS kernels and 4/4 cubins byte-identical; test-sm121-mmio exact LDG.E.MMIO.GPU with EF refused).
Note: GB10 device/cache behavior is NOT_RUN, so this should not be cited as on-device proof of MMIO ordering. Optional: attach a full make test log on cccf95a (the final replay skipped it; CI host-tests passed).
drakestapleton
left a comment
drakestapleton
left a comment
Contributor
There was a problem hiding this comment.
Classification: ready-to-merge.
- All checks green on cccf95a; not draft; mergeable, clean.
- The encoding fix (
LDG.E.MMIO.GPU, EF refused) is backed by host evidence: 773/773 corpus kernels byte-identical, Omega/C parity, nvdisasm 13.0.85 cross-check, negative controls. - No GPU-behavior claim is made; GB10 device/cache behavior is stated NOT_RUN, so this should not be cited as on-device proof of MMIO ordering.
- Optional: attach a full
make testlog on cccf95a (the final replay was targeted; CI host-tests passed).
— Grokbot
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
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.
Scope
Fixes #171. Separate from Ampere #170. Canonical baseline is
c6977b4906229fe83f41103b7eff65b64a0986d6(merged #169). Semantic implementation is3c163287d27b6e5a5f5b90c15b61ac83d8f26e2d; final host-evidence/docs commit is885e3201.BW_IR_LDG_MMIOpromisesLDG.E.MMIO.GPU; EF is not an alias. The Omega encoder, declarative specification and retained C migration oracle agree on the corrected policy. The seeded Omega fixture owns the exact mnemonic, register, descriptor and address expectation plus wrong-EF control. Historical receipts stay immutable.Final canonical HOST replay
tools/sm121dis/corpus.shcaptures: 773 / 773 SASS kernels, all byte-identical. No added, removed or changed kernels. Every producer succeeded.test-native-artifactcaptures: four / four cubins, all byte-identical, including the copied tampering-control bundle. Both runs: 47 checks, zero failed.test-sm121-mmio: PASS. ExactLDG.E.MMIO.GPU R4, desc[UR2][R8.64]; otherwise-identical EF refused with exit 4.test-om-units: PASS, all committed hosted/freest bindings regenerate byte for byte against the canonical bridge pin.test-om-encoder: PASS; Omega/C parity on 68 golden instructions and 14 fixture programs; flipped-bit negative refused.test-om-spec-encoder: PASS, including semantic and raw refusal controls. Initial invocation refused an incorrectly resolved sibling generator; explicitOMEGA_BRIDGEenvironment fixed the invocation without changing code or pins. Both logs retained.Compiler SHA256
2c82480596bbdc50cda9626866209ef662d5834f278f839ff4ce0250c495029f(committed seed origin omega-mve5b32567a7aff3ce594e3254a622a846a6cd190b). Local matching binary resides in an omega checkout atf1b9ecb5d13f3b7303c055a187fff06e802eb952.Canonical bridge source
0c4cb9ff980af278890b231b994eaede4dc7621c; native generator SHA256d3cfe0561696aad92164ddf11340522645cf6daf7af238835c3d3c61ecdef656.Read-only NVIDIA nvdisasm 13.0.85, binary SHA256
bc40070d596fa49b81c0905ca1d05e457aaec071280f742997d4a0b511781b25.Evidence and review
evidence/CUDAPP-SM121-MMIO/c697-final-host/corpus-impact.jsonrecords exact source hashes, complete inventories, archive identities, commands/statuses and log digests. Neighboring logs contain the final execution evidence;review.jsonis the scoped source review (not an independent review).The original
evidence/CUDAPP-SM121-MMIO/corpus-impact.jsonremains unchanged historical e843 evidence and is not relabeled canonical. Repository-local sm121dis README updated. No Omega harness guide affected: no language, compiler, ABI, bridge behavior, ownership or authority change.Limits
Existing corpus does not cover op21; the exact-op fixture covers that gap. Generated binding/host archive bytes change; no captured existing kernel/cubin changes. This is not a claim about arbitrary external callers.
Complete
make testsuite: NOT_RUN in this targeted final replay. Remote CI is separate. GB10 cache behavior/device execution and RTX execution: NOT_RUN. No GPU opened, no live model or service touched, no deployed artifact replaced. Draft only; no merge or deployment requested/performed.