Skip to content

ATOMS.INC / ATOMS.DEC are a compare-and-swap loop; ATOMS.CAS gives the old value - #1413

Closed
Copak731 wants to merge 1 commit into
seedds:mainfrom
Copak731:shared-atomics
Closed

ATOMS.INC / ATOMS.DEC are a compare-and-swap loop; ATOMS.CAS gives the old value#1413
Copak731 wants to merge 1 commit into
seedds:mainfrom
Copak731:shared-atomics

Conversation

@Copak731

@Copak731 Copak731 commented Nov 5, 2026

Copy link
Copy Markdown
Contributor

What and why

  • ATOMS.INC / ATOMS.DEC (operations 3 and 4, CUDA's atomicInc / atomicDec): AAtoms hands them to EmitAtomicOp, which has no case for them; it logs "Invalid atomic operation" and gives the constant 0, with no access to shared memory. They are now a compare-and-swap loop (INC: old >= b ? 0 : old + 1; DEC: old == 0 || old > b ? b : old - 1), Rd the old value.
  • ATOMS.CAS's result. AAtomsCas gave every compare-and-swap CAST.SPIN's 1 / 0. nvdisasm names the forms by bit 87: clear is CAS (CUDA's atomicCAS, PTX atom.cas: Rd the old value); set is CAST, and CAST.SPIN with bit 88 too. CAST and CAST.SPIN keep 1 / 0, so Fire Emblem Fortune's Weave's retry loop is unchanged.

Reference entry: docs/reference/shaders.md, "ATOMS.INC / ATOMS.DEC, and ATOMS.CAS's result".

Checked

  • Tests: DMSH_TEST_XLATE=<translator> python3 dm/gfx/tests/test_dmsh_ampere_alu.py, new SharedAtomics (3, all fail on main). All test_dmsh_*.py pass.
  • GPU cross-check: 400 LDS / STS / ATOMS / ATOMS.CAS from the corpus, each as a one-thread compute shader (inputs and outputs through an NVN2 storage buffer, LDC.64 c[0][0x10] then LDG.E / STG.E; only the 16-byte block the access touches is set and compared), production options, MSL 2.3: 84 mismatches on main, all INC, DEC and CAS, to none; the same on encodings from real compiled code (CUDA sm_80 / sm_75 binaries). Everything else (LDS of every size, STS, ADD / MIN / MAX / AND / OR / XOR / EXCH) already matched.
  • Devices and OS: MacBook, Apple M5, macOS 26.6.2, Xcode 27.2.
  • Games run: none (no game files were used).

Docs

  • The finding is in its docs/reference/shaders.md entry, and python3 dm/tools/docs_symptoms.py was run (the symptom index)
  • No game files, keys or data extracted from a game in the change or the attachments

🤖 Generated with Claude Code

@Copak731 @claude
ATOMS.INC / ATOMS.DEC are a compare-and-swap loop; ATOMS.CAS gives th…
923e491 
…e old value

AAtoms handed operations 3 and 4 to EmitAtomicOp, which has no case for them
(the constant 0, no access). AAtomsCas gave every compare-and-swap CAST.SPIN's
1 / 0; nvdisasm names three forms by bit 87, and a plain CAS (bit 87 clear)
gives the old value. CAST and CAST.SPIN keep 1 / 0.

New SharedAtomics tests in test_dmsh_ampere_alu.py (three fail on main). A
GPU cross-check of 400 shared accesses as one-thread compute shaders with the
production options: 84 mismatches (INC, DEC, CAS) to none.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@github-actions github-actions Bot added nvn NVN on Metal and its tests (dm/gfx/nvn, dm/gfx/tests) docs Documentation labels Nov 5, 2026
fpedrolucas95 pushed a commit to fpedrolucas95/nxnative that referenced this pull request Nov 6, 2026
@seedds @fpedrolucas95
Merge seedds#1413 into upstream-prs: ATOMS.INC / ATOMS.DEC are a comp…
3d6cfbb 
…are-and-swap loop; ATOMS.CAS gives the old value
fpedrolucas95 pushed a commit to fpedrolucas95/nxnative that referenced this pull request Nov 6, 2026
@seedds @claude @fpedrolucas95
The translator rebuilt from the merged source (seedds#1343, seedds#1376…
236cc17 
…, seedds#1410, seedds#1411, seedds#1412, seedds#1413, seedds#1414, seedds#1329 on top of main's translator fixes): chain c9cf9613657e

Publishes Ryujinx.Graphics.Shader, Ryujinx.Common, Spv.Generator and dmsh_xlate (osx-arm64, framework-dependent, the universal launcher kept) from the merged source; the
PRs' own binaries were built without this tree's translator changes (float atomic adds, F2IP, the growing raw pool table). The lock and dmsh_chain.h are written again.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
seedds added a commit that referenced this pull request Nov 7, 2026
@seedds @Copak731
ATOMS.INC / ATOMS.DEC are a compare-and-swap loop, and ATOMS.CAS give…
1a410e6 
…s the old value: shared-memory counters move as on Ampere (chain 1ddb83a9817a, PR #1413)

AAtoms handed operations 3 and 4 (CUDA's atomicInc / atomicDec) to EmitAtomicOp, which has no case for them: it logged "Invalid atomic operation" and gave the constant 0, with no access to shared memory. They are now the compare-and-swap loop ATOMG's INC / DEC already are on main (INC: old >= b ? 0 : old + 1; DEC: old == 0 || old > b ? b : old - 1), Rd the old value. AAtomsCas gave every compare-and-swap CAST.SPIN's 1 / 0; nvdisasm names the forms by bit 87: clear is CAS (atomicCAS, Rd the old value), set is CAST, with bit 88 CAST.SPIN, which keep 1 / 0, so Fire Emblem Fortune's Weave's retry loop is unchanged.

Checked: test_dmsh_ampere_alu SharedAtomics (3 tests, all fail on the translator before) and the 21 translator test files (test_dmsh_pairs' two NameErrors as before), test_nsp_chain, test_xlate_anycpu, test_docs. The author's GPU cross-check: 400 LDS / STS / ATOMS / ATOMS.CAS from the corpus as one-thread compute shaders, 84 mismatches on main (all INC, DEC and CAS) to none, the same on encodings from CUDA sm_80 / sm_75 binaries. No game is known to depend on it. Regression: one Ocarina of Time run on the chain that ends this batch.

Co-authored-by: Copak731 <224629733+Copak731@users.noreply.github.com>
@seedds

seedds commented Nov 7, 2026

Copy link
Copy Markdown
Owner

Landed as 1a410e6 with you as co-author (SharedAtomics fails before, passes after; main already had ATOMG's INC / DEC). Thank you.

@seedds seedds closed this Nov 7, 2026
Aznracer pushed a commit to Aznracer/Mighty-Green-Kokiri-Kid that referenced this pull request Nov 8, 2026
@seedds @Contributor12
ATOMS.INC / ATOMS.DEC are a compare-and-swap loop, and ATOMS.CAS give…
05a4cb7 
…s the old value: shared-memory counters move as on Ampere (chain 1ddb83a9817a, PR seedds#1413)
Know-It-All-Brother pushed a commit to Know-It-All-Brother/nxnative that referenced this pull request Nov 9, 2026
@seedds @Copak731
ATOMS.INC / ATOMS.DEC are a compare-and-swap loop, and ATOMS.CAS give…
d520bd8 
…s the old value: shared-memory counters move as on Ampere (chain 1ddb83a9817a, PR seedds#1413)

AAtoms handed operations 3 and 4 (CUDA's atomicInc / atomicDec) to EmitAtomicOp, which has no case for them: it logged "Invalid atomic operation" and gave the constant 0, with no access to shared memory. They are now the compare-and-swap loop ATOMG's INC / DEC already are on main (INC: old >= b ? 0 : old + 1; DEC: old == 0 || old > b ? b : old - 1), Rd the old value. AAtomsCas gave every compare-and-swap CAST.SPIN's 1 / 0; nvdisasm names the forms by bit 87: clear is CAS (atomicCAS, Rd the old value), set is CAST, with bit 88 CAST.SPIN, which keep 1 / 0, so Fire Emblem Fortune's Weave's retry loop is unchanged.

Checked: test_dmsh_ampere_alu SharedAtomics (3 tests, all fail on the translator before) and the 21 translator test files (test_dmsh_pairs' two NameErrors as before), test_nsp_chain, test_xlate_anycpu, test_docs. The author's GPU cross-check: 400 LDS / STS / ATOMS / ATOMS.CAS from the corpus as one-thread compute shaders, 84 mismatches on main (all INC, DEC and CAS) to none, the same on encodings from CUDA sm_80 / sm_75 binaries. No game is known to depend on it. Regression: one Ocarina of Time run on the chain that ends this batch.

Co-authored-by: Copak731 <224629733+Copak731@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

docs Documentation nvn NVN on Metal and its tests (dm/gfx/nvn, dm/gfx/tests)

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@Copak731 @seedds