Repository navigation
fix(recompiler): defer the TYPE check of a guarded read-only static V# - #6412
Draft
fdfawef324 wants to merge 5 commits into
Draft
fdfawef324 wants to merge 5 commits into
fdfawef324 wants to merge 5 commits into
Conversation
11 tasks done
|
Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.
Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.
|
11 tasks done
11 tasks done
11 tasks done
12 tasks done
12 tasks done
Open
11 tasks done
11 tasks done
fdfawef324
force-pushed
the
claude/agc-defer-static-vsharp-type
branch
from
October 16, 2026 19:52
6e89aba to
6b9410b
Compare
Contributor
Author
|
We rebased this PR onto
Tests, CPU: GPU, on the RTX 3060 (driver 591.86, Windows Server 2025), run on the build of 677b9af (before the label commit, which touches no GPU path): AI-assisted: yes |
Contributor
|
Coordination question from #5879: the Kena capture is a fragment-stage case with stale, unwritten user-data registers; #6412 is explicitly compute-only and its current eligibility excludes pixel programs. I found no other open PR that defers this path. If #6412 lands, would you prefer a separate follow-up to extend the planner/materializer/fault handling to fragment shaders, or are you already considering that scope? I will avoid opening overlapping work until the boundary is clear. The fragment case would need its own regression for an inactive guarded read and an active read that faults, and must preserve the current failure behavior for stores/atomics and unsupported formats. |
This was referenced Oct 17, 2026
12 tasks done
github-actions
Bot
added
the
unblocks title
Contributor
Author
|
Please keep the fragment-stage/Kena case from #5879 as a separate follow-up. #6412 remains compute-only: changed-EXEC plain buffer loads, no LDS/barrier, and the existing BDA fault and dispatch-reuse restrictions. I am keeping that eligibility boundary while rebasing this PR. You can proceed with the fragment work independently. It needs its own active/inactive-lane and completion-fault coverage; the compute tests here do not establish fragment behavior. AI-assisted: yes |
A compute shader can leave a V# whose TYPE (word3[31:30]) is not 0 in a static buffer slot that it reads only behind a lane-dependent guard, for example v_cmpx followed by s_cbranch_execz. The capture refused every such V# with "buffer descriptor uses an unsupported type", so the whole dispatch threw even when no lane runs the read. A static buffer is now eligible for a deferred TYPE check when: - the shader is a compute shader with no LDS and no s_barrier; - it reads the buffer only through plain buffer_load_dword, _dwordx2, _dwordx3 or _dwordx4, under an EXEC the program changed; - no such load is in a loop's continue block; - the target has the BDA fault ABI. The capture plan and the compiled variant both decide eligibility from the IR. Planning-only scalar SRT reads are skipped, as ResourceTracker skips them, and a program that cannot defer skips the walk. For an eligible V# of any nonzero TYPE, capture records an UnsupportedType status instead of throwing. The status is carried into the binding plan, its key and the snapshot hash. On each lane that runs the load with EXEC set, the load records an InvalidDescriptor fault (reason 8) and stops the invocation. Lanes with EXEC clear, and loads that are branched over, do nothing. Scalar, formatted, typed and sub-dword loads, stores, atomics, and loads under the entry EXEC still throw at capture. The RDNA 2 ISA requires TYPE 0 for a buffer (section 8.1.8, Table 37) and does not say what a load through any other TYPE does. None of this is measured on AMD hardware; the shader debt entry records it. CompiledShaderArtifact::deferredTypeFaultsOnly marks a fault buffer that exists only for these checks. The disk cache moves to format 37. Co-Authored-By: Codex <noreply@openai.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…onzero For a guest buffer element whose captured V# has the UnsupportedType status, the driver: - binds the empty placeholder buffer; - requires the compute fault contract; - turns fault checks on, so a lane that ran the load fails the dispatch at completion; - makes the resource object never reusable and refuses to move its buffers; - adds the status to the content key. A shader whose fault buffer exists only for deferred TYPE checks (deferredTypeFaultsOnly) leaves fault checks off while every captured V# has TYPE 0, unless APS5_LOOP_GUARD is set. Those dispatches keep GPU-signalled labels and the dispatch-recipe path. Co-Authored-By: Codex <noreply@openai.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- agc_shader_memory (verifyDeferredStaticBufferTypes): a buffer_load_dword behind v_cmpx/s_cbranch_execz through a V# of TYPE 1, 2 or 3 captures with the UnsupportedType status, a fault buffer and a deferred-only marker. A TYPE 0 capture carries no status. The same read under the entry EXEC, or on a target without the BDA fault ABI, is still refused. - agc_shader_deferred_static_planning_reads: a planning-only scalar read of register quad 3 or 0 neither indexes past the static buffers nor keeps the read-only buffer from deferring. - agc_driver_guarded_null_pointer runs the shader on the GPU. With TYPE 0, the lane that passes the guard reads the buffer. With TYPE 2 or 3 and no lane passing, the dispatch completes without a fault. With one lane passing, completion reports "BDA access failed" with reason 8 at the load's pc. - agc_driver_graphics (BdaResources): a deferred-only fault buffer is polled only when a captured V# has a nonzero TYPE; a shader with other fault sources is still polled. - agc_shader_disk_cache round-trips the binding status, the buffer flag and the artifact and info flags. The shader debt entry cites RDNA 2 ISA sections 8.1.8 and 8.2.6 and records what is not measured on AMD hardware: whether an executed load through such a V# faults, returns zeros or reads through the base, and whether lanes with EXEC clear are unaffected. The driver debt entry records the cost of the fault buffer. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ShaderResources::ContentKey writes each binding's bufferDescriptorStatus count and entries between the bufferWritten bits and the sampler count, but LabelShaderKey, which ShaderResources::DiffContentKeys uses to name what changed between two cache keys, did not know that word. It read the status count as the sampler count and the sampler count as the shape count, so every binding after the first was parsed from the wrong offset and a real key came back unlabelled or with the wrong component. LabelShaderKey now labels the status count and its entries as Other, like the neighbouring bufferWritten bits. KeyMiss.cpp builds its hand-built keys with the status word, and a new test checks them word for word against ShaderResources::ContentKey for dispatch, table and status-carrying bindings, so the hand-built layout cannot drift from the real one again. Another new test checks that a changed status entry is labelled Other and that the V#, image and sampler words of the bindings around status entries keep their own labels, for dispatch and draw keys. Without the parser change the first labelled key already fails with "equal keys were not comparable". Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Parse the sampler-coordinate flag vector introduced on main before the deferred descriptor status block. Keep the synthetic key fixture aligned with ContentKey. Co-Authored-By: Codex <noreply@openai.com>
11 tasks done
12 tasks done
fdfawef324
marked this pull request as draft
October 18, 2026 23:31
fdfawef324
force-pushed
the
claude/agc-defer-static-vsharp-type
branch
from
October 18, 2026 23:31
6b9410b to
9585842
Compare
Contributor
Author
|
Rebased the compute-only change onto Fresh results: all twelve Release targets built, nine CPU/mock CTests passed in 19.35 seconds, and eight real Vulkan registrations passed in 24.39 seconds with Vulkan required and no skips. The latter includes skipped/taken invalid static-TYPE loads and expected fault completion. Root hash-checked the exact executables and runtime dependencies before and after. Conventions pass and the merge into main @MarcZX10, fragment-stage/Kena #5879 remains a separate follow-up as discussed. The PR is a draft while the #6177 shared-area landing order is coordinated; that future fast-path composition must preserve active-lane evaluation of the faulting descriptor. No combined runtime result is claimed. AI-assisted: yes. Co-Authored-By: Codex noreply@openai.com |
Contributor
|
I opened #7689 as the separate fragment-stage follow-up for #5879, as you suggested. It adds fragment active/inactive EXEC coverage. I listed #6412 as a dependency because the descriptor-fault API and completion path share implementation points; after #6412 merges, I will rebase the fragment change and preserve the cache format version. |
This branch has not been deployed
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
3 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.
What
A guarded read-only static V# of nonzero TYPE currently rejects a whole compute dispatch during capture even when no lane runs its load. This change captures that descriptor with a status and faults only an active lane that executes the invalid load, using the existing InvalidDescriptor completion contract.
Eligibility remains limited to compute shaders without LDS/shared accesses or barriers, plain DWORD loads under changed EXEC, no continue-block load, and the BDA fault ABI. Stores, atomics, scalar/formatted/typed reads and entry-EXEC loads retain capture rejection. Valid descriptors keep existing completion and recipe paths; invalid captured descriptors force fault polling and prevent resource reuse/moved-buffer recipes. Fragment/Kena #5879 remains a separate follow-up.
This rebase retains current main's plan and serializer fields, moves the cache to format 39 (pinned main 38), and preserves the ten-vector shader-key layout including samplerCoordinatesIgnored before the added descriptor-status block. Active invalid-TYPE hardware behavior remains unmeasured and documented as debt.
Nearby #6177 has a source-supported future-composition risk: its robust-direct fast path can evaluate a faulting descriptor outside the active-lane guard. This is not a reproduced defect in #6177 alone. Combining the branches must retain that guard or exclude deferred descriptors from the fast path and rerun skipped/taken fixtures in robust mode. Proposed order is this PR first; coordination is pending, so this update is a draft. No combined build or landing-order agreement is claimed. Other shared cache/reuse changes must retain the descriptor status in identity and the deferred-fault nonreuse guards.
Tested
Exact candidate
9585842db6b7d62a387fe7d696652fef69050513on pinned mainfc623324a2ab3443cb69a6d0199141b8b2fee604, Windows Server 2025 10.0.26100, MinGW-w64 GCC Release-O3.Nine CPU/mock CTest cases passed in 19.35 seconds:
agc_shader_memory,agc_shader_bindless,agc_shader_deferred_static_planning_reads,agc_shader_disk_cache,agc_shader_control_flow(ordinary, iteration-selection-off and forced-fallback),agc_driver_key_miss, andagc_driver_graphicsincluding BdaResources mocks. No case skipped. All twelve requested build targets pass in Release-O3; the final incremental build took 128.44 seconds. Conventions and diff checks pass against the pinned base.Independent adversarial source review passed at this exact head. The original CPU red mutations and GPU results recorded at the older public stack are historical evidence; they were not rerun on this rebased head. Synthetic project fixtures are retained; no game shader payload is attached.
The previous public body at
6b9410b92bbd9d74bf3f32ebbbd2b91a0be08f18records a historical offline comparison on mainb3f9f89e62979d7328ff0dec2d14b006f8c2c0ab: main refused capture, while an earlier PR stack materialized successfully in three runs. That report does not record the full integration-stack head, and used a replay-tool option from #5959 that is not part of this PR. It is historical preparation/materialization evidence, not current-head or GPU verification. No request or shader payload is attached.Fresh current-head managed-slot Vulkan verification passed all eight CTest registrations in 24.39 seconds:
guarded_null_pointer,bda_device,bda_span_reads,buffer_descriptor_runtime,gated_dispatch_capture(ordinary and short-timeout),loop_ending_exitsandsubvector_loop. All seven unique executables ran withANYPS5_REQUIRE_VULKAN=1; no case or reported subcase skipped. The null-pointer fixture includes both skipped and taken invalid static-TYPE loads, with the expected completion faults. Actual executables and runtime dependencies were hash-checked before and after. A clean merge into main84c1a4d9awas checked separately; tests ran on the pinned candidate. No current title milestone or performance improvement is claimed.Checklist
main; no other open PR implements the same functionsNotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file.mdfiles or images added to the repository (attach them to this PR)🤖 Generated with Claude Code
Co-Authored-By: Codex noreply@openai.com