Program 203 reuses prefixSumCache for a second subgroup reduction before every workgroup invocation has consumed the first result. Add a fingerprint-gated SPIR-V pass that inserts the missing Workgroup acquire-release barrier while preserving native subgroup operations.
Keep the repair opt-in behind MOBILEGL_ITERATIONRP_FIX_BARRIER, cover insertion, pass-through, and idempotence, and enable it together with the existing iterationRP subgroup repairs for the matching Linux and Android CI retraces.
The previous commit's fingerprint was pinned to one array's incidental
dimensions - workgroup exactly 32x16x1, element exactly vec2, length
exactly 32 - which is the auto-exposure reduction and nothing else. The
pack ships the same idiom twice:
- auto-exposure: 32x16 (512 invocations), shared vec2 prefixSumCache[32]
- RTW warp: 1024 invocations, shared float prefixSumCache[64]
so the warp kept writing 128 subgroups into 64 entries on an 8-lane
device and the retrace stayed bit-identically wrong (ssim 0.027902).
Key the fingerprint on the pack's idiom instead of one array's shape: a
workgroup array of 32-bit floats indexed by gl_SubgroupID, fed by a
subgroup scan, whose declared length is below ceil(invocations / native
width). Three properties keep that a targeted repair rather than a
general array resizer:
- the index must BE gl_SubgroupID (through OpCopyObject, a signedness
OpBitcast, or a spill whose every store is that id), so an index
masked or clamped into range is left alone;
- the >= 16-lane early-out is retained, so every module on the devices
the pack was written for passes through byte-identical;
- growth is certified against maxComputeSharedMemorySize using a
natural-alignment layout model, and declined outright when a
declaration cannot be sized, so a patched module can never fail
pipeline creation where the original would not have.
Verified against the shaders the CI trace actually contains: of the 14
compute modules in the fixture exactly these two change, the other
twelve are byte-identical, and all fourteen pass spirv-val. The
integration scenario grows a second case for the 1024-invocation shape;
both abort with heap corruption when the patch is disabled.
iterationRP's Program 203 declares shared vec2 prefixSumCache[32] for a
512-invocation workgroup indexed by gl_SubgroupID; any device narrower
than 16 lanes partitions into more than 32 subgroups and the pack writes
shared memory out of bounds (heap corruption on lavapipe's CPU
rasterizer, ssim 0.028 on the CI retrace). Fix it where the fault lies -
in the fixture - and keep the GL contract sound everywhere else:
- FixIterationRPSubgroupScratchPass: fingerprint-gated SPIR-V pass that
grows exactly that array to ceil(invocations/width) entries on sub-16-lane devices; every other module passes through byte-identical.
- DeriveNumSubgroupsPass stays default-on for the Adreno topology bug
and is made spec-sound: pipelines request REQUIRE_FULL_SUBGROUPS
whenever the workgroup shape makes the flag legal (computeFullSubgroups
enabled, local_size_x a multiple of the native width, subgroup count
within maxComputeWorkgroupSubgroups).
- EmulateSubgroupsPass: 32-lane virtual-subgroup lowering kept in-tree
as a last resort, enabled only by MOBILEGL_MAGMA_EMULATE_SUBGROUP=1 on
devices with no native subgroup support; fails closed on extended
subgroup instructions and on modules whose added scratch would exceed
maxComputeSharedMemorySize.
- IterationRPFirstReductionScenario skips gracefully outside the pack's
16..256-lane source domain; the new IterationRPScratchFixScenario runs
the fixture-shaped reduction on any width and asserts the exact
width-independent total. DriverPost keeps reporting FAIL on
out-of-domain devices.
- Program203 -> IterationRP rename throughout; the per-trace
num_subgroups_quirk plumbing is removed from the trace replayer, JNI
chain, and CI workflows.
glCompileShader with MOBILEGL_ASYNC_SHADER_COMPILE=1 snapshots its inputs on
the GL thread (source SharedPtr, CompileEnv, cache handle) and runs the whole
pure pipeline - preprocess, validators, extractors, glslang parse - as a
ShaderCompileTask on the worker pool, returning immediately. Every read of
compile-produced state joins through the single Compiled() gate; links stay
synchronous this stage and join their attached shaders at the top of the
body. Flag off, the path is the same code run inline.
Mechanics: the job node owns all its inputs (no back-pointer, no lifetime
tie to the shader object), so re-sourcing or deleting a pending shader is
cancel-and-drop, never a wait; glslang worker hygiene is a TLS-allocator
scope guard plus GL-thread builtin prewarm (gated on the flag, latch reset
on Destroy so re-initialization re-warms); worker-side diagnostics defer
through the job and replay on the GL thread at the join, enforced by
IsPoolThread asserts in RecordError and an empty-deferred-errors tripwire.
A body that throws publishes a COMPLETE failed compile (status false, real
info log) rather than an abandoned node, and never memoizes away the retry;
a failed enqueue (OOM) cancels the node instead of stranding the joiner -
including inside the dispatch loop, where the in-flight slot is repaid.
The pool StopAndDrains from an atexit sentinel too: workers still inside
glslang parse while exit() ran static destructors was a real 2-in-5 SIGSEGV,
reproduced and fixed (15/15 clean after).
Backend-internal shader objects (default FS, DirectVulkan blit/mipmap) are
cache-less and always compile inline - compile-and-read-in-one-breath needs
no round trip.
Gates: unit suite 488/488 with the flag off AND on (x5); AsyncCompileTest
(12 e2e cases: pending re-source/delete/recompile, byte-identical failure
logs across modes, 48-compile cache stress) x10 repeats clean both modes;
full NVIDIA DirectGLES retrace identical result sets flag off/on (zero new
deltas); compile-phase timing flat as designed (links still serial - the
parallel win arrives with stage 4's async link + stage 5's
KHR_parallel_shader_compile).
Still fully synchronous - EnsureLinkJoined()/EnsureCompileJoined() are empty
inline no-ops (verified to fold away at every one of the ~1200 call sites;
this project builds without LTO) - but every read of link- or compile-produced
state now goes through a private accessor the compiler enforces, so when
stage 4 moves the bodies onto pool workers, 'which reads must join' is a
type-system fact instead of a 400-line audit.
- ProgramObject: the 31 fields ResetLinkArtifacts clears plus the 5 link
outputs it forgot (infoLog, linkedFragData{Location,Index}, the geometry
strip-capture pair) move into a nested LinkArtifacts behind Artifacts().
ResetLinkArtifacts is now a worker-safe pure clear; the link-observable
version bumps (backendState/link/uboContent) move to a GL-thread-only
BumpLinkObservableVersions() called once from Link()'s prologue and from
glProgramBinary's mandated failure - the link body never writes them, so
a stage-4 worker cannot lose an invalidation against the draw path.
- ShaderObject: compile artifacts (TShader, preprocessed source, side-channel
maps, status/log, consume-once flag) behind Compiled(); the P0b layer-1
memo trio deliberately stays outside as the future non-joining
COMPLETION_STATUS_KHR fast path.
- CompileEnv (new): a GL-thread snapshot of everything the compile pipeline
used to read live from the backend mid-parse - compute limits (the
GetIntegeri_v reach-back is gone from the worker path), advertised
extensions, device quirks, TBuiltInResource inputs. Captured lazily per
backend activation; the consume-once re-parse now runs against the same
env as the original parse.
- The GL-thread prologue / worker-body boundary is marked in Link() where
the stage sort ends; everything below is a pure function of the snapshot.
Public getter signatures unchanged - MG_Impl and both backends compile
untouched. Unit 476/476, Program suites 117/117, DirectGLES retrace 38/39 on
llvmpipe (the one failure is the known pre-existing non-CI iterationrp case;
the NVIDIA userspace driver was updated out from under the running kernel
module mid-session, so GLX there is down until a reboot).
RenameBuiltinShadowingFunction probed the whole source ten times per compile
(5 names x mask + scan, each a full-text pass) and still had two blind spots:
a 5-name list and single-line-definition-only detection. On Complementary-scale
packs (4.5MB of sources) that was ~68% of the compile phase.
The rename is now split by FAILURE LAYER, both halves sharing one name table
header so they cannot drift:
- A SPIR-V OpName pass in SanitizeAndOptimizeBinary covers the full ESSL 3.20
builtin set (~146 names). Renaming a function id is safe by construction:
builtin calls are GLSL.std.450 instructions and can never resolve to a user
OpFunction, overloads are distinct ids (a helper overload delegating to the
real builtin keeps working), dead preprocessor branches never reach SPIR-V,
and macro-expanded definitions are covered. ESSL 3.x is the only consumer
that forbids the redefinitions, and this pass runs before its transpile.
- A lexical pass covers only the 5 names whose exact-signature redefinitions
glslang's relaxed parse rejects outright (never producing SPIR-V for the
backstop): the historical fma/max3/min3/round/tanh. One TokenizeCode pass;
definition detection requires brace depth 0, a type-identifier previous
token that is neither a statement keyword nor a directive tail, and skips
files whose token-level braces do not balance (preprocessor-asymmetric
arms) - over-detection is unrecoverable, so every ambiguity falls through
to the backstop.
Measured on the compile phase (prefix-diff, 3-run medians, Espryt/NVIDIA):
complementary-reimagined 20.0s -> 5.5s, BSL 2.14s -> 1.85s. bliss (the pack
that ships from-scratch fma/tanh helpers) stays at SSIM 0.999962.
Tests: end-to-end ESSL assertions for the multiline-definition and
new-overload shapes, the three adversarial-review reproductions (statement-
keyword call under asymmetric braces, dead-#if compat shim, overload
delegating to the shadowed builtin), and a source-level assertion pinning
the lexical half specifically.
swung0x48
Swung0x48