[minseobshin11]

Motivation

Range replay restores the coarse-grained device snapshot before every pass. The documented scope already excludes unified/managed, host, fine-grained and virtual-memory-mapped allocations, and said such ranges are "not detected — see the divergence check". But the divergence check hashes only snapshot regions, so it cannot see this memory. Those ranges replayed with REPLAYED, and a kernel writing that memory applied its writes once per pass. A counter incremented 4 times in the application's run read 10 after 3 passes, with divergence 0.

Technical Details

When a dispatch is recorded inside a range, its explicit kernel arguments are checked. The trailing 256-byte HIP hidden-argument block is skipped. An argument that points at GPU-writable memory outside the snapshot declines the range with the existing SNAPSHOT_FAILED, and a warning names the memory kind. This covers host-pinned, managed, locked, fine-grained, another agent's memory, virtual-memory mappings, and pageable host memory when the runtime reports XNACK enabled. The application's own run is unchanged.

What the opt-in divergence check (ROCPROF_RANGE_REPLAY_VERIFY=1) catches on the PR head, compared with this guard:

Case	Divergence check (PR head)	This guard
Kernel writes hipHostMalloc memory	Missed (divergence 0, result 10 instead of 4)	Declines
Kernel writes managed memory	Missed (divergence 0)	Declines
Kernel writes hipMallocAsync memory	Missed (divergence 0)	Declines
Kernel writes fine-grained device memory	Missed (divergence 0)	Declines
Kernel writes another GPU's memory	Missed: not in this agent's snapshot	Declines
OpenMP zero-copy under XNACK	Missed (divergence 0, application check fails)	Declines
Small hipMemcpyAsync into pinned host memory	Missed: host memory is not hashed	Declines
A wrong untracked value later written into a hipMalloc buffer	Caught: that buffer's hash differs	Declines, if the untracked memory was a kernel argument
Pointer to untracked memory stored inside another buffer	Caught only if the effect reaches snapshot memory	Not seen
Host recomputes an input inside the range	Caught only if the effect reaches snapshot memory	Not seen

The first six rows were measured with the check on. The others follow from what each check reads. The guard acts before any replayed pass. The divergence check reports after the passes, when the application's result is already wrong, so it remains the diagnostic for what the guard cannot see.

No new public status, environment variable, kernel replay change, or memory tracker change. The SNAPSHOT_FAILED comment, the header warning and the conceptual docs describe the check.

Device-side malloc replays correctly: HIP's device heap is allocated through the tracked allocator, so the snapshot covers it. Growing the heap inside a range declines ALLOCATION_CHANGED_IN_RANGE through the existing allocation check.

Test Plan

• New GPU-free unit test range-replay-untracked-memory-test: memory classification, hidden-argument skip, and decline decisions.
• Replay CTests on MI300X (gfx942).
• Small HIP programs that increment a buffer inside a range, run with 3 passes and checked against the run without the tool: hipMalloc, hipHostMalloc, hipMallocManaged, hipMallocAsync, fine-grained and uncached device memory, another GPU's fine-grained memory, small hipMemcpyAsync in both directions, device malloc within and beyond the initial heap, and OpenMP zero-copy under XNACK.
• 12-application correctness matrix (HIP and OpenMP builds of BabelStream, XSBench, RSBench and miniBUDE).

Test Result

Case	PR head	This change
hipMalloc	Replays, correct	Replays, correct
Device malloc within HIP's heap	Replays, correct	Replays, correct
Device malloc that grows the heap	Not run	Declines ALLOCATION_CHANGED_IN_RANGE, correct
hipHostMalloc, hipMallocManaged, hipMallocAsync	Replays, wrong	Declines, correct
Fine-grained and uncached device memory	Replays, wrong	Declines, correct
Fine-grained memory on another GPU	Replays, wrong	Declines, correct
Small hipMemcpyAsync (4 KiB, both directions, pageable and pinned)	Replays; pinned device-to-host wrong	Declines, correct
OpenMP zero-copy under XNACK	Replays, application check fails	Declines, application check passes
OpenMP device malloc	Process aborts after the 60 s drain timeout	Same, unchanged
PyTorch inference step, 2 or more passes	Process aborts with a GPU fault	Same, unchanged

• Replay CTests: 189 of 189 pass. The kernel-replay validate tests flake under ctest -j8 on the PR head too and pass when rerun alone.
• Application matrix: identical statuses to the PR head, except BabelStream (HIP) with dot, which now declines because its partial-sum buffer is hipHostMalloc memory. Every application's own check passes.
• The two aborts reproduce on the PR head and are not addressed here.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.