[danielbolsson]

Motivation

Addresses @LeoZhao-BJ's review comment on #10225.

Region materialisation was wired into kernel-argument translation only. replay_memcpy_impl called translate_ptr directly, so a sidecar-declared allocation whose first touch was a captured H2D, D2H or D2D copy failed as unmapped in the memcpy handler before materialisation could run.

His "breaks the stated behavior" framing is right, and the promise being broken is our own. The comment above advance_to in dispatch_event says the region timeline is advanced there so that memcpys and every other handler get the same view. The advance was shared; the materialisation was not.

Technical Details

The three translate_ptr call sites in replay_memcpy_impl now go through a small translate_or_materialize helper, which mirrors what the whole-pointer kernel-argument path already does: translate, and on failure back the declared segment and translate into it.

The regression test needed real HSA memory

Every other bypassed-allocation test in this file fakes the address — the workload passes probe + kProbeFarOff, which nothing allocated, and no thread dereferences it. That cannot be reused here: a memcpy has to succeed at capture for the event to be recorded at all. The memory has to genuinely exist while genuinely being invisible to HRR.

So the new hidden workload allocates from an HSA coarse-grained device pool. The buffer is real and GPU-reachable but never crosses the HIP dispatch table, so nothing in the archive can explain the address — exactly the property the assertion needs. Two details worth knowing:

• Access is granted to every agent, not just the owning GPU. The copy is serviced by a DMA engine, and on an unshared coarse-grained allocation that engine faults exactly like a stray kernel would. My first attempt granted access to one agent and the capture died with a GPU memory fault.
• The segment is deliberately not passed to any kernel. A kernel argument would let the existing whole-pointer materialisation resolve it, and the test would then pass with or without this fix.

Because an HSA address appears in no hipMalloc record, find_alloc_base cannot recover it, so the child reports the base on stdout and the parent builds the sidecar from that.

This is the first test under catch to include HSA headers, so it adds an optional hsa-runtime64 link. Where the package is absent, the workload and its test compile out and the test reports itself skipped rather than failing an otherwise fine build.

Issue Tracking

AIRUNTIME-2700

Test Plan

The test is only worth anything if it fails without the fix, so that was verified explicitly rather than assumed: the fix was stashed, the suite rebuilt, and the new test run against the unfixed memcpy path.

Test Result

Without the fix — Unit_HRR_Regions_MemcpyFirstTouchMaterialization fails. The sidecar loads and is visible in the log, but the memcpy handler never consults it:

[HRR] H2D dst 0x7fbe01400000 not mapped (size=262144 blob_sz=262144)
[HRR] Fatal: T68 Event 4 (hipMemcpy) returned 1 (invalid argument) — aborting replay
[HRR] Replay aborted due to fatal HIP error — exiting
[HRR] Regions : 1 records from 1 stream(s)

No Region segs line is printed at all.

With the fix — the test passes: Region segs : 1 materialised, replay exits 0, and the D2H blob matches byte for byte.

Full suite: 78/78 passed, 0 failed. GPU-free pass PASS, GPU pass PASS. All five region tests green, including the new one. One skip (Unit_HRR_MemcpyPeerRoundtrip, needs a second device).

Submission Checklist

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

Made with Cursor

[LeoZhao-BJ]

The new code materializes dst_rec unconditionally at the top of replay_memcpy_impl():

That is correct for H2D and D2D device destinations, but for D2H/H2H the destination is a host pointer. Unlike the old translate_ptr() lookup, translate_or_materialize() has side effects: if ctx.regions_enabled and the host pointer value happens to fall inside an annotated SEGMENT range, replay can incorrectly create/register a materialized device allocation and skew alloc_map / Region segs accounting.

Suggested fix: only call translate_or_materialize() inside branches where the operand is known to be a device pointer: H2D dst, D2D dst/src, and D2H src.

[Copilot]

🟡 Changes recommended

The new HSA test helper leaves HSA runtime initialization unbalanced and a new build comment is misleading about “runtime skip” behavior when HSA is absent.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR fixes a replay correctness gap in HRR where bypassed-region materialisation was only triggered via kernel-argument translation, causing captured memcpys (H2D/D2H/D2D) to fail when the first touch of a sidecar-declared segment was a copy rather than a launch.

Changes:

• Route memcpy pointer translation through a new translate_or_materialize helper so sidecar segments can be backed on-demand during memcpy replay.
• Add an HSA-backed regression test that exercises “memcpy first touch” on a real device-accessible allocation that HIP/HRR never observed.
• Document the broader “fails to translate at point-of-use” materialisation rule in the HRR design notes and register the new tests.

File summaries

File	Description
projects/hip-tests/catch/unit/hrr/hrr_regions_test.cc	Adds an HSA-backed hidden workload and a regression test covering memcpy-first-touch segment materialisation.
projects/hip-tests/catch/unit/hrr/CMakeLists.txt	Optionally links hsa-runtime64 and enables the HSA-based HRR region test when available.
projects/hip-tests/catch/config/configs/unit/hrr.yaml	Registers the new direct workload and the memcpy-first-touch materialisation test.
projects/clr/hipamd/src/hrr/playback/hip_playback.cpp	Adds translate_or_materialize and uses it for memcpy replay pointer translation.
projects/clr/hipamd/src/hrr/DESIGN.md	Updates the design doc to reflect materialisation being triggered by translation failures at any use site (including memcpys).

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[systems-assistant]

TheRock Submodule Bump Activity

Newest first

• 2026-09-18 — Included in TheRock via ROCm/TheRock#8300.