[iassiour]

Motivation

HostBlitManager::fillBuffer is slow for large fills because it calls memcpy() once per pattern element. The host fill path is taken when the destination is directly accessible from the host, for example:

• HIP on Windows: hipMemset/hipMemsetAsync on hipHostMalloc memory.
• OpenCL: clEnqueueFillBuffer on host-accessible buffers.

Technical Details

The pattern size is a runtime value, so each loop iteration is a real libc memcpy() call that copies a few bytes. The per-call overhead dominates, and fill time scales with the element count rather than the byte count. A 64 MiB hipMemset makes about 67M calls.

The fill now uses wide copies:

• 1-byte patterns use a single memset().
• Wider patterns are replicated into a 4 KiB stack staging block, which is then copied out in block-sized memcpy() calls. The block is built by doubling and capped at the fill size, so small fills stay cheap.
• Fill patterns are capped at FillMemoryCommand::MaxFillPatterSize (128 bytes), so a pattern always fits in the staging block. A static_assert enforces this.
• The destination is only ever written, never read back, since it may be write-combined memory.

The result of the fill is unchanged, including the existing behaviour when the size is not a multiple of the pattern size.

Issue Tracking

JIRA ID: ROCM-32043

Test Plan

• Standalone host microbenchmark comparing the old and new fill loops on regular host memory, for 1/2/4-byte patterns and fills from 4 B to 1 MiB.
• Byte-for-byte comparison against the old loop for every pattern size from 1 to 128 bytes, across 25 fill sizes from 0 to 1 MiB (including sizes that are not a multiple of the pattern), with an unaligned destination and guard bytes checked for overruns.
• hip-tests memset tests, using a local build in which HIP 1D memsets take the host fill path.

Test Result

Microbenchmark, time per fill (old loop -> new):

Pattern	64 B	4 KiB	64 KiB	1 MiB
1 byte	84 ns -> 2 ns	4.9 us -> 24 ns	78 us -> 0.3 us	1.25 ms -> 5 us
2 bytes	46 ns -> 20 ns	2.5 us -> 66 ns	39 us -> 0.43 us	627 us -> 11.5 us
4 bytes	22 ns -> 16 ns	1.3 us -> 60 ns	21 us -> 0.42 us	333 us -> 11.4 us

The output matched the old loop in all 3,200 comparison cases.

All 41 positive 1D memset tests in MemoryTest1/MemoryTest2 pass with the final code (D8/D16/D32, sync and async, unaligned, small and zero size). The 2D/3D memset, pitch and graph-capture memset tests also passed on the first commit; the second commit only restructures the staging loop.

Not tested on Windows or with OpenCL.

Submission Checklist

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

🤖 Generated with Claude Code

[therock-pr-bot]

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	⚠️ Warning	Error: Source/code files changed without an accompanying unit test. Expected: add at least one test file named like test_<name>.py / test_<name>.cpp (or <name>_test.*). Current: code file(s) changed: projects/clr/rocclr/device/blit.cpp; no test file found
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🚫 Please fix the failed policies before requesting reviews.

The following policy checks failed:

• ❌ PR Description

The Not ready to Review label has been added to this PR.
Once all policies pass, the label will be removed automatically.

[Copilot]

Copilot review overview

🟢 Approval recommended

The bounded staging algorithm preserves existing semantics and avoids reading write-combined destinations.

Review effort: Balanced
Findings: None

What changed in this PR

Optimizes host-accessible buffer fills by reducing small memcpy() calls.

Changes:

• Uses memset() for single-byte patterns.
• Replicates wider patterns into a 4 KiB staging block for bulk copies.
• Preserves existing partial-pattern behavior.

File	Description
projects/​clr/​rocclr/​device/​blit.cpp	Implements wide-copy host buffer filling.

---

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

[mangupta]

🤖 AI-assisted pre-review (for @mangupta)

Re-reviewed the updated head (d878e22424f9) — no new findings at or above the posting threshold.

[systems-assistant]

TheRock Submodule Bump Activity

Newest first

• 2026-10-09 — Included in TheRock via ROCm/TheRock#8885.