redline

KO | EN

gitlite — search

#hacktoberfest #mcp #artificial-intelligence #http #openai #typescript #rust #react #svg #chatbot

(★ 9)

Lightning-fast kernel dispatch for ROCm. Retained command replay over public ROCr/HSA queues.

File Explorer

ZIP Download

- - ci.yml
  - publish-pypi.yml
- - microwave_pack_proof.hip
- - analyze_dispatch_queues.py
  - apu_membw.cpp
  - aql_dispatch_floor.cpp
  - explicit_multistream.cpp
  - graph_dependency_fencing.cpp
  - llama_decode_shape.cpp
  - queue_release_busy.cpp
  - rocm_ident.cpp
  - rocr_queue_retire.cpp
- - graph_execupdate_oom.cpp
  - queue_signal_reuse_stress.cpp
  - vmm_4k_kernel_access.cpp
  - vmm_chunk_scale.cpp
  - vmm_expandable_canary.cpp
  - vmm_failed_setaccess.cpp
  - vmm_fd_ceiling.cpp
  - vmm_fd_conditional.cpp
  - vmm_fd_kind.cpp
  - vmm_granularity_matrix.cpp
  - vmm_reserve_alignment.cpp
  - vmm_setaccess_rule.cpp
- acc_kernel.hip
- contention_load.hip
- decode_chain.hip
- floor_hipgraph.hip
- floor_kernel.hip
- floor_kernel_ctr.hip
- phased_pm4_smoke.hip
- - Cargo.toml
  - README.md
- - Cargo.toml
  - cbindgen.toml
- - Cargo.toml
- - build.rs
  - Cargo.toml
  - README.md
  - redline_hipgraph.map
- - Cargo.toml
- - Cargo.lock
  - Cargo.toml
  - pyproject.toml
  - README.md
- - Cargo.toml
  - PROVENANCE.md
- - 2026-07-31-rocm-6529-address-zero-sqc-fault.md
  - 2026-08-01-hipgraph-handle-ownership-redesign.md
  - 2026-08-16-upstream-posts.md
  - 2026-08-23-dispatch-floor-rocm714.md
  - 2026-08-29-upstream-plan-rejected.md
  - 2026-09-01-hipfire-vmm-pm4-fault.md
  - claims-discipline.md
  - posted-2026-08-23-rocm-10021.md
  - posted-2026-08-23-rocm-10604-issue.md
  - posted-2026-08-23-rocm-9360.md
  - posted-2026-08-24-rocm-6409-correction.md
  - posted-2026-08-24-rocm-6409-update-rocm10.md
  - posted-2026-08-24-rocm-6409.md
  - posted-2026-08-27-rocm-6409-graph-shape.md
- devices.toml.example
- DISPATCH-FLOOR.md
- INTEGRATION.md
- - demo.py
  - gmb_aco_style.s
  - gmb_buffer.hip
  - gmb_direct_srd.s
  - gmb_noop.hip
  - hipgraph_baseline.hip
  - README.md
- - dispatch_matrix.py
  - hipgraph_timing_override.hpp
  - hsaco_manifest.py
  - normalize_redline.py
  - README.md
  - redline_hip_timing.py
  - redline_timing_override.hpp
  - run_matrix.py
  - run_python_runner.py
  - summarize_results.py
  - test_redline_adapters.py
- - README.md
  - run_ab.sh
- - build.rs
  - Cargo.lock
  - Cargo.toml
  - join_arms.py
  - radiowave-plain-recipes.json
  - README.md
- - build.rs
  - Cargo.lock
  - Cargo.toml
  - README.md
  - summarize.py
- - ab.sh
  - build.sh
  - graph_demo.cpp
  - graph_demo_mod.cpp
  - README.md
- 6529-contention-ab.sh
- 6529-preemption-probe.sh
.gitignore
Cargo.lock
Cargo.toml
LICENSE
NOTICE
README.md

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<img width="1536" height="1024" alt="image" src="https://github.com/user-attachments/assets/db0984c5-c03c-4ee3-bc0d-a21b48e5898c" />

**Lightning-fast kernel dispatch for ROCm.** Record a kernel DAG once, derive
only the required fences from declared memory access, and replay it as one
retained submission over the public ROCr/HSA ABI — avoiding HIP's per-launch
submission-and-fence floor.

Redline targets the well-documented HIP dispatch-floor gap
([ROCm/ROCm#6409](https://github.com/ROCm/ROCm/issues/6409)): at tiny-dispatch
workloads, Vulkan/RADV command-buffer replay is faster than `hipGraph` because
HIP fences and submits more work per launch than the data flow requires.

## What you get

Four integration paths, one retained-PM4 engine:

| Path | Use when |
| --- | --- |
| **Explicit C/C++ ABI** (`redline-capi`) | Your engine owns its HIP allocations, code objects, and fixed launch sequence. |
| **Python** (`redline-py`) | You want graph authoring or retained module load/build/replay from Python. |
| **`hipGraph*` preload** (`redline-hipgraph`) | An existing application already uses compatible HIP graph capture; unsupported operations fall through to HIP. |
| **Rust** (`redline-dispatch`) | You want the native graph, public-AQL, HIP multistream, or direct retained-PM4 APIs. |

Start with the choose-your-path **[Using Redline](docs/INTEGRATION.md)** guide.

Radiowave is the compiler **policy** layer (not a compiler fork): it drives
installed LLVM/hipcc, inspects the AMDGPU code object, certifies VMEM-only
mutable reads for the narrow RMW boundary, and emits a hashed build manifest.
Scalar or ambiguous consumers fail closed to the broader same-agent boundary.

## Architecture support

| Path | Coverage |
| --- | --- |
| Public ROCr/AQL replay | Architecture-neutral; exercised on gfx1010, gfx1030, gfx1100, gfx1151, gfx1201 |
| Retained direct PM4 | Family-specific GFX10/GFX11 and GFX12 encoders; device-family mismatch fails closed |
| Multi-queue independent IBs | Auto policy: Q2 (gfx1100 and gfx12), Q4 (other measured gfx11), Q1 (unswept gfx10); serial RMW stays single-queue |

Legacy direct PM4 supports zero-scratch HSA kernels whose implicit user data is
the optional private-segment buffer plus kernarg pointer. Unsupported scratch,
queue, dispatch, or flat-scratch contracts fail closed.

**Requires ROCm Core SDK ≥ 7.14** (TheRock layout, typically `/opt/rocm/core`).

## Current results (ROCm 7.14)

Current headline — clean, same-commit three-card comparison at
`bb612d14a95c92c4bf20f492be28f7ae11cb51f7`: the full 240-row HipEngine-shaped
matrix, Redline/HIP/Vulkan, 3 warmups, 7 samples, and CPU-oracle correctness
gating. **All 720 rows matched; zero were rejected.**

<img width="1693" height="929" alt="image" src="https://github.com/user-attachments/assets/3b9bc818-9131-4967-a431-8958efc6eef7" />
<p align="center">
  <a href="examples/hipfire-6409/results/gfx1201/2026-07-23-rocm714-three-way/REPORT.md">
    <img alt="RX 9070 XT Report" src="https://img.shields.io/badge/RX%209070%20XT-gfx1201%20Report-red?style=for-the-badge">
  </a>
  <a href="examples/hipfire-6409/results/gfx1100/2026-07-23-rocm714-three-way/REPORT.md">
    <img alt="RX 7900 XTX Report" src="https://img.shields.io/badge/RX%207900%20XTX-gfx1100%20Report-red?style=for-the-badge">
  </a>
  <a href="examples/hipfire-6409/results/gfx1151/2026-07-23-rocm714-three-way/REPORT.md">
    <img alt="Strix Halo Report" src="https://img.shields.io/badge/Strix%20Halo-gfx1151%20Report-red?style=for-the-badge">
  </a>
</p>
Median speedup is backend time divided by Redline time. On gfx1100, `auto` is
capped at Q2: clean explicit-Q3 and Q4 controls both reproduced retained-PM4
timeouts in adjacent memory-waitcnt rows. Those queue counts remain diagnostic
overrides, not supported defaults.

Earlier controls remain available for comparison:

- Four-backend gfx1201 retest:
  [`2026-07-22-rocm7.14-retest`](examples/hipfire-6409/results/gfx1201/2026-07-22-rocm7.14-retest/REPORT.md) —
  192/240 Redline firsts; Redline beat Vulkan and HipGraph/HIP in 192 and
  222 rows respectively.
- Cross-RDNA ROCm 7.2-era native PM4:
  [`2026-07-14-rdna-rocr-native`](examples/hipfire-6409/results/2026-07-14-rdna-rocr-native/REPORT.md) —
  537/960 firsts, Redline > Vulkan 606/960, and 960/960 correct across gfx1010,
  gfx1030, gfx1100, and gfx1151.
- HipEngine pristine 224-row harness:
  [gfx1201](examples/hipengine-6409/results/gfx1201/2026-07-22-714-bench/REPORT.md)
  197/224, [gfx1151](examples/hipengine-6409/results/gfx1151/2026-07-22-714-bench/REPORT.md)
  164/224, and [gfx1100](examples/hipengine-6409/results/gfx1100/2026-07-22-714-bench/REPORT.md)
  127/224 Redline wins over Vulkan.

Raw JSON records keep their original dirty-tree flags, empty/`hipcc` capture
fields, and absolute runner paths where the harness wrote them. Product docs
label that provenance; they do not rewrite the artifacts.

Harness details and reproduce commands:
[`examples/hipfire-6409`](examples/hipfire-6409/README.md),
[`examples/hipengine-6409`](examples/hipengine-6409/README.md).
Dispatch-floor methodology (historical ROCm 7.2 microbench):
[`docs/DISPATCH-FLOOR.md`](docs/DISPATCH-FLOOR.md).

## Start here

The canonical **[Using Redline](docs/INTEGRATION.md)** guide covers prerequisites,
the shared kernel/kernarg contract, verified C, Python, preload, and Rust
walkthroughs, [device-anchored GPU selection](docs/INTEGRATION.md#device-anchored-gpu-selection),
failure behavior, and current package availability.

Common source-build setup:

```bash
export PATH=/opt/rocm/core/bin:/opt/rocm/core/lib/llvm/bin:$PATH
export ROCM_PATH=/opt/rocm/core HIP_PATH=/opt/rocm/core

cargo build --release -p redline-dispatch -p redline-capi -p redline-hipgraph
```

Redline is currently distributed from source. The guide distinguishes commands
that work now from the intended PyPI wheel and versioned C SDK release.

C header: [`crates/redline-capi/include/redline_dispatch.h`](crates/redline-capi/include/redline_dispatch.h).
Python reference: [`crates/redline-py`](crates/redline-py/README.md).
Preload reference: [`crates/redline-hipgraph`](crates/redline-hipgraph/README.md).

## Limitations (honest)

- **Not a full HIP runtime replacement.** The preload path covers supported
  `hipGraph*` / module-loaded kernels; everything else falls through to real HIP.
- **Retained PM4 is family-specific.** Wrong encoder for the device is rejected.
- **Residual Vulkan wins are mostly codegen.** Packed-dot / VOPD-class rows and
  some production shapes remain RADV/ACO advantages on the shared-HSACO control;
  transport is not the whole story.
- **The current three-card comparison is clean and same-commit.** Earlier
  2026-07-22 controls retain their original dirty-tree flags, empty `hipcc`
  capture fields, and absolute runner paths. Those artifacts remain useful
  controls but are not the current certification.
- **Leverage A/B is secondary.** The earlier 194/240 partitioned run is
  dirty-tree non-regression evidence; roctx/amd-smi observation was not
  uniformly available across arms.
- **Historical dispatch-floor µs tables** in `docs/DISPATCH-FLOOR.md` are ROCm
  7.2 methodology on R9700; current product validation is the retained ROCm
  7.14 set above.

## Crate map

| Crate | Role |
| --- | --- |
| **`radiowave`** | HIP compiler policy: lowering helpers, wave/target selection, hipcc invocation, CO inspection, manifests |
| **`redline-dispatch`** | Record/replay core, hazard checks, minimal fences, backends, Rust `hipgraph` adapter |
| **`redline-rocr`** | Public ROCr/HSA ABI, AQL, queue publication ([provenance](crates/redline-rocr/PROVENANCE.md)) |
| **`redline-capi`** | C ABI for engines |
| **`redline-py`** | Python bindings |
| **`redline-hipgraph`** | `hipGraph*` ABI + `LD_PRELOAD` interposer |
| **`redline-observe`** | Optional ROCm 7.14 observability hooks (roctx / amd-smi when present) |

## License

Licensed under the **Apache License, Version 2.0** — see [`LICENSE`](LICENSE).
Redistributions must retain [`NOTICE`](NOTICE) per section 4(d).

"Redline" is a trademark of Kaden Schutt. As stated in section 6 of the License,
the Apache-2.0 grant does not include trademark rights: you may use, modify, and
redistribute this code, but not the "Redline" name to identify your own fork or
product.

`/opt/rocm/core` is the expected TheRock layout on the machines that produced
the retained results; adjust `PATH` / `ROCM_PATH` if your install differs.

redline

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README