Skip to content

ADR-0468: HIP float_adm real kernel (ninth HIP consumer)

  • Status: Accepted
  • Date: 2026-05-16
  • Deciders: lusoris
  • Tags: hip, build, feature-extractor

Context

The CUDA backend ships a fully implemented float_adm_cuda.c / float_adm_score.cu (ADR-0192 / ADR-0202) that covers the four-stage DWT+CSF+CM ADM pipeline across all four DWT scales. The HIP backend had scaffold stubs for eight other feature extractors but was missing float_adm — the most signal-rich extractor in the VMAF model. Closing this gap brings the HIP backend to feature parity with CUDA for ADM scoring, which is the dominant contributor to the VMAF score.

Decision

Port float_adm_cuda.c and float_adm/float_adm_score.cu to HIP as feature/hip/float_adm_hip.c and feature/hip/float_adm/float_adm_score.hip. Wire the HSACO compilation target into src/meson.build's hip_kernel_sources dict and add the .c host TU to hip/meson.build's hip_sources list. Register vmaf_fex_float_adm_hip in the #if HAVE_HIP block of feature_extractor.c.

Key HIP adaptations from the CUDA twin:

  • Warp size 64 (GCN/RDNA): __shfl_down without mask; shared-mem warp-partial arrays sized at FADM_WARPS_PER_BLOCK = 4 (vs. 8 on CUDA).
  • Buffer alloc: hipMalloc / hipMemsetAsync / hipMemcpyAsync replace vmaf_cuda_buffer_alloc + cuMemsetD8Async.
  • Module API: hipModuleLoadData / hipModuleGetFunction / hipModuleLaunchKernel.
  • HSACO binary embedded via xxd -i in the hip_hsaco_sources custom_target pipeline (ADR-0372 pattern).

Alternatives considered

Option Pros Cons Why not chosen
Keep scaffold-only (ENOSYS) No work HIP ADM gap persists indefinitely Gap is the stated motivation
Share CUDA PTX via HIP's CUDA compat layer Single kernel source Compat layer is experimental on ROCm 6+; warp-size mismatch (32 vs 64) would silently corrupt reductions Correctness risk outweighs convenience

Consequences

  • Positive: HIP backend now supports float_adm and all VMAF model components that depend on it (adm2 score + 4 scale subscores). cross-backend-diff can be run against HIP ADM.
  • Negative: Ninth HSACO compilation target; adds ~2 s to enable_hipcc=true builds.
  • Neutral: Same scaffold posture as all prior HIP consumers — init() returns -ENOSYS without enable_hipcc=true; no user-visible change on CPU-only builds.

References

  • ADR-0192 / ADR-0202: CUDA float_adm implementation.
  • ADR-0372: HIP HSACO kernel build pipeline.
  • ADR-0468 closes the last major HIP feature-parity gap for standard VMAF.
  • req: "Port CUDA float_adm_cuda.c to HIP float_adm_hip.c + float_adm/*.hip kernels."