Skip to content

perf(hip): upload native samples and convert on device in psnr_hvs - #1658

Merged
lusoris merged 2 commits into
masterfrom
perf/hip-psnr-hvs-device-convert
Oct 1, 2026
Merged

perf(hip): upload native samples and convert on device in psnr_hvs#1658
lusoris merged 2 commits into
masterfrom
perf/hip-psnr-hvs-device-convert

Conversation

@lusoris

@lusoris lusoris commented Oct 1, 2026

Copy link
Copy Markdown
Contributor

Summary

Ports the ADR-1369 psnr_hvs design to the HIP backend (closing T-HIP-PSNR-HVS-HOST-CONVERT-2026-09-29):

  • Eliminates host-side float conversions and discards unused pinned host staging allocations (h_uint_ref and h_uint_dist).
  • Uploads native raw samples directly via pinned staging (vmaf_hip_picture_upload() / stage plane).
  • Reads and converts raw samples directly on device for 8 bpc (uint8_t) and wide 9–12 bpc (uint16_t) in psnr_hvs_score.hip.
  • Fuses all plane dispatches into a single kernel (n_dispatches_per_frame = 1), reducing 4K frame time from 221.79 ms to 18.90 ms/frame on AMD gfx1036.
  • Resolves a latent scaling bug on 9-bit and 11-bit depths, now verified by test_psnr_hvs_deep_parity.

Type

  • perf — performance improvement
  • hip / cuda / simd — backend-specific

Checklist

  • Commits follow Conventional Commits (the commit-msg hook enforces this).
  • make format && make lint is green locally.
  • Unit tests pass: python3 scripts/ci/run_meson_test.py -- -C build.
  • If I touched any SIMD/GPU code path, I ran /cross-backend-diff and the worst ULP is ≤ 2.
  • If I touched a feature extractor with SIMD/GPU twins, I either updated every twin or listed the gap under "Known follow-ups" below.
  • If I added a new .c / .cpp / .cu / .h / .hpp, it has the appropriate license header (see CONTRIBUTING.md).
  • If this is a breaking change, the commit message uses ! or BREAKING CHANGE: and the migration path is documented below.
  • If this PR adds an ADR, the ADR row lives in docs/adr/_index_fragments/<NNNN-slug>.md and the slug is appended to docs/adr/_index_fragments/_order.txt — do not edit docs/adr/README.md directly (regenerated by scripts/docs/concat-adr-index.sh; see ADR-0221).

Bug-status hygiene (ADR-0165)

  • docs/state.md updated in this PR with a row in the appropriate section (Open / Recently closed / Confirmed not-affected / Deferred), OR no state delta: REASON.

Netflix golden-data gate (ADR-0024)

  • I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.
  • If I believe a golden value must change, I have explained why below AND pinged @lusoris for a CODEOWNERS exception.

Cross-backend numerical results

psnr_hvs delta vs CPU on AMD gfx1036:
- Netflix 576x324 (48 frames): max abs diff 8.37e-05 dB (within 5e-4 tolerance)
- BBB 3840x2160 (50 frames): max abs diff 0.01099 dB (within area-scaled tolerance, ADR-1361)

Performance (if perf or feat)

Measured on AMD gfx1036 ((t(22) - t(2)) / 20 ms/frame, median of 3):

  • 576x324: 0.52 ms/frame (CPU 16t: 0.13 ms/frame)
  • 3840x2160: 18.90 ms/frame (down from 221.79 ms/frame; CPU 16t: 6.04 ms/frame)

Deep-dive deliverables (ADR-0108)

  • Research digest — no digest needed: architecture directly follows SYCL ADR-1369.
  • Decision matrix — captured in the corresponding ADR's ## Alternatives considered (or in the digest), OR "no alternatives: only-one-way fix".
  • AGENTS.md invariant note — added to the relevant package's AGENTS.md, OR "no rebase-sensitive invariants".
  • Reproducer / smoke-test command — pasted below under "Reproducer".
  • CHANGELOG fragment — a new file under changelog.d/<section>/<topic>.md (added / changed / deprecated / removed / fixed / security). Do not edit CHANGELOG.md directly — scripts/release/concat-changelog-fragments.sh renders the Unreleased block from the fragment tree (see ADR-0221).
  • Rebase note — entry added to docs/rebase-notes.md under a new ID, OR no rebase impact: REASON.

Reproducer

python3 scripts/dev/speed_gpu_parity.py --backend hip --feature psnr_hvs --max-abs-diff 0.02 --vmaf $(pwd)/build/tools/vmaf --netflix-dir python/test/resource/yuv --bbb-dir testdata/bbb

Known follow-ups

None. Closes T-HIP-PSNR-HVS-HOST-CONVERT-2026-09-29.

Breaking changes / migration

None.

@github-actions github-actions Bot added the type:perf Performance improvement label Oct 1, 2026
@lusoris lusoris mentioned this pull request Oct 1, 2026
Merged
17 of 26 tasks
lusoris added a commit that referenced this pull request Oct 1, 2026
@lusoris
fix(cuda): reproduce the CPU's running float sum in psnr_hvs_cuda
f2b637e 
psnr_hvs_cuda now returns the CPU extractor's scores bit for bit. Before,
it was up to 1.7e-2 dB from --backend cpu at 3840x2160, beyond the
ADR-1361 parity tolerance (3.34e-3 dB).

calc_psnrhvs() adds every masked coefficient error of a plane into one
running float, so its result depends on the order of the additions. The
twin summed the 64 terms of a block on the device and the blocks on the
host. Per maintainer decision (ADR-1397, amending ADR-1361) the twins copy
the CPU's accumulation; a double accumulator on the CPU was ruled out
because it moves a Netflix golden value past places=4.

- The kernel stores the 64 terms of every block, computed in the CPU's
  arithmetic: masking table and threshold in double, integer coefficient
  difference, fatbin built with --fmad=false.
- psnr_hvs_score.c (new) adds a plane's terms in the CPU's order and forms
  the combined score and the dB value with the CPU's expressions. It is
  built with the strict floating-point arguments of the scalar reference.
- The parity gates compare a CPU and psnr_hvs_cuda cell with tolerance 0
  at --precision max (EXACT_TWINS). The SYCL cells keep ADR-1361.
- test_cuda_psnr_hvs_parity asserts equality on all four outputs,
  including two 3840x2160 cases that fail on master by 1.6e-2 dB.
  test_psnr_hvs_score and test_psnr_hvs_twin_exact_sum_contract.py are
  device-free.

Measured on an RTX 4090 at --precision max: every frame of the Netflix
576x324 pairs (8, 10, 12 bits, 4:2:2), the 1920x1080 checkerboard pairs
and BBB 1920x1080 / 3840x2160 (8 and 10 bits) equals the CPU. CPU scores
are unchanged.

The exact sum costs throughput: 12.2 ms per 3840x2160 frame instead of
2.4 ms (1920x1080: 3.1 instead of 0.6; 576x324: 0.29 instead of 0.09),
and 256 bytes of readback per block. Tuning is tracked as
T-CUDA-PSNR-HVS-EXACT-SUM-THROUGHPUT-2026-10-01. The HIP and SYCL twins
are unchanged until #1658 and #1657 land
(T-HIP-PSNR-HVS-EXACT-SUM-2026-10-01, T-SYCL-PSNR-HVS-EXACT-SUM-2026-10-01).
Closes T-PSNR-HVS-CPU-FLOAT-SUM-4K-2026-09-30.
@lusoris lusoris mentioned this pull request Oct 1, 2026
Closed
@lusoris
lusoris force-pushed the perf/hip-psnr-hvs-device-convert branch from a731cee to f2881b1 Compare October 1, 2026 08:18
lusoris added 2 commits October 1, 2026 10:23
@lusoris
perf(hip): upload native samples and convert on device in psnr_hvs
6a114c8 
Port ADR-1369 design to the HIP psnr_hvs feature extractor, closing
T-HIP-PSNR-HVS-HOST-CONVERT-2026-09-29.

- core/src/feature/hip/integer_psnr_hvs_hip.c: replace host float
  conversion loops with native sample upload via vmaf_hip_picture_upload().
  Discard unused pinned staging allocations (h_uint_ref/dist).
- core/src/feature/hip/integer_psnr_hvs/psnr_hvs_score.hip: take raw sample
  planes and convert to integers directly on device. Resolves a latent
  scaling bug where 9-bit and 11-bit depths were multiplied by 16.
- core/test/test_hip_psnr_hvs_parity.c: add test_psnr_hvs_deep_parity
  covering 9, 10, 11, and 12-bit input parity against CPU reference.
- docs/state.md: close T-HIP-PSNR-HVS-HOST-CONVERT-2026-09-29 with measured
  parity and timing evidence on ryzen-4090-arc.
@lusoris
perf(hip): fuse psnr_hvs planes into single-dispatch kernel
640e5c0 
Fuse all plane dispatches into a single kernel launch in psnr_hvs_score.hip
(n_dispatches_per_frame = 1), matching the SYCL twin's ADR-1369 architecture:
- Pass plane descriptors and buffer pointers via PsnrHvsHipKernelArgs
- Launch one grid covering all blocks across all 3 planes
- Reduce 4K frame time from 221.79 ms to 18.90 ms/frame on AMD gfx1036
- Keep 576x324 delta within 8.37e-05 dB and 4K delta within area-scaled tolerance
- Update docs, rebase notes, CHANGELOG and docs/state.md
@lusoris
lusoris force-pushed the perf/hip-psnr-hvs-device-convert branch from f2881b1 to 640e5c0 Compare October 1, 2026 08:29
@lusoris
lusoris merged commit 57b3f21 into master Oct 1, 2026
37 of 39 checks passed
@lusoris
lusoris deleted the perf/hip-psnr-hvs-device-convert branch October 1, 2026 08:29
@lusoris lusoris mentioned this pull request Oct 1, 2026
Merged
lusoris added a commit that referenced this pull request Oct 1, 2026
@lusoris
fix(cuda): reproduce the CPU's running float sum in psnr_hvs_cuda
9c300cd 
psnr_hvs_cuda now returns the CPU extractor's scores bit for bit. Before,
it was up to 1.7e-2 dB from --backend cpu at 3840x2160, beyond the
ADR-1361 parity tolerance (3.34e-3 dB).

calc_psnrhvs() adds every masked coefficient error of a plane into one
running float, so its result depends on the order of the additions. The
twin summed the 64 terms of a block on the device and the blocks on the
host. Per maintainer decision (ADR-1397, amending ADR-1361) the twins copy
the CPU's accumulation; a double accumulator on the CPU was ruled out
because it moves a Netflix golden value past places=4.

- The kernel stores the 64 terms of every block, computed in the CPU's
  arithmetic: masking table and threshold in double, integer coefficient
  difference, fatbin built with --fmad=false.
- psnr_hvs_score.c (new) adds a plane's terms in the CPU's order and forms
  the combined score and the dB value with the CPU's expressions. It is
  built with the strict floating-point arguments of the scalar reference.
- The parity gates compare a CPU and psnr_hvs_cuda cell with tolerance 0
  at --precision max (EXACT_TWINS). The SYCL cells keep ADR-1361.
- test_cuda_psnr_hvs_parity asserts equality on all four outputs,
  including two 3840x2160 cases that fail on master by 1.6e-2 dB.
  test_psnr_hvs_score and test_psnr_hvs_twin_exact_sum_contract.py are
  device-free.

Measured on an RTX 4090 at --precision max: every frame of the Netflix
576x324 pairs (8, 10, 12 bits, 4:2:2), the 1920x1080 checkerboard pairs
and BBB 1920x1080 / 3840x2160 (8 and 10 bits) equals the CPU. CPU scores
are unchanged.

The exact sum costs throughput: 12.2 ms per 3840x2160 frame instead of
2.4 ms (1920x1080: 3.1 instead of 0.6; 576x324: 0.29 instead of 0.09),
and 256 bytes of readback per block. Tuning is tracked as
T-CUDA-PSNR-HVS-EXACT-SUM-THROUGHPUT-2026-10-01. The HIP and SYCL twins
are unchanged until #1658 and #1657 land
(T-HIP-PSNR-HVS-EXACT-SUM-2026-10-01, T-SYCL-PSNR-HVS-EXACT-SUM-2026-10-01).
Closes T-PSNR-HVS-CPU-FLOAT-SUM-4K-2026-09-30.
lusoris added a commit that referenced this pull request Oct 1, 2026
@lusoris
fix(cuda): reproduce the CPU's running float sum in psnr_hvs_cuda
c09c0a5 
psnr_hvs_cuda now returns the CPU extractor's scores bit for bit. Before,
it was up to 1.7e-2 dB from --backend cpu at 3840x2160, beyond the
ADR-1361 parity tolerance (3.34e-3 dB).

calc_psnrhvs() adds every masked coefficient error of a plane into one
running float, so its result depends on the order of the additions. The
twin summed the 64 terms of a block on the device and the blocks on the
host. Per maintainer decision (ADR-1397, amending ADR-1361) the twins copy
the CPU's accumulation; a double accumulator on the CPU was ruled out
because it moves a Netflix golden value past places=4.

- The kernel stores the 64 terms of every block, computed in the CPU's
  arithmetic: masking table and threshold in double, integer coefficient
  difference, fatbin built with --fmad=false.
- psnr_hvs_score.c (new) adds a plane's terms in the CPU's order and forms
  the combined score and the dB value with the CPU's expressions. It is
  built with the strict floating-point arguments of the scalar reference.
- The parity gates compare a CPU and psnr_hvs_cuda cell with tolerance 0
  at --precision max (EXACT_TWINS). The SYCL cells keep ADR-1361.
- test_cuda_psnr_hvs_parity asserts equality on all four outputs,
  including two 3840x2160 cases that fail on master by 1.6e-2 dB.
  test_psnr_hvs_score and test_psnr_hvs_twin_exact_sum_contract.py are
  device-free.

Measured on an RTX 4090 at --precision max: every frame of the Netflix
576x324 pairs (8, 10, 12 bits, 4:2:2), the 1920x1080 checkerboard pairs
and BBB 1920x1080 / 3840x2160 (8 and 10 bits) equals the CPU. CPU scores
are unchanged.

The exact sum costs throughput: 12.2 ms per 3840x2160 frame instead of
2.4 ms (1920x1080: 3.1 instead of 0.6; 576x324: 0.29 instead of 0.09),
and 256 bytes of readback per block. Tuning is tracked as
T-CUDA-PSNR-HVS-EXACT-SUM-THROUGHPUT-2026-10-01. The HIP and SYCL twins
are unchanged until #1658 and #1657 land
(T-HIP-PSNR-HVS-EXACT-SUM-2026-10-01, T-SYCL-PSNR-HVS-EXACT-SUM-2026-10-01).
Closes T-PSNR-HVS-CPU-FLOAT-SUM-4K-2026-09-30.
lusoris added a commit that referenced this pull request Oct 1, 2026
@lusoris
fix(cuda): reproduce the CPU's running float sum in psnr_hvs_cuda (#1666
5627800 
)

* fix(cuda): reproduce the CPU's running float sum in psnr_hvs_cuda

psnr_hvs_cuda now returns the CPU extractor's scores bit for bit. Before,
it was up to 1.7e-2 dB from --backend cpu at 3840x2160, beyond the
ADR-1361 parity tolerance (3.34e-3 dB).

calc_psnrhvs() adds every masked coefficient error of a plane into one
running float, so its result depends on the order of the additions. The
twin summed the 64 terms of a block on the device and the blocks on the
host. Per maintainer decision (ADR-1397, amending ADR-1361) the twins copy
the CPU's accumulation; a double accumulator on the CPU was ruled out
because it moves a Netflix golden value past places=4.

- The kernel stores the 64 terms of every block, computed in the CPU's
  arithmetic: masking table and threshold in double, integer coefficient
  difference, fatbin built with --fmad=false.
- psnr_hvs_score.c (new) adds a plane's terms in the CPU's order and forms
  the combined score and the dB value with the CPU's expressions. It is
  built with the strict floating-point arguments of the scalar reference.
- The parity gates compare a CPU and psnr_hvs_cuda cell with tolerance 0
  at --precision max (EXACT_TWINS). The SYCL cells keep ADR-1361.
- test_cuda_psnr_hvs_parity asserts equality on all four outputs,
  including two 3840x2160 cases that fail on master by 1.6e-2 dB.
  test_psnr_hvs_score and test_psnr_hvs_twin_exact_sum_contract.py are
  device-free.

Measured on an RTX 4090 at --precision max: every frame of the Netflix
576x324 pairs (8, 10, 12 bits, 4:2:2), the 1920x1080 checkerboard pairs
and BBB 1920x1080 / 3840x2160 (8 and 10 bits) equals the CPU. CPU scores
are unchanged.

The exact sum costs throughput: 12.2 ms per 3840x2160 frame instead of
2.4 ms (1920x1080: 3.1 instead of 0.6; 576x324: 0.29 instead of 0.09),
and 256 bytes of readback per block. Tuning is tracked as
T-CUDA-PSNR-HVS-EXACT-SUM-THROUGHPUT-2026-10-01. The HIP and SYCL twins
are unchanged until #1658 and #1657 land
(T-HIP-PSNR-HVS-EXACT-SUM-2026-10-01, T-SYCL-PSNR-HVS-EXACT-SUM-2026-10-01).
Closes T-PSNR-HVS-CPU-FLOAT-SUM-4K-2026-09-30.

* docs: regenerate the indexes and the citation map after rebasing
@lusoris lusoris mentioned this pull request Oct 1, 2026
Merged
17 of 26 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

type:perf Performance improvement

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@lusoris