refactor(vulkan): T-GPU-DEDUP-22 — float_adm_vulkan migrated to kernel_template + _add_variant by lusoris · Pull Request #288 · lusoris/vmaf

lusoris · 2026-05-02T13:13:51Z

Summary

Twin migration to adm_vulkan (T-GPU-DEDUP-21 / PR #287);
same 16-pipeline 2-D [stage][scale] array (4 stages × 4 scales).

State collapses dsl + pipeline_layout + shader + desc_pool to
VmafVulkanKernelPipeline pl. pipelines[0][0] aliases
s->pl.pipeline (template's base); the other 15 entries are
sibling pipelines via _add_variant().

Test plan

ninja -C build-vulkan clean
meson test test_vulkan_smoke test_vulkan_async_pending_fence test_vulkan_pic_preallocation — all green
pre-commit run --files <touched> clean

Base branch

This PR targets refactor/vulkan-template-add-variant (#272).

Six deep-dive deliverables (ADR-0108)

(1) Research digest: no digest needed: routine refactor
(2) Decision matrix: no alternatives: only-one-way fix
(3) AGENTS.md invariant note: see docs/rebase-notes.md entry 0122 — variants destroyed before bundle + the (0,0) destroy-skip invariant (pipeline aliasing)
(4) Reproducer / smoke-test command: see Test plan above
(5) CHANGELOG fragment: entry under ### Changed
(6) Rebase note: 0122

🤖 Generated with Claude Code

Copilot

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Refactors the Vulkan float_adm feature extractor to use the shared kernel_template infrastructure, deduplicating pipeline/layout/shader/descriptor-pool management and creating the 16 stage/scale pipelines via template variants.

Changes:

Migrates float_adm_vulkan pipeline creation/destruction to VmafVulkanKernelPipeline + vmaf_vulkan_kernel_pipeline_add_variant().
Keeps the existing 16-entry [stage][scale] pipeline table while making [0][0] alias the template base pipeline.
Documents the migration and invariants in CHANGELOG.md and docs/rebase-notes.md.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
libvmaf/src/feature/vulkan/float_adm_vulkan.c	Replaces manual Vulkan pipeline/DSL/pool management with `kernel_template` + variant pipelines.
docs/rebase-notes.md	Adds entry 0122 describing the migration and the destroy-skip invariant for `(0,0)`.
CHANGELOG.md	Records the refactor under “Changed” with scope and test validation.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

+    VkComputePipelineCreateInfo cpci = {
+        .stage =
+            {
+                .pName = "main",
+                .pSpecializationInfo = &spec_info,
+            },
    };
-    if (vkCreatePipelineLayout(dev, &plci, NULL, &s->pipeline_layout) != VK_SUCCESS)
-        return -ENOMEM;
+    return vmaf_vulkan_kernel_pipeline_add_variant(s->ctx, &s->pl, &cpci, out_pipeline);


+    map[0] = (VkSpecializationMapEntry){.constantID = 0, .offset = 0, .size = sizeof(int32_t)};
+    map[1] = (VkSpecializationMapEntry){.constantID = 1, .offset = 4, .size = sizeof(int32_t)};
+    map[2] = (VkSpecializationMapEntry){.constantID = 2, .offset = 8, .size = sizeof(int32_t)};
+    map[3] = (VkSpecializationMapEntry){.constantID = 3, .offset = 12, .size = sizeof(int32_t)};
+    map[4] = (VkSpecializationMapEntry){.constantID = 4, .offset = 16, .size = sizeof(int32_t)};


…_template Twin migration to adm_vulkan (T-GPU-DEDUP-21); same 16-pipeline 2-D [stage][scale] shape (4 stages * 4 scales). State collapses dsl + pipeline_layout + shader + desc_pool to VmafVulkanKernelPipeline pl. The VkPipeline pipelines[4][4] 2-D lookup is preserved so the per-stage dispatch path stays clean, but pipelines[0][0] aliases s->pl.pipeline (template's base) and the other 15 entries are sibling pipelines via _add_variant(). close_fex destroys the 15 sibling variants before calling vmaf_vulkan_kernel_pipeline_destroy() on the bundle. The (0, 0) skip is essential because pipelines[0][0] aliases s->pl.pipeline. Validated by meson test test_vulkan_smoke test_vulkan_async_pending_fence test_vulkan_pic_preallocation (all green) + clean compile. Numerical contract unchanged. Based on refactor/vulkan-template-add-variant (PR #272); will rebase clean once #272 lands. CHANGELOG.md + docs/rebase-notes.md (entry 0122) updated per ADR-0108 deep-dive deliverables. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

float_adm_vulkan (T-GPU-DEDUP-22) binds 9 SSBOs (src + 2× dwt_tmp + 2× band + 2× csf + accum + dis) and was tripping the template's max-8 guard at init() time, causing every extract() call to fail with "problem reading pictures" / "problem flushing context". The new cap of 16 is purely additive: existing consumers (PSNR 3 SSBOs, moment 4, ciede 3, motion 2, ssim 6, vif 6, etc.) are unaffected. Vulkan's portability minimum for maxDescriptorSetStorageBuffers is 96 on conformant devices, so 16 is well within the safe range. Smoke: float_adm Netflix-pair (576×324×48f) now reports adm2 mean 0.934515; smoke tests + lint clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Per CLAUDE.md §12 r13, every PR that closes a bug, opens a bug, or rules a Netflix upstream report not-affecting-the-fork updates docs/state.md in the same PR. Today's session shipped 7 such PRs without state.md updates (most were drafted in subagent-driven runs where the bookkeeping slipped). This is the catch-up. Recently closed (added): - vmaf_tiny_v1.onnx external-data ref fix (PR #296) - kernel_template SSBO cap 8->16 (PR #288 + named constant in #292) - deliverables-check.sh backslash strip (PR #292) - CI workflows draft-skip guard (PR #300) - CLAUDE.md §12 r14 ffmpeg-patches gate fix (PR #297) - ADR slug-drift cleanup (PR #304) - 1.07e-3 CPU vmaf_v0.6.1 score drift bisect (PR #305) — confirmed inherited from upstream a44e5e6 motion edge-mirror fix; not a fork regression. Snapshot regen in separate PR aligns with fork. No code changes. State.md only. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…307) * docs(state): record 2026-05-02 session bug-status changes (7 closes) Per CLAUDE.md §12 r13, every PR that closes a bug, opens a bug, or rules a Netflix upstream report not-affecting-the-fork updates docs/state.md in the same PR. Today's session shipped 7 such PRs without state.md updates (most were drafted in subagent-driven runs where the bookkeeping slipped). This is the catch-up. Recently closed (added): - vmaf_tiny_v1.onnx external-data ref fix (PR #296) - kernel_template SSBO cap 8->16 (PR #288 + named constant in #292) - deliverables-check.sh backslash strip (PR #292) - CI workflows draft-skip guard (PR #300) - CLAUDE.md §12 r14 ffmpeg-patches gate fix (PR #297) - ADR slug-drift cleanup (PR #304) - 1.07e-3 CPU vmaf_v0.6.1 score drift bisect (PR #305) — confirmed inherited from upstream a44e5e6 motion edge-mirror fix; not a fork regression. Snapshot regen in separate PR aligns with fork. No code changes. State.md only. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci: re-trigger after (6) opt-out fix --------- Co-authored-by: Lusoris <lusoris@pm.me> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Per ADR-0165 / CLAUDE.md §12 r13. Bookkeeping refresh of `docs/state.md` to close out the busy 2026-05-02/03 session: - Header date bumped 2026-04-29 → 2026-05-03. - Closed Issue #239 (FFmpeg `libvmaf_vulkan` wall-clock serialisation): row moved Open → Recently closed; cited PR #241 / commit `e266bf8e` and ADR-0251 (renumbered from 0235 by PR #310 dedup sweep). Closure verified by the v2 async pending-fence ring's `v2 ≤ 0.7 × v1` measurement gate flipping ADR-0251 to Accepted. - New Open-bugs row: `y4m_convert_411_422jpeg` heap-buffer-overflow on 4:1:1 with `dst_c_w == 1` (PR #348 fuzz). Reproducer parked at `libvmaf/test/fuzz/y4m_input_known_crashes/y4m_411_w2_h4_oob_dst.y4m`; fix follow-up PR TBD. - Recently closed audited for stale drafts: six rows updated with merged commit SHAs (#288, #292, #296, #297, #300, #304, #305) and the kernel-template citation corrected to ADR-0246 (post-dedup; ADR-0221 is now `changelog-adr-fragment-pattern.md`). - Netflix#955 deferred row: last-checked stamp refreshed to 2026-05-03; Netflix#1494 still `state=OPEN` per gh API. - "Update protocol" untouched; no row removed below its closure threshold; "Confirmed not-affected" entries unchanged. Six ADR-0108 deliverables: 1. Research digest — no digest needed: state-md bookkeeping. 2. Decision matrix — no alternatives: only-one-way fix per ADR-0165 / CLAUDE.md §12 r13. 3. AGENTS.md invariant note — no rebase-sensitive invariants. 4. Reproducer / smoke command — `mkdocs build --strict` + view the file; PR #348's harness reproduces the new Open bug. 5. CHANGELOG fragment — `changelog.d/changed/state-md-refresh-2026-05-03.md`. 6. Rebase note — no rebase impact: fork-local doc bookkeeping. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

#352) Per ADR-0165 / CLAUDE.md §12 r13. Bookkeeping refresh of `docs/state.md` to close out the busy 2026-05-02/03 session: - Header date bumped 2026-04-29 → 2026-05-03. - Closed Issue #239 (FFmpeg `libvmaf_vulkan` wall-clock serialisation): row moved Open → Recently closed; cited PR #241 / commit `e266bf8e` and ADR-0251 (renumbered from 0235 by PR #310 dedup sweep). Closure verified by the v2 async pending-fence ring's `v2 ≤ 0.7 × v1` measurement gate flipping ADR-0251 to Accepted. - New Open-bugs row: `y4m_convert_411_422jpeg` heap-buffer-overflow on 4:1:1 with `dst_c_w == 1` (PR #348 fuzz). Reproducer parked at `libvmaf/test/fuzz/y4m_input_known_crashes/y4m_411_w2_h4_oob_dst.y4m`; fix follow-up PR TBD. - Recently closed audited for stale drafts: six rows updated with merged commit SHAs (#288, #292, #296, #297, #300, #304, #305) and the kernel-template citation corrected to ADR-0246 (post-dedup; ADR-0221 is now `changelog-adr-fragment-pattern.md`). - Netflix#955 deferred row: last-checked stamp refreshed to 2026-05-03; Netflix#1494 still `state=OPEN` per gh API. - "Update protocol" untouched; no row removed below its closure threshold; "Confirmed not-affected" entries unchanged. Six ADR-0108 deliverables: 1. Research digest — no digest needed: state-md bookkeeping. 2. Decision matrix — no alternatives: only-one-way fix per ADR-0165 / CLAUDE.md §12 r13. 3. AGENTS.md invariant note — no rebase-sensitive invariants. 4. Reproducer / smoke command — `mkdocs build --strict` + view the file; PR #348's harness reproduces the new Open bug. 5. CHANGELOG fragment — `changelog.d/changed/state-md-refresh-2026-05-03.md`. 6. Rebase note — no rebase impact: fork-local doc bookkeeping. Co-authored-by: Lusoris <lusoris@pm.me> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Copilot AI review requested due to automatic review settings May 2, 2026 13:13

Copilot AI reviewed May 2, 2026

View reviewed changes

lusoris force-pushed the refactor/vulkan-template-add-variant branch from 0cad1f3 to af40030 Compare May 2, 2026 13:26

lusoris force-pushed the refactor/float-adm-vulkan-template branch from d0d9a19 to 6981c64 Compare May 2, 2026 13:31

lusoris marked this pull request as draft May 2, 2026 14:29

lusoris force-pushed the refactor/vulkan-template-add-variant branch from af40030 to 2ed2ec8 Compare May 2, 2026 15:00

Base automatically changed from refactor/vulkan-template-add-variant to master May 2, 2026 15:21

lusoris marked this pull request as ready for review May 2, 2026 20:29

lusoris force-pushed the refactor/float-adm-vulkan-template branch from 6981c64 to 5c5b553 Compare May 2, 2026 20:29

lusoris marked this pull request as draft May 2, 2026 20:40

lusoris marked this pull request as ready for review May 2, 2026 20:45

lusoris force-pushed the refactor/float-adm-vulkan-template branch from c67bddb to 500f6a9 Compare May 2, 2026 21:14

lusoris marked this pull request as draft May 2, 2026 21:26

lusoris marked this pull request as ready for review May 2, 2026 21:42

Lusoris and others added 2 commits May 2, 2026 23:42

lusoris force-pushed the refactor/float-adm-vulkan-template branch from 500f6a9 to 2353f2d Compare May 2, 2026 21:42

lusoris merged commit bb9d772 into master May 2, 2026
54 checks passed

lusoris deleted the refactor/float-adm-vulkan-template branch May 2, 2026 22:25

lusoris mentioned this pull request May 2, 2026

docs(state): record 2026-05-02 session bug-status changes (7 closes) #307

Merged

8 tasks

lusoris mentioned this pull request May 3, 2026

docs(state): refresh 2026-05-03 — close #239, add Y4M OOB, audit slugs #352

Merged

13 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

refactor(vulkan): T-GPU-DEDUP-22 — float_adm_vulkan migrated to kernel_template + _add_variant#288

refactor(vulkan): T-GPU-DEDUP-22 — float_adm_vulkan migrated to kernel_template + _add_variant#288
lusoris merged 2 commits intomasterfrom
refactor/float-adm-vulkan-template

lusoris commented May 2, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

lusoris commented May 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Base branch

Six deep-dive deliverables (ADR-0108)

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

lusoris commented May 2, 2026 •

edited

Loading