perf(hevc): decode ready pictures concurrently

.github/workflows/hevc-conformance.yml

@@ -70,7 +70,20 @@ jobs:
           toolchain: stable
       - run: cargo test --locked -p coreav-hevc -p coreav-kernels-hevc -p coreav-backend-hevc --no-default-features
       - run: cargo test --locked -p coreav-hevc -p coreav-kernels-hevc -p coreav-backend-hevc -p coreav-thumbnail --features native-hevc --all-targets
+      - name: Exercise the real native reference driver on generated pixels
+        run: |
+          set -euo pipefail
+          cargo test --locked -p coreav-backend-hevc --features native-hevc,test-support --example reference | tee "$RUNNER_TEMP/native-reference-tests.log"
+          grep -F 'test tests::native_reference_checks_generated_depths_at_every_worker_boundary ... ok' "$RUNNER_TEMP/native-reference-tests.log"
       - run: cargo test --locked -p coreav-hevc -p coreav-kernels-hevc --features native-hevc,scalar --all-targets
+      - name: Check diagnostic-only extraction and row reads
+        if: runner.os == 'Linux'
+        run: |
+          set -euo pipefail
+          RUSTFLAGS='--cfg hevc_thread_diagnostics' cargo test --locked -p coreav-hevc -p coreav-kernels-hevc --features native-hevc --lib | tee "$RUNNER_TEMP/hevc-diagnostic-tests.log"
+          grep -Fx 'test decoder::scheduler::tests::header_only_dependency_replay_distinguishes_chains_and_siblings ... ok' "$RUNNER_TEMP/hevc-diagnostic-tests.log"
+          grep -Fx 'test decoder::scheduler::tests::row_dependency_trace_observes_generated_predictions ... ok' "$RUNNER_TEMP/hevc-diagnostic-tests.log"
+          grep -Fx 'test inter::tests::diagnostic_read_bounds_match_interpolation_support ... ok' "$RUNNER_TEMP/hevc-diagnostic-tests.log"
       - run: cargo clippy --locked -p coreav-hevc -p coreav-kernels-hevc -p coreav-backend-hevc -p coreav-thumbnail --features native-hevc --all-targets -- -D warnings
       - name: Check the opt-in CLI registration with and without default features
         run: |
@@ -86,6 +99,10 @@ jobs:
           if grep -E 'coreav-(kernels-)?hevc' "$RUNNER_TEMP/backend-off.txt"; then exit 1; fi
           cargo tree --locked -p coreav-cli -e normal,build > "$RUNNER_TEMP/default-consumer.txt"
           if grep -E 'coreav-(kernels-|backend-)?hevc' "$RUNNER_TEMP/default-consumer.txt"; then exit 1; fi
+          for feature in release-default release-full; do
+            cargo tree --locked -p coreav-cli --no-default-features --features "$feature" -e normal,build,features > "$RUNNER_TEMP/$feature.txt"
+            if grep -E 'native-hevc-private|coreav-(kernels-|backend-)?hevc' "$RUNNER_TEMP/$feature.txt"; then exit 1; fi
+          done
       - name: Check Linux acquisition and subprocess isolation
         if: runner.os == 'Linux'
         run: python3 -m unittest discover -s scripts/tests -p 'test_hevc*.py'

.github/workflows/hevc-native-rows.yml

@@ -0,0 +1,136 @@
+name: hevc-native-rows
+run-name: Native HEVC rows ${{ inputs.candidate }} (${{ inputs.mode }})
+# workflow-check: not required
+
+on:
+  workflow_call:
+    inputs:
+      candidate:
+        required: true
+        type: string
+      mode:
+        type: string
+        default: all
+      only:
+        type: string
+        default: ''
+      frames:
+        type: number
+        default: 32
+  workflow_dispatch:
+    inputs:
+      candidate:
+        description: Exact decoder commit
+        required: true
+        type: string
+      mode:
+        description: Native reference calls, complete checks, or both
+        type: choice
+        options: [all, reference, checks]
+        default: all
+      only:
+        description: Optional comma-separated non-wavefront input IDs for reference calls
+        type: string
+        default: ''
+      frames:
+        description: Maximum pictures per reference input
+        type: number
+        default: 32
+
+permissions:
+  contents: read
+
+env:
+  CANDIDATE: ${{ inputs.candidate }}
+  ONLY: ${{ inputs.only }}
+  FRAMES: ${{ inputs.frames }}
+
+jobs:
+  reference:
+    if: inputs.mode == 'all' || inputs.mode == 'reference'
+    name: hevc-native-rows-reference (${{ matrix.shard }})
+    runs-on: ubuntu-24.04
+    timeout-minutes: 180
+    strategy:
+      fail-fast: false
+      matrix:
+        shard: ${{ fromJSON(inputs.only != '' && '[0]' || '[0,1,2]') }}
+    steps:
+      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683
+        with:
+          persist-credentials: false
+          ref: ${{ inputs.candidate }}
+          fetch-depth: 0
+      - name: Admit exact candidate and bounded reference selection
+        run: |
+          set -euo pipefail
+          [[ "$CANDIDATE" =~ ^[0-9a-f]{40}$ ]]
+          test "$(git rev-parse HEAD)" = "$CANDIDATE"
+          test -n "$(git branch -r --contains "$CANDIDATE")"
+          python3 -I - "$ONLY" "$FRAMES" <<'PY'
+          import sys
+          sys.path.insert(0, 'scripts')
+          from hevc_threading import ROW_INPUT_IDS
+          selected = set(sys.argv[1].split(',')) if sys.argv[1] else set(ROW_INPUT_IDS)
+          if not selected or not selected <= set(ROW_INPUT_IDS) or not 1 <= int(sys.argv[2]) <= 32:
+              raise SystemExit('invalid_native_rows_selection')
+          PY
+      - uses: dtolnay/rust-toolchain@686976e191b89faba57d3206551f0f330d8cb249
+        with:
+          toolchain: stable
+      - name: Check the actual native driver before release building
+        run: |
+          set -euo pipefail
+          cargo test --locked -p coreav-hevc --features native-hevc-rows-private,test-support --lib decoder::scheduler::tests
+          cargo test --locked -p coreav-backend-hevc --features native-hevc,test-support --example reference
+          cargo build --locked --release -p coreav-backend-hevc --features native-hevc,test-support --example reference
+      - name: Install the pinned binary oracle
+        env:
+          R2_ENDPOINT: ${{ secrets.COREAV_CORPUS_R2_ENDPOINT }}
+          R2_ACCESS_KEY_ID: ${{ secrets.COREAV_CORPUS_R2_ACCESS_KEY_ID }}
+          R2_SECRET_ACCESS_KEY: ${{ secrets.COREAV_CORPUS_R2_SECRET_ACCESS_KEY }}
+          R2_BUCKET: ${{ vars.COREAV_CORPUS_R2_BUCKET }}
+        run: scripts/install-oracle-ffmpeg.sh "$RUNNER_TEMP/oracle"
+      - name: Verify every worker count before paired native reference calls
+        env:
+          R2_ENDPOINT: ${{ secrets.COREAV_CORPUS_R2_ENDPOINT }}
+          R2_ACCESS_KEY_ID: ${{ secrets.COREAV_CORPUS_R2_ACCESS_KEY_ID }}
+          R2_SECRET_ACCESS_KEY: ${{ secrets.COREAV_CORPUS_R2_SECRET_ACCESS_KEY }}
+          R2_BUCKET: ${{ vars.COREAV_CORPUS_R2_BUCKET }}
+        run: |
+          set -euo pipefail
+          umask 077
+          mkdir -p "$RUNNER_TEMP/private"
+          python3 -I scripts/hevc_threading.py native-reference \
+            --native "$PWD/target/release/examples/reference" --oracle-bin "$RUNNER_TEMP/oracle" \
+            --out "$RUNNER_TEMP/private/native" --candidate "$CANDIDATE" \
+            --shard-index '${{ matrix.shard }}' --shard-count '${{ inputs.only != '' && 1 || 3 }}' \
+            --only "$ONLY" --frames "$FRAMES" > "$RUNNER_TEMP/private/native.log" 2>&1
+      - name: Retain private reference results
+        if: always()
+        run: |
+          set -euo pipefail
+          cd "$RUNNER_TEMP"
+          [ -d private ] || exit 0
+          tar -czf records.tgz private
+          mkdir -p sealed
+          openssl rand -hex -out key.txt 32
+          openssl enc -aes-256-cbc -pbkdf2 -salt -pass file:key.txt -in records.tgz -out sealed/records.tgz.enc
+          openssl pkeyutl -encrypt -pubin -inkey "$GITHUB_WORKSPACE/bench/hevc-conformance/evidence-public.pem" -pkeyopt rsa_padding_mode:oaep -pkeyopt rsa_oaep_md:sha256 -in key.txt -out sealed/key.txt.enc
+          shred -u key.txt records.tgz
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
+        if: always()
+        with:
+          name: hevc-native-rows-reference-${{ matrix.shard }}-${{ inputs.candidate }}
+          path: ${{ runner.temp }}/sealed
+          if-no-files-found: error
+          retention-days: 7
+
+  checks:
+    if: inputs.mode == 'all' || inputs.mode == 'checks'
+    uses: ./.github/workflows/hevc-threading.yml
+    with:
+      candidate: ${{ inputs.candidate }}
+      baseline: ''
+      frames: 32
+    secrets: inherit

.github/workflows/hevc-reference.yml

@@ -15,7 +15,7 @@ on:
         required: true
         type: string
       baseline:
-        description: Optional full baseline SHA for a comparison
+        description: Optional full baseline SHA, or native for same-head row evidence
         required: false
         default: ''
         type: string
@@ -24,6 +24,21 @@ on:
         required: false
         default: false
         type: boolean
+      thread_diagnostics_only:
+        description: Run only untimed 4T picture-job diagnostics on all 30 inputs
+        required: false
+        default: false
+        type: boolean
+      thread_dependencies_only:
+        description: Replay only header reference dependencies on all 30 inputs
+        required: false
+        default: false
+        type: boolean
+      thread_rows_only:
+        description: Decode only CTU-row reference reads on the 24 non-WPP inputs
+        required: false
+        default: false
+        type: boolean
       only:
         description: Optional comma-separated input IDs
         required: false
@@ -54,7 +69,7 @@ jobs:
     secrets: inherit
 
   reference:
-    if: ${{ !inputs.inverse && inputs.baseline == '' && !inputs.sanitizer_only }}
+    if: ${{ !inputs.inverse && inputs.baseline == '' && !inputs.sanitizer_only && !inputs.thread_diagnostics_only && !inputs.thread_dependencies_only && !inputs.thread_rows_only }}
     runs-on: ubuntu-24.04
     timeout-minutes: 150
     env:
@@ -128,12 +143,25 @@ jobs:
           retention-days: 30
 
   threading:
-    if: ${{ !inputs.inverse && (inputs.baseline != '' || inputs.sanitizer_only) }}
+    if: ${{ !inputs.inverse && inputs.baseline != 'native' && (inputs.baseline != '' || inputs.sanitizer_only || inputs.thread_diagnostics_only || inputs.thread_dependencies_only || inputs.thread_rows_only) }}
     uses: ./.github/workflows/hevc-threading.yml
     with:
       candidate: ${{ inputs.candidate }}
       baseline: ${{ inputs.baseline }}
       sanitizer_only: ${{ inputs.sanitizer_only }}
+      thread_diagnostics_only: ${{ inputs.thread_diagnostics_only }}
+      thread_dependencies_only: ${{ inputs.thread_dependencies_only }}
+      thread_rows_only: ${{ inputs.thread_rows_only }}
+      only: ${{ inputs.only }}
+      frames: ${{ fromJSON(inputs.frames) }}
+    secrets: inherit
+
+  native-rows:
+    if: ${{ !inputs.inverse && inputs.baseline == 'native' }}
+    uses: ./.github/workflows/hevc-native-rows.yml
+    with:
+      candidate: ${{ inputs.candidate }}
+      mode: ${{ inputs.only != '' && 'reference' || inputs.diagnose_only && 'checks' || 'all' }}
       only: ${{ inputs.only }}
       frames: ${{ fromJSON(inputs.frames) }}
     secrets: inherit