Broadcast Kernel Fusion

.buildkite/pipeline.yml

@@ -8,12 +8,23 @@ steps:
   - group: ":julia: Julia"
     key: "julia"
     steps:
-      - label: "GPU - Julia {{matrix.julia}}"
+      # Fusion prefs are compile-time (CNPreferences). Write LocalPreferences.toml and
+      # set both values before instantiate/test so each job precompiles cuNumeric
+      # with the intended fusion configuration.
+      # - fusion=on:  FUSE_BROADCAST_EXPRS=true, MIN_OPS=1 (fuse single-op exprs)
+      # - fusion=off: FUSE_BROADCAST_EXPRS=false (MIN_OPS irrelevant)
+      - label: "GPU - Julia {{matrix.julia}} (fusion {{matrix.fusion}})"
         plugins:
           - JuliaCI/julia#v1:
               version: "{{matrix.julia}}"
           - jquick/pre-hook#v1.2.0:
-              command: "julia --project -e 'using Pkg; Pkg.resolve(); Pkg.instantiate()'"
+              command: |
+                if [ "{{matrix.fusion}}" = "on" ]; then
+                  printf '%s\n' '[CNPreferences]' 'FUSE_BROADCAST_EXPRS = true' 'FUSE_BROADCAST_MIN_OPS = 1' > LocalPreferences.toml
+                else
+                  printf '%s\n' '[CNPreferences]' 'FUSE_BROADCAST_EXPRS = false' > LocalPreferences.toml
+                fi
+                julia --project -e 'using Pkg; Pkg.resolve(); Pkg.instantiate()'
           - JuliaCI/julia-test#v1:
               test_args: "--quickfail"
           - JuliaCI/julia-coverage#v1:
@@ -42,13 +53,47 @@ steps:
               - "1.10"
               - "1.11"
               - "1.12"
+            fusion:
+              - "on"
+              - "off"
             arch:
               - "x64"
           # adjustments:
           #   - with:
           #       julia: "1.12"
           #     soft_fail : true
 
+  - group: ":hammer: Developer"
+    key: "developer"
+    steps:
+      - label: "GPU - Julia 1.11 (source wrapper, fusion {{matrix.fusion}})"
+        plugins:
+          - JuliaCI/julia#v1:
+              version: "1.11"
+        command: ".buildkite/run_developer_ci.sh"
+        artifact_paths:
+          - "deps/build.log"
+          - "deps/*.log"
+          - "deps/*.err"
+          - "*.log"
+        agents:
+          queue: "cuda"
+        if: build.message !~ /\[skip tests\]/
+        timeout_in_minutes: 60
+        env:
+          LD_LIBRARY_PATH: ""
+          LEGATE_AUTO_CONFIG: "0"
+          LEGATE_SHOW_CONFIG: "1"
+          GPUTESTS: "1"
+          LEGATE_TEST: "1"
+          LEGATE_CONFIG: "--cpus 1 --gpus 1 --utility 1 --fbmem 500 --logging legate=info,level=2 --log-to-file 1"
+          CUNUMERIC_FUSION: "{{matrix.fusion}}"
+        matrix:
+          setup:
+            fusion:
+              - "on"
+              - "off"
+
   # - group: ":computer: CUDA Versions (Julia 1.11)"
   #   key: "cuda"
   #   depends_on: "julia"

.buildkite/run_developer_ci.sh

@@ -0,0 +1,81 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+CI_PROJECT="$(mktemp -d)"
+
+case "${CUNUMERIC_FUSION:-}" in
+    on | off) ;;
+    *)
+        echo "CUNUMERIC_FUSION must be 'on' or 'off'" >&2
+        exit 2
+        ;;
+esac
+
+CMAKE_VERSION="3.30.7"
+CMAKE_ROOT="$(mktemp -d)"
+CMAKE_INSTALLER="$CMAKE_ROOT/cmake-installer.sh"
+curl --fail --silent --show-error --location \
+    --output "$CMAKE_INSTALLER" \
+    "https://github.com/Kitware/CMake/releases/download/v$CMAKE_VERSION/cmake-$CMAKE_VERSION-linux-x86_64.sh"
+sh "$CMAKE_INSTALLER" --skip-license --prefix="$CMAKE_ROOT"
+export PATH="$CMAKE_ROOT/bin:$PATH"
+cmake --version
+
+LEGATE_BRANCH_INPUT="${BUILDKITE_MESSAGE:-}"
+if [[ "${BUILDKITE_PULL_REQUEST:-false}" =~ ^[0-9]+$ ]]; then
+    echo "Reading Legate branch override from pull request #$BUILDKITE_PULL_REQUEST"
+    PR_BODY="$(
+        curl --fail --silent --show-error --location \
+            --header "Accept: application/vnd.github+json" \
+            "https://api.github.com/repos/JuliaLegate/cuNumeric.jl/pulls/$BUILDKITE_PULL_REQUEST" |
+            python3 -c 'import json, sys; print(json.load(sys.stdin).get("body") or "")'
+    )"
+    LEGATE_BRANCH_INPUT+=$'\n'"$PR_BODY"
+fi
+
+shopt -s nocasematch
+LEGATE_BRANCH=""
+if [[ "$LEGATE_BRANCH_INPUT" =~ legate[-_]branch:[[:space:]]*([A-Za-z0-9._/-]+) ]]; then
+    LEGATE_BRANCH="${BASH_REMATCH[1]}"
+fi
+shopt -u nocasematch
+
+if [[ -n "$LEGATE_BRANCH" ]]; then
+    LEGATE_CHECKOUT="$(mktemp -d)/Legate.jl"
+    echo "Using Legate.jl branch override: $LEGATE_BRANCH"
+    git clone --depth 1 --branch "$LEGATE_BRANCH" \
+        https://github.com/JuliaLegate/Legate.jl.git "$LEGATE_CHECKOUT"
+    git -C "$LEGATE_CHECKOUT" log -1 --format="Legate checkout: %D (%H)"
+    (
+        cd "$LEGATE_CHECKOUT"
+        julia --color=yes --project="$CI_PROJECT" -e '
+            using Pkg
+            Pkg.develop(PackageSpec(path = "lib/LegatePreferences"))
+            Pkg.develop(PackageSpec(path = "."))
+            using LegatePreferences
+            LegatePreferences.use_developer_mode()
+            Pkg.build("Legate")
+        '
+    )
+else
+    echo "No Legate.jl branch override - using JLL"
+fi
+
+julia --color=yes --project="$CI_PROJECT" -e '
+    using Pkg
+    Pkg.develop(PackageSpec(path = "lib/CNPreferences"))
+    using CNPreferences
+    CNPreferences.use_developer_mode()
+    CNPreferences.set_broadcast_fusion!(ENV["CUNUMERIC_FUSION"] == "on")
+    CNPreferences.set_broadcast_fusion_min_ops!(1)
+    Pkg.develop(PackageSpec(path = "."))
+    Pkg.build("cuNumeric")
+'
+
+cp "$CI_PROJECT/LocalPreferences.toml" test/LocalPreferences.toml
+
+julia --color=yes --project="$CI_PROJECT" -e '
+    using Pkg
+    Pkg.test("cuNumeric"; test_args = ["--quickfail"])
+'

.github/workflows/TagBot.yml

@@ -28,3 +28,5 @@ jobs:
       - uses: JuliaRegistries/TagBot@v1
         with:
           token: ${{ secrets.GITHUB_TOKEN }}
+          # SSH deploy key so the tag push can trigger Docs (GITHUB_TOKEN pushes cannot).
+          ssh: ${{ secrets.DOCUMENTER_KEY }}

.github/workflows/docs-tags.yml

@@ -0,0 +1,43 @@
+# Deploy docs for version tags. Kept separate from docs.yml so tag pushes are
+# not subject to the docs/README path filter (a tag tip often only bumps
+# Project.toml).
+name: Docs (tags)
+on:
+  push:
+    tags:
+      - 'v*'
+jobs:
+  docs:
+    name: Documentation
+    permissions:
+      actions: write
+      contents: write
+      pull-requests: read
+      statuses: write
+    runs-on: ubuntu-latest
+    env:
+      LEGATE_AUTO_CONFIG: 0
+    steps:
+      - uses: actions/checkout@v4
+      - uses: julia-actions/setup-julia@v2
+        with:
+          version: '1.11'
+      - uses: julia-actions/cache@v2
+
+      - name: Setup cuNumeric.jl
+        run: |
+          julia --color=yes --project=docs -e '
+              using Pkg
+              Pkg.develop(PackageSpec(path = "lib/CNPreferences"))
+              Pkg.develop(PackageSpec(path = "."))
+          '
+      - name: Instantiate docs environment
+        run: |
+          julia --color=yes --project=docs -e '
+            using Pkg
+            Pkg.instantiate()'
+      - name: Build documentation
+        run: julia --color=yes --project=docs docs/make.jl
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          DOCUMENTER_KEY: ${{ secrets.DOCUMENTER_KEY }}

.github/workflows/docs.yml

@@ -1,44 +1,21 @@
 name: Docs
 on:
   push:
-    paths:
-      - '.github/workflows/docs.yml'
-      - 'src/**'
-      - 'docs/src/**'
-      - 'docs/make.jl'
-      - 'docs/Project.toml'
-      - 'README.md'
-    tags:
-      - 'v*'
+    # After merge: deploy live docs. Feature branches only run via pull_request.
     branches:
       - main
   pull_request:
     paths:
-    - '.github/workflows/docs.yml'
-    - 'src/**'
-    - 'docs/src/**'
-    - 'docs/make.jl'
-    - 'docs/Project.toml'
-    - 'README.md'
+      - '.github/workflows/docs.yml'
+      - '.github/workflows/docs-tags.yml'
+      - 'src/**'
+      - 'docs/**'
+      - 'README.md'
+      - 'lib/CNPreferences/**'
+  workflow_dispatch:
 jobs:
-  check_changes:
-    name: Check for wrapper changes
-    runs-on: ubuntu-latest
-    outputs:
-      wrapper_changed: ${{ steps.filter.outputs.wrapper }}
-    steps:
-      - uses: actions/checkout@v4
-      - uses: dorny/paths-filter@v3
-        id: filter
-        with:
-          filters: |
-            wrapper:
-              - 'lib/cunumeric_jl_wrapper/**'
-
   docs:
-    name : Documentation
-    needs: check_changes
-    if: ${{ github.base_ref == 'main' || needs.check_changes.outputs.wrapper_changed != 'true' }}
+    name: Documentation
     permissions:
       actions: write
       contents: write

.gitignore

@@ -18,8 +18,9 @@ debug
 debug/*
 
 # benchmark outputs
-benchmark/results
+benchmark/results**
 benchmark/results/*
+benchmark/plots**
 
 compile_wrapper.sh
 
@@ -28,8 +29,8 @@ compile_wrapper.sh
 build
 build/*
 build.log
-*.log
-*.err
+*.log**
+*.err**
 kernel.ptx
 *.perf
 *.folded
@@ -94,3 +95,8 @@ node_modules
 *.out
 *.app
 .githash
+
+# VitePress caches/output (keep docs/src/.vitepress/config.mts + theme/)
+docs/.vitepress/
+docs/src/.vitepress/cache/
+docs/src/.vitepress/dist/

Project.toml

@@ -4,8 +4,12 @@ version = "0.1.1"
 
 [deps]
 CNPreferences = "3e078157-ea10-49d5-bf32-908f777cd46f"
+CUDACore = "bd0ed864-bdfe-4181-a5ed-ce625a5fdea2"
+CUDATools = "9ec180c6-1c07-47c7-9e6e-ebefa4d1f6d0"
 CxxWrap = "1f15a43c-97ca-5a2a-ae31-89f07a497df4"
+ExpressionExplorer = "21656369-7473-754a-2065-74616d696c43"
 JuliaFormatter = "98e50ef6-434e-11e9-1051-2b60c6c9e899"
+KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 Legate = "1238f2cf-6593-4d60-9aca-2f5364e49909"
 LegatePreferences = "8028f36a-2b64-49e9-aa04-2d0933fd2ed9"
 Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
@@ -20,17 +24,14 @@ cunumeric_jl_wrapper_jll = "49048992-29d2-5fd1-994f-9cecf112d624"
 cupynumeric_jll = "2862d674-414d-5b0b-a494-b21f8deca547"
 libcxxwrap_julia_jll = "3eaa8342-bff7-56a5-9981-c04077f7cee7"
 
-[weakdeps]
-CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
-
-[extensions]
-CUDAExt = "CUDA"
-
 [compat]
 CNPreferences = "0.1.2"
-CUDA = "5.9"
+CUDACore = "6.0.0"
+CUDATools = "6.0.0"
 CxxWrap = "0.17"
+ExpressionExplorer = "1.1.4"
 JuliaFormatter = "2.3.0"
+KernelAbstractions = "0.9.41"
 Legate = "0.1.2"
 LegatePreferences = "0.1.6"
 MacroTools = "0.5.16"