Merge branch 'tilereduce' of https://github.com/arhik/cuTile.jl into tilereduce

Author: arhik

.github/workflows/CI.yml

@@ -0,0 +1,55 @@
+name: CI
+
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+
+jobs:
+  test:
+    name: Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }}
+    runs-on: ${{ matrix.runner }}
+    timeout-minutes: 15
+    strategy:
+      fail-fast: false
+      matrix:
+        version:
+          - '1.11'
+          - '1.12'
+        os:
+          - Linux
+          - Windows
+          - macOS
+        arch:
+          - x64
+          - aarch64
+        exclude:
+          - os: Windows
+            arch: aarch64
+          - os: macOS
+            arch: x64
+        include:
+          - os: Linux
+            arch: x64
+            runner: ubuntu-latest
+          - os: Linux
+            arch: aarch64
+            runner: ubuntu-24.04-arm
+          - os: Windows
+            arch: x64
+            runner: windows-latest
+          - os: macOS
+            arch: aarch64
+            runner: macos-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: julia-actions/setup-julia@v2
+        with:
+          version: ${{ matrix.version }}
+          arch: ${{ matrix.arch }}
+      - uses: julia-actions/cache@v2
+      - uses: julia-actions/julia-buildpkg@v1
+      - uses: julia-actions/julia-runtest@v1
+        with:
+          test_args: '--quickfail'

Project.toml

@@ -18,6 +18,7 @@ IRStructurizer = {path = "IRStructurizer"}
 CUDAExt = "CUDA"
 
 [compat]
+julia = "1.11"
 CUDA_Compiler_jll = "0.4"
 CUDA_Tile_jll = "13.1"
 

src/language/operations.jl

@@ -235,7 +235,7 @@ Indices are 1-indexed. Index tiles are broadcast to a common shape.
     stride1 = broadcast_to(stride1_0d, S)
 
     # Compute linear index = idx0 * stride0 + idx1 * stride1
-    linear_idx = idx0_i32 * stride0 + idx1_i32 * stride1
+    linear_idx = idx0_i32 .* stride0 + idx1_i32 .* stride1
 
     # Compute pointer tile
     ptr_tile = Intrinsics.offset(array.ptr, linear_idx)
@@ -317,7 +317,7 @@ Indices are 1-indexed. Index tiles and value tile must broadcast to same shape.
     stride1 = broadcast_to(stride1_0d, S)
 
     # Compute linear index = idx0 * stride0 + idx1 * stride1
-    linear_idx = idx0_i32 * stride0 + idx1_i32 * stride1
+    linear_idx = idx0_i32 .* stride0 + idx1_i32 .* stride1
 
     # Compute pointer tile
     ptr_tile = Intrinsics.offset(array.ptr, linear_idx)

test/Project.toml

@@ -1,4 +1,5 @@
 [deps]
+cuTile = "0dea8319-8c4a-4662-a73d-20234d115b9a"
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 FFTW = "7a1cc6ca-52ef-59f5-83cd-3a7055c09341"
 FileCheck = "4e644321-382b-4b05-b0b6-5d23c3d944fb"