windows: disabled cuda on ci to avoid incompatible driver issue. Fixing a bad test mem allocation

.github/workflows/cmake-windows.yml

@@ -75,51 +75,6 @@ jobs:
     - name: Build
       run: cmake --build ${{github.workspace}}/build --config ${{env.BUILD_TYPE}} --target ALL_BUILD
 
-    # Diagnostic for test_util3d_features hanging on the CUDA build.
-    # Compares startup time with/without the GPU visible and lists the DLL
-    # graph against a reference test (test_localgrid) that runs fine.
-    # Non-fatal: keeps going even if a step times out so the timings still
-    # show up in the log.
-    - name: Diagnose test_util3d_features startup (CUDA only)
-      if: matrix.build_name == 'windows-2022-cuda'
-      shell: pwsh
-      working-directory: ${{github.workspace}}/build
-      continue-on-error: true
-      env:
-        PYTHONHOME: ${{env.VCPKG_EXPORT_PATH}}/installed/x64-windows-release/tools/python3
-        PYTHONNOUSERSITE: 1
-      run: |
-        function Measure-Listing($exe, $envOverride) {
-          $job = Start-Job -ScriptBlock {
-            param($exe, $envOverride)
-            if ($envOverride) { $env:CUDA_VISIBLE_DEVICES = $envOverride }
-            $sw = [System.Diagnostics.Stopwatch]::StartNew()
-            & $exe --gtest_list_tests *> $null
-            $sw.Stop()
-            [pscustomobject]@{ Seconds = $sw.Elapsed.TotalSeconds; ExitCode = $LASTEXITCODE }
-          } -ArgumentList $exe, $envOverride
-          if (Wait-Job $job -Timeout 120) {
-            $r = Receive-Job $job
-            "{0,7:F2}s exit={1}" -f $r.Seconds, $r.ExitCode
-          } else {
-            Stop-Job $job; Remove-Job $job -Force
-            ">120s (timed out)"
-          }
-        }
-        $exe = "${{github.workspace}}\build\bin\test_util3d_features.exe"
-        $ref = "${{github.workspace}}\build\bin\test_localgrid.exe"
-        Write-Host "=== --gtest_list_tests timing ==="
-        Write-Host ("test_util3d_features (GPU visible)    : " + (Measure-Listing $exe $null))
-        Write-Host ("test_util3d_features (CUDA hidden)    : " + (Measure-Listing $exe ""))
-        Write-Host ("test_localgrid       (GPU visible)    : " + (Measure-Listing $ref $null))
-        Write-Host ("test_localgrid       (CUDA hidden)    : " + (Measure-Listing $ref ""))
-        Write-Host ""
-        Write-Host "=== DLL graph delta (util3d_features minus localgrid) ==="
-        $u = & dumpbin /dependents $exe | Select-String -Pattern '\.dll' | ForEach-Object { $_.ToString().Trim().ToLower() } | Sort-Object -Unique
-        $l = & dumpbin /dependents $ref | Select-String -Pattern '\.dll' | ForEach-Object { $_.ToString().Trim().ToLower() } | Sort-Object -Unique
-        $onlyU = Compare-Object $u $l | Where-Object { $_.SideIndicator -eq '<=' } | ForEach-Object { $_.InputObject }
-        if ($onlyU) { $onlyU | ForEach-Object { Write-Host "  +$_" } } else { Write-Host "  (identical)" }
-
     - name: Test
       working-directory: ${{github.workspace}}/build
       # PYTHONHOME points the embedded interpreter at vcpkg's bundled
@@ -135,6 +90,13 @@ jobs:
         PYTHONHOME: ${{env.VCPKG_EXPORT_PATH}}/installed/x64-windows-release/tools/python3
         PYTHONNOUSERSITE: 1
       run: |
+        # The Windows CUDA runner ships a CUDA driver older than the runtime
+        # OpenCV/cv::cuda was built against. Lazy CUDA-init on first cv::cuda
+        # symbol call stalls some unit tests (test_util3d_features) for >300s.
+        # Hide the GPU so the runtime short-circuits before any blocking call.
+        if ('${{ matrix.build_name }}' -eq 'windows-2022-cuda') {
+          $env:CUDA_VISIBLE_DEVICES = ""
+        }
         ctest -C ${{env.BUILD_TYPE}} --output-on-failure --timeout 300
         ctest -C ${{env.BUILD_TYPE}} --output-on-failure --timeout 300 --rerun-failed
 

corelib/test/test_util3d.cpp

@@ -1280,9 +1280,10 @@ TEST(Util3dTest, GetMinMax3D) {
 
 // Test with a laser scan that contains only 2D data (no Z values)
 TEST(Util3dTest, GetMinMax3DCVPoint3f2DData) {
-    // Create a 2x3 CV matrix of type CV_32FC2 (2D points)
+    // Create a 1x2 CV matrix of type CV_32FC2 (2D points)
     cv::Mat laserScan(1, 2, CV_32FC2);
-    float data[6] = {1.0f, 2.0f, 4.0f, 5.0f};  // No Z values
+    // 2 points * 2 channels = 4 floats.
+    float data[4] = {1.0f, 2.0f, 4.0f, 5.0f};  // No Z values
     memcpy(laserScan.data, data, sizeof(data));
 
     cv::Point3f min, max;