Fix Gaussian splat Vulkan render target setup (#7525)

Author: jeevan6996

CHANGELOG.md

@@ -5,6 +5,7 @@
 -   Add vcpkg support for easier dependency management (PR #7386)
 -   Exposed advanced parameters (`full_depth`, `samples_per_node`, `point_weight`) for Poisson surface reconstruction in `TriangleMesh.create_from_point_cloud_poisson` (PR #7430) (issue #7248)
 -   Add SYCL tensor backends for HashMap, nearest-neighbor search (KNN, fixed-radius, hybrid), geometry transforms, and registration / odometry / feature pipelines (parity with CUDA paths where applicable).
+-   Fix Gaussian Splat rendering on Vulkan by disabling view MSAA before binding a render target with sampleable depth (issue #7495).
 -   Add SYCL ML ops: continuous convolution and sparse convolution (forward/transpose/backprop-filter, sycl-tla GEMM-accelerated), voxelize, voxel pooling, invert neighbors list, reduce-subarrays-sum, and ragged-to-dense, plus contrib NMS, RoI pool, ball query, IoU (BEV + 3D), trilinear devoxelize, and three-interpolate, all exposed via the PyTorch XPU op library.
 -   Hybrid nearest-neighbor search on SYCL now supports the L1 and Linf metrics (previously L2-only), matching fixed-radius search and CUDA.
 -   Convolution ops' `allow_tf32` acceleration on SYCL requires channel counts and leading dimensions to be divisible by 4; falls back to IEEE float32 with a one-time warning otherwise. See `docs/sycl.rst`.

cpp/open3d/visualization/rendering/gaussian_splat/GaussianSplatVulkanBackend.cpp

@@ -179,12 +179,14 @@ class GaussianSplatVulkanBackend final : public GaussianSplatRenderer::Backend {
             targets.render_target =
                     resource_mgr.CreateRenderTarget(view_color, targets.depth);
         }
-        view.SetRenderTarget(targets.render_target);
-
         auto* native = view.GetNativeView();
         auto msaa = native->getMultiSampleAntiAliasingOptions();
         msaa.enabled = false;
         native->setMultiSampleAntiAliasingOptions(msaa);
+
+        // Filament rejects sampleable depth targets while MSAA is enabled, so
+        // update the view before binding the shared depth render target.
+        view.SetRenderTarget(targets.render_target);
         view.SetPostProcessing(false);
 
         return static_cast<bool>(targets.render_target);