shader-slang / slang-rhi / 26593948464

push

github

vulkan: gate NVX and ray tracing extensions behind opt-out flags (#760)

* vulkan: gate NVX and ray tracing extensions behind opt-out flags

Add two new DeviceDesc flags, both default true to preserve historical
behavior:

* enableCUDALaunchFromGfx gates VK_NVX_binary_import and
  VK_NVX_image_view_handle, which expose vkCmdCuLaunchKernelNVX
  (launching CUDA-style kernels from inside a Vulkan command buffer).
  Enabling these on the NVIDIA Linux driver brings up a dual-purpose
  CUDA primary context that has been observed to poison cuDNN's MHA
  dispatcher on Blackwell sm_120 + driver 595.x: torch's
  _scaled_dot_product_cudnn_attention starts returning
  mha_graph.execute(...).is_good() == false on the first attention call.

* enableRayTracing gates VK_KHR_acceleration_structure and dependent
  RT extensions (ray_tracing_pipeline, ray_query, position_fetch,
  NV_ray_tracing_linear_swept_spheres, NV_cluster_acceleration_structure).
  Enabling these initializes a CUDA-adjacent BVH/RT execution path that
  exhibits the same cuDNN MHA poisoning on the same hardware.

Applications that don't use vkCmdCuLaunchKernelNVX or ray tracing can
opt out by setting these flags to false. Default-true preserves existing
behavior for all other consumers.

Confirmed on RTX PRO 6000 Blackwell + Ubuntu 26.04 + driver 595.71.05
via a minimal repro that creates a Vulkan device with subsets of
slang-rhi's extension set; the bug reproduces only when NVX_binary_import
or acceleration_structure are present in ppEnabledExtensionNames.

* Updated comments

* Address review: simplify docs and drop NVX_image_view_handle gate

12069 of 20476 branches covered (58.94%)

Branch coverage included in aggregate %.

2 of 2 new or added lines in 1 file covered. (100.0%)

35667 of 49080 relevant lines covered (72.67%)

278182.46 hits per line