CUDA: Fixes for NVRTC 12.x and warp mask ambiguity; adds CC 8.x warp reduction intrinsics. (#3314)

* CUDA: Fixes for NVRTC 12.x, warp mask ambiguity; add reduction partial specializations.

* Fixes running NVRTC on CUDA 12 without a specified profile (used in testing, e.g. `slang-test -api cuda -category wave`)
* Fixes mask ambiguity between getting the lane index from threadId.x and a full mask of threads.
* Adds partial specializations for compute capability 8.x warp reduction intrinsics.

* Fix formatting

1 files changed, 11 insertions, 5 deletions

diff --git a/source/compiler-core/slang-nvrtc-compiler.cpp b/source/compiler-core/slang-nvrtc-compiler.cpp
index daa392120..b84b1a403 100644
--- a/source/compiler-core/slang-nvrtc-compiler.cpp
+++ b/source/compiler-core/slang-nvrtc-compiler.cpp
@@ -820,15 +820,21 @@ SlangResult NVRTCDownstreamCompiler::compile(const DownstreamCompileOptions& inO
 
     {
         // The lowest supported CUDA architecture version supported
-        // by NVRTC is `compute_30`.
+        // by any version of NVRTC we support is `compute_30`.
         //
         SemanticVersion version(3);
 
-        // Newer releases of NVRTC only support `compute_35` and up
-        // (with everything before `compute_52` being deprecated).
-        //
-        if( m_desc.version.m_major >= 11 )
+        // Newer releases of NVRTC only support newer CUDA architectures.
+        if ( m_desc.version.m_major >= 12 )
+        {
+            // NVRTC in CUDA 12 only supports `compute_50` and up
+            // (with everything before `compute_52` being deprecated).
+            version = SemanticVersion(5, 0);
+        }
+        else if ( m_desc.version.m_major == 11 )
         {
+            // NVRTC in CUDA 11 only supports `compute_35` and up
+            // (with everything before `compute_52` being deprecated).
             version = SemanticVersion(3, 5);
         }