[NVPTX] Add support for PTX 8.6 and CUDA 12.6 (12.8) (#123398)

Add CUDA versions 12.7, 12.8, 12.9 which support PTX8.6+ (enables using Blackwell-specific instructions).

Author: sergey-kozub

clang/include/clang/Basic/BuiltinsNVPTX.def

@@ -28,7 +28,9 @@
 #pragma push_macro("SM_90")
 #pragma push_macro("SM_90a")
 #pragma push_macro("SM_100")
-#define SM_100 "sm_100"
+#pragma push_macro("SM_100a")
+#define SM_100a "sm_100a"
+#define SM_100 "sm_100|" SM_100a
 #define SM_90a "sm_90a"
 #define SM_90 "sm_90|" SM_90a "|" SM_100
 #define SM_89 "sm_89|" SM_90
@@ -1091,6 +1093,7 @@ TARGET_BUILTIN(__nvvm_getctarank_shared_cluster, "iv*3", "", AND(SM_90,PTX78))
 #pragma pop_macro("SM_90")
 #pragma pop_macro("SM_90a")
 #pragma pop_macro("SM_100")
+#pragma pop_macro("SM_100a")
 #pragma pop_macro("PTX42")
 #pragma pop_macro("PTX60")
 #pragma pop_macro("PTX61")

clang/include/clang/Basic/Cuda.h

@@ -44,9 +44,12 @@ enum class CudaVersion {
   CUDA_124,
   CUDA_125,
   CUDA_126,
+  CUDA_127,
+  CUDA_128,
+  CUDA_129,
   FULLY_SUPPORTED = CUDA_123,
   PARTIALLY_SUPPORTED =
-      CUDA_126, // Partially supported. Proceed with a warning.
+      CUDA_129, // Partially supported. Proceed with a warning.
   NEW = 10000,  // Too new. Issue a warning, but allow using it.
 };
 const char *CudaVersionToString(CudaVersion V);
@@ -80,6 +83,7 @@ enum class OffloadArch {
   SM_90,
   SM_90a,
   SM_100,
+  SM_100a,
   GFX600,
   GFX601,
   GFX602,

clang/lib/Basic/Cuda.cpp

@@ -44,6 +44,9 @@ static const CudaVersionMapEntry CudaNameVersionMap[] = {
     CUDA_ENTRY(12, 4),
     CUDA_ENTRY(12, 5),
     CUDA_ENTRY(12, 6),
+    CUDA_ENTRY(12, 7),
+    CUDA_ENTRY(12, 8),
+    CUDA_ENTRY(12, 9),
     {"", CudaVersion::NEW, llvm::VersionTuple(std::numeric_limits<int>::max())},
     {"unknown", CudaVersion::UNKNOWN, {}} // End of list tombstone.
 };
@@ -98,6 +101,7 @@ static const OffloadArchToStringMap arch_names[] = {
     SM(90),                          // Hopper
     SM(90a),                         // Hopper
     SM(100),                         // Blackwell
+    SM(100a),                        // Blackwell
     GFX(600),  // gfx600
     GFX(601),  // gfx601
     GFX(602),  // gfx602
@@ -227,8 +231,8 @@ CudaVersion MinVersionForOffloadArch(OffloadArch A) {
   case OffloadArch::SM_90a:
     return CudaVersion::CUDA_120;
   case OffloadArch::SM_100:
-    return CudaVersion::NEW; // TODO: use specific CUDA version once it's
-                             // public.
+  case OffloadArch::SM_100a:
+    return CudaVersion::CUDA_127;
   default:
     llvm_unreachable("invalid enum");
   }

clang/lib/Basic/Targets/NVPTX.cpp

@@ -285,13 +285,16 @@ void NVPTXTargetInfo::getTargetDefines(const LangOptions &Opts,
       case OffloadArch::SM_90a:
         return "900";
       case OffloadArch::SM_100:
+      case OffloadArch::SM_100a:
         return "1000";
       }
       llvm_unreachable("unhandled OffloadArch");
     }();
     Builder.defineMacro("__CUDA_ARCH__", CUDAArchCode);
     if (GPU == OffloadArch::SM_90a)
       Builder.defineMacro("__CUDA_ARCH_FEAT_SM90_ALL", "1");
+    if (GPU == OffloadArch::SM_100a)
+      Builder.defineMacro("__CUDA_ARCH_FEAT_SM100_ALL", "1");
   }
 }
 

clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp

@@ -2277,6 +2277,7 @@ void CGOpenMPRuntimeGPU::processRequiresDirective(const OMPRequiresDecl *D) {
       case OffloadArch::SM_90:
       case OffloadArch::SM_90a:
       case OffloadArch::SM_100:
+      case OffloadArch::SM_100a:
       case OffloadArch::GFX600:
       case OffloadArch::GFX601:
       case OffloadArch::GFX602:

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -89,6 +89,12 @@ CudaVersion getCudaVersion(uint32_t raw_version) {
     return CudaVersion::CUDA_125;
   if (raw_version < 12070)
     return CudaVersion::CUDA_126;
+  if (raw_version < 12080)
+    return CudaVersion::CUDA_127;
+  if (raw_version < 12090)
+    return CudaVersion::CUDA_128;
+  if (raw_version < 12100)
+    return CudaVersion::CUDA_129;
   return CudaVersion::NEW;
 }
 
@@ -682,6 +688,9 @@ void NVPTX::getNVPTXTargetFeatures(const Driver &D, const llvm::Triple &Triple,
   case CudaVersion::CUDA_##CUDA_VER:                                           \
     PtxFeature = "+ptx" #PTX_VER;                                              \
     break;
+    CASE_CUDA_VERSION(129, 87);
+    CASE_CUDA_VERSION(128, 87);
+    CASE_CUDA_VERSION(127, 86);
     CASE_CUDA_VERSION(126, 85);
     CASE_CUDA_VERSION(125, 85);
     CASE_CUDA_VERSION(124, 84);

clang/test/Misc/target-invalid-cpu-note/nvptx.c

@@ -27,6 +27,7 @@
 // CHECK-SAME: {{^}}, sm_90
 // CHECK-SAME: {{^}}, sm_90a
 // CHECK-SAME: {{^}}, sm_100
+// CHECK-SAME: {{^}}, sm_100a
 // CHECK-SAME: {{^}}, gfx600
 // CHECK-SAME: {{^}}, gfx601
 // CHECK-SAME: {{^}}, gfx602

llvm/lib/Target/NVPTX/NVPTX.td

@@ -39,6 +39,7 @@ foreach sm = [20, 21, 30, 32, 35, 37, 50, 52, 53,
   def SM#sm: FeatureSM<""#sm, !mul(sm, 10)>;
 
 def SM90a: FeatureSM<"90a", 901>;
+def SM100a: FeatureSM<"100a", 1001>;
 
 foreach version = [32, 40, 41, 42, 43, 50, 60, 61, 62, 63, 64, 65,
                    70, 71, 72, 73, 74, 75, 76, 77, 78,
@@ -74,6 +75,7 @@ def : Proc<"sm_89", [SM89, PTX78]>;
 def : Proc<"sm_90", [SM90, PTX78]>;
 def : Proc<"sm_90a", [SM90a, PTX80]>;
 def : Proc<"sm_100", [SM100, PTX86]>;
+def : Proc<"sm_100a", [SM100a, PTX86]>;
 
 def NVPTXInstrInfo : InstrInfo {
 }