[CUDA] Add sm_121/Blackwell to known target

Add an initial NVIDIA GB10 / sm_121 CUDA target description.

The CUDA execution limits are based on local cudaDeviceProp results from
an sm_121 device. Existing NVIDIA MMA ops are reused as a conservative
baseline until Blackwell-specific MMA intrinsics are modeled.

Signed-off-by: Charlie-Tsai1123 <charlie1123tsai@gmail.com>

Author: Charlie-Tsai1123

compiler/src/iree/compiler/Codegen/Dialect/GPU/TargetUtils/KnownTargets.cpp

@@ -913,6 +913,32 @@ StringRef normalizeARMGPUTarget(StringRef target) {
 // cooperative matrix layouts are opaque. We need to create NVIDIA specific WMMA
 // intrinsics if we need to have explicit layout analysis and register mapping.
 
+// Reports initial NVIDIA Blackwell 12.1 target capabilities for GPU target
+// selection. CUDA execution limits are based on sm_121 GB10 device properties.
+const WgpDetails *getBlackwellWgpDetails() {
+  static const MMAIntrinsic mmaOps[] = {
+      MMAIntrinsic::NV_MMA_SYNC_F32_16x8x16_F16,
+      MMAIntrinsic::NV_MMA_SYNC_F16_16x8x16_F16,
+      MMAIntrinsic::NV_MMA_SYNC_F32_16x8x16_BF16,
+      MMAIntrinsic::NV_WMMA_F32_16x16x16_F16,
+      MMAIntrinsic::NV_WMMA_F16_16x16x16_F16,
+  };
+  static const WgpDetails blackwellWgp = {allComputeBits,
+                                          allStorageBits,
+                                          allSubgroupOps,
+                                          allDotProductOps,
+                                          std::size(mmaOps),
+                                          mmaOps,
+                                          0,
+                                          nullptr,
+                                          {32, 32},
+                                          {1024, 1024, 64},
+                                          1024,
+                                          99 * 1024,
+                                          {0x7fffffff, 0xffff, 0xffff}};
+  return &blackwellWgp;
+}
+
 // Reports Ampere-class NVIDIA tensor core capabilities for GPU target
 // selection.
 const WgpDetails *getAmpereWgpDetails() {
@@ -1000,6 +1026,7 @@ const WgpDetails *getPascalWgpDetails() {
 
 // Maps NVIDIA target aliases to the GPU capability model used by codegen.
 std::optional<TargetDetails> getNVIDIAGPUTargetDetails(StringRef target) {
+  const WgpDetails *blackwellWgp = getBlackwellWgpDetails();
   const WgpDetails *ampereWgp = getAmpereWgpDetails();
   const WgpDetails *turingWgp = getTuringWgpDetails();
   const WgpDetails *voltaWgp = getVoltaWgpDetails();
@@ -1038,6 +1065,10 @@ std::optional<TargetDetails> getNVIDIAGPUTargetDetails(StringRef target) {
       .Case("rtx3070ti", TargetDetails{ampereWgp, &rtx3070tiChip})
       // https://www.techpowerup.com/gpu-specs/geforce-rtx-3070.c3674
       .Case("rtx3070", TargetDetails{ampereWgp, &rtx3070Chip})
+      // Initial support for sm_121 / GB10. Other Blackwell compute
+      // capabilities, including sm_120, are intentionally left for follow-up
+      // validation.
+      .Case("sm_121", TargetDetails{blackwellWgp, nullptr})
       .Cases({"ada", "sm_89"}, TargetDetails{ampereWgp, nullptr})
       .Cases({"ampere", "sm_80", "sm_86", "sm_87"},
              TargetDetails{ampereWgp, nullptr})