[XLA:GPU] Switch from specifying maxntid to reqntid.

maxntid specifies the max number of threads in a block, whereas reqntid
says that we will use *exactly* this many threads in a block.

This doesn't have any effect on the benchmarks I ran, but we might as
well do it in case it helps ptxas make a better decision at some point
on some GPU.  At least it will prevent the next person to come along
from doing this same investigation I just did.  :)

PiperOrigin-RevId: 177851116

tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc

@@ -123,10 +123,12 @@ void UpdateLaunchDimensions(const LaunchDimensions& launch_dims, Thunk* thunk,
   llvm::ConstantInt* threads_per_block_ir_value = llvm::ConstantInt::get(
       llvm::IntegerType::get(llvm_context, /*NumBits=*/32),
       launch_dims.threads_per_block());
+  // Our launch bounds are exact, so we can specify them as reqntidx rather than
+  // maxntidx.
   nvvm_annotations_node->addOperand(llvm::MDNode::get(
       llvm_context,
       {llvm::ConstantAsMetadata::get(ir_kernel),
-       llvm::MDString::get(llvm_context, "maxntidx"),
+       llvm::MDString::get(llvm_context, "reqntidx"),
        llvm::ConstantAsMetadata::get(threads_per_block_ir_value)}));
 }
 }  // namespace