diff --git a/src/target/llvm/codegen_amdgpu.cc b/src/target/llvm/codegen_amdgpu.cc
index d95f985fe63f..0ab4771f6c69 100644
--- a/src/target/llvm/codegen_amdgpu.cc
+++ b/src/target/llvm/codegen_amdgpu.cc
@@ -187,7 +187,8 @@ class CodeGenAMDGPU : public CodeGenLLVM {
       }
     }
     llvm::Function* f = llvm::Intrinsic::getDeclaration(module_.get(), intrin_id);
-    return builder_->CreateCall(f, {});
+    llvm::Value* result = builder_->CreateCall(f, {});
+    return this->CreateCast(DataType::Int(32), iv->var->dtype, result);
   }
 
   llvm::Value* CreateStorageSync(const CallNode* op) final {
diff --git a/src/target/llvm/intrin_rule_rocm.cc b/src/target/llvm/intrin_rule_rocm.cc
index 17baaf3e657a..d25126f5d828 100644
--- a/src/target/llvm/intrin_rule_rocm.cc
+++ b/src/target/llvm/intrin_rule_rocm.cc
@@ -89,7 +89,7 @@ inline PrimExpr DispatchShuffle(const PrimExpr& e) {
     index = self + delta;
     index = Select((self & (width - 1)) + delta >= width, self, index);
   }
-  PrimExpr res = Call(var.dtype(), builtin::call_pure_extern(),
+  PrimExpr res = Call(DataType::Int(32), builtin::call_pure_extern(),
                       {StringImm("llvm.amdgcn.ds.bpermute"), index << 2, var});
   return res;
 }