[Runtime] CutensorMap support (#18097)

This PR introduces Cutensor map support in the runtime module. It enables calling kernels whose arguments are cuTensorMap, these arguments are passed as handle(address) and associated with arg_extra_tags that indicate indicate it is tensor map. The TensorMap is allocated on stack with a runtime API

Author: spectrometerHBH

include/tvm/ir/type.h

@@ -330,5 +330,34 @@ class FuncType : public Type {
   TVM_DEFINE_OBJECT_REF_METHODS(FuncType, Type, FuncTypeNode);
 };
 
+/*!
+ * \brief The type of tensor map.
+ * \sa TensorMapType
+ */
+class TensorMapTypeNode : public TypeNode {
+ public:
+  void VisitAttrs(AttrVisitor* v) { v->Visit("span", &span); }
+
+  bool SEqualReduce(const TensorMapTypeNode* other, SEqualReducer equal) const {
+    return equal(span, other->span);
+  }
+
+  void SHashReduce(SHashReducer hash_reduce) const { hash_reduce(span); }
+
+  static constexpr const char* _type_key = "TensorMapType";
+  TVM_DECLARE_FINAL_OBJECT_INFO(TensorMapTypeNode, TypeNode);
+};
+
+/*!
+ * \brief Managed reference to TensorMapTypeNode.
+ * \sa TensorMapTypeNode
+ */
+class TensorMapType : public Type {
+ public:
+  TVM_DLL TensorMapType(Span span = Span());
+
+  TVM_DEFINE_OBJECT_REF_METHODS_WITHOUT_DEFAULT_CONSTRUCTOR(TensorMapType, Type, TensorMapTypeNode);
+};
+
 }  // namespace tvm
 #endif  // TVM_IR_TYPE_H_

include/tvm/script/ir_builder/tir/ir.h

@@ -452,6 +452,8 @@ inline Var Handle(runtime::DataType dtype = runtime::DataType::Void(),
   return is_size_var ? tvm::tir::SizeVar("", type_annotation) : tvm::tir::Var("", type_annotation);
 }
 
+inline Var TensormapHandle() { return tvm::tir::Var("", PointerType(TensorMapType())); }
+
 #define TVM_TIR_IR_BUILDER_DEF_DTYPE_CAST(FuncName, DType)                                     \
   inline PrimExpr FuncName(Optional<PrimExpr> expr = std::nullopt, bool is_size_var = false) { \
     DataType dtype = DType;                                                                    \

python/tvm/ir/type.py

@@ -107,3 +107,19 @@ def __init__(self, arg_types, ret_type):
             arg_types,
             ret_type,
         )
+
+
+@tvm.ffi.register_object("TensorMapType")
+class TensorMapType(Type):
+    """TensorMapType used in the low-level TIR.
+
+    Parameters
+    ----------
+    span : tvm.ir.Span
+        The span information.
+    """
+
+    def __init__(self, span=None):
+        self.__init_handle_by_constructor__(
+            _ffi_api.TensorMapType, span  # pylint: disable=no-member
+        )

python/tvm/script/ir_builder/tir/ir.py

@@ -1566,6 +1566,8 @@ def handle(
     res : PrimExpr
         The new tir.Var with type handle or casted expression with type handle.
     """
+    if dtype == "tensormap":
+        return _ffi_api.TensormapHandle()  # type: ignore[attr-defined] # pylint: disable=no-member
     is_unknown_type = dtype is None
     if dtype is None:
         dtype = "void"

src/ir/type.cc

@@ -89,4 +89,16 @@ TVM_FFI_REGISTER_GLOBAL("ir.TupleType").set_body_typed([](Array<Type> fields) {
   return TupleType(fields);
 });
 
+TVM_FFI_REGISTER_GLOBAL("ir.TensorMapType").set_body_typed([](Span span) {
+  return TensorMapType(span);
+});
+
+TensorMapType::TensorMapType(Span span) {
+  ObjectPtr<TensorMapTypeNode> n = make_object<TensorMapTypeNode>();
+  n->span = std::move(span);
+  data_ = std::move(n);
+}
+
+TVM_REGISTER_NODE_TYPE(TensorMapTypeNode);
+
 }  // namespace tvm

src/runtime/cuda/cuda_device_api.cc

@@ -357,5 +357,211 @@ TVM_DLL int GetCudaDeviceCount() {
 
 TVM_FFI_REGISTER_GLOBAL("runtime.GetCudaDeviceCount").set_body_typed(GetCudaDeviceCount);
 
+/**
+ * \brief FFI wrapper for cuTensorMapEncodeTiled.
+ *
+ * This function registers a global function `runtime.cuTensorMapEncodeTiled` that can be
+ * called from other parts of the TVM runtime (e.g., Python). It wraps the CUDA Driver API
+ * function `cuTensorMapEncodeTiled`, which initializes a tensor map descriptor (CUtensorMap).
+ *
+ * \param tensor_map (handle): A `void*` pointer to the CUtensorMap object to be initialized.
+ * \param tensor_dtype (DataType): The TVM data type of the tensor.
+ * \param tensor_rank (int): The rank (number of dimensions) of the tensor.
+ * \param tensor_ptr (handle): A `void*` pointer to the start of the tensor in global memory.
+ * \param global_shape (int...): `tensor_rank` integer arguments for the global tensor dimensions.
+ * \param global_strides (int...): `tensor_rank - 1` integer arguments for the global tensor
+ * strides. The stride for the innermost dimension is not provided as it's assumed to be contiguous.
+ * \param shared_shape (int...): `tensor_rank` integer arguments for the shape of the tile (box)
+ * in shared memory.
+ * \param shared_strides (int...): `tensor_rank` integer arguments for the strides of the tile (box)
+ * in shared memory.
+ * \param interleaved_kind (int): An integer corresponding to the CUtensorMapInterleave enum.
+ * \param swizzle_kind (int): An integer corresponding to the CUtensorMapSwizzle enum.
+ * \param l2_promotion_kind (int): An integer corresponding to the CUtensorMapL2promotion enum.
+ * \param oob_fill_kind (int): An integer corresponding to the CUtensorMapFloatOOBfill enum.
+ */
+TVM_FFI_REGISTER_GLOBAL("runtime.cuTensorMapEncodeTiled")
+    .set_body_packed([](ffi::PackedArgs args, ffi::Any* rv) {
+      CHECK_GE(args.size(), 4) << "init_cuTensorMap expects at least 4 arguments";
+      size_t arg_cnt = 0;
+      CUtensorMap* tensor_map = static_cast<CUtensorMap*>(args[arg_cnt++].cast<void*>());
+      runtime::DataType tensor_dtype = args[arg_cnt++].cast<runtime::DataType>();
+      uint32_t tensor_rank = static_cast<uint32_t>(args[arg_cnt++].cast<int32_t>());
+      void* tensor_ptr = static_cast<void*>(args[arg_cnt++].cast<void*>());
+
+      CHECK_EQ(args.size(), 4 + tensor_rank * 4 + 3)
+          << "cuTensorMapEncodeTiled expects " << 4 + tensor_rank * 4 + 3 << " arguments"
+          << "tensor_map, tensor_dtype, tensor_rank, tensor_ptr, global_shape(" << tensor_rank
+          << "), global_strides(" << tensor_rank - 1 << "), shared_shape(" << tensor_rank
+          << "), shared_strides(" << tensor_rank << "), interleaved_kind, swizzle_kind"
+          << ", l2_promotion_kind, oob_fill_kind";
+
+      std::vector<cuuint64_t> global_shape(tensor_rank);
+      std::vector<cuuint64_t> global_strides(tensor_rank);
+      std::vector<uint32_t> shared_shape(tensor_rank);
+      std::vector<uint32_t> shared_strides(tensor_rank);
+      for (size_t i = 0; i < tensor_rank; ++i) {
+        global_shape[i] = static_cast<cuuint64_t>(args[arg_cnt++].cast<int64_t>());
+      }
+      for (size_t i = 0; i < tensor_rank - 1; ++i) {
+        global_strides[i] = static_cast<cuuint64_t>(args[arg_cnt++].cast<int64_t>());
+        CHECK_EQ(global_strides[i] % 16, 0) << "global strides must be multiple of 16";
+      }
+      for (size_t i = 0; i < tensor_rank; ++i) {
+        shared_shape[i] = static_cast<uint32_t>(args[arg_cnt++].cast<int32_t>());
+        CHECK_GE(shared_shape[i], 0) << "boxDim must be non-negative";
+        CHECK_LE(shared_shape[i], 256) << "boxDim must be less than or equal to 256";
+      }
+      for (size_t i = 0; i < tensor_rank; ++i) {
+        shared_strides[i] = static_cast<uint32_t>(args[arg_cnt++].cast<int32_t>());
+      }
+      auto interleaved_kind = static_cast<CUtensorMapInterleave>(args[arg_cnt++].cast<int>());
+      auto swizzle_kind = static_cast<CUtensorMapSwizzle>(args[arg_cnt++].cast<int>());
+      auto l2_promotion_kind = static_cast<CUtensorMapL2promotion>(args[arg_cnt++].cast<int>());
+      auto oob_fill_kind = static_cast<CUtensorMapFloatOOBfill>(args[arg_cnt++].cast<int>());
+
+      ICHECK_EQ(tensor_dtype.lanes(), 1)
+          << "Expect tensor_dtype to have lanes=1, but get " << tensor_dtype;
+      CUtensorMapDataType cu_dtype;
+      switch (tensor_dtype.code()) {
+        case DataType::kInt:
+          // int
+          switch (tensor_dtype.bits()) {
+            case 8:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_UINT8;
+              break;
+            case 32:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_INT32;
+              break;
+            case 64:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_INT64;
+              break;
+            default:
+              LOG(FATAL) << "Unsupported data type " << runtime::DLDataTypeToString(tensor_dtype);
+          }
+          break;
+        case DataType::kUInt:
+          // unsigned int
+          switch (tensor_dtype.bits()) {
+            case 8:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_UINT8;
+              break;
+            case 16:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_UINT16;
+              break;
+            case 32:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_UINT32;
+              break;
+            case 64:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_UINT64;
+              break;
+            default:
+              LOG(FATAL) << "Unsupported data type " << runtime::DLDataTypeToString(tensor_dtype);
+          }
+          break;
+        case DataType::kFloat:
+          // float
+          switch (tensor_dtype.bits()) {
+            case 16:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_FLOAT16;
+              break;
+            case 32:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_FLOAT32;
+              break;
+            case 64:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_FLOAT64;
+              break;
+            default:
+              LOG(FATAL) << "Unsupported data type " << runtime::DLDataTypeToString(tensor_dtype);
+          }
+          break;
+        case DataType::kBFloat:
+          // bfloat
+          switch (tensor_dtype.bits()) {
+            case 16:
+              cu_dtype = CU_TENSOR_MAP_DATA_TYPE_BFLOAT16;
+              break;
+            default:
+              LOG(FATAL) << "Unsupported data type " << runtime::DLDataTypeToString(tensor_dtype);
+          }
+          break;
+        case DataType::kFloat8_e4m3fn:
+          // NV float8 e4m3
+          cu_dtype = CU_TENSOR_MAP_DATA_TYPE_UINT8;
+          break;
+        case DataType::kFloat8_e5m2:
+          // NV float8 e5m2
+          cu_dtype = CU_TENSOR_MAP_DATA_TYPE_UINT8;
+          break;
+        default:
+          LOG(FATAL) << "Unsupported data type " << runtime::DLDataTypeToString(tensor_dtype);
+      }
+
+      // sanity checks per cuTensorMapEncodeTiled requirements
+      // see
+      // https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__TENSOR__MEMORY.html#group__CUDA__TENSOR__MEMORY_1ga7c7d2aaac9e49294304e755e6f341d7
+      CHECK_EQ((reinterpret_cast<uint64_t>(tensor_ptr) & 0b1111), 0);    // 16-byte alignment
+      CHECK_EQ((reinterpret_cast<uint64_t>(tensor_map) & 0b111111), 0);  // 64-byte alignment
+      CHECK_LE(tensor_rank, 5) << "cuTensorMapEncodeTiled only supports up to 5D tensors";
+
+      if (swizzle_kind == CU_TENSOR_MAP_SWIZZLE_32B) {
+        CHECK_LE(shared_shape[0] * tensor_dtype.bytes(), 32)
+            << "CU_TENSOR_MAP_SWIZZLE_32B implies the bounding box inner dimension will be <= 32.";
+      } else if (swizzle_kind == CU_TENSOR_MAP_SWIZZLE_64B) {
+        CHECK_LE(shared_shape[0] * tensor_dtype.bytes(), 64)
+            << "CU_TENSOR_MAP_SWIZZLE_64B implies the bounding box inner dimension will be <= 64.";
+      } else if (swizzle_kind == CU_TENSOR_MAP_SWIZZLE_128B) {
+        CHECK_LE(shared_shape[0] * tensor_dtype.bytes(), 128)
+            << "CU_TENSOR_MAP_SWIZZLE_128B implies the bounding box inner dimension will be <= "
+               "128.";
+      }
+
+      const cuuint64_t* global_shape_ptr = global_shape.data();
+      const cuuint64_t* global_strides_ptr = global_strides.data();
+      const uint32_t* shared_shape_ptr = shared_shape.data();
+      const uint32_t* shared_strides_ptr = shared_strides.data();
+
+      CUresult res =
+          cuTensorMapEncodeTiled(tensor_map, cu_dtype, tensor_rank, tensor_ptr, global_shape_ptr,
+                                 global_strides_ptr, shared_shape_ptr, shared_strides_ptr,
+                                 interleaved_kind, swizzle_kind, l2_promotion_kind, oob_fill_kind);
+      const char* errstr;
+      cuGetErrorString(res, &errstr);
+      if (res != CUDA_SUCCESS) {
+        // get error string
+        const char* error_string = nullptr;
+        cuGetErrorString(res, &error_string);
+        std::cerr << "Error in cuTensorMapEncodeTiled: " << error_string << std::endl;
+        std::cout << "cu_dtype: " << cu_dtype << "\n";
+        std::cout << "TMA Desc Addr:   " << tensor_map << "\n";
+        std::cout << "TMA Interleave:  " << interleaved_kind << "\n";
+        std::cout << "TMA L2Promotion: " << l2_promotion_kind << "\n";
+        std::cout << "TMA OOBFill:     " << oob_fill_kind << "\n";
+        std::cout << "SMEM Swizzle:    " << swizzle_kind << "\n";
+        std::cout << "tensor rank: " << tensor_rank << "\n";
+        std::cout << "global prob shape: ";
+        for (size_t i = 0; i < tensor_rank; i++) {
+          std::cout << global_shape[i] << " ";
+        }
+        std::cout << "\n";
+        std::cout << "global prob stride: ";
+        for (size_t i = 0; i < tensor_rank; i++) {
+          std::cout << global_strides[i] << " ";
+        }
+        std::cout << "\n";
+        std::cout << "smem box shape: ";
+        for (size_t i = 0; i < tensor_rank; i++) {
+          std::cout << shared_shape[i] << " ";
+        }
+        std::cout << "\n";
+        std::cout << "smem box stride: ";
+        for (size_t i = 0; i < tensor_rank; i++) {
+          std::cout << shared_strides[i] << " ";
+        }
+        std::cout << "\n";
+        CHECK_EQ(res, CUDA_SUCCESS) << "Error in cuTensorMapEncodeTiled: " << errstr;
+      }
+    });
+
 }  // namespace runtime
 }  // namespace tvm

src/runtime/cuda/cuda_module.cc

@@ -266,7 +266,7 @@ ffi::Function CUDAModuleNode::GetFunction(const String& name,
   const FunctionInfo& info = it->second;
   CUDAWrappedFunc f;
   f.Init(this, sptr_to_self, name, info.arg_types.size(), info.launch_param_tags);
-  return PackFuncVoidAddr(f, info.arg_types);
+  return PackFuncVoidAddr(f, info.arg_types, info.arg_extra_tags);
 }
 
 Module CUDAModuleCreate(std::string data, std::string fmt,

src/runtime/file_utils.cc

@@ -45,6 +45,11 @@ void FunctionInfo::Save(dmlc::JSONWriter* writer) const {
   writer->WriteObjectKeyValue("name", name);
   writer->WriteObjectKeyValue("arg_types", sarg_types);
   writer->WriteObjectKeyValue("launch_param_tags", launch_param_tags);
+  std::vector<int> iarg_extra_tags(arg_extra_tags.size());
+  for (size_t i = 0; i < arg_extra_tags.size(); ++i) {
+    iarg_extra_tags[i] = static_cast<int>(arg_extra_tags[i]);
+  }
+  writer->WriteObjectKeyValue("arg_extra_tags", iarg_extra_tags);
   writer->EndObject();
 }
 
@@ -56,6 +61,12 @@ void FunctionInfo::Load(dmlc::JSONReader* reader) {
   helper.DeclareOptionalField("launch_param_tags", &launch_param_tags);
   helper.DeclareOptionalField("thread_axis_tags",
                               &launch_param_tags);  // for backward compatibility
+  std::vector<int> iarg_extra_tags;
+  helper.DeclareOptionalField("arg_extra_tags", &iarg_extra_tags);
+  arg_extra_tags.resize(iarg_extra_tags.size());
+  for (size_t i = 0; i < arg_extra_tags.size(); ++i) {
+    arg_extra_tags[i] = static_cast<ArgExtraTags>(iarg_extra_tags[i]);
+  }
   helper.ReadAllFields(reader);
   arg_types.resize(sarg_types.size());
   for (size_t i = 0; i < arg_types.size(); ++i) {
@@ -67,12 +78,14 @@ void FunctionInfo::Save(dmlc::Stream* writer) const {
   writer->Write(name);
   writer->Write(arg_types);
   writer->Write(launch_param_tags);
+  writer->Write(arg_extra_tags);
 }
 
 bool FunctionInfo::Load(dmlc::Stream* reader) {
   if (!reader->Read(&name)) return false;
   if (!reader->Read(&arg_types)) return false;
   if (!reader->Read(&launch_param_tags)) return false;
+  if (!reader->Read(&arg_extra_tags)) return false;
   return true;
 }
 

src/runtime/meta_data.h

@@ -59,6 +59,9 @@ struct FunctionInfo {
   std::vector<DLDataType> arg_types;
   std::vector<std::string> launch_param_tags;
 
+  enum class ArgExtraTags : int { kNone = 0, kTensorMap = 1 };
+  std::vector<ArgExtraTags> arg_extra_tags;
+
   void Save(dmlc::JSONWriter* writer) const;
   void Load(dmlc::JSONReader* reader);
   void Save(dmlc::Stream* writer) const;